doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.ppat.1003366 | Cis and Trans Acting Factors Involved in Human Cytomegalovirus Experimental and Natural Latent Infection of CD14 (+) Monocytes and CD34 (+) Cells | The parameters involved in human cytomegalovirus (HCMV) latent infection in CD14 (+) and CD34 (+) cells remain poorly identified. Using next generation sequencing we deduced the transcriptome of HCMV latently infected CD14 (+) and CD34 (+) cells in experimental as well as natural latency settings. The gene expression profile from natural infection in HCMV seropositive donors closely matched experimental latency models, and included two long non-coding RNAs (lncRNAs), RNA4.9 and RNA2.7 as well as the mRNAs encoding replication factors UL84 and UL44. Chromatin immunoprecipitation assays on experimentally infected CD14 (+) monocytes followed by next generation sequencing (ChIP-Seq) were employed to demonstrate both UL84 and UL44 proteins interacted with the latent viral genome and overlapped at 5 of the 8 loci identified. RNA4.9 interacts with components of the polycomb repression complex (PRC) as well as with the MIE promoter region where the enrichment of the repressive H3K27me3 mark suggests that this lncRNA represses transcription. Formaldehyde Assisted Isolation of Regulatory Elements (FAIRE), which identifies nucleosome-depleted viral DNA, was used to confirm that latent mRNAs were associated with actively transcribed, FAIRE analysis also showed that the terminal repeat (TR) region of the latent viral genome is depleted of nucleosomes suggesting that this region may contain an element mediating viral genome maintenance. ChIP assays show that the viral TR region interacts with factors associated with the pre replication complex and a plasmid subclone containing the HCMV TR element persisted in latently infected CD14 (+) monocytes, strongly suggesting that the TR region mediates viral chromosome maintenance.
| Human cytomegalovirus (HCMV) is a ubiquitous herpesvirus where infection is usually subclinical. HCMV initial infection is followed by the establishment of latency in CD34 (+) myeloid cells and CD14 (+) monocytes. Primary infection or reactivation from latency can be associated with significant morbidity and mortality can occur in immune compromised patients. Latency is marked by the persistence of the viral genome, lack of production of infectious virus and the expression of only a few previously recognized latency associated transcripts. Despite the significant interest in HCMV latent infection, little is known regarding the mechanism involved in establishment or maintenance of the viral chromosome. We have now identified the transacting factors present in latently infected CD14 (+) monocytes and CD34 (+) progenitor cells as well as identification of a region of the HCMV genome, the terminal repeat locus that mediates viral DNA maintenance. This is a major step toward understanding the mechanism of HCMV latent infection.
| Human cytomegalovirus (HCMV) is a ubiquitous herpesvirus that infects 60–90% of the population and is usually subclinical, however virus infection can cause severe disease and mortality in immune compromised patients [1]. Disease manifestations include retinitis, pneumonia and hepatitis [2].
HCMV lytic phase of infection is typically studied in cell culture using human fibroblasts (HFs) and viral encoded genes are expressed in a temporally regulated manner. HCMV lytic DNA replication requires cis and trans acting factors and results in the production of infectious virus [3]–[9]. Recent high-resolution transcriptome mapping during a lytic HCMV infection, revealed a complex pattern of transcription [10]. This analysis also showed that during lytic infection most viral RNA production is concentrated in four long non-coding RNAs (lncRNAs), RNA2.7 (also known as ß2.7), RNA1.2, RNA4.9, and RNA5.0 [10]. The high expression level of viral encoded lncRNAs suggests that that these transcripts may be significant factors for regulating viral and cellular processes required for efficient viral replication.
Herpesvirus latency is defined as the persistence of the viral genome in the absence of production of infectious virus. Certain properties of latency have emerged from study of the gamma herpesviruses including maintenance of the viral chromosome is as a circular episome [11], which is controlled by virus-encoded proteins that interact with viral and host cell chromatin [12]–[23]. With respect to maintenance and replication of the HCMV DNA episome, previous studies have quantified the number of genomes during experimental and natural infection [24], [25]. However the cis acting element required for maintenance of the viral genome is unknown. Lifelong HCMV latency is established in myeloid lineage, from bone marrow-derived CD34 (+) progenitors through peripheral blood to CD14 (+) monocytes [26]–[33]. Latently infected cells contain HCMV DNA without supporting lytic replication although virus can be reactivated and recovered through differentiation [34]–[40]. The regulation and maintenance of latency is poorly understood although reactivation of latent virus is a major source of virus associated with serious disease and mortality in immunocompromised hosts, hence a better understanding of HCMV latency may lead to treatments to resolve latent HCMV genomes from infected cells. Differentiation of CD34 (+) or CD14 (+) cells, to macrophages or dendritic cells through the use of various cytokines results in reactivation and subsequent production of infectious virus [41]–[45]. Several in vitro experimental systems have been developed to study HCMV latency. These systems use either cultured CD14 (+) monocytes or CD34 (+) hematopoietic stem cells. CD14 (+) or CD34 (+) cells cultured in vitro can be infected with HCMV clinical strains, resulting in latent infection and efficient reactivation of latent virus [26], [37], [43], [46], [47].
The mechanism of HCMV latent infection is unknown and the mechanism involved in the establishment and maintenance of the latent virus genome has not been addressed in either natural or experimental models of latency. In experimental models, as well as during natural latency in CD34 (+) and CD14 (+) cells, a limited number of latency specific HCMV transcripts have been identified [41], [42], [48], [49]. Three consistently identified viral genes expressed during latency are UL138, a TNF modulator, UL111A, a variant cmvIL-10 cytokine and the UL81–82 antisense transcript encoding LUNA [41], [50], [51]. HCMV UL138 is expressed during, though dispensable for HCMV lytic replication such that UL138 mutant viruses altered latency [41], [52], [53]. This may result from the fact that UL138 has been shown to upregulate TNFR1 and sensitizes infected cells to TNFα [54], [55]. UL138 has been shown to interact with proteins encoded by nearby genes within the UL133–UL138 locus [52]. HCMV latency UL111A transcript encodes a 139 aa protein, which is a homolog to cellular IL-10 [51], [56]. Recombinant viruses that lack the ability to express UL111A are capable of establishing latency and efficiently reactivate, however latently infected cells express higher levels of surface MHC class II in the absence of UL111A expression and suggests that UL111A may function to inhibit the recognition of latently infected cells by CD4 (+) T cells [57]. The expression of LUNA during lytic infection is regulated by IE72 expression and the interaction with hDaxx [58]. LUNA has recently been implicated as having a role in virus reactivation [59].
To elucidate the factors involved in HCMV latency, we infected CD14 (+) monocytes or CD34 (+) progenitor cells with an HCMV clinical isolate [37], [46]. We have now used these experimental latency protocols, coupled with next generation sequencing, to reveal the complete high-resolution HCMV transcriptome (RNA-Seq) during early and latent infections. RNA-Seq analysis shows that during HCMV experimental and natural latency in CD14 (+) monocytes, viral transcripts encoding UL44, UL84, UL95, UL87, UL52, UL50, LUNA, UL138 and the lncRNAs 2.7 and 4.9 were detected. For latently infected CD34 (+) cells, all of the mRNAs detected in latently infected CD14 (+) cells were present, however additional transcripts encoding UL28/29, UL37/38, UL114, UL133/135 and US17 were also detected. Using chromatin isolation by RNA purification (ChIRP), lncRNA4.9 was shown to physically interact with the HCMV major immediate early promoter region and results in an enrichment of the repressive H3K27me3 mark at the MIEP during latency suggesting that HCMV latent genomes are silenced by PRC2 interaction.
To address the HCMV cis requirements for latency, we show that a plasmid containing the terminal repeat (TR) element persisted in latently infected cells, strongly suggesting that this element mediates viral genome maintenance.
Cord blood was received from the Colorado Cord Blood Bank (Univ. of Colorado). Pooled peripheral whole blood for natural infection studies was obtained from Renown Medical Center (Reno, NV) and processed with in 5 hours to isolate CD14 (+) or CD34 (+) cells. All protocols to obtain blood products were approved by IRB and Office of Human Research Protection. The pooled blood samples represented approximately 25 individual HCMV seropositive donors. Cells were isolated using human cord blood CD34 positive selection kit (Stemcell technologies) according to manufacturer's instructions. Briefly, samples were incubated with a pre-enrichment cocktail containing antibodies directed against CD66b and glycophorin A. This was step was performed for negative selection of granulocytes and erythrocytes. CD34 selected cells were retained from the non-selected cells by the use of the EasySep magnet. Cells were resuspended in culturing media or for natural infection studies immediately processed to extract total RNA. Human CD14 (+) were isolated using positive selection MACS bead and LS columns (Miltenyi Biotec) according to manufacturer's instructions and cells were resuspended in culturing media or immediately processed to extract total RNA.
Purity of isolated CD14 (+) and CD34 (+) was assessed by flow cytometry using antibodies for human CD14-FITC or CD34-FITC (Miltenyi Biotec), along with human CD45-Pacific Blue (BioLegend) for total cell staining. Isotype controls included IgG2a-FITC (Miltenyi Biotec) and IgG1-Pacific Blue (BioLegend). Approximately 0.5×106 cells were stained for 30 minutes at 4°C with the fluorescently labeled antibody, washed once and resuspended in 1× PBS with 0.5%FBS, 2 mM EDTA and 1% methanol-free formaldehyde. Cells were determined to be free from red blood cells and >95% CD34 (+) or CD14 (+).
HCMV natural infection was evaluated from pooled blood from 25 seropositive individual donors. From 120 ml of pooled blood approximately 175,000 CD34 (+) cells and 32×106 CD14 (+) were isolated, total RNA was extraction and DNase treated, resulting in 300 ng of RNA for CD34 (+) and 28.5 mg of RNA for CD14 (+). The RNA was used generate a sequencing library as described below.
CD34 (+) cells were cultured in IMDM supplemented with 10% BIT serum substitute (StemCell Technologies), 2 mM L-glutamine, 20 ng/ml low-density lipoprotein (Sigma Aldrich), 50 mM 2-mercaptoethanol, 10 ng/ml Stem Cell factor, 10 ng/ml IL-3, 10 ng/ml G-CSF (R & D Systems) [26], [41], [60] or cultured in X-Vivo 15 (lonza) [27]. Media was refreshed every three days until latency had been established and verified. To reactive latent virus, CD34 (+) cells were stimulated to proliferate and differentiate with the addition of 10 ng/ml GM-CSF, 10 ng/ml Flt-3 ligand, 10 ng/ml TPO, and 10 ng/ml TNF for three days, followed by the addition of 50 ng/ml lipopolysaccharide (LPS) (Sigma-Aldrich) for an additional four days.
Human CD14 (+) cells purchased (Lonza and ReachBio) or isolated from cord blood were maintained in Iscove DMEM (Hyclone) supplemented with 20% heat-inactivated FBS (Atlanta Biologicals), 50 ng/mL M-CSF, 50 ng/mL stem cell factor (SCF), 50 ng/mL G-CSF, 50 ng/mL GM-CSF, 50 ng/mL IL-3 (R&D Systems) at a density of 1×106 cells/mL on low cell-binding plates (Nunc Hydrocell). Medium was replaced every 3 days. BAC-derived FIX strains were propagated in human foreskin fibroblasts (HF) cells. After 12–14 days post infection cells were scraped and subjected to a freeze-thaw to release virus from the cells. Virus titer was determined with standard plaque assay on HF cells. Unless otherwise stated, infections of CD14 (+) cells were done at a multiplicity of 5 pfu/cell. Cells were incubated with virus for 1 hr. Cells were then washed twice with Hanks Balanced Salt solution (HBSS).
HCMV infected CD14 (+) or CD34 (+) cells were maintained and monitored for HCMV gene expression by qPCR analysis. Latency was determined to be established when no IE2 gene expression was detected by qPCR, detection of the virus genome and expression of LUNA and UL138. For reactivation of latent virus, CD14 (+) cells were differentiated by adherence to plastic tissue culture dishes supplemented with 100 ng/mL IL-6 (R&D Systems) at 16–18 days post infection. The reactivation from latency in either CD14 (+) or CD34 (+) was monitored for IE2 gene expression using qPCR.
For UV inactivation, 120 mls of 5×106 pfu/ml of FIX BAC virus was evenly distributed in a thin layer onto a 150 mm tissue culture dish. The virus was irradiated on ice in a Stratalinker 2400 (Stratagene) for 4 min at 9.9×105 mJ. UV inactivation was confirmed by the absence of immediate-early (IE) gene expression in infected human foreskin fibroblasts using qPCR.
Human foreskin fibroblast cells, maintained in DMEM supplemented with 10% FBS, were plated in a 6-well tissue culture plate at 0.1×106 cells per well. 24 hours after plating half of the media was removed and replaced with reactivated CD34 (+) cells (approximately 0.5×106) along with its' media. Cells were co-cultured together for 10–12 days and monitored by the appearance of green plaque formation in the HF cells.
Total RNA from experimentally or naturally infected (as determined above) CD14 (+) or CD34 (+) cells was isolated using PureLink RNA mini kit (Life Technology) followed by removal of genomic DNA using Turbo DNA-free (Life Technology). Poly-A RNA was enriched from total RNA by Dynabeads oligo (dT)25 (Life Technology) according to manufacturers instructions. The resulting Poly-A RNA was used in dUTP based NEXTflex Directional RNA-Seq Kit with Illumina compatible adaptors (Bioo Scientific) according to manufacturers instructions. The resulting libraries were verified on a Bioanalyzer High sensitivity DNA chip (Agilent) and quantified with real-time PCR using Illumina compatible kit and standards (KAPA). Final libraries were sequenced using an Illumina MiSeq instrument. HCMV transcript discovery and alignment was performed using CLC Genomics Workbench software and strand specific RNA-Seq parameters. Transcripts were aligned to the Fix strain (VR1814) reference genome.
Tiling of RNA 4.9 with biotin labeled DNA probes retrieves specific RNA 4.9 bound proteins and DNA sequences. Original protocol is from Chu et al [61]. All probes were biotinylated at the 3′ end with an 18-carbon spacer arm; probes were designed against RNA 4.9 full-length sequence using an online designer at http://www.singlemoleculefish.com, and synthesized at Protein and Nucleic Acid Facility (Standford University). Samples were processed as described previously [62]. Eluted DNA was resuspended in 50 µl of water and used for end point PCR. Primers for the PCR include MIEP-1 forward: GTGTTTGTCCGAAATACGCG, reverse: GCCTCATATCGTCTGTCACC; MIEP-2 forward: GTTACATAACTTACGGTAAATGGCC, reverse: CCAAAACCGCATCACCATG; MIEP-3 forward: GATTTCCAAGTCTCCACCCC, reverse: GCGGTACTTACGTCACTCTTG; MIEP-4 forward: CCCCGCTTCCTTATGCTATAG, reverse: AAGAACCCATGTCCGGAAC; MIEP-5 forward: CTCCTTGCTCCTAACAGTGG, reverse: GTACTGCTCAGACTACACTGC; UL19 forward: CCTGTATGAGCTGTTTCGACG, reverse: GACTCACATCTAGCTCGTCTTC. ChIRP PCR primers are shown in Table S2 in Text S1.
For CD14 (+) and CD34 (+) culturing media was changed every three days in order to replenish nutrients and cytokines necessary to maintain the cell in culture. The supernatant that was removed from the cells was collected and used for quantitative real-time PCR to obtain relative values for the amount of virus being produced. The supernatant was subjected to two low speed spins, 400×g for 10 minutes in a table top centrifuge to remove any residual cells. For the real-time PCR reaction 5 µl of supernatant was used in a total reaction volume of 20 µl using Taqman Universal 2× master mix (Life Technology) and 20× primer-probe (IDT). Relative Ct values of viral supernatant DNA were correlated to standard curve from a known concentration of purified FIX-BAC DNA.
Cell viability, density and growth was monitored by using trypan blue exclusion test. Briefly, cells were gently resuspended and a small aliquot was removed. A 1∶1 dilution of cell suspension and 0.4% Trypan Blue solution (MP Biomedicals) was counted in a hemacytometer to determine and monitor total number of live cells.
Evaluation of changes in histone modifications was performed as previously described [62]. Fold-enrichment of histone marks at various genomic loci was calculated as IgG-subtracted %Input of the locus normalized by the IgG-subtracted %Input of the reference gene GAPDH. 3 separate experiments were performed. PCR primers are listed in Table S4 in Text S4.
5×106 latently infected CD14 (+) cells were transfected with 2.5 µg each of JMJD3-HA and UTX-HA expression plasmids (Addgene), or 5 milligrams of GFP-control plasmid, and a non-transfected control group. The cells were transfected using Nucleofector device and Amaxa Human Monocyte Nucleofector kit (Lonza) according to manufactures instructions. Transfection efficiency was monitored by GFP expression. 72 hours post transfection RNA and protein was harvested using RNA/DNA/Protein purification kit (Norgen BioTek Corp.). Total RNA was subjected to removal of genomic DNA using Turbo DNA-free (Life Technology) according to manufacturers instructions. The purified RNA was then used for cDNA synthesis as previously described. The resulting cDNA was quantified using real-time PCR and Taqman primers targeting specific HCMV transcripts.
Expression of UTX and JMJD3 was confirmed by Western blot where total cell protein extracts were resolved by SDS-PAGE gel that was subsequently transferred to a polyvinylidene difluoride (PVDF) membrane, blocked with 5% nonfat dry milk powder in 1× TBST buffer and reacted with antibodies specific for HA-tag (Sigma-Aldrich). After one hour incubation with the primary antibody the membrane was washed in 1× TBST three times 5 minutes each. The secondary antibody donkey anti-mouse IgG conjugated to Alexa Fluor-680 (Life Technology) was diluted in blocking buffer and added to the membrane for 30 minutes as which time it was again washed with 1× TBST three times 5 minutes each. Specific proteins bands were visualized using the Odyssey by LI-COR.
Total RNA was isolated from cells using PureLink RNA mini kit (Life Technologies), followed by removal of genomic DNA using Turbo DNA-free (Life Technologies). cDNA was synthesized from 1 µg of total RNA in the presence of random hexamers, dNTPs, and Superscript III reverse transcriptase (Life Technologies). The resulting cDNA was then used along with Taqman Universal PCR Master Mix (Life Technologies) and specific primers and FAM labeled probes (IDT) in an Eppendorf RealPlex. The following real-time PCR program was used: one cycle 95°C hot start for 5 minutes, and forty cycles of 95°C for 15 seconds and 60°C for 1 minute. Primers used for detection of specific gene expression are shown in Table S3 in Text S1.
4–5×106 latently infected CD14 (+) monocytes were cross-linked by the addition of formaldehyde to a final concentration of 1%, incubated at room temperature for 10 minutes and quenched with 0.125 M glycine. After washing with ice cold PBS, cells were lysed in 1 ml/5×106 cells ice-cold lysis buffer (0.5% NP-40, 150 mM NaCl, 50 mM Tris, pH 7.4, 1 mM EDTA, and protease inhibitors). Lysed cells were dounced followed by collection of the crude nuclear extract by centrifugation. The nuclear pellet was resuspended in 1 ml of RIPA (50 mM Tris, pH 7.4, 150 mM NaCl, 1% Triton-X, 0.1% SDS, 0.5% sodium deoxycholate, 1 mM EDTA) and sonicated with a Fisher Scientific Sonic Dismembrator and micro-tip at 40% amplitude for 40 cycles of 20 s on followed by 20 s off in a wet ice bath. The sonicated chromatin was collected by centrifugation at 20,000×g for 15 min at 4°C to remove cellular debris. Chromatin shearing to 150–200 bp fragments was confirmed by agarose gel electrophoresis. 100 ul of chromatin was reserved for input library preparation and the remainder was pre-cleared for 1 hour at 4°C with 100 ul normal mouse IgG sepharose beads followed by incubation overnight with 20 ug UL84 or 100 ug UL44 mAb. 200 ul of Active Motif magnetic IgG coated beads were prepared by blocking overnight with 5 mg/ml BSA and 200 ug/ml yeast tRNA (Ambion) in PBS. Beads were washed once with 5 mg/ml BSA in PBS and added to immunoprecipitated chromatin. Immunoprecipitation was performed for 4–5 hours at 4°C. The beads and immunoprecipitated complexes were washed at room temperature with rotation twice for 1 min with RIPA, five times for 5 minutes each with LiCl2 buffer (500 mM LiCl2, 100 mM Tris pH 7.4, 1% NP-40, 1% sodium deoxycholate), and one brief wash with TE (10 mM Tris, pH 8.1, 1 mM EDTA). DNA was eluted from the beads by the addition of 200 ul elution buffer (1% SDS, 0.1 M NaHCO3) and reversed cross-linked overnight at 65°C. Reserved input DNA was also reverse cross-linked and treated the same as ChIP samples from this point forward. Reversed cross-linked UL84 ChIP DNA or UL44 Chip DNA and respective Input DNA were purified with a Qiagen min-Elute kit. Purified DNA was quantified using an Invitrogen Qubit Flourometer and sequencing libraries were created from 10 ng of ChIP or Input DNA with a Bioo Scientific NEXTflex ChIP-Seq library preparation kit (#5143-01). Library integrity was analyzed with an Agilent 2100 Bioanalyzer and quantified using a KAPA q-PCR kit. 12 pM each of UL84 ChIP or UL44 ChIP library and 6 pM of respective Input library were sequenced on an Illumina MiSeq. ChIP-Seq data analysis, including read mapping to FIX genome (VR1814) and peak calling was performed using CLC Genomics Workbench with a maximum false discovery (FDR) rate of 5%, a window size of 250 bp, and reads were shifted based on a length of 200 bp. The data is the result of four separate experiments.
10×106 CD14 (+) or CD34 (+) cells were infected with FIX virus and cells were monitored for IE gene expression using qPCR. Once latency had been established, approximately 18 dpi, total RNA was isolated from cells using PureLink RNA mini kit (Life Technologies), followed by removal of genomic DNA using Turbo DNA-free (Life Technologies). The resulting RNA was used with Qiagen OneStep RT-PCR kit with primers designed to mRNA shown in Table S3 in Text S3.
20×106 CD14 (+) cells were infected with FIX virus. 18 days post infection cells were harvested and processed as described previously [62]. RNA-protein complexes were precipitated by adding 300 µl lysate, 5 µl antibody, UL84, UL44, SUZ12 (Active motif, cat. # 39357), EZH2 (Active motif, cat. # 39875) or GAPDH (Abcam, ab128915), 25 µl Protein G magnetic beads (Active Motif), and 1 µl RNase Out. Primers used for PCR are shown in Table S1 in Text S1.
CD14 (+) monocytes were infected with BAC-derived FIX virus at a multiplicity of 5 pfu/cell. Cells were maintained and monitored for HCMV gene products by qPCR analysis. Latency was determined to be established when no IE gene expression was detected by qPCR analysis. After established latency pGEM7zf(−), pGEM-oriLyt, or pGEM-TR plasmids (4 ug) were added to 3×106 CD14 (+) latent or control non-infected cells using Nucleofector 2b device and Human Monocyte Nucleofector kit (Lonza) according to manufacture's instructions for program Y-001 on the device. Transfected cells were immediately transferred to 12-well low cell-binding plates (Nunc Hydrocell) containing 37°C prewarmed media and placed back into 37°C/5% CO2 incubator. Cells were evaluated for viability using trypan blue and for transfection efficiency using a spike of a plasmid expressing EGFP. Cells were determined to be 80% viable and a transfection efficiency of over 70% was achieved. For analysis of transfection input, 24 hrs post transfection total DNA from 2×105 cells was harvested using Norgen DNA/RNA/Protein purification kit according to manufacture's instructions. DNA was eluted from the column in 100 mls of DNA elution buffer, 50 mls was then used for EcoRI (New England Biolabs) restriction digest. The total sample of EcoRI digested DNA was loaded onto a 0.8% 1× TAE (40 mM Tris-HCl, 20 mM Acetic Acid, 1 mM EDTA) agarose gel and electrophoresed in 1× TAE buffer until the dye front reached the bottom of the gel. For analysis of plasmid maintenance, approximately 2 million latently infected and transfected (15 dpt) CD14 (+) cells/well were prepared, loaded, and electrophoresed on Gardella gels as previously described [63]. DNA was transferred to Zeta-Probe membrane (Bio-Rad) by the alkaline transfer method according to the manufacturer's instructions and hybridized to using a 32P-labled pGEM probe. Membranes were hybridized with 5 ng of probe in 10 ml of hybridization buffer (1.5× SSPE, 1 mM EDTA, 7% sodium dodecyl sulfate [SDS], 10% [wt/vol] polyethylene glycol) for 16 hrs at 65°C in a hybridization oven (Robbins Scientific). Post-hybridization washed were performed with 2× SSC and 0.1% SDS twice for 15 minutes each at 65°C, and then with 0.1× SSC and 0.1% SDS twice for 30 minutes at 65°C. Southern blot were exposed and visualized using the Storm Scanner imaging system (GE Healthcare).
FAIRE was performed as previously described [64]. Briefly, 3×106 CD14 (+) monocytes infected with HCMV strain FIX at 4 dpi or 18 dpi were cross-linked with formaldehyde and lysed, followed by sonication to shear chromatin. 10% of the sheared cross-linked chromatin was reserved for input and the remainder was extracted twice with phenol∶chloroform to remove regions of DNA associated with nucleosomes. The aqueous layer containing open regions of the chromatin corresponding to nucleosome depletion was reserved, reverse cross-linked, and further purified by ethanol precipitation. Input DNA was also reverse cross-linked and purified by ethanol precipitation. Sequencing library preparation was performed using an Illumina Tru-Seq DNA library kit. Input and FAIRE libraries were sequenced by the University of California-Irvine High Throughput Genomics Facility on an Illumina Hi-Seq 2000. Approximately 60 million reads per library were generated. Read alignment with FIX strain (VR1814) was done using CLC Genomics Workbench.
PCR primers were designed to specifically amplify the terminal repeat region using FIX BAC DNA as a template and PrimeSTAR GXL DNA polymerase (Takara). After PCR was completed the product was precipitated with 1/10 volume 3 M NaAcetate and 2.5 X volume 100% ethanol. The DNA pellet was resuspended in 100 mls of water and 100 mls of 2× Easy-A Master mix (Stratagene) was added. The DNA was incubated at 70°C for 30 minutes and then precipitated with 1/10 volume 3 M NaAcetate and 2.5 X volume 100% ethanol. The DNA pellet was resuspended in 20 µl of water and 3 µl was used in pGEM T-easy (Life Technology) cloning reaction according to manufacturers instructions. The resulting clones were subjected to restriction enzyme digestion and DNA sequencing to determine clones with the correct insert.
Each ChIP was performed with chromatin from approximately 200,000 latently infected CD14 (+) monocytes. DNA-protein complexes were immunoprecipitated using CDT1 (Abcam, ab70829), MCM3 (Abcam, ab4460), or an isotype control antibody. PCR for TR region was performed using PrimeSTAR GXL DNA Polymerase (Takara). Primers to detect the TR region were: Set #1 Forward 5′-ACA CCT CCG ACG TCC ACT ATA TAC CA-3′/Reverse 5′-CGT CCA CAC ACG CAA CTC CAA TTT-3′ or Set #2 Forward 5′- GGT GAC GTC GGA GAC AGG G-3′/Reverse 5′- TGC TGT GGT GTA AGG GTA AGG TGT-3′.
Genomic DNA was isolated from approximately 2 million mock infected, 18 dpi, and 20 dpi FIX BAC latent CD14 (+) monocytes using a Norgen DNA/RNA/Protein purification kit. Primers amplifying a region within the TR or IR were Forward: 5′-aacgacagacgaagtacggcacaa-3′ and Reverse 5′-acaaacaccgcagaactccttgacg-3′. Takara PrimeSTAR GXL polymerase was used for PCR.
Although previous studies investigated viral gene expression in HCMV latently infected CD14 (+) and primary hematopoietic progenitor cells [26], [41], [42], [46], [48], [49], [65], [66], new technologies now allow for the high-resolution evaluation of the HCMV latent virus transcriptome. While there is little or no understanding with respect to the mechanism involved to maintain or replicate the virus genome in a latent environment, we speculate that viral encoded factor(s) play a significant role in viral DNA maintenance. This assumption is solidly based on other herpesvirus systems, the most prominent being the gamma herpesviruses where at least one virus encoded factor is required for viral genome maintenance and replication [14], [15], [67]–[71]. Hence, the first step in identifying likely candidates involved in viral genome maintenance/replication is to identify all viral encoded transcripts present during latent infection.
CD14 (+) monocytes were cultured as previously described using specific growth media and culture plates that did not allow attachment/differentiation of cells [46]. Undifferentiated CD14 (+) monocytes were infected with the FIX BAC clinical isolate virus strain and viral immediate early (IE1/IE2), LUNA and UL138 transcript levels were measured at various days post infection using qPCR (Fig. 1). Immediate early mRNA was detected at days 2, 4, 7 and 11 days post infection (Fig. 1). Immediate early gene expression was no longer detected by 14 days post infection (Fig. 1) suggesting that HCMV virus genomes were not producing immediate early proteins and lytic replication had concluded. LUNA and UL138 transcripts were still detected, indicating that the virus entered the latent phase (Fig. 1). HCMV genomic DNA was measured from infected cells at 5 and 18 days post infection by qPCR. The HCMV genome was present in latently infected cells at approximately 4 copies per cell at 18 days post infection. Viral DNA was also detected using traditional PCR at 5 and 18 days post infection (Fig. 1, inset). We also measured viral genome copy number over a 26-day period of infection (Fig. 1B). Viral genomes were present at approximately 4 copies per cell during latent infection at 16 days post infection (Fig. 1B). Cells were monitored during infection and analyzed to ensure the absence of macrophage or dendritic cell markers (CD11c+, CD11b+, CD141+, CD303+, CD11b+, CD68+, CD80+).
For RNA-Seq, total cellular RNA from latently infected CD14 (+) monocytes was extracted at 5 and 18 days post infection. Total cellular RNA was also extracted from HCMV latently infected CD14 (+) monocytes, after an 18-day incubation subsequent to treatment with IL-6 and re-plating of cells on a surface that allowed attachment. This lytic virus reactivated sample, along with the 5 and 18 day samples were subjected to next generation sequencing (RNA-Seq). RNA-Seq was performed using an Illumina HiSeq 2000 instrument (54 million reads per sample, paired end directional sequencing). Data was analyzed using CLC Bio Genomics Workbench software (RNA-Seq analysis) using mock-infected RNA as a control, 4 independent biological replicates were used for analysis. Figure 2 is a “peak map” showing the location and relative abundance of transcripts from the HCMV FIX BAC viral genome identified by RNA-Seq. The height of the peaks represents the relative number of reads for the transcripts detected and correlates to the relative abundance of the mRNA in cells. At 5 days post infection almost all (99%) of the HCMV ORFs present within the viral genome were expressed (Fig. 2A). Interestingly, and partly consistent with what was previously reported from an RNA-Seq analysis of HCMV lytic infection, the highest amount of transcript accumulation after a 5-day infection of CD14 (+) monocytes occurred from the expression of UL22A, lncRNAs 2.7 and 4.9 loci (Fig. 2A). UL22A is expressed in HCMV infected dendritic cells during lytic infection but has not been previously described in early infection of CD14 (+) monocytes. UL22A is responsible for immune evasion and selectively blocks CCL5 [72]. Immediate early transcripts encoding IE1/2 (UL122–123) were also detected as well as all early transcripts encoding replication proteins (UL44, UL70, UL105, UL102, UL54 and UL57) (Fig. 2A). Late transcripts encoding viral glycoproteins and capsid proteins were also present at 5 days post infection.
At 18 days post infection of CD14 (+) monocytes the transcriptome analysis showed that only a subset of viral-encoded RNAs were present in latently infected cells and confirmed that neither immediate early IE2 or IE1 mRNAs were expressed, nor were any other immediate early transcripts detected (Fig. 2B). Transcripts were detected encoding LUNA, UL95, UL138, UL87, UL84, UL52, UL50 and UL44 along with transcripts for the long noncoding RNAs RNA2.7 and RNA4.9 (Fig. 2B). Very low levels of UL111A mRNA were detected, which cannot be seen on the peak map shown (cutoff was 20 reads). A list of all transcripts detected during latent infection is shown in table 1 along with the number of reads from the RNA-Seq analysis. These data are the first reporting of the total virus encoded transcriptome from latently infected CD14 (+) monocytes using next generation sequencing. To confirm the presence of RNA transcripts we performed reverse transcriptase PCR (RT-PCR) using RNA isolated from 5 and 18 day infected CD14 (+) monocytes and primers specific for each of the transcripts detected as well as control transcripts. All transcripts identified from RNA-Seq analysis were detected using RT-PCR, whereas we were unable to detect IE2, UL13 or US21 mRNAs (Fig. 2B, Inset figure), confirming the results from the next generation sequencing.
Lytic reactivation of latent virus was also performed by incubating latently infected (18 days post infection) CD14 (+) monocytes with IL-6 and re-plating in tissue culture flasks that allowed for cell attachment. After a 7-day incubation with IL-6, total cellular RNA was harvested and subjected to next generation sequencing. RNA-Seq analysis showed that HCMV gene expression was observed from almost the entire viral genome consistent with lytic virus replication (Fig. 2C). Together these data show that HCMV latent infection in CD14 (+) monocytes is characterized by the absence of transcripts encoding immediate early proteins and the expression of a several specific transcripts, some of which were normally associated with lytic DNA replication and expression of two lncRNAs.
To test for the production of virus, supernatants were collected at 18 days post infection from reactivated samples and measured the amount of HCMV DNA by qPCR. In reactivated cells, supernatant virus was detected indicating that lytic virus infection was efficiently reactivated in these cells (Fig. 3).
For CD34 (+) cells, we evaluated mRNA expression at 3 days post infection as well as at a time point when IE2 transcripts were no longer detected. RNA was harvested at 3 and 10 days post infection and qPCR was performed. We measured the transcript abundance of IE2, UL44, UL84, RNA 2.7 and RNA4.9. At 10 days post infection of CD34 (+) cells, no IE2 mRNA was detected, however transcripts for UL44, UL84 and the two lncRNAs were detected (Fig. 4A). Based on the results of the qPCR showing that latency associated genes were expressed in the absence of detectable IE2, we harvested total cellular RNA and performed next generation sequencing. The results of the RNA-Seq are shown in figure 4B. For HCMV latent infection of CD34 (+) cells we harvested RNA at 11 days post infection and the RNA-Seq peak map shows the presence of the same transcripts detected in latently infected CD14 (+) cells (Fig. 4A, 11 day latent). Interestingly, several additional transcripts were detected. RNA-Seq of latently infected CD34 (+) cells shows that the mRNAs for UL28/29, UL37/38, UL114, IE1 UL133/135, UL111A and US17 were detected (Fig. 4A). Hence these data show that there are common transcripts associated with HCMV latent infection in both CD14 (+) and CD34 (+) cells, including two lncRNAs and mRNAs that encode UL84 and UL44. At 3 days post infection less than half of the total HCMV genes were expressed and at a very low levels (Fig. 4B, 3 day PI). In reactivated cells, the entire HCMV genome was activated and the relative amount of transcription was about 5-fold higher (Fig. 4B, reactivated). Cells were co-cultured with HFFs and cells were observed for the formation of green plaques. Green plaques were readily visible after a 12-day incubation indicating that efficient reactivation from latent infection was achieved (Fig. 4B, reactivated inset image of green plaques on HFF cells).
We also evaluated transcripts present in naturally latent infected CD14 (+) monocytes and CD34 (+) cells isolated from whole blood from pooled HCMV positive donors (Renown Hospital, Reno NV). Cells were isolated, RNA was extracted and next generation sequencing libraries were generated. RNA-Seq analysis performed on naturally latent infected cells showed that, although the relative abundance of individual transcripts varied between experimental and natural infection, the transcripts present were mostly consistent with experimental latency models (tables 1 and 2, tables 3 and 4). These data strongly suggest that the most of the transcripts detected during experimental latency in both CD14 (+) and CD34 (+) cells mirrors an HCMV natural latent infection and consequently validates in vitro HCMV latency infection protocols used for this study.
Since we observed robust transcription at 5 days post infection of CD14 (+) monocytes, we wanted to determine if virion associated transcripts could be detected from UV inactivated virus at 5 days post infection and evaluate the ability of virus encoded mRNAs to persist in CD14 (+) monocytes. It was previously described that HCMV virions contain virus-encoded mRNAs [73], [74]. Hence we infected CD14 (+) monocytes with UV inactivated virus and performed qPCR at 5 and 10 days post infection. We evaluated infected cells for the presence of transcripts detected during latency as well as IE2 mRNA. As expected all transcripts were detected at 1 hr post infection of CD14 (+) monocytes, suggesting that these transcripts are packaged within the HCMV virion (Fig. 5, 1 hr UV). At 5 and 10 days post infection the amount of fold increase for all transcripts from cells treated with non-UV inactivated virus was consistent with what was observed in the RNA-Seq, strongly suggesting lytic infection (Fig. 5, Day 5 and Day 10 UV). However in UV inactivated samples, at 5 or 10 days post infection, all transcripts were below 1-fold increase compared to mock infected samples suggesting that RNA transcripts observed during latency are not the result of input virus or from initial transcription observed at 5 days post infection (Fig. 5, Day 5 and 10 UV). These results indicated that there is a detectable amount of input mRNAs from virions in infected CD14 (+) monocytes, however these input transcripts are short lived and absent by 5-days post infection.
The RNA-Seq data identified the presence of two transcripts in HCMV latently infected cells that encode UL84 and UL44, two proteins that participate in virus lytic DNA replication in human fibroblasts [3], [4]. UL84 and UL44 interact with oriLyt and other regions of the HCMV genome during lytic infection [5], [6], [75], [76]. Also, UL84 interacts with UL44, the DNA polymerase processivity factor in infected cells [77]. Since UL44 and UL84 were the only two transcripts encoding apparent DNA binding proteins observed in the RNA-Seq analysis in latently infected cells, it allows for the possibility that these two proteins may participate in HCMV latent DNA replication. The presence of these two transcripts in latently infected cells suggests that maintenance of the latent viral genome may involve some of the lytic virus machinery. Hence we investigated if these two proteins interacted with the viral genome under latent conditions in CD14 (+) monocytes.
ChIP-Seq analysis for UL84 and UL44 was performed in latently infected CD14 (+) cells to determine if these proteins interact with the latent HCMV genome in CD14 (+) cells. Ten million CD14 (+) monocytes were infected with wt FIX BAC virus and mRNA levels were monitored until no IE gene expression was detected (18 days). Infected cells were then processed for ChIP-Seq and DNA-protein complexes were immunoprecipitated using antibodies specific for UL84 or UL44. Immunoprecipitated DNA was used to generate libraries for next generation sequencing using input DNA as a reference. ChIP-Seq was performed using Illumina MiSeq instrument. Data was analyzed and peaks were identified using CLC Bio Genomics Workbench software using the ChIP-Seq analysis package compared to input DNA.
ChIP-Seq analysis identified 5 major peaks corresponding to the binding domains for UL84 in the FIX BAC virus genome during latent infection (Fig. 6A, Red peaks). Two of the peaks mapped to the promoter regions for UL84 and UL44 and suggests that UL84 may regulate the expression of these genes during latent infection (Fig. 6A, Red Peaks.). Another peak was localized to the promoter region for UL112/113 and the upstream regions regulating LUNA and MIE gene expression. Also, the region containing oriLyt and encoding the lncRNA4.9 showed specific binding to UL84 in latently infected cells (Fig. 6A, Red peaks). For UL44, many of the binding domains in latently infected monocytes overlapped with those identified for UL84. UL44 interacted most prominently with the upstream region encoding UL95 (Fig. 6A, Blue peaks), one of the transcripts identified from the RNA-Seq analysis from latently infected CD14 (+) monocytes. Interestingly, UL44 also interacted with the region of the genome encoding lncRNA2.7 and, along with UL84, the region upstream of its own coding sequence (Fig. 6A, Blue peaks). Table 5 is shows the number of reads for UL44 and UL84 ChIP-Seq experiments corresponding to the peak map.
These data indicate that UL84 and UL44 interact with the HCMV genome during latent infection in CD14 (+) monocytes and suggests these proteins may activate or suppress several genes during latency and allows for the possibility that viral encoded replication proteins may participate in viral genome maintenance.
Two HCMV transcripts identified in latently infected CD14 (+) monocytes and CD34 (+) cells were long noncoding RNAs. Our previous studies investigating Kaposi's sarcoma-associated herpesvirus (KSHV) lncRNA PAN, showed that herpesvirus encoded lncRNAs can be regulators of both viral and cellular gene expression [62]. lncRNAs associate with chromatin modifying complexes [78] and interact with components of the polycomb repressive complex 2 (PRC2). The PRC2 complex is composed of EZH2, SUZ12 and EED-1. EZH2 is a protein that adds three methyl groups to lysine 27 of histone 3 [79]. SUZ12 is a protein that contains a zinc finger domain that is the point of contact with RNA [80]. EED-1 interacts with HDAC1 and histone deacetylase and various other proteins to mediate gene repression [81]. These PRC2 proteins can mediate changes in histone modifications (methylation) and subsequent repression of gene expression from various genetic loci. Hence the interaction of lncRNAs with PRC2 can globally influence gene expression. Since the RNA-Seq identified HCMV encoded lncRNAs 2.7 and 4.9 during latent infection, we investigated if the newly discovered RNA4.9 could interact with components of the PRC2. Also, since it was previously described that HCMV UL84 is an RNA binding protein [82] and UL84 is also present during latent infection, we also evaluated if lncRNA4.9 interacted with UL84 in latently infected CD14 (+) monocytes.
HCMV latently infected CD14 (+) monocytes were fixed and RNA crosslinking immunoprecipitation (rCLIP) was performed using antibodies specific for EZH2, SUZ12 and UL84. Also, as controls we used an isotype control antibody for immunoprecipitations. Immunoprecipitated RNA-proteins complexes were reverse cross-linked and cDNA was generated. cDNA was used for PCR amplification using primers specific for RNA4.9, cyclophilin A or UL138 RNA. PCR amplification products were observed for immunoprecipitations using UL84, EZH2 and SUZ12 antibodies and PCR primers specific for RNA4.9 (Fig. 6B). No PCR amplification product was detected when the isotype control antibodies or an antibody to the cellular protein GAPDH was used or when using PCR primers specific for UL138 or cyclophilin A (Fig. 6B). Also, no PCR product was observed when the reverse transcriptase was omitted from the PCR protocol (Fig. 6B, RNA4.9, no RT). These results show that HCMV RNA4.9 interacts with viral encoded UL84 and components of the PRC2. Hence, RNA4.9 has the potential to function as a regulatory RNA in the context of HCMV latent infection of CD14 (+) monocytes.
Since it was demonstrated that RNA4.9 interacts with components of the PRC2 we wanted to evaluate if these same factors bound to the MIEP under latent conditions. We performed ChIP assays using latently infected CD14 (+) cells and immunoprecipitated protein-DNA complexes with antibodies specific for EZH2 or SUZ12. Immunoprecipitated DNA was subjected to PCR using primers specific for the core promoter region for the major immediate early gene locus (PCR primer set MIEP-3, see figure 7 and Table S4 in Text S1). Both SUZ12 and EZH2 were shown to interact with the MIEP region under latent conditions (Fig. 6C). We also evaluated the LUNA promoter during latency for the presence of SUZ12 and EZH2. Consistent with the constitutive expression of LUNA during latency no specific interaction of PRC2 components was observed (Fig. 6C, LUNA-Pr). Control immunoprecipitations using isotype specific antibodies shown no PCR product. These experiments, coupled with the observation that RNA4.9 interacts with polycomb proteins, suggest that RNA4.9 could mediate gene suppression at the MIEP during latency.
As mentioned previously, one of activities of lncRNAs is to mediate changes in gene expression by an interaction with PRC and chromatin. Since we observed the presence of RNA4.9 during latent infection and the transcript was shown to interact with UL84 and PRC2 factors, it is logical to assume that RNA4.9 is involved in regulation of gene expression during latency. The ChIP-Seq analysis showed that the major interaction domain for UL84 was at the MIE promoter (MIEP) region. Therefore we investigated if RNA4.9 interacted with the latent HCMV genome at the MIEP. We performed chromatin isolation by RNA purification (ChIRP) using biotinylated oligonucleotides specific for RNA4.9 to determine if the transcript interacts with the MIEP region during both latent and initial infection of CD14 (+) monocytes. We evaluated the ability of RNA4.9 to interact with the MIEP across 5 regions that mapped to the unique, enhancer, core promoter, exon 1 and the first intron of the IE gene (Fig. 7A). Five and 18 day infected CD14 (+) monocytes were used for ChIRP analysis of RNA4.9 binding to the MIEP. At 5 days, the ChIRP analysis showed only a slight interaction of RNA4.9 with the MIEP and enhancer region (Fig. 7B CD14 (+) 5 dpi, lanes RNA4.9 ChIRP). However during latent infection a strong interaction with the HCMV MIEP, enhancer and a lesser interaction with the first exon was detected (Fig. 7B CD14 (+) 18 dpi, lanes RNA4.9 ChIRP). This interaction, coupled with the observation that RNA4.9 interacts with components of the PRC2 strongly suggest that one mechanism for repression of immediate early gene expression during latent infection is the epigenetic modification of chromatin regulating MIEP region mediated by the virus encoded lncRNA4.9.
Since the interaction of PRC2 with chromatin is associated with the increase in the H3K27me3 repressive mark, we investigated the relative marking of this histone modification at the MIEP region. CD14 (+) cells were infected with FIX BAC virus and cells where harvested at 3-days post infection or during latency at 15 days post infection. ChIP assays were performed using antibodies specific for H3K27me3 or control proteins. The amount of H3K27me3 mark was evaluated for enrichment at MIEP-1, MIEP-2 and MIEP-3 regions (Fig. 7A) as well as at the LUNA and GAPDH promoters. During latency, there was a significant increase in enrichment of the repressive H3K27me3 mark at regions MIEP-2 and MIEP-3 of the MIEP region (Fig. 7C). These loci correspond to the regions where RNA4.9 binding is the most robust (Fig. 7B). The enrichment of H3K27me3 at the LUNA promoter during latency changes only slightly during latent infection, which is consistent with the constitutive expression of LUNA during lytic and latent phases of infection (Fig. 7C). We also evaluated the enrichment of the transcriptional activation H3K4me3 mark at the MIEP region. This mark does decrease during latent infection, consistent with less gene activation (Fig. 7C). The results of these experiments suggest that RNA4.9 interacts with the MIEP region and mediates the enrichment of the repressive H3K27me3 mark to repress HCMV transcription of IE2.
Since we observed the enrichment of the repressive H3K27me3 mark at the MIEP region we investigated if the over expression of specific H3K27me3 demethylases could have an affect on reactivation, specifically the release of repression of IE2 gene expression. Plasmids expressing UTX and JMJD3 were cotransfected into latently infected CD14 (+) monocytes and 72 h post transfection total cellular RNA was harvested and qPCR was performed. We examined mRNA accumulation for IE2, UL84, RNA2.7 and UL54 (polymerase). As a control, we transfected a plasmid that expressed EGFP. mRNA expression levels were compared to mock transfected latently infected CD14 (+) cells. In cells cotransfected with UTX and JMJD3 expression plasmids there was a 130 fold increase in IE2 mRNA accumulation (Fig. 7D). Transcripts encoding RNA2.7, UL84 and UL54 also increased markedly in cells transfected with demethylases expressing plasmids (Fig. 7D). Latently infected CD14 (+) cells transfected with the EGFP expressing plasmid show no significant increase in mRNA accumulation (Fig. 7D). These experiments strongly suggest that latent HCMV genomes are silenced, at least partly, by repressed chromatin marked with H3K27me3. Also, our data suggests that RNA4.9 participates in transcriptional suppression of MIE gene expression.
We postulate that the region(s) of the viral genome that could potentially serve as latent replication/maintenance elements should be depleted of nucleosomes and hence would define the major cis regulatory elements within the latent HCMV chromosome. This assumption is based on previous studies showing that origins of replication are depleted of bulk nucleosomes [68], [83]–[86]. It was further demonstrated in Saccharomyces cerevisiae that nucleosomes are positioned as flanking replication origins [87]. We employed the method Formaldehyde Assisted Isolation of Regulatory Elements (FAIRE) to determine which regions within the latent HCMV chromosome are depleted of nucleosome structures [88]–[90]. FAIRE is a powerful approach to identify genome-wide active regulatory elements in vivo. FAIRE involves formaldehyde crosslinking of cells, followed by shearing of chromatin and subsequent phenol/chloroform extraction. The procedure is based on the fact that histones will crosslink with DNA and nucleosome free chromatin will be preferentially partitioned into the aqueous phase. These genomic regions are then mapped back to the HCMV genome using next generation sequencing (Fig. 8A).
To study the HCMV chromosome during latent infection we used the CD14 (+) monocyte experimental model [46]. In this experimental latency model, CD14 (+) monocytes are cultured using specific growth media and culture plates that retain CD14 (+) monocytes in an undifferentiated state. Cells were infected with the HCMV clinical isolate FIX BAC strain and monitored for the expression of immediate early gene expression and the presence of the previously described latency associated transcripts UL138 and LUNA [41], [53], [91]. At 14 days post infection of CD14 (+) monocytes, the level of IE1/2 mRNA was undetectable by qPCR, however UL138 and LUNA transcripts were still present (not shown).
Latently infected cells were harvested and subjected FAIRE followed by next generation sequencing. CD14 (+) monocytes were infected with FIX BAC virus and cells were harvested at 4 and 18 days post infection. At 18 days post infection cells expressed latency-associated transcripts and were absent for expression of immediate early transcripts. Cells were subjected to FAIRE and DNA was analyzed using next generation sequencing (60 million reads per sample, paired end sequencing). FAIRE data was analyzed using CLC Bio Genomics Workbench software (ChIP-Seq analysis). At 4 days post infection of CD14 (+) monocytes over 130 peaks or nucleosome-depleted regions were elucidated (Fig. 8B). These regions were distributed across the HCMV genome with a prominent peak at the major immediate early promoter region (Fig. 8B). These “open” active regions across the HCMV genome are consistent with robust transcription and replication during lytic infection. The FAIRE profile was significantly different during a latent HCMV infection at 18 days. Nucleosome depleted regions of the viral genome were consistent with what was observed from the RNA-Seq analysis (Fig. 2B), in that active regions were specific for loci of the viral chromosome where latent viral transcription was observed (Fig. 8C). One obvious exception was the high degree of sequence reads detected for the inverted and terminal repeat region (TR) of the genome (Fig. 8C, Red Arrow). Although nucleosome depleted regions were identified at both IR and TR regions of the genome, next generation sequencing could not distinguish between the two regions because of the high degree of homology. Also, although a portion of the IR region is not present in FIX BAC, the IR region was identified because the FAIRE-Seq reads were mapped to the annotated parent virus strain (VR1814), hence the nucleosome depleted regions were localized only to the TR regions of the genome. The presence of these highly nucleosome depleted loci, coupled with the observation that no viral transcripts were detected from this region by RNA-Seq during latent infection, strongly suggests that these regions of the genome could act as elements that mediate DNA replication/maintenance.
Since FAIRE data strongly suggested that the TR region was depleted of nucleosomes we investigated if the TR region associated with the protein components of the pre-replication complex (pre-RC). HCMV latently infected CD14 (+) monocytes were fixed and a chromatin immunoprecipitation (ChIP) assay was performed using antibodies specific for pre-RC proteins MCM3 and CDT1 proteins. We also performed immunoprecipitations using isotype control antibodies and an antibody to GAPDH. Figure 9A is a schematic showing the HCMV genome and the location of the terminal repeat (TR) region of the viral DNA. Also shown are primer sets used for amplification of DNA immunoprecipitated from the ChIP assay. Immunoprecipitated protein-DNA complexes were amplified using PCR primers designed to amplify the TR (Fig. 9A, set 1 and 2) of the HCMV genome. As a control we also used primers specific for the UL25 genomic locus. PCR products were observed for ChIP samples using CDT1 or MCM3 specific antibodies when primer set 1 was used in the PCR mixture (Fig. 9B). No specific amplification product was observed from samples using isotype control or GAPDH antibodies, or using primers designed to amplify the UL25 region (Fig. 9B). These data strongly suggest that the TR region of the HCMV genome interacts with factors involved in cellular DNA replication and is consistent with previous results from other herpesvirus systems and latent origins.
Since our data suggested that the terminal repeat region of the HCMV genome was involved in replication/maintenance of the latent viral genome, our next step was to develop an assay to evaluate the ability of the TR element to mediate genome maintenance in latently infected cells. We postulate that if the TR element mediates viral chromosome maintenance then a plasmid containing the TR element would persist in latently infected cells. To this end we subcloned the HCMV FIX BAC TR element, as it would exist in the circular genome form (Fig. 10A), since previous data indicated that latent HCMV genomes exist as a circular episome [92]. This TR subclone, pTR, would be used to transfect HCMV latently infected CD14 (+) monocytes where required viral and cellular encoded factors would be supplied in trans from resident viral DNA and the host chromosome (Fig. 10B). Latently infected CD14 (+) monocytes were transfected with pTR using Amaxa nucleofector and cells were processed as shown in figure 7C. As controls, we also transfected the parent vector pGEM and a plasmid that contains the HCMV origin of lytic DNA replication, oriLyt. After 15 days post transfection (26 days post infection) cells were prepared and DNA was resolved using a Gardella gel [63]. The gel was transferred to a nylon membrane and hybridized to a radiolabeled pGEM probe. Several bands were detected in the lane that contained cells that were transfected with the TR-containing plasmid whereas lanes that contained samples that were transfected with either pGEM, cloned oriLyt or mock infected TR transfected cells failed to show a hybridization product (Fig. 10D). Two bands were detected on the Southern blot from cells containing pTR. These two bands could be due to variable lengths of the repeat sequence present with the TR region of the genome. We also evaluated 26 day latently infected CD14 (+) that were transfected with oriLyt or TR containing plasmids. All transfected cells expressed latency-associated transcripts, in the absence of IE2 gene expression, detected in the RNA-Seq analysis at 26 days post infection (Fig. 10D, inset graph).
These data show that the cloned HCMV TR element is capable of persisting in latently infected CD14 (+) monocytes. This is the first report describing the existence of a DNA element within the HCMV genome that mediates maintenance of the viral chromosome during latency.
To date HCMV latency is defined as the lack of production of infectious virus, absence of immediate early gene expression and the presence of a few specific latency-associated transcripts, as well as the ability to reactivate latent resident viral DNA. Previous studies have identified the expression of some latency-associated transcripts in both naturally and experimentally infected CD14 (+) and CD34 (+) cells [41], [42], [48], [49].
We utilized two experimental systems to study HCMV latency; the first system developed for CD14 (+) monocytes is where cells are cultured in an undifferentiated state by using specific cytokines and a growth surface that retards cell attachment [46]. Using CD14 (+) monocytes cultured under these specific conditions resulted in the presence of the HCMV genome in the absence of immediate early gene expression after approximately 18 days post infection. For the HCMV latency infection protocol used here, we treated CD14 (+) monocytes with media formulation previously shown to maintain monocytes in an undifferentiated state [46]. Although it was previously reported that IE gene expression was not detected at 11 days post infection, in our hands the loss of IE gene expression required the culturing of cells for approximately 14 days. Hence, to ensure that viral infection was indeed in a latent phase, we evaluated transcription at 18 days post infection and only when cells were negative for IE gene expression. RNA-Seq has major advantages over other previously employed methods to identify HCMV latency associated transcripts. RNA-Seq allows for a quantitative unbiased evaluation of transcripts present in infected cells. After a five-day infection, RNA-Seq showed the presence of transcripts originating from most of the HCMV genome. Although this observation suggests that a lytic infection precedes the establishment of latency, more experiments are needed to confirm this finding. Previous studies demonstrated that viral gene expression was required for establishment of latency and several virus-encoded genes were identified [39], [48]. Although at present we do not know the significance of wide spread gene expression in CD14 (+) at early time points post infection, a lytic-type infection upon initial infection of primary cells was observed previously for KSHV and EBV [93]–[95]. In those systems, this initial burst of lytic replication may be critical for subsequent establishment of latency. The observed lytic infection for HCMV in monocytes warrants further investigation and will be the focus of future studies.
The second experimental latency system used infection of CD34 (+) cells where we isolated CD34 (+) cell populations, since evidence suggests that these cell types support HCMV latent virus [96]. RNA-Seq showed that infected CD34 (+) cells did not undergo the same robust HCMV viral gene expression pattern as observed with initial infection of CD14 (+) monocytes. At three days post infection of CD34 (+) cells only a subset of genes were expressed, whereas in CD14 (+) cells almost the entire HCMV genome showed active transcription. Although the same core transcripts, including the lncRNAs RNA2.7 and RNA4.9, were detected in latently infected CD34 (+) cells and CD14 (+) monocytes, some differences were also observed. One of the most notable was the presence of IE1 mRNA in latently infected CD34 (+) cells, which was also confirmed in naturally infected CD34 (+) cells. The presence of IE1 mRNA in latently infected CD34 (+) cells was observed previously [26] and allows for the possibility that this protein could tether the HCMV chromosome to genomic DNA since it was previously demonstrated that IE1 interacts with cellular DNA [97], [98]. The lack of detection of IE2, late gene expression or early genes involved in lytic DNA replication argue against the possibility that that a population of cells are undergoing lytic replication.
Previous studies have examined the expression patterns in HCMV latently infected cells. All of these previous studies used microarrays to evaluate the presence of mRNAs [26], [41], [48]. Although these studies yielded important information regarding gene expression during latent infection, RNA-Seq has several advantages over microarrays [99]–[101]. One major advantage is that RNA-Seq is unbiased in that no prior knowledge about mRNA sequence is required. One example of this demonstrated in the present study is that most microarrays target known ORFs, hence the presence of expressed lncRNAs or antisense transcripts may be overlooked. Also, recent studies indicate that RNA-Seq is more sensitive than conventional microarrays [102], [103].
One of the most striking findings from the study presented here is the observation that transcripts that encode proteins associated with lytic DNA replication were expressed during latency. These transcripts were present even though immediate early mRNAs were not expressed after approximately 14 days post infection of CD14 (+) monocytes. Hence the observed expression of transcripts identified was not the result of transactivation from IE2. One other explanation for the presence of apparent lytic mRNAs is the detection of transcripts with long half-lives that were still present from the initial infection and lytic replication. This is unlikely since infection of CD14 (+) monocytes with UV-inactivated virus indicated that, although all latent RNAs were detected at 1 hr post infection due to apparent packaging in virions, we were unable to detect these transcripts at 5 or 10 days post infection. This suggests that these mRNA half-lives are less than 5 days in CD14 (+) monocytes.
The RNA-Seq analysis showed that mRNAs encoding ORFs UL50 and 52 were present in latently infected cells. The expression of these ORFs during latency is interesting since UL52 is implicated in cleavage-packaging. UL52 a protein required for virus growth in human fibroblasts is localized to the nucleus and appears to enclose replication compartments. Although implicated in encapsidation/cleavage of virus DNA, UL52 was not associated with other proteins known to perform these functions [104]. However, UL52 is quite unique with respect to other proteins involved in cleavage and packaging in that it is localized to the nucleus and found in replication compartments [104]. Hence, it could be that UL52 supplies a function in latency that is related to replication of the virus genome during latent infection.
For UL50, this protein is implicated in nuclear egress and is associated with the nuclear lamina as part to the nuclear egress complex (NEC) [105]. UL50 associates with the cellular factors p32 and protein kinase C (PKC). UL50 is localized to the inner nuclear membrane and associated with the nuclear lamina along with UL53 [105], [106]. UL50 was shown to associate with BiP and this interaction was essential for phosphorylation of the nuclear lamina [106]. UL50 along with UL53 may act to remodel the nuclear lamina [107]. We speculate that UL50, like UL52, may play a role in maintaining the integrity of the HCMV latent virus genome.
In both latently infected CD14 (+) and CD34 (+) cells the transcripts encoding UL95 and UL87 were among the most abundant. This observation also occurred in naturally infected cells. UL87 and UL95 encode essential early transcripts that produce proteins that apparently affect late gene expression including the UL44 late kinetic transcription and colocalize with UL44 prior to initiation of viral DNA synthesis [108]. Interestingly, the pre expression of UL87 inhibited the expression of MIE genes and virus DNA replication [108]. UL87 and UL95 are recruited to replication compartments during lytic infection [108]. Hence, during latent infection UL87 and UL95 could serve to enhance the expression of UL44 and/or UL50 and 52. Another possibility is that UL87 and UL95 may serve to help suppress the expression of IE2 during a latent infection.
UL84 is the putative initiator protein for lytic DNA replication and interacts with UL44 [5], [76], [82], [109], [110]. UL84 is a phosphoprotein that exhibits nucleocytoplasmic shuttling that is required for function [82], [111]. UL44 is the DNA polymerase processivity factor, however these herpesvirus proteins have recently been shown to have diverse functions with respect to regulation of gene expression and initiation of DNA synthesis [112]–[117]. The observation that transcripts encoding UL44 and UL84 are produced in latently infected cells suggests that the maintenance of the HCMV virus genome may involve a mechanism that utilizes some of the HCMV lytic DNA synthesis machinery. UL84 interacts with both Ku70 and Ku80 and the Ku70/80 complex is involved in DNA repair [76]. Ku70/80 is and ATP dependent helicase that is involved with DNA repair and interacts directly with the RNA component (hTR) of telomerase [118], [119]. These interactions suggest that stability of the HCMV genomic DNA in latently infected cells may also involve DNA repair enzymes and telomerase. We also demonstrated that UL84 protein interacted with the HCMV latent genome. We speculate that UL84 may act to suppress IE gene expression, however UL84 may also mediate its own expression and that of UL44 and LUNA. The interaction of UL84 with oriLyt could mean that the protein acts to suppress the activation of the lytic replication. However, it is just as plausible that UL84 (possibly in cooperation with cellular factors) activates the promoter region within oriLyt to facilitate the expression of RNA4.9. Another possibility could be that UL84 is acting in the capacity of a replication factor at oriLyt to replicate the latent genome. Interestingly, UL84 is dispensable for lytic DNA replication in one clinical isolate, BACmid clone (TB40/E) [120]. Hence one possibility is that in clinical isolates UL84 is not required for lytic replication, but may mediate latent genome maintenance.
LncRNAs have emerged as significant regulators of gene expression in human cells. Although the mechanisms used by lncRNAs vary, one well-defined mechanism involves their interaction with chromatin modifying complexes. lncRNAs can act as molecular scaffolds to mediate epigenetic changes in histones, which results in activation or suppression of gene expression [121]–[124]. Hence, lncRNAs can globally affect gene expression patterns. We have recently defined one mechanism of action for KSHV PAN RNA in lytically infected cells. PAN RNA is a highly abundant transcript and we show that expression of PAN RNA can result in the disregulation of genes involved in immune response and cell cycle [62]. This suppression of gene expression is most likely achieved by the interaction of PAN RNA with polycomb proteins. Activation of KSHV and cellular gene expression by PAN RNA appears to involve its interaction with the demethylases UTX and JMJD3, which remove the repressive H3K27me3 mark [62]. Therefore evidence shows that viral lncRNAs are significant regulators of both cellular and viral gene expression.
In HCMV, the lncRNA2.7 was shown to regulate the apoptosis pathway during lytic infection [125]. Since we also observed the expression of RNA2.7 in latently infected cells it is likely that this transcripts plays the same role in latently infected CD14 (+) cells. Early studies did not detect the presence of this transcript in latently infected CD14 (+) cells [35]. One possibility for the discrepancy between this early study and the present one is that we used highly sensitive RNA-Seq, where the early studies used traditional RT-PCR.
The lncRNA4.9 was recently discovered by next generation sequencing of the HCMV transcriptome during lytic infection [10]. RNA4.9 initiates within oriLyt and extends upstream and terminates just downstream of UL69 [10]. Hence, since of the location of the 5 prime start of transcription for RNA4.9 is within oriLyt this suggests that the promoter region is also within oriLyt. We previously identified a bidirectional promoter within oriLyt just upstream of the RNA4.9 transcriptional start site [126]. We now show data that strongly suggests that RNA4.9 may act as a regulatory RNA with the potential to control cellular and viral gene expression during HCMV latency in CD14 (+) monocytes. This is the first reporting of an HCMV encoded RNA that interacts with the PRC. Using ChIRP we show that RNA4.9 physically interacts with the HCMV latent genome in the MIEP region. This observation, combined with the data showing that RNA4.9 interacts with PRC proteins that are also bound to the same locus suggests that this transcript may repress the expression of IE2 during latency. Previous studies have demonstrated epigenetic regulation of IE gene expression [27], [37], [47], [127]. These earlier reports investigated the deposition of the acetylated H4 mark within the MIEP. PRC2 is associated with the repressive H3K27me3 mark; hence we examined the presence of this mark at the MIEP. Recent studies have shown that the presence of H3K27me3 is associated with herpesvirus latent genomes and repression of lytic gene expression [128]–[131]. One of the key factors for mediating gene silencing is lncRNAs [132], [133]. Data presented here strongly implicates lncRNA4.9 as a targeting factor for suppression of IE gene expression during latent infection in CD14 (+) cells.
For CD34 (+) latently infected cells, the presence of UL28/29 and UL37/38 is interesting. It was recently reported that this locus encodes a spliced transcript and stimulate the accumulation of immediate early RNAs [134]. UL28/29 proteins interact with nucleosome remodeling and deacetylase protein complex, NuRD along with UL38 and UL28/29 enhanced activity of the MIEP [135]. Recently, UL28/29 was shown to interact with UL84, p53 and suppresses cellular gene expression during lytic infection [136]. Hence the expression of UL28/29 during latency suggests viral control of p53-regulated genes in the context of a latent infection.
Although RNA-Seq showed some differences between the CD34 (+) and CD14 (+) latent transcriptome, many transcripts were common to both systems including the presence of lncRNAs, UL84 and UL44. One explanation for the presence of different factors could be that specific viral factors may be required to maintain genome latency in a cell type specific manner.
Our RNA-Seq evaluation of HCMV naturally latently infected CD14 (+) and CD34 (+) cells isolated from seropositive donors showed an almost exact match to the transcripts observed from experimental latency, although relative abundances differed. Since pooled blood from HCMV seropositive donors was used, we do not know the prevalence or degree of transcript expression in individual donors. Nevertheless, data shows that experimental latency closely matches natural latent infection. This is the first evaluation of natural latent infection by next generation RNA sequencing. Interestingly, RNA-Seq detected the previously transcript originating from UL126a in naturally infected cells [137], [138]. This transcript was not detected in our experimental latency model. One explanation for the discrepancy is that some differences between natural and experimental latency may exist. This could be due the fact that natural infection data was obtained from pooled donors and hence variations in expression levels are reflected in the RNA-Seq analysis. Nevertheless, overall the transcripts present during natural and experimental infection vary little. This is a very significant finding in that HCMV CD14 (+) and CD34 (+) experimental latency systems are widely used. Our data suggests a strong link between transcripts expressed during HCMV natural latent infection and those that occur in cell culture.
It was previously demonstrated that herpesvirus latent origins are marked by lack of nucleosome structure and the presence of cellular factors involved in DNA replication or licensing [139]. Also, it was shown that nucleosome assembly proteins interact with herpesvirus-encoded proteins to regulate replication at the latent origins replication [140], [141]. For EBV, EBNA-1 was shown to destabilize nucleosomes at the latent origin [142]. Also, for EBV, cellular factors play a significant role in replication of the latent virus origin [143]–[145]. Hence previous evidence indicates that specific herpesvirus and cellular proteins bind to DNA domains to mediate latent viral genome replication and maintenance. Currently little data exists regarding the mechanism involved in maintenance of the viral chromosome in HCMV latently infected cells. Previous studies indicated the HCMV latent genome is in a circular form [25]. Based on the latency models from the gamma herpesviruses, it is assumed that the maintenance and replication of the HCMV latent viral genome requires trans and cis acting factors. In this report we show that, consistent with the presence of active chromatin, that the TR region of the HCMV genome are depleted of nucleosomes during latent infection of CD14 (+) monocytes. We developed a latent replication/maintenance assay where CD14 (+) monocytes harboring the latent HCMV chromosome were transfected with the TR containing plasmid. Control plasmids, including a plasmid that contained oriLyt, failed to persist in latently infected cells. The TR element of HCMV in the circular form contains two “a” sequences and many repetitive DNA sequence motifs. Hence, during replication/maintenance there is a potential for variability of these repeat sequences. The fact that we observed two bands in the Gardella gel may indicate the presence of variable repeated regions. This was also postulated for the KSHV latent origin, which also contains several repeated DNA sequences [15]. Although the Gardella gel clearly shows that the TR plasmid can persist in latently infected cells, further studies are needed to demonstrate if the TR element can direct plasmid replication. Nevertheless, this is the first reporting of a region of the HCMV chromosome that mediates viral DNA maintenance.
The identification of cis and trans acting factors involved in HCMV latency in CD14 (+) monocytes and CD34 (+) cells now allows for a more in depth analysis of the factors required for viral chromosome maintenance/replication.
|
10.1371/journal.pbio.0060289 | Gamma-Secretase Represents a Therapeutic Target for the Treatment of Invasive Glioma Mediated by the p75 Neurotrophin Receptor | The multifunctional signaling protein p75 neurotrophin receptor (p75NTR) is a central regulator and major contributor to the highly invasive nature of malignant gliomas. Here, we show that neurotrophin-dependent regulated intramembrane proteolysis (RIP) of p75NTR is required for p75NTR-mediated glioma invasion, and identify a previously unnamed process for targeted glioma therapy. Expression of cleavage-resistant chimeras of p75NTR or treatment of animals bearing p75NTR-positive intracranial tumors with clinically applicable γ-secretase inhibitors resulted in dramatically decreased glioma invasion and prolonged survival. Importantly, proteolytic processing of p75NTR was observed in p75NTR-positive patient tumor specimens and brain tumor initiating cells. This work highlights the importance of p75NTR as a therapeutic target, suggesting that γ-secretase inhibitors may have direct clinical application for the treatment of malignant glioma.
| Despite technical advances, clinical prognosis of patients with malignant glioma, with an average survival of less than one year, has not changed. The highly invasive nature of these tumors, together with the recently identified brain tumor-initiating cells, provide disease reservoirs that render these tumors incurable by conventional therapies. Here, we present the first evidence to our knowledge that regulated intramembrane proteolysis of the neurotrophin receptor p75NTR is a critical regulator of glioma invasion. Inhibition of this process by clinically relevant γ-secretase inhibitors dramatically impairs the highly invasive nature of genetically distinct glioblastomas and brain tumor-initiating cells and prolongs survival. These data highlight regulated intramembrane proteolysis as a therapeutic target of malignant glioma and implicate the application of γ-secretase inhibitors in the treatment of these devastating tumors.
| Human malignant glioma (MG) is one of the most common primary central nervous system tumors in adults. These tumors are diffuse, highly invasive, with dismal prognosis, and long-term survivors are rare [1,2]. MG extend tendrils of tumor several centimeters away from the main tumor mass. These, as well as the recently identified brain tumor-derived stem-like cells [3–6], herein called brain tumor-initiating cells (BTICs), act as “disease reservoirs,” rendering these tumors refractory to available treatments such as surgery or radiotherapy [7,8]. The highly invasive nature of these tumors is the result of genotypic and phenotypic changes that result in the activation of a number of coordinate cellular programs, including those necessary for migration (e.g., motility) and invasion (e.g., extracellular matrix [ECM] degradation) [9] and changes in pathway signaling that impart resistance to conventional treatments by reducing proliferation and increasing resistance to apoptosis [8,10,11]. A detailed understanding of the mechanisms underlying this invasive behavior is essential for the development of effective therapies.
Several genes, including those that encode uPA/uPAR, ephrinB3/EphB2, matrix metalloproteinases (MMPs), a disintegrin and metalloproteases (ADAMs), cathepsins, and integrins, have previously been implicated in glioma invasion [12]. More recently, gene expression profiling identified several subclasses of gliomas that separate tumors into good and poor prognosis groups of which diffuse infiltrative gliomas are divided into four such subclasses [13]. One of these four subclasses, designated hierarchical cluster 2B (HC2B), was found to include several genes with specific roles in cell migration and invasion, and membership in this group was found to strongly correlate with poor patient survival. Our understanding of the proteins that initiate, and the pathways that regulate, glioma invasion is continually expanding, such as the recent discovery that CD95 via the activation of the PI3K/Akt/glycogen synthetase kinase (GSK3β) pathway regulates glioma invasion [14]. However, despite recent advances and efforts to target these specific molecules or pathways, no clinically relevant agents have been identified as yet. Using a discovery-based approach and a series of functional, biochemical, and clinical studies, we have identified the p75 neurotrophin receptor (p75NTR) as a critical regulator of glioma invasion [15]. We found that p75NTR, through a neurotrophin-dependent mechanism, dramatically enhanced migration and invasion of genetically distinct glioma and that robust expression of p75NTR was detected in the highly invasive tumor cell population from p75NTR-positive glioblastoma patient specimens. In this current study, we investigated the mechanism by which p75NTR imparts this highly invasive behavior to malignant glioma, and assessed the use of a clinically applicable agent in abrogating this invasive behavior.
p75NTR elicits a large array of diverse biological responses that are regulated by a complex layer of mechanisms. These intricate layers of control have been proposed to explain the variety of cellular effects triggered by p75NTR activation. Key p75NTR signaling pathways already identified include Ras homolog gene family, member A (RhoA), Jun N-terminal kinase (JNK), mitogen-activated protein kinase (MAPK), and nuclear factor κ B (NFkB) [16]. These pathways are believed to be activated by upstream proteins that directly associate with various regions of the p75NTR intracellular domain (ICD). These proteins include guanine nucleotide dissociation inhibitor (RhoGDI), ribosome-inactivating protein-2 (RIP-2), and p75NTR-associated cell death executor (NADE) [17–20], which associate with a region referred to as the “death domain”; Schwann cell factor-1 (SC-1); neurotrophin receptor-interacting MAGE homolog (NRAGE); tumor necrosis factor (TNF) receptor-associated factor (TRAF), and neurotrophin receptor interacting factor (NRIF) [21–23], which associate with the juxtamembrane region of p75NTR; and a PDZ-containing protein Fas-associated phosphatease-1 (FAP-1), which associates with the C-terminal Ser-Pro-Val (SPV) [24]. What proteins or biological process are activated by p75NTR, however, is highly cell context specific. In addition to associating with other signaling molecules, p75NTR, similar to amyloid precursor protein (APP) and Notch, has been shown to undergo regulated α-secretase and γ-secretase cleavage, referred to as regulated intramembrane proteolysis (RIP). Cleavage of several type-1 transmembrane receptors has been implicated and shown to be necessary in eliciting some downstream biological responses [25–28]. α-Secretase cleavage of full-length p75NTR by a sheddase liberates the extracellular domain (ECD), leaving an unstable membrane-bound C-terminal fragment (CTF) that is cleaved by the γ-secretase complex to release an ICD with potential signaling capability [26,29]. Here, we show for the first time to our knowledge that regulated intramembrane proteolysis of p75NTR is a requirement for the highly invasive behavior of p75NTR-positive malignant glioma, and designate RIP as a clinical target for the treatment of invasive malignant glioma.
In a previous study, our laboratory identified p75NTR as a potent mediator of invasion in human glioma using a novel invasive glioma mouse model generated by serial in vivo selection [15]. In that study, we found that p75NTR was expressed in 22% mid-grade astrocytomas (two of nine) and 85% of glioblastoma multiforme (GBM) specimens (17 of 20), and that the p75NTR-positive glioma cells in the patient tumor cell population were more migratory than the p75NTR-negative glioma cells. Here, we investigate the mechanism underlying this p75NTR-induced invasion. In neurons, p75NTR is a substrate for sequential α- and γ-secretase–mediated intramembrane proteolysis generating 24 kDa CTF and 19 kDa ICD fragments, respectively, and the generation of these fragments are required for some of its biological functions [26,28,30–34]. We therefore sought to determine whether intramembrane proteolysis of p75NTR occurred in malignant glioma patient specimens. To do this, we assessed whether the generation of the 24-kDa CTF and the 19-kDa ICD occurred in a panel of surgically resected human glioma specimens and normal human brain. Tumor and normal tissue taken at the time of surgery were immediately snap frozen in liquid nitrogen and stored at −80 °C. Frozen tumor tissue was digested in lysis buffer and analyzed by western blots using a p75NTR cytoplasmic-specific antibody that not only detected the full-length p75NTR protein, but also detected p75NTR-positive fragments migrating at 24 and 19 kDa, respectively, in the p75NTR-positive specimens (eight of nine GBMs and two of five Grade III glioma) (Figure 1A). Hence, p75NTR processing occurs in human glioma tumors, and this suggested the possibility that p75NTR processing is required for glioma invasion.
To address the possible role(s) of p75NTR proteolytic processing in glioma cells, we assessed whether the appearance of the p75NTR-positive fragments at 24 and 19 kDa was the result of proteolytic processing of the full-length p75NTR in invading glioma cells. We have previously established the highly invasive glioma cell lines U87R and U251R for which p75NTR accounts for their invasive behavior [15]. The U87R and U251R invasive glioma cells were grown in the absence or presence of the proteasome inhibitor epoxomicin, a compound used to inhibit rapid degradation of proteins often associated with RIP-mediated proteins [26,35]. Western blot analysis showed that in addition to the 75-kDa full-length p75NTR protein, 24-kDa and 19-kDa fragments (Figure 1B, lane 1 and 5) were present and stabilized in the presence of 1 μM epoxomicin (Figure 1B, lane 2 and 6). These results are in agreement with the model that the full-length p75NTR protein is cleaved, releasing the ECD, CTF, and intracellular fragments. Next, we verified that the appearance of the 24-kDa and 19-kDa fragments was the result of sequential cleavage of p75NTR by an α-secretase and then a γ-secretase. First, we determined whether treatment of p75NTR glioma cells (U87p75NTR) using the TNF-α protease inhibitor (TAPI)-2, known to inhibit metalloproteases and ADAMs such as tumor necrosis factor-α converting enzyme (TACE) [36–38] and previously shown to inhibit the proteolytic processing of p75NTR in neurons [26,32,33,39], could inhibit p75NTR processing in glioma cells. TAPI-2 inhibited the proteolytic processing of p75NTR as indicated by the lack of CTF and ICD, and abrogated p75NTR-mediated invasion (Figure S1). Since TAPI-2 has broad specificity, and glioblastomas are known to produce high levels of many proteases, including members of the MMP, ADAM, and ADAMTS families, leaving the exact identity of the α-secretase unclear, we focused our efforts on the second cleavage event. To determine whether the generation of the 19-kDa fragment was the result of cleavage of p75NTR by a γ-secretase, U87R and U251R cells were treated with 2 μM Compound X (Calbiochem), a specific inhibitor of γ-secretase, for 4 h in the absence or presence of epoxomicin. Western blot analysis of p75NTR revealed that in the presence of the γ-secretase inhibitor, an accumulation of the 24-kDa fragment occurred without subsequent cleavage to the 19-kDa ICD, consistent with the release of the ICD of p75NTR by γ-secretase (Figure 1B, lanes 3, 4, 7, and 8). The role of processing of p75NTR was not limited to a single glioma cell line and was a general mechanism observed in glioma cells established from genetically distinct individuals (U87p75, U251p75, U343p75, and U118p75).We found that in all p75NTR-positive glioma cell lines, full-length p75NTR was cleaved to generate two fragments of 19 and 24 kDa: ICD and CTF, respectively (Figure 1C). These results demonstrate that regulated intramembrane proteolysis of p75NTR is a global event occurring in highly invasive p75NTR-positive human glioma cells.
In neurons, ectodomain shedding of p75NTR by α-secretase and then γ-secretase cleavage to generate an ICD fragment can result in the activation of downstream events [26–28,30–34]. To test whether the processing of p75NTR resulting in the release of the ICD fragment has a functional role in glioma invasion, we analyzed in vitro migration and invasion of U87R, U251R, U87p75NTR, and U251p75NTR glioma cell lines using circular monolayer migration assays (Figure 2A and 2B) and 3D-collagen invasion assays (Figure 2C and 2D) in the absence and presence of the γ-secretase inhibitor, Compound X. p75NTR-mediated glioma migration and invasion were significantly inhibited in the presence of Compound X. In contrast, when the proteasome inhibitor epoxomicin was used to stabilize p75NTR-ICD, a significant increase in migration and invasion was seen (unpublished data), consistent with increased invasion observed when a cDNA construct mimicking the ICD fragment was ectopically expressed in U87 glioma cells (Figure 3Aand 3B). To determine whether γ-secretase inhibition was confined to glioma invasion or had effects on other biological processes, we assessed the effect of γ-secretase inhibition on survival and proliferation of p75NTR-positive glioma cells. No significant change was observed on either survival or proliferation in vitro (Figure S2).
It is well known that γ-secretase has many substrates [40,41]. To directly test the role of p75NTR processing in glioma invasion, we constructed cleavage-resistant chimeric proteins of p75NTR by replacing either the transmembrane (p75FasTM) or the extracellular stalk domain of p75NTR (p75FasS) with equivalent domains from the Fas receptor [39] (Figure 3C). Both p75NTR and Fas receptors are members of the TNF receptor superfamily, and although they each contain similar domains, Fas, unlike p75NTR, does not undergo RIP. Since ectopic expression of p75NTR in the human glioma cells lines U87 and U251 was sufficient to mediate glioma invasion [15], these cell lines were used as a model system to assess the p75NTR chimeric mutants. U87 and U251 were therefore stably transfected with the cleavage-resistant p75NTR constructs (p75FasS and p75FasTM). To ensure proper function of all p75NTR protein constructs, we assessed their location, topography, and ability to bind neurotrophin in both U87 and U251 glioma cell lines. Receptor orientation and localization at the plasma membrane was confirmed by flow cytometric analysis using a monoclonal antibody specific to the ECD domain of p75NTR (Figure S3A). As expected, all p75NTR constructs were expressed at the plasma membrane with the correct topography. Next, we assessed whether the chimeric constructs could still bind neurotrophin. Previously, we demonstrated that in the absence of p75NTR, glioma cells secrete high levels of brain-derived neurotrophic factor (BDNF) protein into the culture medium in vitro. When these same cells express p75NTR, the majority of the BDNF is found to be cell associated, presumably bound to p75NTR [15]. To confirm that the p75NTR cleavage-resistant chimeric forms retained the ability to bind neurotrophin, ELISA assays were performed to detect BDNF expression in the conditioned medium and total cell lysates of U87 and U251 cells expressing p75FL, p75FasTM, p75FasS, and p75CRD130 (Figure S3B). p75CRD130 is a neurotrophin-binding mutant created by inserting four amino acids after amino acid residue 130 [15,42–47]. Expression of the chimeric p75NTR proteins (p75FasTM and p75FasS), just like the p75 wild type, resulted in a shift in BDNF localization from the conditioned medium to the cell lysate. This was in contrast to the cells expressing the neurotrophin-binding mutant (p75CRD130) or the empty vector (pcDNA) where the bulk of BDNF was detected in the culture medium. These data demonstrate that p75 NTR cleavage-resistant chimeric constructs p75FasTM and p75FasS retained their ability to bind neurotrophin (Figure S3C). Once we confirmed the correct expression and binding of the various p75NTR constructs, western blots using a p75NTR cytoplasmic domain-specific antibody were performed to evaluate proteolytic processing of the various p75NTR receptors (Figure 3D). In cells expressing p75FasS, only the full-length protein was detected, consistent with inhibition of the α-secretase cleavage, whereas the full-length 75 kDa and the 24-kDa fragment were detected in cells expressing the p75FasTM construct corresponding to the ectodomain shedding of p75NTR by α-secretase but with inhibition of the γ-secretase cleavage. Moreover, in the presence of epoxomicin, no additional p75NTR fragments were observed (Figure S4). These results demonstrate the cleavage-resistant chimeric p75NTR alleles were expressed with correct biochemical characteristics in U87 and U251 glioma cells. In addition, and consistent with the hypothesis that proteolytic processing of p75NTR is required for glioma invasion, only the full-length 75 kDa band was detected in lysates from U87 cells expressing p75CRD130, a p75NTR construct that was unable to induce glioma invasion [15].
Since we have shown that neurotrophin binding is required for p75NTR-mediated glioma invasion [15], and the neurotrophin-binding mutant p75CRD130 does not undergo RIP, it would appear that RIP of p75NTR is required for glioma invasion. To determine whether this is in fact true, U87 and U251 cells expressing the p75NTR cleavage-resistant constructs were assessed for their invasive ability using 3D-collegen invasion assays. We found that expression of cleavage-resistant forms of p75NTR (p75FasS, p75FasTM, and p75CRD130), which prevented receptor proteolysis, blocked p75NTR-mediated glioma invasion (Figure 4A and 4B), providing evidence to support a role for γ-secretase–dependent release of p75NTR ICD in mediating glioma invasion.
To determine whether p75NTR processing was required for glioma invasion in vivo, U87 glioma cell lines ectopically expressing p75FasS and p75FasTM were implanted into the brains of immunocompromised (SCID) mice. U87 glioma cells expressing full-length p75NTR (U87p75) or control vector (U87pcDNA) were used for comparison. Twenty-eight days after implantation, the mice were sacrificed, and frozen brain sections were stained with antibodies against human nuclei, to visualize all glioma cells (Figure 4C, upper panel) or with anti-human p75NTR (Figure 4C, bottom panel). Implantation of U87 glioma cells stably transfected with the control pcDNA vector led to the formation of well-circumscribed tumors, while U87 glioma cells ectopically expressing p75NTR formed tumors with highly infiltrative edges. In sharp contrast to the p75NTR-expressing U87 tumors, tumors expressing either p75FasTM or p75FasS formed well-circumscribed tumors similar to the p75NTR-negative tumors (U87pcDNA). Comparable results were seen in three independent experiments. In conjunction with the in vitro data, these data suggest that proteolytic processing of p75NTR is required for glioma invasion in vivo (Figure 4C).
It is well known that the microenvironment of tumors can change the biochemical characterization and function of cells. We have demonstrated in vitro that glioma cells expressing p75NTR undergo proteolytic processing to generate first the 24-kDa CTF and then the 19-kDa ICD. To provide evidence that RIP of p75NTR occurs in vivo, 7–9-μm cryosections from mice implanted for 3–4 wk with in vivo–selected U87R and U251R, or ectopically expressing p75NTR, p75FasS, p75FasTM, and pcDNA, were assessed for p75NTR processing by western blot. The 24- and 19-kDa fragments were found in the highly invasive glioma cells, U87R, U251R, and U87p75NTR. In contrast, neither the 24-kDa nor the 19-kDa fragment was seen in cells expressing p75FasS, and as expected, only the 24-kDa fragment was detected in cells expressing the p75FasTM, consistent with their in vitro characterization (Figure 4D). These data further support a role for RIP of p75NTR in glioma invasion.
Our data demonstrated that 85% of GBM specimens (17/20) express p75NTR, that the p75NTR-positive glioma cells in the original patient tumor cell population were more migratory [15], and that 24- and 19-kDa p75NTR-positive fragments are present in p75NTR-positive primary Grade III and GBM patient specimens (Figure 1). We therefore wanted to determine whether the appearance of these fragments in the malignant glioma patient specimens was the result of intramembrane proteolysis of p75NTR. To do this, we established primary cultures from human glioma patient tumors. The recent discovery that human stem cell-like tumor cells, termed BTICs, retain characteristics that closely recapitulate the original patient tumor [3–6,48,49] prompted us to establish the primary patient tumor cultures under neural stem cell–promoting conditions. BTICs share characteristics with neural stem cells (NSCs) such as continuous self-renewal, extensive brain parenchymal migration and infiltration, and potential for full or partial differentiation, properties not found in established glioma cell lines [50,51]. Operative samples of human GBM were obtained at the time of surgery, and brain tumor sphere cultures were established in NS-A basal medium plus epidermal growth factor (EGF) and basic fibroblast growth factor (bFGF) (EF medium). Immunocytochemical analysis of BTICs established from five glioma patients expressed the early neural cell progenitor proteins Nestin, Musashi, hSox2, and CD133 (J. J. P. Kelly, S. Weiss, P. A. Forsyth, and D. L. Senger; unpublished data). In addition, four out of five glioma tumor progenitor cells in vitro expressed high levels of p75NTR as detected by immunocytochemistry (Figure 5A) and western blot (Figure 5B). To determine whether p75NTR expressed on the BTICs undergoes RIP, BTICs were grown in the absence and presence of γ-secretase and/or 2 μM epoxomicin. Similar to the glioma cell lines, full-length p75NTR, CTF, and ICD were detected, and the 19-kDa ICD fragment was dependent on γ-secretase cleavage (Figure 5B). We next determined whether the BTICs retained their expression of p75NTR in vivo: BTIC cells were implanted into the brains of SCID mice and allowed to establish for 4–8 wk. Animals were sacrificed, and frozen sections were stained with either an anti-human nuclear antibody (to identify the human BTICs) or anti-p75NTR (Figure 5C). All tumors established from BTICs cells showed highly infiltrative tumors, and consistent with the in vitro data, four out of five tumors showed high expression of p75NTR in vivo.
p75NTR as a substrate for γ-secretase [26,33,39] adds to a growing list of proteins shown to be substrates for γ-secretase cleavage, including APP [52–54], Notch [55,56], and Notch ligands Delta1 and Jagged2 [57], ErbB4 [58], CD44 [59,60], and E-cadherin [61,62]. Our in vitro data and the recent application of γ-secretase inhibitors (egs.LY-450139 and LY-411575) in advanced clinical trials for Alzheimer disease [63–65] prompted us to investigate the use of γ-secretase inhibitors to treat highly invasive gliomas. Using an intracranial glioma model, we assessed the therapeutic potential of γ-secretase inhibitors. Parallel experiments were performed using the genetically distinct U87p75NTR and U251p75NTR glioma cell lines and the p75NTR-positive BTICs established from a patient GBM. Cells were implanted intracerebrally into SCID mice; and 3 d after implantation for U87p75NTR and U251p75NTR or 5 d after for BTICs, mice were administered subcutaneously (s.c.) either 10 mg/kg γ-secretase inhibitor or vehicle control (corn oil) once/day for 2–3 wk (three to five mice/group). Tumors were allowed to grow for a total of 4–6 wk, at which time, all animals were sacrificed, their brains removed, frozen in O.C.T. compound, and sectioned. Immunohistochemical staining of the frozen brain sections for anti-human nuclei (unpublished data) or p75NTR (Figure 6A–6C) showed that animals implanted with U87p75NTR-, U251p75NTR-, or p75NTR-positive BTICs formed tumors with highly infiltrative edges (Figure 6A–6C; upper panels). In sharp contrast, animals implanted with U87p75NTR-, U251p75NTR-, or p75NTR-positive BTICs and given the γ-secretase inhibitor DAPT developed localized tumors with highly demarcated edges (Figure 6A–6C; lower panels). These results strongly suggested γ-secretase inhibition that results in blocking the generation of the p75NTR ICD substantially inhibited the invasive ability of glioma cells in vivo. To establish whether the effects of the γ-secretase inhibition were confined to glioma invasion or had consequences on other biological functions, we assessed the effect of γ-secretase inhibition on glioma cell proliferation in vivo and found no significant change (Figure S5). The γ-secretase inhibitor had negliable effects on the self-renewal capability of the BTIC line used in Figure 5 (unpublished data).
To confirm that administration of DAPT inhibited p75NTR ICD generation in vivo, 7–9 μm cryosections were lysed as described previously, and western blots for p75NTR were performed (Figure 6D). These experiments show the presence of full-length p75NTR, CTF, and ICD in control animals that received s.c. administration of vehicle (corn oil) alone. In contrast, western blots of cryosections from animals that received daily injections of the γ-secretase inhibitor DAPT detected only the full-length and CTF fragments of p75NTR. Subcutaneous administration of γ-secretase inhibitor DAPT inhibited the generation of the p75ICD and resulted in a visible accumulation of the CTF. In addition and most importantly, animals bearing U87p75NTR orthotopic xenographs and given daily s.c. injections of the γ-secretase inhibitor DAPT survived significantly longer (p < 0.0001) than control animals (Figure 6E), further highlighting the potential use of γ-secretase inhibitors in the clinical treatment of malignant glioma.
The p75NTR signaling cascade is a complex signaling axis that depends on numerous factors, including cellular context and specific protein interactions that influence biological outcomes to regulated intramembranous proteolysis of p75NTR. For example, a recent report showed that in primary cultures of cerebellar neurons, p75NTR ectodomain shedding and subsequent γ-cleavage is necessary for the growth inhibitory signal of myelin associated glycoprotein (MAG) [33]. Conversely, in retinal ganglion cells, neurotrophin-induced activation of p75NTR was shown to promote neurite growth in a RIP-dependent manner [66]. Ligand-dependent induction of p75NTR cleavage has also been reported in other cell systems, including sympathetic neurons and glial cells [28,30,34]. Similarly, we have shown that glioma migration is neurotrophin dependent [15] and that this neurotrophin-induced invasion is dependent on RIP of p75NTR.
To assess the role of RIP in p75NTR-mediated glioma invasion, we used a pharmacological and molecular approach both in vitro and in vivo to demonstrate that (1) p75NTR proteolytic processing occurs in glioma cell lines, surgically resected tumor specimens, and BTICs isolated from patient specimens; (2) cleavage-resistant alleles of p75NTR are insufficient to mediate glioma invasion; and (3) pharmacological inhibition with a clinically applicable γ-secretase inhibitor results in a dramatic decrease of glioma invasion both in vitro and in vivo and significantly prolonged survival of animals bearing p75NTR-positive intracranial tumors. Together, these data highlight the potential of using pharmacological inhibition to interfere with RIP as a therapeutic intervention for highly infiltrative p75NTR-positive gliomas.
One of the initial steps in regulating RIP is the shedding of the ECD by an α-secretase. This shedding event is required for subsequent cleavage of the CTF to generate an ICD. In order to show that both of these proteolytic events were important in the processing of p75NTR, we made a series of chimeric molecules with the Fas receptor, a related family member that does not generate CTF and ICD fragments [39,67]. The means by which the ECD is shed from the full-length p75NTR protein and what its biological role is are not fully understood [68–70]. Glioma cells are known to express many proteases, including serine, cysteine, and metalloproteinases that are involved in invasion and tumor progression. The ADAM metalloproteinase disintegrins, including ADAM17 and ADAM10, have been described as prominent sheddases for p75NTR as well as other transmembrane type-1 receptors such as APP [71–73], with recent in vivo evidence establishing a correlation with glioma invasion and an increase in ADAM17 under hypoxic stress [74,75]. The use of inhibitors targeting these proteinases may thus result in preventing RIP of p75NTR. Here, we have shown that the broad-range metalloproteinase inhibitor TAPI-2 was able to prevent both proteolytic processing of p75NTR in glioma and p75NTR-mediated invasion. Although a possible therapeutic strategy for highly invasive p75NTR-positive tumors, previous clinical attempts to inhibit the protease-rich environment of tumors using broad-spectrum MMP inhibitors have so far proven to be ineffective as anti-cancer agents, with phase II and III trials failing to show efficacy or survival benefit [76,77]. The reason for this lack of efficacy may result in part from the fact that glioblastomas produce high levels of proteases, many of which have been suggested to help facilitate tumor cell survival and invasiveness [74,78–83]. Attempts, therefore, to inhibit the ectodomain shedding of p75NTR in a clinical setting may prove difficult.
The second proteolytic event with possible direct therapeutic importance is mediated by the γ-secretase complex, which is composed of several proteins including presenilin, nicastrin, APH-1, and presenilin enhancer 2 (PEN-2) [84]. This protein complex is known to be essential in the normal processing of amyloid β-peptides from β-APP. Abnormal accumulation of amyloid β-peptides with the formation of plaques is believed to be the pathogenesis of Alzheimer disease. Given the connection between Alzheimer disease and γ-secretase, there has been great interest in developing compounds that can inhibit this protein complex with some of these compounds already in phase II/III clinical trials [85]. The exact molecular mechanism(s) by which the ICD fragment of p75NTR exerts the invasive behavior of glioma cells is unknown. As is the case with the Notch signaling pathway [41,86,87], there have been recent studies to suggest that the ICD fragment can translocate to the nucleus, but whether it acts as part of a transcriptional complex is unclear [28,30]. In addition, myelin-associated glycoprotein binding to cerebellar neurons induces α- and γ-secretase proteolytic cleavage of p75NTR, and the resulting ICD fragment is necessary for both the activation of the small molecular weight GTPase, RhoA, and inhibition of neurite outgrowth [33]. Whether these processes or others are regulated by the p75NTR ICD fragment in glioma cells remains to be determined.
In our present study, we show that neurotrophin-induced p75NTR proteolytic processing is required for p75NTR-mediated glioma invasion in vitro and in vivo. Furthermore, daily administration of the γ-secretase inhibitor DAPT to animals bearing p75NTR-postive intracranial tumors significantly prolonged survival. These results are intriguing and support the possible clinical application of γ-secretase inhibitors for the treatment of these deadly tumors. We cannot, however, exclude the possibility that we are inhibiting the processing of other proteins that may be involved in glioma invasion since γ-secretase is known to mediate the proteolytic processing of several transmembrane proteins [52,53,55,57–61]. The biochemical evidence presented here, however, supports the hypothesis that the anti-invasive effect of γ-secretase inhibition is at least in part the result of inhibition of p75NTR RIP. Moreover, the fact that we did not observe any significant effects on proliferation or survival of the human glioma cells in vivo suggests that the dominant mechanism of activity is the inhibition of p75NTR-mediated glioma invasion.
Excitingly, we also found that a large percentage (four out of five) BTICs from primary glioma patient tissue express high levels of p75NTR. This rare population of cells with stem-like properties and the ability to repopulate the tumor [3–6] have been shown to be resistant to our current therapies (radiation and temozolomide) and thus may represent a “disease reservoir” for these devastating tumors [88,89]. Unlike the U87 parental cells, these cells are highly invasive in vivo, and treatment with a γ-secretase inhibitor dramatically blocked their invasive nature (Figure 6). Several recent studies have demonstrated strong similarities between BTICs and neural stem and progenitor cells [4,90,91]. However, whether human glioblastoma stem cells arise from mutated neural stem cells or a more mature cell type that acquires self-renewal capacity remains to be determined. Interestingly, a small population of cells (0.3%) within the stem cell niche of the adult rat subventricular zone has neurosphere-forming capacity, express p75NTR [92], and appear to be maintained from birth through adulthood [93–95]. In addition, the more migratory p75NTR glioma cell population in clinical glioblastoma patient specimens also represents a small percentage of the main tumor mass [15]. It is intriguing to note that glioma cells that express high levels of p75NTR seem to possess many characteristics of BTICs, including self-renewal, extensive brain parenchymal migration, and potential for differentiation (J. J. P. Kelly and S. Weiss, unpublished data). Whether p75NTR is an early brain tumor stem cell marker, at least for some GBMs, remains to be determined.
In a previous study, we postulated that p75NTR itself may be a valid target for the treatment of glioma, and now we propose that abrogation of the cellular processing of p75NTR represents an additional therapeutic target. Although these inhibitors may have application in malignant glioma, they may have an even broader application for cancer, as p75NTR has also been implicated in other cancers, including melanoma, specifically the more aggressive melanomas that metastasize to the brain [96–98]. Thereby, therapies that target the processing of p75NTR may also be beneficial for other metastatic cancers.
The human glioma cell lines U87, U118, and U343 were obtained from the American Type Culture Collection. The human glioma cell line U251N was a kind gift from V. W. Yong (University of Calgary, Calgary, Alberta, Canada). All cells were maintained in complete media (Dulbecco's modified eagle's medium [DMEM] F12 supplemented with 10% heat-inactivated fetal bovine serum [FBS], 0.1 mM nonessential amino acids, 2 mM l-glutamine, and 1 mM sodium pyruvate [Gibco BRL, http://www.invitrogen.com]) at 37 °C in a humidified 5% CO2 incubator. Cells were passaged by harvesting with trypsin (Gibco BRL) at 80%–90% confluence. Stable transfectants of U87, U251, U118, and U343 cells were maintained in identical media with the exception of the addition of 400 μg/ml of G418 (Invitrogen, http://www.invitrogen.com).
The human p75NTR expression vector was constructed as described previously [98]. The expression plasmids containing the p75NTR mutants were constructed by subcloning of PCR fragments containing the desired p75NTR sequences. Chimeric proteins were created by replacing either the transmembrane (p75FasTM) or the extracellular stalk domain of p75NTR (p75FasS) with equivalent domains from the Fas receptor [Figure 3C] as described by Domeniconi et al. 2005 [33]. The neurotrophin-binding mutant that was mutated by a four-amino acid (ARRA) insertion after amino acid residue130 was termed p75CRD130 [15,42,43,47]. The p75NTR intracellular domain construct was created using amino acids 236–399 of the wild-type receptor plus a methionine at the amino terminus (p75-ICD). The original p75NTR templates were from B. Hempstead (p75WT; Weill Medical College of Cornell University, New York, New York) and M. Chao (pT3/T7-p75; New York University School of Medicine, New York, New York). All constructs were inserted into pcDNA 3.1 expression vectors (Invitrogen). The sequences of all the mutant expression plasmids were confirmed prior to stable transfection.
Transfection of glioma cell lines was performed as described previously [15]. Briefly, cells to be transfected were seeded at 2 × 105 cells/well in six-well plates, and incubated at 37 °C overnight. Vector DNA was introduced to the cells using FuGENE 6 transfection reagent (Roche Diagnostic, http://www.roche.com) according to the manufacturer's instructions. The following day, the medium was changed to fresh complete medium containing the antibiotic G418 (concentration determined by toxicity curve for each cell line) to select for those cells that had taken up the vector. The cells were then grown under antibiotic selection until the cells were at confluence. For U87p75NTR, U251p75NTR, U118p75NTR, U343p75NTR, U87p75FasTM, U87p75FasS, U87p75CRD130, U251p75FasTM, and U251p75FasS transfection, transfected cells were identified by flow cytometry and western blot.
The desired cells were washed in ice-cold PBS and lysed by gentle rocking in lysis buffer (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 10 mM NaF, 0.02% NaN3, 0.5% sodium deoxycholate, 0.1% SDS, 1% Nonidet P-40, 1% Triton X-100, 1 mM EDTA, 60 mM β-octyl glucoside, 25 μg/ml aprotinin, 10 μg/ml leupeptin, 3 mM sodium orthovanadate, 1 mM PMSF) at 4 °C. Cellular debris was removed by centrifugation, and protein quantification was performed using the bicinchoninic acid (BCA) assay (Pierce Biotechnology, http://www.piercenet.com). Proteins were resolved on 12% SDS-PAGE gels, and western blots were performed using the following primary antibodies: rabbit polyclonal anti-human p75NTR intracellular domain (Promega, http://www.promega.com), mouse monoclonal anti-human p75NTR ECD (Upstate Biotechnology), mouse monoclonal anti-β-tubulin (Sigma-Aldrich, http://www.sigmaaldrich.com), or mouse monoclonal anti-β-actin (Cell Signaling Technology, http://www.cellsignal.com). The appropriate HRP-conjugated secondary antibody (Pierce Biotechnology) was used, and blots were visualized using enhanced chemiluminescence (Amersham Biosciences).
A total of 1 × 106 cells stably transfected with p75NTR wild type or p75NTR cleavage-resistant constructs p75FasTM and p75FasS were collected using Puck's EDTA at 37 °C and then washed in PBS containing 1 mM EDTA (PBS/EDTA). Cells were exposed to the monoclonal anti-p75NTR, clone ME20.4 (which recognizes the extracellular domain; Upstate Biotechnology, http://www.upstate.com), diluted 1:250 in PBS/EDTA for 30 min on ice. Cells transfected with pcDNA vector alone were used as negative controls. After washing with PBS/EDTA, cells were incubated with Alexa-488 conjugated goat anti-mouse IgG (Invitrogen/Molecular Probes, http://www.probes.invitrogen.com) diluted 1:500 in PBS/EDTA for 30 min on ice. Cells were then washed with PBS/EDTA, resuspended in PBS/EDTA, and analyzed using a FACScan flow cytometer (Becton, Dickinson and Company, http://www.bdbiosciences.com).
U87 and U251 glioma cells stably transfected with p75NTR wild type or the cleavage-resistant constructs were allowed to condition culture medium for 5 d. The conditioned medium was then collected, centrifuged, and filtered through a 0.2-μm syringe filter (VWR International, http://www.vwr.com). The remaining cells were washed with ice-cold PBS, and total cellular lysates were extracted as described for western blot. Protein quantification was performed using the BCA assay (Pierce Biotechnology), and BDNF, nerve growth factor (NGF), or neurotrophic factor 3 (NT-3) ELISA (R&D Systems, http://www.rndsystems.com) was performed as per the company protocol. Briefly, MaxiSorp ELISA plates (Nalge Nunc International, http://www.nalgenunc.com.com) were coated with monoclonal anti-human BDNF, NGF, or NT-3 (R&D Systems), nonspecific binding was blocked, and then the standards of serial dilutions of recombinant human BDNF, NGF, or NT-3 (Sigma-Aldrich) and equal volumes of conditioned medium or equal quantities of lysate were added. Bound antigen was detected using the corresponding biotinylated antibody, streptavidin HRP, and a tetramethylbenzidine substrate (R&D Systems). Absorbance was measured at 450 nm.
For in vitro p75NTR cleavage assessment, the desired cells were treated for 4 h at 37 °C and 5% CO2 with the proteasome inhibitor epoxomicin (1 μM) (Calbiochem, http://www.emdbiosciences.com) and/or γ-secretase inhibitor Compound X (2 μM) (Calbiochem). DMSO was used as the vehicle control. Cells were then washed one time with cold PBS on ice, lysed in lysis buffer (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 10 mM NaF, 0.02% NaN3, 0.5% sodium deoxycholate, 0.1% SDS, 1% Nonidet P-40, 1% Triton X-100, 1 mM EDTA, 60 mM β-octyl glucoside, 25 μg/ml aprotinin, 10 μg/ml leupeptin, 3 mM sodium orthovanadate, 1 mM PMSF) at 4 °C with protease inhibitors, and centrifuged for 5 min at 14,000×g; supernatants were quantified by BCA assay (Pierce Biotechnology) for use in SDS-PAGE. The CTF (24 kDa) and ICD (19 kDa) fragments were detected using an antibody specific to the p75NTR ICD (Promega).
Migration assays were performed using a microliter-scale radial monolayer migration assay as described previously [15]. Briefly, ten-well Teflon-masked microscope slides were coated with 20 μg/ml laminin, followed by the addition of 50 μl of medium to each well. Sedimentation manifold (Creative Scientific Methods, http://www.creative-sci.com) was placed over the laminin-coated slide. Cells were seeded through the central lumen of the cell sedimentation cylinder at 2,000 cells/well (five wells per cell type/condition) to establish a circular 1-mm diameter confluent monolayer. Slides were placed on ice for 60 min and then incubate at 37 °C for approximately 6 h. Attachment of the cells was confirmed prior to removing the sedimentation manifold. Once the sedimentation manifolds were removed, cells were given complete medium containing the γ-secretase inhibitor (Compound X, 2 μM). A digital image of the cells was taken (before migration = 0 h) using a Zeiss Axiovert 200M inverted fluorescent microscope (Carl Zeiss, http://www.zeiss.com). The cells were then incubated in a humidified chamber at 37 °C and 5% CO2, and a second digital image was taken 48 h later. Best-fit circles were drawn around the area covered by the cells at the 0 h and 48 h time points, and the actual cell area was determined using Axiovision 4.2 imaging software (Carl Zeiss). Quantitative migration scores were calculated as the increase in the area covered by the cells beyond the initial area of the cells.
To test the invasive ability of the p75NTR cleavage-resistant constructs, actively growing glioma cells U87 and U251 stably transfected with p75NTR and p75NTR cleavage-resistant constructs were suspended in a collagen gel solution and plated in transwell chambers with 8.0-μm pore size polycarbonate membrane (Costar, http://www.costar.com). The collagen gel was prepared by mixing collagen solution (Chemicon International 3D Collagen cell culture system, Cat# ECM675, http://www.chemicon.com) with 5× DMEM F12 medium on ice. Neutralization solution (40:1) and extracellular matrix (ECM) proteins were added at a concentration of 10 μg/ml (laminin, fibronectin, chondroitin sulfate proteoglycan, Chemicon). A total of 1 × 105 cells were suspended in 350 μl of collagen gel solution, and 70 μl of the collagen/cell mixture was pipetted into the Transwell chambers (five chambers for each cell line). Chambers were immediately transferred to a 37 °C incubator for 60 min to allow the matrix to polymerize. Once polymerized, 100 μl of serum-free DMEM was added to the upper chamber and 1.0 ml of 10% FBS complete medium with the γ-secretase inhibitor Compound X (2 μM) was added to the lower chamber. Transwell chambers were kept at 37 °C for 6 h, at which time the chamber was washed with PBS, fixed with acid-alcohol for 15 min at room temperature, and then stained with hematoxylin. Any cells remaining in the top chamber were removed, and membranes were mounted on glass slides. Four different fields were counted for each membrane.
Six- to 8-wk-old female SCID mice were purchased from Charles River Laboratories (http://www.criver.com). The animals were housed in groups of three to five and maintained on a 12-h light/dark schedule with a temperature of 22 °C ± 1 °C and a relative humidity of 50% ± 5%. Food and water were available ad libitum. All procedures were reviewed and approved by the University of Calgary Animal Care Committee.
Actively growing glioma cells stably transfected with p75NTR and p75NTR cleavage resistant constructs were harvested by trypsinization, washed, and resuspended in sterile PBS(137 mM NaCl, 8.1 mM Na2HPO4, 2.68 mM KCl, and 1.47 mM KH2PO4 [pH 7.5]). These cells were implanted intracerebrally into the right putamen of SCID mice (1 × 105 cells/mouse) at a depth of 3 mm through a scalp incision and a 0.5-mm burr hole made 1.5–2 mm right of the midline and 0.5–1 mm posterior to the coronal suture. All mice were anaesthetized by intraperitoneal administration of ketamine (85 mg/kg) plus xylazine (15 mg/kg) (MTC Pharmaceuticals). The stereotactic injection used a 5-μl syringe (Hamilton Co., www.hamiltoncompany.com) with a 30-g needle mounted on a Kopf stereotactic apparatus (Kopf Instruments). After withdrawal of the needle, the incision was sutured. Animals were sacrificed at specific time points (generally weekly, from 2–6 wk postinjection) or when they lost 20% of their body weight or had difficulty ambulating, feeding, or grooming. For some experiments, BrdU was given by intraperitoneal injection 24 h prior to sacrifice. Following sacrifice, the brains were removed, frozen in Tissue-Tek O.C.T. compound (Electron Microscopy Sciences, http://www.emsdiasum.com), and cryo-sectioned into 7–9-μm sections for examination by immunohistochemistry and in vivo p75NTR proteolysis assessment.
Frozen sections were air-dried at room temperature, fixed with cold acetone, and then rinsed with PBS. Endogenous peroxidases in the sections were inactivated with 0.075% H2O2/methanol, and nonspecific binding was blocked with 10% normal goat serum in PBS. The sections were incubated with rabbit polyclonal anti-human p75NTR ICD antibody (Promega) or mouse monoclonal anti-human nuclei (Chemicon) in blocking buffer overnight at 4 °C. Following washing with PBS, the appropriate biotinylated secondary antibody (Vector Laboratories, http://www.vectorlabs.com) was applied. Avidin-biotin peroxidase complexes were then formed using the VECTASTAIN Elite ABC kit (Vector Laboratories) and detected by addition of SIGMAFAST DAB (3,3′-diaminobenzidine tetrahydrochloride) (Sigma-Aldrich), which was converted to a brown reaction product by the peroxidase. Toluidine blue (for frozen sections) was used as a nuclear counterstain. Sections were then dehydrated in an ethanol/xylene series and mounted with Entellan (Electron Microscopy Sciences).
For detection of p75NTR proteolytic processing in vivo, the desired cells were implanted intracerebrally into SCID mice as described previously. Mice were sacrificed 3–4 wk later. Following sacrifice, the brains were removed, frozen in Tissue-Tek O.C.T. compound (Electron Microscopy Sciences), cryosectioned into 7–9-μm sections, and alternating sections were stained with toluidine blue. Based on the size of tumor, cryosections were lysed in 2× loading buffer (0.1 M Tris-HCl [pH 6.8], 4% SDS, 20% glycerol, 10% β-mercaptoethanol, 0.02% bromphenol blue). Proteins were resolved on 12% SDS-PAGE gels, and western blots were performed using an anti-p75NTR cytoplasmic specific antibody (Promega).
Tumor and normal tissues were obtained from the Canadian Brain Tumor Tissue Bank in London, Ontario, and the Brain Tumor Tissue Bank at the Clark Smith Brain Tumor Center within the Southern Alberta Cancer Research Institute. Briefly, tissue was taken during surgery while patients were under a general anesthetic, and was placed immediately in liquid nitrogen and stored at −80 °C. An institutional ethics board approved the collection and use of all of the surgical tissue used, and all of the patients gave signed informed consent. Frozen sections of patient tumor tissue were lysed in lysis buffer (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 10 mM NaF, 0.02% NaN3, 0.5% sodium deoxycholate, 0.1% SDS, 1% Nonidet P-40, 1% Triton X-100, 1 mM EDTA, 60 mM β-octyl glucoside, 25 μg/ml aprotinin, 10 μg/ml leupeptin, 3 mM sodium orthovanadate, 1 mM PMSF) with protease inhibitors on ice using a homogenizer (Life Technologies). Lysates were centrifuged for 5 min at 14,000×g to remove debris, and supernatants were quantified by BCA assay (Pierce Biotechnology). In this study, based on our previous immunohistochemistry results of p75NTR expression in GBM patient specimens, western blot analysis was performed on nine GBMs and five mid-grade glioma samples using an antibody specific for the intracellular domain of p75NTR.
Tumor and normal tissues were obtained from the Tumor Tissue Bank in Foothills Hospital, Calgary, Alberta. Operative samples of human gliomas were obtained during brain tumor surgery and transported to the laboratory in serum-free DMEM-F12. Primary cultures of brain tumor-initiating cells (BTICs) were established. Briefly, necrotic and connective tissue and any blood clots were removed using forceps, and the remaining tissue was washed in PBS and cut into pieces of approximately 1 mm3. The tissue was then incubated for 5–10 min at 37 °C in an enzyme cocktail of trypsin (0.25%) and DNase I (10 μg/ml) in PBS. The digested tissue was strained through a 100-μm mesh and washed with PBS. Following lysis of the red blood cells, the remaining cells were washed with PBS and strained through a 40-μm mesh. After spinning at 1,000 rpm for 5 min, cells were resuspended in stem cell medium (M medium) or stem cell medium plus EGF and bFGF (EF medium). M medium is NeuroCult NS-A basal medium (human) 450 ml plus NeuroCult NS-A Proliferation Supplements (human) 50 ml (StemCell Technologies). EF media is M media plus human recombinant EGF (20 ng/ml; Sigma) and bFGF (20 ng/ml; Chemicon).
Eight-well LAB-TEK chamber slides (Nalgel Nunc, http://www.nuncbrand.com) were coated with poly-l-ornithine (Sigma) and incubated at 37 °C for 1 h. The desired BTICs were plated into chambers with stem cell culture medium to equilibrate overnight at 37 °C, 5% CO2. Chambers were then rinsed with PBS, fixed in 3.7% paraformaldehyde diluted in PBS for 20 min, and rinsed twice with PBS. Anti-p75NTR cytoplasmic specific antibody (1:3,000, Promega) and antibodies to the progenitor markers: Nestin (1:1,000), hSOX2 (1:5,000), and mushashi (1:200) (R&D System) were diluted in 0.3% Triton X-100/PBS/10% goat serum, and 200 μl of these solutions were added to each chamber, incubated overnight at 4 °C. Following washing with PBS, the appropriate Cy-3– or FITC-conjugated secondary antibodies (1:2,000) (Cedarlane, http://www.cedarlanelabs.com) were applied and incubated for 30 min in the dark at room temperature. The chambers were removed, and slides were mounted with DAPI counterstained mounting medium (Vector Laboratories) and imaged with an Olympus IX70 Delta Vision RT microscope and the SoftWoRx software package.
Based on the immunocytochemistry staining results, BTICs were collected, dissociated by polished glass pipette, aliquoted into six-well plates, and treated with the proteasome inhibitor epoxomicin (1 μM) (Calbiochem) and/or the γ-secretase inhibitor Compound X (2 μM) (Calbiochem), or DMSO vehicle alone, for 4 h at 37 °C and 5% CO2. Cells were then washed twice with cold PBS on ice, lysed in lysis buffer with protease inhibitors, and quantified by BCA assay. Western blots for p75NTR using a p75NTR cytoplasmic-specific antibody were performed.
Actively growing glioma cell line U87p75 and BTICs were implanted intracerebrally into SCID mice as described previously. Three days later for the U87p75 cell line, and 5 d later for BTICs, mice were administered s.c. vehicle (corn oil) or 10 mg/kg γ-secretase in corn oil once/day for 2 wk (U87p75NTR) or for 3 wk (BTIC). At 3 wk for animals bearing U87p75NTR tumors or 4 wk for animals bearing BTIC tumors, animals were sacrificed, the brains were removed, frozen in Tissue-Tek O.C.T. compound, and cryosectioned into 7–9-μm sections. The cryosections were used for tumor immunohistochemistry staining and for assessment of in vivo p75NTR cleavage. For survival studies (eight animals per group), U87p75NTR glioma cells were implanted intracerebrally into SCID mice. Animals were given daily s.c. injections of 10 mg/kg γ-secretase inhibitor DAPT or vehicle (corn oil) alone, once/day beginning on day 3. Animals were followed until sacrifice was required.
Statistical analysis of data was performed using GraphPad Prism software (GraphPad Software). Survival curves were generated using the Kaplan-Meier method. The log-rank test was used to compare the distributions of survival times. A p-value of less than 0.05 was considered statistically significant.
|
10.1371/journal.pntd.0005900 | Genotypic characterization directly applied to sputum improves the detection of Mycobacterium africanum West African 1, under-represented in positive cultures | This study aimed to compare the prevalence of Mycobacterium tuberculosis complex (MTBc) lineages between direct genotyping (on sputum) and indirect genotyping (on culture), to characterize potential culture bias against difficult growers.
Smear-positive sputa from consecutive new tuberculosis patients diagnosed in Cotonou, (Benin) were included, before patients had started treatment. An aliquot of decontaminated sputum was used for direct spoligotyping, and another aliquot was cultured on Löwenstein Jensen (LJ) medium (90 days), for indirect spoligotyping. After DNA extraction, spoligotyping was done according to the standard method for all specimens, and patterns obtained from sputa were compared versus those from the derived culture isolates. From 199 patient’s sputa, 146 (73.4%) yielded a positive culture. In total, direct spoligotyping yielded a pattern in 98.5% (196/199) of the specimens, versus 73.4% (146/199) for indirect spoligotyping on cultures. There was good agreement between sputum- and isolate derived patterns: 94.4% (135/143) at spoligotype level and 96.5% (138/143) at (sub)lineage level. Two of the 8 pairs with discrepant pattern were suggestive of mixed infection in sputum. Ancestral lineages (Lineage 1, and M. africanum Lineages 5 and 6) were less likely to grow in culture (OR = 0.30, 95%CI (0.14 to 0.64), p = 0.0016); especially Lineage 5 (OR = 0.37 95%CI (0.17 to 0.79), p = 0.010). Among modern lineages, Lineage 4 was over-represented in positive-culture specimens (OR = 3.01, 95%CI (1.4 to 6.51), p = 0.005).
Ancestral lineages, especially M. africanum West African 1 (Lineage 5), are less likely to grow in culture relative to modern lineages, especially M. tuberculosis Euro-American (Lineage 4). Direct spoligotyping on smear positive sputum is effective and efficient compared to indirect spoligotyping of cultures. It allows for a more accurate unbiased determination of the population structure of the M. tuberculosis complex.
ClinicalTrials.gov NCT02744469
| The vast majority (95%) of tuberculosis (TB) patients worldwide live in low-income countries, including in West-Africa. Typing the bacteria responsible for TB (tuberculosis; Mycobacterium tuberculosis complex) is important for targeted TB control. Typing is usually performed on isolates obtained after the culture isolation of TB bacteria in the sputa from patients. However, cultures can be false negative, and some ‘ancestral’ strains, only found in West-Africa (Mycobacterium africanum), require more time (90 days versus the usual 56 days) to grow in culture. To characterize potential culture bias against such “difficult growers”, we compared the performance of direct typing (on sputum) relative to its yield on culture isolates. We found that ancestral types of TB bacteria were significantly less likely to grow in culture despite the 90-day incubation. This suggests that typing results of cultured isolates are not representative of the diversity in the population of TB bacteria causing disease in patients. Typing sputum directly is effective and can be used for a more precise, unbiased determination of the proportion of different TB bacteria in a population.
| Tuberculosis (TB), caused by bacteria of the Mycobacterium tuberculosis complex (MTBc), remains a public health problem. Globally, over 8 million new patients with TB disease arise each year, including 2 million deaths. The vast majority (95%) of global TB is detected in limited-resource countries [1], including West-Africa. Each year in Benin, over 4000 cases of TB are detected, and the incidence of smear-positive pulmonary TB is 39 per 100000 inhabitants.
Genotypic characterization is important in order to understand the population structure of the MTBc for better insights into endemic- and epidemic strains and to identify instances of nosocomial transmission or laboratory contamination. M. tuberculosis sensu stricto and M. africanum sub-species within the MTBc have been subdivided into 7 main lineages of human importance [2,3]. These 7 MTBc lineages are classified as ancestral (or ‘ancient’) (Lineages 1, 5, 6) [4,5], intermediate (Lineage 7) [3,4] and modern lineages (Lineages 2,3,4) [4]. Lineage 5 (M. africanum West African 1) and Lineage 6 (M. africanum West African 2) are only found in West- and Central Africa, where they cause up to 40% of all TB [6,7]. Recent reports suggested a decrease in prevalence of M. africanum in some West-African countries [8–10]. Whether methodological issues explain the apparent disappearance of M. africanum has not been excluded to date.
For the determination of the population structure of the MTBc, genotyping is usually applied on culture isolates [11]. M. africanum grows significantly slower than the other members of the MTBc (M. tuberculosis sensu stricto) [12] and cultures should be incubated for 90 days rather than the usual 56 days, before reporting a negative result [13]. However, even this extended incubation time may not permit recovery of M. africanum isolates at the same rate as M. tuberculosis, and thus bias the population structure derived from cultured isolates, especially in settings where M. africanum is endemic. Differences in expression of genes involved in metabolism pathways of the various MTBc lineages may also affect their growth in culture, as recently reported for M. africanum Lineage 6 which has an under-expression for the gene (Dos R) involved in adaptation to lower oxygen tension relative to Lineage 4 [14]. For isolation, of some MTBc species, including M. africanum, the need for pyruvate to support growth in culture [15] has been known for a long time [16].
Few studies evaluated genotyping, such as spoligotyping, directly on clinical specimens such as sputa [17,18], sputum smears [19], paraffin wax-embedded tissues [20] or mummified remains of human [20]. Only one study from Brazil, where M. africanum is not endemic, compared spoligotyping on sputum to spoligotyping from the respective isolates [21]. Moreover, to the best of our knowledge, no study has investigated whether the proportional prevalence of MTBc lineages differs among specimens with a positive culture versus culture-negative specimens.
In this study, we determined the performance of spoligotyping on sputum (‘direct spoligotying’) relative to its yield on culture (‘indirect spoligotyping’) for genotypic characterization of MTBc, and evaluated for a potential culture bias against difficult-growers, even when incubation was prolonged to enhance detection of M. africanum.
This study is part of the BeniDiT study that has been approved by the national ethics committee of Benin, the Institutional Review Board of the Institute of Tropical Medicine of Antwerp, Belgium and the ethics committee of the University of Antwerp. It is registered on ClinicalTrials.gov under the registration number NCT02744469. All sputa were anonymized before laboratory analyses.
Smear-positive sputa from consecutive new TB patients diagnosed in the Centre National Hospitalier Universitaire de Pneumo-Phtisiologie in Cotonou, (Benin) were prospectively included (Fig 1), before patients initiated TB treatment. Laboratory analyses were conducted in the National Reference Laboratory for Mycobacteria (Laboratoire de Référence des Mycobactéries) in Cotonou, Benin.
Spoligotype patterns were recorded in an Excel file using a binary code (1 for presence of a given spacer and 0 for the absence of a given spacer). Entered profiles were verified and validated by an independent person. The persons who typed and validated the data were blinded to the spoligotype pattern of the corresponding sputum or isolate. The Excel file was loaded into the TBlineage database http://tbinsight.cs.rpi.edu/run_tb_lineage.html [26] for lineage assignment. Sub-lineages (spoligotype families) were obtained by loading the Excel file with the spoligotype patterns in the SPOTCLUST database http://tbinsight.cs.rpi.edu/run_spotclust.html [27].
Data was analyzed using the statistical software Stata/IC 12.0 (StataCorp). The two-group proportion test or the Fisher Exact test was used to analyze independent data. Mc Nemar Chi2 test was used to compare paired proportions. Two-sided p-values were calculated and for differences in proportion, odds ratios were calculated along with 95% confidence interval. Differences were considered statistically significant when p<0.05.
From the 199 recruited TB patients and their sputum samples, 146 (73.4%) yielded a positive culture, whereas 36 (18.1%) remained negative and 17 (8.5%) were contaminated. Spoligotype patterns were obtained for all the 146 culture isolates, and for 196 of the 199 sputa, yielding an overall success for direct spoligotyping of 98.5%. All of the extraction controls and amplification/hybridization controls yielded expected results, and repeat spoligotyping for discordant results between sputum and culture confirmed the original patterns. Stratified by culture result, direct spoligotyping reached a success of 100% (53/53) for culture-negative or contaminated sputa, and 98% (143/146) for culture-positive sputa. Microscopy was negative in 6 sediments after decontamination, while all the others had positive microscopy. Of the 6 microscopy negative sediments, 3 failed direct spoligotyping and 2 others had a negative culture.
Spoligotype patterns were available for 98.5% of sputa versus 73.4% of cultures (Table 1).
Comparison between respective direct and indirect spoligotypes showed 94.4% (135/143) agreement. In total three types of discrepancies were observed (Fig 2): mixed infection with one pattern found in sputum and the other found in the culture isolate (n = 3, discrepancy 5–7), mixed infection with overlapping spoligotype patterns in sputum (n = 2, discrepancies 1 and 4), and false negative (missing) spacers in sputum (n = 3, discrepancies 2, 3 and 8). Five (5) of these patterns led to inter-lineage discrepancies, and three (3) to intra-lineage discrepancies.
For inter-lineage discrepancies, sub(lineages) observed in isolates are shaded grey. The five inter-lineage discrepant pairs (discrepancy 4–8) showed patterns suggestive of a simultaneous presence of ancestral and modern lineages, while these yielded only the ancestral lineage in sputum and only the modern lineage in culture. Three (discrepancy 5–7) of these five inter-lineage pairs showed this (ancestral M. africanum in sputum and modern Lineage 4 in culture), without any other possible explanation, while the other two (discrepancy 4 and 8) can also be interpreted as follows. Inter-lineage pair 8 and intra-lineage pairs 2 and 3 showed patterns suggestive of false negative spacers in sputum (spacer present in isolate but absent in sputum). Intra-lineage pair 1 and inter-lineage pair 4 showed patterns suggestive of overlapping spoligotype signatures in sputum (discrepancy 1 and 4) and/or in isolate (discrepancy 1). Discrepancy 4 suggested an overlapping of Lineages 2 and 4 signatures in sputum, with only the Lineage 2 grown in culture. Discrepancy 1 was suggestive of overlapping spoligotype signatures in sputum and in culture isolate that could be a mixture of Lineages 2 and 4.
The distribution of lineages in culture-positive sputa versus directly in sputum with unsuccessful culture differed, with Lineage 5 (M. africanum West African 1) being significantly less prevalent among culture-positive sputa (OR = 0.48 95%CI (0.24 to 0.94) p = 0.033, Table 2). This association became more significant when contaminated cultures were excluded from the analysis (OR = 0.37, 95%CI (0.17 to 0.8), 21% vs 41.7%, p = 0.011, Table 2). Ancestral lineages (Lineages 1, 5 and 6) were significantly less present among culture-positive sputa (OR = 0.33, 95%CI (0.16 to 0.7), 37.1% vs 63.9%, p = 0.004, Table 2). Lineage 4 (M. tuberculosis Euro-American), a modern lineage, was most overrepresented in culture-positive sputa (OR = 2.81, 95%CI (1.30 to 6.03)55.2% vs 30.5%, p = 0.008, Table 2).
Excluding discrepant spoligotypes between direct and indirect spoligotype analysis, the association gained further statistical significance. The odds of detecting ancestral lineages in positive-cultures was 0.30 fold (95% CI (0.14 to 0.64); p = 0.0016) less in positive-cultures relative to negative cultures, especially Lineage 5 (OR = 0.37 95%CI (0.17 to 0.79); p = 0.010) (S1 Table). Modern lineages were inversely more represented in positive-culture specimens (OR = 3.31, 95%CI (1.57 to 6.99), p = 0.0016), especially Lineage 4 (OR = 3.01, 95%CI (1.4 to 6.51), p = 0.005) (S1 Table).
The prevalence of L1, L5, L6 tended to be higher among culture-negative specimens (respectively 8.3%, 41.7%, 13.9%; S1 Table) than in culture-positive specimens (7.4%, 20.7%, 6.7%; S1 Table). In contrast the prevalence of L2, L3, L4 tended to be lower among culture-negative specimens (5.6%, 0%, 30.5%) than in culture-positive specimens (6.7%, 1.5%, 57.0%; S1 Table). This justified the analysis in subgroup of ancestral and modern lineages. The distribution of sub-lineages (families) within Lineage 4 showed that LAM 10, LAM 9, LAM 1, T1, T2, Haarlem 1, Haarlem 2, Haarlem 3, X3 families were present in new TB patients in Cotonou. This distribution of Lineage 4 families did not differ significantly in culture-positive versus culture-negative sputa (S2 Table).
Almost all positive cultures were positive within 8 weeks of incubation, while prolonged incubation only yielded one additional positive culture. This was a Lineage 5/ M. africanum West African 1 strain.
Among positive cultures, over half (5/9: 55.5%) of the Lineage 6/ M. africanum West African 2 cultures became positive between 6 to 8 weeks of incubation, whereas most of positive cultures from other lineages specimens were positive within 6 weeks: 10/11 (90.9%) for Lineage 1, 10/10 (100%) for Lineage 2, 2/2 (100%) for Lineage 3, 83/85 (97.6%) for Lineage 4 and 28/29 (95.6%) for Lineage 5.
Despite the prolonged incubation period, over a third of specimens from each M. africanum lineage remained culture negative (34.1% for Lineage 5 and 35.7% for Lineage 6), while for other lineages, none (Lineage 3) or fewer specimens (21.4% for Lineage 1, 16.7% for Lineage 2, 11.5% for Lineage 4) remained negative (Table 3). The sediment smear of the culture negative specimens from Lineage 1 and 2 had low AFB-grading or were negative whereas nearly all (14/15) the culture negative specimens from Lineage 5 had high smear grading (S3 Table).
Our results show that indirect spoligotyping provided spoligotype profiles for all 146 culture-positive specimens (73.4%), while direct spoligotyping provided spoligotyping profiles for 50 more sputa (+ 25.1% of all 199 specimens, 95% CI (18.1% to 32.1%)) that would not otherwise be genotyped in the absence of an isolate. Direct spoligotyping on sputum after semi-automated DNA extraction using Maxwell DNA tissue purification kit, has a high sensitivity (98.5% (196/199)) to detect MTBc genotypes. The 98.5% (196/199) overall availability of spoligotype profiles in our study is higher than the 90.9% (159/175) found on smear-positive sputa by Goyal et al. in Ghana (p = 0.001) [18] and the 49.1% (28/57) found by Heyderman et al. in Zimbabwe [17]. This could be explained by the variability of methods used for DNA extraction from sputa and/or the variability in PCR reagents mix. The overall availability of spoligotype profiles on sputa in our study (98.5%) is also higher than the 77.7% (41/53) found by Suresh et al. and 90.5% (19/21) by Zanden et al. on smears [19,28], which likely have less mycobacterial DNA than a 200 μL sputum sample.
The fact that- within mixed infections- ancestral lineages are found with direct spoligotyping on sputum, suggests that the load of ancestral lineage bacilli in vivo exceeds the load of the modern lineage bacilli, with subsequent out-competition in culture by the latter. Sarkar et al. also found that Lineage 4 grows more rapidly (in liquid medium) than other lineages including Lineage 1, an ancestral lineage [29]. Moreover, Gehre et al. found that Lineage 6, another ancestral lineage, grows more slowly than MTBc lineages other than M. africanum in liquid medium [12].
Sputum provided the most representative population distribution of lineages of the MTBc in new TB patients in Cotonou, with more TB due to ancestral lineages, including M. africanum. This distribution did not alter when the three isolates which sputum failed direct spoligotyping were added (two from Lineage 4 and one from Lineage 5; Table 2). The ‘most true’ distribution is the one combining profiles obtained directly from sputum, complemented by profiles on isolates from failed direct spoligotyping, and includes: 8.0% (16/199) for Lineage 1, 5.6% (11/199) for Lineage 2, 1% (2/199) for Lineage 3, 51.8% (103/199) for Lineage 4, 25.1% (50/199) for Lineage 5, 8.5% (17/199) for Lineage 6, or 41.7% for ancestral lineages, and 33.7% for M. africanum (Table 2). This distribution would have been different if smear-negative specimens were also genotyped, as it had been previously reported that M. africanum is more likely to be found in lower grade smear-positive specimens [30], and Lineage 6 is associated with HIV infection [31], which is in turn associated with smear-negativity [32–34].
The comparison of the distribution of MTBc lineages in a similar population, also consisting of consecutive smear-positive new pulmonary TB patients aged at least 15 years old of Cotonou in year 2005–2006 on cultured isolates [9,35], to the one obtained in our study indirectly on cultured isolates from similar patients in Cotonou 10 years later, showed that the previous prevalence of Lineage 1 (7.7%), Lineage 2 (10.3%), Lineage 3 (0%), Lineage 6 (6.2%) are similar to our findings in this study (respectively: 7.5%, 6.8%, 1.4% and 6.2%). Yet the prevalence of Lineage 4 (42.3% in year 2005–2006) has increased to 58.2% (difference: +15.9%), and Lineage 5 prevalence (30.9% in year 2005–2006) has decreased to 19.9% (difference: -11.0%). While we demonstrate that the present L5 prevalence of 19.9% on indirect genotyping is an underestimate, even the present ‘true’ L5 prevalence of 25.1% on direct genotyping would constitute a decline from the L5 prevalence of 30.9% on indirect genotyping in 2005–2006. Other authors also reported a decrease of M. africanum [8,9].
Our results show that rates of culture isolation from smear-positive pulmonary TB patients are lower for Lineages 5 and 6 of the MTBc, despite prolonged incubation of cultures for 90 days [13]. Extending the incubation time beyond 6 weeks enhanced isolation of Lineage 6 (between 6–8 weeks) yet did not further augment the isolation rate. Ancestral lineages, especially Lineage5/M. africanum West African 1 are ‘difficult-growers’ in culture relative to modern lineages, such as Lineage 4. The decreased odds of ancestral lineages to grow in culture could partly be due to culture procedures (culture medium or decontamination method) that were originally developed for modern lineages prevalent in Europe. Ofori-Anyinam et al. reported that Lineage 6 as compared to Lineage 4, is more adapted to microaerobic growth [14] which may be the reason for its impaired growth on solid media such as LJ used in this study. Furthermore Gehre et al. found that Lineage 6 has mutations in genes that lead to its attenuated growth in vitro[12]. Such genetic analyses need to be conducted on Lineage 5 in order to understand the reasons for its difficult growth in vitro. Further studies should also be conducted on other lineages to find out the genetic basis of their in vitro growth pattern. To the best of our knowledge, this is the first demonstration that ancestral lineages are underrepresented in positive cultures. Direct spoligotyping is thus more appropriate for unbiased determination of MTBc population structure in settings where ancestral lineages, including M. africanum, are common.
The implications of our findings also affect MTBc population structures generated with different typing methods, including whole genome sequencing. Such studies tend to be culture-based, given the ongoing limitations of sequencing entire MTBc genomes directly from clinical material. While direct genome sequencing is technically feasible given sufficient coverage, in practice the associated costs are prohibitive. Studies to date have shown limited coverage, precluding SNP cut-offs for molecular epidemiological studies [36]. Optimized methods to sequence genomes directly from clinical material are thus urgently needed.
One strength of this study is the prolonged incubation time, to maximize the yield of M. africanum in culture. Other strengths include the paired design for the comparison of direct spoligotyping versus indirect spoligotyping and the inclusion of multiple controls and blinding of operators, and the fact that the study was conducted in a setting where M. africanum is prevalent. A limitation is that only LJ medium was used, and we do not know whether other medium, such as liquid medium (known to enable the growth of more non-tuberculous mycobacteria) may also favor the growth of ancestral MTBc lineages. This study was conducted only on fresh unshipped acid-fast bacilli positive sputa from new TB patients. Culture positivity may be worse if sputa had to be shipped from peripheral laboratories to a reference or central laboratory where spoligotyping can be done. Another limitation is that the number of specimens with Lineages 1, 2, 3, 6 among culture negative specimens under-powered the estimation of any difference in the prevalence of these individual lineage among culture-negative versus culture-positive specimens. So, although no evidence of such difference in prevalence among culture-negative versus –positive specimens was found in Lineages 1, 2, 3, 6 in the present study, such difference could be tested for in settings with higher prevalence of these lineages.
In conclusion, ancestral lineages especially M. africanum West African 1 (Lineage 5), are less likely to grow in culture, unlike modern lineages especially M. tuberculosis Euro-American (Lineage 4). Direct spoligotyping on sputum is effective, and saves effort and time compared to indirect spoligotyping of cultures. It has an important gain in sensitivity, especially for ancestral lineages that may not yield a positive culture, allowing a more precise unbiased determination of the population structure of the MTBc. It can also be used for specimens from patients under TB treatment and other specimens in which culture may be negative or contaminated. While differences in culture isolation technique and reliance on indirect spoligotyping may partially account for the reduction in the prevalence of M. africanum observed in several West African countries [8,9], comparison of our findings with the genotyping study from Cotonou 10 years ago suggests that the decline in M. africanum is not explained by the lower sensitivity of culture isolation. The potential decline of M. africanum lineages will be addressed in more depth in a larger ongoing study on the population structure of the M. tuberculosis complex in Benin, in which direct genotyping will be applied, given the findings presented in this manuscript. Further studies must be conducted to investigate whether culture procedures (medium, decontamination) can be optimized for growth of ancestral lineages. Additional studies should address the frequency and role, if any, of a mixed infection between an ancestral- and modern lineage in the faster spread of modern lineages [4] and disappearance of ancestral lineages [8,9].
|
10.1371/journal.pcbi.1004749 | The Role of Adherence and Retreatment in De Novo Emergence of MDR-TB | Treatment failure after therapy of pulmonary tuberculosis (TB) infections is an important challenge, especially when it coincides with de novo emergence of multi-drug-resistant TB (MDR-TB). We seek to explore possible causes why MDR-TB has been found to occur much more often in patients with a history of previous treatment. We develop a mathematical model of the replication of Mycobacterium tuberculosis within a patient reflecting the compartments of macrophages, granulomas, and open cavities as well as parameterizing the effects of drugs on the pathogen dynamics in these compartments. We use this model to study the influence of patient adherence to therapy and of common retreatment regimens on treatment outcome. As expected, the simulations show that treatment success increases with increasing adherence. However, treatment occasionally fails even under perfect adherence due to interpatient variability in pharmacological parameters. The risk of generating MDR de novo is highest between 40% and 80% adherence. Importantly, our simulations highlight the double-edged effect of retreatment: On the one hand, the recommended retreatment regimen increases the overall success rate compared to re-treating with the initial regimen. On the other hand, it increases the probability to accumulate more resistant genotypes. We conclude that treatment adherence is a key factor for a positive outcome, and that screening for resistant strains is advisable after treatment failure or relapse.
| Our ability to treat and control acute pulmonary tuberculosis (TB) is threatened by the increasing occurrence of multi-drug-resistant tuberculosis (MDR-TB) in many countries around the globe. It is not clear whether MDR-TB occurs predominantly due to transmission, or whether there is a substantial contribution due to de novo emergence during treatment. Understanding the underlying mechanisms that are involved in the emergence of MDR-TB is important to develop countermeasures. We use a computational model of within-host TB infection and its subsequent treatment to qualitatively assess the risks of treatment failure and resistance emergence under various standard therapy regimes. The results show that especially patients with a history of previous TB treatment are at risk of developing MDR-TB. We conclude that de novo emergence of MDR-TB is a considerable risk during treatment. Based on our findings, we strongly recommend widespread implementation of drug sensitivity tests prior to the initiation of TB treatment.
| Tuberculosis (TB) is a key challenge for global health [1,2]. At present about one third of the global population is latently infected [3] and every year about 1.7 million people die of tuberculosis. A large number of patients live in resource-limited settings with restricted access to health-care. It is imperative that standard treatment measures are assessed for their efficacy and reliability.
Understanding the driving forces behind therapy failures is challenging. This is to a large extent the case because of the complex life cycle and population structure of TB: The typical sequence of events leading to acute pulmonary tuberculosis occurs as follows [1,4–7]. Upon inhalation, TB bacilli reach the pulmonary alveoli of the lung. There they are assimilated by phagocytic macrophages. In most cases the bacteria are being killed continuously by phagocytosis while the cell-mediated immunity develops. More rarely, they may persist in an inactive state, which is considered a latent infection. Infected macrophages may aggregate and form granulomas by recruiting more macrophages and other cell types. Inside granulomas, increased necrosis of macrophages can lead to the formation of a caseous core. In latently infected hosts, an equilibrium establishes where the immune system prevents further growth but the bacteria persist in a dormant state [8,9]. However, especially in patients with a compromised immune system, the bacteria may continue or resume growth [4,6]. In this case, the bacterial population steadily increases until the granuloma bursts into the bronchus forming an open cavity. Mycobacterium tuberculosis is an aerobic organism and depends on the availability of oxygen to promote its growth. Because the oxygen levels inside macrophages and granulomas are low, the growth rate is reduced [6,9–13]. In open cavities, oxygen supply is not limiting anymore and the population size increases rapidly. The extracellular bacteria in the cavities may also spread to other locations in the lung where they are again combated by the dendritic cells of the immune system. Some bacteria can be expelled with sputum and be transmitted to other individuals or they may enter a blood vessel and cause lesions in other organs.
The standard treatment is a six-month short-course regimen [1,14–17], consisting of two months of combination therapy with isoniazid, rifampicin, pyrazinamide and ethambutol followed by a continuation phase of four months with isoniazid and rifampicin only [18]. According to tuberculosis treatment guidelines all drugs are taken daily during the first two months. During the following four months isoniazid and rifampicin are administered three times a week with a 3-fold increased isoniazid dose [15]. For patients with previous TB treatments the WHO recommends a 8-month retreatment regimen containing additionally streptomycin [17].
In recent years, the problem of drug resistance has increased in severity due to the emergence and spread of multi-drug-resistant tuberculosis (MDR-TB) [19–21], where MDR-TB is defined as infection by M. tuberculosis strains conferring resistance to at least isoniazid and rifampicin. Resistant TB is assumed to emerge at least in part due to inappropriate treatment or suboptimal adherence to the treatment regimen [22]. Poor compliance has been associated with treatment failure and the emergence of resistance in previous studies [23–27]. Multi-drug-resistance usually develops in a step-wise manner. These steps are thought to include functional monotherapy; either due to different drug efficacies among certain bacterial populations or due to different pharmacokinetics [28,29]. Prevalence data of MDR-TB in Europe (see Fig 1) show that patients who have previously received treatment are on average six times more likely to suffer from MDR-TB than patients who are newly diagnosed. There are several possible explanations for this observation. Individuals who are infected with MDR-TB are more likely to have a treatment failure or a later relapse [30–33], especially if they are not properly diagnosed. These patients could then come under more accurate scrutiny and eventually be reported as MDR-TB patients with previous treatment history. Another more direct possibility is that a considerable fraction of patients who have contracted susceptible TB develop de novo MDR-TB during the first therapy [34].
The goal of this study is to assess the factors that determine the de novo acquisition of drug resistance and to get a better insight in the underlying dynamics. Specifically, we want to study the contribution of imperfect compliance and retreatment regimens. In some areas, second-line drugs are not easily accessible. Moreover, drug-susceptibility tests may not be performed due to the lack of required infrastructure or questionable reliability of patient treatment history [37]. Hence, we assess the impact of a retreatment that is identical to the first therapy as well as a retreatment that follows the WHO recommendation [17]. To achieve this goal we develop a computational model of a within-host TB infection and its consecutive treatment with currently recommended first-line regimens. The model framework encompasses the population dynamics of various M. tuberculosis genotypes with different resistance patterns in three pulmonary compartments as well as the pharmacodynamics and the pharmacokinetics of the drugs that are used for treatment. The aim is to provide qualitative insights into the infection dynamics of tuberculosis. The parameterization is based on the most recent concepts and individual experimental results found in the literature. Given the current lack of a good animal or in vitro model for TB, a computational model, may help to bridge the gaps arising from the inaccessibility of TB in experimental model systems and allow the hypothetical assessment of treatment scenarios, which would be otherwise ethically inadmissible in patient trials. In particular, problems resulting from imperfect therapy adherence can be usefully addressed with a computational model.
In the following section we present the basic framework of the computational model, the parameterization and key aspects of our simulations. In essence, our model consists of coupled logistic-growth models that are connected such that they capture the basic population structure (compartments) of TB (see Fig 2). The action of TB-drugs is included in this model via realistic pharmacokinetics / pharmacodynamics functions. Resistance to these drugs is modeled by distinguishing between up to 32 genotypes (all combinations of 5 mutations) with varying resistance patterns. Since mutations are generated at low frequencies and numbers (due to the low bacterial mutation rate), chance events are essential in the dynamics of this system and hence we consider a stochastic version of the model. In the following we provide a detailed description of the model; the model equations and further details can be found in S1 Text.
Our model describes pulmonary tuberculosis and assesses the emergence of resistance during multi-drug therapy. A graphical illustration of the model is provided in Fig 2. The model reflects the compartmentalization of the bacteria into three distinct subpopulations as described by Grosset [5]: intracellular bacteria within macrophages (M), bacteria within the caseating tissue of granulomas (G) and extracellular bacteria which mostly reside in open cavities (OC). The compartments differ in their maximum population sizes as well as the bacterial replication rates that they allow. The base replication rate r is modified by a factor γ, which reflects the compartment specific conditions that influence the replication rate. Bacteria have a natural density-dependent death rate in each compartment. The constant replication rate and the density-dependent death rate constitute a logistic growth model that was assumed to describe the basic population dynamics. Bacteria also migrate unidirectionally at a rate m from one compartment to another. Offspring bacteria have a certain chance to acquire or lose a mutation that confers resistance to one out of up to five drugs that may be administered during treatment. Every resistance mutation confers a fitness cost which affects the reproductive success of its carrier. This means that the bacterial population inside a compartment comprises of up to 32 genotypes, which differ in their drug resistance pattern as well as their relative fitness.
To outline the population dynamics within a single compartment we describe them first in the form of a deterministic differential equation. The dynamical equation is given by
dNc,gdt=r⋅γc⋅ωg⋅Nc,g−mc⋅NcKc⋅Nc,g+mc′⋅Nc′Kc′⋅Nc′,g−(dc+κc,g)⋅Nc,g
(1)
Here Nc,g is the number of bacteria of a specific genotype g in a specific compartment c. The parameter r is the base replication rate of M. tuberculosis and γc is a factor, which modifies the replication rate according to the different metabolic activities in each compartment. ωg represents the relative fitness of the specific genotype. mc is the rate with which bacteria migrate to the subsequent compartment. The migration rate is multiplied by the ratio between the total population size Nc and the carrying capacity Kc. This reflects the increased migratory activity that takes place during an acute infection. Nc´, Kc’ and mc´ correspond to the overall bacterial population including all genotypes of the supplying compartment, its carrying capacity and its migration rate, respectively. The last term reflects the density-dependent death rate dc and the drug induced genotype specific killing rate κc,g. The bactericidal effects of the drugs contribute additively to the killing rate κc,g (see S1 Text for further details).
The dynamics of the bacterial population in the model are actually simulated as stochastic processes. For this reason we translated the underlying deterministic differential equations into a corresponding stochastic framework by applying Gillespie’s τ-leap method [38].
The parameter estimates used in this model are whenever possible drawn or derived from experimental results in the literature. To account for the diversity of infection and treatment courses in different patients we allow some parameters to vary within a certain range. Parameters are summarized in Table 1.
The basic growth dynamics rest upon the replication rate and the carrying capacity of the compartments. Based on recent studies [39–41] we assume a maximum bacterial load between 105 and 107 bacteria each for the macrophage and the granuloma compartment and 108 to 1010 bacteria for the extracellular compartment. Under optimal conditions M. tuberculosis has a replication time of 20h, hence we set the maximum replication rate in the model to 0.8 d-1 [5].
Every new bacteria cell has at birth the chance to acquire or lose one or multiple resistance mutations and therefore get a genotype, which is different from the mother cell. The frequency of specific resistance mutations and therefore the mutation rate for the main first-line drugs have been first estimated by David in 1970 [42] to be around 10−7–10−10. However, more recent observations suggest considerably higher frequencies in the order of 10−6 to 10−8 [5,43]. A possible reason for this discrepancy between these estimates are varying mutation rates in in vitro experiments compared to the conditions encountered in vivo due to stress-induced mutagenesis mechanisms or variations among strains [44–46]. Furthermore, we assume that mutations only occur during proliferation while mutations during the stationary phase could serve as an additional source of resistance mutations [47]. Therefore, we choose to allow for patients with the more recent higher mutation rates because this will yield more conservative estimates (see Table 2). Our model incorporates backwards mutations from the resistant to the sensitive phenotype, which also restore the reproductive fitness. However, we consider a reversion to be ten times less likely than the original forward mutation because the occurrence of any additional mutation within a gene to be an exact reversion is more infrequent.
When assessing the prevalence of certain genotypes, fitness costs that come with resistance mutations have to be considered. The cost of resistance against anti-tuberculosis drugs appears generally to be low [48–51]. Drug-resistant mutants isolated in patients have even been found to be on par with susceptible wild type strains regarding their infectivity and replicative potential. Since cost-free resistance mutations are rather rare, the high fitness of resistant strains that have been found in clinical isolates [48,49] is assumed to arise due to the acquisition of secondary site mutations which minimize the fitness costs (so-called compensatory mutations) [50]. However, there is evidence that at least initially newly acquired drug resistance confers some physiological cost [52]. Because our model simulates the de novo acquisition of resistance mutations and because the time frame of a single patient treatment is rather short we assign a small fitness cost to every possible mutation and neglect the counterbalance of fitness costs by compensatory mutations.
The effect of administered drugs depends on the pharmacokinetics and pharmacodynamics of these drugs (see Table 1). Both influence the killing rate κc,g at any given time point during treatment. While pharmacokinetic parameters describe the course of the drug concentration in the target tissue, pharmacodynamic parameters characterize the effect the drugs have at a given concentration. The minimal inhibitory concentration (MIC) describes the minimal drug concentration at which bacterial growth is reduced by at least 99%. Additionally, the EC50 describes at which drug concentration the half-maximal effect (commonly, bacterial killing) is observed, while the Emax indicates the maximal effect of the drug. These pharmacodynamic parameters are obtained by fitting the drug action model to killing curves found in the literature [53,54] (see S1 Text). The specific efficacy of most drugs in the different compartments is typically not quantified. There are several studies that tried to assess the bactericidal activity inside macrophages [55–59]. Unfortunately, these estimates are highly variable and sometimes even contradictory [55,58]. In addition to these experimental difficulties, it is possible that the pharmacodynamics of anti-tuberculosis drugs are again different in the human body [60–64]. To reflect this uncertainty we assign compartment efficacies from a range of values which corresponds to the most recent estimates [56–59,65–70].
To investigate the role of treatment adherence on patient outcome, we followed disease progression starting with the infection of macrophages until all compartments approximately reached their maximum bacterial load. For each parameter set, we simulate the outcome of 10’000 patients who vary both in their pharmacokinetic and–dynamic characteristics as well as compartmental attributes. Parameters are generally picked from a normal distribution. If only a range is known the parameters are chosen from a uniform distribution. To measure the actual treatment efficacy we let every patient develop an acute tuberculosis infection during 360 days. This allows for the emergence of mutants prior to treatment initiation and provides enough time for the establishment of an equilibrium in the bacterial population composition. After this period we start the standard short course therapy regimen with four drugs being taken daily for two months followed by four months in which just isoniazid and rifampicin are taken three times per week. If the infection is not completely sterilized after the first treatment we schedule a retreatment. Since the model does not cover the possibility of dormant bacteria the population recovers rather quickly after an unsuccessful treatment. Hence, we begin the retreatment 30 days after completion of the previous treatment. After such a time span the population reaches a bacterial load where acute symptoms would be again suspected. If not stated otherwise the retreatment corresponds to the WHO recommendation for retreatments [31,71]. The WHO recommendations include streptomycin, which is used together with the original four first-line drugs during the first two months. Afterwards the therapy is being continued for another month without streptomycin and during the last five months only isoniazid, rifampicin and ethambutol are administered. All drugs are being taken daily during the whole retreatment.
The 95% confidence intervals (CI) of patient outcomes in the figures is calculated by picking the value for a two-sided 95% confidence limit with n– 1 degrees of freedom from a t-distribution table where n is the number of patients. This value is then multiplied with the standard deviation σ and divided by the square root of n. The resulting value is then added and subtracted from the mean to get the actual confidence interval.
The impact of treatment on the net growth rate of wild-type or MDR bacteria differs strongly between compartments (Fig 3): Before treatment starts, the growth rates in macrophages and granulomas are lower than in the open lung cavities due to hypoxia and a generally adverse environment for bacterial growth in these compartments. Since we assume that the drug concentration immediately reaches the maximum the impact of combination therapy on growth rate is immediately apparent after the administration of the first dose of drugs. In all compartments the drugs are able to keep the wild-type populations from regrowth during the following days. Especially in granulomas pyrazinamide is able to diminish the population over a long period due to its relatively long half-life. MDR-TB is substantially less affected by the combination therapy because only pyrazinamide and ethambutol are effective. This means that in macrophages or open lung cavities the multi-drug-resistant population remains constant at best or is even able to slowly grow. Only in the granulomas where mostly pyrazinamide is active (see Table 1) the loss of effectiveness of isoniazid and rifampicin is less prominent.
The compliance of a patient with the prescribed drug regimen is a key factor for a successful treatment outcome. For the assessment of treatment success we monitor for every patient three different nested treatment outcomes. Firstly, we define treatment failure as the incomplete sterilization of the lung at the end of the therapy. Secondly, the emergence of MDR-TB is defined in our simulations as 10% or more [72] of the remaining bacterial population after treatment failure being resistant against at least isoniazid and rifampicin. Finally, emergence of full resistance (FR) is defined as 10% or more of the population being resistant against all drugs that were used in the treatment regimen (either 4 drugs for first treatment or up to 5 drugs for retreatment).
Adherence in our simulations refers to the probability with which the patient takes the prescribed drugs at any given day. We assume that failure to take drugs on a given day always affects all drugs of the prescribed regimen.
In our simulations, the level of adherence has a strong but complex impact on treatment success (Fig 4A). Under perfect adherence the model shows a very low failure rate. However, if adherence decreases the probability for treatment failure increases rapidly. Between 40% and 80% adherence there is also a small fraction of patients that fail treatment due to the emergence of MDR-TB. Furthermore, at these adherence levels the model also shows only limited treatment success. Thus, failure decreases monotonically with adherence while MDR is maximized at intermediate levels. Patients who fail on the first treatment and who undergo retreatment (Fig 4B) have a failure rate of 20% at 80% adherence. However, the probability for treatment failure increases to about 50% under perfect adherence. Patients who fail the first treatment despite high adherence may often have disadvantageous combinations of PK/PD parameters, which also decrease their success probabilities during the retreatment. In Fig 4B, 4C and 4D the number of patients per adherence level undergoing retreatment decreases strongly as can be seen from the frequency of treatment failure in Fig 4A. When comparing Fig 4A and 4E, which shows the combined outcome probabilities for both treatments, we see that the retreatment reduces the probability of failure over the upper half of the adherence spectrum.
The additional treatment success of retreatment regimens depends on adherence and the addition of streptomycin to the regimen (Fig 4B). In our model, even under perfect adherence the chance of treatment failure remains substantial, and in the majority of patients who fail under retreatment MDR-TB emerged de novo. Furthermore, at suboptimal adherence levels a considerable proportion of patients even carry strains that are not susceptible to any of the five administered drugs. The outcome of retreatment depends crucially on whether MDR was acquired during initial treatment: Because the majority of patients who fail the first treatment do not carry MDR-TB their outcome probabilities for the retreatment are almost identical to the overall cohort of failed patients (Fig 4C). Even though the vast majority of patients who failed the first treatment did not develop MDR-TB, a substantial fraction of patients who also failed the second treatment harbor MDR or FR strains. This occurs due to increased subpopulations of monoresistant bacteria that accumulate during the first treatment and that are by itself not sufficient to be diagnosed as MDR-TB. When comparing patients who are diagnosed with MDR-TB after the first treatment (Fig 4D) and those who are not (Fig 4C) we see that patients who develop MDR-TB are very likely to fail the retreatment as well. At higher adherence levels the majority of those patients develops full resistance against all five drugs (Fig 4D). When considering the outcome for both treatments combined (Fig 4E) it becomes more evident that the addition of streptomycin and the more intense retreatment has a beneficial effect on the overall success rate but patients who also fail the retreatment are more likely to carry multidrug-resistant TB strains.
When second-line drugs are not available or susceptibility test are not performed, it may occur frequently that a previously treated patient is retreated with the first line treatment. Our results in Fig 5 show that such a retreatment with the first line drugs has almost no additional treatment success beyond the initial treatment. Patients all across the spectrum of adherence experience treatment failure. The identical first-line retreament only increases the chances for the bacteria to accumulate resistance mutations and leads between 50% to 100% adherence to nearly all uncleared patients harboring MDR-TB or worse. This outcome is standing out when comparing the cumulative treatment success in Fig 5D with the results after the first treatment. While the overall success curve did not change the fraction of MDR-TB patients over a large adherence range increased substantially.
The aim of this study is to elucidate the effects of treatment adherence and retreatment on the emergence of resistance in TB. The model explicitly incorporates the pharmacodynamics and pharmacokinetics of all drugs that are used for standard therapy and the WHO retreatment recommendation. Depending on the compartment in the lung in which the bacteria reside (macrophages, caseous centers of granulomas or open cavities), M. tuberculosis has different stages of infection and drug-susceptibilities. Therefore, we explicitly include these different compartments to be able capture the effect of heterogeneous selection pressure. Because not all of the parameters used in our model have been quantified with high accuracy, we do not claim that the model has quantitative predictive power. Rather, it aims to qualitatively demonstrate the underlying dynamics of a tuberculosis infection.
Our results suggest that poor adherence is a major cause for treatment failure. When considering the predicted rates of treatment failure one also has to take into account that our definition of treatment failure is probably rather conservative. We do not include the possibility of remaining dormant bacteria, which might increase the likelihood of treatment failure or relapse. On the other hand, we also neglect the possibility of the infection being contained at a later time point by the immune system, thus probably underestimating the chance of success. It is also noteworthy that even at perfect adherence some patients may have a negative treatment outcome. This is most likely due to a random aggregation of very adverse pharmacokinetic parameters and unfavorable infection attributes in some patients. Such outcomes due to pharmacokinetic variability and despite good adherence have been predicted in an in vitro study [73]. Furthermore, our results show that over a certain range of adherence a small fraction of patients develop MDR-TB. At intermediate adherence these patients also have a low likelihood of being treated successfully. Thus, good adherence to therapy is crucial: Not only does it increase treatment success, it also decreases the probability for the emergence of MDR-TB.
According to our model, the WHO recommendation for retreatment is somewhat of a double-edged sword. While at high adherence levels the recommended treatment is able to cure the majority of patients who failed the first line therapy, it also increases the fraction of patients harboring drug resistant strains across almost the whole spectrum of adherence. Previous studies already raised concerns about the possible amplification of resistance [71,74–77]. In the WHO treatment guidelines it is recommended that drug susceptibility test results should be taken into account when deciding upon the retreatment regimen [17]. However, the vast majority of patients in our model would probably not have been diagnosed with MDR-TB after the first regimen even though they may still harbor increased subpopulations of monoresistant bacteria. Therefore it is conceivable that many would have been treated with the WHO recommended regimen. A large fraction of patients who failed this retreatment eventually developed MDR-TB. Considering the results from our model further clinical studies are needed which analyze the treatment success rates and the accompanying risks of the standard retreatment regimen.
Retreating failed patients with an identical short course therapy leads to poor outcome in our simulations. A lower success rate for MDR-TB patients treated with the standard short-course therapy has been confirmed in a large cohort study [37]. In our simulations it is rare that patients who failed the previous treatment are cured after undergoing the same therapy again provided that adherence remains unchanged. Retreatment with the same regimen only generates more opportunities for single resistant mutants that emerged during the first treatment to accumulate further mutations, thus minimizing the number of future treatment options.
These findings are in accordance with previous studies which found a positive correlation between previous treatment and the occurrence of resistance [78–81]. This might be an indicator that de novo resistance on an epidemiological scale occurs at a significant frequency and that the main contributor to the frequency of MDR-TB is not necessarily the mere transmission of such strains.
In summary our data show that patient adherence is a crucial component of treatment success. The probably cheapest and most effective way to ensure a positive treatment outcome while also minimizing the risk for the emergence of MDR-TB is to maintain proper patient compliance with the treatment. This supports the Directly Observed Treatment, Short-Course (DOTS) strategy of the WHO, which includes healthcare workers or community health workers who directly monitor patient medication. If treatment fails, thorough tests of drug susceptibility of the remaining infecting population, would be of considerable value. According to our results a retreatment regimen including streptomycin has the potential to increase the overall cure rate, but also increases the fraction of patients who carry drug-resistant strains. A common principle of physicians is to “never add a single drug to a failing regimen” [82] this principle is often not followed in retreatment. A preceding drug sensitivity test might show existing drug resistances and the retreatment regimen could be adapted accordingly. Nonetheless the standard retreatment regimen is still superior to a retreatment with the identical first-line drugs. Such a retreatment is unlikely to achieve a higher overall cure rate and dramatically increases the probability for the emergence of MDR-TB, which reduces further treatment options. This shows that a dependable patient treatment history that is available to the responsible health professional is also important before initiating a treatment regimen.
|
10.1371/journal.pntd.0002696 | A Monoallelic Deletion of the TcCRT Gene Increases the Attenuation of a Cultured Trypanosoma cruzi Strain, Protecting against an In Vivo Virulent Challenge | Trypanosoma cruzi calreticulin (TcCRT) is a virulence factor that binds complement C1, thus inhibiting the activation of the classical complement pathway and generating pro-phagocytic signals that increase parasite infectivity. In a previous work, we characterized a clonal cell line lacking one TcCRT allele (TcCRT+/−) and another overexpressing it (TcCRT+), both derived from the attenuated TCC T. cruzi strain. The TcCRT+/− mutant was highly susceptible to killing by the complement machinery and presented a remarkable reduced propagation and differentiation rate both in vitro and in vivo. In this report, we have extended these studies to assess, in a mouse model of disease, the virulence, immunogenicity and safety of the mutant as an experimental vaccine. Balb/c mice were inoculated with TcCRT+/− parasites and followed-up during a 6-month period. Mutant parasites were not detected by sensitive techniques, even after mice immune suppression. Total anti-T. cruzi IgG levels were undetectable in TcCRT+/− inoculated mice and the genetic alteration was stable after long-term infection and it did not revert back to wild type form. Most importantly, immunization with TcCRT+/− parasites induces a highly protective response after challenge with a virulent T. cruzi strain, as evidenced by lower parasite density, mortality, spleen index and tissue inflammatory response. TcCRT+/− clones are restricted in two important properties conferred by TcCRT and indirectly by C1q: their ability to evade the host immune response and their virulence. Therefore, deletion of one copy of the TcCRT gene in the attenuated TCC strain generated a safe and irreversibly gene-deleted live attenuated parasite with high immunoprotective properties. Our results also contribute to endorse the important role of TcCRT as a T. cruzi virulence factor.
| Trypanosoma cruzi is a protozoan parasite which infects 9 million people in Latin America. Currently there is no vaccine to prevent this disease. Therefore, different approaches or alternatives are urgently needed to identify new protective immunogens. Live vaccines are likely to be most effective in inducing protection; however, safety issues associated with their use have been raised. Hence, we genetically manipulated an attenuated strain of T. cruzi as a safety device to rule out the possibility of reversion to the virulent phenotype. The genetically modified parasites were highly susceptible to killing by the complement machinery and presented a reduced propagation and differentiation rate. We have extended these studies to assess, the virulence, immunogenicity and safety of the mutant as an experimental vaccine. Accordingly, we show that genetically modified parasites present attenuated virulence in mice. The genetic alteration was stable and, after long term infection, it did not revert back to wild type form. Furthermore, after challenge with a virulent T. cruzi strain, mutant immunization induces a highly protective response evidenced by significantly lowered parasite density, mortality, spleen weight index and tissue inflammatory response. Our study provides new insights into the host-pathogen interactions and into the use and evaluation of irreversibly gene-deleted live attenuated parasites to protect against Chagas disease.
| Chagas disease is a neglected tropical ailment caused by the flagellate protozoan Trypanosoma cruzi. It is estimated that 12–20 million people are infected worldwide causing 10–50,000 deaths/year [1]. Vector control strategies were not entirely successful mainly due to the inaccessibility and the vast distances that separate endemic areas. Transmission, despite the spraying of insecticides, has been increasing in parts of Argentina, Venezuela and Brazil [2]. In addition, the cases of Chagas disease have raised in many parts of South America and have spread globally because of immigration into non-endemic areas in developed countries [3], [4], [5]. Drugs used for treatment have serious adverse effects and do not cure the chronic stage [6]. However, vaccination to protect the 40–100 million individuals at risk of acquiring this serious disease has not been well developed or entered in human trials.
Considering T. cruzi complexity, with a genome of more than 12,000 genes and four distinct life stages, DNA and peptide vaccination for Chagas disease is insufficient and has, so far, not been reported to induce sterile immunity after challenge [7]. Currently, there is an increased interest in the development of irreversibly gene-deleted live attenuated parasites, as a possible mechanism to reduce the risk of reversion to virulence. There is considerable evidence in genetically modified organisms such as Toxoplasma, Plasmodium and Leishmania, which argues for the usefulness and effectiveness of these parasites as promising immunogens [8], [9], [10], [11], [12], [13], [14], [15], [16]. In T. cruzi, unfortunately, there are so far only five studies of vaccination using genetically attenuated strains [17], [18], [19], [20], [21]. The advantages of using this kind of immunogens are: (1) They can provide the full spectrum of relevant native epitopes and immune stimulating molecules, such as Toll-like receptors organized together, which would generate a high immunogenicity, unlike other types of vaccines that offer only a restricted spectrum of immunogens. (2) They can be manipulated to develop multiple genetic modifications. (3) They undergo antigen processing and presentation as in the case of virulent infection. (4) They generate, after inoculation, a strong and long lasting protective response compared with other experimental T. cruzi vaccines [7]. (5) They can be grown in axenic conditions with a lower economic production cost than other vaccine strategies [22].
The TCC wild type strain does not produce considerable tissue lesions or bloodstream parasite levels detectable by fresh blood mounts in rats [23]. Immunization with TCC provided, after a virulent challenge, a strong immune protection against virulent T. cruzi infections [24], [25], also evidenced when the challenge was performed using 17 isolates of T. cruzi obtained in an extensive endemic area of the Province of Salta, Argentina [26]. A strong control of parasitemia and tissue damage was observed in mice challenged a year after immunization [27], [28]. The protective effect of TCC was extended to field experiments in guinea pigs [25] and dogs [29]. Unfortunately, the TCC attenuation is genetically undefined and the possibility of reversion to the virulent phenotype cannot be excluded. In order to add a safety mechanism to prevent this reversion, in a previous work, we generated and characterized a TCC clonal cell line that lacks a TcCRT allele (TcCRT+/–) and another clone overexpressing it (TcCRT+). TcCRT is a T. cruzi virulence factor, that after being translocated from the endoplasmic reticulum (ER) to the area of flagellum emergence, can hijack the complement C1 component, inhibiting the activation of the classical and lectin complement pathways at their earliest stages [30], [31], [32] and producing pro-phagocytic signals increasing parasite infectivity [33]. Recently, an important role of TcCRT in the C1-dependent T. cruzi infectivity of human placenta explants has been determined in one of our laboratories, thus providing a plausible mechanism for congenital transmission of this infection [34]. In our previous work, we determined that the TcCRT+/– mutant contained about 6-fold less TcCRT polypeptide than wild type parasites [35]. Moreover, parasites overexpressing TcCRT contained about 2-fold more TcCRT polypeptide than wild type parasites. Consequently, monoallelic mutant parasites were significantly more susceptible to killing by the complement machinery. On the contrary, TcCRT+ parasites showed higher levels of resistance to killing by the classical and lectin but not by the alternative complement activation pathways. The involvement of surface TcCRT in depleting C1 was confirmed through restoration of serum killing activity by addition of exogenous C1. In axenic cultures, a reduced propagation rate of TcCRT+/– parasites was observed. Moreover, TcCRT+/– parasites presented a reduced rate of differentiation in in vitro and in vivo assays [35].
The previous studies led us to the objective of this report, to detect whether the TcCRT monoallelic deletion caused changes in the infectivity and immunoprotective behavior of the attenuated TCC strain.
All animal protocols adhered to the National Institutes of Health (NIH) ‘‘Guide for the care and use of laboratory animals’’ and were approved by the Animal Ethics Committee of the School of Health Sciences, National University of Salta (N° 014-2011) [36].
A T. cruzi clone derived from the attenuated TCC strain [37], designated here as wild type, was used. Also, we used a clonal cell line lacking one TcCRT allele (TcCRT+/−) and a recombinant T. cruzi clone that overexpresses the TcCRT polypeptide (TcCRT+) [35]. Epimastigotes were grown at 28°C in liver infusion-tryptose medium (LIT) supplemented with 10% fetal bovine serum decomplemented at 56°C for 60 min., 20 µg hemin (Sigma, St. Louis, MO, USA),100 IU of penicillin and 100 µg/ml streptomycin. To obtain metacyclic trypomastigotes, epimastigote forms were allowed to differentiate by adding 10% w/v triatomine gut homogenate to the cultures [38]. The percentage of metacyclic forms was recorded daily in a Neubauer chamber. In addition, we used the Tulahuén strain and a highly infective isolate recently characterized [39].
Hemocultures were performed by seeding 200 µl of heparinized blood into 2 ml of LIT under sterile conditions; the cultures were incubated at 28°C and scanned for motile parasites under an inverted microscope on days 15, 30, 45, and 60. PCR for T. cruzi detection was also performed. Briefly, 700 µl of blood from each inoculated animal was processed. Kinetoplast DNA was amplified using primers 121 (5′-AAATAATGTACGGGTGAGATGCATGA-3′) and 122 (5′-GTTCGATTGGGGTTGGTGTAATATA-3′). Sample storage, DNA extraction, amplification, electrophoresis and staining were performed as previously described [40].
To assess the stability of the mutation, we used TcCRT+/– and TCC wild type parasites recovered from hemocultures performed on nude mice on day 90 post-infection (p.i.) after immunosuppression with cyclophosphamide. These parasites were grown and expanded in LIT medium. Genomic DNA was purified using the phenol–chloroform method. Diagnostic PCR analysis confirmed sequences corresponding to TcCRT and HYG gene. Primers used were: Pair 1, to amplify the entire TcCRT CDS (1.2 Kb), CRT1 (5'-GCCAGATATCATGAGGAGAAATGACATAAA-3') which anneals into the TcCRT initiation codon and CRT2 (5'-TCCTCTCGAGTCAAAACTTTCCCCACCGAA-3'), for the stop codon. Pair 2, to amplify the CDS of HYG gene (0.96 kb), H1 (5'-CGTCTGTCGAGAAGTTTCTG-3') which anneals into the HYG initiation codon and H2 (5'-GAAGTACTCGCCGATAGTG-3') for the stop codon. Pair 3, CRT 7 (5'-CCTTCCGATGGCATTAGC-3') which anneals upstream of TcCRT gene plus primer H4 (5'-CTCGCTCCAGTCAATGACC-3') for the HYG sequence (1.4 kb). Pair 4, CRT93 (5'-ATTCCAAACAACATTGCCGT-3') which anneals downstream of TcCRT gene plus H6 (5'-GGACCGATGGCTGTGTAGAAGTACTCGCCGATAGTGG-3') for the HYG sequence (1.4 kb).
Total Immunoglobulin G antibodies against T. cruzi were measured by Enzyme-linked Immunosorbent Assay (ELISA) using T. cruzi epimastigote homogenates (2 µg/well) as antigens. Dilutions of sera, anti-mouse IgG as a secondary antibody (Sigma, St. Louis, MO, USA) and conjugate were 1/100; 1/2,500 and 1/16,000 respectively. The antibody concentration was expressed as the optical density at 490-nm wavelength.
Male Balb/c inbred or athymic nude (nu/nu) immunodeficient mice (about 1 month old) were inoculated intra-peritoneally (i.p.) with 5×105 metacyclic TCC TcCRT+/–; TcCRT+ and wild type trypomastigotes. Balb/c mice were subjected to PCR (15, 30, 90, 180 and 220 days p.i.), hemoculture (15, 30, 90 and 220 days p.i.) and serological determination of antibody levels (20, 47, 60, 90 and 165 days p.i.) as described above. Nude mice were examined by PCR and hemoculture on day 15, 30 and 90 p.i. To improve the detection of latent infections, the last sample of both, Balb/c and nude mice, were obtained after immunosuppressive treatment with cyclophosphamide. The immunosuppression regimen is based on 5, 250 mg/kg cyclophosphamide doses administered during 10 days. Samples were collected 10 days after the last dose.
To test whether mutant T. cruzi clones induced immunological protection, groups of 6 Balb/c mice, about 1 month old, were inoculated i.p. with 5×105 metacyclic TCC TcCRT+/–; TcCRT+ and wild type trypomastigotes. A control group was inoculated with 100 µl of PBS (day 0). On day 15 a boost similar to the initial inoculation was administered. On day 30, antibody levels from immunized mice were determined and, on day 120, all groups were challenged with 104 blood trypomastigotes of a highly virulent T. cruzi TcVI isolate, recently characterized [39]. Blood was drawn from the tail tip of mice, under slight ether anesthesia using heparinized, calibrated capillary tubes. Ten microliters of blood were placed between slide and cover slip and the number of parasites per 100 fields was recorded microscopically (40X) twice a week. Then, the number of parasites per 100 fields (parasitemia) was recorded from fresh blood mounts under microscope (40X). Finally, on day 60 post-challenge, surviving animals were sacrificed, spleen index and the presence of histological damage was measured in tissue samples.
Tissue samples from heart and quadriceps muscle were fixed in 10% formaldehyde and processed using routine histological techniques. Serial histological hematoxylin-eosin-stained sections (3–5 µm thick) were studied. We searched for lymphocytic infiltrates in areas averaging 53 mm2 for heart and 38 mm2 for quadriceps muscle, scanning at least three sections per organ. Quantification of the inflammatory response was scored blindly as severe (+++: presence of foci containing numerous inflammatory cells covering at least half of the sections surface), moderate (++: large inflammatory foci covering up to ¼ of the section surface), slight (+: presence of small and isolated inflammatory foci) or absent (–: no presence of foci or inflammatory cells).
Body and spleen weight were determined to calculate the spleen index (spleen index = spleen weight X 100/body weight) as an indirect effect of infection severity.
The Mann-Whitney U tests and one-way variance analysis (ANOVA) of the GraphPad Prism version 5.0 software were used. Values are expressed as means ± standard error of mean of at least three separate experiments. P values equal to or minor that 0.05 were considered as significant.
To determine whether the TCC mutant parasites were capable of infecting and survive for long periods of time in the host, we monitored their in vivo infectivity and persistence. TCC TcCRT+/–; TcCRT+ and wild type epimastigotes were transformed into metacyclic trypomastigotes and inoculated (5×105) in Balb/c and nude mice. Since the T. cruzi TCC strain is attenuated it is not possible to detect circulating parasites in blood samples by fresh blood mounts. Therefore, infection was detected by more sensitive methods (hemoculture and PCR). No positive hemocultures were obtained from any immunocompetent Balb/c mice inoculated with the three clones at any of the evaluated time points (Table 1). However, PCR determinations showed different infection patterns. No positive PCR was detected in TcCRT+/– immunocompetent inoculated mice throughout the follow-up, beyond 200 days, and even after immunosuppression. In contrast, positive reactions were found in all animals infected with wild type parasites. After immunosuppression, 2/3 of wild type inoculated mice were positive. All mice infected with TcCRT+ parasites presented a behavior similar to the wild type strain (Table 1). Thus, the attenuated TCC T. cruzi strain could be rendered even less virulent than wild type via the targeted deletion of one TcCRT allele.
Using nude mice we detected an increased rate of infection by PCR and hemoculture in all three experimental groups. No differences were found between the strains at any time p.i. As expected, after immunosuppression we detected a high mortality in all groups (Table 1).
We determined serum antibody levels in BALB/c mice infected with the three parasite clones on acute and chronic stages of infection and disease developmenRTt. TcC+/– infected mice showed undetectable antibody levels (p = 0.01) contrasting with mice inoculated with both wild type and TcCRT+ parasites. In fact, the TcCRT+/– values were comparable to those obtained from the PBS-inoculated, negative controls. No differences between TcCRT+ and wild type were found (p = 0.84). As previously described [27], mice inoculated with the Tulahuén virulent strain (positive control) showed about six-fold higher antibody levels than those obtained with any of the TCC strains (Fig. 1).
To exclude the possibilities of cross-contamination, reversion of the genetic mutation or TcCRT locus instability, we determined whether the parasites isolated from hemocultures after long term infection in mice corresponded to mutant parasites. Genomic DNA was extracted from TcCRT+/– and wild type parasites grown on hemocultures at day 90 p.i. (Table 1). We amplified sequences corresponding to the TcCRT coding sequence (CDS) and the hygromycin phosphotransferase (HYG) marker gene. The sizes of amplified fragments in the DNA of the recovered parasites corresponded to those predicted for the replacement of TcCRT by the HYG gene (Fig. 2).
Additionally, the antibiotic resistance of TcCRT+/– and wild type parasites were tested. Only TcCRT+/– parasites survived in the presence of 300 µg/ml Hygromicin B mediated by the HYG resistance gene at the deleted TcCRT allele (data not shown). Thus, this evidence showed that TcCRT+/– parasites conserved the targeted allele introduced by homologous recombination, that there is no cross-contamination and that the locus remained stable throughout the infection cycle in the mammalian host.
To assess the immunoprotective capacity of mutant parasites against a subsequent reinfection with virulent parasites, groups of six BALB/c mice were inoculated with 5×105 metacyclic trypomastigotes of each of the three clones plus a naive, sham-preinoculated control group. At day 15 a booster similar to the initial inoculation was administered. To determine whether this immunization regimen induces an immune response, blood samples were taken during the immunization phase on day 30 post-priming. After 120 days, these mice together with controls were challenged with 10,000 bloodstream trypomastigotes of a virulent T. cruzi isolate [39]. The protective response generated by immunizing with TcCRT+/– and wild type was significantly higher than in the non-immunized group (p = 0.0001). Mice immunized with TcCRT+/– and wild type parasites showed, after challenge, reduced levels of circulating parasites in peripheral blood, ranging between 0–3 parasites per 100 microscopic fields throughout follow-up, demonstrating the protection afforded by immunization. Parasitemia curves between wild type and TcCRT+/– immunized groups are not significantly different (p = 0.22). These results showed that deletion of a TcCRT allele does not modify the protective response induced by TCC wild type parasites. In contrast, mice immunized with TcCRT+ did not afford protection (Fig 3A). Non-immunized control mice presented high parasitemia with peaks between days 13 and 16, at a time when there was 50% mortality. As expected, these mice showed high parasitemia before death, thus explaining the wide dispersion of the data at that time (Fig 3A). In contrast, in the remaining experimental groups no mortality was recorded. TcCRT+/– immunized and boosted mice showed undetectable specific anti-T. cruzi antibody levels (similar to those obtained in the non-immunized, negative controls) compared to the levels found in both wild type and TcCRT+ (p = 0.004) and clearly different from those obtained from mice infected with Tulahuén parasites (Fig 3B).
Autopsies were performed on mice 4 months after priming with TCC TcCRT+/–; TcCRT+ or wild type trypomastigotes and 2 months after a virulent T. cruzi challenge. Non-immunized, wild type and TcCRT+ mice presented severe inflammatory response throughout the heart tissue, however, this response was extensively reduced in TcCRT+/– immunized mice (p = 0.002) (Fig. 4A), thus confirming the protective effect conferred by previous immunization with these parasites. The same effect was observed in muscle tissue: non immunized, wild type and TcCRT+ mice presented moderate to slight cellular damage that was reduced in TcCRT+/– immunized mice (p = 0.0007) (Fig. 4B).
Splenomegaly is a macroscopic manifestation of the expansion of B- and T-lymphoid cell populations produced by the infection of mice with T. cruzi [41]. Thus, the spleen index represents an indirect effect of infection severity. Spleen index at day 60 post-challenge was significantly decreased in TcCRT+/– and wild type immunized mice compared to that in the non-immunized controls (p = 0.02 for both cases). However, TcCRT+ immunized mice presented no differences (p = 0.14) with non-immunized controls (Fig. 4G).
In a previous work, we have characterized a mutant cell line that lacks a TcCRT allele (TcCRT+/–), with bases on the attenuated TCC T. cruzi strain. We showed that TcCRT+/– epimastigotes contained about 6-fold less TcCRT polypeptide than wild type parasites. Moreover, they were significantly susceptible to killing by the complement machinery and presented a reduced in vitro propagation and differentiation rate. In addition, we generated another clonal cell line that over-expresses TcCRT (TcCRT+) and showed high resistance levels to complement attack [35]. Furthermore, it was not possible to generate biallelic TcCRT–/– null mutant clones, perhaps a reflection of the essential character of the TcCRT protein for parasite survival.
TCC wild type infection is hardly detected in immunocompetent animal models due to the attenuation of this strain. The use of highly sensitive methods such as immunosupression regimens followed by PCR and hemoculture is usually required. When inoculated in Balb/c mice, and during a 6-month follow-up period, mutant TcCRT+/– parasites were not detected by either of these techniques, even after immunosupression (Table 1). TcCRT is highly immunogenic in different animal species [42]. Most humans infected with T. cruzi possess anti-TcCRT antibodies [43]. However, levels of specific antibodies in TcCRT+/– inoculated mice were even more reduced as compared to mice inoculated with wild type or TcCRT+ and, as described [27], with the highly infective Tulahuén strain (Fig. 1).
The increased virulence attenuation of TcCRT+/– in mice could probably be related to increased complement susceptibility and to the deposition of C1q on the parasite surface, configuring a strategy called "apoptotic mimicry". In infective trypomastigotes, TcCRT is translocated from the ER to the area of flagellum emergence where it could hijack C1q resulting in an increased affinity for host cells [32], [33], [44]. Previous reports affirm that the C1q binding on the T. cruzi trypomastigote surface increases parasite infectivity [45] and thus, any disruption of TcCRT/C1q interaction may result in a reduction of infectivity both, in vitro and in vivo [33], [34]. Furthermore, apoptotic mammalian cells express surface ligands with high C1q affinities, among them, the calreticulin orthologue. C1q-coating over apoptotic cells produces pro-phagocytic “eat me” signals that promote clearance of apoptotic bodies conducted by phagocytic cells [46], [47]. One of our laboratories [33], proposed that T. cruzi expressing TcCRT mimic the “eat me” signals, promoting C1q coating, phagocytic cell chemotaxis and increasing parasite infectivity in the early stages of infection. In our work, the TcCRT allele deletion and synthesis reduction [35], possibly generated a lower capacity to capture C1 thereby inducing lower pro-phagocytic signals and reduced infectivity of phagocytic cells in the early stages of infection. In a negative feedback, the limited invasion of phagocityc cells would help TcCRT+/– parasites to stay free, for a longer period of time, and exposed to the complement lytic action in the bloodstream system of the host. These properties may have contributed to the important TcCRT+/– infectivity attenuation (Table 1).
In addition, antibodies aggregated to the T. cruzi surface antigens (including the anti-TcCRT antibodies) through their Fc regions have a high affinity for C1q [33]. Thus the apparent paradox that C1q-fixing antibodies, rather than preventing parasite replication, contribute to increase their infectivity, is explained. Thus, pretreatment with anti-TcCRT (Fab')2 fragments (which lack the Fc fragment of C1q binding) produces the disruption of TcCRT/C1q with serious negative impact on the in vivo e in vitro infectivity [33]. As expected, the TcCRT+/– attenuated line did not produce detectable specific anti-T. cruzi antibodies (Fig.1) probably causing a limited C1q deposit on the parasite surface which, in turn, would contribute to diminish phagocytic signals and hence parasite infectivity. In contrast, mice inoculated with TcCRT+ and wild type showed an increased level in antibody titers compared to TcCRT+/–, which would generate a denser C1q coating. This phenomenon may explain the divergences in infectivity of TcCRT+/–, TcCRT+ and wild type parasites.
We were unable to recover infecting parasites from immunocompetent mice by hemoculture. However, we could detect parasite DNA by PCR in those mice infected with TCC wild type and TcCRT+ parasites (Table 1). This is probably a consequence of both a lower density of circulating parasites and a greater PCR sensitivity for detection of T. cruzi in mouse blood (X 20) as compared with hemoculture [27].
In agreement with our hypothesis, mice inoculated with TcCRT+ and wild type parasites infected a high percentage of mice, although without detectable differences between these groups.
It is unclear whether TcCRT+/– parasites did infect. However, the possibility that infection occurs is favored by the fact that an adaptive protective status was verified when the animals were challenged 4 months after a primary infection. Since only marginal antibody levels were occasionally detected, protection maybe cellular rather than humoral, issues now under investigation in our laboratories.
Using immunedeficient nu/nu mice, infections caused by the three parasite populations could be detected in a high proportion of mice and even in hemocultures (Table 1). These results confirm previous studies from our laboratory, showing that the TCC wild type strain infects immature or immunocompromised animals [24], [48]. These results suggest that although TcCRT+/– infectivity is attenuated, the suppression of host immunity allows the replication and persistence of these parasites in animals. A similar behavior was observed in the dhfr-ts (dihydrofolate reductase-thymidylate synthase) single mutant, also developed on the TCC T. cruzi strain. This mutant showed a reduced infectivity in immunocompetent mice and as in this work, no mutant parasites could be recovered from hemocultures [20].
The virulence reduction in genetically modified parasites in mice models has previously been reported for genes Tc52 [49] and oligopeptidase B [50]. In our laboratory, this phenomenon was observed working with mutant gp72 genes [19], cub (calmodulin-ubiquitin) [17], lyt1 [18] and dhfr-ts [20].
We have extensively studied the TCC T. cruzi strain as a live attenuated experimental vaccine [28], [29], [51]. The molecular basis of the TCC attenuation is unknown. Thus, we incorporated a rational attenuation mechanism (targeted gene deletion) as a safety device to eliminate the possibility of reversion to a virulent phenotype. In this regard, we tried to rule out the possible reversion of the TcCRT+/– genetic modifications during the chronic stage of the disease in mice. We recovered TcCRT+/– parasites from nude mice at day 90 p.i. (Table 1) and detected sequences corresponding to the TcCRT locus engineering. Thus, the TcCRT+/– mutation is genetically stable in chronically infected mice and there is no reversion to the TCC wild type genotype (Fig.2). Furthermore, the same experiment ruled out strain cross contamination during handling in the laboratory.
Moreover, we tested whether the TcCRT+/– attenuation affects the protective capacity of the TCC wild type parasites against a virulent challenge. Our results suggest that the deletion of one TcCRT allele did not change the already reported immunoprotection induced by TCC wild type parasites [28]. Using TcCRT+/– immunized mice we did not obtain, after a virulent challenge, a sterilizing protective response, although, we achieved low parasite density, mortality (Fig.3A) and a significantly reduced tissue inflammatory response (Fig.4A–F) and spleen index (Fig.4G). Infection with the parental TCC clone has been shown to be protective, in spite of the fact that it generates inflammatory foci in cardiac tissue [23]. When we inactivated one of the TcCRT alleles, a significant decrease in local inflammation is recorded, perhaps a reflection of an impaired virulence. Certainly, the most important fact was that the protective response was achieved at the cost of a possible primary infection with attenuated TcCRT+/– parasites which could not be detected by our most sensitive methods during a six month follow-up and even after immunosuppression of the infected mice (Table 1). It is crucial for vaccinating parasites not to persist in the organism and to discontinue the transmission cycle in peridomestic animals from endemic areas. This could impact over the Chagaś disease infection incidence.
In a previous work [33], mice immunization with TcCRT induced the generation of specific anti-TcCRT antibodies resulting in increased parasitemia of the T. cruzi-challenged mice. Most likely, as mentioned above, immunization with TcCRT induces C1q binding anti-TcCRT antibodies thus increasing the parasite infectivity in the challenged animals. According to this hypothesis, TcCRT+ immunized mice showed higher levels of specific anti-T. cruzi antibodies (Fig 3B) inducing an elevated parasitemia after challenge (Fig 3A). On the contrary, the TcCRT+/– attenuated line did not produce detectable antibodies (Fig 3B) or parasitemia post-challenge (Fig 3A).
Wild type TCC parasites are not detectable by direct blood examination, however, they could be detected by PCR in cyclophosphamide treated chronically infected mice (Table 1) or after hemoculture recovery. In this regard, the attenuated biological behavior of the TcCRT+/– mutants is interesting because if employed as live immunogens, an eventual (natural or induced) immunosuppression of the host should not produce the reactivation of the vaccinating parasites.
Inoculation and eventual infection with TcCRT+/– parasites did not induce detectable antibodies levels (Fig.1 and 3B). However, protection from T. cruzi infection is considered at present to be mediated primarily by cytotoxic T cells [52] and not by antibodies.
In summary, our results show that TcCRT+/– clones were restricted in two important properties conferred by TcCRT and indirectly by C1q: the ability to evade the host immune response, and their virulence status. Therefore, deletion of one copy of the TcCRT gene in the attenuated TCC strain resulted in the generation of a safe and irreversibly gene-deleted live attenuated parasite with high experimental immunoprotective properties.
|
10.1371/journal.ppat.1007348 | Streptococcal Lancefield polysaccharides are critical cell wall determinants for human Group IIA secreted phospholipase A2 to exert its bactericidal effects | Human Group IIA secreted phospholipase A2 (hGIIA) is an acute phase protein with bactericidal activity against Gram-positive bacteria. Infection models in hGIIA transgenic mice have suggested the importance of hGIIA as an innate defense mechanism against the human pathogens Group A Streptococcus (GAS) and Group B Streptococcus (GBS). Compared to other Gram-positive bacteria, GAS is remarkably resistant to hGIIA activity. To identify GAS resistance mechanisms, we exposed a highly saturated GAS M1 transposon library to recombinant hGIIA and compared relative mutant abundance with library input through transposon-sequencing (Tn-seq). Based on transposon prevalence in the output library, we identified nine genes, including dltA and lytR, conferring increased hGIIA susceptibility. In addition, seven genes conferred increased hGIIA resistance, which included two genes, gacH and gacI that are located within the Group A Carbohydrate (GAC) gene cluster. Using GAS 5448 wild-type and the isogenic gacI mutant and gacI-complemented strains, we demonstrate that loss of the GAC N-acetylglucosamine (GlcNAc) side chain in the ΔgacI mutant increases hGIIA resistance approximately 10-fold, a phenotype that is conserved across different GAS serotypes. Increased resistance is associated with delayed penetration of hGIIA through the cell wall. Correspondingly, loss of the Lancefield Group B Carbohydrate (GBC) rendered GBS significantly more resistant to hGIIA-mediated killing. This suggests that the streptococcal Lancefield antigens, which are critical determinants for streptococcal physiology and virulence, are required for the bactericidal enzyme hGIIA to exert its bactericidal function.
| The human immune system is capable of killing invading bacteria by secreting antimicrobial proteins. Cationic human Group IIA secreted phospholipase A2 (hGIIA) is especially effective against Gram-positive bacteria by degrading the bacterial membrane. HGIIA requires binding to negatively charged surface structures before it can penetrate through the thick peptidoglycan layer and reach the target phospholipid membrane. HGIIA is constitutively expressed at high concentrations at sites of possible bacterial entry, e.g. in tears, skin and small intestine. In serum, normal concentrations are low but can increase up to 1,000-fold upon inflammation or infection. In vitro, ex vivo and in vivo experiments suggest an important role for hGIIA in defense against two human pathogens, Group A and Group B Streptococcus (GAS, GBS). We demonstrate that the Lancefield cell wall polysaccharides that are expressed by these bacteria, the Group A Carbohydrate (GAC) for GAS and the Group B Carbohydrate (GBC) for GBS, are required for optimal hGIIA bactericidal efficacy by facilitating penetration through the peptidoglycan layer. Given the increased hGIIA resistance of antigen-modified or antigen-deficient streptococci, it will be of interest to determine potential regulatory mechanisms regarding expression of streptococcal Lancefield polysaccharides.
| Many important human bacterial pathogens are also common colonizers of mucosal barriers. Occasionally, such pathogens penetrate these physical barriers to invade the underlying tissues and cause infections. Antimicrobial molecules, sometimes also referred to as ‘endogenous antibiotics of the host’, are a critical part of the innate immune response to eradicate these intruders and clear the infection. In humans, one of the most potent bactericidal molecules against Gram-positive bacteria is the secreted enzyme human Group IIA phospholipase A2 (hGIIA) [1,2].
HGIIA belongs to a family of 11–12 secreted phospholipase A2 enzymes, which are structurally related and hydrolyze various phospholipids [2–5]. In non-inflamed conditions, hGIIA serum levels are low and not sufficient to kill most Gram-positive bacteria [6]. However, sterile inflammation or infection increases hGIIA expression with concentrations reaching up to 1 μg/ml in serum [7], which is sufficient to kill most Gram-positive pathogens in vitro. A unique feature of hGIIA compared to other secreted phospholipase A2 family members is its high cationic charge, which is required for binding to negatively-charged surface structures and for penetration of the thick peptidoglycan layer surrounding Gram-positive bacteria [2,8,9]. The potent bactericidal activity of hGIIA against Gram-positive bacteria has been demonstrated in vitro, using recombinant hGIIA, and is suggested by infection experiments that show increased protection from infection using hGIIA transgenic mice [10–16].
To counter the bactericidal effects of hGIIA, pathogens have evolved different resistance mechanisms, for example by suppressing hGIIA expression [17,18] or by increasing the net positive charge of surface structures and membrane. The surface modifications include the addition of positively-charged D-alanine moieties to teichoic acid polymers by the highly conserved dlt operon to repulse hGIIA [8] and other cationic antimicrobials [19–22]. In addition, Staphylococcus aureus (S. aureus) modifies the charge of its bacterial membrane through the molecule MprF [23,24] by adding the cationic amino acid lysine to phosphatidylglycerol (PG), resulting in lysyl-PG [25]. In Group A Streptococcus (GAS), the enzyme sortase A (SrtA), a conserved enzyme in Gram-positive bacteria that recognizes proteins with an LPXTG motif and covalently attaches them to peptidoglycan [26,27], was shown to increase hGIIA resistance [12].
Studies with recombinant hGIIA have highlighted differences in intrinsic hGIIA susceptibility between different Gram-positive species, where Bacillus subtilis is killed in the low ng/ml concentration range [28,29], and GAS is one of the most resistant species known to date [12]. Interestingly, this high resistance is not a common trait of streptococcal pathogens since Group B Streptococcus (GBS) is killed by concentrations that are approximately 500 times lower compared to those required to kill GAS [11,12]. Streptococci are historically classified by the expression of structurally different Lancefield antigens [30]. Lancefield antigens are cell wall polysaccharides making up approximately 50% of the dry cell wall mass [31]. All GAS serotypes express the Lancefield Group A carbohydrate (GAC), which consists of a polyrhamnose backbone with alternating N-acetylglucosamine (GlcNAc) side chains [31], which are important for virulence [32]. In contrast, all GBS serotypes express the Lancefield Group B carbohydrate (GBC), a multi-antennary structure, containing rhamnose, galactose, GlcNAc, glucitol, and significant amounts of phosphate [33]. Both streptococcal species are important human pathogens as they can cause systemic infections associated with high mortality and morbidity [34–36]. Mouse infection models and ex vivo studies on human serum from infected patients suggest the importance of hGIIA in defense against lethal infections with GAS and GBS [11,12]. Given the importance of hGIIA in host defense against streptococci, we set out to identify the molecular mechanisms that confer resistance to hGIIA using a comprehensive and unbiased approach.
A previous study found that GAS strains are among the most resistant Gram-positive bacteria regarding hGIIA-mediated killing [12]. Mutation of srtA in the GAS strain JRS4, an emm6 serotype, increased hGIIA susceptibility by about 50-fold [12]. GAS M1T1 is a globally-disseminated emm1 clone that is most often responsible for invasive GAS infections in industrialized countries [37,38] and was not included previously in hGIIA studies [12]. GAS strain 5448, a representative M1T1 isolate, showed concentration-dependent killing by recombinant human hGIIA, with an LD50 of 0.05 μg/ml (S1 Fig). Also, GAS M1T1 resistance mechanisms against hGIIA at least partially overlap with GAS JRS4 emm6, since mutation of srtA rendered GAS M1T1 approximately 35-fold more susceptible to hGIIA (S1 Fig) [12].
We set out to identify additional genes that affect hGIIA susceptibility of GAS M1T1 using the GAS Krmit transposon mutant library [39]. To ensure complete coverage of the library in our experiment, we optimized our hGIIA killing assay to support an inoculum of 107 CFU, using a final concentration of 0.125 μg/ml hGIIA. The Tn-seq experiment with the GAS Krmit transposon mutant library consisted of four non-exposed control samples and four hGIIA-treated samples. Each sample contained on average approximately 30 million reads, of which over 90% of the reads aligned once to the GAS M1T1 5448 reference genome (S1 Table) [40]. To quantify the number of transposon insertions per gene, we divided the reference genome into 25 nucleotide windows, resulting in 73,182 windows, and mapped each read to a specific window. More than 48% of the windows had at least one read aligned. We identified one gene with an exceptionally high number of transposon insertions at a specific part of the gene (M5005_Spy_1390), suggesting biased insertion of the transposon (S2 and S3 Tables and S2 Fig). This gene was therefore excluded from further analysis. No other biased transposon insertion sites were observed.
We identified 16 genes that contained a significantly different number of transposon insertions after exposure to hGIIA as indicated by P-value of <0.05 (Benjamini-Hochberg (BH) corrected; Fig 1A, S3 Fig, and S2–S4 Tables). Nine of the 16 genes (56%) showed a decrease in transposon insertions compared to untreated controls, indicating that the products of the disrupted genes provide resistance against hGIIA-mediated GAS killing (Fig 1A, S3 Fig, S3 Table). Three susceptibility genes are located within the dlt operon (M5005_Spy_1070, M5005_Spy_1072, M5005_Spy_1073), which is responsible for D-alanylation of teichoic acids [41]. Consistently, this operon was previously linked to GAS resistance against other cationic antimicrobials, such as LL-37 and hGIIA [8,42]. The other six genes with significant fold decrease in transposon insertions are annotated as hypothetical proteins (M5005_Spy_0918 and M5005_Spy_1794), a lactoylglutathione lyase (M5005_Spy_0876), LytR (M5005_Spy_1474) of the LytR/CspA/Psr protein family, the transcriptional regulator FabT (M5005_Spy_1495), and the NAD glycohydrolase inhibitor (M5005_Spy_0140) (S3 Fig and S2 and S3 Tables).
Seven genes showed a relative increase in the number of transposon insertions after hGIIA exposure, indicating that the products of these genes are important for hGIIA to exert its bactericidal effect (Fig 1A, S3 Fig, S4 Table). Five of the six genes (83%) mapped to two gene clusters; one gene cluster is annotated as an ABC transporter (M5005_Spy_0939, M5005_Spy_0940, M5005_Spy_0941) and the other gene cluster is the previously identified 12-gene cluster responsible for biosynthesis of the Group A carbohydrate (GAC) (Fig 1B) [32]. Within the GAC gene cluster, gacI and gacH (M5005_Spy_0609 and M5005_Spy_0610) showed significantly increased number of transposon insertions. The small downstream gene gacJ (M5005_Spy_0611) also demonstrated a 3-fold increase, however, the BH corrected P-value is slightly above 0.05. Other genes within the GAC gene cluster are essential or crucial as described previously [39,43]. Finally, guaB (M5005_Spy_1857) and the IIC component of a galactose-specific PTS system (M5005_Spy_1399) were identified as their mutation may confer increased resistance to hGIIA (S3 Fig and S2 and S4 Tables). Overall, the transposon library screen identified genes that confer resistance or are important for the mechanisms of action of hGIIA.
To validate the Tn-seq findings, we confirmed the involvement of three genes (dltA, lytR, and gacI) by comparing hGIIA-mediated killing of WT GAS with previously generated GAS mutants [32,42,44]. Deletion of dltA and lytR indeed increased GAS susceptibility to hGIIA-mediated killing by 45-fold and 35-fold, respectively (Fig 2A and 2B). The dltA defect could be restored by re-introduction of the gene on a plasmid (Fig 2A).
In contrast to dltA and lytR, mutation of gacI, which results in loss of the GAC GlcNAc side chain [45], increased GAS resistance to hGIIA by approximately 10-fold compared to the parental or gacI-complemented (gacI*) strain (Fig 2C). The GAC is conserved in all GAS serotypes. We therefore questioned whether deletion of gacI would have a similar effect on the bactericidal efficacy of hGIIA in four other GAS serotypes (M2, M3, M4, M28). In all serotypes, deletion of gacI increased resistance of GAS to hGIIA by 5- to 50-fold (Fig 2D), indicating that hGIIA requires the GAC GlcNAc side chain for optimal bactericidal efficacy in all genetic backgrounds tested.
To study the activity of hGIIA in a more physiological setting, we spiked pooled normal human serum with different concentrations of recombinant hGIIA. As described previously [32,46], GAS grows in human serum, a trait that is not influenced by the presence of endogenous hGIIA since addition of the hGIIA-specific inhibitor LY311727 [47] did not affect GAS growth in serum (S4A Fig). Addition of recombinant hGIIA to human serum potentiated its bactericidal effect compared to the purified assay as reflected by a 5-fold lower LD50 (0.01 ug/ml; Fig 3A versus Fig 2). Interestingly, heat-inactivation of serum reduced the ability of hGIIA to kill GAS by 10-fold compared to active serum, indicating that there are heat-labile factors in serum that potentiate hGIIA efficacy (Fig 3A).
We determined how the addition of serum would affect the efficacy of hGIIA to kill the mutants with altered hGIIA susceptibility. We first compared bacterial survival of the WT strain and the individual mutants in normal serum (S4A Fig). Interestingly, the lytR and srtA mutant already showed a significant loss of fitness in non-inflamed serum, which is not attributed to the presence of endogenous hGIIA as addition of LY311727 did not restore their survival (S4A Fig). Both ΔsrtA and ΔdltA bacteria remained more susceptible to hGIIA-mediated killing in serum (Fig 3B and 3C), whereas the ΔlytR and ΔgacI mutants were now equally resistant to WT GAS (Fig 3D and 3E). These results reflect the multitude of effects that occur simultaneously in a complex environment such as serum. More specifically, serum likely contains factors that have an opposite effect to hGIIA on lytR and gacI mutants, such that the net survival of these mutants is equal to WT. Finally, we compared the effect of serum heat-inactivation on hGIIA efficacy in the context of individual mutants (S4B–S4D Fig). Similar to WT GAS, heat inactivation of serum reduced the efficacy of hGIIA to kill ΔsrtA, ΔdltA and ΔlytR, suggesting that the hGIIA-potentiating factor(s) is required to kill all mutants in our panel.
Our observation that GAS ΔgacI is more resistant to hGIIA implies that the GAC GlcNAc moiety is important for the function of hGIIA. To assess whether loss of the GAC GlcNAc side chain affected hGIIA binding to bacteria, we first analyzed binding of hGIIA by fluorescence microscopy using a phospholipase A2-specific antibody (Fig 4A). A visual quantification of hGIIA-stained bacteria indicated reduced binding of hGIIA in the absence of GAC GlcNAc (Fig 4C). In addition, we observed that the localization of hGIIA on the bacterial surface was affected, where hGIIA predominantly localized to the GAS cell poles in WT bacteria (Fig 4A and 4B), but distribution became more disperse upon mutation of gacI (Fig 4A and 4B). Since fluorescence microscopy did not allow for more extensive binding assessments, we also quantified binding of recombinant hGIIA to GAS by flow cytometry. At concentrations up to 1 μg/ml, we did not observe differences in hGIIA binding to the three strains (Fig 4D). Only at concentrations of 5 μg/ml, hGIIA showed reduced interaction with the gacI mutant compared to GAS WT and gacI*-complemented strains (Fig 4D). The contribution of differential hGIIA binding to GAS is therefore only relevant to specific locations such as in tears which contain up to 30 μg/ml hGIIA [28]. Since hGIIA binding is charge-dependent, we analyzed whether reduced binding at high hGIIA concentrations could be caused by difference in surface charge. Using the highly cationic protein cytochrome C, we indeed observed that the gacI mutant has a reduced negative surface charge compared to GAS WT and the gacI*-complemented strain (S5 Fig), which could likely explain the reduced binding of hGIIA to this mutant.
Cell wall architecture can significantly affect hGIIA cell wall penetration [2]. To assess how absence of the GAC GlcNAc side chain affected hGIIA cell wall penetration, we measured changes in membrane depolarization over time using the fluorescent voltage-sensitive dye DiOC2(3) [48]. In this assay, membrane depolarization results in reduced red fluorescence. HGIIA required at least 5 minutes to penetrate the GAS cell wall since no changes in red fluorescence signal were observed at this time point for any of the strains (S6A Fig). At 30 minutes (S6B Fig), membrane depolarization occurred as visualized by diminished red fluorescence at hGIIA concentrations of 0.1 μg/ml in the GAS WT and the gacI*-complemented strain. Compared to these two strains, the gacI mutant exhibited limited effects on membrane potential at all time points and all hGIIA measured (Fig 4E and S6 Fig). These data suggest that hGIIA reaches the membrane faster in the presence of GAC GlcNAc moieties.
Membrane depolarization likely precedes more pronounced hGIIA-mediated disruption of the membrane that would allow influx of the fluorescent DNA dye SYTOX green, which can only enter damaged membranes [49]. As expected, hGIIA increased the SYTOX signal in GAS WT and GAS gacI* in both a time and concentration-dependent manner (Fig 4F and S7A–S7E Fig). Importantly, addition of LY311727 completely prevented SYTOX influx (S7F Fig), confirming that our assay indeed reflects hGIIA phospholipase activity on the bacterial membrane. In sharp contrast, SYTOX intensity in GAS ΔgacI increased at a much slower rate and never reached the levels of GAS WT and GAS gacI* after two hours. The observed differences in kinetics and severity of hGIIA on membrane depolarization and SYTOX influx in GAS ΔgacI compared to GAS WT suggest that the GAC GlcNAc side chain is essential for efficient trafficking of hGIIA through the GAS cell wall.
A recent study demonstrates that GacI is a membrane protein that is required for the intracellular formation of undecaprenyl-P-GlcNAc [45]. Therefore, loss of GacI could alter membrane composition and fluidity to impact the activity of hGIIA on the membrane. To analyze whether phospholipid hydrolysis is affected in GAS ΔgacI, we performed the SYTOX influx assay on protoplasts [50]. Unlike the previous SYTOX results with intact bacteria, protoplasts from WT, ΔgacI and gacI* strains all became SYTOX positive (Fig 4G and S8 Fig), underlining our conclusion that the presence of the cell wall in the ΔgacI limits access of hGIIA to the streptococcal membrane. Nonetheless, the significantly lower SYTOX in the ΔgacI protoplasts compared to the WT and gacI*-complemented protoplasts (Fig 4G and S8 Fig), suggests that the absence of GacI has a minimal impact on hGIIA degradation. To further reinforce this conclusion, we determined the levels of phosphatidylglycerol (PG) in bacteria and protoplasts after treatment with hGIIA (Fig 4H). PG levels were significantly higher in GAS ΔgacI after hGIIA treatment compared to WT, whereas equal PG levels were observed in GAS ΔgacI and WT after hGIIA treatment (Fig 4H). We therefore conclude that cell wall trafficking and not cell membrane differences are the major determinant of susceptibility differences between GAS WT and ΔgacI mutant.
We investigated whether the importance of the GAC for hGIIA activity could be extended to other streptococci such as GBS. As previously described, GBS are generally more sensitive to hGIIA compared to GAS [12]. Indeed, killing of GBS strain NEM316 occurred at substantially lower concentrations of hGIIA compared to GAS M1T1 (compare Figs 5A and 2), also in the presence of serum (S9 Fig). We confirmed that killing depends on the catalytic activity of the enzyme since introduction of an inactivating point mutation (H48Q; Fig 5B) or addition of LY311727 abrogated all killing (Fig 5C). Just as the GAC is the molecular signature for GAS, GBS uniquely express another Lancefield antigen, known as the Group B Carbohydrate (GBC). The GBC is a more complex structure compared to the GAC and contains significant amounts of phosphate that introduce a negative charge. Unfortunately, there are currently no GBS mutants available with specific structural variations in the GBC. Instead, we assessed the effect of the complete GBC, through deletion of gbcO [33], on susceptibility of GBS to hGIIA. Deletion of gbcO rendered GBS at least 100-fold more resistant to hGIIA compared to GBS WT (Fig 5A–5C), and the phenotype is restored upon complementation with gbcO on a plasmid (Fig 5A). We could reproduce the ΔgbcO phenotype by treating WT GBS with tunicamycin, an inhibitor of gbcO-type transferases (Fig 5D) [33,51]. Finally, as observed in GAS, fluorescence microscopy demonstrated that hGIIA bound to the poles of GBS WT (Fig 5E and 5F). Unlike to GAS, we did not observe that loss of GBC expression reduced binding of hGIIA at higher concentration of hGIIA as assessed by flow cytometry (S10 Fig). In conclusion, these results highlight a key role for streptococcal Lancefield antigens in the bactericidal effect of hGIIA.
Intrinsic resistance to acute phase protein hGIIA varies among Gram-positive bacteria, including among closely-related streptococcal species. GAS, an important cause of lethal infection worldwide, is among the most resistant bacteria, whereas GBS, an important cause of neonatal sepsis and meningitis, is killed by hGIIA at concentrations that are approximately 500-fold lower [12]. For GAS, we confirmed the role of Sortase A and DltA and identified LytR as hGIIA resistance factors. Despite the differences in cell wall composition, i.e. cell wall crosslinking, cell wall associated proteins and membrane physiology, the streptococcal Lancefield antigens are structural requirements for the activity of hGIIA in both GAS and GBS.
HGIIA is approximately 10-fold more effective against GAS when spiked into normal serum compared to heat-inactivated serum, and 5-fold more effective compared to our ‘purified’ system. This corresponds to a previous observation where hGIIA activity was approximately 10-fold greater in serum or plasma than in the protein-depleted serum in studies using S. aureus as the target pathogen [52]. This suggests the presence of a heat labile protein in serum that facilitates hGIIA-mediated killing of Gram-positive bacteria. Heat-inactivation of serum is a well-established method to study the influence of the complement system and also abolishes hGIIA activity in acute phase serum [53]. Since the low basal levels of hGIIA in normal human serum are not sufficient to affect GAS viability, the enhancement could indicate a synergistic effect between hGIIA and the complement system. A recent study shows formation of the Membrane Attack Complex (MAC) on the GAS surface without affecting bacterial viability [54]. It is therefore tempting to speculate that MAC is deposited on Gram-positive bacteria so that bactericidal enzymes, such as hGIIA, can reach the bacterial membrane more easily. Such a cooperative effect between different innate defense mechanisms would not be surprising, since previous studies have already observed that hGIIA synergizes with neutrophil oxygen-dependent mechanisms to kill S. aureus [55,56]. Finally, the concentrations of hGIIA that are measured in human serum are likely underestimating the true availability of this bactericidal enzyme since hGIIA attaches to surfaces of blood vessels due to its hydrophobic nature. We speculate that vessel-attached hGIIA may help prevent bacterial dissemination to other tissues, an effect that has not yet been addressed experimentally.
Sortase A, an enzyme that links LPXTG-containing proteins to peptidoglycan, was previously described as a hGIIA resistance factor in GAS serotype M6 [12]. We confirmed that deletion of srtA in a GAS M1T1 background similarly sensitizes GAS to hGIIA both in a ‘purified’ as well as a serum environment. Whether a single or multiple LPXTG proteins confer resistance is an unresolved question. Our study suggests that Sortase A-mediated resistance is not caused by a single LPXTG protein since we did not identify a single LPXTG-encoding gene in the Tn-seq screen (S5 Table). Possibly, the underlying mechanism is similar to the SrtA-dependent resistance of GAS to the antimicrobial peptide cathelicidin [46], which depends on the accumulation of sorting intermediates at the bacterial membrane. SrtA itself was not identified in the transposon library screen since the mutants are lost in the competitive environment likely due to inherent defects in growth [39].
We identified and confirmed a role for the protein LytR in GAS hGIIA resistance. LytR is a member of the LytR-CpsA-Psr (LCP) protein family, a conserved family of cell wall assembly proteins in Gram-positive bacteria [57]. The GAS genome encodes two members of this family, lytR (M5005_Spy_1474) and psr (M5005_Spy_1099). The fact that we only identified LytR suggests that these proteins have non-redundant, but as yet unidentified, functions. In several Gram-positive pathogens, including Streptococcus pneumoniae, S. aureus and Bacillus anthracis, LCP proteins anchor cell wall glycopolymers such as wall teichoic acid (WTA), lipoteichoic acid (LTA) and capsular polysaccharides to the cell envelope and are therefore critical for cell envelope assembly and virulence [57–62]. Additionally, lytR homologues in Bacillus subtilis and Streptococcus mutans contribute to cell wall remodeling by increasing autolysin activity [63,64]. Previously, hGIIA activity has been linked to autolysins; autolysin-deficient mutants are more resistant to hGIIA than their parent strain [65]. A suggested mechanism is that hGIIA displaces positively-charged autolysins from negatively-charged WTA and LTA, resulting in localized peptidoglycan digestion and facilitated movement of hGIIA through the cell wall. Currently, the role of LytR either in GAS cell wall assembly or in the regulation of autolysin activity is not known, but LytR-deficient GAS display altered membrane integrity and potential [66], which could impact hGIIA susceptibility. Moreover, lytR has been linked to GAS virulence in two different studies. In the first study, lytR mutants in two different GAS M1 backgrounds showed a more virulent phenotype in a subcutaneous murine model of infection, which was suggested to be a result of increased SpeB activity [66]. LytR-mediated regulation of SpeB is unlikely to play a role in hGIIA-mediated resistance in our experiments, since we used washed bacteria. In a more recent study, lytR mutants in GAS 5448 M1T1 showed a competitive disadvantage for fitness in vivo upon mixed subcutaneous infection [44]. Unfortunately, there is no information regarding pathology or survival of the mice upon infection with the lytR mutant added alone [44].
We also identified genes that render GAS more susceptible to hGIIA. GacH, gacI, and gacJ are located in the biosynthesis gene cluster of the GAC, which may suggest that the GAC is a target for hGIIA on the GAS surface. Mutation of gacI and gacJ results in loss of the GAC GlcNAc side chain [32,45], whereas mutation of gacH does not affect side chain formation [32]. We therefore hypothesize that the GAC provides hGIIA resistance through two distinct mechanisms. First, a gacI/J-dependent mechanism that works through the GAS GlcNAc side chain as important for binding and penetration of hGIIA to the cell membrane. The second mechanism involves GacH but the underlying molecular aspects remain to be determined. The first mechanism seems to conflict with our previous observations that GlcNAc-deficient GAS have decreased virulence capacity due to increased neutrophil killing and increased susceptibility to antimicrobials in serum including LL-37 [32]. However, hGIIA would not have contributed to in vitro assays since we used non-inflamed serum or plasma where basal hGIIA concentrations are too low to affect GAS viability [32]. The fact that gacI mutants demonstrate reduced survival in vivo suggests that the benefits of expressing the GlcNAc side chain outweigh the increased susceptibility to hGIIA. Since GAS already shows high intrinsic resistance towards hGIIA there is no pressure to lose the GlcNAc side chain. It might even be detrimental since it makes GAS more vulnerable to effects of other antimicrobials or yet unidentified host defenses. In contrast to the GAC [31,67], the GBC is a multi-antennary structure and contains anionic charge due to the presence of phosphate [33]. For GBS, the increased hGIIA resistance in GBC-negative gbcO mutants is therefore likely explained by the loss of negatively charged groups on the surface. This corresponds to previous observations in S. aureus, where loss of the secondary cell wall glycopolymer WTA, increased resistance to several antimicrobial proteins, including hGIIA [10].
Binding of hGIIA to streptococci was reduced when bacteria expressed a modified GAC or lacked complete expression of GBC but these differences were only apparent using high hGIIA concentrations. However, these findings need to be interpreted with caution since possibly only a small portion of the bound hGIIA is required for the bactericidal action of the enzyme. Therefore, even small fluctuations in binding might result in meaningful functional differences. We are currently not able to analyze hGIIA binding at a more sensitive level.
Contrary to our expectations, fluorescence microscopy analysis showed that hGIIA bound to the cell poles of both GAS and GBS. However, the observed binding pattern does not correspond to the reported localization of the GAC and GBC, which are distributed over the entire cell wall as shown by early electron microscopy studies [68,69]. Binding at the septa of dividing bacteria seems a preferred binding site for bactericidal agents due to a high turnover of peptidoglycan which would make penetration easier [70,71]. In addition, the septum is rich in anionic phospholipids [72], a likely target for cationic hGIIA. Finally, the GAS ExPortal, a unique microdomain in the GAS membrane that is enriched in anionic lipids, would be another favored location of binding for the cationic hGIIA [73]. However, the ExPortal is distributed asymmetrically across the GAS surface and not at the cell poles [73]. The fact that we observe a similar binding pattern to GBS and GAS, may indicate that GAS and GBS express a similar protein that localizes at the cell poles and is used by hGIIA as an initial docking site. Importantly, localization became more disperse upon deletion of gacI in GAS, possibly suggesting a redistribution of hGIIA-interacting structures. Identification of hGIIA susceptible and resistant GBS mutants using a Tn-seq mutant transposon library may help identify such conserved or homologous hGIIA targets in the GAS and GBS cell wall.
Lack of the GAC GlcNAc side chain most profoundly affected penetration of hGIIA through the cell wall, a mechanism that depends on charge [2,9]. Indeed, membrane depolarization and permeabilization occurs at a much slower rate in the gacI mutant compared to WT and complemented strains. This implies that the GAC GlcNAc side chain facilitates penetration of hGIIA through the cell wall in what is referred to as an ‘anionic ladder process’ [2]. Interestingly, the GAC does not contain any charged structures. Therefore, the underlying mechanism may be linked to the previously mentioned autolysin displacement from interaction with the GAC.
In conclusion, we show that the bactericidal agent hGIIA is able to kill GAS in a complex serum environment. However, modification or removal of the Lancefield antigen renders GAS more resistant to the bactericidal activity of hGIIA. Similarly, removing the Lancefield antigen from GBS renders this species also more resistant to the bactericidal activity of hGIIA. The Lancefield antigens, previously thought to be solely involved in physiology, are thus critical cell wall structures for hGIIA to exert its bactericidal effect. The Tn-seq data discussed in this paper provide exciting new insights into the resistance mechanisms of GAS and encourage similar experiments in other streptococci species. Disrupting the resistance mechanisms with therapeutic agents could possibly be sufficient to provide our own immune system the upper hand in clearing invading streptococcal pathogens.
The GAS M1T1 5448 strain was used in this study unless stated otherwise. The 5448ΔgacI knockout and gacI* complemented strain [32], the 5448ΔlytR [44] and the GAS serotypes M2, M3, M4, and M28 and corresponding ΔgacI knockouts [74] were described previously. Preparation and characterization of the GAS M1T1 5448 transposon library was described previously by Le Breton et al., 2015 [39]. All GAS strains were grown in Todd-Hewitt broth (Becton Dickinson) supplemented with 1% yeast extract (Oxoid; THY) as static cultures at 37°C. Kanamycin (Sigma-Aldrich) was used at a concentration of 300 μg/ml when appropriate. GBS NEM316 WT, ΔgbcO and the complemented strains ΔgbcO/pTCV were kindly provided by Dr. Mistou [33].
Unless stated otherwise, overnight cultures of GAS were diluted and re-grown to mid-log phase (OD600nm = 0.4), washed and resuspended in HEPES solution (20 mM HEPES, 2 mM Ca2+, 1% BSA [pH 7.4]) solution at OD600nm = 0.4 (~1x108 CFU/ml). For GBS strains, overnight cultures of NEM316 WT, ΔgbcO and the complemented strains ΔgbcO/pTCV were diluted in TH broth and grown to mid-log phase (OD620nm = 0.4 for WT and complemented strains, 0.25 for ΔgbcO mutant). Bacteria were then diluted in HEPES solution and pushed rapidly through a 27-gauge needle, a process repeated three times, to disrupt bacterial aggregates. Normal human serum and heat-inactivated serum was obtained from healthy volunteers as described previously [54].
Recombinant hGIIA was produced as described previously [75]. The GAS M1T1 Krmit transposon mutant library was grown to mid-log phase in 100 ml THY containing Km and resuspended in HEPES solution to OD600nm = 0.4. Four experimental replicates of 100 μl (~ 1x107 CFU) were subsequently incubated in HEPES solution with or without 125 ng/ml hGIIA for 1 hour at 37°C. After incubation, 3 ml THY was added to all samples and incubated at 37°C until the mid-log phase was reached (recovery step). Cultures were collected by centrifugation and used for isolation of genomic DNA (gDNA). gDNA was isolated by phenol-chloroform extraction. Samples were barcoded and prepared for Tn-seq sequencing as described previously [76]. Tn-seq sequencing was performed on Illumina NextSeq500 (Sequencing facility University Medical Center, Utrecht, The Netherlands).
Tn-seq data analysis was performed as previously described [76]. In short, barcodes were split using the Galaxy platform [77] and sequences were mapped to the GAS M1T1 5448 genome [40] using Bowtie 2 [78]. The genome was subsequently divided in 25-bp windows and each alignment was sorted and indexed by IGV [79]. Insertions were counted per window and then summed over the genes. Read counts per gene were adjusted to cover only the first 90% of the gene since transposon insertions in the final 10% potentially do not cause a knock-out phenotype. Then, read counts were normalized to the total number of reads that mapped to the genome in each replicate, by calculating the normalized read-count RKPM (Reads Per Kilobase per Million input reads; RKPM = (number of reads mapped to a gene x 106) / (total mapped input reads in the sample x gene length in kbp)). Cyber-T [80] was used to perform statistical analysis on the RKPM values. Genes that contributed to either hGIIA susceptibility or hGIIA resistance were determined when the Benjamini-Hochberg (BH) corrected p-value was <0.05. Illumina sequencing reads generated for the Tn-seq analysis were deposited in the European Nucleotide Archive under the accession number PRJEB27626.
Mid-log streptococcal suspensions were diluted 1,000 times in HEPES solution and 10 μl was added to sterile round-bottom 96 well plates (triplicates). Recombinant hGIIA or catalytically-deficient hGIIA mutant enzyme H48Q was serially diluted in HEPES solution or human serum and 10 μl aliquots were added to bacteria-containing wells. For hGIIA inhibition experiments, 50 μM LY311727 was added to the HEPES solution or serum. For GAS, samples were incubated for 2 hours at 37°C, without shaking, PBS was added and samples were 10-fold serially diluted and plated on THY agar plates for quantification. For GBS, bacteria were incubated with hGIIA at 37°C for 30 minutes, the samples were diluted in PBS and plated onto blood agar plates. After overnight incubation 37°C, colony forming units (CFU) were counted to calculate the survival rate (Survival (% of inoculum) = (counted CFU * 100) / CFU count of original inoculum or Survival (%) = (counted CFU * 100) / CFU count at 0 μg/ml hGIIA). For pharmacological inhibition of GBC expression, NEM316 WT bacteria were grown to mid-log phase (OD620nm = 0.4) in the presence of 0.5 mg/ml tunicamycin (Sigma) and used in killing assays as described above.
Changes in hGIIA-dependent membrane potential were determined using the membrane potential probe DiOC2(3) (PromoKine) [48,81]. Bacterial suspensions (OD600nm = 0.4) were diluted 100 times (~1x106 CFU/ml), 100 μl aliquots were divided into eppendorf tubes and incubated with serial dilutions of hGIIA. After incubation at 37°C, 3 mM DiOC2(3) was added and incubated at room temperature for 5 minutes in the dark. Changes in green and red fluorescence emissions were analyzed by flow cytometry.
Bacterial staining with the DNA stain SYTOX Green (Invitrogen) is a measurement for membrane permeabilization and an indication of bacterial cell death [49]. Serial dilutions of hGIIA in HEPES solutions were added to wells of a sterile flat-bottom 96 well plate. Bacteria were resuspended in HEPES solution containing 1 μM SYTOX green (OD600nm = 0.4) and added to hGIIA dilutions in a final volume of 100 μl. For hGIIA inhibition experiments, 500 μM LY311727 was added. Fluorescence over time was recorded using FLUOstar OPTIMA (green fluorescence 530 nm emission and excitation 488 nm) at 37°C.
Bacterial surface charge was determined as previously described [81]. Briefly, exponential phase bacteria (OD600nm = 0.4) were washed twice in 20 mM MOPS buffer [pH 7.0] and adjusted to OD600nm = 0.7. After a 10-fold concentration step, 200 μl bacterial aliquots were added to 200 μg cytochrome c (from Saccharomyces cerevisiae, Sigma-Aldrich) in a sterile 96-well round-bottom plate. After 10 minutes at room temperature in the dark, the plate was centrifuged, the supernatant was transferred to a sterile 96 well flat-bottom plate and absorbance was recorded at 530 nm. The percentage of bound cytochrome c was calculated using samples containing MOPS buffer only (100% binding) and samples containing MOPS buffer and cytochrome c (0% binding).
To determine hGIIA surface binding, 12.5 μl of bacterial cultures in mid-log phase (OD600nm = 0.4 and 0.25 for GBS ΔgbcO) were added to wells of a sterile 96-well round-bottom plate (triplicates). hGIIA was serially diluted in HEPES solution without Ca2+ and added to the bacteria at indicated concentrations. After 30 minutes incubation at 4°C, bacteria were collected by centrifugation and resuspended in HEPES solution without Ca2+ containing 1:300 dilution of anti-phospholipase A2 antibody (Merck Millipore) [28]. After incubation at 4°C for 30 minutes, the samples were washed and incubated with a 1:1,000 dilution of FITC-labeled goat-anti-mouse IgG (SouthernBiotech) or a 1:500 dilution of Alexa Fluor 647 conjugated goat-anti-mouse IgG (Jackson Immuno Research). After washing with HEPES solution without Ca2+, samples were fixed with 1% paraformaldehyde and fluorescence was recorded by flow cytometry (FACSVerse, BD Biosciences).
To analyze hGIIA surface localization by microscopy, bacteria were grown in 10 ml broth to mid-log phase and washed with 0.1 M NaHCO3 [pH 9]. For GBS, the bacterial septa were stained by addition of a 1:1 mixture of Vancomycin bodipy FL conjugate (Invitrogen, V34550) and vancomycin (Sigma) at a final concentration of 1.25 μg/ml, during the last generation time of growth. The surface of GBS was stained with Alexa Fluor 350 Carboxylic acid Succinimidyl ester (Molecular Probes by Life Technologies, A10168) for 1 hour in room temperature. Bacteria were then resuspended in 500 μl HEPES solution and the suspension was divided over two tubes. A final concentration of 10 μg/ml hGIIA H48Q was added to one tube and HEPES solution to the other before a 30 min incubation at room temperature. The samples were washed and resuspended in 200 μl HEPES solution, then again divided to two tubes. A mouse anti-human hGIIA monoclonal antibody (Clone SCACC353 Cayman Chemical) or an IgG1 isotype control (mouse anti human IgA clone 6E2C1, DAKO) was added to a final concentration of 10 μg/ml to the bacterial suspensions and incubated at RT. After washing, the samples were incubated with 8 μg/ml of Alexa Flour 594 goat anti-mouse IgG1 (Molecular Probes by Life Technologies, A21125). After 30 min incubation, the samples were washed in HEPES solution and fixed in 4% paraformaldehyde. Ten μl of bacterial suspension were mounted onto microscopic slides (VWR) using MOWIOL (Sigma) mounting medium before viewing the samples using Zeiss Axiovert 200M microscope. Pictures were captured using a 63× objective and the AXIOVISION 4.8 software.
To determine hGIIA efficacy in hydrolyzing membrane phospholipids, the membrane permeabilization assay was modified for protoplasts. Mid-log bacterial suspension were prepared in in protoplast buffer (20% sucrose, 20 mM Tris-HCl, 10 mM MgCl2, 2 mM CaCl2 [pH 7.4]) containing 1.4 units/μl mutanolysin (Sigma-Aldrich) [50,82,83]. After incubation for 1 hour at 37°C, protoplasts were collected by centrifugation (1,200 rpm 15 minutes) and resuspended in protoplast buffer to an OD600nm = 0.4. Pore formation by hGIIA was monitored using SYTOX Green as described above.
Approximately 3*107 CFU from a mid-log bacterial suspension in HEPES solution, or protoplasts in protoplast buffer, were exposed to 2 μg/ml hGIIA for 30 minutes. Afterwards, bacterial suspensions were centrifuged at 140,000 rpm for 4 minutes and bacterial pellets were resuspended in MeOH. The protoplast suspensions were mixed with MeOH 1:1. Bacterial lipids were extracted under acidic conditions in the presence of 10 pmol PG standards (PG 14:1/14:1, PG 20:1/20:1 and PG 22:1/22:1) as described [84]. Lipid extracts were resuspended in 60 μl methanol and diluted 1:10 in 96 wells plates (Eppendorf twintec 96, colorless, Sigma, Z651400-25A) prior to measurement. Measurements were performed in 10 mM ammonium acetate in methanol. Samples were analyzed on an AB SCIEX QTRAP 6500+ mass spectrometer (Sciex, Canada) with chip-based (HD-D ESI Chip, Advion Biosciences, USA) electrospray infusion and ionization via a Triversa Nanomate (Advion Biosciences, Ithaca, USA) as described [84]. PG species were measured by neutral loss scanning selecting for neutral loss of m/z 189. Data evaluation was done using LipidView (ABSciex).
GraphPad Prism 6 was used to perform statistical analysis. An unpaired two-tailed Student’s t-test was used to compare the means of two groups. A 2-way ANOVA with Bonferroni multiple comparison test was used to compare multiple groups. Data shown are mean ± SD.
|
10.1371/journal.ppat.1003916 | A Gammaherpesvirus Bcl-2 Ortholog Blocks B Cell Receptor-Mediated Apoptosis and Promotes the Survival of Developing B Cells In Vivo | Gammaherpesviruses such as Epstein-Barr virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV, HHV-8) establish lifelong latency in their hosts and are associated with the development of several types of malignancies, including a subset of B cell lymphomas. These viruses are thought to co-opt the process of B cell differentiation to latently infect a fraction of circulating memory B cells, resulting in the establishment of a stable latency setpoint. However, little is known about how this infected memory B cell compartment is maintained throughout the life of the host. We have previously demonstrated that immature and transitional B cells are long-term latency reservoirs for murine gammaherpesvirus 68 (MHV68), suggesting that infection of developing B cells contributes to the maintenance of lifelong latency. During hematopoiesis, immature and transitional B cells are subject to B cell receptor (BCR)-mediated negative selection, which results in the clonal deletion of autoreactive B cells. Interestingly, numerous gammaherpesviruses encode homologs of the anti-apoptotic protein Bcl-2, suggesting that virus inhibition of apoptosis could subvert clonal deletion. To test this, we quantified latency establishment in mice inoculated with MHV68 vBcl-2 mutants. vBcl-2 mutant viruses displayed a marked decrease in the frequency of immature and transitional B cells harboring viral genome, but this attenuation could be rescued by increased host Bcl-2 expression. Conversely, vBcl-2 mutant virus latency in early B cells and mature B cells, which are not targets of negative selection, was remarkably similar to wild-type virus. Finally, in vivo depletion of developing B cells during chronic infection resulted in decreased mature B cell latency, demonstrating a key role for developing B cells in the maintenance of lifelong latency. Collectively, these findings support a model in which gammaherpesvirus latency in circulating mature B cells is sustained in part through the recurrent infection and vBcl-2-mediated survival of developing B cells.
| Gammaherpesviruses such as Epstein-Barr virus and Kaposi's sarcoma herpesvirus are widespread pathogens that establish lifelong infections in a dormant state termed latency. Although most gammaherpesvirus infections are asymptomatic, infection of some individuals leads to the development of B cell lymphoma or other cancers. It is well known that during latency these viruses reside in mature B cells of the immune system; however, little is known about how this reservoir is maintained for life. Using murine gammaherpesvirus 68 infection of mice as a model to study gammaherpesvirus infections inside a living host, we have previously demonstrated that gammaherpesviruses can infect early precursors of B cells. In normal situations, the differentiation of such precursors into mature B cells is a tightly regulated process that leads to the death of cells that react inappropriately to host tissues. Here though, we demonstrate that a gammaherpesvirus protein called vBcl-2 can block the death of infected precursor B cells, and that vBcl-2 is critical for infection of these cells. Finally, we show that depleting precursor B cells reduces mature B cell latency. Together, these data suggest that vBcl-2 proteins play a key role in lifelong gammaherpesvirus latency and may be a potent target for future drug development.
| The human gammaherpesviruses, Epstein-Barr virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV, HHV-8), and the genetically- and pathogenically-related murine gammaherpesvirus 68 (MHV68, γHV68, MuHV-4), establish lifelong latent infections in circulating B cells. B cells are a crucial component of the adaptive immune response as they are capable of mounting responses to an enormous range of antigens through the production of antibodies and the establishment of immunological memory. Hence, maintaining a fully functional and diverse B cell population is critical for protection against a variety of bacterial and viral infections. Although gammaherpesvirus infections have been linked with the development of a considerable number of malignancies including B cell lymphomas and Kaposi's sarcoma, such pathogenic outcomes occur rarely in healthy hosts and have vastly increased prevalence in immunosuppressed populations [1]–[3]. Thus, gammaherpesviruses have evolved a symbiotic relationship with the host immune system in which they are able to maintain lifelong infection in B cells without significantly altering normal B cell function or repertoire.
The most widely held model for latency establishment posits that gammaherpesviruses have evolved mechanisms to mimic natural B cell activation pathways, such that infection of naïve follicular B cells results in their activation and subsequent differentiation to memory B cells [4]. The model contends that lifelong infection is maintained because latent virus is indefinitely retained in this long-lived pool of circulating, resting memory B cells. Work from Thorley-Lawson's group has provided important in vivo support for this concept by demonstrating that in chronically infected individuals EBV genome is maintained in a frequency of circulating memory B cells that, while variant among individuals, remains stable over time, suggesting that B cell homeostatic mechanisms maintain a lifelong latency setpoint [5]. Similarly, during chronic infection MHV68 is primarily restricted to class-switched memory B cells [6], [7] and is maintained at a stable frequency over time [8].
While work with both EBV and MHV68 support the basic concept that virus-driven mature B cell differentiation contributes to lifelong latency, it remains unclear how memory B cell infection is maintained at a steady setpoint. The two most prevalent hypotheses hold that maintenance of the infected memory B cell pool occurs via reactivation of latent virus and reseeding naïve B cells, with subsequent virus-driven differentiation to memory B cells [9], [10], or via homeostatic proliferation, with virus episome replication and segregation to daughter cells [5]. However, one intriguing alternate possibility is that lifelong latency is facilitated by continual infection of newly generated developing B cells, which subsequently follow normal B cell maturation pathways. In support of this concept, newly formed splenic CD21−CD23− B cells have been reported to carry MHV68 genome [11], [12], and we have recently demonstrated that developing B cells harboring MHV68 genome are present in both the bone marrow (pro-B/pre-B and immature B) and the spleen (transitional B) throughout chronic infection [13]. However, the lifespan of these cells is only 24 to 72 hours; thus, these findings suggest that (a) MHV68 recurrently infects developing B cells or (b) MHV68 indefinitely extends the life of developing B cells. Because hematopoiesis results in the daily generation of new immature B cells which in turn maintain the mature B cell population [14]–[18], recurrent or stable infection of these early stage cells could allow gammaherpesviruses to continually access the memory B cell compartment.
During B cell maturation, the stochastic process of V(D)J recombination results in randomly generated B cell receptors (BCRs) on developing B cells. To guard the host against the generation of functional autoreactive mature B cells, immature and transitional B cells must navigate through multiple negative selection checkpoints. In the processes of central tolerance in the bone marrow and peripheral tolerance in the spleen, B cells that react with self-antigen are eliminated through apoptotic clonal deletion, are made anergic, or are subjected to further BCR editing [19]–[25]. BCR binding to self-antigen triggers the apoptotic death of immature and transitional B cells due in part to the low expression of host anti-apoptotic proteins Bcl-2, Bcl-XL and A1 in these cells [26]–[31]. Consistent with this, enforced expression of host Bcl-2 or Bcl-XL in vivo allows the survival of autoreactive immature and transitional B cells [25], [32]–[34].
Notably, several gammaherpesviruses encode orthologs of host anti-apoptotic proteins. For example, EBV BHRF1, EBV BALF1, and KSHV ORF16 all encode proteins with homology to Bcl-2, and KSHV K13 encodes a protein with homology to the human FLICE inhibitory protein (FLIP) [35]–[39]. Similarly, MHV68 M11 encodes a Bcl-2 ortholog (vBcl-2) [40], [41] that is expressed during latency [12], [42]. While the specific molecular role that MHV68 vBcl-2 plays in B cell infection has not been determined, it is capable of blocking Fas-, TNFα-, Sindbis virus- and dexamethasone-induced apoptosis [41], [43]–[45], as well as rapamycin- and starvation-mediated autophagy [46], [47]. In vivo, MHV68 M11 mutants have demonstrated no significant defects during acute infection and only minor defects during latency [43], [48]–[50]. Thus, EBV, KSHV and MHV68 all encode vBcl-2 orthologs that retain anti-apoptotic activity. However, it is unknown whether any of these proteins play a role in promoting the survival of developing B cells.
To further define the role that gammaherpesvirus Bcl-2 orthologs play during chronic infection, we tested whether MHV68 vBcl-2 can promote the survival of developing B cells that undergo BCR-mediated selection. Using MHV68 mutant viruses, we found that vBcl-2 played a critical role in infection of immature and transitional B cells in vivo which could be complemented by host Bcl-2. Further, we found that ectopically-expressed vBcl-2 protected immature B cells from BCR-mediated apoptosis. Finally, by depleting developing B cells during chronic MHV68 infection in vivo, we uncovered a role for developing B cells in the lifelong maintenance of MHV68 latency in circulating mature B cells.
Clonal deletion of self-reactive developing B cells is one of the core mechanisms utilized to enforce B cell tolerance. Transitional B cells are thought to be the primary target of clonal deletion in vivo, with this process resulting in the apoptotic death of a significant portion of the transitional B cell population [15], [24], [27], [30], [51]. Previous work from our laboratory has demonstrated that, following inoculation of wild-type mice, MHV68 maintains latency in a stable frequency of transitional B cells throughout chronic infection [13]. Because these cells have a high rate of turnover and undergo BCR-mediated negative selection, we questioned whether the virus could provide surrogate anti-apoptotic signals to facilitate the survival of infected cells. To test whether the MHV68 vBcl-2 protein played a role in transitional B cell infection, C57BL/6J (B6) mice were inoculated with either wild-type MHV68 or a MHV68 mutant virus deficient in vBcl-2 expression (MHV68.vBcl2stop), and the frequency of latently infection transitional B cells was assessed by PCR. At day 15 post-intranasal (i.n.) inoculation, spleens were harvested and total CD19+AA4+ transitional B cells were isolated using flow cytometry (Fig. 1A). Although the transitional B cell population in wild-type mice can be further subclassified into T1, T2 and T3 B cells based on surface CD21 and CD23 expression [15], we have previously demonstrated that all three populations are infected by MHV68 [13]; thus for experiments here we examined total transitional B cell infection. At 15 days post-inoculation, both the percentage of transitional B cells in the spleen and the absolute number of splenocytes were similar between the two viruses (Table 1). To determine the frequency of cells harboring viral genome, we performed limiting dilution nested PCR analysis, which allows the specific detection of a single copy of viral genome in a background of up to 50,000 uninfected cells [13], [52], [53]. Strikingly, mice inoculated with MHV68.vBcl2stop displayed a 23-fold decrease in the frequency of infected transitional B cells compared to mice inoculated with wild-type MHV68 (MHV68 1 in 590; MHV68.vBcl2stop 1 in 13,500) (Fig. 1B). This phenotype was not confined to early latency, as MHV68 vBcl-2 mutants were also attenuated in transitional B cells during long-term latency (Fig. S1).
Although we postulated that the large reduction in MHV68.vBcl2stop-infected transitional B cells was due to the inability of this mutant virus to block BCR-mediated apoptosis, an alternative hypothesis was that the reduction instead resulted from a low level of virus reactivating from latency – a process that is known to be reduced in MHV68 vBcl-2 mutants [49], [50] and could conceivably be critical for re-seeding the transitional B cell population. To control for this possibility, we performed identical experiments using a MHV68 mutated in the viral cyclin D ortholog (vCycD). Like the vBcl-2 mutant, the vCycD mutant virus (MHV68.vCycD.LacZ) undergoes normal acute replication and latency establishment in whole splenocytes and peritoneal cells, but exhibits a low efficiency of reactivation from latently infected cells [50], [54]. However, in contrast to the vBcl-2 mutant virus, the frequency of transitional B cells infected with MHV68.vCycD.LacZ was similar to that of wild-type MHV68 (MHV68 1 in 590; MHV68.vCycD.LacZ 1 in 620) (Fig. 1B), demonstrating that the decreased frequency of infected transitional B cells in the absence of vBcl-2 is not due to decreased infection secondary to reactivation.
Together these data demonstrated that MHV68 vBcl-2 is critical for latent infection of transitional B cells in vivo, and suggested the possibility that vBcl-2 could promote the survival of transitional B cells that are induced to undergo apoptosis as a result of BCR-mediated negative selection. To further explore this possibility, we determined whether loss of MHV68 vBcl-2 expression similarly reduced infection of other B cells that are subjected to BCR-mediated selection events. During early bone marrow hematopoiesis, pro-B and pre-B cells undergo immunoglobulin gene rearrangement and thus do not express a completed cell surface BCR and are not subjected to BCR-mediated selection. In contrast however, in the final stage of bone marrow hematopoiesis immature B cells, which have completed the process of immunoglobulin gene rearrangement, express complete cell surface BCRs and are required to pass a key central tolerance selection checkpoint in which cells that recognize self-antigen are susceptible to clonal deletion [17]. To determine whether reduced MHV68.vBcl2stop infection was indeed a feature of cells undergoing BCR-mediated clonal deletion, we quantified infection in B cell subsets that do (immature B, transitional B) or do not (pro-B/pre-B, mature B), undergo selection. Fifteen days after inoculation of wild-type B6 mice with MHV68 or MHV68.vBcl2stop virus, bone marrow cells and splenocytes were harvested, and purified populations of B cells were isolated using flow cytometric sorting (Fig. 2A). Pro-B/pre-B cells (CD19+AA4+IgM−) and immature B cells (CD19+AA4+IgM+) were isolated from the bone marrow, and transitional B cells (CD19+AA4+) and mature B cells (CD19+AA4−) were isolated from the spleen. Total bone marrow and splenocyte cell numbers and percentages of each B cells subset were similar for wild-type and vBcl-2 mutant virus infections (Table 1). The frequency of cells harboring viral genome in each sorted population was determined using limited dilution nested PCR analyses (Figs. 2B and 2C). While loss of vBcl-2 expression had no apparent effect on infection of the pro-B/pre-B cell population, infection of immature B cells and transitional B cells was significantly attenuated in mice inoculated with the vBcl-2 mutant virus (immature 4.3-fold reduced, transitional 23-fold reduced). Furthermore, infection of the bulk mature B cell population in the spleen was not significantly altered in the absence of vBcl-2 expression. Notably, the significantly reduced frequencies of infected immature B cells in the bone marrow and transitional B cells in the spleen was not a reflection of decreased total infection, as the vBcl-2 mutant virus displays near wild-type virus frequencies of infection in bulk splenocytes [50] and bulk bone marrow cells (Fig. S2). Thus, these data demonstrate that vBcl-2 plays a key role specifically in B cell populations that are susceptible to BCR-mediated clonal deletion. Further, these results suggest the possibility that vBcl-2 could promote the survival of MHV68-infected developing B cells and thereby allow those cells to bypass key tolerance selection checkpoints.
Based on the preferential requirement of vBcl-2 in B cells susceptible to clonal deletion, we hypothesized that vBcl-2 blocks BCR-mediated induction of the pro-apoptotic pathway. However, it is notable that in addition to its anti-apoptotic functions, vBcl-2 binds with high affinity to Beclin-1 [46], [47] and blocks the induction of autophagy [46], [47], [49]. Thus, it was conceivable that either or both vBcl-2 functions played crucial roles during infection of developing B cells. Previous work using mutagenesis screens defined independent domains within vBcl-2 that are critical for each function – including a BH2 domain required for anti-apoptotic function and an α1 domain critical for anti-autophagic function – and facilitated the generation of specific loss-of-function MHV68 mutants [49]. To define the requirement of each activity in transitional B cells, we inoculated B6 mice with MHV68, MHV68.vBcl2.ΔBH2 (loss of anti-apoptosis function, normal anti-autophagy function) or MHV68.vBcl2.Δα1 (loss of anti-autophagy function, normal anti-apoptosis function) and performed limiting dilution nested PCR assays on sorted transitional B cells (CD19+AA4+) and control mature B cells (CD19+AA4−) at 16 days post-inoculation (Figs. 3A and 3B). In support of our hypothesis, MHV68.vBcl2.ΔBH2 infection of transitional B cells was reduced 21-fold compared to mice infected with wild-type MHV68. Strikingly though, the frequency of transitional B cell infection was similarly reduced (14-fold) in mice infected with MHV68.vBcl2.Δα1. In contrast, the frequencies of mature B cells that harbored viral genome were remarkably similar for all infection groups. Importantly, no significant differences in spleen cell numbers or percentages of B cell populations were observed among groups (Table 1). Thus, these data demonstrate that both the BH2 domain and the α1 domain of vBcl-2 are important for latent infection of transitional B cells, and suggest that actively blocking both apoptosis and autophagy in cells susceptible to clonal deletion is a key facet of MHV68 infection in vivo.
The work described above supported the conclusion that a primary function of vBcl-2 during MHV68 infection is to block the apoptosis of developing B cells. This supposition is consistent with data demonstrating that developing B cells are selectively susceptible to apoptosis due to low levels of host Bcl-2 expression [28], [55]. However, one alternative possibility was that MHV68 vBcl-2 mutant viruses are, either directly or indirectly, comprised in their ability to infect developing B cells. To distinguish these two possibilities, we performed complementation experiments in genetically mutated mice that express host Bcl-2 at a higher level than wild-type mice in developing B cells. New Zealand Black (NZB) mice spontaneously develop a lupus-like syndrome, characterized by the production of pathogenic auto-antibodies, in large part because transitional B cells from these mice are resistant to BCR-mediated apoptosis, allowing autoreactive B cells to breach B cell tolerance checkpoints [55], [56]. The resistance of autoreactive transitional B cells to clonal deletion has been shown to correlate with elevated levels of Bcl-2 expression [55]. To confirm this phenotype, we sorted transitional (CD19+AA4+) and mature (CD19+AA4−) B cells from the spleens of naïve wild-type B6 and NZB mice and western blotted for host Bcl-2 (Fig. 4A). Indeed, transitional B cells isolated from NZB mice expressed Bcl-2 at a substantially higher level than transitional B cells from B6 mice, and at a level equivalent to B6 and NZB mature B cell populations, which are both resistant to BCR-mediated clonal deletion.
To test whether increased expression of host Bcl-2 in transitional B cells complemented the loss of vBcl-2 anti-apoptotic activity, we performed limiting dilution nested viral genome PCR assays on transitional and mature B cells isolated from the spleens of NZB mice 16 days after inoculation with MHV68 or MHV68.vBcl2.ΔBH2 (Fig. 4B). Interestingly, following inoculation of NZB mice, the frequency of transitional B cells carrying MHV68.vBcl2.ΔBH2 genome was not significantly different from that of wild-type MHV68. Similar levels of infection between the two groups were also observed in mature B cells. These results stand in stark contrast to those from wild-type B6 mice (summarized in Fig. 4C): While the frequency of transitional B cells harboring MHV68.vBcl2.ΔBH2 genome was reduced 21-fold in B6 mice (1 in 920 for MHV68, 1 in 19,860 for MHV68.vBcl2.ΔBH2), the frequency of infection was negligibly reduced in NZB mice (1 in 100 for MHV68, 1 in 190 for MHV68.vBcl2.ΔBH2). It is notable that, consistent with previous reports examining MHV68 infection of lupus-prone mice [57], the frequency of viral genome positive cells was higher for both the transitional and mature B cell populations isolated from NZB mice as compared to B6 mice (Fig. 4C). However, preformed infectious virus was undetectable in splenocytes from NZB mice, demonstrating that this result is not a reflection of enhanced lytic replication in these mice (Fig. S3). Collectively, these results demonstrate that the attenuation of an MHV68 vBcl-2 BH2 mutant virus in transitional B cells can be rescued by host Bcl-2 expression, and accordingly, that the vBcl-2 BH2 mutant is competent for transitional B cell infection. Thus, these results strongly support the hypothesis that the MHV68 vBcl-2 specifically promotes the survival of developing B cell populations that are susceptible to clonal deletion.
To more directly test whether vBcl-2 could block BCR-mediated apoptosis of immature B cells, we generated stable immature B cells lines that expressed MHV68 vBcl-2 or host Bcl-2. WEHI-231 is a murine IgM+ B cell line that displays the phenotype of immature B cells and has been widely used for in vitro studies of B cell selection and tolerance mechanisms, including the induction of BCR-mediated apoptosis [58], [59]. To generate WEHI-231 cell lines that ectopically expressed vBcl-2 or host Bcl-2, we engineered murine stem cell viruses (MSCV) that carried genes encoding full-length murine Bcl-2 or full-length MHV68 vBcl-2 fused with a C-terminal HA tag. The MSCV retroviral expression system has been extensively utilized to transduce mammalian cells with target genes of interest [60], [61]. Following generation of recombinant MSCV stocks, viral particles were applied to cultured WEHI-231 cells, and stable cell lines carrying empty vector (EV), host Bcl-2 (Bcl-2), or MHV68 vBcl-2 (M11.1 and M11.2) were generated by antibiotic selection and subcloning. To verify the expression level of transduced genes in the engineered cell lines, we performed western blots on whole cell lysates from each line using antibodies directed toward murine Bcl-2 (Fig. 5A) or HA (Fig. 5B). While the WEHI.Bcl-2 cell line expressed an increased level of Bcl-2 compared to control WEHI-231 cells, Bcl-2 expression in the other cell lines was unchanged, indicating that introduction of other MSCV vectors had no effect on host Bcl-2 expression. Both WEHI.M11.1 and WEHI.M11.2 effectively expressed HA-tagged vBcl-2, although the expression level was slightly increased in the M11.2 line. Similar results were obtained by immunofluorescent microscopy (Fig. 6), demonstrating greatly enhanced expression of host Bcl-2 in the WEHI.Bcl-2 line and ectopic expression of vBcl-2 in both WEHI.M11 lines. Importantly, both Bcl-2 and vBcl-2 localized to mitochondrial compartments, as indicated by co-staining with MitoTracker Red.
To test whether vBcl-2 could block the BCR-mediated apoptosis pathway, we cultured each cell line with or without anti-IgM for 16 hours then assayed the cleavage-based activation of the key pro-apoptosis enzymes caspase-9 (Fig. 7A), caspase-6 (Fig. 7A), and caspase-3 (Fig. 7B). Although faint levels of the active, cleaved forms of all three caspases were detectable in unstimulated WEHI cells, their levels were enhanced nearly 10-fold in cells treated with anti-IgM. Because cleaved caspase-9 is a direct downstream product of Apaf-1 oligomerization and apoptosome formation, these results demonstrate that the pro-apoptotic Apaf-1 pathway was induced in BCR-stimulated WEHI cells. Identical results were obtained in the control WEHI line carrying empty vector (EV). In contrast, activation of all three caspases was completely blocked in WEHI cells over-expressing host Bcl-2. Similarly, all three cleaved caspase products were vastly reduced in both WEHI lines expressing vBcl-2 (M11.1, M11.2). As expected, activated caspases were not induced in control A20 mature B cells that express IgG. These results demonstrate that MHV68 vBcl-2 can block BCR-mediated induction of the pro-apoptotic apoptosome/effector caspase pathway in immature B cells.
To further confirm these results, we used annexin V staining to test the ability of vBcl-2 to block the induction of apoptosis at a cellular level. In viable cells, phosphatidlyserine (PS) localizes to the inner face of the plasma membrane, but in the early stages of apoptosis, PS translocates to the outer face of the membrane. Because annexin V binds to PS, the use of fluorescently-labeled annexin V, in conjunction with a dye to detect membrane permeability, provides a convenient means to detect cells undergoing apoptosis. To determine whether vBcl-2 could block the induction of immature B cell apoptosis following BCR signaling, we cultured WEHI cell lines with or without anti-IgM for 16 hours, and then co-stained with annexin V and the nucleic acid stain SYTOX Blue. Actinomycin D (ActD) treatment of WEHI.EV cells was also included as a positive control for induction of apoptosis. ActD inhibits RNA synthesis and is thus a potent inducer of apoptosis via induction of p53 and disruption of mitochondrial membrane potential [62]–[65]. For all cell samples, the percent of cells in early stage apoptosis was quantified by flow cytometric analysis (Fig. 8A). Cells were considered to be in the early stage of apoptosis if they retained an intact membrane (SYTOX Blue−) but displayed PS on the cell surface (annexin V+). As expected, all nonstimulated WEHI cell lines displayed a low background level of apoptosis (<10%), as indicated by negative staining for annexin V (Fig. 8B). In contrast, following IgM stimulation 40% of WEHI cells and 52% of control WEHI.EV cells were annexin V+ and SYTOX Blue−, indicating that they were in the early apoptotic stage. These results were similar to the percent of WEHI.EV cells in early apoptosis following ActD treatment (67%), demonstrating that BCR stimulation strongly induced cellular apoptosis in both the parental and empty vector control WEHI cell lines. WEHI cells that over-expressed host Bcl-2 were completely protected from BCR-induced apoptosis (8% for nontreated, 6% for IgM-treated). Similarly, WEHI cells that expressed vBcl-2 were almost completely protected from apoptosis (4% and 5% for untreated M11.1 and M11.2, 9% for IgM-treated). Thus, these data directly demonstrate that MHV68 vBcl-2 can block the induction of BCR-mediated apoptosis in immature B cells.
Previous work from our laboratory demonstrated that developing B cells carry latent MVH68 throughout chronic infection, implicating this population as a previously unrecognized reservoir for long-term gammaherpesvirus latency [13]. Work presented here further demonstrates that MHV68 vBcl-2 plays a key role in developing B cell infection and that it can block BCR-mediated apoptosis of immature B cells. Because these cells are short-lived and have a high rate of turnover, these findings cumulatively suggest that MHV68 may actively promote the survival of developing B cells in order to take advantage of the normal homeostatic mechanisms that maintain the mature circulating B cell population. In theory, such a strategy would facilitate the recurrent generation of new latently infected mature B cells and thus serve to maintain lifelong latency in the mature B cell compartment. To determine whether recurrent infection of developing B cells is critical for the maintenance of lifelong MHV68 latency, we undertook experiments to deplete developing B cells at the beginning of, or during, a course of MHV68 infection. Because no cell surface markers are known to be solely expressed on developing B cells, we utilized the in vivo administration of anti-IL-7 antibody as a means to transiently deplete these cells during MHV68 infection. Interleukin-7 (IL-7) is required for B cell development in the mouse, and in the absence of IL-7 developing B cells do not progress past the pro-B cell stage [66], [67]. In contrast, mature B cells do not require IL-7 for their survival. Thus, transient in vivo antibody neutralization of IL-7 results in a marked reduction of developing B cells, including transitional B cells in the spleen, with little or no effect on mature lymphocyte populations [66], [68].
We first examined the contribution of developing B cells to the establishment phase of latency in the mature B cell compartment. Previous reports have shown that 14 day intraperitoneal (i.p.) administration of the murine IgG2b M25 clone of anti-IL-7 neutralizing antibody is sufficient to significantly deplete developing B cells in vivo [66]. Thus, for initial experiments, naïve B6 mice were injected i.p. with 2 mg of anti-IL-7 every other day for 14 days. Control mice were injected with 2 mg isotype control antibody (FLAG-M1 IgG2b) or PBS. To confirm that developing B cells were effectively depleted but that adaptive immune cells remained, splenocytes from control and anti-IL-7-treated mice were analyzed at the end of the depletion period. While the percentage of mature B (CD19+AA4−) and mature T (CD4+ or CD8+) cells were remarkably similar across treatment groups, we observed an 84% depletion (3.8% isotype, 0.6% anti-IL-7) of the transitional B cell population following two weeks of anti-IL-7 treatment (Fig. S4). At this time point, the remaining mice in each group were inoculated i.n. with 104 PFU MHV68. To prevent renewed generation of developing B cells following infection, every other day anti-IL-7 treatments were continued for the final 15 days of the experiment. By 15 days post-inoculation, lytic replication is no longer detectable and latency establishment is at its peak (Fig. 9A).
At experiment termination, splenocytes were harvested and pooled from 3 mice per treatment group, and flow cytometric sorting was performed to isolate naïve (CD19+AA4−IgM+), germinal center (CD19+AA4−IgM−CD38lo), and memory (CD19+AA4−IgM−CD38hi) B cells for MHV68 latency analyses. Pooled cell suspensions were simultaneously analyzed to confirm developing B cell depletion. At 15 days post-virus inoculation (a total of 29 days of antibody administration), greater than 91% of splenic transitional B cells were depleted (Table S1), while mature T cell populations (CD4+ T cells, CD8+ T cells) and mature B cell populations (naïve B cells, germinal center B cells, memory B cells) remained normal (Tables S1 and S2). Subsequently, naïve, germinal center, and memory B cell populations were sorted (Fig. S4) and analyzed for the presence of viral genome by limiting dilution nested PCR (Fig. 10A). Despite the nearly complete depletion of developing B cells during early infection, the frequencies of naïve, germinal center and memory B cells harboring viral genome in the anti-IL-7 treatment group were nearly identical to that of mock and isotype control groups (Fig. 10B). Because we did not observe 100% depletion of developing B cells, we cannot preclude with certainty the possibility that a low level of developing B cells is sufficient to impact mature B cell latency. However, the finding that early mature B cell latency was completely unaltered in the absence of greater than 95% of the developing B cell population strongly suggests that developing B cells do not play a significant role in the establishment phase of MHV68 latency. Furthermore, these experiments demonstrate that transient IL-7 depletion does not significantly impact T cell control of MHV68 latency, since it is well-established that loss of T cell effector function results in increased numbers of latently infected cells [52], [69]–[73].
Previous work from our laboratory and others has demonstrated that the establishment phase of latency is fundamentally different from the maintenance phase of latency with regard to infected cell composition and, presumably, the molecular profile of viral gene expression [6]–[8]. Notably though, MHV68 infection is maintained in a stable frequency of immature and transitional B cells over time [13], suggesting that infection of these cells may play a key role in facilitating lifelong latency. To determine whether developing B cells contribute to the maintenance phase of latency in mature B cells, we performed IL-7 depletion experiments after the establishment of chronic infection. For these experiments, B6 mice were infected i.n. with 104 PFU of MHV68, then housed untouched for 28 days, allowing time for the virus to set up stable latency. Beginning on day 28, anti-IL-7 was administered every other day for 30 days (Fig. 9A). At 58 days post-inoculation, splenocytes were harvested and stained, and flow cytometric analysis and sorting was performed. Following this 30 day depletion regimen, the percentage of transitional B cells was reduced greater than 87% (4.0% mock and isotype, 0.5% anti-IL7), but importantly, the percentages and absolute numbers of mature B cells and T cells were unaffected (Fig. 9B and Tables S1, S2). To determine whether developing B cell depletion altered the maintenance of mature B cell latency, we performed limiting dilution nested viral genome PCR on sorted naïve, germinal center and memory B cells from each treatment group (Fig. 10A). As expected, in the control groups the overall frequency of infection in each population dropped significantly from 15 days to 58 days (Fig. 10B), owing to the contraction of the early expansion phase of latency and the establishment of stable long-term infection [7]. Interestingly, short-term depletion of the developing B cell population during the stable maintenance of latency led to a statistically significant decline in the overall frequencies of infection in the naïve and germinal center B cell populations (2.8-fold for naïve, 4.2-fold for germinal center) and a more subtle decline of infection in the memory B cell population (2.0-fold). These data stand in clear contrast to depletions during the establishment phase of latency, suggesting that the maintenance of long-term latency in mature B cells requires a fundamentally different virological process than early stage infection. These results for the first time provide a clear link between infection of developing B cell populations and latency in the mature B cell compartment.
The means by which gammaherpesviruses such as EBV maintain lifelong infection at a stable setpoint in circulating memory B cells is poorly understood. Although the vast majority of human gammaherpesvirus research effort has focused on infection of circulating mature B cells, several clinical studies have reported the presence of EBV and KSHV in bone marrow [74]–[78], suggesting the possibility that gammaherpesviruses can infect B cells that are still in the developmental stage. While these reports have mostly attributed the presence of virus in bone marrow to underlying disease states [77]–[84], one alternate possibility is that gammaherpesviruses utilize infection of bone marrow hematopoietic progenitor cells as one facet of their natural life cycle. A key observation in support of this possibility is our previous finding that stable fractions of short-lived immature and transitional B cells carry MHV68 throughout chronic infection [13]. Based on this finding, we have hypothesized that recurrent infection of developing B cells, and their subsequent differentiation along normal maturation pathways contributes to the maintenance of latency in a stable fraction of memory B cells (Fig. 11).
While such a linear B cell infection and differentiation model is an intriguing possibility, this strategy would seem inefficient unless the virus could provide critical signals to promote the survival of those cells that would otherwise undergo clonal deletion. It is perhaps expedient then that all gammaherpesviruses encode anti-apoptotic proteins, including the Bcl-2 ortholog expressed by MHV68. To test whether the MHV68 vBcl-2 plays a key role in developing B cells, we analyzed in vivo infection of developing B cell subsets using MHV68 vBcl-2 mutants and determined whether vBcl-2 alone could block BCR-mediated apoptosis of immature B cells. In work described here, we demonstrate that efficient MHV68 infection of immature and transitional B cells requires vBcl-2 expression. Interestingly though, this gene was not required for infection of the pro-B/pre-B or mature B cell populations, which are not subjected to clonal deletion checkpoints, strongly suggesting that a key role of vBcl-2 in vivo is to promote the survival of B cell populations that are susceptible to BCR-mediated selection. Importantly, increased host Bcl-2 expression in transitional B cells complemented the lack of vBcl-2 expression during MHV68 infection, demonstrating that (a) MHV68 vBcl-2 mutants are competent for developing B cell infection, and (b) vBcl-2 carries out functions that are normally associated with host Bcl-2 activity. Consistent with this conclusion, BCR-mediated clonal deletion of developing B cells has been attributed to low Bcl-2 expression [27], [28], [31], [55], [85], and we demonstrate here that expression of vBcl-2 in WEHI-231 immature B cells blocked BCR-induced cleavage of caspases -9, -6 and -3, and induction of cellular apoptosis. Finally, transient in vivo depletion of developing B cells during the maintenance phase of long-term latency resulted in reduced numbers of mature B cells carrying viral genome, strongly suggesting that ongoing infection of developing B cells is intimately linked to maintaining latent infection in a stable fraction of circulating mature B cells. Collectively, these studies provide evidence that MHV68 utilizes a Bcl-2 ortholog to promote infection and survival of developing B cells, and point to a key role for developing B cells in the sustenance of lifelong latency (Fig. 11).
Host proteins of the Bcl-2 family play a major role in regulating B cell survival, especially during clonal deletion. The pro-apoptotic Bcl-2 family proteins Bak, Bax, and Bim have been shown to mediate the apoptotic death of self-reactive B cells following BCR stimulation with antigen [8], [13], [86]–[88]. Conversely, overexpression of anti-apoptotic proteins Bcl-2 or Bcl-XL prevents BCR-mediated cell death [25], [32]–[34], [55]. Thus, our observations that the MHV68 vBcl-2 mutants displayed a significant defect in immature and transitional B cell infection are consistent with the conclusion that vBcl-2 plays an anti-apoptotic role similar to host Bcl-2 in these cell populations. This conclusion is further supported by our complementation experiments in NZB mice, which showed that increased expression of host Bcl-2 in developing B cells completely negated the attenuated phenotype of the vBcl-2 BH2 mutant virus in these cells. Finally, our experiments using WEHI-231 B cells directly demonstrated for the first time that vBcl-2 could block the apoptosis of immature B cells that were stimulated through the BCR. Thus, we speculate that the decline in frequency of genome positive immature and transitional B cells that occurs in the absence of vBcl-2 expression is due to a decrease in the number of infected cells surviving clonal deletion.
It is noteworthy that our findings do not rule out a role for the anti-autophagy function of vBcl-2 in developing B cells. Indeed, our experiments with the vBcl-2 α1 mutant strongly suggest that MHV68 vBcl-2 blockade of autophagy also plays a critical role in infection of developing B cells. This is consistent with previous reports demonstrating that autophagy of WEHI-231 cells and primary splenic B cells is induced by BCR stimulation [89], and that the autophagy pathway may serve as a backup mechanism for cell death when apoptosis is blocked [90]. A natural extension of this conclusion is that autophagy plays a key and underappreciated role in B cell selection. Thus blocking both apoptosis and autophagy may be requisite for survival of cells that are otherwise destined for clonal deletion.
Our previous demonstration that immature and transitional B cells carry viral genome throughout chronic infection [13] strongly supported a key role for developing B cells during lifelong latency, but did not address whether a linkage exists between developing B cell infection and the peripheral mature B cell latency reservoir. Does infection of developing B cells directly contribute to the dynamic maintenance of latency in circulating mature B cells? Or instead is it an autonomous event, unrelated to peripheral B cell infection? In work presented here, we gained insight into this question by depleting developing B cells in vivo then assessing the extent of latent infection in mature B cell subpopulations. Interestingly, developing B cell depletion from 28 to 58 dpi resulted in a pronounced and statistically significant decrease in naïve and germinal center B cell infection, demonstrating that developing B cells are required for the maintenance of peripheral mature B cell latency. Depleted mice also demonstrated a highly reproducible, but smaller decrease in the frequency of infected memory B cells, likely owing to the relative longevity of memory B cells [91]–[94]. Importantly, these results provide the first clear demonstration of a potential direct link between the infection of developing B cells and stability of the major latency reservoir of circulating mature B cells, and strongly suggest that the maintenance of lifelong latency is a dynamic process that involves constant reseeding of the mature B cell reservoir.
Interestingly, depletion of developing B cells prior to and during the first 16 days of infection had no effect on latency in any of the mature B cell subpopulations. These results demonstrate that developing B cells do not contribute to the early establishment of latency in the mature B cell compartment. This finding is notable because it strongly implies that different virological mechanisms operate in vivo during establishment versus maintenance of latency. Consistent with this conclusion, the frequency of latent MHV68 infection peaks at 16–20 dpi during the “expansion phase” of latency, then gradually decreases until it reaches a stable level at 42–49 dpi [52], [53], [95], [96]. Similarly, the percentage of latently infected cells that reactivate ex vivo is highest at 16 dpi and decreases over time [52], [53]. In light of our results, it is reasonable to speculate that this transition from an active form of latency to a more quiescent form of latency may reflect a conversion from a majority of mature B cells that were directly infected by free virus during acute replication to mature B cells that arose from differentiation of infected developing B cells. Notably, Thorley-Lawson and colleagues have proposed an analogous concept for EBV, wherein direct infection of memory B cells during acute EBV results in an active latency growth program and subsequent cytotoxic T cell targeting, whereas virus-driven maturation of naïve B cells sets up a quiescent latency program that facilitates life-long infection [4]. While further experiments will be required to unravel these complexities, our results clearly indicate that the stable maintenance of long-term latency is a dynamic process that is distinct from early latency and requires an ongoing contribution from developing B cells.
To date, it is unclear whether human gammaherpesviruses infect developing B cell populations as part of their natural life cycle in healthy individuals. However, both EBV and KSHV genomes have been detected in the bone marrow and progenitor cells of humans in the context of disease. For example, EBV has been detected in the bone marrow of patients with EBV-associated hemophagocytic lymphohistiocytosis (EBV-HLH) [77], [78] and both EBV and KSHV have been detected in the bone marrow of AIDS patients [74], [75]. Further, EBV-associated lymphoproliferative disease following allogeneic bone marrow hematopoietic stem cell (HSC) transplantation is almost always of donor origin [82], [97], [98]. Consistent with the possibility of a progenitor cell source of gammaherpesvirus infection, KSHV has been detected in circulating human CD34+ hematopoietic progenitor cells (HPCs) of KS patients [99] and in morphologically immature cells in the bone marrow of transplant recipients [76]. Additionally, several reports have demonstrated the presence of EBV+ B cells, presumed to be of progenitor cell origin, arising from long-term human bone marrow cultures of healthy donors [100], [101] and hematologic patients [102]. Cumulatively, these reports provide substantial support for the concept that the human gammaherpesviruses can infect developing B cells in the bone marrow or in circulation. Nevertheless, a great deal of additional work will be required to comprehensively define the role of precursor B cells during a normal course of EBV or KSHV infection.
It is also noteworthy that several recent reports have correlated high numbers of circulating transitional B cells with high EBV loads in patients at risk for the development of EBV-associated B cell lymphomas. For example, it is now widely recognized that EBV and malaria co-infection correlate with a high incidence of endemic Burkitt's B cell lymphoma in African children [103], [104]. Although the synergistic interplay of these two pathogens during oncogenesis is poorly understood, a recent report demonstrated that infants from a malaria-endemic region of Kenya display normal levels of naïve (IgD+CD27−) and classical memory (IgD−CD27+) B cells, reduced numbers of non-class switched memory (IgD+CD27+) B cells, but expanded numbers of immature transitional (CD10+CD34−) B cells [105]. Interestingly, this population of children also exhibits increased EBV loads in accordance with earlier ages of infection [106]. Likewise, patients with chronic HIV infections frequently display increased EBV loads [107] and are at high risk for the development of EBV-associated B cell lymphoma, and a recent study linked high EBV loads in chronic HIV patients with an increased frequency of circulating immature or transitional B cells [108]. At minimum, these studies provide correlative evidence of a link between high numbers of developing B cells and enhanced EBV infection, and may suggest that transitional B cells serve as a conventional EBV reservoir that greatly expands during particular types of immune dysfunction.
Our finding that vBcl-2 promotes the survival of immature and transitional B cells during MHV68 infection may provide an important clue to the long-speculated potential link between gammaherpesvirus infections and autoimmune disease. For example, numerous groups have published reports providing circumstantial evidence of a causal relationship between EBV infection and the development of, among others, multiple sclerosis and systemic lupus erythematosus (reviewed in [109], [110]). Nevertheless, this relationship has been seriously questioned due to the incongruous ubiquity of EBV with the relatively rarity of autoimmune diseases. However, on a teleological basis it is reasonable to speculate that if indeed gammaherpesviruses promote survival and maturation of B cells with autoreactive BCRs, then these viruses would also have a means to prevent autoreactive BCRs from signaling as a means to simultaneously protect the host and facilitate long-term latency. Thus, as with gammaherpesvirus-associated tumors, the development of autoimmune disease may represent an anomalous consequence of gammaherpesvirus infection, likely resulting from synergism with disease-promoting secondary factors such as host genetics or pathogen co-infection. In support of the multifactorial nature of any potential link between gammaherpesviruses and autoimmune disease, several conflicting reports have indicated that MHV68 both suppresses [57], [111], [112] and exacerbates [113], [114] murine autoimmune diseases. In light of our demonstration that MHV68 (a) blocks BCR-mediated apoptosis of immature B cells and (b) promotes the survival of developing B cell subpopulations that are known to undergo autoreactive BCR-mediated clonal deletion, a potential link between gammaherpesvirus infection and the survival of B cells with autoreactive BCRs warrants further exploration.
The work presented here represents a substantial step forward in the understanding of the in vivo role of a gammaherpesvirus-encoded Bcl-2 ortholog. The finding that viruses deficient in vBcl-2 function were most significantly attenuated in those developing B cell populations that are required to surmount tolerance selection checkpoints strongly suggests that the virus alters normal B cell development outcomes as a means to promote long-term survival. Consistent with this conclusion, in vivo depletion of developing B cells during long-term latency resulted in reduced infection in mature B cells, supporting the possibility of a direct link between precursor B cell infection and the stability of lifelong latency in circulating mature B cells. A great deal of further work will be required to determine whether a normal course of gammaherpesvirus infection promotes the simultaneous survival and inactivation of autoreactive B cells, and whether in rare scenarios co-factors can play the role of key intermediary between gammaherpesvirus infection of developing B cells and development of autoimmune disease.
Wild-type C57Bl/6J or NZB mice age 7–10 weeks purchased from Jackson Laboratory (Bar Harbor, Main) were used for experiments presented here. Mice were housed at University of Florida in accordance with all federal and university guidelines. Mice were anesthetized with isofluorane and infected intranasally (i.n.) with 104 PFU of virus in 30 µl serum-free DMEM. MHV68 strain WUMS (ATCC VR1465), MHV68.vBcl2stop [50], MHV68.vBcl2ΔBH [49], and MHV68.vBcl2Δα1 [49] were used for inoculations. At indicated time points mice were sacrificed by exposure to inhalation anesthetic.
All animal experiments were performed in strict accordance with Federal and University guidelines. Specifically, we adhered to the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and the American Veterinary Medical Association Guidelines on Euthanasia. The animal protocol was approved by the Institutional Animal Care and Use Committee at the University of Florida (study number 201105767).
To obtain splenocyte and bone marrow cell suspensions, spleens, femurs, and tibias were harvested and single cell suspensions were prepared by homogenizing spleens and flushing each bone with 5 mls of DMEM. For spleen samples red blood cell lysis (144 mM NH4Cl and 17 mM Tris, pH 7.2) was carried out for 5 minutes at 37°C prior to staining. Harvested cells were then suspended in blocking buffer (PBS, 5% bovine serum albumin, 10% normal rat serum, and purified anti-mouse CD16/CD32 [Fc block; clone 2.4G2; BD Biosciences]) for 30 minutes on ice prior to staining with antibodies.
Limiting-dilution assays to determine the frequency of transitional and mature B cells reactivating from latency or containing performed infectious virus were performed as previously described [13], [52], [88]. Briefly, transitional and mature B cell populations were serially diluted two-fold and plated onto MEF monolayers in 96-well plates. Twelve dilutions were plated per sample, and 24 wells were plated per dilution. Wells were microscopically scored for cytopathic effect (CPE) following a three week incubation period. To detect preformed infectious virus, parallel samples of mechanically disrupted cells were plated onto MEF monolayers. This process kills >99% of live cells, while leaving preformed infection virus intact. Data points are the average of two or three independent experiments, each consisting of pooled spleens from three mice, and are presented as the percentage of wells per dilution that scored positive for viral CPE+/− standard error.
LDPCR was used to determine the frequencies of cells positive for MHV68 genome, as previously described [13], [52], [115]. Briefly, cell samples were serially diluted threefold in a background of uninfected RAW 264.7 murine macrophages. A total of 1×104 or 5×104 cells were plated in a 96-well PCR plate at 12 wells per dilution. 10, 1, or 0.1 copies of a MHV68 ORF72 plasmid in a background of RAW 264.7 cells, and RAW 264.7 cells only were included on all plates for controls. Cells were lysed with proteinase K at 56°C for 8 hours. Two rounds of nested PCR were then performed using primers specific for MHV68 ORF72, and 195 bp bands resolved using a 3% agarose gel. Unless otherwise indicated, data points are the average of three to five independent experiments, each consisting of pooled spleens from three to five wild-type C57Bl/6 mice. Data are presented as the mean percentage of wells per dilution that were positive for viral genome +/− standard error. On graphs, the dashed line at 63.2% indicates the point at which one viral genome-positive cell per reaction is predicted to occur. The x axis shows the numbers of cells per reaction; the y axis shows the percentages of 12 reactions positive for viral genome.
B cell lines WEHI-231, WEHI empty vector, WEHI Bcl-2, WEHI M11.1, WEHI M11.2, and A20 were maintained in RPMI-1640 containing 10% fetal bovine serum (FBS), 100 U/mL penicillin, 100 mg/mL streptomycin, and 0.05 mM 2-mercaptoethanol with 10% CO2, at 37°C. 0.025 µg/mL puromycin was added to culture media of WEHI-231 cell lines carrying Murine Stem Cell Virus (MSCV) vectors in order to maintain selection. The MSCV retro viral vector packaging cell line BOSC23 was cultured in complete Dulbecco's modified Eagle's medium (DMEM) supplanted with 10% FCS, 100 U/mL penicillin, and 100 mg/mL streptomycin and incubated at 5% CO2, 37°C.
vBcl-2 and host Bcl-2 were cloned into murine stem cell virus (MSCV) retroviral vectors (Clontech). To detect M11 expression, a hemagglutinin (HA) epitope tag was added to the C-terminus of M11. The MSCV Retroviral Expression System (Clontech) was used to generate WEHI empty vector (WEHI.EV), WEHI M11 (WEHI.M11.1 and WEHI.M11.2), and WEHI Bcl-2 (WEHI.Bcl2) cell lines according to the manufacturer's protocols. Briefly, retroviral stocks were generated by transfecting BOSC23 packaging cells with 8 µg DNA (pMSCV M11, pMSCV bcl-2, or pMSCV empty vector) using Lipofectamine2000 (Invitrogen). Viral supernatants were collected at 48 hours and 72 hours post-transfection and filtered through a 0.45 µm nylon filter. The viral supernatants were treated with 6 µg/mL polybrene and were then added to WEHI-231 cells seeded in 60 mm dishes at 1×106 cells per dish. Twenty four hours post-infection 0.025 µg/mL puromycin was added to cultures and selection was carried out continuously. Western Blots and immunofluorescence microscopy, as described below, were used to confirm expression of HA (M11) and Bcl-2.
For immunofluorescence assays, 1×105 WEHI cells were washed with PBS and incubated with MitoTacker Red for 30 min at 37°C. Following incubation, cells were washed twice with PBS. Cells were then fixed with ice cold MeOH for 15 min at −20°C, then pelleted and resuspended in 300 µl PBS. The cell suspension was then fixed to microscope slides via cytocentrifugation using a Cytopro cytospin at 400 RPM for 5 min at room temperature. Next, cells were blocked with 5% normal goat serum (NGS) for 1 hour at room temperature to prevent non-specific antibody binding, then incubated with the primary antibodies rabbit anti-BCL-2, rabbit anti-HA or the IgG isotype control in 2% NGS in PBS overnight. Cells were washed three times with PBS and then incubated with secondary antibody Alexa Fluor 488 goat anti-rabbit IgG (Molecular Probes, Inc.) for 1 hour at room temperature, then washed and then treated with DAPI. Images were captured with Leica software using a Leica laser confocal microscope (TCS SP2 ABOS laser scanning spectral confocal) with 63× objective at 4 times zoom.
Cell lysates were prepared using a 1∶1 ratio of PBS and Laemmli buffer with 2-mercaptoethanol and heating for 5 minutes in a boiling water bath. Total protein per µL of sample was quantitated using Thermo Scientific NanoDrop 2000. Equal amounts of total cellular protein from each sample were separated using SDS-PAGE on 15% acrylamide gels, followed by immunoblotting. Membranes were blocked in a 10% milk–TBS–Tween 20 solution, followed by incubation with primary antibody (anti-HA-tag [Cell Signaling Technology], anti-Bcl-2 [Cell Signaling Technology], rabbit anti-caspase-3, -6, or -9 [Cell Signaling Technology], or anti-actin clone c4 [Millipore]). Blots were then incubated with goat anti-mouse IgG HRP (Millipore) or goat anti-rabbit IgG HRP (Abcam). Antibody-labeled protein bands were detected using Western Lightning Plus-ECL Enhanced Chemiluminescence Substrate (Perkin Elmer). Densitometry was performed using Bio-Rad Gel Doc XR system using Quantity One 4.6.9 software (Bio-Rad).
For flow cytometric annexin V assays, WEHI cell lines and A20 cells were plated at 2.5×105 cells/mL with 2 mL total volume in 6 well plates. Cells were incubated for 16 hours in the absence or presence of 20 µg/ml goat anti-mouse IgM, μ chain specific (Jackson ImmunoResearch Laboratories) or 40 nM Actinomycin D (Sigma). Following a 16 hour incubation, cells were washed in cold PBS and resuspended in 100 µl binding buffer, then incubated on ice for 15 minutes with 10 µl of Annexin V-FITC (BD Biosciences). Following Annexin V staining, cells were washed and resuspended in 400 µl binding buffer and 0.5 µl SYTOX Blue (Life Technologies). Cells were incubated an additional 5 minutes at room temperature before analysis on a LSR II (Becton Dickinson) flow cytometer. Data were analyzed using FACS Diva software (Becton Dickinson).
FACS data were analyzed using FACSDiva (BD Biosciences) and FloJo. All other data were analyzed using GraphPad Prism software (GraphPad Software, San Diego, CA). The frequencies of cells positive for viral genome, reactivating ex vivo, and containing preformed virus were determined from the nonlinear regression analysis of sigmoidal dose response best-fit curve data. Based on Poisson distributions, the frequency at which at least one event in a given population is present occurs at the point where the regression analysis line intersects 63.2%. Calculation of statistical significance was determined by Student's t test of paired cell dilution results.
|
10.1371/journal.pntd.0000608 | Field Validation of a Transcriptional Assay for the Prediction of Age of Uncaged Aedes aegypti Mosquitoes in Northern Australia | New strategies to eliminate dengue have been proposed that specifically target older Aedes aegypti mosquitoes, the proportion of the vector population that is potentially capable of transmitting dengue viruses. Evaluation of these strategies will require accurate and high-throughput methods of predicting mosquito age. We previously developed an age prediction assay for individual Ae. aegypti females based on the transcriptional profiles of a selection of age responsive genes. Here we conducted field testing of the method on Ae. aegypti that were entirely uncaged and free to engage in natural behavior.
We produced “free-range” test specimens by releasing 8007 adult Ae. aegypti inside and around an isolated homestead in north Queensland, Australia, and recapturing females at two day intervals. We applied a TaqMan probe-based assay design that enabled high-throughput quantitative RT-PCR of four transcripts from three age-responsive genes and a reference gene. An age prediction model was calibrated on mosquitoes maintained in small sentinel cages, in which 68.8% of the variance in gene transcription measures was explained by age. The model was then used to predict the ages of the free-range females. The relationship between the predicted and actual ages achieved an R2 value of 0.62 for predictions of females up to 29 days old. Transcriptional profiles and age predictions were not affected by physiological variation associated with the blood feeding/egg development cycle and we show that the age grading method could be applied to differentiate between two populations of mosquitoes having a two-fold difference in mean life expectancy.
The transcriptional profiles of age responsive genes facilitated age estimates of near-wild Ae. aegypti females. Our age prediction assay for Ae. aegypti provides a useful tool for the evaluation of mosquito control interventions against dengue where mosquito survivorship or lifespan reduction are crucial to their success. The approximate cost of the method was US$7.50 per mosquito and 60 mosquitoes could be processed in 3 days. The assay is based on conserved genes and modified versions are likely to support similar investigations of several important mosquito and other disease vectors.
| Once infected with dengue virus, a female Aedes aegypti mosquito must survive longer than twelve days before it can transmit the virus to an uninfected person. New dengue control strategies therefore aim to circumvent dengue transmission using entomopathogenic microorganisms that shorten mosquito lifespan. Accurate methods to determine the age of individual mosquitoes are required for these and other mosquito control interventions. We have previously shown that mosquito age can be predicted from the transcription of specific genes. Here we demonstrate that this can be achieved under natural conditions when mosquitoes are uncaged and free to engage in natural behavior. To do this, we produced “free-range” female mosquitoes by releasing 8007 mosquitoes at an isolated location and recapturing the females of known ages. We developed an age prediction model from gene transcription measures of mosquitoes maintained in small “sentinel cages” maintained onsite. We then used this model to predict the ages of the free-range mosquitoes, based on their own transcription measures. Age predictions were robust to physiological changes associated with blood feeding and egg development. We show that the technique could be applied to identify a 50% reduction in mosquito population survival that is expected from infection with entomopathogenic Wolbachia bacteria.
| The survival of mosquitoes to a relatively old age is required for the transmission of mosquito-borne diseases. Mosquito-borne pathogens such as dengue viruses and malaria parasites require a period of development or multiplication inside the mosquito (extrinsic incubation period; EIP) before transmission can occur. Female mosquitoes ingest the pathogen when taking a blood meal from an infected host. The pathogen must then penetrate the midgut, escape from the midgut, multiply and disseminate through the mosquito before infecting the salivary glands. Transmission may then occur when the female subsequently bites a naïve host. For many of the world's most important mosquito-borne diseases (malaria, dengue and lymphatic filariasis), the EIP of the parasite or virus is long relative to the lifespan of the mosquito vector. The EIP of the dengue viruses in the primary mosquito vector, Aedes aegypti, is approximately 12–16 d [1]. Landmark epidemiological studies identified mosquito survival as a target for the prevention of mosquito borne disease and a sensitive indicator of disease activity [2]–[4]. However, the importance of mosquito longevity has seldom been directly tested because few tools exist that can accurately determine the age of wild caught mosquitoes.
Mosquito control strategies adopting various microbial agents aim to reduce mosquito longevity to impact disease transmission [5]–[9]. Successful infection of Ae. aegypti with a life-shortening strain of the intracellular bacteria Wolbachia has recently been reported [9]. The infection causes a 50% reduction in Ae. aegypti longevity, is maternally inherited and has the capacity to be driven through wild mosquito populations through the mechanism of cytoplasmic incompatibility. Implementation of these strategies will require rapid and high throughput age determination of the targeted mosquito vectors to evaluate the efficacy of control. Traditional dissection based methods of age grading mosquitoes fall well short of the required accuracy and throughput required, and biochemical approaches such as the measurement of cuticular hydrocarbons lose accuracy beyond 15 d old [10]–[13].
We previously reported a method of predicting the age of adult female Ae. aegypti mosquitoes using the transcriptional profiles of age responsive genes [14],[15]. Quantitative Reverse Transcriptase PCR (qRT-PCR) was used to measure the transcriptional profiles of these genes from the head and thorax of individual mosquitoes. Age predictions were then derived from the transcriptional profiles using multivariate calibration. An initial validation of the method was performed on mosquitoes that had been maintained inside 13 m3 capacity dome tents positioned inside and adjoining an elevated residence in Cairns, northern Australia. An age prediction model incorporating eight genes was reported that facilitated age predictions of female Ae. aegypti to within ±5 d of their actual age up to 19 d of age. For large scale studies, a reduced model incorporating three genes was recommended. 74.99% of the variation in the transcription of these genes was explained by mosquito age. However, the activity of mosquitoes in cages may have been reduced, affecting their capacity to seek out preferred micro-habitats with unknown consequences for the aging assay. Similarly, the influence of major physiological changes during the mosquito blood feeding - egg development cycle is unknown. Physiological processes associated with the mosquito gonotrophic cycle involve extensive changes in gene expression [16]–[19].
Here we report the evaluation of transcriptional age grading on Ae. aegypti that were entirely uncaged as adult mosquitoes (called free-range mosquitoes). By conducting the experiment at an isolated homestead without resident Ae. aegypti, a large cohort of mosquitoes could be released unmarked and recaptured at known ages to 29 d old. To facilitate high throughput age grading, a multiplexed assay incorporating Taqman probes [20] was used for quantitative PCR analysis. Several physiological parameters (including body size, blood digestion and ovary development) were investigated as possible sources of variance in our age predictions. We have demonstrated that mosquito age estimates generated from transcriptional profiles under natural conditions can be applied to identify changes in mean mosquito population life expectancy. By incorporating Taqman probes, the method could be scaled-up to facilitate the expected increases in requirements for mosquito age grading when new mosquito control strategies are implemented.
Human ethics approval for allowing colonized (dengue-free) mosquitoes to feed on the investigators was obtained from James Cook University (Human ethics approval H2250). Blood feeding was considered to cause a medium risk of allergic reaction and provision was in place that individuals were excluded if they reacted strongly to bites. Written consent was obtained acknowledging the right to refuse or withdraw.
Aedes aegypti were collected as eggs from ovitraps set at Machans Beach (16° 51′ 14″ S, 145° 44′ 55″ E), a suburb of Cairns, Queensland, Australia. G1 eggs were hatched in hay-infused water and reared on a diet of dry adult cat food (Friskies; North Ryde, Australia) under ambient conditions and low densities to ensure synchronous development. Pupae were transported to a homestead in an isolated rainforest located 13 km from Cairns. The homestead consisted of a cluster of three buildings; a single story, one bedroom, unscreened wooden house and two open-sided shelters. Ae. aegypti were known to be absent from the site and this was confirmed by attempts to trap Ae. aegypti using three adult mosquito traps (BG-Sentinel traps; Biogents, Regensburg, Germany) at the site over a four week period before release. The pupae were randomly divided into a “free-range” group for release and a “sentinel-cage” group. A third group was used to determine the sex ratio (n = 83). The free-range pupae were transferred to open 9 L release containers that were maintained at the site for 24 hr. Over this time the adults emerged and dispersed. The number of adult Ae. aegypti released was estimated by counting the pupal exuviae in the release containers at the end of the release period. Mosquitoes were allowed free movement around the property and volunteers residing at the site provided blood meals at 2 d intervals. The property was supplemented with additional larval habitats (tyres, buckets and pot-plant bases) and all larval habitats/potential oviposition sites were flushed-out and cleaned every 5 to 6 d to prevent emergence of any adult mosquitoes. Fifteen resting adult females were recaptured at 2 d intervals from 1 to 29 d from various sites around the field house using mechanical aspirators. Temperature and humidity was recorded throughout the experiment using Hobo data loggers (Onset Computer Corporation, Pocasset, U.S.A). The sentinel cage group was placed into two cages (450×450×450 mm) that were maintained on-site for the duration of the experiment and 10 mosquitoes were sampled from these at 4 d intervals from 1 to 29 d. At the time of capture, mosquitoes were briefly anaesthetized at −20°C for 5 min. The heads and thoraces of individual mosquitoes were dissected from abdomens, wings and legs on a glass slide using fine tweezers and a scalpel. Heads and thoraces were rapidly placed in 300 µl of RNAlater (Ambion, Austin, U.S.A) and stored as per the manufacturer's protocol. Abdomens of recaptured mosquitoes were stored at −20°C until dissections were performed to determine physiological status. At the conclusion of sampling, mosquitoes were exhaustively removed by aspirator collections and using four BG-Sentinel traps and 18 sticky-ovitraps positioned around the homestead over a three week period. No Ae. aegypti were collected after the second week or have been observed since, indicating that the mosquitoes did not become established.
Various physiological characteristics of the free-range females were determined by dissection. Mating status was determined from the presence or absence of sperm in the spermathecae. The midgut was observed and females were classified as to whether blood was present or absent. Ovarian development was graded according to Christophers' stages [21] and parity by the presence or absence of ovary tracheolar skeins [22]. Wing length was used as a proxy for body size and was measured as the distance from the axial notch to the wing tip, excluding the fringe scales [23].
Samples were removed from RNAlater and transferred to 1.5 ml screw-capped plastic vials with a single 3 mm silica glass bead and 0.5 ml Trizol reagent (Invitrogen, Carlsbad, U.S.A.). Samples were mechanically homogenized in a Minibeadbeater (Biospec) for 1.5 min, transferred to 1.5 ml microfuge tubes and centrifuged at 17,000×g for 10 min at 4°C to pellet the chitinous material. The supernatant was transferred to a new 1.5 ml microfuge tube. Trizol extraction was performed according to manufacturer's instructions; however, isopropanol precipitation was performed overnight at −30°C. Total RNA pellets were reconstituted in 20 µl RNAse-free water. Total RNA was quantified by absorbance readings using a Nanodrop spectrophotometer (Biolab, Scoresby, Australia). Total RNA was standardized at 500 ng and treated with 0.2 U recombinant RNAse-free DNase (Roche) as per the manufacturer's protocol.
A previous report [14] proposed a reduced set of three gene expression (GE) measures (Aedes aegypti calcium binding protein [Ae-15848; XM_001653412], Aedes aegypti pupal cuticle protein 78E [Ae-8505; XM_001656550] and Aedes aegypti cell division cycle 20 [cdc20; fizzy][Ae-4274; XM_001664201]) all normalized to a housekeeping gene Aedes aegypti 40S ribosomal protein S17 [Ae-RpS17; AY927787]). Primer and dual-labelled Taqman probe sets (Table S1) were designed for this gene set, using web-based assay design software RealTimeDesign (http://www.biosearchtech.com/products/probe_design.asp; Biosearch Technologies Novato, CA), to allow multiplex qRT-PCR assays to be developed. Gene-specific Taqman probes were labeled with different fluorophores that had minimal spectral overlap to minimize “cross-talk” between color channels. Four different fluorophores were used to label Taqman probes specific to each gene. Gene-specific labeling was incorporated into the Taqman assay to allow for the possibility of a triplex, excluding the housekeeping gene, or quadraplex assay to be designed. However, these initial attempts to co-amplify three or four PCR products were unsuccessful as discussed below. Instead, two duplex assays were optimized to co-amplify: (1) Ae-RpS17 and Ae-15848, and (2) Ae-4274 and Ae-8505. Multiplex reactions were validated by comparing the Ct values obtained from the duplex and single-plex assays across a 107-fold dilution series (100–107 copies). The dilution series was constructed using a mixture of linearized plasmids containing inserts for each gene of interest. For the construction of plasmids, PCR products were amplified from a pool of Ae. aegypti cDNA, gel purified, ligated into pGEM-Teasy (Promega, Madison, U.S.A) and transformants cultured. Mini-preps were digested with AatII for 2 h to linearize plasmids, which were then quantified and serially diluted. All qRT-PCR assays contained 500 nM of each primer, 200 nM Taqman probe, 6 mM MgCl2 and 2 µl template cDNA, and were amplified with the following cycling conditions: 50°C, 2 min; 95°C, 2 min; then 50 cycles of 95°C for 10 s; 60°C for 20 s; fluorescence acquisition.
All qRT-PCR assays were run in triplicate on the Corbett Rotorgene 6000 real-time PCR platform (Corbett Research, Sydney, Australia). Ct values were calculated as the second derivative maximum of the fluorescence curve using the comparative quantification analysis module in the Rotorgene software (Corbett Research, version 1.7). Mean Ct values were calculated from replicate reactions and used to construct standard curves for single and duplex reactions. Single and duplex standard curves were analyzed with the ANCOVA procedure in SAS (version 9, SAS Institute, Cary, U.S.A.) to determine that they were comparable in terms of slope.
Reverse transcription was performed using 500 ng of DNAse-treated total RNA, anchored oligo(DT)20 priming, 20 U RNaseOut (Invitrogen) and 100 U Superscript III reverse transcriptase (Invitrogen) based on the manufacturer's protocols. cDNA was diluted 5-fold to minimize the influence of PCR inhibitors. A random sample of 15 RT reactions were re-synthesized as negative RT controls (no reverse transcriptase). These were screened for genomic DNA contamination by standard PCR with Ae-RpS17 primers (95°C, 3 min; 95°C, 30s; 60°C, 30s; 72°C, 1 min; 35 cycles; 72°C, 10 min). All RT negative controls tested negative. Transcriptional profiling of Ae-RpS17, Ae-15848, Ae-8505 and Ae-4274 was performed using the multiplex qPCR assay described above.
ANOVA was applied to investigate the effects of adult mosquito age and confinement (sentinel cage versus free-range) on total RNA yield (µg) from the head and thorax of all Ae. aegypti females. Age was log10 transformed to account for curvature in the change in total RNA yield at younger ages. The influence of mosquito physiological parameters on total RNA yield was evaluated for free-range females. ANOVA was performed on total RNA yield with presence of blood in the midgut (no blood or some blood) and Christophers' ovarian development stage (stage I to V; stage G females were omitted due to the disproportionate statistical influence of this group) as factors and log-age and wing length as covariates. Estimated means for Christophers' ovarian stage groups were then calculated and compared. ANOVA was implemented in SPSS (SPSS Inc., Chigago, U.S.A).
Gene transcription measures for Ae-15848, Ae-8505 and Ae-4274 were normalized to the expression of the housekeeping gene (Ae-RpS17) by calculating log contrast values for each gene [15] (log10 of the ratio of the Ct value to the Ct value of Ae-RpS17). The effects of adult mosquito age, grouping, wing length, blood presence and ovary development on the log contrast gene transcription measures was determined using ANOVA as described above for total RNA yield.
An age prediction model was constructed using a multivariate analysis procedure that extracts a linear variable from multiple gene transcription measures [14],[15]. Briefly, log contrast values of test mosquitoes (sentinel cage females) were entered into canonical redundancy analysis to reduce the dimensionality of the data by creating new variables called redundancy variates. The age prediction calibration model was created by regression of the first redundancy variate against adult mosquito age for the sentinel cage females.
A non-parametric bootstrapping procedure was used to predict the age of each free-range female. This analysis was implemented in SAS (version 9.1; SAS Institute, Cary, U.S.A) using a SAS editor syntax that is provided in Cook et al. [15]. The sentinel cage dataset was input as the training dataset (n = 72) and log contrast values for all free-range females were input as the test dataset (n = 145). The mean of 1000 bootstrap age predictions for each free-range female was reported as its predicted age. Alternative models were investigated for the prediction of mosquito age from the transcriptional measures. A Poisson regression model with a logarithmic link function was applied because predictions were constrained to positive values. Mosquito age was analyzed so that the regression coefficients described the log of the relative risk of the independent variables (log contrast values for Ae-15848, Ae-8505 and Ae-4274 with or without total RNA yield). Poisson regression with logarithmic link models were implemented in WinBUGS [24]. In addition, the redundancy variate model was repeated as described but with total RNA yield as an additional independent variable. Sources of variance in age prediction accuracy of the redundancy variate three gene model were investigated by performing ANOVA on the age prediction residuals (predicted age minus actual age) with blood digestion and ovary development as factors and log-age and wing length as covariates.
The accuracy of mean life expectancy (ex) estimates that would be derived by applying our grading technique was investigated using Monte Carlo simulations. In particular, we tested our age grading method for the ability to differentiate two mosquito populations with a two-fold difference in mean life expectancy. This difference was chosen to test the ability to detect a 50% lifespan reduction induced by Wolbachia wMelpop infection. Two populations of 10,000 mosquitoes were created for which survival was described by exponential mortality models with ex set at 5 (probability of daily mortality [α] of 0.1) and 10 days (α = 0.2), respectively. Samples of 100, 200, 300, 400 and 500 mosquitoes of defined ages were randomly removed from the population. The age of each mosquito was predicted by randomly sampling from normal cumulative distributions of ages described by the mean and standard deviation of the experimentally predicted age estimates at each age. The distributions for even-aged mosquitoes were defined by the interpolated mean and standard deviations from the adjacent experimentally determined values. Mortality rates were estimated by calculating the regression coefficient of the natural log of the proportion of the population predicted to be within 24 hr age intervals against the age of each class. Age classes containing <3 mosquitoes were omitted from the calculations. Estimated ex values were then calculated (1/- αestimated). Sampling, age predictions and ex predictions were iterated 999 times using the Monte Carlo simulation function in the PopTools add-in in Microsoft Excel (http://www.cse.csiro.au/poptools/). Ninety five percent confidence intervals for ex were determined from the 2.5 and 97.5 percentiles of the resulting distributions.
Approximately eight thousand newly emerged Ae. aegypti adults (4804 female, 3203 male) were released inside and around an isolated homestead near Cairns, north Queensland, Australia. These free-range mosquitoes were free to engage in natural mosquito behavior including human blood feeding, mating, oviposition and harborage in natural micro-habitats. Females from this cohort were recaptured at known ages (2 d intervals from 1 to 29 d) and in varied physiological states, which provided an ideal sample to validate transcriptional mosquito age grading. Transcriptional profiles from the sentinel-caged mosquitoes were used to produce an age prediction model for the free-range females in an approach that may be applied to determine the age structure of wild Ae. aegypti populations. Mild conditions prevailed during the experiment, with average ambient temperatures of 20.9°C (range 14.7–26.5°C) in the house and 20.3°C (13.2–27.5°C) in an open sided shelter and average relative humidity of 86.2% (51.2–99.7%). Exhaustive sampling at the conclusion of the experiment collected 286 female and 11 male Ae. aegypti.
Dissection of free-range mosquitoes to determine physiological status showed that no female had mated by 1 d old (n = 10), 90% females had mated by 3 d (n = 10) and all females ≥5 d old were mated (n = 120). The first peak in blood feeding activity occurred when females were 30 hr old (±12 hr). This was reflected by high percentages of free-range females containing blood when recaptured at 3 and 5 d old (Figure S1A). The percentage of females with blood varied between 10% and 90% for subsequent samples. In some females, ovary development had advanced as far as Christophers' stage IV by 3 d old (28 hr after the first blood feeding peak; Figure S1B). The proportion of gravid females (those containing a clutch of mature, stage V ovaries) was 10% at 5 d, increased to 80% by 9 d and fell in subsequent days as females oviposited and blood feeding recommenced. For the females in which ovarian skeins could be visualized, all females ≤3 d old were nulliparous (n = 14) and all older females were parous (n = 48). The mean wing length of the free range female Ae. aegypti was 3.01 mm, SE = 0.01, significantly larger than previously recorded for wild Ae. aegypti collected under equivalent seasonal conditions in Cairns (2.85 mm, SE = 0.02; F = 71.40, df = 256, P<0.001).
A preliminary step in the transcriptional age grading assay is the isolation of total RNA from the mosquito head and thorax and reverse transcription of a standard quantity of total RNA to cDNA. Early indication that a specimen was relatively young was gained at this stage because total RNA yield decreased with age in Ae. aegypti females from 1 to 5 d old, with RNA levels stabilizing in later samples (Figure 1). The total RNA yield was not significantly different between sentinel cage and free range females (ANOVA, n = 214, P = 0.13) but was strongly influenced by log-age (P<0.001). We then examined potential physiological factors that may influence mosquito total RNA quantity and found that as well as log-age, the yield was highly influenced by ovary development (ANOVA, n = 145, P<0.001) and wing length, measured as a proxy for body size (P = 0.002), but was not influenced by the presence of blood in the midgut (P = 0.61). The model was fitted with an interaction between blood presence and ovarian development that was not significant (P = 0.15). A comparison of the main effects of ovarian development showed that egg maturation was associated with an increase in total RNA quantity in the head and thorax, with the rate steadily increasing from stage I to stage IV before total RNA levels dropped at the completion of egg development at stage V (Figure S2).
Having previously identified a set of four genes that facilitated age predictions of Ae. aegypti (three age responsive genes; Ae-15848, Ae-8505 and Ae-4274 and a housekeeping gene, Ae-RpS17 [14]), we applied Taqman-probes targeting transcripts of these genes to streamline the determination of mosquito age. Ae-RpS17 and Ae-15848 are more highly expressed than Ae-4274 and Ae-8505, with cycle threshold (Ct) values differing by approximately 10 cycles. Attempts were made to primer limit the amplification of Ae-RpS17 and Ae-15848, and allow for the efficient amplification of the other two amplicons in later cycles. However, the co-amplification of all amplicons could not be achieved despite efforts to optimize MgCl2 and primer concentration. The similarity of transcript abundance between Ae-RpS17 and Ae-15848, and Ae-4274 and Ae-8505 allowed for the development of a duplex Taqman assay without the obstacle of preferential amplification of highly abundant cDNA templates. Duplex assays were validated across a 107-fold dilution series of template abundance. Standard curves were constructed for the single and duplex Taqman assays, by plotting linear regressions of Ct value against the log concentration of linearized plasmid template. The slopes of the regression (PCR efficiency) for the single and duplex standard curves were determined to be equivalent by ANCOVA. Ae-4274 was the only amplicon where the duplex reactions were significantly different from the single-plex reactions across the 107-fold dynamic range examined (n = 53, df = 1, P<0.05).
Age related variation was evident from the transcriptional profiles of the age responsive genes in Ae. aegypti head and thorax tissue (Figure S3). Transcription is represented as the log contrast of the qPCR Ct values of the gene relative to the housekeeping gene (Ae-RpS17). Log contrasts are inverse measures of transcript abundance, increasing as transcription is decreasing. Log contrast values describing the transcription of Ae-15848 showed a four-fold increase from 1 to 29 d old (Figure S3A). The effect of log-age was highly significant (ANOVA, n = 217, P<0.001) but there were no differences between sentinel cage and free-range females (P = 0.73). For Ae-8505, log contrast values increased rapidly from 1 to 3 d and increased at a more gradual rate with age in older females (Figure S3B). Similarly, the effect of log-age was highly significant (P<0.001) and differences between caged and free-range mosquitoes were not significant (P = 0.23). Log contrast values for Ae-4274 decreased gradually with age (Figure S3C); however the effect of log-age was highly significant (P<0.001). No significant differences were observed between caged and free-range mosquitoes (P = 0.65). We then examined the influence of the measured physiological factors on the transcription of the genes as a further test of their robustness as biomarkers of age. We analyzed the effects of log-age, wing length, presence of blood in the midgut, ovary development stage and an interaction between blood presence and ovary development on the log contrast values of Ae-15848, Ae-8505 and Ae-4274. The effect of log-age was highly significant for all genes, however; none of these factors or the interaction was significant (Table S2).
The sentinel cage females were used as training samples to establish an age-prediction calibration model. The log contrast values for each female were entered into canonical redundancy analysis to produce a single redundancy variate. The analysis indicated that 68.8% of the variance in the gene transcription measures was explained by age. An age prediction calibration model was constructed from the regression of the redundancy variate for each female against adult age (Figure S4). An important outcome was that there was a strong linear component to the model (Linear regression R2 = 0.688, n = 72, P<0.001).
The free-range Ae. aegypti females were treated as age-blinded test specimens. Transcription was quantified and log contrasts were calculated as for the sentinel cage mosquitoes. Canonical redundancy analysis was used to derive a redundancy variate for each individual and age was predicted using a bootstrap procedure that applied inverse regression of the sentinel cage model from this redundancy variate. The predicted ages were then compared to the actual ages of these females, known from the time of recapture. This comparison showed a strong, near-linear relationship between the predicted ages and the actual ages of the free-range females (Figure 2; R2 = 0.62, n = 145, P<0.001). Negative ages were predicted, because the normal distribution used to model age is not constrained to positive values. A Poisson model with logarithmic link function was applied to the training and test datasets because predicted values were constrained to positive values; however no overall gains in precision or accuracy were made over the redundancy variate model (Fig. S5A). Interestingly, inclusion of total RNA yield as a predictor variable increased the fit of the Poisson regression model (Fig. S5B); however no gains were made by including total RNA in the redundancy variate model (Fig. S5C). Negative predictions were therefore manually reset to zero days. Moderate accuracy of age predictions was achieved, with 31.0% of ages predicted to within 2 d of the actual age, 55.9% to within 4 d and 77.2% to within 6 d. However, 8.3% of predictions were greater than 10 d from the actual age. Two of these samples, obvious from Fig. 2, were 7 d old females that had log contrast values for Ae-4274 that were greater than three standard deviations from the mean of all free-range females. None of the physiological factors measured (blood digestion, ovary development and body size) had a significant effect on the age prediction accuracy (Table S3). However, slight age prediction bias was observed as indicated by a significant effect of age on the error (P<0.001). Inspection of age prediction residuals against predicted age (Fig. S6) showed no clear trend indicating that an appropriate model had been fitted.
Monte Carlo simulation enabled us to model the application of transcriptional age grading to the estimation of population ex values using the experimental error distributions obtained from the release-recapture experiment (Fig. 3A). First, we determined the error distribution that would be expected through sampling error alone, assuming that all mosquitoes sampled were age graded with 100% accuracy (Fig. 3B). As expected, very little bias was observed in the estimates of ex. The estimates for the two populations were clearly differentiated and the precision of the estimates increased with increasing sample sizes. Second, we examined the estimates of ex from these same populations that would result if age estimates were derived by applying transcriptional age grading. For the population with ex of 10 d, the predicted ex values were not significantly different from the actual ex for sample sizes of 100–300 mosquitoes but an underlying bias towards underestimation of ex became evident at greater sample sizes (Fig. 3C). However, estimates for the ex = 5 days population were significantly greater than the actual ex values at all sample sizes. An important outcome was that the two populations could be significantly differentiated from each other when estimates of ex were based on the age predictions of >200 mosquitoes.
We have shown that age grading mosquitoes based on gene transcription can be successfully applied to adult female Ae. aegypti maintained in the wild. The ability to determine the age of female Ae. aegypti to an accuracy of ±6 d for 72.2% of females under field conditions is a valuable asset for investigations of mosquito population age structure. One of many applications of mosquito age grading is for the assessment of the efficacy of mosquito control interventions, whether testing the capacity of an entomopathogenic micro-organism to shorten the mean life expectancy of a mosquito population, or to test the fitness of a transgenic mosquito with impaired ability to transmit pathogens in comparison to wild type mosquitoes. We demonstrated the capacity for transcriptional age grading to differentiate between two populations of mosquitoes having mean life expectancies of 5 and 10 d. A 2-fold difference was chosen to test the capacity of the model to identify a 50% lifespan reduction of Ae. aegypti females that is expected to result from a dengue control intervention based on Wolbachia intracellular bacteria [6],[8],[9]. Although bias was evident from absolute estimates of ex derived from transcriptional age estimates, a relative comparison of the predicted life expectancies enabled successful differentiation of the populations.
We have increased the throughput of transcriptional age grading by applying duplex Taqman probe assays to the quantification of gene transcripts which is highly desirable for investigations of mosquito population age structure in which hundreds of specimens potentially require age grading. We applied a stringent test for validation of transcriptional age grading by releasing free-range Ae. aegypti, thereby allowing mosquitoes to engage in normal mosquito behavior, including blood feeding, dispersal in search of natural oviposition and resting sites and thermoregulation through harborage in typical microhabitats. As a result, females were recaptured in various physiological states at each age. However, of several physiological factors assessed by dissection (presence of blood in the midgut, ovary development and body size), none affected the transcriptional profiles of the age responsive genes or age prediction accuracy. Similarly, transcriptional profiles did not differ between females from the sentinel cages or the free-range females which is important from an applied perspective as it indicates that age prediction models for wild caught mosquitoes can be calibrated on known-age mosquitoes maintained in captivity.
There was some decrease in accuracy when compared to our previous application of the three gene age prediction model to Ae. aegypti maintained in field cages up until 19 days old [14]. However, we have increased the sample size of test specimens in the present study (30 to 145) and have extended the maximum age of samples to 29 d here. A Poisson log link model and alternative combinations of genes and total RNA yield predictor variables were tested; however, no improvements in accuracy were achieved over the three gene redundancy variate model. Outliers were attributed to aberrant gene transcription values, in some cases greater than three standard deviations from the mean of the group. The reasons for these extreme observations could not be determined. However, these outliers comprised less than 5% of all predictions. Environmental variation or other physiological factors not measured could account for the differences. The free-range females were larger than wild specimens collected under equivalent seasonal conditions; however, investigations in the laboratory have shown that transcription of the age responsive genes is robust to body size variation induced by varying the quantity of food provided to larvae (LEH, unpublished data).
We have also shown that a yield of total RNA from the head and thorax of an adult female above a threshold (4.4 µg in our experiments) provides early indication that the specimen is a newly emerged, <3 d old adult (a teneral adult). A 40% decrease in total RNA abundance with age from 1 to 7 d old has been previously observed from Ae. aegypti females [20]. High levels of total RNA in 1–2 d old adults followed a peak in total RNA abundance during the pupal stage, and were probably a residual effect of high transcription rates during metamorphosis. However, age related changes to total RNA yields have not previously been utilized for mosquito age grading assessments. In D. melanogaster, total RNA abundance decreased by 60% at a constant rate from 2 to 40 d of age [21]. Total ribosomal RNA, transfer RNA and mRNA levels decrease rapidly from emergence to 10 d in Drosophila and are thought to be due to down-regulation of RNA polymerase I, II and III mediated transcription [22]. Total RNA yield is dependent on the method of extraction used and we have shown that total RNA yield in the adult head and thorax increases with body size and during ovary development. These factors should be standardized in age grading assessments of wild mosquitoes based on total RNA yield.
The ability to accurately determine the ages of wild caught mosquitoes is crucial for investigations into the population dynamics and vulnerabilities of important mosquito vectors. In particular, the capacity to differentiate two populations of mosquitoes based on changes to mean life expectancy will be critical for evaluating new dengue control interventions targeting mosquito longevity. We estimate that adopting the Taqman probe multiplexing approach saves approximately 20% in reagent and 30% time savings when compared to the equivalent Sybr green based approach. The cost of reagents required to derive age predictions was approximately US$7.5 per mosquito. Improvements in accuracy and throughput can be expected if additional age responsive genes are identified and included in the model. Optimized Taqman reactions could enable these transcripts to be measured in triplex or quadriplex qPCR assays. The genes on which our age prediction assay is based are conserved and therefore there is a large potential for the development of transcriptional age grading methods for other insect vectors of tropical diseases.
|
10.1371/journal.pntd.0002681 | Zika Virus in Gabon (Central Africa) – 2007: A New Threat from Aedes albopictus? | Chikungunya and dengue viruses emerged in Gabon in 2007, with large outbreaks primarily affecting the capital Libreville and several northern towns. Both viruses subsequently spread to the south-east of the country, with new outbreaks occurring in 2010. The mosquito species Aedes albopictus, that was known as a secondary vector for both viruses, recently invaded the country and was the primary vector involved in the Gabonese outbreaks. We conducted a retrospective study of human sera and mosquitoes collected in Gabon from 2007 to 2010, in order to identify other circulating arboviruses.
Sample collections, including 4312 sera from patients presenting with painful febrile disease, and 4665 mosquitoes belonging to 9 species, split into 247 pools (including 137 pools of Aedes albopictus), were screened with molecular biology methods. Five human sera and two Aedes albopictus pools, all sampled in an urban setting during the 2007 outbreak, were positive for the flavivirus Zika (ZIKV). The ratio of Aedes albopictus pools positive for ZIKV was similar to that positive for dengue virus during the concomitant dengue outbreak suggesting similar mosquito infection rates and, presumably, underlying a human ZIKV outbreak. ZIKV sequences from the envelope and NS3 genes were amplified from a human serum sample. Phylogenetic analysis placed the Gabonese ZIKV at a basal position in the African lineage, pointing to ancestral genetic diversification and spread.
We provide the first direct evidence of human ZIKV infections in Gabon, and its first occurrence in the Asian tiger mosquito, Aedes albopictus. These data reveal an unusual natural life cycle for this virus, occurring in an urban environment, and potentially representing a new emerging threat due to this novel association with a highly invasive vector whose geographic range is still expanding across the globe.
| Not previously considered an important human arboviral pathogen, the epidemic capacity of Zika virus (ZIKV, a dengue-related flavivirus) was revealed by the Micronesia outbreak in 2007, which affected about 5000 persons. Widely distributed throughout tropical areas of Asia and Africa, ZIKV is transmitted by a broad range of mosquito species, most of which are sylvatic or rural, Aedes aegypti, an anthropophilic and urban species, being considered the main ZIKV epidemic vector. In a context of emerging arbovirus infections (chikungunya (CHIKV) and dengue (DENV)) in Gabon since 2007, we conducted a retrospective study to detect other, related viruses. In samples collected during the concurrent CHIKV/DENV outbreaks that occurred in the capital city in 2007, we detected ZIKV in both humans and mosquitoes, and notably the Asian mosquito Aedes albopictus that recently invaded the country and was the main vector responsible for these outbreaks. We found that the Gabonese ZIKV strain belonged to the African lineage, and phylogenetic analysis suggested ancestral diversification and spread rather than recent introduction. These findings, showing for the first time epidemic ZIKV activity in an urban environment in Central Africa and the presence of ZIKV in the invasive mosquito Aedes albopictus, raise the possibility of a new emerging threat to human health.
| Zika virus (ZIKV) is a mosquito-borne flavivirus phylogenetically related to dengue viruses. Following its first isolation in 1947 from a sentinel monkey placed in the Zika forest in Uganda [1], serological surveys and viral isolations (reviewed in [2]) suggested that ZIKV (i) ranged widely throughout Africa and Asia, and (ii) circulated according to a zoonotic cycle involving non-human primates and a broad spectrum of potential mosquito vector species.
In Africa, ZIKV has been isolated from humans in western and central countries such as Senegal, Nigeria, Central African Republic and Uganda [3]–[7]. Serological surveys (reviewed in [2]) suggested that its geographic range might extend not only to other West and Central African countries (Sierra Leone, Cameroon, Gabon), but also to eastern (Ethiopia, Kenya, Tanzania and Somalia) and northern Africa (Egypt). ZIKV has also been isolated from mosquitoes collected in Senegal, Ivory Coast, Burkina Faso, Central African Republic and Uganda [1], [6], [8], [9]. These mosquitoes mainly belonged to sylvan or rural species of the genus Aedes, and more precisely to the Aedimorphus, Diceromyia and Stegomyia subgenera. The virus has also been isolated in West Africa (Burkina Faso, Senegal and Ivory Coast) [6], [9] and Asia [10] from Aedes aegypti, a species being considered the main ZIKV epidemic vector outside Africa [11]. Moreover, Ae. aegypti was shown experimentally to be an efficient ZIKV vector [12]–[14].
Despite its apparent broad geographic distribution in Africa and Asia, only sporadic cases of human ZIKV infection have been reported. This virus received little attention until its sudden emergence in Yap Island (Micronesia) in 2007, which involved about 5000 persons [15], [16], revealing its epidemic capacity. Patients develop a mild dengue-like syndrome, including fever, headache, rash, arthralgia and conjunctivitis. This clinical similarity with other, more commonly diagnosed arboviral infections such as chikungunya (CHIKV) and dengue (DENV), might delay the diagnosis and/or lead to underestimation of ZIKV infections.
Here, we report the first direct evidence of ZIKV epidemic activity in Central Africa, and its occurrence in an urban environment during concomitant CHIKV/DENV outbreaks in Libreville, the capital of Gabon, in 2007. We also report the first detection of ZIKV in the Asian tiger mosquito, Ae. albopictus. These findings, together with the global geographic expansion of this invasive species and its increasing importance as epidemic vector of arboviruses as exemplified by CHIKV adaptation, suggest that the prerequisites for the emergence and global spread of Zika virus may soon be satisfied.
In 2007 and 2010, Gabon recorded simultaneous outbreaks of CHIKV (genus Alphavirus) and DENV (genus Flavivirus) infections. The 2007 outbreaks primarily affected Libreville, the capital of Gabon, and subsequently extended northwards to several other towns [17], while the 2010 outbreaks occurred in the south-eastern provinces [18]. To detect other circulating arboviruses, we conducted a retrospective study based on molecular screening of 4312 sera from symptomatic patients presenting to healthcare centers; 24.7% of the samples were obtained during the 2007 outbreaks, 9.7% during the inter-epidemic period, and 65.5% during the 2010 outbreaks (data not shown). We also analyzed a collection of 4665 mosquitoes captured during the same period and split into 247 pools according to the species, date and sampling site (Table 1, see [18] and [19] for the details of the methodology used for mosquito trapping).
The Centre International de Recherches Médicales de Franceville (CIRMF) and the Gabonese Ministry of Health cooperated in the 2007 and 2010 outbreak response and management, that included blood sampling for laboratory diagnostic and epidemiological survey. The study was approved by our Institutional review board (Conseil scientifique du CIRMF).
Symptomatic patients presented to health care centers for medical examination. All patients were informed that blood sampling was required for laboratory diagnosis of suspected acute infections, such as malaria, dengue or chikungunya fever. During the two outbreaks, given the urgency of diagnosis, only oral consent was obtained for blood sampling and was approved by the institutional review board. However during the active surveillance study that was performed between the two outbreaks (described in reference [18]), written consent could be obtained.
Primary molecular screening was based on hemi-nested reverse-transcription PCR (hnRT-PCR) with the generic primers PF1S/PF2Rbis/PF3S targeting highly conserved motifs in the flavivirus polymerase (NS5) gene (280-bp) [20]. Yellow fever virus RNA (vaccinal strain 17D) was used as a positive control. A second screening was performed with a ZIKV-specific real-time PCR method using the primers-probe system ZIKV-1086/ZIKV-1162c/ZIKV-1107-FAM [16], also targeting a short sequence (160 bp) of the NS5 gene.
Virus isolation was attempted on the Vero and C6/36 cell lines but was unsuccessful, presumably because of low viral titers (despite two patients presenting only 1 and 4 days after symptom onset), and unsuitable initial storage conditions. To further characterize the Gabonese ZIKV strains, partial envelope (E) (841 bp) and NS3 (772 bp) gene sequences were amplified by conventional nested RT-PCR with specific primers derived from published ZIKV sequences. The primer pairs targeting the E gene were ZIK-ES1 (TGGGGAAAYGGDTGTGGACTYTTTGG)/ZIK-ER1 (CCYCCRACTGATCCRAARTCCCA) and ZIK-ES2 (GGGAGYYTGGTGACATGYGCYAAGTT)/ZIK-ER2 (CCRATGGTGCTRCCACTCCTRTGCCA). The primer pairs for NS3 amplification were ZIK-NS3FS (GGRGTCTTCCACACYATGTGGCACGTYACA)/ZIK-NS3FR (TTCCTGCCTATRCGYCCYCTCCTCTGRGCAGC) and ZIK-X1 (AGAGTGATAGGACTCTATGG)/ZIK-X2 (GTTGGCRCCCATCTCTGARATGTCAGT).
The E and NS3 sequences obtained from one Gabonese patient were concatenated and analyzed using a set of previously published ZIKV sequences. Phylogenetic relationships were reconstructed with the maximum likelihood algorithm implemented in PhyML [21] (available at http://www.atgc-montpellier.fr/phyml/) with best of NNI (Nearest Neighbor Interchange) and SPR (Subtree Pruning and Regrafting) criteria for tree topology searching, and the GTR model of nucleotide substitutions. The Gamma distribution of rate heterogeneity was set to 4 categories, with a proportion of invariable sites and an alpha parameter estimated from the dataset. Branch support was assessed from 100 bootstrap replicates. Tree reconstructions were also performed by Bayesian inference with MrBayes v3.2 [22] under the GTR+I+G model of nucleotide substitutions, and with the distance neighbor-joining method [23] implemented in MEGA5 [24] with confidence levels estimated for 1000 replicates. To test for phylogenetic discrepancies, tree reconstructions were also performed independently from the envelope dataset and the NS3 dataset with PhyML according to the parameters described above.
The resulting trees were visualized with the FigTree software (Available at: http://tree.bio.ed.ac.uk/software/figtree/), and rooted on midpoint for clarity. The Genbank accession numbers for the Gabonese ZIKV strain are KF270886 (envelope) and KF270887 (NS3).
The NS5 PCR products were sequenced, resulting in the first ZIKV RNA detection in a human sample (Cocobeach) and in two Ae. albopictus pools (Libreville) collected during the 2007 outbreaks. Real-time PCR was then performed, leading to the detection of four additional positive human samples, collected in 2007 in four suburbs of Libreville (Diba-Diba, Nzeng-Ayong, PK8, PK9) (Figure 1). No ZIKV was detected during the inter-epidemic period or during the 2010 outbreaks.
Clinical information was available for only one ZIKV-positive patient, who had mild arthralgia, subjective fever, headache, rash, mild asthenia, myalgia, diarrhea and vomiting. No information was available on this patient's outcome. Cycle threshold values for human blood samples were high (>37 cycles), suggesting low viral loads (data not shown).
Aedes albopictus was the predominant species collected, accounting for 55.4% of the mosquito pools, while Aedes aegypti accounted for 18.2% (Table 1). The other mosquito species consisted of members of the Aedes simpsoni complex, Anopheles gambiae, Mansonia africana, Mansonia uniformis, Culex quinquefasciatus, Eretmapodites quinquevittatus and unidentified Culex species. Positive mosquito pools were captured from two suburbs (Nzeng-Ayong and Alenkiri) where Aedes albopictus was the predominant species (Figure 1, Table 1).
As isolation on the Vero and C6/36 cell lines failed, the Gabonese ZIKV strain was further characterized by partial sequencing of the E and NS3 genes. Phylogenetic analysis was performed on concatenated E and NS3 sequences from one Cocobeach serum sample. The resulting tree topology (Figure 2) was similar to that previously obtained from the complete coding sequences, corroborating Asian and African distinct lineages [2]. The African lineage was further split into two groups, one containing the genetic variants of the MR766 strain (Uganda, 1947) and the second including West African strains (Nigeria, 1968; Senegal, 1984) and the new ZIKV sequence from Gabon, at a basal position. Phylogenetic trees derived from the E and NS3 partial sequences resulted in a similar topology, apart from the weakly supported branching pattern for the MR766 variant DQ859059, oscillating between the two African sister groups (Supporting Figure S1). The deletions in potential glycosylation sites previously reported for the Nigerian ZIKV strain and two variants of the Ugandan strain MR766 (sequences AY632535 and DQ859059) [2] were absent from the Gabonese ZIKV sequence.
Evidence of human ZIKV infections in Central Africa is limited to one isolate from RCA in 1991 [6] and two serological surveys performed 50 years ago in Gabon [25], [26]. No report of human ZIKV infections was made in other countries of the Congo basin forest block, despite probable circulation through a sylvan natural cycle. We provide here the first direct evidence of human ZIKV infections in Gabon, as well as its occurrence in an urban transmission cycle, and the probable role of Ae. albopictus as an epidemic vector.
Our phylogenetic results are in agreement with the tree topology previously obtained with complete coding sequences of ZIKV strains, showing an African lineage and an Asian lineage [2]. The branching pattern obtained here suggests that ZIKV emergence in Gabon did not result from strain importation but rather from the diversification and spread of an ancestral strain belonging to the African lineage. The identification of ZIKV in two different localities of Gabon (Cocobeach and Libreville) suggests that the virus was widespread rather than restricted to a single epidemic focus. The simultaneous occurrence of human and mosquito infections in Libreville also suggests that the virus circulated in 2007 in an epidemic cycle rather than as isolated cases introduced from sylvan cycles.
Of note, ZIKV transmission occurred here in a previously undocumented urban cycle, supporting the potential for urbanization suggested in 2010 by Weaver and Reisen [27]. While some mosquito species (including Ae. aegypti) previously found to be associated with ZIKV, were captured and tested here, only Ae. albopictus pools were positive for this virus. Moreover, this species largely outnumbered Ae. aegypti in the suburbs of Libreville where human cases were detected, suggesting that Ae. albopictus played a major role in ZIKV transmission in Libreville.
The ratio of ZIKV-positive Ae. albopictus pools is similar to that reported for DENV-positive pools, suggesting that these two viruses infect similar proportions of mosquitoes. The small number of recorded human ZIKV cases, by comparison with DENV cases, may be due to the occurrence of subclinical forms of ZIKV infections that did not required medical attention. Thus, an underlying ZIKV epidemic transmission might have been masked by concomitant CHIKV/DENV outbreaks.
The natural histories of CHIKV and ZIKV display several similarities. Before the large Indian Ocean outbreaks in 2005–2007, chikungunya fever was a neglected arboviral disease. Both viruses are phylogenetically closely related to African viruses [28]–[30] suggesting they probably originated in Africa, where they circulated in an enzootic sylvan cycle involving non-human primates and a wide variety of mosquito species, human outbreaks presumably being mediated by Ae. aegypti [5], [31]. In Asia, both viruses are thought to circulate mainly in a human-mosquito cycle involving Ae. aegypti [11], [14], [31]. Together with the recent Yap Island outbreak, this prompted some researchers to re-examine the susceptibility of Ae. aegypti to ZIKV infection [14]. However, it must be noted that the vector of the Yap Island outbreak was not definitely identified since the predominant potential vector species Aedes hensilli remained negative [15], and that ZIKV has been isolated only once from Ae. aegypti in Asia [10], so that its vector status in natura is not confirmed. Additionally, a ZIKV enzootic transmission cycle involving non-human primates in Asia and sylvatic vectors cannot be ruled out as suggested by serologic studies carried on orangutang [32], [33]. Finally both CHIKV and ZIKV have shown their ability to adapt to a new vector, Ae. albopictus, upon introduction in an environment where their primary vector was outnumbered. This mosquito species being native to South-East Asia, our findings may help to explain human ZIKV transmission in Asia.
Aedes albopictus was first introduced in Africa in 1991 [34] and detected in Gabon in 2007, where its invasion likely contributed to the emergence of CHIKV and DENV in this country [17]–[19], [35]. Multiple lines of evidence supporting its increasing impact as an arboviral vector have also been obtained during CHIKV outbreaks in the Indian Ocean region (2005–2007) and in Italy (2007) [36], [37] through viral evolutionary convergence of Ae. albopictus adaptive mutations [38]–[41]. Whether or not the transmission of ZIKV in Central Africa was also link to such an adaptative mutation of ZIKV to Ae. albopictus cannot be answered at this stage. Wong and colleagues [42] have just demonstrated experimentally that Ae. albopictus strains from Singapore were orally receptive to the Ugandan strain of ZIKV sampled in 1947, suggesting that this virus-vector association in Africa may have been previously prevented because the required ecological conditions did not yet exist. However, given the relatively low ZIKV viral loads previously reported in patients - with an order of magnitude of 105 copies/ml compared to 107 to 109 copies/ml for CHIKV [16], [18], [40] - the oral infectivity for Ae. albopictus may seem at least as critical as it was for CHIKV in establishing this new human-mosquito cycle.
Why ZIKV has not yet been detected in the areas where DENV and CHIKV have already spread via Ae. albopictus is unclear, but it may be an ongoing process which we are just starting to detect. The spread of CHIKV reflects the ability of arboviruses to adapt to alternative hosts, and the resulting public health concerns in both developed and developing countries. Is ZIKV the next virus to succeed CHIKV as an emerging global threat? The increasing geographic range of Ae. albopictus in Africa, Europe, and the Americas [34], [36], [43], [44], together with the ongoing ZIKV outbreak in French Polynesia at the time of writing [45] suggest this possibility should be seriously considered. Analysis of sylvan and urban transmission cycles, together with viral genetics and vector competence studies, are now required to assess (i) how ZIKV is able to establish a sustainable transmission cycle involving this new vector in Central Africa, (ii) vector(s)-virus relationships in Asia, and (iii) the risk of importation and spread to new areas where Ae. albopictus occurs as well.
|
10.1371/journal.pntd.0005325 | Evaluation and Monitoring of Mycobacterium leprae Transmission in Household Contacts of Patients with Hansen's Disease in Colombia | Leprosy in Colombia is in a stage of post elimination—since 1997, prevalence of the disease is less than 1/10000. However, the incidence of leprosy has remained stable, with 400–500 new cases reported annually, with MB leprosy representing 70% of these case and 10% having grade 2 disability. Thus, leprosy transmission is still occurring, and household contacts (HHCs) of leprosy patients are a population at high risk of contracting and suffering from the effects of the disease during their lifetime. We performed a cross-sectional study with the aim of evaluating leprosy transmission within Family Groups (FGs) from four Colombian departments: Antioquia, Bolívar, Córdoba and Sucre. This study included 159 FGs formed by 543 HHCs; 45 FGs were monitored twice, first in 2003 and again in 2012. Migration, forced displacement by violence, loss of contact with the health center and the lack of an agreement to participate in the second monitoring were the primary reasons not all FGs were tested a second time. In each HHC, a clinical examination was performed, epidemiological data recorded, the bacillary index determined, DNA was isolated for M. leprae detection by nested PCR and IgM anti-phenolic glycolipid-I (PGL-I) titers were inspected. Further, DNA from M. leprae isolates were typed and compared among FGs. Twenty-two (4.1%) of the 543 HHCs had IgM anti-PGL-I positive antibody titers, indicating infection. Nasal swabs (NS) taken from 113 HHCs were tested by RLEP PCR; 18 (16%) were positive for M. leprae DNA and two new leprosy cases were detected among the HHCs. Of the confirmed HHCs with leprosy, it was possible to genotype the bacterial strains from both the index case and their HHCs. We found that the genotype of these two strains agreed at 9 markers, showing the individuals to be infected by the same strain, indicating familiar transmission. HHCs of leprosy patients not only are a high-risk population for M. leprae infection, they can act as M. leprae carriers and therefore serve as sources for transmission and infection. Our results confirm familiar leprosy transmission and suggest that follow-up of HHCs is a good strategy for early diagnosis of leprosy and to monitor its transmission.
| Leprosy in Colombia is considered in a post-elimination stage, as prevalence of the disease is less than 1/10000 since 1997. However, leprosy transmission is still common with 400 to 500 new cases reported each year—70% of them multibacillary (MB) and 10% with grade 2 disability, demonstrating late diagnosis. HHCs of leprosy patients are a population at risk for infection by M. leprae and the subsequent development of leprosy. However, the M. leprae incubation period is long, and measures to follow-up with this population are difficult and not included in leprosy control programs in Colombia. We performed this survey with the aim to evaluate leprosy transmission in family groups of leprosy patients from four Colombian departments: Antioquia, Bolívar, Córdoba and Sucre. Volunteers (n = 713), 170 (24%) leprosy patients and 543 (76%) HHCs belonging to 159 family groups (FG) were included after informed consent was given. Of these volunteers, 225 (31.5%) were monitored two times: 45 leprosy patients and 180 HHCs. Volunteers were given a clinical examination and epidemiological data was recorded. Skin biopsies, nasal swabs and slit skin samples from patients were taken for bacillary index determination and Mycobacterium leprae genotyping. Nasal swabs and slit skin samples from HHCs were tested by nested PCR. Additionally, serum samples were tested for IgM anti-phenolic glycolipid-I (PGL-I) titers. Twenty-two (4.1%) of the 543 HHCs had IgM anti-PGL-I positive antibody titers, indicating infection. PCR of DNA isolated from nasal swabs was positive for M. leprae in 18 (16%) HHCs, suggesting the presence of carriers. Two new leprosy cases were detected among the HHCs. We found three leprosy cases in one family group, two of them multibacillary. The M. leprae genotype of these two strains agreed at 9 markers, showing these individuals are infected by the same M. leprae strain, indicative of familiar transmission. Our results confirm that leprosy transmission is active in a country where leprosy is in post-elimination stage.
| Leprosy, also known as Hansen's disease, is an infectious and chronic disease caused by Mycobacterium leprae [1]. The mode of transmission of M. leprae has not yet been demonstrated, although entry through the nasal passages is a commonly accepted potential mechanism [2]. While humans are the main reservoir of M. leprae, nine-banded armadillos are also known to serve as reservoirs of this bacterium [2]. It is estimated that about 2 million people worldwide have some type of disability due to leprosy [3]. While multidrug therapy (MDT) has been highly effective in treating leprosy infections, treating nerve damage that results from the disease has proven more difficult [3].
In Colombia, the detection of new leprosy cases decreased in 2009 and 2010. However, the number of new cases remained stable in 2011 (434 cases), 2012 (364 cases), 2013 (433 cases), 2014 (370), and again in 2015 (349) [4, 5]. These data suggest that the transmission of leprosy in Colombia continues despite the country classified as being in a period of post-elimination. It has been observed that the global decrease in leprosy prevalence has not been accompanied by a decrease in the incidence of the disease [6]. The late diagnosis of leprosy in Colombia is evident by the proportion of multibacillary (MB) to paucibacillary (PB) leprosy cases of 70/30, with 10% of MB patients having grade 2 disability. Thus, the prevention of transmission has not been achieved despite the implementation of MDT programs. Further complicating matters is the under-reporting of the disease [6]. Prominent reasons why the incidence of the disease continues in endemic countries appears to be the presence of reservoirs within infected populations—sub-clinical leprosy or non-human environmental sources that have not been detected [7–9].
In comparison with the general population, household contacts (HHCs) of leprosy patients are a population at high risk of contracting the disease and suffering the effects of M. leprae infection during their lifetimes. Studies have demonstrated that most new leprosy patients have had contact with another patient [10,11]. Due to the long and imprecise incubation period of leprosy, it cannot be determined which HHC will ultimately develop leprosy. Further, Colombian health programs do not regularly monitor HHCs of leprosy patients. Using enzyme-linked immunosorbent assays (ELISA) to determine M. leprae infection, the phenolic glycolipid-1 (PGL-I) has been found to be specific to M.leprae [10]. While MB patients generate antibodies against PGL-I, PB patients do not. In HHCs of leprosy patients, detection of these antibodies may be indicative of infection but offer no protection against the disease [10, 12–14].
One form of protection for HHCs used in some countries is the Calmette-Guerin Bacillus (BCG) vaccine [15], recognized for its protection against Mycobacterium tuberculosis infection. The protective effect of the BCG vaccine to non-infected persons ranges from 10–80% [16], with the vaccine considered a stimulus to the immunologic reactivity of the HHCs of leprosy patients. It is possible that the combination of medication and BCG vaccine may facilitate elimination of M. leprae in the host (by increasing TNF-α and IL12 levels and activating macrophages), decrease relapse rates and shorten the positivity of the bacilloscopy [16,17].
Molecular tests have been developed that detect specific M. leprae nucleic acids with high sensitivity and specificity and are used to confirm the diagnosis of leprosy in PB patients and to detect the bacterium in asymptomatic HHCs [18,19]. Likewise, advances in the genotyping of M.leprae based on insertions, deletions, Single Nucleotide Polymorphisms (SNPs) and Short Tandem Repeats (STRs) have revolutionized our understanding of leprosy’s origins, its patterns of migration and propagation and the disease’s resistance to drug treatment [20]. These tools can be used to better understand areas where the disease is present and its means of transmission [18–21].
Leprosy control programs in Colombia include a clinical review of HHCs immediately after diagnosis of the index case has occurred [22]. This vigilance is important, but not sufficient because leprosy has a variable period of latency; clinical follow-ups for several years are necessary to detect the early stages of the disease. Additionally, a clinical exam is not a good tool to detect subclinical cases of the disease [23].
In the current study, we monitored leprosy transmission in HHCs of patients with leprosy. We examined clinical, bacteriologic, and immunologic changes in the HHCs. We also monitored genetic markers in the bacterium, which may improve early detection and improve knowledge about transmission of the disease, thereby avoiding late diagnosis and preventing permanent damage resulting from the disease.
This study was approved as the minimal risk by the ethical committee of the Instituto Colombiano de Medicina Tropical–Universidad CES. An informed consent form was signed by patients, HHCs, and parents or tutors of children under 18 years of age.
A cross sectional survey was performed in the leprosy cases, and their HHCs, registered from 2003 to 2012 in the Colombian departments of Antioquia, Bolívar, Córdoba and Sucre. Leprosy patients and their HHCs were monitored once or twice by examining their epidemiological, clinical, bacteriologic, and IgM PGL-I antibody titer changes. The first monitoring was performed in 2003, the second in 2012. All volunteers, parents or tutors of children signed a consent form to participate in this survey.
For each index case (leprosy patient) and HHC (family member or any person that lived under the same roof with the index case for more than six months), a clinical record was filed which included medical and epidemiological data. Age, sex, the relationship with the index case, and detection of a BCG vaccination scar were recorded. Finally, clinical symptoms were recorded as well as data regarding the treatment stage according to each individual.
Each HHC was examined for signs and symptoms of leprosy. This included the detection of areas of hypoesthesia or anesthesia, thermic sensibility to cold and heat, palpation of the nerve trunks, presence of hypo- or hyper-pigmented lesions, unnoticed burns or wounds, nodules, atrophy, contractures, anomalous positions of the fingers, loss of muscular strength and an alteration of motion. A HHC was classified as symptomatic when he or she presented at least one of these symptoms.
The classification of leprosy was performed according to World Health Organization (WHO) recommendations. Patients classified with MB leprosy had a positive bacillary index (BI) and more than five skin lesions. Patients classified with PB had a negative BI and less than five skin lesions [6]. For the prescription of treatment, clinical classification by Ridley and Joplin [8] was also used.
A descriptive and bivariate analysis of the data was performed using Statistical Package for the Social Sciences program (SPSS Inc, Chicago, IL) PASW Statistics 18. The odds ratio (95% CI) was calculated and a P value < 0.05 was considered significant.
A total of 159 FGs comprised of 713 individuals were included in this study: 170 leprosy patients (24%) and 543 HHCs (76%). A total of 225 individuals corresponding to 32% of the study population were monitored twice: 46 leprosy patients (20.4%) and 180 HHCs (44.8%). Table 2 describes the characteristics of the leprosy patients and their family groups.
Of the 170 leprosy patients, 135 (79.4%) were MB and 35 PB (20.6%). A higher frequency of leprosy was found in men than in women with a ratio of 1 woman per 3.4 men.
The average age of the patients was 53 years with a variation of 17.6 years. Half of the patients were over 51 years old with a variation of 28 years; the interquartile range (IQR) was 40–68.3. The minimum age was 5 years and the maximum age was 90 years.
In this study we found four leprosy patients undergoing treatment younger than 18 years of age: a 5 year old girl with MB leprosy who had an uncle as a family contact, a 9 year old boy with MB leprosy whose mother was in treatment for MB at the moment of his diagnosis and two young boys of 14 and 16 years old, both with MB leprosy. Primary school was the highest educational level achieved by 83 (48.8%) of the leprosy patients studied.
BCG scars were evident in 27 (15.9%) of the patients. 114 (67.1%) of patients did not have the scar and in 29 (17.1%) it was not possible determine if a vaccination was carried out. We found a statistically significant relationship between positive BCG scars and not having leprosy (p = 0.0001), OR: 0.131, IC (95%): 0.083–0.207. We did not observe a statistically significant relationship between receiving the vaccination and MB vs. PB leprosy (p = 0.2615), OR: 1.867, IC (95%) 0.619–5.627.
Average age of the HHCs was 32 ± 20.3. Half of the HHCs were over 27 years old with a variation of 32 (IQR was of 15–47). The minimum age was 1 year and the maximum was 90 years. The 543 HHCs and the 170 leprosy patients belonged to 154 family groups. Table 2 shows the families characteristics.
M. leprae typing of the index case (in the FG where the two new cases were detected), and one of the new cases (a MB patient), confirmed familiar transmission. Table 3 shows the genotypes of both M.leprae isolates.
This study describes leprosy transmission from index cases to their family groups in the Colombian departments of Antioquia, Bolívar, Córdoba and Sucre from 2003 to 2012. Clinical exams, bacillary index, RLEP PCR, IgM anti-PGL-I titers and M. leprae genotyping were performed to determine leprosy transmission.
Of the leprosy patients and HHCs monitored, it was possible to contact 225 of them (32%) a second time. Due to the lack of follow up from Hansen’s programs after treatment, it is difficult to contact patients and their families again.
We found a greater incidence of the disease in men (77.1%) compared to women, which coincides with other studies [25,26]. We observed no relationship between gender and the MB or PB status (p>0.05) [26].
Leprosy in Colombia is considered to be in post elimination phase [27]. However, the four children under 14 years of age that are currently undergoing treatment and the new case (a 5 year old) diagnosed during this study are important epidemiological reminders that should be considered indicators of the prevalence of the disease in the general population as well as a sign of ongoing transmission. Some studies suggest that the long incubation period of the disease affects children in the age range of 10–14 years old; nevertheless, the affected children between 1 and 9 years old likely reflect their early exposure to active cases of the disease [28,29] and/or to areas of transmission within communities [28].
The socio-economic status of the leprosy population was revealed during this study: 48.8% of the patients only had a primary school level of education and 27.1% did not receive any type of scholarly study, results in accordance with other reports [30]. One of the difficulties we encountered during the socio-epidemiologic survey was the lack of information from patients regarding their age, their knowledge of the disease, any previous MDT treatment, the number of supplied doses of MDT, complementary treatments, the date of diagnosis of the disease and their current treatment status. This reflects the patient’s lack of education regarding leprosy.
The BCG vaccine is known to protect against Mycobacterium tuberculosis infection; cross protection of the vaccine for M. leprae ranges from 10 to 80% [16]. In this study, the BGC vaccine showed a protective effect of 87% (OR: 0.13, IC95%: 0.08–0.21). However, to confirm this result, a follow up of the same population must be performed. Being a protective measure, we found a high percentage of leprosy patients that had not received the BCG vaccine, primarily due to the fact that leprosy patients in Colombia are not vaccinated at an early age—the majority of those vaccinated are in their adult years—or their access to the vaccine was limited or unavailable in the areas where they live [16]. Leprosy control programs in Brazil recommend the BCG vaccination to all the healthy persons who are in contact with leprosy patients [15]. In Colombia, BCG vaccination of HHCs has been established for their protection [31]. Our results show that 60% of HHCs had evidence of receiving the vaccine in the form of a scar while 33.3% did not, indicating that the Hansen’s disease programs of these departments do not implement revaccination to 100% of HHCs after diagnosis of the index case.
Use of the BCG vaccine is considered a stimulus for immunological reactivity, possibly due to the fact that the combination of MDT and the BCG vaccine may facilitate the elimination of M.leprae from the patient (increasing the TNF alpha, IL12 and activating macrophages), decreasing the rates of relapse and reducing the positivity of the BI [16,17].
Numerous studies have demonstrated that leprosy appears to have a relationship between the clinical outcome of the disease and a familiar relationship with a leprosy patient. Correa et al [30] found that the reports of family leprosy are of first and second grade consanguinity. The current study found that 495 (91.2%) of the HHCs have had an exposure time of years with the index case, while only 2 (0.4%) had occasional contact, suggesting that HHCs are exposed for a prolonged time to BI positive patients without a diagnosis.
The presence of PGL-I antibodies in the HHCs of patients has been widely studied. Nevertheless, few studies have performed long-term monitoring of HHCs [10]. The current study shows IgM anti-PGL-I in 4.1% of HHCs. However, we did not find a statistically significant relationship between the time of exposure of the HHC and positive IgM anti-PGL-I (p > 0.05). The positive IgM anti-PGL-I in non-symptomatic HHCs suggests infection without the disease; follow-up of these HHCs is needed to determine if these HHCs eventually develop the disease.
Klatser et al [32] found M. leprae DNA in nasal swabs in 7.8% of 1228 samples from an endemic population. In this study of 113 HHCs, 16% showed a positive PCR. These results suggest that HHCs may act as hosts of M.leprae and therefore could be a source of infection and transmission. Thus, it is necessary to perform periodic clinical examinations and complementary exams to diagnose the disease early in high-risk populations.
Leprosy detection in two symptomatic HHCs of the same family whose index case was an MB patient confirms the transmission of leprosy between family members, which has been considered a main mode for the propagation of the infection in BI positive patients without treatment [33]. The index case corresponds to the father and the HHC to the son. The genotype of these two strains agreed at 9 markers; two markers did not amplify and one marker did not agree between the two strains (AC8a), which is highly polymorphic. Genotype comparisons will allow monitoring of the circulating strains in the region in general, and in the affected homes in particular. However, it is necessary to take samples from the index case prior to treatment for M.leprae to allow comparisons with the new isolates from family cases or contacts.
Only a small minority of the human population develop leprosy because M. leprae infection is unlike the universal susceptibility to other members of the Mycobacteriaceae family. It’s accepted that the majority of the humans are immune to leprosy through an as yet defined mechanism [33]. That a small minority of persons who succumb to the disease are diagnosed late leads to the acquisition of disabilities that alter their familial, social and occupational environment. An early diagnosis that includes the correct monitoring of the index case and HHCs would assure cutting the chains of transmission in both the family and the community.
Follow up of HHCs is a public health decision that can improve leprosy control. The presence of anti-PGL-I antibodies and M. leprae DNA in HHCs can suggest infection and source of infection and transmission of leprosy. The genotyping of M. leprae strains between family members allowed us to establish the source of transmission and make comparisons between the circulating M. leprae strains of a specific region.
Follow-up of HHCs using clinical exams to detect skin or peripheral nervous system symptoms of the disease, and the detection of infection using anti-PGL-I antibodies and M. leprae DNA immediately upon diagnosis of the index case may allow us to establish better methods to control the transmission of the infection.
|
10.1371/journal.pntd.0004481 | Identification and Analysis of the Role of Superoxide Dismutases Isoforms in the Pathogenesis of Paracoccidioides spp. | The ability of Paracoccidioides to defend itself against reactive oxygen species (ROS) produced by host effector cells is a prerequisite to survive. To counteract these radicals, Paracoccidioides expresses, among different antioxidant enzymes, superoxide dismutases (SODs). In this study, we identified six SODs isoforms encoded by the Paracoccidioides genome. We determined gene expression levels of representative isolates of the phylogenetic lineages of Paracoccidioides spp. (S1, PS2, PS3 and Pb01-like) using quantitative RT-PCR. Assays were carried out to analyze SOD gene expression of yeast cells, mycelia cells, the mycelia-to-yeast transition and the yeast-to-mycelia germination, as well as under treatment with oxidative agents and during interaction with phagocytic cells. We observed an increased expression of PbSOD1 and PbSOD3 during the transition process, exposure to oxidative agents and interaction with phagocytic cells, suggesting that these proteins could assist in combating the superoxide radicals generated during the host-pathogen interaction. Using PbSOD1 and PbSOD3 knockdown strains we showed these genes are involved in the response of the fungus against host effector cells, particularly the oxidative stress response, and in a mouse model of infection. Protein sequence analysis together with functional analysis of knockdown strains seem to suggest that PbSOD3 expression is linked with a pronounced extracellular activity while PbSOD1 seems more related to intracellular requirements of the fungus. Altogether, our data suggests that P. brasiliensis actively responds to the radicals generated endogenously during metabolism and counteracts the oxidative burst of immune cells by inducing the expression of SOD isoforms.
| Paracoccidioidomycosis is a health-threatening human systemic mycosis, endemic to some Latin America countries. The disease is caused by species belonging to the Paracoccidioides genus. Once inside the human host, Paracoccidioides must face the host innate immune system, escaping from the cytotoxic capacity of innate immune cells (ROS production and liberation of polypeptide antibiotics). To do so, they express and synthetize superoxide dismutases (SODs). We aimed to identify and characterize the SOD isoforms present in the Paracoccidioides genome. We identified six isoforms, among which we found an increased expression of PbSOD1 and PbSOD3 during the transition-to-yeast process, exposure to oxidative agents and interaction with phagocytic cells. Additionally, we found that PbSOD3 expression might be linked with a pronounced extracellular activity while PbSOD1 and the other isoforms seem more related to intracellular requirements of the fungus. We propose that the defence against endogenous-produced ROS may depend on intracellular Sods (mostly SOD1, but possibly also SOD2, SOD4 and SOD5), but defence against extracellular ROS (produced during host-pathogen interactions) might rely to a greater extent on SOD3, which is endowed with an extracellular activity.
| Dimorphic fungal pathogens are exposed to reactive oxygen species (ROS) from both internal and external sources. ROS include superoxide anion (O2.-), hydroxyl radical (.OH), and hydrogen peroxide (H2O2), among others. The superoxide anion radical is the first product of oxygen reduction. This radical is mediated by a variety of electron carriers [1] and it is considered the precursor of most other ROS [2]. Internally, ROS are mostly produced in fungal mitochondria as a by-product of aerobic cellular respiration [3]; externally, reactive oxygen and nitrogen species (ROS/RNS) can be produced by host cells during fungal infections [4]. The latter represent an important line of defense and one of the primary effector mechanisms of the host’s immune system aimed at controlling fungal infections [5,6]. At high concentrations, ROS can be extremely harmful at levels that exceed the defense mechanism of the fungal cell, generating an oxidative-stress state that can lead to oxidation of proteins, lipids and DNA, and ultimately to cell death [7]. However, at low concentrations ROS also function as critical second messengers in a variety of intracellular signaling and regulation pathways [8], and have also been correlated with life-span regulation and cell differentiation in microbial eukaryotes [9–12]. Understanding this dual role/effect of ROS is important when defining the components of the fungal antioxidant response and trying to understand the cellular and molecular changes required for adaption to these internal or external stimuli.
To counteract these reactive species, fungal pathogens, such as those belonging to the Paracoccidioides genus, are equipped with an antioxidant system that prevents ROS-damaging effects. This antioxidant system includes enzymes such as catalases, peroxidases and superoxide dismutases (SODs) [13]. Additionally, certain metabolic pathways are set in motion in order to supply reducing power, such as the pentose phosphate pathway and the thioredoxin and glutathione redox systems [1,13].
Thermally dimorphic fungi belonging to the genus Paracoccidioides are the etiological agents of paracoccidioidomycosis (PCM), a neglected health-threatening human systemic mycosis endemic to Latin America where up to ten million people appear to be infected. The fungus is thought to exist in nature in the mycelial form at environmental temperatures, while within the human host or at 37°C, it grows as the yeast form [14,15]. Paracoccidioides species belong to the largest group of dimorphic fungal pathogens, which includes species from the genus Histoplasma, Blastomyces and Coccidioides in the order Onygenales. Within the Paracoccidioides genus, there are four well-characterized phylogenetic lineages, all of them capable of infecting humans and causing PCM: three P. brasiliensis lineages (S1, PS2 and PS3) and one P. lutzii lineage (Pb01-like) [16,17].
As other fungal pathogens, Paracoccidioides spp. are exposed to different aggressions once inside the human host. The higher temperature within the human host has been shown to induce an increase in fungal metabolism leading to a greater oxygen consumption and ROS production in fungi such as Cryptococcus neoformans, Saccharomyces cerevisiae and Schizosaccharomyces pombe [9,18,19]. In addition, Paracoccidioides will also encounter host effector cells, which produce ROS during the oxidative burst in the phagolysosome through the activation of the NADPH-oxidase complex, generating superoxide radical as the first intermediate product [20]. Paracoccidioides yeast cells cope with these radicals within the phagocytes through the expression of proteins from the antioxidant system, such as SOD enzymes, to neutralize superoxide radicals and convert them into less damaging molecules, namely hydrogen peroxide and oxygen molecules [13]. SODs are metallo-proteins, which are classified on the basis of the metals located in their active sites (Fe, Mn, Ni and Cu/Zn) [21–23]. These proteins have been shown to contribute to the virulence of some pathogenic fungi, namely Candida albicans [24,25], C. neoformans [26], Aspergillus fumigatus [27], and H. capsulatum [20], all of which are capable, to a certain extent, of neutralizing the toxic levels of ROS generated by the host. However, the mechanism by which SOD contributes to the defensive mechanism of Paracoccidioides against exogenous and endogenous oxidative damage remains elusive.
In the present study, we sought to identify and characterize the SOD isoforms encoded by the Paracoccidioides’ genome, with the purpose of better understanding their function during the morpho-physiological transformation inherent to its dimorphic life cycle as well as during host-pathogen interactions. To pursue these goals, we first identified SOD proteins encoded by the Paracoccidioides spp. genome and determined gene expression of one representative isolate of each one of the phylogenetic lineages of Paracoccidioides spp. Assays were carried out to analyze SODs gene expression of yeast, mycelia cells and conidia undergoing the transition (mycelia-yeast and conidia-yeast) and the germination (yeast-mycelia and conidia-mycelia) processes, as well as under treatment with oxidative agents and during interaction with phagocytic cells (e.g. human PMNs and alveolar macrophages). Antisense RNA technology was also employed to further evaluate the role of specific isoforms upon exposure to oxidative stress-inducing agents, interaction of yeast cells with phagocytic cells and in a mouse model of infection.
In order to identify the SOD isoforms encoded be the Paracoccidioides genome and classify its SOD’s orthologs, we used bioinformatics tools based on similarity and orthology analyses. We used bidirectional BLAST analysis version 2.2.28+ with default parameters to identify sequence similarities [28]. Orthology analysis was performed using OrthoMCL version 1.4 with a Markov inflation index of 1.5 and a maximum e-value of 1e-5 [29].
We studied one strain representing each of the lineages of Paracoccidioides. Sequences of Pb18, Pb03 and Pb01 strains were chosen to represent S1, PS2 and Pb01-like lineages, respectively [30,31]. For the PS3 lineage, we used paired-end reads and reference assembly of the Pb60855 strain (ATCC60855; PbWT60855; ongoing genomic project). For all the identified SOD isoforms, protein domain conservation analyses were done using InterProScan [32], by sequence comparison with InterPro collection of protein signature databases in the EMBL-EBI (http://www.ebi.ac.uk/interpro/). Multiple sequence alignments were constructed using Muscle [33], and phylogenetic trees were constructed employing a distance computation method (Neighbor joining) [34]. Other features in the gene/protein sequences and annotations of the SOD isoforms, were identified using SignalP 4.1 server (http://www.cbs.dtu.dk/services/SignalP/) [35], TargetP 1.1 server (http://www.cbs.dtu.dk/services/TargetP/) [36] and PredGPI predictor (http://gpcr.biocomp.unibo.it/predgpi/pred.htm).
A representative isolate from each one of the phylogenetic lineages of Paracoccidioides spp. was used. P. brasiliensis knockdown strains used in this study were derived from the wild-type strain Pb60855. Strains and isolates used in this study are listed in Table 1. Paracoccidioides cells were maintained by sub-culturing in brain heart infusion (BHI) supplemented with 1% glucose (Beckton Dickinson and Company, Sparks, MD), under constant agitation at 36°C for the yeast form, and at 20°C for the mycelia form, unless otherwise indicated. P. brasiliensis conidia were produced and purified using the glass-wool filtration protocol, as described previously by Restrepo et al. [37]. In order to induce and evaluate the transition processes (conidia to yeast (C-Y); mycelia to yeast (M-Y)) and the germination process (yeast to mycelia (Y-M)), cells were incubated at 36°C or at 20°C, respectively, under constant agitation in a flask containing BHI [38,39]. During the morphological switch several fungal morphotypes coexist. The different stages of the dimorphic transition are fully established (for more information see Nunes et. al, 2005 and Hernández et. al 2011). Cultures during M-Y transition are characterized by the presence of hyphae, differentiating hyphae (chlamydospore-like cells), transforming yeast (production of multiple buds by the chlamydospore) and mature, multibudding yeast [40]. During C-Y transition, cultures are characterized by the presence of conidia, intermediate cells and yeast cells. After 12 h, intermediate cells are present, from 48 h onwards it is possible to observe the characteristic yeast cells, although, they are more abundant at 72 h. Regarding to the C-M germination, conidia began to germinate approximately at 24 h and the formation of branched mycelia occurs after 96 h [41,42]. Samples were collected for RNA extraction and quantification of gene expression analyses during evaluated time points.
Samples were collected for RNA extraction and quantification of gene expression analyses during evaluated time points.
Agrobacterium tumefaciens strain LBA1100 [43] was used as the recipient for the binary vectors constructed in this study (Table 1). Bacterial cells were maintained at 28°C in Luria–Bertani (LB) medium containing kanamycin (100 mg/ml). Escherichia coli DH5α was grown at 37°C in LB medium supplemented with appropriate antibiotics and used as the host for plasmid amplification and cloning.
Total RNA was obtained from Paracoccidioides cells using the Trizol reagent (Invitrogen). Total RNA was treated with DNase I (Thermo Scientific) and tested using a conventional PCR with β-tubulin primers to confirm the absence of chromosomal DNA contamination. cDNA was synthesized using 2 μg of total RNA with Maxima First Strand cDNA synthesis kit for RT-qPCR, according to the manufacturer’s instructions (Fermentas).
Real-time PCR was carried out using a Maxima SYBR Green/Fluorescein qPCR Master Mix (2X; qRT-PCR) kit with SYBR green, according to the manufacturer’s instructions (Fermentas). The CFX96 real time PCR detection system (Bio-Rad, Hercules, CA) was used to measure gene expression level of SOD isoforms present in the four phylogenetic lineages (S1, PS2, PS3 and Pb01-like) encoded into the Paracoccidioides’ genome. Primers were designed in order to anneal properly to each SOD transcript of the four phylogenetic lineages (S1 Table). Additionally, in Pb60855, SOD isoforms were evaluated in cells undergoing the transition (C-Y and M-Y) and germination processes (C-M and Y-M). We also evaluated gene expression in knockdown strains for PbSOD1 and PbSOD3 genes, and in P. brasiliensis cells carrying the empty binary vector as a control (PbEV60855). Melting curve analysis was performed after the amplification phase to eliminate the possibility of non-specific amplification or primer-dimer formation. β-tubulin gene (housekeeping gene) was used in order to normalize the expression value of each SOD isoform. Each experiment was done in triplicate, and the expression level was measured three times. We also compared the elongation factor 3 (TEF3, PABG_05066) as normalizer for the expression experiments. We saw no differences by using TUBE3 or TEF3 as normalizers. Accordingly, the calibrator gene used along the expression experiments was the TUBE3 gene (S1 Fig).
Yeast cell samples of Pb60855 were collected at 48 hours of growth for protein extraction. Briefly, after a 3000 rpm centrifugation during 15 minutes, the supernatant was collected and subsequently concentrated through a dialysis membrane (Spectra/Por 1, SPECTRUM. MWCO 6–8,000). Cells were washed twice in PBS, then resuspended in 2 ml of 0.05 M Tris-HCl pH 8.5, cocktail protease inhibitors and disrupted using glass beads (0.5 mm diameter) through vortexing. The cell wall fraction was removed by low speed centrifugation (3000 g, 5 min at 4°C) [27], and incubated in a 50 mM NaOAc and 5 mM NaN3 pH 5.6 solution at 36°C under agitation for 12 h. Protein contents of the extracellular, cell wall and intracellular (cytoplasmic) fractions were quantified with the Bradford reagent (Bio-Rad) [44], using bovine serum albumin as the standard and stored at -20°C.
To detect Sod enzymatic activity, protein samples were separated by 9% native acrylamide gel electrophoresis (1D-PAGE; running buffer: 0.025 Tris-HCl, 0.192 M Glycin, pH 8.3), without SDS to keep the proteins activity. Electrophoresis was carried out at 120 V at 4°C. Sod enzymatic activity was visualized as the inhibition of the reduction of nitro blue tetrazolium (NBT; Sigma) according to the method of Beauchamp and Fridovich (1971) [45]. Here, two reactions occur, the first one is the autoxidation from riboflavin and the second one is the riboflavin/NBT reduction, using NBT as chromogenic substrate. The Sod activity is determined as an achromatic zone, since the enzyme inhibits NBT reduction [45]. Following electrophoresis, the gel is washed twice during 10 min on ice-cold water, soaked in 2.45 mM NBT solution for 20 min in the darkness. This was followed by a further incubation in 0.028 mM riboflavin and 0.028 M tetramethylethylenediamine (TEMED) in PBS 1X, pH 7.8, for 15 min in the darkness. Upon illumination, an achromatic band indicating zones of activity appear in the region of gel where Sod proteins are present. For the 2D-PAGE, we followed the method described by Niyomploy et al. [46], with minor modifications. Briefly, cytoplasmic crude extracts were loaded onto 7-cm, pH 3–10 IPG gel strips (Bio-RAD), and left overnight at room temperature for the rehydration process. The isoelectrofocusing [47] was performed as described by Niyomploy et al.; thereafter, strips were rinsed in running buffer, equilibrated in equilibration buffer (0.05 M Tris-HCl pH6.8, 30% glycerol) for 10 min at room temperature, following by a rinse with running buffer, and incubation with equilibration buffer containing 2.5% iodoacetamide (IAA). Rinsed once again with running buffer and proceeding with the second dimension and zymogram as described above.
For these assays, we employed the human lung epithelial cell line A549 (ECACC No. 86012804), corresponding to type II epithelial cells from an adenocarcinoma cell line, which was obtained from the European Collection of Cell Cultures (ECACC). Cells were grown in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS).
We also employed mouse alveolar macrophages transformed with SV40 (MH-S cell line), obtained from the European Cell Cultures Collection (ECACC No. 95090612). Cells were grown in RPMI 1640 medium plus 2mM glutamine (Invitrogen) + 0.05 mM 2-mercaptoethanol (Sigma Aldrich, USA) + 10% fetal bovine serum (Invitrogen).
Polymorphonuclear neutrophils (PMNs) were isolated from human blood samples taken from healthy volunteers. We used whole blood treated with anticoagulant EDTA. Briefly, a layer 5.0 ml of non-coagulated whole blood was placed over 5.0 ml of Polymorphprep in a 12 ml centrifuge tube. Centrifuge samples for 450 x g for 30 min. The polymorphonuclear fraction was then washed with Hanks' Balanced Salt Solution and centrifuged for 400 x g for 10 min. Finally, PMNs were resuspended in Dulbecco's Modified Eagle Medium (DMEM; Gibco), enumerated in hemacytometer and cell viability was determined using trypan blue [48]. PMNs were seeded into 24-well tissue culture plate and allowed to adhere for 20 min at 36°C with 5% of CO2. For inhibition of NADPH-oxidase, 10μM diphenylene iodinium (DPI; D2926, Sigma) was added to PMNs 20 min before infection.
For all assays, we used a ratio of 1:5 for P. brasiliensis yeasts: host cells. The interaction was kept at 36°C with 5% of CO2, during 3h. After these interactions, SODs gene expression and colony forming units were determined to establish the percentage of viability [49,50].
DNA from Pb60855 was extracted from yeast cells cultures during exponential growth. We employed a Platinum high-fidelity Taq DNA polymerase (Invitrogen) to amplify aRNA oligonucleotides, designed on the sequences identified as PADG_07418 for PbSOD1 gene [51] and PADG_02842 for PbSOD3 gene. Gene sequences of the strain Pb03 were obtained from the Paracoccidioides genome database [30,31].
P. brasiliensis plasmid construction for aRNA and Agrobacterium tumefaciens-mediated transformation (ATMT) were performed as previously described [52]. Briefly, the amplified PbSOD1 and PbSOD3 aRNA oligonucleotides were independently inserted into the pCR35 plasmid under the control of the Calcium Binding Protein 1 (CBP-1) promoter region from H. capsulatum [53]. The pUR5750 plasmid was used as a parental binary vector to harbor this aRNA cassette within the transfer DNA (T-DNA). Constructed binary vectors were introduced into A. tumefaciens LBA1100 ultra competent cells by electroporation as described previously [54], and isolated by kanamycin selection (100 mg/ml).
In P. brasiliensis yeast cells (Pb60855), ATMT was done using A. tumefaciens cells harboring the desired binary vector, as described previously by Almeida et al. (2007) in order to obtain the knockdown strains. A 1:10 P. brasiliensis/A. tumefaciens ratio was employed during the 3 days period of co-culture at 28°C. Selection of P. brasiliensis transformants was performed in BHI solid media containing hygromycin B (Hyg; 200mg/ml) over a 15 days incubation period at 36°C. Randomly selected Hyg resistant transformants were tested for mitotic stability. This was determined by analyzing the stability of hygromicin B resistance. PbSOD1, PbSOD3-aRNA and PbEV strains were successively sub-cultured on BHI without hygromicin B (three consecutive rounds). Later on, we put them again under the selective pressure of the hygromicin B and analyzed them via amplification of the HPH cassette, verifying in this way the presence of the T-DNA. We analyzed more than one PbSOD1-aRNA and PbSOD3-aRNA strains, although results presented throughout the work refer to one selected aRNA strain per gene, and the same phenotypic characteristics were observed. P. brasiliensis yeast cells were also transformed with the empty vector pUR5750 (PbEV) as a control. In order to confirm the presence of the hygromycin B resistance cassette (HPH), PCR analysis was carried out to detect an HPH 1000-bp amplification product in PbEV, PbSOD1- and PbSOD3-aRNA strains.
Growth curves were performed in BHI broth (100 mL). We adjusted the inoculum to an OD of 0.4 for PbWT, PbEV and PbSOD1- PbSOD3-aRNA yeast cells. Then, cultures were incubated at 36°C and samples were collected at specific time points to determine growth curves by spectrophotometric analysis (OD600 nm; SmartSpec Plus (Bio-Rad, Hercules, CA).
Vitality was evaluated following the protocol reported by Hernandez et al. [55]. This corresponds to the ability of yeast cells to metabolize glucose upon late activation of a cell membrane proton pump and subsequent acidification of the medium due to H+ release [56]. Briefly, P. brasiliensis yeast cells were cultivated in BHI liquid medium and collected at exponential phase growth (72 h). Then, washed twice with sterile water pH 7.0. A 3 ml bottom was suspended in a final volume of 8 ml of water (pH 7.0) to obtain the yeast concentrated solution (YCS). Two milliliters of YCS were added to a beaker containing 38 ml of water pH 7.0. When the pH became stable (pH 5.5 to 6), 10 ml of 20% glucose were added. The pH was recorded every three min for 30 min, in order to evaluate changes in the pH of the medium. Also, as a negative control for the assay, PbWT yeast cells previously treated with 16 μg/ml of amphotericin B during 4h (Fungizone, Bristol-Myers Squibb Pharmaceuticals, England) were used [57]. The assays were performed in triplicates.
For the phenotypic analysis, we tested the sensitivity of the PbSOD1 and PbSOD3 knockdown strains to conditions inducing oxidative stress. The sensitivity to exogenous ROS of SOD mutants was analyzed using hydrogen peroxide, xanthine oxidase and menadione. Hydrogen peroxide sensitivity was tested using H2O2-saturated filter disks. 1×105 P. brasiliensis yeast cells were spread on BHI plates. After inoculation, sterile filter disks were loaded with 20 μl of PBS1X as a control, and 0.5, 1, 2, 4 and 8 M of H2O2 (02194057, MP Biomedicals). Plates were incubated at 36°C with 5% CO2 and after eight days the area lacking P. brasiliensis growth was measured. For xanthine oxidase induced-oxidative stress, cells were incubated in 50 mM Tris pH 8, 100 μM hypoxanthine (H9636, Sigma) and 5 mU/ml xanthine oxidase (X4500, Sigma). Yeasts were incubated for 4 h at 36°C with shaking [20]. After this time, we plated serial dilutions of the experiment on Kurita’s medium in order to determine viable CFUs. Finally, menadione (M5625, Sigma) sensitiveness was evaluated in a 96-well plate. 1×104 P. brasiliensis yeast cells were inoculated in each well, containing 0.5, 5, 10, 20, 40, 80 and 160 μM menadione. Plates were incubated at 36°C during eight days, after incubation time, the plate results were recorded.
Isogenic 6 to 8-week-old BALB/c male mice, obtained from the breeding colony of the Corporación para Investigaciones Biológicas (CIB), Medellín, Colombia, were used in assays and were kept with food and water ad libitum [58]. All animals were handled according to the national (Law 84 of 1989, Res No. 8430 of 1993) and international (Council of European Communities and Canadian Council of Animal Care, 1998) guidelines for animal research. The CIB research ethics committee approved the experimental protocols.
P. brasiliensis yeast cells were collected from exponentially growing batch cultures in BHI medium and counted using a hemacytometer. Animals (n = 5 per isolate) were infected with P. brasiliensis yeast by intranasal delivery of 1.5×106 cells suspended in PBS buffer from PbWT, PbEV or PbSOD1- PbSOD3-aRNA strains. Mice were monitored daily for survival, weight loss and symptoms of disease. At 12 days post-infection, mice were euthanized and lung, liver and spleen tissues were homogenized in 2 mL PBS. Homogeneous suspensions were diluted (1:10, 1:100 and 1:1000) and 0.1 ml of each dilution was plated on Kurita’s medium [49] in order to determine the fungal burden in each organ. Plates were incubated at 36°C, 5% CO2. CFU counts were assessed ten days after cultivation. The data was transformed into Log10 CFU/g of tissue.
Data are either the means or representative results of at least three similar repetitions, each performed in triplicate. Statistical analysis and comparisons were performed using paired Student’s t tests.
By means of sequence similarity and orthology analyses (BLAST and OrthoMCL, respectively), and using documented superoxide dismutase [59] sequences of reference fungal genomes, such as those of Candida albicans, Aspergillus fumigatus and H. capsulatum, we identified six conserved SOD homologs in the reference genomes of Paracoccidioides spp. (Pb18, Pb03 and Pb01) [30,31] as well as in Pb60855 (representing the PS3 lineage). Table 2 describes the Sod proteins encoded by the genome of Paracoccidioides spp. and the corresponding orthologs for each PbSOD in the Pb18, Pb03 and Pb01 genomes.
This set of homologs encompasses the well-studied putative Sods, which include Sod1 (PbSOD1; PADG_07418; cytosolic), Sod2 (PbSOD2; PADG_01755; mitochondrial) and Sod3 (PbSOD3; PADG_02842; extracellular). Sequence analyses revealed three additional SOD homologs in Paracoccidioides (PADG_01400, PADG_01954, PADG_01263), which encoded putative Sods. These new genes were designated as PbSOD4, PbSOD5 and PbSOD6 respectively. PbSOD4 shared sequence homology with the copper/zinc-dependent superoxide dismutases PbSOD1 and PbSOD3 genes. PbSOD5 and PbSOD6 shared sequence homology with the iron/manganese-dependent superoxide dismutase PbSOD2 gene. Fig 1A shows the relationship between SOD homologs encoded by the genome of Paracoccidioides spp.
Protein family domain analysis using InterProScan showed that three of the Sod proteins (PbSOD2, PbSOD5 and PbSOD6) had the Fe/Mn protein domains (PF00081: alpha N-terminal domain; PF02777: alpha/beta C-terminal domain). Unlike PbSOD2 and PbSOD5, the PbSOD6 protein sequence only had the Fe/Mn C-terminal domain divided into two intervals by 53 amino acids. Protein sequences of both PbSOD2 and PbSOD5 have both the Fe/Mn N-terminal and C-terminal domains conserving the Fe/Mn Sods putative structure (Fig 1A). In A. fumigatus, the PbSOD6 homologous is a predicted mitochondrial protein that has an essential function to promote survival of the fungus, and as in the case of Paracoccidioides, it also lacks the N-terminus consensus sequence [27]. The three remaining identified Sod proteins (PbSOD1, PbSOD3 and PbSOD4) have the Cu/Zn protein domain (PF00080; Cu/Zn superoxide dismutase; SODC). In addition to the Cu/Zn domain, the PbSOD4 has a heavy metal-associated domain (HMA; PF00403) that is related to a metal iron transporter. In the case of PbSOD3 we found a N-terminal signal peptide sequence (from 1 to 20 aa) using SignalP 4.0 [35], as long as an extracellular location pattern that was predicted using TargetP 1.1 [36]. There is also a C-terminal GPI-anchor/transmembrane signal (210–226 aa; Fig 1A) suggesting that this protein is associated with the cell surface. A previous study showed that SOD3 of the dimorphic fungus H. capsulatum is part of the extracellular proteins produced by the pathogenic yeast phase and its localization allows HcSOD3 to protect yeast cells specifically from exogenous superoxide [20]. Both results suggest that PbSod3p, as well as the HcSod3p are secreted proteins. Unlike PbSOD3, PbSOD2 and PbSOD5 have both a putative mitochondrial targeting signal as reported for other Mn/Fe SOD2 [25,27]. Also, PbSOD1, PbSOD4 and PbSOD6 lack secreted-targeting sequences or mitochondrial targeting signals suggesting that these proteins are likely to be cytosolic. The identified SOD isoforms, their protein sequence domains and functional annotations are well conserved in the genomes of Pb18, Pb03, Pb60855 and Pb01, indicating that these proteins are part of the core genes of the Paracoccidioides genera.
We analyzed the gene expression level of each isoform in one representative isolate of the phylogenetic lineages of Paracoccidioides spp. (S1, PS2, PS3, and P. lutzii), in both mycelial and yeast cells during their exponential growth phase in batch culture. We observed that in PS3 isolates (Pb60855 and PbBAC) PbSOD3 was the gene with the highest expression level in both phases, albeit higher in yeast cells (Fig 1B). PbSOD1 was predominant in the mycelia phase in the S1, PS2 and P. lutzii strains. In Pb01, the expression level of all isoforms was lower when compared to that of P. brasiliensis isolates. Additionally, using a Sod zymogram we aimed at detecting if these isoforms were active and functional. First, we used a non-denaturing polyacrylamide gel (1D-PAGE) with crude extracts of P. brasiliensis, ATCC 60855: the cytoplasmic, the cell wall and extracellular fractions. The detection was performed following the method created by Beauchamp and Fridovich (1971). Despite all efforts, we only observed a minimum of two bands, which could be explained in part because the 1D-PAGE does not discriminate between isoforms with different molecular weight or isoelectric point (S2 Fig, top). To solve this, and in order to obtain a better resolution, we decided to carry out a 2D-PAGE (S2 Fig, bottom). Due to the characteristics of the technique, it was possible to observe three spots indicating the activity of a Sod isoform, without however specifically determining which achromatic zone corresponded to which isoform.
P. brasiliensis cells respond to conditions such as i) variations in temperature (which induces morphological changes) and ii) continuous ROS production (a step required to carry out biological functions) with the induction of heat shock proteins [41,42,60] and enzymes that counteract ROS. As cell differentiation has been shown to be triggered by oxidative stress in different microbial eukaryotes, such as C. albicans, A. nidulans and Neurospora crassa, inter alia [61] and in order to better understand the possible role of SOD isoforms in P. brasiliensis, we analyzed the expression profile of one representative strain (P. brasiliensis ATCC 60855) in yeast and mycelia phases and during the morphological switch. P. brasiliensis conidia, mycelia and yeast cells were placed at either 36°C or 20°C in order to induce the transition (mycelia-yeast and conidia-yeast) and the germination (yeast-mycelia and conidia-mycelia) processes and a kinetic analysis of the expression of SOD isoforms was performed. In both the yeast and mycelia phases, PbSOD1 and PbSOD3 gene expression was higher during all evaluated time points, while PbSOD2, PbSOD4 and PbSOD5 were the less expressed genes (S3A Fig). PbSOD3 was significantly expressed at higher levels in the yeast phase when compared to mycelia. On the other hand, PbSOD1 was expressed at slightly higher levels in mycelia when compared to the yeast phase (S3A Fig). Regarding the morphological shift, PbSOD1 and PbSOD3 showed increased expression during the Y-M and M-Y morphological shifts (S3B Fig). PbSOD4 and PbSOD5 were the less expressed genes during all evaluated time points and conditions. During the Y-M germination and M-Y transition, PbSOD1 and PbSOD3 were the most expressed isoforms (S3B Fig). Overall, expression levels in experiments involving morphological shifts from conidia were low for all SOD isoforms with the exception of during the C-M germination in which PbSOD2 gene expression increased throughout time, peaking at 96 hours, and the initial hours (3 and 12) of the C-Y transition in which PbSOD6 increased its expression.
We determined the expression of SOD isoforms during the interaction with pneumocytes (cell line A549), since these cells would be the first barrier faced by Paracoccidioides cells when trying to adhere to the host lung tissue and thus establishing the infection [55,62]. We found that the interaction of P. brasiliensis yeast cells with A549 cells induced a slight increase in PbSOD3 gene expression (Fig 2). The gene expression of the other isoforms remained unaltered when compared to the unchallenged control. Subsequently, we analyzed gene expression of SOD isoforms during the interaction with phagocytic cells [alveolar macrophages (cell line MH-S) and human PMNs] since after the initial establishment of the infectious process, Paracoccidioides cells must counteract the phagocytic arsenal elicited by macrophages and PMNs in order to establish itself within the host. During the interaction of P. brasiliensis yeast cells with human PMNs we observed an increase of PbSOD1 gene expression of almost 2-fold, when compared to the unchallenged control. However, upon interaction with human PMNs, PbSOD3 gene expression was much higher (almost 4-fold) when compared to unchallenged yeast cells (Fig 2). Gene expression of the other isoforms remained unaltered.
To further analyze the function of SOD1 and SOD3 in P. brasiliensis, we used an antisense-RNA (aRNA) knockdown strain for SOD1 (PbSOD1-aRNA), which was previously constructed [51] (S4 Fig) and a PbSOD3-aRNA knockdown strain constructed in this study using ATMT. Briefly, one region of the gene was selected in order to design two different aRNA oligonucleotides and generate knockdown strains. Both of them were directed at exon 1 (107-bp for PbSOD3AS1 and 104-bp for PbSOD3AS2; Fig 3A). We selected a mitotically stable isolate with the highest decrease in PbSOD3 gene expression, which ranged between 50 to 70% (strain PbSOD3-aRNA; Fig 3B). The insertion of the transfer DNA (tDNA) into P. brasiliensis’ genome was confirmed via amplification of the HPH cassette (S5 Fig).
We initially analyzed possible alterations during batch culture of the knockdown strains (PbSOD1-aRNA and PbSOD3-aRNA). No major changes were detected during growth curve analysis of either knockdown strain when compared to PbWT or PbEV. However, a decreased capacity to metabolize glucose in both PbSOD1-aRNA and PbSOD3-aRNA strains as compared to PbWT and PbEV yeast cells was observed, detected as an increase in the pH, (Fig 3C), indicating reduced yeast cell vitality of the knockdown strains.
To further evaluate the role of PbSod1p and PbSod3p, we analyzed the effect of oxidative stress-inducing agents (hydrogen peroxide, xanthine oxidase and menadione) on fungal growth of the knockdown strains. PbSOD1-aRNA and PbSOD3-aRNA yeast cells were exposed to hydrogen peroxide using a disk-diffusion assay with increasing concentrations of the compound (0.5, 1, 2, 4 and 8 M). Knockdown strains showed a higher sensitivity to H2O2 with the largest clearing areas (Fig 4A and 4B), although no major differences were detected between PbSOD1-aRNA and PbSOD3-aRNA.
Taking into account that the main substrate for SOD1 and SOD3 is the superoxide anion [63], we also induced oxidative stress using xanthine oxidase and menadione, compounds known to be superoxide-generating agents. Menadione generates ROS through redox cycling [64], while xanthine oxidase generates ROS catalyzing the oxidation of substrates, such as purines (xanthine and hypoxanthine), and a variety of electron acceptors such as O2 and NAD+, which react with the enzyme [65]. Both PbSOD1-aRNA and PbSOD3-aRNA yeast cells were less resistant to menadione when compared to the control (Fig 4C). Furthermore, we used xanthine oxidase 5 mU/ml and hypoxanthine 100 μM to generate superoxide in vitro. Interestingly, PbSOD3-aRNA yeast cells were more sensitive to this oxidative stress-inducing agent than the PbSOD1-aRNA and PbWT strains (Fig 4D).
To test the involvement of Sod1p/Sod3p in phagocyte-defense, human PMNs were incubated with both knockdown and wild-type yeast cells. As the fungicidal effect of PMNs on fungal cells depends mostly on the production of ROS, we also investigated the effect of the suppression of NADPH oxidase, required for the activation of the oxidative burst and subsequent generation of ROS in PMNs. In order to inhibit NADPH oxidase we treated human PMNs with DPI prior to challenging with yeast cells. PbEV had the same behavior as PbWT yeast cells. It is important to note that PbSOD3-aRNA yeast cells were more susceptible to PMN’s fungicidal ability than PbSOD1-aRNA. Furthermore, treatment with DPI considerably reduced PMN’s ability to kill PbSOD3-aRNA strain (from 25 to 50%; Fig 5A). Evaluation of PbSOD1 and PbSOD3 gene expression in P. brasiliensis yeast cells challenged with human PMNs was also performed. As shown previously, the interaction with PMNs led to increased expression of PbSOD1 and PbSOD3, both in the presence–to a greater extent–and absence of DPI (S6 Fig). PbSOD1-aRNA and PbSOD3-aRNA strains showed a similar behavior but with reduced expression levels most likely due to knockdown of the gene expression.
To further analyze the involvement of Sod1p and Sod3p in P. brasiliensis virulence, mice were infected intranasally with 1.5×106 PbSOD1-aRNA or PbSOD3-aRNA yeast cells. CFU in mouse lungs, kidneys and spleen were determined on the twelfth day after infection, which matches with the onset of the cell-mediated immune response [20]. We found active infection in lungs of PbWT, and PbEV, while in the PbSOD1-aRNA strain there was a significantly lower fungal burden. Moreover, regarding the PbSOD3-aRNA strain, no CFUs were detected (Fig 5B). CFUs were recovered from the liver of mice infected with PbWT and PbEV strains, but not from tissues of mice infected with the PbSOD1-aRNA or PbSOD3-aRNA strain. CFUs were not recovered from the spleen in any of the infected mice, irrespective of the strain used.
This is the first study on the SOD family, the largest antioxidant gene family identified so far in the Ajellomycetaceae, specifically in the Paracoccidioides genus. We report the existence of six SOD isoforms encoded by the Paracoccidioides’ genome. Despite our efforts and due to technical difficulties, we were only able to detect biochemical evidence for the presence of three isoforms (S2 Fig). Additionally, we could not identify to which isoform corresponded each achromatic zone, due to the as-of-date lack of methods to generate knockout strains in P. brasiliensis. Through quantitative RT-qPCR, we determined the profile expression of each isoform in different conditions. PbSOD2, PbSOD4, PbSOD5 and PbSOD6 did not show differential expression in any phase (yeast or mycelia), morphological shift or during interaction with host cells, while PbSOD1 and PbSOD3 were differentially expressed depending on the growth conditions and stimuli.
PbSOD3 gene expression was predominant in the yeast phase of the PS3 isolates, unlike S1, PS2 and P. lutzii isolates, where there was no such a differential expression of this isoform (Fig 1B). On the other hand, PbSOD1 and PbSOD3 were similarly expressed during the yeast phase, while in mycelia PbSOD1 expression was slightly higher for S1, PS2 and P. lutzii isolates. Importantly, differences in the metabolic profile among the members of the Paracoccidioides genus have been detected [66], similar to as shown in the Histoplasma genus [67], which could underlie distinct capabilities to survive phagocytic microbicidal attack set by the host. In the case of H. capsulatum, strain G217B is equipped with an improved antioxidant defense system while in G186A the evasion of phagocyte detection is critical for virulence [67]. Regarding Paracoccidioides spp., a recent study demonstrated that P. lutzii has a more active anaerobic metabolism than P. brasiliensis [66]; which correlates with the lower expression level of the studied SOD isoforms in P. lutzii. Since these cells have a reduced oxygen consumption they would most likely not have to synthesize such an elevated level of proteins related to oxidative stress defense as P. brasiliensis, particularly during host-pathogen interaction at the onset of the infection. Additionally, we should considerer that although both species and isolates from all phylogenetic lineages can infect humans and produce Paracoccidioidomycosis (PCM), they can also vary in virulence and induce different immune responses [68]. This issue is critical and needs to be elucidated in order to better understand the pathobiology of PCM and how it may relate to the different species and lineages of Paracoccidioides spp.
Regarding to the morphological switch we detected that during the M-Y transition and in the yeast phase, PbSOD3 was expressed at higher levels (S3A and S3B Fig). This suggests that the gene may display a phase-dependent expression and possibly be required for the defense against ROS produced endogenously (as a consequence of the high temperature and increased metabolism) and exogenously (as a consequence of the infectious process). Importantly, Paracoccidioides, H. capsulatum and Blastomyces SOD3 have a conserved predicted signal peptide and a glycophosphatidyl inositol (GPI) anchor (Fig 1A) [15, 28], which could allow the cell to excrete the protein and to associate with the cell surface, suggesting its role as an extracellular Sod. Overall, these data might suggest a role in protecting Paracoccidioides yeast cells against the conditions faced once within the human host (e.g. increase in temperature) and also when defending against phagocytes-induced ROS. To test our hypothesis, we used host cells in order to investigate which SODs were involved during the initial host-pathogen interactions. A549 and MH-S cells did not significantly trigger SOD induction, but human PMNs induced the expression of only two of the six SODs, PbSOD1 -to a lesser extent- and PbSOD3, the latter at higher levels (Fig 2). P. brasiliensis yeast cells posses some mechanisms in order to evade MH-S cells, including the production of melanin (in vivo and in vitro) [69], which can interfere directly with the binding of the fungi to phagocytic cells [70]. In addition to the melanin, P. brasiliensis also produce an extracellular antigen (Gp43) that is able to interact with host cells, inhibiting both the phagocytosis and the releasing of nitric oxide (NO) by macrophages [71,72]. Therefore, macrophages are not able to produce ROS in response to P. brasiliensis infection [71]. This could explain the low induction of SOD isoforms during interaction with both, A549 (pneumocytes) and MH-S cell lines.
Studies in P. brasiliensis have focused on studying the role of neutrophils during the initial stages of the infection and during the development of PCM. Neutrophils play a relevant role in host defense and the resistance mechanisms against PCM, exerting an immunoregulatory role in antibodies and cytokine secretion during the course of the disease [73]. These cells are most frequently associated with extracellular killing mechanisms that involve the release of large amounts of ROS and granule components in the extracellular medium [74,75]. This was demonstrated in vitro for P. brasiliensis. Human PMNs ingest yeast through phagocytosis but when these are too large they form an extracellular vacuole in order to kill both ingested and extracellular cells [76]. These facts together with the presence of high activities of NADPH oxidase during the phagocytosis of P. brasiliensis by neutrophils [76] are in line with our results regarding the induction of antioxidant enzymes SOD1 and SOD3 during PMN-P. brasiliensis interaction.
Accordingly, we generated the knockdown strains for either gene and observed that both knockdown strains had reduced vitalities during batch culture growth (Fig 3D). The high concentration of glucose used during the vitality test accelerates glycolysis in the cells, leading to ATP synthesis (necessary for metabolizing glucose) and the consequent production of ROS [77]. We attributed these low vitalities to the fact that PbWT could more readily counteract ROS produced as a consequence of the glucose metabolism, contrarily to the aRNA strains. The decreased cell vitality could be indicating metabolic alterations in PbSOD1-aRNA and PbSOD3-aRNA yeast cells, that could also affect the performance of the pathogen either during batch culture growth or during the host-pathogen interactions.
Furthermore, when we challenged knockdown strains with H2O2, menadione and xanthine oxidase, we found that both knockdown strains were both similarly more susceptible to H2O2- and menadione-induced oxidative stress than the wild-type strain (Fig 4A, 4B and 4C). Although H2O2 is not a Sod substrate, it has been previously established that treatment with this compound in Saccharomyces cerevisiae cells induces the transcription of SODs [11,47], implying in some way an indirect involvement of these enzymes in defense against H2O2. We also measured the gene expression of SODs after the interaction with this compound, and found that both PbSOD1 and PbSOD3 genes were induced (S7 Fig). Additionally, H2O2 is able to inactivate SOD activity, in a concentration and time-dependent fashion [78]. It has also been proven that cytosolic, extracellular and mitochondrial SODs have peroxidase properties [79,80], which would enable the enzyme to directly interact with its product H2O2. In order to prove this in P. brasiliensis cells, we determined the ability of the wild-type strain and the aRNA strains to eliminate H2O2 in vitro. In agreement with the evidence that suggests that SOD3 acts at an extracellular level, PbSOD3-aRNA strain had a decreased ability to destroy H2O2 in the extracellular and cell wall fractions. PbSOD1-aRNA strain had a decreased ability to degrade H2O2 in the intracellular fraction. PbWT and PbEV strains had similar behaviors in all evaluated cellular extracts, with a higher capability to destroy this compound when compared to aRNA strains (S8 Fig). These data may further support the involvement of PbSOD1 and PbSOD3 in defending P. brasiliensis cells not only against superoxide anions but also against peroxide-induced oxidative stress.
Menadione induces endogenous oxidative stress generating superoxide anions through redox cycling. Notably, we found that PbSOD1-aRNA and oddly PbSOD3-aRNA were similarly susceptible to this compound. In this respect, it was demonstrated for S. cerevisiae cells that superoxide anions generated by menadione also acts and are formed at the outside of the cell, and consequently, the addition of SOD into the incubation buffer (acting as an extracellular enzyme) protected cells from cytotoxic effects of the compound [81]. This fact attests to the relevance of the extracellular enzyme in defending the cell against superoxide anions generated by menadione. As the anions generated extracellularly do not readily diffuse across the plasma membrane, the toxic effects might also occur in the outside of the cell; which is aligned with our finding of PbSOD3-aRNA strain being as sensitive as PbSOD1-aRNA to menadione-induced oxidative stress, supporting that both, extracellular and intracellular enzymes are required for defending yeast cells against menadione-induced oxidative stress. Accordingly, we observed a higher susceptibility of PbSOD3-aRNA yeast cells to xanthine oxidase-induced oxidative stress than in PbSOD1-aRNA (Fig 4D). As in the menadione-induced oxidative stress, the superoxide anions generated by xanthine oxidase and the lack of diffusion of them across membranes may maintain higher levels of ROS outside of the fungal cell, and consequently Paracoccidioides yeast cells should require SOD enzymes capable of acting in an extracellular level in order to counteract the detrimental effects of this ROS. This result may be related to the fact that Sod3p most likely counteracts exogenous superoxide and that in the PbSOD3-aRNA strain this activity is partially lost, presenting increased susceptibility to xanthine oxidase-induced oxidative stress. Moreover, despite the fact that Paracoccidioides cells have and express other Sods (Sod1, Sod2, Sod4, Sod5, and Sod6) it is possible that Sod3 is potent enough to counteract exogenously-produced ROS under the studied conditions. In summary, defense against endogenous-produced ROS may depend on intracellular Sods (mostly Sod1p, but also could be involved Sod2p, Sod4p, Sod5p and Sod6p), but defense against extracellular-produced ROS (produced during host-pathogen interactions) might rely mostly on Sod3p [20,25,27]. Nonetheless, this needs to be further elucidated.
We employed yeast co-cultured with human PMN and a murine model of infection to study the involvement of PbSOD1 and PbSOD3 isoforms in Paracoccidioides virulence. We observed that P. brasiliensis cells are highly resistant to the action of phagocytes and that the PbSOD3-aRNA strain was significantly more susceptible to the action of PMNs than PbSOD1-aRNA and the control strains (PbWT and PbEV; Fig 5A). Later on, we confirmed that the increased susceptibility of PbSOD3-aRNA strain to PMNs accounted for the oxidative mechanism and not for the polypeptide antibiotics delivered from lysosomal granules [82] using DPI to specifically inhibit NADPH-oxidase. These results showed a reduction in the killing by PMNs in PbSOD3-aRNA, likely related to the inhibition of the oxidative burst in PMNs and the reduction in the expression of the extracellular Sod (PbSOD3; Figs 5A and S6). Furthermore, results using a mouse model of infection indicated that PbSOD3 seems to be more relevant in the establishment and development of the PCM, since yeast cells with decreased PbSOD3 gene expression were unable to infect lung tissues. Meanwhile, the PbSOD1-aRNA strain was able to infect the lungs, but unable to disseminate to other organs (Fig 5B). In agreement with this, H. capsulatum knockout cells for SOD3 showed that this gene is required for full virulence in vivo, while the absence of the SOD1 gene did not impair lung infection [20]. In addition, in Blastomyces it has been also shown the up-regulation of SOD3 during the interaction with macrophages and in a mouse model of infection [83]. Thus, our results may suggest that Sod3p may play an important role in P. brasiliensis virulence, either the establishment of the infectious process or dissemination, while Sod1p, although not essential for establishing the respiratory disease, still might be required for fungal dissemination.
Based on our results, we postulate that P. brasiliensis actively responds to the radicals generated endogenously during metabolism and counteracts the oxidative burst of immune cells by inducing the expression of SOD isoforms. In the former case, they could specifically induce PbSOD3 gene expression endowed with an extracellular activity and therefore might be considered an essential gene during the events underlying the host-pathogen interaction.
|
10.1371/journal.ppat.1007212 | Pulmonary influenza A virus infection leads to suppression of the innate immune response to dermal injury | The innate immune system is responsible for many important functions in the body including responding to infection, clearing cancerous cells, healing wounds, and removing foreign substances. Although many of these functions happen simultaneously in life, most laboratory studies of the innate immune response focus on one activity. How the innate immune system responds to concurrent insults in different parts of the body is not well understood. This study explores the impact of a lung infection on the cutaneous wound healing process. We used two complimentary models of injury: the excisional tail wound and subcutaneous implantation of polyvinyl alcohol (PVA) sponges. These models allow for assessment of the rate of closure and measurement of cellular and cytokine responses during acute wound healing, respectively. When mice with these healing wounds were infected with influenza A virus (IAV) in the lung there was a delay in wound healing. The viral lung infection suppressed the innate immune response in a healing wound, including cellular infiltrate, chemokines, growth factors, and cytokines. However, there was not a global immune suppression as there was an increase in inflammation systemically in mice with both infection and healing wounds compared to mice with only wounds or IAV infection. In addition, the lung immune response was largely unaffected indicating that responding to a lung infection is prioritized over a healing wound. This study introduces the concept of immune triage, in that when faced with multiple insults the immune system prioritizes responses. This paradigm likely applies to many situations that involve the innate immune system, and understanding how the innate immune system handles multiple insults is essential to understanding how it can efficiently clear pathogens while responding to other inflammatory events.
| In a natural setting, the innate immune system is frequently faced with multiple insults, against which it must mount overlapping inflammatory responses. We are interested in how the innate immune system deals with multiple, simultaneously occurring inflammatory insults, and if the response to one will take priority. For example, the innate immune system is essential in mediating both the early control of pathogen replication in infected tissue and in the early stages of wound healing. Pulmonary infections occur frequently in injured patient populations; therefore, we set out to determine the impact of a respiratory infection on a healing wound. To examine this, mice with healing dermal wounds were infected with influenza A virus (IAV), a common cause of viral pneumonia. We found that the innate immune response to the lung infection took priority at the expense of the healing wound, in that the initial control of viral replication in the lung was not impacted, while the wound healing response was suppressed. Very little work has been done examining how the immune response can respond to overlapping inflammatory insults. Our work shows that not all immune responses are created equal, and that the cells of the innate immune system are preferentially routed towards fighting a lung infection rather than the healing dermal wound. This apparent prioritization of the innate immune response opens up a new direction of study. It is relevant to many fields where competing insults may alter the disease state.
| The immune system plays roles in multiple processes in the body including the response to infection, tissue repair, development, cancer immunosurveillance, and maintenance of homeostasis [1–4]. Dysregulation of these processes can lead to a disease state [5–8]. Most studies of the immune response focus on just one of the functions of the immune system. In reality, however, the immune system, especially the innate immune system, must be able to respond to multiple insults at the same time. Understanding how the innate immune system is equipped to handle simultaneous and disparate inflammatory events will provide greater insight into the complexity of the immune response. While some progress has been made in understanding how the innate immune system is altered when faced with multiple infections [9–16], how the onset of an infection alters the immune response to an ongoing non-infectious insult, such as a cutaneous wound, has not been well explored.
Wound healing is an essential process in human health, and its proper progression is critical to a full recovery from trauma and surgery. Innate immunity plays an important role in initiating wound repair. Neutrophils and monocytes infiltrate the wound from the periphery based on signaling from chemokines, cytokines, and interactions with adhesion molecules on the activated endothelium [2,17–22]. The earliest phase of cutaneous wound repair is characterized by an inflammatory cytokine milieu, including TNF-α, IL-6, IL-1α, IL-1β, and IFN-γ, as well as leukocyte-attracting chemokines such as CXCL1, CXCL5, CXCL10, CCL2, CCL3, CCL4, and CCL5 [18,19,23–28]. Neutrophils and inflammatory monocytes are rapidly attracted to wounds following the release of DAMPs and chemokines. [2,18–20,24,28]. Together, these earliest-responding cells aid in the clearance of damaged tissue and coordinate downstream responses required for tissue repair, and disruption of this acute phase negatively impacts downstream healing. Once in the wound, monocytes differentiate into repair macrophages; as a consequence, impairment of monocyte trafficking to wounds through loss of CCL2 or CCL3 signaling disrupts angiogenic and fibrotic responses required for proper wound healing [3, 29–31]. Similarly, blockade of neutrophil trafficking after myocardial infarction has been shown to impair cardiac repair [32]. Efferocytic clearance of apoptotic neutrophils by wound macrophages dampens inflammatory responses and promotes the transition to a repair phenotype, and loss of this interaction can prolong the inflammatory phase of wound healing [23, 32–34].
In addition to the complications that can arise from impaired wound healing, many co-morbidities, including infection of the wound, metabolic disorders, increased age and other environmental and genetic factors, can delay wound healing [8,17,21, 35–37]. There is also evidence that systemic disorders, such as sepsis, suppress wound healing [38,39]; however, sepsis is an extreme condition that causes widespread changes in physiology including decreased oxygenation and altered immune responses [38]. One area that is understudied is the effect of a distal infection, such as pneumonia, on the dermal wound healing process.
Pulmonary infections frequently occur after injury or trauma, and patients with infections have longer hospital stays and increased morbidity and mortality [40,41]. A recent report provides a link between the influenza season and increased risk of complications after surgery [42]. While it is generally thought that trauma impacts the ability to respond to a pulmonary infection [43], there are fewer studies that investigate the impact that lung infections, in particular viral infections, have on the wound healing response. Those that do exist mainly focus on viral infections of the wound site itself [44,45]. It remains an open question how a viral lung infection that is contained in one region of the body impacts the host’s ability to respond to tissue damage in a distinct region.
The acute cellular and cytokine responses to pulmonary influenza infection and dermal wounds share many common features. Recruitment and activation of innate immune cells such as neutrophils and monocytes by DAMPs, PAMPs, cytokines, and chemokines, are essential for the early control of pulmonary viral infection [21, 22, 27, 46]. Depletion of neutrophils and monocytes during influenza infection is linked to increased viral replication, although the accumulation of inflammatory monocytes after infection has also been shown to contribute to lung injury [27, 28, 46–50]. Given the overlap in innate immune cells and factors that are important in both wound healing and the early control of viral infection, we hypothesized that initiation of a pulmonary viral infection would disrupt the innate immune response at a distal cutaneous wound.
To explore the interconnectedness of the innate immune response when simultaneously responding to infection and injury, we used a viral lung infection model combined with two complementary murine wounding models. The first model is an excisional skin wound to the tail that allows for assessment of the rate of wound closure. The second model, subcutaneous implantation of polyvinyl alcohol (PVA) sponges, allows for the measurement of cellular and cytokine responses during acute wound healing. Wounded mice were subsequently infected with influenza A virus (IAV), which remains confined to the lung and does not spread systemically. Excisional skin wounds experienced delayed closure in mice that also had an IAV pulmonary infection. Examination of acute healing responses using the sponge implantation model revealed that wounds from mice with a pulmonary IAV infection had reduced cellularity and cytokine concentrations and a resulting decrease in the reparative growth factor, VEGF-A. These data demonstrate that the presence of a pulmonary infection disrupts the inflammatory phase of wound healing by altering the innate immune responses that drive the initial stages of repair. This impairment was not a result of systemic immunosuppression, as circulating cytokine and cellular responses remained intact. Investigating the interconnectedness of the early cellular and cytokine responses to these concurrent insults is paramount to a complete understanding of the immune response. This new area of study has exciting implications in the field of innate immunity, and has relevance for many disease models. How immune responses to infection, injury, development, or cancer influence each other and the ability to maintain a healthy organism is an important new area of future work.
For all experiments, mice were first wounded by tail skin excision or PVA sponge implantation, and infected with IAV 24 hours later. The treatments were timed in this way to mimic the onset of these inflammatory events in the clinical setting. Wounding by PVA sponge implantation causes measureable systemic inflammation and depletion of inflammatory monocytes from the circulation as they marginate to the wound within 24 hours [19]. IAV was administered at this time point to maximize the overlap of systemic responses from wounding and pulmonary infection.
To determine the effect of pulmonary IAV infection on the rate of wound healing, mice were subjected to excisional tail wounding and uninfected or infected with IAV 24 hours later (Fig 1A). In this model, a 1.0x0.3cm portion of skin is excised 0.5cm from the base of the tail. Murine tail skin wounds heal primarily by re-epithelization and provide a better model of human wound healing than dorsal skin punch biopsy wounds, which heal primarily by contraction [51]. The total wound area was measured every other day for a period of 14 days. The wounds from IAV-infected mice had delayed closure compared to those from control mice (Fig 1B and 1C). The initial healing rate was similar between the two groups, however by 7 days after wounding, the wounds from mice with IAV infection were significantly larger in area and healed at a slower rate from day 7 to day 11 (Fig 1B and 1C).
In order to examine changes in the wound healing response at the cellular level, we used the PVA sponge implantation model. In this model PVA sponges are surgically implanted subcutaneously along the mouse dorsum. This wound model follows the inflammatory, angiogenic, and fibrotic stages of acute wound healing within the first two weeks of sponge implantation [24,25,29]. It allows for the retrieval of wound fluids and infiltrating leukocytes from implanted sponges by mechanical disruption. The sequence of inflammatory and repair responses mimics cutaneous wound models up to two weeks post-implantation, after which the non-resolving fibrotic response models a sterile foreign body reaction [18,19,23,25,31,52,53]. During the first seven days after sponge implantation, innate leukocytes, in particular monocytes and neutrophils, dominate the cellular wound healing response [18,19,23]. Wound day 7 is an inflection point representing the transition from the early inflammatory wound healing phase to the later reparative stages. At this time, inflammatory cytokines such as TNF-α and IL-6 are contracting in the wound, and pro-angiogenic and pro-fibrotic factors such as VEGF and TGF-β, are emerging [24,28].
For these studies, wounds were analyzed up to seven days after sponge implantation to assess the inflammatory phase of repair [19]. Wounded mice were uninfected or infected intranasally with IAV 24 hours after PVA sponge implantation. Half of the implanted sponges were removed at indicated time points (2, 4, and 7 days after wounding), and wound cells were collected for analysis (Fig 1D). The other half of the implanted sponges were used for wound fluid collection for analysis of cytokine and chemokine content in the healing wound environment. This enables a complete analysis of the early stages of the wound healing response mediated by innate immune cells. The cellularity of PVA sponge wounds in uninfected mice increased steadily over the first 7 days post-implantation. In mice that were infected with IAV 24 hours after wounding, the wound cellularity measured at days 2 and 4 after wounding was similar to uninfected control mice (Fig 1E). However, while the number of wound cells in uninfected mice approximately doubled between days 4 and 7 post-wounding, the number of wound cells in mice with concurrent infection did not increase. Correlating with the decrease in wound infiltrate, there was a greater than fifty percent decrease in the pro-angiogenic growth factor VEGF-A in the day 7 PVA sponge wound fluid (Fig 1F). VEGF-A is made locally in the wound primarily by repair macrophages and fibroblasts, and is essential in coordinating wound angiogenesis. In the PVA sponge wound model, it is expressed during the reparative stages of healing and is detectable in the wound fluid beginning around day 7 [19,54].
To determine whether the decrease in wound cellularity observed in IAV-infected mice at day 7 after wounding was attributable to changes in specific wound cell infiltrates, wound innate immune cell populations were analyzed by flow cytometry. The complete gating strategy for wound cells is shown in S1 Fig. On a percentage basis, there was a decrease in Ly6G+Siglec-F−cells, which are predominately neutrophils, in wounds from mice that had pulmonary IAV infection (Fig 1G). There was also a decrease in the percentage of Ly6ChighF4/80+ monocytes and a relative increase in Ly6ClowF4/80+ macrophages in mice with pulmonary viral infection (Fig 1G). The percentage of Siglec-F+Ly6G– eosinophils also increased in mice with IAV infection (Fig 1G). However, when total cell numbers were calculated there was a significant decrease in all of these cell subsets at day 7 after wounding with the exception of eosinophils, which remained unchanged in number (Fig 1H). There was a larger decrease in the number of neutrophils and monocytes than in macrophages. Neutrophils and inflammatory monocytes traffic into wounds early, and the latter mature into macrophages over time in the PVA sponge wound environment [19]. This suggests that lung infection may block monocyte and neutrophil trafficking into the wound.
The absence of innate immune cell subsets in the wounds of IAV-infected mice suggested that the inflammatory environment that drives acute wound healing was disrupted [8, 36,53–55]. To assess this, we measured the levels of inflammatory cytokines and chemokines in the wounds of uninfected and infected mice. Soluble factors were extracted from sponges by centrifugation. An array of inflammatory factors involved in acute wound healing was assayed in the wound fluid including monocyte chemoattractants (CCL2, CCL3, CCL4, and CXCL10), neutrophil chemoattractants (CXCL1 and CXCL5), and cytokines (IL-1α, IL-1β, IL-6, IFN-α, IFN-γ, TNF-α, and GM-CSF). A number of these chemokines have overlapping functions in coordination of the innate immune response and angiogenic processes, which are critical for proper wound healing [55]. The chemokines CCL2, CCL3, CCL4, CCL5, CXCL1, CXCL5, and CXCL10, as well as the cytokines IL-6, TNF-α, IL-1α, IL-β, and IFN-γ were detectable in wound fluid within the first seven days post-wounding (Fig 2). Many inflammatory cytokines and chemokines were expressed at high levels one day after wounding and declined over time (Fig 2). CCL5, CXCL10, and IL-1α peaked at day 4 while CXCL5 and IL-1β had biphasic peaks (Fig 2). In mice that were infected with IAV, there were changes in cytokines and chemokine levels over the course of the healing wound. The chemokine CCL2 was decreased in wounds from mice infected with IAV as early as 2 days after wounding (Fig 2A). CXCL10 and CCL5 were decreased in wound fluids from IAV-infected mice 4 days after wounding (Fig 2A). This is remarkable as these decreases in chemokines and cytokines occurred before a significant change in innate immune cellularity, suggesting that the low cellularity could be partially due to a reduced chemotactic signal (Fig 1E). There was a decrease in the concentration of all detectable chemokines except CXCL1 in the day 7 wound fluids collected from IAV-infected mice, as compared to uninfected controls (Fig 2A). The proinflammatory cytokines IL-6 and TNF-α, as well as the alarmin IL-1α were all decreased in healing wounds at days 4 and 7 after wounding (Fig 2B). The overall modulation of cytokine and chemokine expression indicates that a lung infection can induce changes in the dermal wound environment.
Some studies have indicated that trauma suppresses the immune response in the lung [43]. Our PVA sponge model did not impact the initial control of IAV replication in the lung (Fig 3A). This could be because the PVA sponge wound is restricted to the subcutaneous space, does not involve blood loss, and therefore does not induce the high degree of trauma-induced immunosuppression that major surgery does. Wounding also did not impact the influx of immune cells into the lung, as wounded and unwounded mice responding to IAV infection had the same number of immune cells in the bronchoalveolar lavage fluid (BALF) (Fig 3B) and the lung parenchyma (Fig 3C). Since innate immune myeloid cells were decreased in the wounds of IAV infected mice, we investigated whether these same cells were also impacted 7 days after wounding in the lungs of mice that were responding to the two insults. The full gating strategy for BALF and lung cells can be seen in S2 Fig. To differentiate lung myeloid cell subsets, CD11c combined with F4/80 was used to define macrophages and monocytes. Neutrophils were identified by Ly6G expression. In the lung, alveolar macrophages express Siglec-F; therefore, this marker was not used to identify eosinophils in this compartment. Wounding did not alter the types of innate leukocytes, including Ly6G+ neutrophils and F4/80+Ly6Chi monocytes, which infiltrated the BALF and lung in response to IAV infection 7 days after wounding (Fig 3D and 3E). However, in mice that were wounded and infected with IAV there was an increase in the number of CD11c+F4/80+Ly6C- macrophages 7 days after wounding (Fig 3F and 3G), indicating a small change in the cellularity of the interstitial lung space.
The cellular innate immune response and the ability to clear IAV were not severely impacted in mice that were responding to the PVA sponge implantation wound. However, since cytokine and chemokine levels were impacted in the wounds prior to changes in cellularity, we examined these factors in the BALF from control, IAV-infected, wounded, and IAV-infected and wounded mice. Wounding alone did not significantly alter the levels of cytokines and chemokines in the BALF compared to unwounded and uninfected control mice (wounding day 0 in Fig 4). Infection with IAV caused a gradual increase in cytokine and chemokine concentrations as the cellular immune response increased in the lung. The cytokines IL-6, TNF-α, IL-1α, and IFN-γ were all increased in the BALF of mice that were responding to dermal wounds and infected with IAV, compared to mice infected with IAV alone (Fig 4A). Interestingly, the alarmin IL-1α was elevated in infected and wounded animals compared to infected mice alone as early as day 4 after wounding (Fig 4A). Responding to pulmonary IAV infection and dermal wounds simultaneously caused a decrease in the chemokine CXCL5 at day 4 after wounding, and an increase in CCL2, CCL4, CXCL1, and CXCL10 at day 7 after wounding, compared to IAV infection alone (Fig 4B). While there was a mild increase in cytokine- and chemokine-mediated inflammation in wounded and infected animals as compared to infected mice alone, this did not correlate with a loss of pulmonary vascular permeability, as assayed by total protein content in cell-free BALF (S3 Fig) [56–59].
We next examined the circulation to determine if local changes in the innate immune response to wounding and pulmonary pathogens reflected systemic changes in immune function. The cellularity of the blood was impacted primarily by infection, although there were distinct differences in wounded and infected animals (Fig 5A). PVA sponge wounding alone caused only slight modulations in blood cellularity over time. In contrast, as the innate immune system responded to IAV infection alone, there was a large concurrent decrease in the number of cells in the blood by day 2 (Fig 5A). Interestingly, the number of cells in the blood of mice that were infected and wounded initially dropped similarly to infected mice, but the recovery was quicker starting at day 4, and by day 7 the number of cells was close to that of uninfected wounded mice (Fig 5A).
The distribution of innate immune cell subsets in the blood was assessed by flow cytometry analysis (Fig 5B). The complete gating strategy for blood cells is shown in S4 Fig. Wounding and infection both caused an increase in the percentage of inflammatory cells including Ly6G+ neutrophils and F4/80+ monocytes in the blood compared to uninfected and unwounded control animals (wounding day 0) (Fig 5B). The most notable difference was that mice that were both infected and wounded had increased neutrophil numbers compared to wounded mice alone and control mice (wounding day 0) (Fig 5C).
Using the PVA sponge wound model, plasma from all four groups of mice was analyzed for cytokine and chemokine concentrations. Wounding caused an initial systemic spike of the inflammatory cytokines IL-6 and TNF-α at wound day 1 (Fig 5D). Plasma levels of IL-1β were slightly elevated later in the wound healing response, while IFN-γ was not impacted by the response to a healing wound (Fig 5D). The systemic cytokine and chemokine responses measured from days 4 to 7 in wounded and IAV-infected mice were similar to those measured in unwounded, IAV-infected mice. This suggests that the later systemic response in wounded and infected mice is dominated by the effects of pulmonary IAV infection. With the exception of IFN-γ, which was elevated at day 7 in the plasma of mice that had a wound and lung infection compared to IAV infection alone, the dual insults did not alter the profiles of the other cytokines (Fig 5D).
Mice that were infected with IAV after wounding also had elevated plasma chemokine concentrations (Fig 5E). In particular, CCL2, CXCL5, and CXCL10 had increased levels early after wounding and infection (Fig 5E). This augmentation of the response in the dual insult model was abrogated by day 7 after wounding (day 6 after infection), at which point the levels were close to that measured in mice infected with IAV alone (Fig 5E). This is in contrast to the changes in cytokine concentrations and blood cellularity, which were observed at later times during the response to infection and the healing wound. Overall these data indicate that there is a pattern of mild persistent systemic hyper-inflammation in mice that are responding to two insults.
The early innate immune response to both wound healing and infection share many properties [60]; we therefore hypothesized that the outcome of wound repair would be negatively influenced by the presence of an infection at a distal site. Previous observational studies by others demonstrated that both Sendai virus infection and murine hepatitis virus (MHV) infection altered skin wound tensile strength; however, the mechanisms for this altered wound healing were not elucidated [61]. To determine how a pulmonary infection with IAV impacts dermal wound healing we developed models of post-injury lung infection that allowed us to specifically examine the cellular responses of wound repair, as well as the rate of wound healing. As IAV is the most common viral cause of both community- and hospital-acquired pneumonia [62], this was an ideal model viral lung pathogen.
Acute wound healing is a coordinated process that occurs in three phases: inflammation, proliferation, and repair or fibrosis [2]. The initial inflammatory phase is critical to prevent wound infection, clear cell and tissue debris, and coordinate downstream angiogenic and fibrotic responses. Disruption of the inflammatory phase can result in impaired tissue repair responses and delayed healing. This is consistent with findings presented here, which demonstrate that a viral pulmonary infection negatively impacts wound cellularity, dampens wound cytokine and chemokine responses, and delays the rate of wound closure. Interestingly, these effects on the wound environment were uncoupled from the systemic response, in which we observed elevated concentrations of circulating chemokines and proinflammatory cytokines in mice responding to both injury and lung infection. This is distinct from the immune suppression observed in the systemic response to bacterial infection after pulmonary IAV infection [14]. In addition, the innate immune system was able to mediate early control of pulmonary IAV infection, as the viral load was not impacted in mice that had an ongoing wound healing response. Our results demonstrate that when the innate immune system is activated by consecutive challenges with contemporaneous responses, the response may be biased towards one particular site. Here we examined a coincident lung infection and a sterile injury of the skin, which demonstrated that the response was biased towards prioritizing the lung infection in detriment to the wound. This may be evidence for an immune triage mechanism offering the best chances of overall survival of the host.
The cells that were impacted the most by simultaneous responses to lung infection and wounding were neutrophils and monocytes/macrophages. These cells are important innate immune drivers of both acute wound healing and the early pulmonary antiviral response, which we hypothesized would affect their ability to respond to one or both of these successive and concurrent inflammatory insults. Prior studies with the PVA sponge model, as well as other wound models, have established that neutrophils are recruited to the injury site very early [2,3,18,19,24]. They are followed by Ly6Chi inflammatory monocytes, and together they orchestrate the early inflammatory phase of wound repair. Monocytes mature over time into macrophages that contribute to wound repair responses such as angiogenesis and fibrosis [2, 19, 24, 53]. Our study shows that accumulation of neutrophils and monocytes is suppressed in PVA sponge wounds in mice with a lung infection. Work done by others has demonstrated that depleting monocytes/macrophages at early stages of wound repair negatively impacts healing [8,53]. Interestingly, the decrease in the proportion of inflammatory monocytes in the wounds of infected mice was accompanied by a relative increase in the proportion of wound macrophages, which could suggest an accelerated transition of monocytes to wound repair macrophages. When examining absolute numbers, however, both populations were reduced, suggesting that the loss of infiltrating monocytes may instead lead to a reduced number of monocyte-derived macrophages recovered from the wounds of infected mice [19].
In addition to decreased inflammatory cells, the decrease of cytokines such as IL-6 has important implications in the wound healing process [26, 63–66]. Our data from studies using the PVA sponge model demonstrated suppressed acute wound cellular and cytokine responses including IL-6 in the presence of a simultaneous lung infection. IL-6 has been shown to control levels of adhesion molecules, growth factors, and chemokines that are essential to the healing wound [29,30,67]. Additionally, the production of the angiogenic growth factor, VEGF-A, was also suppressed in mice with concurrent pulmonary infection suggesting downstream consequences for the initiation of the repair stage of wound healing.
Coordinated chemokine expression is essential to orchestrating the various phases of wound repair. We measured decreased expression of CCL2, CCL5, and CXCL5 at early, mid, and late acute stages of healing, respectively, in the wound fluids of IAV-infected mice. These chemokines recruit innate immune populations such as neutrophils and inflammatory monocytes, and are pro-angiogenic, suggesting that their sequential disruption may inhibit the progression through the early stages of wound repair. CXCL10 levels were also reduced in the wound fluids of IAV-infected mice at 4 days post-wounding. In the wound, CXCL10 is an angiostatic signal that promotes the transition from granulation to resolution phases, further suggesting that dysregulated chemokine signaling negatively impacts wound healing.
CXCL10 is normally induced by IFN-γ, however examination of local and systemic IFN-γ production suggested that wound CXCL10 was induced independently. For example, wound fluid IFN-γ levels were very low at all time points examined, measuring only 10 pg/mL at its peak one day after wounding, and this IFN-γ peak preceded the height of CXCL10 expression in the wound fluid by three days. Furthermore, there were no differences in the wound fluid concentration of IFN-γ between uninfected mice and mice with IAV infection, whereas CXCL10 expression was diminished in IAV-infected mice. A recent study showed that CXCL10 can be induced independently of the IFN response [68]. In particular this has been shown in response to viruses, although it has yet to be explored, to our knowledge, in wound healing responses. Some studies have demonstrated the ability of oral keratinocytes and fibroblasts to produce CXCL10 transcripts in vitro in response to TNF-α, IL-1α, and IL-4 stimulation. Keratinocytes and fibroblasts also have roles in healing cutaneous wounds, suggesting that IFN-γ-independent CXCL10 expression could occur in the wound [69–71]. In addition, impaired chemokine levels may stem from numerical deficits or functional defects in their numerous cellular sources, such as platelets, endothelial cells, and leukocytes. Systemic IFN-γ could also induce local wound production of CXCL10. IFN-γ was detected in the plasma, but its systemic peak succeeded the peak of both wound and plasma CXCL10, suggesting that, in general, the modulation of CXCL10 expression in IAV-infected mice was independent of IFN-γ signaling. Interestingly, plasma CXCL10 was induced by IAV infection more rapidly than in the BALF, suggesting an extrapulmonary site of production, such as the liver, which has been shown to become activated during influenza infection [14, 72–73].
In contrast to what was observed in the healing wound, there was a mild increase in systemic inflammation in mice with both wounds and infection. IAV infection with and without wounding resulted in a drop in circulating leukocytes 4 days after wounding and 3 days after IAV infection, likely due to extravasation to the infected lung. At 7 days after wounding, wounded mice had the highest number of circulating leukocytes while blood cellularity was lowest in IAV-infected mice. Mice with wounds and infection had intermediate levels of blood cellularity, suggesting that circulating cells were influenced by both inflammatory sites. This trend was also reflected in circulating neutrophils and monocytes. There were also increased concentrations of systemic cytokines and chemokines measured in the plasma of wounded and IAV-infected mice. The trends in plasma cytokine content reflected those measured in the BALF at 7 days post-wounding, suggesting that IAV infection was the major driver of systemic inflammation in wounded and infected mice at later time points. At day 4 after wounding there was an increase in the chemokines CCL2 and CXCL10 in the serum from mice that had been both wounded and infected. Both CXCL10 and CCL2 are important in the recruitment of many cell types, including monocytes, to inflamed tissue. These synergistic inflammatory cytokine and chemokine responses may suggest that the presence of dual inflammatory insults signals to increase leukocyte mobilization from sites of hematopoiesis to the circulation for increased availability to the inflamed periphery.
There was also an increase in markers of inflammation in the infected lungs of mice with PVA sponge wounds. As observed in the plasma, BALF recovered from the lungs of wounded and infected mice had increased levels of the chemokines CCL2 and CXCL10. BALF concentrations of IL-6, TNF-α, IL-1α, and IFN-γ were also increased in the BALF of wounded and infected mice when compared to IAV-infected mice alone. Macrophage numbers were slightly elevated in the lung and could be responsible for this mild inflammatory response. Alternatively, this could be due to increased cytokine production on a per-cell basis. Activated non-immune cells, such as the epithelium or endothelium, may also contribute to increased inflammation in the lung. Together these data suggest that, while the innate immune response is able to control the early stages of viral infection, it appears that distal wounding causes dysregulation of certain aspects of pulmonary inflammatory immune responses during IAV infection. However, this increase in pulmonary inflammation in wounded mice does not appear to contribute to additional lung damage compared to IAV-infected mice without wounds.
Our study shows that when confronted with both a pulmonary infection and a dermal wound the immune response is impacted in all compartments; however, the impact on the wound healing response is the greatest. The data presented here suggest that a distal infection disrupts the inflammatory phase of wound repair, resulting in delayed healing. Surprisingly the innate immune mediated control of the lung infection was not impacted by the injury. This study specifically addresses the effect that lung infection has on cutaneous and subcutaneous wounds, but it could also impact the healing of internal injuries. In the patient population, the complications that arise from the immune response combating simultaneous inflammatory events have important consequences. Altered wound healing could increase susceptibility to further complications leading to increased patient morbidity [17,36,37]. This proof of concept study opens a new area of research that aims to understand the intricacies underlying the innate immune response in complex biological systems facing multiple inflammatory insults. How the immune response is directed towards the lung at the expense of the healing wound and how this prioritization can be altered are important areas of future study. Given the numerous important functions of the innate immune response, these results have implications for many diseases, which remain to be explored. The ultimate aspiration is to advance the clinical outcomes of patients with complex disease sequelae.
All animal studies were approved by the Brown University Institutional Animal Care and Use Committee and carried out in accordance with the Guide for the Care and Use of Animals of the National Institutes of Health. Brown University adheres to the “U.S. Government Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training”, “PHS Policy on Humane Care and Use of Laboratory Animals”, “USDA: Animal Welfare Act & Regulations”, and “the Guide for the Care and Use of Laboratory Animals”. The University is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC). Brown University’s PHS Assurance Number: A3284-01, expiration date: July 1, 2018. The USDA Registration Number is 15-R-0003. Brown University IACUC was approved on October 8, 2013, and the animal protocol number is 1308000011. The de novo renewal was approved on September 28,2016 and the animal protocol number is 1608000222.
Group sizes of studies were determined by power analysis. To have confidence in our data we aimed to have a p value (alpha) of .05, and a power of .80 (beta of 20%) in a 1-way ANOVA. All power analysis was done using Power Analysis in R package: pwr.anova.test(k = 4, f = .25, sig.level = .05, power = .8 n = 4). Means and shared sigma values (standard deviation) were used from biological data generated in preliminary experiments, and from published work from lung infections. Exclusion of mice from data collection was determined based on extensive experience with murine pulmonary infection and sterile wound models. No mice were excluded from these studies according to the following criteria: 1) mice were not infected as determined by observation of physical appearance and measurement of viral titers and lung cellularity and 2) mice displayed overt wound infection, indicated by swelling, redness, pus, and irritation of the surrounding skin. Study endpoints were determined prospectively based on prior experience with mouse models of pulmonary infection and sterile wound healing. Each experiment was repeated a minimum of three times, with a minimum of three mice per group.
C57BL/6J mice were purchased from The Jackson Laboratory. As is consistent with previously published work with the wound model and to prevent gender specific complications, male mice 8–12 weeks of age were used in all experiments [74,75].
Mice under anesthesia and analgesia by ketamine (60–80 mg/kg) and xylazine (30–40 mg/kg) injection were administered IAV intranasally in a volume of 30 μL using a sterile saline vehicle. Mice were infected with 500 PFU influenza A virus (A/WSN/33 (H1N1)) strain. Influenza A virus was obtained from Akiko Iwasaki at Yale University. It was propagated using MDCK cells using standard procedures as described [12]. Mice were monitored daily for a minimum of three days, and every other day for the remainder of the experiment.
Polyvinyl alcohol (PVA) sponge implantation surgeries were performed under anesthesia and analgesia by ketamine (60–80 mg/kg) and xylazine (30–40 mg/kg) injection. Backs were shaved and cleaned with povidone-iodine solution and isopropyl alcohol. Six 1cm×1 cm×0.3 cm sterile PVA sponges (Ivalon, PVA Unlimited, Inc.) were placed into subcutaneous pockets through a 2cm midline dorsal incision under sterile conditions. The incision was closed with surgical clips. Mice were monitored daily for the first three days following surgery then a minimum of every other day for the remainder of the duration of the experiment.
The tail was cleaned with povidone-iodine solution and isopropyl alcohol. A 1cm x 0.3cm area of the skin was excised using a scalpel 0.5cm from the base of the tail. The wound bed was covered with a spray barrier film (Cavilon, 3M). Wound area was measured using calipers. Length and width measurements were taken at the midpoints of the wound bed. Tail wound images were acquired from a fixed position using a 12-megapixel iSight camera and were analyzed using ImageJ (NIH). All measurements were done in a blinded fashion to prevent bias.
Mice were euthanized by CO2 asphyxiation prior to sponge removal. For wound fluid collection, the three sponges implanted left of the midline were removed and placed in the barrel of a 5mL syringe, which was placed in a tube and centrifuged for fluid collection. The three remaining sponges that were implanted right of the midline were removed from each animal and placed in 1x HBSS medium (1% FCS/penicillin/streptomycin/1M hepes), and the cells were isolated using a Stomacher (Tekmar). Wound cells were washed with 1x HBSS medium and red blood cells were lysed. Cell counts were obtained using a Moxi Z Automated Cell Counter (Orflo).
Blood was collected retroorbitally into heparinized tubes at indicated time points. Each collection time point represents an independent sample group. Plasma, leukocytes, and red blood cells were fractionated by centrifugation in Wintrobe Tubes (CMSLabcraft). Plasma was collected for cytokine analyses. The buffy coat layer containing leukocytes was collected into a fresh tube and residual red blood cells removed by water lysis. Cells were counted with a Moxi Z Automated Cell Counter (Orflo) and used for flow cytometry analyses. The remaining red blood cell layer was discarded.
To collect bronchoalveolar lavage fluid (BALF), the trachea was exposed, and a BD Venflon IV catheter was inserted into the trachea. The needle was removed and a 1ml syringe filled with PBS was inserted. The lung was rinsed with 1ml PBS twice using an attached syringe. Cell-free supernatants were collected for cytokine analyses and protein content quantification. Cells were counted with a Moxi Z Automated Cell Counter (Orflo) and used in flow cytometry analyses.
The concentration of protein in the BALF was determined using the bicinchoninic acid (BCA) assay according to the manufacturer’s instructions (Pierce Chemical Co.). A dilution series was tested for each sample against an Albumin standard.
For isolation of cells from lungs, the right superior and middle lobes were perfused with 20 ml of PBS. The lung tissue was cut into small pieces and incubated for 45min at 37 degrees C in 4ml of media containing type 4 collagenase (Worthington Biochemical Corporation) and DNAse I (Sigma-Aldrich). Afterwards, digested lung tissue was passed through a 70uM cell strainer to make a single cell suspension. After centrifugation the cell pellet was re-suspended in 4ml of 40% Percoll/RPMI and carefully layered over 4ml of 80% Percoll/PBS. The gradient was centrifuged at room temperature for 20 minutes at 600g with minimal acceleration and deceleration. Cells assembled in the interphase were collected, and washed with 10ml RPMI media containing 5% fetal calf serum by centrifugation.
Viral PFU were obtained using the unperfused right inferior lobe. The lobe was homogenized in 1 ml of PBS using the Polytron 2100 homogenizer. Viral plaque forming units (PFUs) were obtained by titration of diluted supernatant on MDCK cells as described elsewhere [14]. Briefly, cell homogenate was diluted in PBS and 100 μl was plated on MDCK cells in 6 well plates. After a 1 hour incubation at 37° C the virus was removed and the cells were covered in 50% media/50% oxoid agar supplemented with DEAE dextran, NaHCO3, and penicillin/streptomycin. After 3 days the agar was removed, the plates were stained with crystal violet, and viral plaques were counted.
The following antibodies were used to identify cell subsets: Ly6C-FITC (AL-21, BD Biosciences), F4/80-PerCP-Cy5.5 or APC eFluor660 (BM8, eBioscience), Siglec-F-PE (E50-2440, BD Biosciences), CD11c-PE or BV711 (HL3, Biolegend), and Ly6G-PerCP eFluor710 or V450 (1A8, eBioscience or BD Biosciences). Dead cells were excluded from analyses using Fixable Viability Dye APC eFluor780 or BV506 (eBioscience).
Surface staining: Cells from all samples were adjusted to an equal concentration, and treated with anti-CD16/CD32 Fc receptor blocking antibody (clone 2.4G2) in 1x PBS (1% FBS) for 10 minutes on ice. Surface staining antibodies were then added and incubated for 15 minutes on ice. Cells were washed with 1x PBS then incubated with Fixable Viability Dye diluted in 1x PBS for 15 minutes on ice. Cells were washed, then fixed with 1% paraformaldehyde for 15 minutes on ice.
Samples were acquired using an Attune NxT Acoustic Focusing Cytometer with Attune Software or a BD FACSAria with FACSDiva Software. Analyses were performed using FlowJo v10 software (Tree Star, Inc.). Gate placement was determined using isotype, FMO, or unstained control samples. Total cell numbers of each cell subset was obtained by using the total cell counts of the compartment as described above, and multiplying by the percent of total viable cells as determined by flow cytometry.
Cytokine concentrations were determined in wound fluid, BALF, and plasma using a custom LEGENDplex bead-based immunoassay (BioLegend) according to manufacturer instructions, except for VEGF-A, CXCL1, and CXCL5. The concentrations of these cytokines were determined using DuoSet sandwich ELISA kits (R&D Systems) according to manufacturer instructions.
Biostatistical analysis was carried out using the GraphPad Prism software package. For comparison of two groups the nonparametric Mann Whitney test was used. To compare 3 or more groups the Kruskal-Wallis one-way analysis of variance was used, with Tukey-Kramer test for post-hoc analysis. All the groups were compared to each other. However, for clarity, the statistics shown are the most relevant for this study. This means a comparison of Wound + IAV with control, Wound, or IAV at all time points. Results are considered statistically significant when the P value ≤ 0.05. Statistically significant changes between control and wound + IAV are denoted by %, between IAV and wound +IAV are denoted by #, wound and wound +IAV are denoted by *.
|
10.1371/journal.pcbi.1006822 | Efficient neural decoding of self-location with a deep recurrent network | Place cells in the mammalian hippocampus signal self-location with sparse spatially stable firing fields. Based on observation of place cell activity it is possible to accurately decode an animal’s location. The precision of this decoding sets a lower bound for the amount of information that the hippocampal population conveys about the location of the animal. In this work we use a novel recurrent neural network (RNN) decoder to infer the location of freely moving rats from single unit hippocampal recordings. RNNs are biologically plausible models of neural circuits that learn to incorporate relevant temporal context without the need to make complicated assumptions about the use of prior information to predict the current state. When decoding animal position from spike counts in 1D and 2D-environments, we show that the RNN consistently outperforms a standard Bayesian approach with either flat priors or with memory. In addition, we also conducted a set of sensitivity analysis on the RNN decoder to determine which neurons and sections of firing fields were the most influential. We found that the application of RNNs to neural data allowed flexible integration of temporal context, yielding improved accuracy relative to the more commonly used Bayesian approaches and opens new avenues for exploration of the neural code.
| Being able to accurately self-localize is critical for most motile organisms. In mammals, place cells in the hippocampus appear to be a central component of the brain network responsible for this ability. In this work we recorded the activity of a population of hippocampal neurons from freely moving rodents and carried out neural decoding to determine the animals’ locations. We found that a machine learning approach using recurrent neural networks (RNNs) allowed us to predict the rodents’ true positions more accurately than a standard Bayesian method with flat priors (i.e. maximum likelihood estimator, MLE) as well as a Bayesian approach with memory (i.e. with priors informed by past activity). The RNNs are able to take into account past neural activity without making assumptions about the statistics of neuronal firing. Further, by analyzing the representations learned by the network we were able to determine which neurons, and which aspects of their activity, contributed most strongly to the accurate decoding.
| Place cells, pyramidal neurons found in CA1 and CA3 of the mammalian hippocampus [1–4], exhibit spatially constrained receptive fields, referred to as place fields. In general, the activity of place cells is considered to be stable [5, 6]; place fields are typically robust to the removal of specific environmental cues [7, 8], persist between visits to a location [9], and in the open field do not strongly depend upon an animal’s behaviour [2, 5]. Upon exposure to a novel enclosure the firing correlates of place cells rapidly ‘remap’; place fields change their firing rate and relative position, forming a distinct representation for the new space [10–12]. For these reasons place cells are widely held to provide the neural basis of self-location, signalling the position of an animal relative to its environment and thus being a necessary element for the control of spatial behaviours, such as navigation, and the retention of spatial memories [2]. Unsurprisingly then, given information about the activity of a population of place cells, it is possible to decode the location of an animal with a relatively high degree of accuracy [13, 14].
However, although place cell activity is strongly modulated by self-location this relationship is non-trivial and not exclusive. For example, during rest and brief pauses, but also during motion, the place code can decouple from an animal’s current location and recapitulate trajectories through the enclosure [15]; ‘replaying’ previous experience [16] or, perhaps, foreshadowing future actions [17]. Similarly, when animals run on linear runways or perform constrained navigational tasks, such as T-maze alternation, place cell activity becomes strongly modulated by behaviour, disambiguating direction of travel [18], prospective and retrospective trajectories [19, 20], and the degree of engagement with a task [21]. Furthermore, although place fields are repeatable they are not static. Even though remapping occurs rapidly in a novel environment, the newly formed firing fields continue to be refined during subsequent experience, a process that appears to persist for several hours [10, 13, 22, 23]. Even in familiar environments, that animals have visited many times, the spatial activity of place cells is known to exhibit incremental changes that can result in the generation of distinct spatial codes [23–26], which might be important for encoding goal locations [27] or other non-spatial variables [28]. As such, although hippocampal activity provides considerable information about an animal’s self-location the representation is dynamic: accumulating changes and sometimes encoding other variables both spatial and non-spatial.
A common approach used to interrogate neural representations, such as that of place cells, is decoding; the accuracy with which a variable, such as self-location, can be decoded from the brain, places a useful lower limit on the amount of information present [13, 14]. In the case of place cells, decoding methodologies typically apply a Bayesian framework to calculate a probability distribution over the the animal’s position, given the observed neural data [14, 16, 29]. Decoding to a specific location is then accomplished via a maximum likelihood estimator applied to the probability distribution. However, the accuracy of Bayesian methods depends on accurate information about the expected activity of neurons. For place cells, activity recorded over the course of tens of minutes is typically used to estimate the firing rate of each cell at different points in the animal’s enclosure, with instantaneous rates assumed to exhibit Poisson dynamics. However, for the reasons outlined above, it is not clear that hippocampal activity can be modelled in this way. Indeed, the variability of place cell firing rates is known to greatly exceed that expected from a Poisson process [30]. As such, it is likely that Bayesian methods, as currently applied, do not provide an accurate reflection of the accuracy with which the hippocampus encodes self-location.
To better understand these constraints, we trained a deep recurrent neural network (RNN) [31–33] to decode rodent location from the firing rates of CA1 neurons. At each time step the network was presented with a vector corresponding to the spike counts of hippocampal cells within a given time window. After accumulating information for 100 time-steps the network was required to predict the animal’s location—supervision being provided in the form of the animal’s true location. We found that decoding with the trained RNN was consistently more accurate than a standard Bayesian approach [14, 29] with flat priors (essentially a MLE) as well as a Bayesian decoder with priors informed by the animals’ historic activity [14, 16]. This demonstrates that RNNs are able to capture the relationship between a temporal sequence of neural activity and an encoded variable without the necessity of explicit assumptions about the underlying noise model or complicated hand-coded priors. Further, inspection of the trained network allowed us to identify both the relative importance of individual neurons for accurate decoding and the locations at which they were most informative. Thus, not only does the accuracy of the RNN set a new limit for the amount of information about self-location encoded by place cells but more generally this work suggests that RNNs provide a useful approach for neural decoding and provide a means to explore the neural code.
In the following Results section we summarize the decoding performance with three decoders. The simple Bayesian decoder uses spike counts from a single time window centered around the location measurement. It combines the likelihood of these spike counts occurring in different locations with a flat prior to make a decision, meaning it is essentially a maximum likelihood estimator (MLE). The Bayesian with memory[14] employs a prior informed from self-location and movement speeds decoded from previous time steps (memory) as well as knowledge about the locations the animal most often visits, and combines it with the likelihood calculated based on the spike counts in the current time window.
The third decoder, recurrent neural network (RNN), is a machine learning algorithm that learns from examples via gradient descent [31]. In particular, by providing the network with a sequence of spike count vectors from 100 consecutive time windows and the location at the center of the last window, the algorithm learns to predict the latter from the former. As the network has access to neural activity from many preceding time windows, it learns to use this as contextual information to improve decoding accuracy. Learning is achieved by incrementally modifying the connection strengths in the network in order to minimize the mean squared error loss. The direction of change can be found by calculating the gradient of the loss with respect to each parameter.
In this work we used a type of recurrent neural network called long-short term memory (LSTM, [34]). All reported results are obtained with a network consisting of two 512-unit LSTM layers followed by a linear output layer (2 values, one for each coordinate). For more detailed description of the network, its inputs and training parameters see the Methods section and the code in GitHub.
To test the RNN’s ability to decode rodent location based on hippocampal activity we first characterized the decoding error for a single animal foraging in a 2D arena (1m x 1m square). Single unit recordings were made using tetrodes from region CA1 of five rats. In all animals less than 10% of the recorded neurons were interneurons, characterized by narrow waveforms and high firing rates. Rat R2192 yielded the greatest number of simultaneously recorded hippocampal neurons (n = 63). Since the number of recorded neurons is expected to correlate with decoding accuracy, we first focused on this particular animal.
Neural data was processed to extract action potentials and these were assigned to individual neurons using the amplitude difference between tetrode channels [35] (see Methods). The input features for the RNN-decoder then consisted of spike counts for each neuron within a set of time windows. The length of time windows was parametrically varied between 200 ms and 4000 ms in 200 ms increments. Each consecutive window started 200 ms later than previous one (this means 0% overlap for 200 ms windows, 50% overlap for 400 ms windows, 80% overlap for 1000 ms windows, etc. See “Feature extraction” in Methods). The network was presented with spike counts from 100 windows before being asked to predict the animal’s location at the center of the latest window.
As the RNN training process is stochastic, 10-fold cross validation (CV) procedure was run multiple times for each window size. For each of these runs we trained 10 models (for each fold of CV) and extracted the mean and median results across the folds. Black dots on Fig 1 correspond to these different realizations of the 10-fold CV procedure (notice multiple dots per window size). 10-fold cross validation was also applied to the Bayesian decoders.
For both the mean (Fig 1a) and median (Fig 1b) of the validation errors, the error curve was convex with lowest errors obtained at intermediate values. Best median decoding accuracy was achieved with time window of 1200 ms (median error = 10.18±0.23 cm). Best mean decoding was achieved for a time window of 1400 ms (mean error = 12.50±0.39) cm). Using longer or shorter time windows lead to higher errors—likely because spike counts from shorter windows are increasingly noisy, while the animal’s CA1 activity is less specific to a particular location for longer windows. For all time windows, the accuracy of the RNN considerably exceeded that of both the simple Bayesian decoder (dashed red line) and the Bayesian decoder with memory (solid red line). The lowest median decoding error achieved with the simple Bayesian decoder was 12.00 cm (17.9% higher than for the RNN; this accuracy was obtained with multiple different window sizes), lowest mean error was 15.83 cm with a 2800 ms window. The Bayesian decoder with priors informed by the animal’s historic activity was more accurate than the naive Bayes approach but was still considerably less accurate than the RNN (lowest mean decoding error was 15.46 cm with 2000 ms windows, and lowest median 11.31 cm with 1400ms windows).
The RNN has the ability to flexibly use information from all 100 input vectors and thus integrates contextual information over time. This results in lower mean and median errors as compared to the two baseline Bayesian approaches. The naive Bayesian method with flat priors does not have access to information about past activity, resulting in lowest accuracy. Equally, the Bayesian decoder with memory incorporates past activity to form an informed prior, but does this in a predefined manner, being less flexible than the RNN. Notice also that the RNN approach achieves its best results for shorter time windows than the Bayesian approaches (see also Table 1 for optimal window size results from other animals). We hypothesize that the RNN’s efficient use of contextual information helps it to overcome the stochastic noise in the spike counting obtained for shorter time windows.
Beyond the global descriptors of mean and median error, we also inspected the distribution of decoding error sizes (Fig 2a).
The RNN error distribution followed a unimodal curve with most predictions deviating from the rat’s true position by 6-8 cm and few errors were larger than 35 cm (1.7% of errors > 35 cm, see Fig 2a). The Bayesian classifiers achieve more very low (<2 cm) errors, but also an abundance of very large (>50 cm) errors (≈8% of errors > 35 cm, ≈2.7%> 50 cm; for both Bayesian classifiers).
In many cases single unit recordings yield fewer than the 63 neurons identified from R2192. We hypothesised that the RNN’s ability to use contextual information would be increasingly important in scenarios where neural data was more scarce. To test this prediction we randomly downsampled the dataset available from R2192, repeating the training and decoding procedure for populations of neurons varying in size from 5 to 55 in increments of 5. As expected we saw that decoding accuracy reduced as the size of the dataset reduced. However the RNN was considerably more robust to small sample sizes, decoding with an error of 30.9 cm with only 5 neurons vs. 46.0 cm error for the Bayesian decoder (Fig 2b).
In total we analyzed recordings from five animals as they foraged in a 2D open field environment (1m x 1m square). For each of these 5 datasets, we determined the best performing time window size (similarly to Fig 1) for the RNN architecture (composed of 2-layers of 512 LSTM units), simple Bayesian decoder (MLE), and Bayesian decoder with memory. The optimal time window sizes for the five 2D foraging datasets are given in top half of Table 1 along with the length of the recording and the number of identified neurons.
In the 2D decoding task, for different animals, the mean error (mean across cross validation folds) ranged between 12.5-16.3 cm and median between 10.3-13.1 cm (Fig 3a and 3b). Interestingly, despite some recordings yielding as few as 26 or 33 cells, the decoding accuracy using RNNs is roughly similar. In all cases the mean and median decoding results from the RNN decoder outperformed both the standard Bayesian approach and Bayesian with memory.
We also performed decoding on 1D datasets recorded while the same 5 animals shuttled back and forwards on a 600 cm long Z-shaped track for reward placed at the corners and ends (Table 1) [36]. As before we applied RNN and Bayesian decoders to 10-fold cross validated data, selecting in each case the optimal time window size (Table 1). The RNN decoder greatly outperformed the two Bayesian decoders in all 5 data sets when comparing mean errors (Fig 3c). In the 2D task the largest possible error was 141.7 cm (if the predicted location is in the corner diagonally opposite to the true location), whereas in 1D task it is 600 cm (if the opposite end of the track is predicted). In the 1D task a small number of extremely large errors will inflate the mean error, whereas the median will be less affected (Fig 3c and 3d). Examining the median errors we found that RNN outperformed the Bayesian decoders in all cases. However for four of the five animals the difference in error was relatively small (Fig 3d). For the fifth rat with the fewest recorded cells (R2117, n = 40), the RNN clearly outperformed the Bayesian approaches, having a median decoding error that was almost half that of what the two types of Bayesian decoders achieved.
Next to understand how behavioural and neural variability influenced decoding accuracy we focused on the results obtained from rat R2192 in the 1m square—the animal with the greatest number of neurons and the lowest decoding error.
First we examined the decoding error as a function of the rat’s location. It is important to note that the animals’ behaviour is non-uniform—the rats visit some parts of the arena more often than others (see Fig 4a). Since more training data is available for frequently visited regions it is expected that any decoding approach would be most accurate in those locations. The spatial distribution of decoding error for R2192 seems to confirm this conjecture—well sampled bins in the center of the enclosure and portions of its borders are more accurately decoded (Fig 4b). To confirm this, we calculated the correlation between the decoding error and the number of training data points located within 10 cm radius of the predicted data point, finding a significant negative correlation (Spearman’s Rank Order, r = −0.16, pval ≪ 0.001, dof = 4412).
Another important factor influencing the decoding accuracy is the distribution of neural activity across the 2D enclosure. In particular, place fields of the recorded hippocampal cells do not cover the enclosure uniformly. Clearly it would be difficult for the algorithm to differentiate between locations where no cell is active. As such, it is likely that areas where more neurons are activated are decoded with higher precision. Our results confirm that the sum of spike counts across neurons at a given location is strongly anti-correlated with the prediction error made at that location (Fig 4c, Spearman’s Rank Order, r = −0.31, pval ≪ 0.001, dof = 4412).
We also inspected the x and y components of the decoding error separately. Previous work suggests that, in the case of grid cells, contact with an environmental boundary results in a reduction of error in the representation of self-location perpendicular to that wall [37]. Such a relationship would be expected if boundaries function as an elongated spatial cue, used by animals to update their representation of self-location relative to its surface. Accordingly, we found that for RNN decoding based on CA1 neurons, the decoding accuracy orthogonal to environmental boundaries increased with proximity to that boundary (Fig 4d, Spearman’s Rank Order between error and distance to wall in the region up to 25cm from the wall, r = 0.31, p ≪ 0.001, dof = 3968). The result also held for x (r = 0.35, p ≪ 0.001, dof = 2101) and y (r = 0.25, p ≪ 0.001, dof = 1855) coordinates separately. Conversely, decoding error parallel to the boundary was not modulated by proximity.
Furthermore, an additional factor that seemed to influence prediction accuracy was the animal’s motion speed. Predictions were more reliable when the rat was moving as opposed to stationary. The mean prediction error for speeds below 0.5 cm/s being 16.5 cm, higher than the 12.1 cm average error for all speeds above 0.5 cm/s (two-sided Welch’s t-test, t = 10.62, p ≪ 0.001, median errors 8.68 cm and 7.74 cm accordingly). It seems plausible that the lower prediction accuracy during stationary periods might be due to place cells preferentially replaying non-local trajectories during these periods [38]. A second interesting observation is that the prediction error does not increase at higher speeds (two-sided Welch’s t-test between errors in data points where speed is in range from 0.5 cm/s to 10.5cm/s and errors in data points with speed above 10.5cm/s, t = 0.31, p = 0.76).
The accuracy of any neural decoder represents a useful lower bound on the information about the decoded state contained by the recorded neurons. Thus, a biologically relevant question is how such information is distributed among the neurons, across space and time. In short we asked which features of the neuronal activity are the most informative at predicting the animal’s position. To this end we conducted two different types of sensitivity analyses to measure robustness to different types of perturbations. For a visualization of the representations learned by the RNN, see the dimensionality reduction analysis (using t-SNE) in S1 Text, S2 and S3 Figs.
We have shown that the sequential processing afforded by an artificial recurrent neural network (RNN) provides a flexible methodology able to efficiently decode information from a population of neurons. Moreover, since a RNN decoder is a neural network, it represents a biologically relevant model of how neural information is processed. Specifically, when applied to hippocampal neural data from freely moving rats [2], the network made use of the past neural activity to improve the decoding accuracy of the animals’ positions. In a 2D open field arena (1m x 1m), the RNN decoder was able to infer position with a median error of between 10.3 cm to 13.1 cm for 5 different rats. These results represented a marked improvement over both a simple Bayesian decoder using a flat prior [14, 16, 29], which bases its decision solely on spike counts from a single time window centered around the moment of position measurement, as well as a Bayesian decoder incorporating priors informed by the animals’ behaviour and recent spiking history [14].
Bayesian methods are known to be optimal decoders when using appropriate priors [41]. However, when applied to neural decoding it is difficult to determine these appropriate priors—as a result sub-optimal approximations are commonly used. Hence we propose that RNNs offer a practical methodology to incorporate sequential context without the need to choose or estimate specific priors over high-dimensional spaces. The improvement in 2D position decoding observed for the RNN was mirrored by similar results from a 1D decoding task using hippocampal recordings made while animals ran on a 6 meter track. Here again, the RNN decoder achieved equal or better results than the Bayesian approaches.
Making use of the past neural activity as contextual information, the RNN seems more robust to noise than the two Bayesian classifiers. In particular when using shorter time windows the spike counts become noisier and the Bayesian models’ prediction accuracy degraded rapidly. In contrast the RNN decoder was more resistant to the variability of spike counts, likely due to its ability to combine information over the complete sequence of past inputs. Similarly, in situations where fewer neurons were available and hence the total amount information was reduced, the RNN exhibited a pronounced advantage over the Bayesian decoders. Equally, in the 1D task the benefit of the RNN was most evident for animal R2217, which had the fewest recorded neurons. Nevertheless notice that fewer recorded neurons does not necessarily mean lower accuracy. As described in Section 2.3.1, the error depends strongly on the amount of training data available (length of recording) and the quality of the cells (amount and location of firing). Taken together these results suggest that RNN decoding of neural data may prove to be particularly useful in situations where large populations of neurons are not available or are difficult to stably maintain.
Beyond quality and amount of data available, the size of error the RNN decoder made was also seen to depend on the distance of the animal from the walls and its instantaneous speed. At higher speeds (above 10.5 cm/s) the decoding accuracy does not decrease, but when the animal is immobile (below 0.5 cm/s) the error was significantly higher than when in motion. We hypothesize that while stationary hippocampal activity may reflect non-local activity associated with sharp-wave ripple states [38].
Beyond providing more accurate decoding, the neural network approach also provides a new means of conducting sensitivity analyses. While knockout-type sensitivity analyses can be applied to both Bayesian and RNN decoders, the latter approach also supports gradient analyses. The two types of sensitivity—knockout and gradient—are correlated, but not identical. By design knockout analyses answers how the system behaves if an input is completely removed, while gradient analyses investigated how the system behaves in response to small perturbations to that input. Having access to the gradients with respect to each spike count makes is possible to pose new questions about the dynamic variability of the information content of individual neurons.
All procedures were approved by the UK Home Office, subject to the restrictions and provisions contained in the Animals Scientific Procedures Act of 1986.
Deep learning is a class of algorithms that learn a hierarchy of representations or transformations of the data that make the problem of classification or regression easier [31, 33]. In particular, deep neural networks, inspired by biological neural circuits, consist of layers of computational units called neurons or nodes. The deepness means that there are multiple “hidden” layers between the input and output. By tuning the connection weights between its layers a neural network can learn to approximate a function from a set of examples, i.e., pairs of related input and output data. In this work we are interested in training a neural network to decode the rat spatial coordinates from the activity recorded from its hippocampal cells.
Whereas feed-forward neural networks learn to predict an output based on a single input, recurrent neural networks (RNNs) can deal with series of inputs and/or outputs [32, 33]. In particular, a recurrent network can preserve information from previous inputs by means of feedback connections (loops between its units). Having access to past information can be useful to minimize errors in certain tasks. Such memory of past inputs also means that the order in which the inputs are presented to the network may change the eventual predictions, and thus integrate contextual information over time. A naive implementation of RNNs can only maintain information from a few past inputs, making it possible for the network to detect only immediate trends, but not long timescale dependencies. Advanced realizations of recurrent networks, such as long-short term memory (LSTM) [34] and gated recurrent units (GRU)[42, 43] have specific architecture and sets of parameters that control to what extent past activity should be remembered or overwritten by a new input [43]. This makes them capable of integrating knowledge over a longer sequence. Through using past inputs as contextual information these networks have achieved outstanding performance with noisy sequential data such as text and speech.
Spatial decoding was also implemented using a Bayesian framework [14, 29, 49] subject to 10-fold cross validation (see also the next subsection). Specifically, for each fold, 90% of the data was used to generate ratemaps for hippocampal neurons—spike and dwell time data were binned into 2 cm square bins, smoothed with a Gaussian kernel (σ = 1.5 bins), and rates calculated by dividing spike numbers by dwell time. Note, for the Z-maze only, positional data was linearised before binning.
Next, with the remaining 10% of the data, using temporal windows (200 ms to 4000 ms) each of which overlapped with its neighbours by half, we calculate the probability of the animal’s presence in each spatial bin given the observed spikes—the posterior probability matrix [14, 16, 29].
Specifically during a time window (T) the spikes generated by N place cells was K = (k1,…, ki,…, kN), where ki was the number of spikes fired by the i−th cell. The probability of observing K in time T given position (x) was taken as:
P ( K | x ) = ∏ P o i s s o n ( k i , T α i ( x ) ) = ∏ i = 1 N ( T × α i ( x ) ) k i k i ! × e - T α i ( x ) ,
where x indexes the 2 cm spatial bins defined on the Z-track/foraging environment and αi(x) is the firing rate of the i − th place cell at position x, derived from the ratemaps.
In the case of the simple Bayes decoder, to compute the probability of the animal’s position given the observed spikes we applied Bayes’ rule, assuming a flat prior for position (P(x)), to give:
P ( x | K ) = R [ ∏ i = 1 N α i ( x ) k i ] × e - T ∑ i = 1 N α i ( x ) ,
where R is a normalizing constant depending on T and the number of spikes emitted. Note in this case we do not use the historic position of the animals’ to constrain P(x|K) thus the probability estimate in each T is independent of its neighbours. Finally, position was decoded from the posterior probability matrix using a maximum likelihood method—selecting the bin with the highest probability value. Decoding error was then taken as the Euclidean distance between the centre of the decoded bin and the centre of the bin closest to the animal’s true location.
Finally for the Bayes decoder with memory we made two further changes. First for each animal P(x), the probability of being at position x, was calculated directly from the experimental data for the entire trial, giving:
P ( x | K ) = R × p ( x ) × [ ∏ i = 1 N α i ( x ) k i ] × e - T ∑ i = 1 N α i ( x ) ,
Second, following [14] we incorporated a continuity constraint such that information about the animal’s decoded position in the previous time step was used to calculate the conditional probability of P(xt | spikest, xt−1).
P ( x t | s p i k e s t , x t - 1 ) = C × P ( x | K ) × n o r m D i s t ( x t - 1 , s i g m a )
Where C is a normalising constant and normDist is a normal distribution centred on the animal’s decoded position in the previous time step with sigma equal to the mean distance travelled per time step in the previous 15 time steps multiplied by a scaling factor which was set to 1 for the open field decoding and 5 for linear track.
The implementations of these two approaches can be found in the Bayesian folder of the GitHub repository https://github.com/NeuroCSUT/RatGPS. Please notice that the data for MLE and Bayesian apporaches must be downloaded and added to the Bayesian/Data folder manually, as the files were too large do be added to GitHub. As instructed in the README files in the repository, the data can be found via DOI (https://doi.org/10.5281/zenodo.2540921).
The reported errors for both Bayesian and RNN approach are measured using a 10-fold cross validation method that divides the D data points between training and validation sets. Due to the overlap between consecutive time windows a random assignment of data points to training and validation sets would imply that for most of the validation data points a highly correlated neighbouring sample can be found in the training set. This would result in an artificially high validation accuracy that does not actually reflect the model’s ability to generalize to new, unseen data.
Instead, in our analysis the first fold in cross validation simply corresponds to leaving out the first 10% of the recording time and training the model on the last 90% of data. The second fold, accordingly, assigns the second tenth of recordings to the validation set, and so on. For RNNs we need to additionally discard 99 samples at each border between training and validation sets. Remind that the input for RNNs is a series of 100 spike count vectors—to avoid any overlap between training and test data we remove validation data points that have at least one shared spike count vector with any training data point.
For each fold we train a model on the training set and calculate the error on the validation set. All reported errors are the validation errors—errors that the models make on the one tenth of data that was left out of the training procedure. To increase the reliability of the results, we perform 10-fold cross validation procedure multiple times and report the mean and median of the errors. This is done only for the RNN decoder, because the Bayesian decoder is deterministic and repeating cross-validation procedure multiple times is not necessary.
|
10.1371/journal.pntd.0005839 | “Koko et les lunettes magiques”: An educational entertainment tool to prevent parasitic worms and diarrheal diseases in Côte d’Ivoire | Integrated control programs, emphasizing preventive chemotherapy along with health education, can reduce the incidence of soil-transmitted helminthiasis and schistosomiasis. The aim of this study was to develop an educational animated cartoon to improve school children’s awareness regarding soil-transmitted helminthiasis, diarrheal diseases, and related hygiene practices in Côte d’Ivoire. The key messages included in the cartoon were identified through prior formative research to specifically address local knowledge gaps.
In a first step, preliminary research was conducted to assess the knowledge, attitudes, practices, and beliefs of school-aged children regarding parasitic worm infections and hygiene, to identify key health messages to be included in an animated cartoon. Second, an animated cartoon was produced, which included the drafting of the script and story board, and the production of the cartoon’s initial version. Finally, the animated cartoon was pilot tested in eight selected schools and further fine-tuned.
According to the questionnaire results, children believed that the consumption of sweet food, eating without washing their hands, sitting on the floor, and eating spoiled food were the main causes of parasitic worm infections. Abdominal pain, diarrhea, lack of appetite, failure to grow, and general fatigue were mentioned as symptoms of parasitic worm infections. Most of the children knew that they should go to the hospital for treatment if they experienced symptoms of parasitic worm diseases. The animated cartoon titled “Koko et les lunettes magiques” was produced by Afrika Toon, in collaboration with a scientific team composed of epidemiologists, civil engineers, and social scientists, and the local school children and teachers. Pilot testing of the animated cartoon revealed that, in the short term, children grasped and kept key messages. Most of the children who were shown the cartoon reported to like it. Acceptance of the animated cartoon was high among children and teachers alike. The messaging was tailored to improve knowledge and practices for prevention of helminthiases and diarrheal diseases through prior identification of knowledge gaps. Integration of such education tools into the school curriculum, along with deworming campaigns, might improve sustainability of control and elimination efforts against helminthiases and diarrheal diseases.
| Soil-transmitted helminthiases, schistosomiasis, and diarrhea remain important public health issues in sub-Saharan Africa. Health educational animated cartoons can help raise awareness and improve hygiene practices, and thus contribute to the control and elimination of these diseases. For the development of an educational animated cartoon, we first evaluated the knowledge, attitudes, practices, and beliefs of school-aged children in Côte d’Ivoire regarding soil-transmitted helminthiases, diarrhea, and schistosomiasis in order to identify setting-specific health messages for our animated cartoon. We found that children believed that the consumption of sweet food, eating without washing their hands, sitting on the floor, and eating spoiled food were the main causes of parasitic worm infections. As a next step, the alpha version of the animated cartoon was produced and given the title: “Koko et les lunettes magiques”. The animated cartoon was pre-tested in eight schools and further developed. Our study found that children could retain most of the information provided by the animated cartoon. It is suggested that such a tool, integrated into the school curriculum, together with deworming campaigns in Côte d’Ivoire, might improve the sustainability of control and elimination efforts against soil-transmitted helminthiases, diarrhea, and schistosomiasis.
| Infections with soil-transmitted helminths (STHs; i.e., Ascaris lumbricoides, Trichuris trichiura, and hookworm) are among the most common neglected tropical diseases (NTDs) [1,2]. Indeed, more than one billion people are infected with at least one species of STHs [3]. Parasitic worm infections are intimately linked to poverty, such as inadequate sanitation and waste disposal, lack of access to clean water, poor hygiene, and limited access to health care and preventive measures [4,5]. Although reinfection can occur rapidly after treatment [6], the main strategy to control morbidity due to parasitic worm infections is preventive chemotherapy, that is the repeated large-scale administration of anthelmintic drugs to high-risk groups, particularly school-aged children [7]. Importantly, STHs are amongst the many pathogens causing diarrhea, whereas the condition remains a leading driver of mortality and morbidity among children under the age of 5 years worldwide and disproportionally affects those from low- and middle-income countries (LMICs) [8]. Inadequate water and sanitation, suboptimal breastfeeding, zinc and vitamin A deficiency, and lack of access to quality health care and timely and effective treatment with oral rehydration solution are reasons for the high global burden of diarrheal diseases [8].
Although health education and sanitation are two important components of primary health care emphasized by the World Health Organization (WHO) as a basis for the prevention and control of communicable diseases [9], there is still limited application of health education in the control and elimination of parasitic worm infection as part of an integrated strategy [4,10]. It has been demonstrated that integrated control programs that combine preventive chemotherapy with health education for prevention of re-infections, can reduce prevalence and morbidity of STH and schistosomiasis [11–16]. In particular, health education offers opportunities for the community to improve their health by increasing knowledge and skill sets [17].
The potential of educational videos is well established, as they can engage and inform at the same time. This is of importance, particularly for younger school-aged children as their attention span is limited. Videos can display real-life situations that children can readily grasp. It is essential that the messaging of videos is tailored to the context of the target population. Hence, a preliminary assessment of the knowledge, attitudes, practices, and beliefs (KAPB) of a population can allow the identification of appropriate messages to be incorporated into a story that entertains and engages the audience, thus rendering the educational tool popular and effective in different age groups [13].
In Côte d’Ivoire, NTDs are of considerable public health relevance [18–20]. Recent data obtained from a national school-based survey with more than 5,000 children aged 5–16 years revealed that 26% of the children were infected with at least one species of parasitic worms. Hookworm was identified as the predominant STH infection (overall prevalence was 17%) with a fairly homogeneous distribution. The other STH infections (A. lumbricoides and T. trichiura) were found with a prevalence of less than 5% each [21].
The aim of the current study was to develop an educational cartoon that might help improve school children’s awareness regarding STHs and diarrheal diseases, and related hygiene practices in Côte d’Ivoire. The study was guided by experiences and lessons from previous investigations in the People’s Republic of China where a health educational video proved useful for lowering the incidence of STH reinfection rates and improving STH case management. In particular, “The magic glasses concept” employed in the educational video proposed by Bieri and colleagues was applied for the current research [12,22].
Ethical clearance for the study was obtained from the ethics committee of Basel (EKBB; reference no. 300/13) and from the ethics committee of the Ministry of Health and Public Hygiene in Côte d’Ivoire (reference no. 76-MSLS-CNER-dkn). Local community and school authorities were visited and informed about the purpose and procedures of the study before commencement of the field activities. Because this study involved the participation of school-aged children, each of them provided a written informed consent signed by their parents or legal guardians prior to enrollment. Children assented orally. Participation was voluntary and children could withdraw any time without further obligations.
The development of “Koko et les lunettes magiques” was done in three distinct phases (Fig 1). The first phase, carried out in four selected villages (two in south-central and two in western Côte d’Ivoire), consisted of preliminary research to assess KAPB of school-aged children (school enrolled and non-enrolled) regarding parasitic worm infections and hygiene, and to identify key messages to be included in the animated cartoon. Phase 2 entailed the cartoon production, which included the drafting of the script and story board and production of the initial version of the animated cartoon. The work was done in the Afrika Toon studios in Abidjan. In phase 3, we focused on school children, and hence, phase 3 took place in eight schools (four in south-central and four in western Côte d’Ivoire). Children from phase 1 were not necessarily part of phase 3. Based on specific feedback received from children, the cartoon was readily adapted and the final version produced. The choice of the two study settings for assessing KAPB and pilot-testing of the cartoon deliberately included different social-ecological contexts and was guided by previous research focusing on NTDs and diarrheal diseases [23–26]. Fig 2 displays the map with the two study settings.
The south-central part of Côte d’Ivoire is culturally very heterogeneous. In the Tiassalé region, there is an indigenous population (Elomoins, Abidjis, Agni, and Abbeys in the sub-prefecture of Tiassalé, and Souamlins, N’gbans, and Didas in the sub-prefecture of Taabo), a foreign-borne population (Baoulés, Attié, Senufo, Gouros, Yacouba, Malinké, and Abrons) and an allogenic population (originally from Burkina Faso, Niger, Nigeria, Ghana, Mali, Togo, and Benin). In the Tiassalé setting, phase 1 of the study took place in two villages, namely Niamoué and Tiassalékro. Phase 3 was carried out in four villages, namely Niamoué, Boussoué, Binao, and Tiassalékro. In this area, the villages have school groups (two schools or more in the same place) but we worked in only one school per village.
In western Côte d’Ivoire four main regions (Cavaly, Guemon, Haut-Sassandra, and Tonkpi), belonging to the district des Montagnes, are populated by four main ethnic groups, including Guéré, Toura, Wobé, and Yacouba. The ethnic composition in western Côte d’Ivoire is more homogeneous compared to the south-central part. Indeed, most people are native ethnic groups of the region. For this study we chose the Man setting in the Tonkpi region that has been the focus of our research activities since the late 1990s. Phase 1 of the study took place in Kogouin and Krikouma, while phase 3 took place in Dompleu, Zê, Krikouma, and Kogouin. Each village is small and has only one school.
The quantitative data collected from the KAPB questionnaires was double entered and cross-checked in EpiInfo version 6.04 (Centers for Disease Control and Prevention; Atlanta, GA, United States of America). Using STATA version 10 (Stata Corporation, College Station, TX, United States of America), frequency tables were generated. Qualitative data gathered from the FGDs were recorded, transcribed into Microsoft Word and then imported into MaxQDA version 1 (VERBI Software Consult; Berlin, Germany) for qualitative data analyses. The data were coded and analyzed to identify the frequency at which coded information and content categories occurred. A triangulation approach was used for analysis of the data derived from the three different methods (questionnaire, FGD, and observation).
The animated cartoon was produced by Afrika Toon, in a transdisciplinary collaboration with the scientific team of the project (epidemiologists, civil engineers, and social scientists) and the communities (school children and teachers; see pilot-testing below). In brief, “Koko et les lunettes magiques” is a 15-minute story about clean water, use of latrines, and hand washing with the aim of improving health- and hygiene-related knowledge and practices towards preventing intestinal worm infections and improving hygiene (see Fig 3 displaying the poster of the animated cartoon). Koko is an 8-year-old boy living in the village Popokro, who is confronted with intestinal worm infections and inappropriate hygiene behavior. Koko himself suffers from intestinal worms. The village doctor provides him magic glasses (lunettes magiques; in French) that allow Koko to see the environment as if he was seeing through a microscope. Hence, Koko sees the parasites and how the transmission to humans occurs (Fig 4). Koko and his friends then decide that they need to take action and inform the community about the dangers related to these infections and come forward with preventive strategies. Toward the end of the animated cartoon, the most important key messages are re-iterated, comparing appropriate and inappropriate behaviors (Fig 5).
Box 1 displays the key messages that were incorporated into the story, following the KAPB results from the preliminary assessment. Particular emphasis was put into using a language adapted to school-aged children. Thus, simple and known terms (e.g., worms, microbes, and parasites) identified during the KAPB assessment, were used instead of less accessible medical terms. The final version of “Koko et les lunettes magiques” is available online (https://www.youtube.com/watch?v=PCNLEK5Ityw).
The pilot testing was conducted in eight schools. Four were in the Tiassalé region (Binao, Boussoué, Niamoué, and Tiassalékro) and four in the Man region (Krikouma, Kogouin, Zê, and Dompleu). All children from grades 3–5 were invited to watch the cartoon. A total of 401 children participated in the KAPB survey.
STH infections, schistosomiasis, and diarrhea are among the most common infectious diseases in LMICs and they cause a considerable global burden [27]. WHO recommends preventive chemotherapy as the key strategy for morbidity control due to STHs and schistosomiasis [7,28]. Indeed, in 2015, an estimated 566.7 million doses of albendazole or mebendazole have been administered to preschool- and school-aged children requiring preventive chemotherapy against STHs and 65.2 million doses of praziquantel have been administered against schistosomiasis [29]. Whenever resources allow, integrated control approaches should be implemented that include access to clean water and improved sanitation, along with health education to change behavior [30,31]. It is conceivable that such integrated control approaches have a considerable impact on diarrheal diseases as well. Building upon the experience from previous work on educational videos that have proven to be effective at improving children’s knowledge and changing their attitudes and behaviors [22,32–34], our aim was to develop a health education tool to improve school children’s awareness regarding STH, schistosomiasis, diarrheal diseases, and hygiene-related practices.
WHO defines health education as follows: “Health education is a process comprising of consciously constructed opportunities for learning and communication designed to improve health information, health literacy, health knowledge and developing life skills which are conducive to the promotion of an individual and community’s health including that of the environment” [35]. The aim of health education is to change human behavior by increasing awareness of the health and social impacts of a disease [17,36]. Of note, the value of moving images in health education has been highlighted by WHO as early as 1988 [37]. When designing a health education tool or intervention, it is essential to understand the baseline knowledge of the target population and to tailor the messaging to local contexts and needs. For characterizing the baseline situation, KAPB studies are a promising approach, using different tools such as questionnaires, FGDs, and direct observations, as done in the current study and elsewhere in Côte d’Ivoire and abroad [13,23]. Findings from KAPB studies allow the identification of appropriate and culturally adapted key messages for integration into health education tools. Thus, the tools can take into account a particular context and the socio-cultural organization of the target population, focusing on precise purposes, e.g., filling identified knowledge gaps in a way that is contextually appropriate and culturally acceptable. For example, although Koko is a school-aged child, we opted to tell a story in the community rather than in a school because a considerable proportion of school-aged children still do not have access to primary school education. Furthermore, still today, the opinions of the family heads and the community elderly are very important in rural Côte d’Ivoire as in many other West African settings. It becomes apparent that telling a story adapted to the local socio-cultural context can maximize the impact of an intervention and is more likely to make change happen. The preliminary assessment allowed us to identify some misconceptions in the school-aged population (e.g., eating sweet fruit as cause of worm disease and diarrhea). Thus, we could address this issue in a key message of the animated cartoon, by emphasizing that consumption of fruit is healthy, although they need to be washed beforehand.
We observed specific differences in the knowledge of children regarding the transmission of parasitic worm infections. Indeed, in some communities, the majority of children believed that they were at risk and were afraid of contracting parasitic worms because of their behaviors (i.e., sweet fruit consumption, drinking unsafe water, eating dirty and spoiled food, and playing outdoors without wearing shoes). In contrast, in other communities, a large proportion of children were not afraid because they reported to not put themselves at risk and if they did contract parasitic worms, it was possible to remedy this situation through access to deworming drugs. The latter part of children were those living in villages close to the towns of Man and Tiassalé, thus we conjecture that they had better access to information and treatment and thus were less afraid of parasitic worm diseases. This emphasizes the importance of assessing the KAPB in a representative sample of the target population before developing a health educational tool.
“Koko et les lunettes magiques” targets a very specific population, namely school-aged children. Indeed, the main character of the animated cartoon is an 8-year-old boy and the story is designed in such a way that the audience (children) can identify themselves with Koko. Nonetheless, other adult characters (Koko’s father, a nurse, a wise doctor, an old man, and Koko’s friends) are also integrated in the story. Hence, adults can also identify themselves with the story told. The animated cartoon was developed for school-aged children and discussions are underway with the education sector how to integrate it into the school curriculum. However, the cartoon can also be screened at public places for entertaining and educating adults, since it simplifies complex issues.
“Koko et les lunettes magiques” can be seen as an entry point of a larger health education intervention package that will allow rapidly attracting children’s (and adults’) attention. However, it is essential that messages are reiterated during discussions, drawing assignments or group work (the type and combination of these methods would depend whether the intervention is school-based or community-based) in order to consolidate the knowledge gained and to make sure that the most important messages are understood correctly. Indeed, the pilot testing revealed that although children retain most of the key messages right after the screening [38], some knowledge gaps still remain. Furthermore, repetition of key messages should ideally be done not only right after the screening but also over time, so that messages can be retained over the longer term. Moreover, it could be of great value to involve teachers by increasing their urge to make sure their schools have available clean water for drinking and hand washing facilities as well as soap and properly maintained hygienic facilities for defecation. Teachers can also be trained to regularly update their students’ knowledge regarding hygiene- and health-related practices. Such interventions could be supplemented by participation of parents and the community, thus encouraging them to take an active role in the health education of their children [13]. When the aforementioned issues are considered, the decision to adopt a good health practice, however, is also matter of individual choice. The health believe model presented by Rosenstock in 1974 assumes that people will only act to prevent a disease if they have minimal knowledge of health, whereas health needs to be an important dimension in their life. The perception of the threat to health and belief in the effectiveness of the action to reduce this threat will determine a preventive action [39–40].
In conclusion, “Koko et les lunettes magiques” was developed to educate school-aged children in an entertaining and context-specific manner. Clearly, children liked the animated cartoon and the pilot testing revealed that children could keep the transmitted information shortly after they were shown the cartoon. In a recent intervention study in 25 schools of western Côte d’Ivoire the cartoon has been tested for its effect on KAPB in the short term and results of the study will be presented elsewhere. An important next step will be to validate the cartoon on its effect on the KAPB regarding STH infections, schistosomiasis, and diarrheal diseases in the mid- and long-term. It is essential to validate such health education tools in different social-ecological contexts in order to create evidence-based and locally adapted designs for sustainable integration of health educational packages into school curricula along with regional deworming programs against STHs, schistosomiasis, and other NTDs.
|
10.1371/journal.pntd.0006846 | Indoor residual spraying for kala-azar vector control in Bangladesh: A continuing challenge | Visceral leishmaniasis (VL) in the Indian subcontinent is a fatal disease if left untreated. Between 1994 to 2013, the Ministry of Health of Bangladesh reported 1,09,266 cases of VL and 329 VL related deaths in 37 endemic districts. Indoor residual spraying (IRS) using dichlorodiphenyltrichloroethane (DDT) was used by the national programme in the 1960s to control malaria. Despite findings of research trials demonstrating that the synthetic pyrethroid deltamethrin 5 WP was very effective at reducing vector densities, no national VL vector control operations took place in Bangladesh between 1999 to early 2012. In 2012, IRS using deltamethrin 5 WP was re-introduced by the national programme, which consisted of pre-monsoon spraying in eight highly endemic sub-districts (upazilas). The present study aims to evaluate the effectiveness of IRS on VL vectors, as well as the process and performance of the spraying activities by national programme staff.
Five highly endemic upazilas of Mymensingh district were purposively selected (Fulbaria, Trishal, Mukthagacha, Gaforgaon and Bhaluka) to conduct the present study using the WHO/TDR monitoring and evaluation tool kit. IRS operations, conducted by 136 squads/teams, and 544 spraymen, were observed using check lists and questionnaires included in the WHO/TDR monitoring and evaluation tool kit. A household (HH) acceptability survey of IRS was conducted in all study areas using a structured questionnaire in 600 HHs. To measure the efficacy of IRS, pre-IRS (two weeks prior) and post-IRS (at one and five months after), vector density was measured using CDC light traps for two consecutive nights. Bioassays, using the WHO cone-method, were carried out in 80 HHs (40 sprayed and 40 unsprayed) to measure the effectiveness of the insecticide on sprayed surfaces.
Of the 544 spraymen interviewed pre-IRS, 60%, 3% and 37% had received training for one, two and three days respectively. During spraying activities, 64% of the spraying squads had a supervisor in 4 upazilas but only one upazila (Mukthagacha) achieved 100% supervision of squads. Overall, 72.8% of the spraying squads in the study upazilas had informed HHs members to prepare their houses prior to spraying. The required personal protective equipment was not provided by the national programme during our observations and the spraying techniques used by all sprayers were sub-standard compared to the standard procedure mentioned in the M&E toolkit. In the HH interviews, 94.8% of the 600 respondents said that all their living rooms and cattle sheds had been sprayed. Regarding the effectiveness measurements (i.e. reduction of vector densities), a total of 4132 sand flies were trapped in three intervals, of which 3310 (80.1%) were P. argentipes; 46.5% (1540) males and 53.5% (1770) females. At one month post-IRS, P. argentipes densities were reduced by 22.5% but the 5 months post-IRS reduction was only 6.4% for both male and female. The bioassay tests showed a mean corrected mortality of P. argentipes sand flies at one month post-IRS of 87.3% which dropped to 74.5% at 4 months post-IRS in three upazilas, which is below the WHO threshold level (80%).
The national programme should conduct monitoring and evaluation activities to ensure high quality of IRS operations as a pre-condition for achieving a fast and sustained reduction in vector densities. This will continue to be important during the maintenance phase of VL elimination on the Indian subcontinent. Further research is needed to determine other suitable vector control option(s) when the case numbers are very low.
| The visceral leishmaniasis (VL) elimination programme was launched in the Indian subcontinent (Bangladesh, India and Nepal) in 2005. Although the integrated vector management (IVM) system is one of the important elements highlighted in the Regional VL elimination strategy, indoor residual spraying (IRS) is the sole intervention practice that has been implemented. In fact, in Bangladesh from 1999 to early 2012, no VL vector control was used at all and pre-monsoon IRS was only re-introduced by the national programme in eight high endemic upazilas (sub-districts) in 2012. The present study monitored IRS operation in five upazilas (Fulbaria, Trishal, Mukthagacha, Gaforgaon and Bhaluka). Monitoring took place with the help of using observation check lists and questionnaires included in the WHO/TDR monitoring and evaluation tool kit. The study identified that training of spraymen was insufficient and a supervisor was not always present during spraying. The spraying techniques by all the sprayers were sub-standard. It was also found that all the required personal protective equipment was not provided by the national programme. It is recommended that the national programme should conduct monitoring and evaluation activities to ensure high quality of IRS operations in order to achieve maximum benefit.
| Visceral leishmaniasis (VL), also known as kala-azar (KA) in the Indian subcontinent (ISC), is a fatal disease if left untreated [1]. In the ISC, the disease is transmitted exclusively by the sand fly vector Phlebotomus argentipes [2]. Leishmania donovani is the protozoan causative agent. Visceral leishmaniasis is affecting marginalized communities worldwide [1,3]. Within the ISC, VL is highly prevalent in Bangladesh, India and Nepal; and some sporadic cases were reported from Bhutan [4]. Among the global burden, 90% of all VL cases are reported by only six countries (India, Bangladesh, Sudan, South Sudan, Ethiopia and Brazil) and it is estimated that 200,000 to 400,000 cases and 20,000 to 40,000 deaths occur annually on a global scal [3]. The first historical report of VL was from Jessore district, currently located in the south-western part of Bangladesh, where an epidemic outbreak killed an estimated 75,000 people between 1824 and 1827 [5]. In the 1960s, a comprehensive malaria eradication programme (MEP) was launched using dichlorodiphenyltrichloroethane (DDT) for indoor residual spraying (IRS) for malaria vector control. Visceral leishmaniasis was virtually eradicated at that time, possibly as a beneficial by-product of the MEP. However, with the discontinuation of MEP, VL re-emerged in the early 1980s [6,7]. DDT was later banned entirely by the Government of Bangladesh in 1998 as it was an environmental hazard [8].
In 2005, a memorandum of understanding (MoU) was signed among three countries (Bangladesh, India and Nepal) in the ISC, with the aim to eliminate VL as a public health problem across the region by 2015 [9]. This date was later extended to 2017 with the inclusion of Bhutan and Thailand in the consortium [10]. The target was to reduce the VL incidence to less than one case per 10,000 of the population annually at the upazila level in Bangladesh [9]. Between 1994 to 2013, the Directorate General of Health Services (DGHS) of Bangladesh reported 1,09,266 cases of VL and 329 VL related deaths from 37 endemic districts [7]; with about 50% of total VL cases being reported from five upazilas in the Mymensingh district. In the period after signing the MoU, extensive research activities were conducted to reduce VL transmission and provide treatment in coordination with the Special Programme for Research and Training in Tropical Diseases (TDR) at World Health Organization (WHO). This lead to a slow, but continued, reduction of VL cases until the elimination target was achieved and the first steps of the maintenance phase were initiated in 2017 [11].
Integrated vector management (IVM) is one of the most important elements in the regional elimination strategy [9]. However, for a significant period during the attack phase of the programme, from 1999 to early 2012, no VL vector control activities were performed in Bangladesh [7,12], and the number of cases increased every year during that time. A research programme in Bangladesh, India and Nepal showed that IRS using the synthetic pyrethroid (SP) deltamethrin 5WP was very effective against the VL vector if performed in accordance with guidelines [13,14]. Based on this, a pilot IRS trial was conducted in 2011 by the national programme using deltamethrin 5 WP, which proved to be successful. In 2012 a scale-up IRS operation was undertaken in eight highly endemic upazilas in four districts [7]. The Ministry of Health and Family Welfare, Government of Bangladesh organized several training courses for spraymen and supervisors. In parallel, WHO/TDR, along with programme managers and local researchers, developed a monitoring and evaluation toolkit for IRS [15]. This comprehensive toolkit should be used in order to provide an evidence base and explain possible reasons affecting the efficacy of IRS [16]. The present study will use this toolkit to monitor and document activities of IRS operations conducted by the national programme in Bangladesh using deltamethrin 5 WP.
In 2012, pre-monsoon (May-June) IRS was conducted in eight highly endemic upazilas, namely: Fulbaria, Trishal, Mukthagacha, Gaforgaon, Bhaluka of Mymensingh district (Fig 1), and Nagarpur, Madarganj, Terokhada upazilas of Tangail, Jamalpur and Khulna districts, respectively. The study was carried out from March to October of 2012 and included the following activities:
All field and laboratory data was checked, verified, cleaned of obvious errors and entered into databases using Microsoft Office Access 2007. The data analyses was performed using SPSS version 25 (SPSS Inc, Chicago, IL) and Stata version 12 (Stata Corp, College Station, TX). Descriptive statistics were used to explore the nature of the data. Frequency and percentages were used to analyze all the datasets (observation of spray team, household acceptability surveys, vector density measurement and bioassay). Due to the over-dispersed vector density measurements, the nonparametric Mann-Whitney U test was used to test whether there was a significance difference between intervention and control groups in between pre-intervention and post-intervention measurements. Wilcoxon Signed Rank test for paired sample was used to test whether there was any significance difference between baseline and follow-up. For each post-IRS timepoint, the percentage reduction in sand fly counts attributable to IRS was calculated as 100 [(mean post-intervention value for the intervention group-mean baseline value for the intervention group)-(mean post-intervention value for the control group-mean baseline value for the control group)]/(mean baseline value for the intervention group). To investigate the effect of vector density reduction we used the following formula:
Reductionrate(%)=(I¯−I¯base)−(C¯−C¯base)I¯base×100
Where, I¯base = average number of baseline sand fly in intervention group; C¯base = average number of baseline sand fly in control group; I¯ = average number of post intervention sand fly in intervention group; C¯ = average number of post intervention sand fly in control group. The effect was negative and positive if the sand fly density decreased or increased post-intervention, respectively.
The study protocol (ID: BMRC/NREC/2010-2013/535) was approved by the ethical review committees of the Bangladesh Medical Research Council (BMRC), Dhaka, Bangladesh. On behalf of the national programme, Director Disease Control, DGHS had kindly given permission to observe spraying activities and interview spraying squads. Written informed consent was obtained from the head of each study household for active participation in the study.
Each squad consists of five people: four spraymen and one squad leader. We observed IRS operations of 136 squads (544 spraymen and 136 squad leaders), in five endemic upazilas of Mymensingh district. Table 1 shows that 59.0%, 3.1% and 37.4% of the 544 spraymen received training for one, two and three days, respectively, before the IRS operation but all 136 squad leaders had training for three days. Only 14.0% of spraymen were involved in previous IRS activities. Regarding the procedures to approach houses for spraying, about 3.0% of squad leaders reported that there were only a few HHs who did not want to get their houses sprayed. Forty nine percent of squad leaders informed that they found some houses locked when they arrived. Almost all squads (98.5%) reported that they washed their spray pump at the end of the day and 72.8% of spraymen knew how to handle left-over insecticide or did not have any left-over (20.6%) [Table 1].
All the respondents in the 600 HHs of the five study upazilas mentioned that their houses had been sprayed with insecticide [Table-3] and 94% said that their living rooms and cattle sheds had also been sprayed. Only 36.2% of interviewees mentioned that they had been informed about IRS in advance. The majority of respondents (85.3%) said that they were happy with the IRS activities. Of the 15% respondents who were unhappy, the main source of complaint (76% of this subgroup) was that they did not recognize that the spraying killed any insects. Although the majority of the respondents (81.7%) received advice to remove or cover clothes, food/utensils, children and animals prior to spraying, only 13.3% reported that they were instructed not to enter the house after spraying was completed. Only 7.2% of respondents had been informed about possible side effects (itching, burning of skin, dizziness, cough, etc.) of the spraying and 91.3% reported that they were not advised to abstain from re-plastering (mud wall) or re-painting (pacca/tin house) their houses after spraying [Table 3].
Table 4 shows the absolute numbers, species, and physiological status of sand flies collected in the households in both treatment groups (IRS and control) of the study. A total of 4132 sand flies were trapped in three time intervals (at baseline, 1 month and 5 months post-intervention), of which 3310 (80.1%) were P. argentipes and 822 other species [Table 4]. Of all P. argentipes, 46.5% (1540) were males and 53.5% (1770) females; including 20.1% (666) gravid females and only 2 blood-fed females.
The average numbers of P. argentipes per household per night per trap in IRS and control arms were 4.36 and 8.31 for both (combined) male and female, 2.29 and 3.47 for only male, 2.71 and 5.12 for only female, 0.60 and 2.19 for gravid female respectively [Table 5]. At baseline (pre-intervention), the number of P. argentipes was significantly lower in the IRS treatment group compared with the control group in both male and female (p < 0.05), only female (p<0.05), and gravid female (p<0.01) whereas no significance difference was observed in only male (p = 0.177). Similarly, at 1 month post-IRS, there was a significant difference in P. argentipes densities between the treatment groups for both male and female (p<0.05) and for only male (p<0.05). But no significant difference was found for only female (p = 0.318) and gravid female (p = 0.106). At 5 months post-IRS, no significant difference in P. argentipes densities was observed between treatment groups for both male and female (p = 0.282), for only male (p = 0.350), for only female (p = 0.261) and for gravid female (p = 0.264). This indicates that IRS is able to control the increase of only male P. argentipes density in the IRS areas up to one month post spraying (at first follow-up). But it fails to control the P. argentipes density in IRS houses at five months (second follow-up). Table 5 shows that one month after IRS the P. agrentipes density was reduced by 22.61% and 118.79% in both (combined) male and female, and only male which dropped to 6.37% and 53.94% at five months respectively. However, no reduction was found in only female and gravid female P. argentipes density.
Fig 2(A) shows that the average number of P. argentipes sand flies per household per night per trap increase in both IRS and control treatment groups after one month post-IRS but the increase was more pronounced in the control group for both (combined) male and female. Similar increases happened in female and gravid female P. argentipes sand flies as well but the increase was not pronounced in both IRS and control treatment groups. But reduction of density was observed only in male P. argentipes sand flies in IRS treatment group and was increased in the control treatment group [Fig 2(A)]. There is no significant difference between the treatment groups compared with their baseline measurements apart from male P. argentipes sand flies. At 5-months post-intervention, there was no significant difference between the two treatment groups from baseline for all categories of P. argentipes sand flies [Fig 2(B)]. On an average, one P. argentipes sand flies increased in the sprayed areas per HH per night per trap compared to baseline at first follow-up which is two in control areas for both (combined) male and female. In sprayed areas, only male P. argentipes sand flies density reduced on an average 0.5 compared to baseline, which was 2.5 increases in the control areas at first month of IRS whereas only female increase was 1.5 and about one increase for gravid female [Fig 2(B)].
Bioassays was conducted in three upazilas, namely: Fulbaria, Mukthagacha and Trishal. The mean corrected mortality of female P. argentipes sand flies was 87.3% (CI = 82.3%–92.3%) and 74.5% (CI = 69.3%–79.7%) at one and four months after spraying [Fig 3] showing that the insecticidal effect rapidly disappeared.
Spraymen and squad leaders are hired temporarily according to the requirements of the sub-district and in accordance with the micro-action plan of IRS. Spraymen training is compulsory prior to the IRS operation. The present study identified that only 37.5% of spraymen had received a full training in accordance with the national guidelines. As a result, the overall performance of house spraying was poor. Issues with the late release of fund, low daily wage of spraymen and squad leaders, and local pressure are hindering the timely recruitment of spraymen and squad leaders. Due to this, the national programme is facing challenges when training the spraymen. As no IRS was conducted for VL vector control from 1998 to early 2012 in the country [7,12], when DDT was banned [8], more than 85% of spraymen were never involved in IRS activities before. Though the acceptance of IRS activities was found to be very high (97%) in the study areas, more than 50% spraying squads found that there was no one home while they were in the village. This points to a communication gap between the community and the service provider. Often, the quality of IRS could not be ensured by proper supervision: the present study established that less than 44% of supervisors were present in Fulbaria and Trishal upazilas during the spraying activities [7]. Over- or under-concentration of the insecticide solution depends on the correct filling of the pumps; proper mixing of the insecticide and correct spraying technique ensures the uniform distribution of insecticides on sprayed surfaces. We found these two indicators to be substandard in all upazilas and extremely poor in two particular upazilas. Personal protective equipment is key to shield a person from any occupational hazard. However, as these were not distributed, we observed that none of the squad members were wearing boots. The use of goggles and gloves were also found to be low in all sites due to weak supervision. Though adequate training and proper supervision are essential for ensuring proper spraying, we observed substandard spraying techniques (ranges from 34.2%– 52.9% correct technique) except for spraying from bottom to top. Marking of sprayed houses is important for their supervision and identification; however, we found poor compliance with the guidelines, apart from mentioning the team number (about 84%).
Despite the many limitations, the communities positively accepted the IRS operations as they had not experienced any IRS or other vector control activities for a long time [7,12]. Understandably, people in the endemic communities do not want to suffer from insect nuisance, particularly VL, so they welcomed IRS activities because ‘something is better than nothing’. A previous study described how community members suffer from loss of earnings and how the symptoms of VL affect their wellbeing; therefore poor householders resorted to desperate actions, like the sale or rent of their assets, taking out loans, etc. to cover the expenditures of diagnostics and treatment [23]. The national programme should ensure that high quality IRS is delivered to get the maximal results. Substandard application of insecticides is also a potential cause of insecticide resistance since sub-lethal doses or poor quality insecticides enhance resistance development. Several studies have reported DDT resistance in P. argentipes [16,24,25] and it is essential that synthetic pyrethroids remain effective.
IRS is highly effective in quickly reducing vector densities underlining the need to continue the operations in all three countries of the ISC even during the consolidation and maintenance phases of VL elimination. Based on our study results the following recommendations can be made:
IRS is a huge operation where each step needs to include training of human resources as well as proper management. The experienced people involved in the malaria eradication programme carried out in the 1970s had all retired prior to the VL elimination programme so the national programme faced challenges to prepare a new cohort of spraymen and supervisors to handle such a complex operation.
In Bangladesh, the Director of the Central Medical Store Deport (CMSD), DGHS under the Ministry of Health and Family Welfare procured the insecticide and distributed the required quantity to the local hospitals (Upazila Health Complexes, UHCs). The Director of Disease Control, DGHS is responsible for conducting the national IRS operation through district managers (Civil Surgeon) and sub-districts managers. Filed operations of IRS in the respective UHC are managed by Upazila Health & Family Planning Officer (UH&FPO).
The responsibility for supervision and monitoring of IRS is to be done by the following officers:
Integrated vector management is one of the most important elements mentioned in the regional VL elimination strategy. IRS is an effective vector control tool but it is an expensive operation. Therefore, it is recommended that the national programme ensure: (1) procurement of quality insecticide, (2) proper training of human resources involved in IRS operation, (3) proper monitoring and supervision during spraying, (4) regular vector surveillance and bioassays on sprayed surfaces, (5) routine testing of vector susceptibility and (6) community sensitization.
Since the launch of the VL elimination programme in 2005, this is the first study conducted in Bangladesh where national IRS operation was monitored by a third party (researcher) and substantial information was generated out of it. Bangladesh, along with other countries in the region, are still heavily relying on IRS so we strongly feel this information can guide national programmes for conducting well coordinated IRS operations. A limitation of the study includes the condition that we had to comply with the micro action plan made by the sub-district managers (UH&FPO) for all monitoring visits as we monitored their IRS activities. In such a case, there could have been a chance of leakage about the monitoring visits to IRS squads from beforehand. To avoid such circumstances we did not disclose our detailed plan of visits to local health authorities so that the quality and validity of data were ensured.
In conclusion, the national programme should routinely apply monitoring and evaluation activities to ensure high quality of IRS operations as a pre-condition for achieving a fast and sustained reduction of vector densities. This will continue to be important during the maintenance phase of VL elimination in the Indian subcontinent. Further research is needed to determine other suitable vector control option(s) when the case numbers are very low.
|
10.1371/journal.pgen.1007829 | SPOC domain-containing protein Leaf inclination3 interacts with LIP1 to regulate rice leaf inclination through auxin signaling | Leaf angle is an important agronomic trait and influences crop architecture and yield. Studies have demonstrated the roles of phytohormones, particularly auxin and brassinosteroids, and various factors in controlling leaf inclination. However, the underlying mechanism especially the upstream regulatory networks still need being clarified. Here we report the functional characterization of rice leaf inclination3 (LC3), a SPOC domain-containing transcription suppressor, in regulating leaf inclination through interacting with LIP1 (LC3-interacting protein 1), a HIT zinc finger domain-containing protein. LC3 deficiency results in increased leaf inclination and enhanced expressions of OsIAA12 and OsGH3.2. Being consistent, transgenic plants with OsIAA12 overexpression or deficiency of OsARF17 which interacts with OsIAA12 do present enlarged leaf inclination. LIP1 directly binds to promoter regions of OsIAA12 and OsGH3.2, and interacts with LC3 to synergistically suppress auxin signaling. Our study demonstrate the distinct effects of IAA12-ARF17 interactions in leaf inclination regulation, and provide informative clues to elucidate the functional mechanism of SPOC domain-containing transcription suppressor and fine-controlled network of lamina joint development by LC3-regulated auxin homeostasis and auxin signaling through.
| Leaf angle is a major trait of ideal architecture of crops that associates with photosynthetic efficiency and yields. Studies of the underlying mechanism will greatly help to improve the crop yield. Phytohormones especially auxin and brassinosteroids play crucial roles in regulating the leaf inclination, however, the upstream regulatory network is still unknown. Here, we functionally characterize a novel SPOC domain-containing protein LC3 (leaf inclination3) in lamina joint development through regulating auxin signaling. LC3 deficiency results in the excessive cell elongation at lamina joint adaxial side and hence the enlarged leaf angle. LC3 acts as a transcription suppressor through interacting with LIP1 (LC3-interacting protein 1, a HIT zinc finger domain-containing transcription factor), which directly binds to the promoters of auxin signaling and homeostasis related genes. Our studies provide new insights in the functional mechanism of SPOC domain-containing proteins and help to elucidate how auxin signaling is regulated during lamina joint development.
| Rice is one of the most important crops in the world and breeding rice varieties with ideal architecture is a vital strategy for improvement of grain yields [1, 2]. Leaf is the main organ for photosynthesis and its development is crucial for the yield potential. Leaf inclination indicates the angle between leaf blade and culm [3], and studies have shown that erect leaf facilitates the penetration of sunlight and enhances the photosynthetic efficiency [4, 5], which is suitable for dense planting. Unbalanced development of collar cells at adaxial or abaxial sides, development of mechanical tissue and mechanical strength, formation of vascular bundle and cell wall compositions also have been pointed out to affect the leaf angle [2, 3, 6, 7].
Recent systemic analysis of the dynamic developmental processes of lamina joint through cytological observation showed that cell differentiation, division and elongation, cell wall thickening, and programmed cell death (PCD), are closely correlated with leaf angle and regulated by a complex network, consisting of various factors, especially protein kinases and hormones [8]. Indeed, studies by using mutants or transgenic approaches indicate that altered biosynthesis or signaling of brassinosteriods (BRs) lead to the change of leaf inclination, such as BR-deficient mutant dwarf4-1 [9], ebisudwarf (d2) [10], dwarf1 (brd1) [5], BR signaling mutant d61-1, 2 (weak mutant alleles of OsBRI1) [11], or rice plants with reduced expression of OsBZR1 [12]. Similarly, plants with suppressed auxin signaling by overexpressing miR393a/b that suppress expression of receptor OsTIR1 [13], or with reduced auxin levels including mutant lc1 [6] or plants overexpressing GH3 family members OsGH3.2, OsGH3.5, and OsGH3.13 [7, 14, 15], present increased leaf inclination. It is noticed that BR stimulates while auxin suppresses the leaf inclination through regulating the cell division or elongation at adaxial side of lamina joint and auxin coordinates with BR to control the lamina joint development [6, 7, 16]. In addition, ethylene may participate in BR-induced leaf inclination [17] and repressed expression of a gibberellin signaling negative regulator, SPINDLY, leads to increased leaf angles [18].
Many transcription factors (TFs) are involved in leaf inclination regulation. OsWRKY1 and MADS-box proteins OsMADS22, OsMADS55 and OsMADS47 negatively regulate leaf inclination [19–21]. Ectopic expression of LAX PANICLE (LAX), a basic helix-loop-helix TF, leads to increased bending of lamina joint [22]. Deficiency of OsLIGULELESS1 (OsLG1, a SBP domain-containing TF) results in defects in auricle, ligule, and lamina joint [23]. Recently, a genome-wide association study shows that rice TFs OsbHLH153, OsbHLH173 and OsbHLH174 involve in flag leaf angle regulation [24]. In addition, the Aux/IAA family members interact with AUXIN RESPONSE FACTOR (ARFs) to suppress auxin signaling [25] and regulate the leaf blades [26]. Overexpression of OsIAA1, OsIAA4, OsARF19 lead to the increased leaf angle [7, 16, 27], while deficiency of OsARF11, an ortholog of Arabidopsis ARF5, results in reduced leaf angle [28]. However, the detailed mechanism, especially how Aux/IAA is regulated during lamina joint development and which distinct Aux/IAA-ARF interaction regulates leaf inclination is unknown yet.
By systemic analysis of a rice mutant with enlarged leaf angle, we showed that leaf inclination3 (LC3), a SPOC domain-containing protein that is speculated to facilitate protein-protein interactions in transcription repression complex [29], interacts with a HIT zinc finger domain-containing TF LIP1 (LC3-interacting protein 1) to suppress the auxin signaling and homeostasis genes, hence to regulate the cell elongation at adaxial side of lamina joint and thus leaf inclination. These results provide informative clues on the fine-controlled network regulating lamina joint development.
Our previous studies by analyzing the global transcriptome of developing lamia joint showed that gene leaf inclination3 (LC3, Os06g39480) is down-regulated from stage 2 to stage 6 during lamina joint development. Further analysis of the corresponding knockout mutant, lc3, revealed the obviously increased leaf angles under LC3 deficiency [8]. LC3 encodes a novel SPOC domain-containing protein and the underlying functional mechanism is thus detailed studied.
Analysis of the transcription pattern of LC3 by quantitative real-time RT-PCR (qRT-PCR) confirms the reduced expression of LC3 along with lamina joint development, while LC3 is relatively highly expressed in pistil, spikelet and seeds at early stage (Fig 1A). Further promoter-reporter gene fusion analysis (a 3-kb promoter regions of LC3 was fused to the β-glucuronidase gene) consistently show that LC3 is transcribed at adaxial side of lamina joint, glume and pollen, pistil and seeds (Fig 1B). Based on the significantly decreased LC3 expression in lc3 [8], transgenic lc3 plants with complemented expression of LC3, driven by its native promoter, was generated (Fig 1C, left panel). Phenotypic observation and measurement show the restored leaf inclination (Fig 1C), which confirms the role of LC3 in regulating leaf inclination.
To clarify the cytological change of lc3 mutants, lamina joint paraffin section was conducted. Observations of the longitudinal sections show the increased cell width, while unaltered cell number and cell length of the second layer parenchyma cells at adaxial side of lc3 lamina joint (Fig 1D). Further observations of transverse sections consistently show the increased cell length and unaltered cell layer numbers at adaxial side (Fig 1E), and no change of cell length and cell layers at abaxial side of lc3 mutant lamina joint (S1 and S2 Figs). Overall, the excessive cell elongation at adaxial side of lamina joint results in the enlarged leaf angle of lc3.
Auxin and brassinosteroids play crucial roles in regulating lamina joint development and thus leaf inclination. To investigate the functional mechanism of LC3, expression level of auxin and brassinosteroids signaling related genes and some reported genes regulating leaf angles were examined. qRT-PCR analysis reveals the decreased level of auxin signaling related genes ARF2, IAA6, IAA9, unaltered expressions of BR-related genes, increased expressions of LAZY1 [30] and TAC1 [31], and interestingly, dramatically increased levels of OsIAA12 and OsGH3.2 in lc3 mutant (Fig 2A). Previous studies showed that overexpression of OsGH3.2 did result in the increased leaf angles [14], similar to LC1 (OsGH3.1) overexpressing plants [6]. We thus focus on the effect of OsIAA12 and relevant regulatory mechanism.
Transgenic rice plants overexpressing OsIAA12 driven by a maize ubiquitin promoter were generated and analysis of the positive lines (Fig 2B, left panel) at 10 days after heading showed that OsIAA12 overexpression indeed leads to the increased leaf inclination (Fig 2B). Analysis of the longitudinal sections of lamina joint reveals the increased cell width of the second layer parenchyma cells at adaxial side of OsIAA12-overexpressing lines (Fig 2C), which is same to that of lc3. Consistently, observations of the transverse sections of flag leaf lamina joint show that though there is no change in cell layer numbers in adaxial or abaxial regions (S3 and S4 Figs), increased length of adaxial cells was detected (Fig 2D). These results indicate that LC3 regulates lamina joint development possibly through OsIAA12 and auxin signaling.
AUX/IAA proteins interact with ARFs to suppress the auxin signaling, and which ARF cooperates with OsIAA12 to regulate the leaf inclination distinctly is studied. Previous studies on the IAAs-ARFs interacting networks indicated the interaction between OsIAA12 and OsARF17 [32], which was confirmed by the yeast two-hybrid assays (Fig 3A). Further analysis by Split-YFP assay through expressing N-terminal YFP fused OsIAA12 (nYFP-OsIAA12) and C-terminal YFP fused OsARF17 (OsARF17-cYFP) in tobacco leaf epidermal confirmed the OsIAA12-OsARF17 interaction in nucleus in planta (Fig 3B). Being consistent, transient expression of OsIAA12-RFP and OsARF17-GFP fusion proteins in rice protoplasts showed that OsIAA12 co-localizes with OsARF17 in nucleus (Fig 3C).
To confirm the role of OsIAA12-OsARF17 interaction in leaf inclination regulation, plants deficiency of OsARF17 were generated by Crispr/Cas9 approach (OsARF17-Cas9). Six independent transgenic lines were obtained and four of them were homozygous with either insertion or deletions at 5’ end of OsARF17 (Fig 3D, upper panel). Phenotypic observations and measurement of T2 generations showed obviously enhanced flag leaf angles (Fig 3D, bottom panel). Analysis of paraffin section revealed similar cytological change as lc3 mutant and transgenic plants overexpressing OsIAA12 (Fig 2C and 2D; S3 and S4 Figs), suggesting that LC3 regulates cell elongation at adaxial side and leaf inclination through suppressing the OsIAA12 expression, which regulates OsARF17 effects by protein interaction. Examination of the transcriptions of OsIAA12 and OsGH3.2 showed the suppressed expression of OsIAA12 and OsGH3.2 under LC3 overexpression and restored expression in lc3 plants with complemented expression of LC3 (Fig 3E), further confirming the regulation of OsIAA12 and OsGH3.2 by LC3. In addition, expression of OsARF17 is unaltered under OsIAA12 or LC3 overexpression (S5 Fig), which is consistent with that Aux/IAA proteins repress ARF transcription factors via direct protein-protein interaction.
LC3 encodes a SPOC-domain containing protein and localizes widely in cells (mainly in nucleus, Fig 4A). Previous reports showed that Spilt ends (Spen) protein family members compose an N-terminal RNA recognition motifs (RRM) domain and a conserved C-terminal SPOC domain [33, 34]. RRM domain regulates chromatin modification by recognizing and binding to DNA/RNAs specifically [34], while SPOC domain is proposed to facilitate protein-protein interactions in the transcription repression complex [29]. However, the underlying mechanism is unclear yet. In animals, Spen family members are reported to involve in neuron development, immune responses [29] and sex determination [35], which is less clarified in plants.
Phylogenetic analysis shows that there are three identified proteins close to LC3 (S6 Fig), including Arabidopsis FPA that controls flowering time [36]. OsRRMh and OsRRM, two rice homologues of AtFPA, that control flowering, fertility, and architecture [37, 38]. Protein structural analysis shows that compared to OsRRM, OsRRMh and AtFPA, the conserved RRM domain is absent in LC3 (Fig 4B), suggesting the distinct function of LC3. It is speculated that LC3 possibly interacts with other factors, which help to recognize DNA or RNA sequence and cooperate with LC3 to repress the transcription of downstream genes.
Yeast two-hybrid screening was thus conducted to isolate the candidate proteins that interact with LC3. Four proteins possibly interacting with LC3 were identified and designated as LIPs (LC3-interacting proteins). Further analysis confirmed the interaction between LC3 and LIP1, a HIT zinc finger domain-containing protein (Fig 4C). Transcription pattern analysis showed that LIP1 presents similar expression pattern as LC3 during lamina joint development (Fig 4D). Observation of fluorescence in rice protoplasts expressing LIP1-GFP/LC3-YFP revealed the similar localization of LIP1 and LC3 (Fig 4E). Furthermore, transient expression of LC3-RFP and LIP1-GFP fusion proteins in rice protoplasts showed that LC3 co-localizes with LIP1 both in nucleus and cytoplasm (Fig 4F). Split-Luciferase assay confirmed the interactions between LIP1 and LC3 in vivo (Fig 4G), indicating that LIP1 may coordinate with LC3 to regulate the leaf inclination.
As LC3 lacks the RRM domain, it is hypothesized that LC3 may repress the downstream genes OsIAA12 and OsGH3.2 through interacting with LIP1, which recognize the binding sequences. Yeast one-hybrid analysis of OsIAA12 promoter (fragments -1710 to 0 before ATG) showed that LIP1 binds to later region (-914 to 0 before ATG) but not the forward one, and LC3 binds to neither region (Fig 5A). In addition, by using 10-day-old transgenic seedlings expressing LC3-GFP, analysis of chosen five fragments in later region of OsIAA12 promoter by quantitative chromatin immunoprecipitation (ChIP)-PCR indicated the enrichment of four DNA fragments (Fig 5B), confirming that LC3 binds to OsIAA12 promoter through interacting with LIP1.
To further confirm the repression effect of LC3-LIP1 on downstream genes, two effector constructs carrying LC3 and LIP1 fusion GFP, were transiently expressed with a luciferase reporter (LUC) construct containing ~2.7-kb promoter of OsIAA12 in rice protoplasts. Measurement showed that LUC expression was significantly reduced in the presence of LC3 or LC3-LIP1, and no differences in the presence of single LIP1 (Fig 5C), suggesting that LIP1 alone does not present inhibition effect, and LC3 and LIP1 cooperatively suppress the expressions of downstream genes. Similarly, ChIP-PCR assays of different DNA fragments in OsGH3.2 promoter showed the enrichment of two DNA fragments (Fig 5D), indicating the binding of LC3 to OsGH3.2 promoter as well. These results suggest that LIP1 orchestrates with LC3 to repress the OsIAA12 and OsGH3.2 expressions.
There is no change of leaf angles under LC3 overexpression (S7 Fig), indicating that LC3 functions to maintain the normal leaf inclination. To testify the function of LIP1 in lamina joint development, plants deficiency of LIP1 in background of LC3 overexpression were generated by Crispr/Cas9 approach (LIP1-Cas9 in LC3-ox). Eighteen independent transgenic lines were obtained and four of them were homozygous with either insertion or deletions at 5’ end of LIP1 (Fig 6A). Observation and analysis of the leaf inclination of plants in fields showed the increased leaf inclination under LIP1 deficiency (Fig 6A), indicating the crucial roles of LIP1 in mediating the LC3-LIP1 effects. In addition, the increased expression of OsIAA12 under LIP1 deficiency (Fig 6B) further demonstrate that LIP1 and LC3 synergistically inhibit the transcription of OsIAA12 expression.
SPOC-domain is speculated to facilitate protein-protein interactions in the transcription repression complex. Although SPOC domain-containing proteins are demonstrated to involve in regulation of various developmental processes, functions of them in plants are rarely reported. On the other hand, though auxin signaling/biosynthesis related genes are shown to affect the lamina joint development, the upstream regulation is still poorly understood. We functionally characterize a novel rice SPOC domain-containing protein leaf inclination 3 (LC3), whose deficiency (lc3 mutant) presents enhanced leaf angle due to the excessive cell elongation at adaxial side of lamina joint, and demonstrate that LC3 controls leaf inclination by regulating auxin signaling through interacting with LIP1, a HIT zinc finger domain-containing transcriptional factor. It is therefore proposed that LC3 interacts with LIP1 to cooperatively suppress the expression levels of OsIAA12 and OsGH3.2, resulting in the suppressed auxin signaling and homeostasis, to maintain the normal lamina joint development (Fig 6C). Our findings not only identify a novel factor regulating leaf inclination through auxin signaling and homoeostasis, but also reveal the function and underlying mechanism of a novel SPOC domain-containing protein.
Previous reports showed that RRM domain of SPOC domain-containing protein functions to recognize and bind to DNA/RNAs. Deficiency of RRM domain suggests that LC3 acts as a transcription repressor through interacting with other factors. Indeed, LIP1, a HIT zinc finger domain-containing TF, recognizes specific DNA sequence and forms a heterodimer with LC3 through interaction to suppress the transcription of downstream genes, especially OsIAA12 and OsGH3.2. These illustrate the mechanism how LC3-LIP1 heterodimer represses the expression of auxin signaling and homeostasis related genes. In addition, it’s the first time to characterize the function and relevant mechanism of a SPOC-domain containing protein lacking RRM domain, which expands the knowledge on regulating the expression of downstream target genes in addition to RRM domain. A mouse Spen-like protein, MINT, binds to homeoprotein Msx2 to co-regulate osteocalcin [34, 39], and our results provide another example showing how SPOC-domain containing protein functions through interacting with a HIT zinc finger domain-containing protein, suggesting that SPOC-domain containing protein may interact with distinct TFs to suppress the transcription of specific genes, which provides novel insights for the functions of SPOC-domain containing proteins.
The underlying mechanism how LC3-LIP1 represses downstream target genes expression and whether there are other factors involving in the regulation, still need further investigations. In human, SHARP (SMRT/HDAC1-associated repressor protein), a spen protein, interacts with co-repressor SMRT (silencing mediator for retinoid and thyroid receptors) and NCoR (nuclear receptor corepressor), and these co-repressors repress transcription by recruiting a large complex containing histone deacetylase (HDAC) activity [29, 40]. In mice, Znhit1 binds to HDAC1 and suppresses CDK6 expression by decreasing the histone H4 acetylation level in its promoter region [41]. Whether LC3 interacts with histone deacetylase or any other factors to repress the downstream gene transcription and hence regulates the distinct developmental processes needs further studies.
Plant phytohormone IAA plays crucial roles in lamina joint development, however, the upstream regulations of the key negative regulator Aux/IAAs during the process and which distinct IAA-ARF interaction is involved in lamina joint development control are unclear. We at first time demonstrate that a SPOC domain-containing protein LC3 regulates auxin signaling by directly suppress OsIAA12 and auxin homeostasis through OsGH3.2. Aux/IAAs bind to ARFs to suppress its function [42] and as a multi-member family (there are 25 ARFs and 31 Aux/IAA proteins in rice), studies have revealed a complex interacting network of Aux/IAAs-ARFs that participate in regulation of various aspects of plant growth and development. Although it is known that each IAA protein can interact with different ARFs and each ARF protein can be suppressed by different IAAs to perform the diverse and specific functions [43], distinct functions of each interaction pair and how IAA-ARF interaction regulates leaf inclination remain to be elucidated. Our studies demonstrate the specific role of OsIAA12-OsARF17 interaction, which will help to illustrate the auxin effects in lamina joint development. Interestingly, other Aux/IAAs proteins (i.e. OsIAA20, OsIAA21, and OsIAA31) interact with OsARF17 besides OsIAA12, and OsIAA12 can also bind with OsARF21 [32], whether other Aux/IAAs-ARFs interactions regulate leaf inclination need further investigation. Further studies of the downstream genes of OsARF17 will expand the understanding of the detailed mechanism of OsIAA12-OsARF17 regulation in lamina joint development. In addition, the expression level of neither LC3 nor LIP1 is influenced by exogenous IAA treatment (S8 Fig), what kind of factors regulate LC3-LIP1 complex and hence lamina joint development will be interesting to be investigated.
GH3 family members encode an indole-3-acetic acid-amido synthetase that conjugates free IAA to various amino acids [44, 45]. In addition to the regulation of GH3 gene by ARFs, which is conserved among dicot and monocot plants [7], our results provide further understanding of GH3 gene regulations by other regulators.
Rice Zhonghua11 (ZH11, Oryza sativa japonica variety) plants, lc3 mutant, and transgenic lines were grown in Shanghai and Lingshui (Hainan Province) under standard paddy conditions. Seedlings used to isolate protoplasts were grown in MS medium at 28°C with 12h-light/12-h dark cycle.
To analyze the expression pattern, lamina joints of flag leaf were collected from 60-, 65-, 70- or 80-day-old plants (stages 2, 4, 5, 6 according to definition by Zhou et al., 2017). Leaf, root, and stem were collected from 7- or 20-day-old seedlings. Seeds (3, 6, or 9 days after pollination), anther, pistil and spikelet were sampled.
For auxin treatment, 7-day-old seedlings were immersed in liquid 1/2 MS (Murashige and Skoog) medium containing IAA (indole-3-acetic acid, 10 μM) for 2 h and the collars were collected for qRT-PCR (quantitative real-time RT-PCR) analysis.
Leaf angle measurement and paraffin section were performed using plants 10 days after heading. Collected leaf lamina joints were photographed, and angle between sheath and leaf was measured by ImageJ program. At least 30 leaf angles of individual plants were measured.
For paraffin section analysis, leaf lamina joints were fixed in FAA solution (45% ethanol, 5% acetic acid, and 12.5% formaldehyde in water) for 24 h and dehydrated in a graded ethanol series and xylene-ethanol solution. Samples were embedded in paraffin (Sigma) for 1 day, then sections were cut (10 mm) and deparaffinized in xylene, hydrated through a graded ethanol series, and stained with Toluidine Blue. Extra stain was flushed and sections were dehydrated by a graded ethanol series again. Sections were microscopically observed and photographed, and cell number and cell size were calculated using ImageJ software.
Entire LC3 gene sequence including the 3-kb promoter region was amplified using primers LC3-3/LC3-4 and subcloned into pCAMBIA2300 for complementation study. Coding sequence of LC3 was amplified by primers LC3-13/LC3-14 and subcloned into pCAMBIA1300 driven by maize ubiquitin promoter for overexpression analysis. LC3 promoter region was amplified using primers LC3-11/LC3-12 and subcloned into pCAMBIA1300+pBI101 vector [46] to drive the β-glucuronidase (GUS) gene. A binary vector pCAMBIA2300 carrying OsIAA12 coding sequence amplified by primers IAA12-3/IAA12-4 and driven by Zea mays ubiquitin promoter was constructed for overexpressing OsIAA12. Transgenic rice with OsARF17 mutation was generated by Crispr/Cas9 [47] with pOs-sgRNA using primers ARF17-1/ARF17-2. The gene editing construct for LIP1 deficiency via Crispr/Cas9 was designed using primers (LIP1-5/LIP1-6 and LIP1-7/LIP1-8) as previously described [48]. Confirmed constructs were transformed into rice by Agrobacterium-mediated transformation. Sequences of used primers were listed in S1 Table.
Various tissues were collected from the confirmed positive transgenic lines and incubated in substrate buffer (pH 7.0 NaH2PO4, 0.1M; EDTA, 10 mM; K4Fe(CN)6, 0.5 mM; K3Fe(CN)6, 0.5 mM; 1% Trition X-100; 40 mg/mL X-Gluc). Examined samples were vacuumed and kept at 37°C overnight, then washed with 75% ethanol and observed.
Total RNAs were extracted by Trizol reagent (Invitrogen) and used to synthesize cDNA through reverse transcription (Toyobo). qRT-PCR was conducted in a total volume of 20 μL containing 10 μL SYBR Premix Ex-Taq, 0.2 μL cDNA, primers (0.2 mM) and 8.3 μL double distilled water. Rice Actin gene was used as an internal control and transcription levels of LC3, LIP1, OsIAA12, OsGH3.2, OsARF17 were examined using primers LC3-1/LC3-2, LIP1-9/LIP1-10, IAA12-1/IAA12-2, GH3.2-1/GH3.2–2, ARF17-9/ ARF17-10. Other primers used in qRT-PCR analysis were listed in S1 Table. All examinations were conducted with three biological and technological replicates.
Coding sequences of LC3, LIP1, OsIAA12, and OsARF17 were amplified using primer LC3-5/LC3-6, LIP1-1/LIP1-2, IAA12-5/IAA12-6, and ARF17-3/ARF17-4 and subcloned into pGADT7 and pGBKT7 vectors respectively (Clontech). Confirmed constructs were co-transformed into yeast AH109 strain. Transformed yeast clones were diluted 101, 102, 103 times, grown on synthetic dropout (SD) medium in the presence or absence of histidine with different concentrations of 3-amino-1, 2, 4-triazole (3-AT) according to the manufacturer’s instructions (Matchmaker user’s manual, Clontech, California), and observed after 4 days.
pGBKT7-LC3 vector was transformed into yeast strain Y2H Gold and used as bait in yeast-two hybrid screening analysis. The prey cDNAs derived from a rice cDNA library constructed from rice seedlings at different stages of ZH11 were transformed into yeast strain Y187. Mating bait and prey plasmid transformants were rotated at low speed for 20 h, then grown on synthetic dropout (SD) medium absence of histidine. Identified candidate prey cDNA was isolated from yeast cells and transformed into Escherichia coli for sequencing. Full cDNA sequence was amplified and cloned into pGADT7, which was co-transformed with pGBKT7-LC3 into yeast strain AH109 to verify the interaction.
Coding sequence of OsIAA12, OsARF17, LC3, LIP1 were amplified using primers IAA12-7/IAA12-8, ARF17-5/ARF17-6, LC3-9/LC3-10, LIP1-3/LIP1-4, LC3-15/LC3-16 and subcloned into pBI221-RFP [49] or pA7 vectors (C-terminus fusion with GFP or YFP) respectively. Resultant constructs expressing OsIAA12-RFP, OsARF17-GFP, LC3-YFP, LIP1-GFP, LC3-RFP fusion proteins were transformed into rice protoplasts and fluorescence was observed by confocal laser scanning microscope (FV10i, OLYMPUS) after 16 h.
Coding sequences of OsIAA12 and OsARF17 were amplified using primers IAA12-9/IAA12-10, ARF17-7/ARF17-8 and subcloned into pCAMBIA1300S-YN and pCAMBIA1300S-YC vector by Infusion kit (Clontech) separately. Resultant constructs were transformed into Agrobacterium tumefaciens strain GV3101, which were used to infiltrate the leaves of 6-week-old tobacco plants. After infiltration for 48 h, YFP fluorescence was observed using a confocal laser scanning microscope (FV10i, OLYMPUS). Fusion proteins nYFP-OsHAL3 and cYFP-OsHAL3 were used as a positive control [50].
Coding sequences of LC3 and LIP1 were amplified and subcloned into the Gateway vector nLUC and cLUC respectively. After infiltration of tobacco leaves for 48 h, excess luciferin was sprayed on leaves and kept in dark for 10 min to eliminate the background fluorescence. Relative LUC activity was measured by a low light cooled CCD imaging apparatus at -70°C. Experiments were repeated three times for each assay.
Coding sequence of LC3 and LIP1 were amplified using primers LC3-7/LC3-8 and LIP1-1/LIP1-2, and subcloned into pGADT7 (Clontech). OsIAA12 promoter regions Pro2 (-1709 to -915 bp before ATG) and Pro3 (-914 bp to 0 before ATG) were amplified using primers IAA12-11/IAA12-12, IAA12-13/IAA12-14 and subcloned into pHIS2 vector. Resultant constructs were transformed into yeast strain Y187. Yeast transformants were grown on synthetic dropout (-Leu/-Trp/-His) medium containing 175 mM 3-AT for 3 days and observed. Experiments were repeated three times.
ChIP-PCR assays were performed according to previous description [51]. Genomic DNAs extracted from 10-day-old transgenic seedling expressing LC3-GFP were digested into small pieces and crosslinked with formaldehyde. Resultant DNA fragments were sonicated to be ~200 bp in length. Chromatin immunoprecipitation were performed using anti-GFP antibody (ab290; Abcam), and Normal rabbit lgG (10500C; Invitrogen) was used as a negative control. Samples collected before immunoprecipitation were ‘input DNA’. Immunoprecipitated and input DNA were purified with PCR purification kit (Qiagen) and amplified using primers covering around 150-bp region of OsIAA12 or OsGH3.2 promoters by qPCR to examine the ChIP enrichment. Sequences of primers (IAA12-17 ~ IAA12-28, and GH3.2–3 ~ GH3.2–12) are listed in S1 Table. Fold-enrichment was calculated by normalizing the amount of a target DNA fragment against the respective input DNA samples. Experiments were repeated three times.
For effector constructs, coding regions of LC3 and LIP1 were amplified using primers LC3-9/LC3-10 and LIP1-3/LIP1-4 and subcloned into vector pA7 (C-terminus fusion with GFP). A ~2.7-kb DNA fragment of OsIAA12 promoter was amplified by primers IAA12-15/IAA12-16 and subcloned into a modified pGreen0800 vector to generate the reporter construct. Effector and reporter constructs were co-transformed intro rice protoplasts. Dual-luciferase transcriptional activity assay was performed as previously described [52]. Experiments were biologically repeated three times.
Ten-day-old ZH11 seedlings were used to isolate protoplasts and 100 μL protoplasts suspension (containing ~2×105 protoplasts) were transfected with plasmid (5–10 μg DNA) and 110 μL PEG solution. Transformation mixture was incubated in darkness for 15 min at 28°C, then diluted by 1 mL W5 solution (NaCl, 154 mM; CaCl2, 125 mM; D-Glucose, 5 mM; KCl, 5 mM; MES, 2 mM, pH 5.7) and centrifuged at 100 g for 2 min. Protoplasts were suspended in WI solution (Mannitol, 0.5 M; KCl, 20 mM, MES, 4 mM, pH 5.7) and transferred into multi well plates and incubated at 28°C for 16 h.
To construct a phylogenetic tree of SPOC-domain protein, homolog sequences in A. thaliana, O. sativa were obtained at the TAIR Web site (http://www.arabidopsis.org) and Rice Genome Annotation Project (http://rice.plantbiology.msu.edu). Alignment of available sequences was performed with CLUSTALX 1.83. The phylogenetic tree was constructed with MEGA 3 [53] using the neighbor-joining algorithm with 1001 bootstrap replicates.
All relevant data are within the paper and its Supporting Information files except for the sequence data which is available from the rice genome database RICEGE (http://signal.salk.edu/cgi-bin/RiceGE) or GenBank databases (https://www.ncbi.nlm.nih.gov/genbank/) under the following accession numbers: LC3 (Os06g0595900), LIP1 (Os10g0520700), OsIAA12 (Os03g0633800), OsARF17 (Os06g0677800).
|
10.1371/journal.pgen.1002861 | Simple Methods for Generating and Detecting Locus-Specific Mutations Induced with TALENs in the Zebrafish Genome | The zebrafish is a powerful experimental system for uncovering gene function in vertebrate organisms. Nevertheless, studies in the zebrafish have been limited by the approaches available for eliminating gene function. Here we present simple and efficient methods for inducing, detecting, and recovering mutations at virtually any locus in the zebrafish. Briefly, double-strand DNA breaks are induced at a locus of interest by synthetic nucleases, called TALENs. Subsequent host repair of the DNA lesions leads to the generation of insertion and deletion mutations at the targeted locus. To detect the induced DNA sequence alterations at targeted loci, genomes are examined using High Resolution Melt Analysis, an efficient and sensitive method for detecting the presence of newly arising sequence polymorphisms. As the DNA binding specificity of a TALEN is determined by a custom designed array of DNA recognition modules, each of which interacts with a single target nucleotide, TALENs with very high target sequence specificities can be easily generated. Using freely accessible reagents and Web-based software, and a very simple cloning strategy, a TALEN that uniquely recognizes a specific pre-determined locus in the zebrafish genome can be generated within days. Here we develop and test the activity of four TALENs directed at different target genes. Using the experimental approach described here, every embryo injected with RNA encoding a TALEN will acquire targeted mutations. Multiple independently arising mutations are produced in each growing embryo, and up to 50% of the host genomes may acquire a targeted mutation. Upon reaching adulthood, approximately 90% of these animals transmit targeted mutations to their progeny. Results presented here indicate the TALENs are highly sequence-specific and produce minimal off-target effects. In all, it takes about two weeks to create a target-specific TALEN and generate growing embryos that harbor an array of germ line mutations at a pre-specified locus.
| Many genes are being discovered solely on the basis of their association with a trait or disease, or their relatedness to other known genes, but nevertheless the precise biological functions of these genes remain mysterious. We need new tools to discover the immediate molecular, cellular, and developmental functions of genes of interest. Increasingly, the zebrafish is being used as a model organism to discover gene functions that are shared among all vertebrates. In this study we develop new, highly efficient, and very easy to apply methods for generating zebrafish that lack the function of any desired gene. We also introduce sensitive and easy-to-apply methods for detecting newly arising mutations. The approach developed here can also be used to quickly eliminate the function of any chosen gene in other animals or in tissue culture cells. In all, we anticipate the methods described here will be widely applied to study gene function in many different contexts.
| The zebrafish has emerged as a leading model organism for the study of vertebrate biology, because of the remarkable cellular resolution with which the embryo can be studied, the ease of assaying its development and physiology in the laboratory, and its amenability to genetic analyses. Forward genetic screens have been used to discover genes that contribute to tissue specification and morphogenesis, cell biology processes including growth regulation and genome maintenance, specificity of neural wiring, metabolism, behavior, and other aspects of the life cycle [1], [2], [3], [4], [5], [6], [7], [8]. Reverse genetics approaches have been used to uncover the biological processes controlled by genes of interest, and thus the zebrafish is being used increasingly to discover the immediate cellular and molecular functions of genes identified by virtue of their association with disease processes [9], [10], [11]. The methods developed here are aimed at improving and simplifying reverse genetics approaches in the zebrafish.
Several methods exist for perturbing the function of selected genes in the zebrafish, but until recently none reliably and efficiently eliminated the function of any specified gene [5], [12]. Gene function can be attenuated in the embryo with antisense morpholino oligonucleotides (MOs) [13], [14], but the rules for designing effective antisense oligonucleotides have not been perfected, the MOs themselves frequently have unintended off-target effects on development, and even when effective, MOs can only disrupt gene expression transiently [5], [15]. Therefore methods to isolate bona fide mutations in pre-selected genes continue to be pursued. Genome-screening methods have been used to identify and recover locus-specific mutations following random mutagenesis [16], [17], but although these methods are effective, they are highly labor-intensive and not very efficient.
In recent years the development of zinc finger nuclease (ZFN) technology portended the ability to induce mutations in any locus of any genome [18], [19], [20]. In this approach a double strand break (DSB) is induced at a unique site in the genome with a synthetic nuclease and host machinery repairs the chromosome break via the error-prone Non-Homologous End-Joining (NHEJ) pathway. Repair of such lesions often produces small insertions and/or deletions (indels) centered at the site of the DSB. Recent studies demonstrated that locus-specific mutations could be readily induced in zebrafish using ZFNs, and it appeared this approach might be applied to any locus [21], [22], [23]. However severe limitations still constrain the ability to generate zinc finger domains that can bind specifically to any desired genomic target sequence. Thus although ZFN-mediated targeted mutagenesis appeared promising, widespread implementation of the strategy requires a new approach for generating nucleases that exhibit very high sequence specificity.
The recent discovery of the Transcription Activator-Like Effector (TALE) proteins produced by plant pathogenic bacteria of the genus Xanthomonas has uncovered a new type of DNA-binding motif that can be used to create peptides that bind DNA with high affinity and sequence selectivity [24], [25], [26], [27]. TALE DNA binding is mediated by arrays of 33–35 amino acid DNA recognition motifs, each of which interacts with a single target nucleotide. As illuminated by X-ray crystallographic analysis of a TALE-DNA complex [28], [29], the process of DNA recognition occurs in a remarkably modular fashion, so that adjacent recognition motifs interact with adjacent nucleotides in a manner that does not appear to be affected by the presence of neighboring motifs. Nucleotide discrimination is determined by a pair of adjacent amino acid residues within the motif, called the Repeat Variable Di-Residue (RVD); hence the recognition motif is referred to as the RVD repeat module. Combining the simple modular TALE recognition cipher with a few empirically based guidelines, web-based algorithms have been established for designing RVD repeat-based DNA binding peptides that can bind genomic target sequences of interest [30].
Fusion of TALE-based DNA binding domains with the sequence-non-specific nuclease domain derived from the type IIS FokI restriction enzyme has been used to create sequence-specific nucleases, called TALENs [27], [31]. Preliminary studies have demonstrated the promise of TALENs for inducing locus-specific mutations in the zebrafish [12], [32]. Here we describe very simple methods, using reagents that are available for research without restriction, for generating TALENs that are extremely effective for inducing mutations at any locus in the zebrafish. The TALE-based DNA binding domains are generated quickly using a strategy that extracts sequences encoding RVD repeat modules from a library of plasmids and joins them in an ordered sequence using the Golden Gate cloning system [30]. We also establish easy and rapid methods for detecting the mutations induced by TALENs. Using the methods presented here, every embryo injected with mRNA encoding a TALEN will acquire mutations at the targeted locus in somatic tissues and approximately 90% of the animals that reach adulthood transmit newly induced specific locus mutations through their germ lines. In all, it takes about two weeks to create target-specific TALENs and generate growing embryos that harbor an array of germ line mutations at a pre-specified locus.
The TALENs we generate function as sequence-specific heterodimer endonucleases. Each monomer component is a chimeric protein composed of the FokI nuclease domain fused with a synthetic DNA binding domain consisting of an array of RVD repeat modules. Nuclease activity requires the binding of the two components on opposing strands of the duplex at a small interval distance. Although the FokI enzyme normally functions as a homodimer, we utilize mutant derivatives of the nuclease domain [23] so that the TALENs function as obligate heterodimers, thus demanding that both ‘Left’ and ‘Right’ monomers simultaneously recognize their cognate binding sites to achieve nuclease activity. The RVD repeat assembly reagents generated by Cermak et al. [30] allow construction of sequences encoding DNA binding domains composed of up to 31 recognition motifs. However, generally we design monomer TALEN components that each contain 16–20 RVD repeats and that, including the DNA interaction function of the N-terminal portion of the TALEN [28], [29], bind a half-target site of approximately 17–21 nt present on opposing strands and spaced apart by 14–17 bp (target site configuration ≥17 bp – N14–17 – ≥17 bp). Applying parameters described in Materials and Methods to the TALEN Targeter program (https://boglab.plp.iastate.edu/node/add/talen), TALENs can be designed that recognize only a single target site in the zebrafish genome. Such unique target sites can be identified in many exons, as well as introns and promoter sequences (see Discussion). The gene sequences targeted and the RVD repeat arrays of the TALENs used in this study are presented in Figure S1.
To generate a plasmid encoding a Left or Right monomer component consisting of n RVD repeats (see Materials and Methods for details), an initial Golden Gate cloning step is used to assemble two arrays, encoding repeats 1–10 and repeats 11 – n-1, as in [30]. The final expression plasmid encoding an entire TALEN monomer is generated in a second Golden Gate cloning assembly, which brings together the two partial arrays, sequences encoding the nth motif, and a modified CS2+ backbone vector, pCS2TAL3RR or pCS2TAL3DD (Figure S2; see Materials and Methods). Assembly results in a fusion gene that encodes: vector-provided N-terminal TALE-derived sequences, an RVD repeat array, C-terminal TALE-derived sequences, and a modified nuclease domain. TALENs can be expressed directly from the CMV promoter resident in these CS2+ vectors. However, for the zebrafish experiments described below, mRNA encoding Left or Right monomer components were generated individually by in vitro transcription of linearized plasmids and equal amounts of each mRNA were co-injected into embryos at the 1 cell stage.
To estimate the efficiency with which targeted mutations can be induced, we measured the ability of TALENs to induce somatic tissue mutations in the golden (gol) gene, which governs pigmentation in the embryo and adult without compromising viability [33], [34]. Mutations at gol are recessive: embryos with at least one WT allele are darkly pigmented at 2 days postfertilization (dpf), whereas homozygous gol mutants appear pigmentless at this stage. To maximize the chance of inducing complete loss-of-function alleles, we chose to develop a TALEN that would generate DSBs within coding sequences residing toward the 5′ end of the gol locus. Using criteria described in Materials and Methods, a potential TALEN target site was identified in the second exon of gol and the gol-ex2 TALEN was designed (Figure 1A, Figure S1). As imprecise repair of targeted DSBs is likely to produce recessive loss-of-function mutations at gol, we injected embryos heterozygous for the golb1 null mutation [33], [34] with gol-ex2 TALEN mRNA and measured the appearance of golb1/* mutant pigmentless cells in the Retinal Pigmented Epithelium (RPE) (Figure 1). At 2 dpf the RPE is a monolayer of approximately 550 pigmented cells that envelops each eye, and the presence of even small clones of pigmentless tissue can be detected easily [35]. Whereas it is extremely rare for pigmentless cells to be found in the RPEs of control golb1/+ heterozygous embryos [36], all but one of 20 TALEN-injected golb1/+ embryos had large patches of gol mutant cells (Figure 1B–1F, Table 1). The TALEN-injected embryos had multiple patches of mutant tissue, indicating they were genetically mosaic.
On average ≥50% of the RPE cells in TALEN RNA-injected embryos were golden (Figure 1C–1F), indicating the majority of genomes in the embryos acquired TALEN-induced mutations. Three experiments demonstrated the new mutations were indeed induced by the gol-ex2 TALEN. First, the gol-ex2 TALEN could induce pigmentless tissue in the RPEs of gol+/+ embryos (Figure 1G–1K), indicating the induction of mutant tissue did not require a pre-existing gol mutant allele. Mutant cells were observed in almost 100% of these injected WT embryos (Table 1) and occasionally even wholly gol embryos were observed (Figure 1L, 1M; see Table S2 for frequencies), highlighting the efficiency with which these TALENs can induce mutations in both genomes of a cell. Second, as discussed below, analysis of exon 2 sequences amplified from gol-ex2 TALEN RNA-injected embryos revealed a diverse set of indel mutations were induced, typical of those produced by NHEJ-mediated repair of DSBs. Third, gol mutant alleles were transmitted to the F1 offspring of TALEN-injected WT embryos (see below). We conclude the gol-ex2 TALEN is extremely effective at inducing mutations at gol.
Whereas newly induced mutations at golden are simple to detect, TALEN-induced mutations at most loci are unlikely to present a phenotype that can be scored in individual somatic cells. As the repair of DSBs can lead to an assortment of indels centered at the TALEN target site, it is desirable to detect induced mutations with a method that can detect any DNA change arising at a selected target site. Furthermore, as DSBs are induced only following translation of injected TALEN mRNAs, an individual embryo may acquire several independent mutations, each of which may arise in a mosaic fashion during development. Therefore a sensitive method is required that can detect mixtures of TALEN-induced DNA polymorphisms at a targeted locus present among the genomes of a single embryo.
We find High Resolution Melt Analysis (HRMA) is a simple, rapid, and sensitive method for detecting TALEN-induced somatic mutations. HRMA has been used in the past to detect known sequence polymorphisms in the zebrafish [37], but it is also useful for discovering polymorphisms of unknown sequence that lie in a small, defined region of the genome. We developed standard conditions for detecting sequence alterations resulting from TALEN activity at targeted loci (Materials and Methods). To detect the occurrence of a DNA polymorphism at a particular locus, short PCR amplicons (90–120 bp) that include the region of interest are generated from a genomic DNA (gDNA) sample, subjected to denaturation and rapid renaturation, and the thermostability of the population of renatured amplicons is analyzed (Materials and Methods). If TALEN-induced polymorphisms are present in the template gDNA, heteroduplex as well as different homoduplex molecules will be formed (Figure 2A). The presence of multiple forms of duplex molecules is detected by HRMA, which records the profile of temperature-dependent denaturation and detects whether duplex melting acts as a single species or more than one species. For example, three prominent duplex types with distinct melting temperatures (Tm's) are evident in the analysis of renatured lef1 amplicons generated from an embryo heterozygous for an intragenic 18 bp deletion in lef1 (Figure 2B).
We determined the sensitivity of HRMA for detecting the presence of a mutant genome mixed among a population of WT genomes using standard conditions of analysis (Materials and Methods). gDNA prepared from an embryo heterozygous for the lef1Δ18 mutation was mixed with differing amounts of WT gDNA, and the ability of HRMA to detect the mutant allele was determined. As the mutant genome is present as a decreasing fraction of all template genomes, the relative abundance among amplicons of homoduplex mutant and heteroduplex populations changes, but the presence of mutant genomes can be detected unambiguously even when mutant genomes represent 1/70th of the total mix (Figure 2C). Figure S3 shows a 4 bp insertion mutation in the lef1 gene can be detected with similar sensitivity. In genetically mosaic embryos, when multiple mutant alleles are present as minor populations within a mixture of genome, the melt curves become more complex, but deflections away from the WT profile are additive and therefore simple to detect (see examples in Figure 3). Given our findings (presented below) that the majority of mutations induced using the methods described here consist of alterations of >4 bp, we estimate our standard conditions of analysis can routinely detect one mutant genome present among 50 WT genomes.
To test the efficacy of our methods for generating and detecting TALEN-induced somatic mutations at any locus, we injected 1 cell embryos with mRNAs encoding the gol-ex2 TALEN or TALENs designed to recognize and cleave sequences in exon 3 of tbx6, exon 5 of ryr3, or exon 6 of ryr1a (Figure S1). Under standard conditions of injection with 100 pg total TALEN RNA, approximately 95% of the embryos developed normally (Table S3). gDNA was prepared from individual 1–2 dpf TALEN-injected or control embryos and analyzed by HRMA for the presence of DNA sequence variants at the targeted loci using primer pairs listed in Table S1. Nearly all TALEN RNA-injected embryos had targeted mutations, including the gol-ex2 TALEN RNA-injected embryos that did not have gol mutant cells in the RPE (Table 2, Figure 3). Sequence analysis of PCR amplicons covering the targeted loci (Figure 3) indicated the induction of a spectrum of indel mutations, consistent with what is expected from NHEJ repair. Analysis of loci amplified from 9 embryos targeted at tbx6, ryr3, or ryr1a indicated approximately half the genomes (26/52) of TALEN-injected embryos harbored targeted mutations.
As gol does not encode an essential function, the induction of gol mutant cells is not detrimental to development. However for other targeted genes it may be advantageous to avoid inducing bi-allelic mutations in many cells of the embryo. The frequency with which mutant cells are induced is dependent on the amount of TALEN RNA that is injected (Table 1, Table S2), so the frequency of cells harboring bi-allelic mutations can be adjusted. As shown in Figure 4, HRMA can be used to determine the dose-dependent induction of mutations that cannot be easily measured in somatic tissues. The amount of TALEN RNA delivered to embryos affects both the fraction of embryos that harbor detectable mutations as well as the average abundance of mutant genomes present in any individual injected embryo.
To determine if induced mutations detected in 1–2 dpf embryos enter the germ line, we raised TALEN-injected (G0) embryos to adulthood and analyzed the transmission of mutations to progeny. Individual adult G0 animals were mated with WT partners and gDNA isolated from individual F1 embryos was analyzed by HRMA. Targeted germ line mutations were transmitted by 51 of the 57 G0 animals (approximately 90%) that had been exposed to TALENs directed at the golden, tbx6, ryr3, or ryr1a genes (complete transmission data is provided in Tables S4, S5, S6, S7; data is summarized in Table 3). The fraction of G0s carrying germ line mutations induced by each TALEN ranged from 77% to 100%. Analysis of the sequence alterations inherited by F1 embryos revealed the range of induced indel mutations among the germ line transmitted mutations mimicked those that had been observed in embryos (Figure 5A). The majority of TALEN-induced mutations identified here in injected G0, 1–2 dpf F1, or adult F1 individuals were sequence changes of 3–20 bp (Figure 3, Figure 5, Table S8). Larger indels were identified, but rarely. Most if not all mutations identified among the 1 dpf F1 offspring were viable in the heterozygous state, as adult F1s descended from G0 founders harbored a distribution of mutations similar to that found among the F1 embryos (Table S8). To date all the F1 fish heterozygous for TALEN-induced mutations at tbx6, ryr1a, and ryr3 that have been bred (n = 16) have transmitted their mutations to the F2 generation. We have no evidence of mutations in the F1 generation that could not be further propagated.
Most G0 animals transmitted multiple mutant alleles and most mutations were transmitted by significantly less than 50% of the gametes (Table 3; Tables S4, S5, S6, S7, S8), indicating the germ lines of G0 founders were mosaic. HRMA analysis of heterozygous F1 offspring was used to distinguish transmitted mutant alleles, and sibling individuals that appeared to carry a common DNA sequence alteration were grouped based on common melt curve shapes. As illustrated in Figure 5B and 5C, individual sibling F1 embryos descended from a single mutagenized G0 founder could harbor different alleles, indicated by the multiple distinct HRMA melting profiles of amplicons derived from sibling F1 offspring. On average, two new alleles were recovered from the germ lines of each mutagenized G0 founder (Table 3).
The finding that the germ lines of most G0 founders were genetically mosaic is consistent with the interpretation that TALENs induce mutations independently in different cells of the embryo and mutations accrue as the embryo develops. Indeed we found that following injection of 1 cell stage embryos with gol-ex2 TALEN mRNAs, the fraction of embryos with detectable levels of targeted mutations increased with developmental time (Figure S4). Whereas 100% of injected embryos had mutations at 6 or 24 hours postfertilization (hpf), only a fraction of the 3 hpf embryos had detectable mutations, and among the early stage embryos, HRMA genotype analyses indicated a relatively low abundance of mutant genomes (Figure S4). Altogether these data indicate numerous targeted indel mutations can be recovered routinely from a small set of adults arising from TALEN mRNA-injected embryos.
The very high frequency with which mutations were induced raised the possibility that the TALENs were simply mutagenic in a sequence-non-specific manner. However, in contrast with published reports [21], [22], [23] and our own experience with ZFNs (data not shown), TALENs exhibit very low toxicity. For example, although injection of only 4 pg total RNA encoding any of the TALENs used here was sufficient to induce mutations in the majority of embryos (Figure 4 and data not shown), approximately 95% of the embryos developed normally even after being injected with 100 pg total TALEN RNA (Table S3). These observations are consistent with the interpretation that random chromosome breaks are generally not produced by the TALENs.
As one test of the specificity of the TALENs used here, we generated TALENs to recognize a specific target sequence and measured the induction of mutations at very closely related sequences present in homologues of the targeted gene. The ryr genes encode related Ryanodine Receptor intracellular calcium release channels. We designed TALENs targeted to the ryr3 or ryr1a genes and measured the induction of mutations in the homologous sequences present in the ryr1a, ryr1b, or ryr3 genes. The ryr3-ex5 TALEN recognizes a Left half-site of 19 nt and Right half-site of 18 nt (Figure S1). As illustrated in Figure 6A, the homologous sites in the ryr1a and ryr1b genes present a 3 base mismatch for the Left monomer and 4 base mismatch for the Right monomer of the ryr3-ex5 TALEN. Although HRMA indicated every embryo injected with ryr3-ex5 TALEN had induced mutations in ryr3, the embryos did not have detectable levels of mutations at the ryr1a or ryr1b loci (Figure 6B).
In a second experiment we analyzed the off-target activity of the ryr1a-ex6 TALEN, which recognizes a Left half-site of 19 nt and Right half-site of 17 nt. As shown in Figure 6C, the homologous site in ryr3 differs at 3 positions in both the Left and the Right half-sites, but the homologous site in ryr1b differs only at 2 positions in the Left half-site and presents a perfect match with the Right half-site. HRMA analysis revealed the ryr1a-ex6 TALEN induced targeted mutations at the cognate locus in 100% of the injected embryos but failed to induce detectable mutations at the homologous ryr3 target (Figure 6D). In contrast, the ryr1a-ex6 TALEN did induce mutations in ryr1b gene of all of the injected embryos (Figure 6D). These results indicate the TALENs display high but not perfect sequence specificity under the conditions described here.
Sequence-specific TALENs that can target almost any zebrafish gene can be generated simply and rapidly using the target site design parameters, the reagents, and the cloning strategy described here. The TALENs generated using these reagents are very effective at inducing DSBs whose repair often produces mutations at the target sites. Using our standard methods to generate mutations at four different loci, we found virtually every embryo injected with TALEN mRNA harbored targeted mutations. As demonstrated with the gol-ex2 TALEN, in some cases over half the genomes in a TALEN RNA-injected embryo may acquire mutations at targeted loci. Almost all treated animals acquire new germ line mutations, which are subsequently inherited in a stable Mendelian fashion. We have found many additional sites in the genome can be targeted with the ease and efficiency of the four genes described here: during the course of the current study we induced mutations at 19 of the 21 sites we attempted to target (90%). The genes that have been successfully targeted are present at many different locations in the genomes (Figure S1) including the gol locus, which is close to a telomere [33]. Further, as shown by the induction of gol mutations in early stage embryos (Figure S4), mutations can be induced in genes that are not being expressed. Given the modular mode of RVD repeat module binding, the flexibility in the design and the length of repeat arrays that can be generated, and the relatively few guidelines that restrict TALEN design, it appears TALENs can be used to induce loss-of-function mutations in almost every zebrafish gene.
The system we present here for constructing TALENs that are effective in the zebrafish is easy to implement: 1) it employs a very simple and rapid cloning strategy, using the Golden Gate cloning method system to assemble RVD repeats; 2) the RVD repeats are available in plasmids that can be used directly for Golden Gate cloning without need to PCR amplify or otherwise alter the repeat sequences; 3) the cloning strategy leads to accurate constructions of arrays (we typically analyze only 2 transformants from each cloning step); 4) the TALENs function as obligate heterodimers, increasing specificity of target sites recognized by the TALENs; and 5) all reagents are readily available through Addgene.
Although the studies presented here focus on induction of mutations in the zebrafish, our preliminary results indicate that TALENs assembled using the reagents described here are also effective at inducing mutations in Drosophila and mammalian cell culture systems.
We developed High Resolution Melt Analysis as a principal method for measuring the induction of mutations with TALENs in zebrafish embryos. HRMA can detect almost any newly arising polymorphism at a pre-specified region of the genome, and thus it allows detection of the large variety of sequence changes that may be produced by NHEJ repair. As HRMA can detect sequence alterations arising at any target site, this method of mutation detection does not bias or affect the choice of a target site and thus is an improvement over previous assays that detected loss of a restriction site. We show HRMA is an extremely sensitive method for detecting mutant genomes among a mixed population of genomes. Hence it is particularly useful for detecting TALEN-induced mutations, any one of which is likely to be present in only a subset of the cells of an embryo injected with TALEN RNAs. Furthermore, because HRMA is sensitive to the total fraction of mutant genomes in a mixture, it is particularly well-suited for detecting the heterogeneous set of mutations that is likely to arise in a single embryo. Finally, HRMA is simple to apply and does not involve manipulation of samples following PCR amplification. The entire procedure for generating amplicons and analyzing the thermostability of the heteroduplexes can be performed in less than 2 hours.
We consider several practical issues concerning implementation of the methods presented here to induce mutations in the zebrafish: 1) the kinds of mutations generated and implications for target design considerations; 2) the frequency with which potential TALEN target sites can be identified; 3) the possibility that TALEN-injected embryos express mutant phenotypes; and 4) the ability of TALENs to induce unintended mutations.
Most of the mutations we have recovered from G0 germ lines are indels that affect ≤20 bp stretches of genomic sequence. Many of these sequence changes cause frameshift mutations. If TALENs are used with the goal of inducing null mutations in a targeted gene, it is best to target a region of the gene in which a frameshift mutation is likely to produce a protein product that lacks important functional elements. We routinely target the second or third exon of a gene. Among 40 genes for which we have designed TALENs, we have always been able to identify targets in the second and/or third exons of the genes using the design parameters presented in the Materials and Methods. Among searches of 116 stretches of (mostly coding) sequences with an average size 238 bp, we identified at least one suitable target in 77% of the sequences (the average size of the sequences without a best-fit target was 158 bp). Using the guidelines we suggest, optimal TALEN target sites can be identified for most genes.
The parameters that govern the specificity and activity of TALENs are not completely understood. It is clear from our studies and those of others that the Left and Right TALEN monomer components can bind at various distances from each other and still cooperate effectively to accomplish target site cleavage [27]. The spacer length for achieving optimal activity has not been determined, and we can only say the 14–17 bp spacer length we routinely use in the design of TALEN target sites consistently allows for effective target cleavage in the zebrafish. Importantly, the extreme minimum or maximum spacer distance at which some cleavage activity may occur has yet to be determined in vivo, an uncertainty that affects the identification of unintended sites in the genome that may be susceptible to TALEN activity.
As expected from previous work on the specificity and selectivity of TALE binding [25], [27], the TALENs function as highly sequence-specific nucleases. Given that we identify potential TALEN target sites on the basis of a reference genome, and as the common laboratory WT strains of zebrafish harbor polymorphisms, we have found it prudent to sequence genomic regions of the zebrafish used in any series of experiments to verify the existence of a presumed target site in the embryos that are injected with TALEN RNA. Furthermore, it should be noted HRMA can detect pre-existing polymorphisms, and thus we choose to inject embryos shown to be free of polymorphisms near the TALEN target site.
We routinely verify the in vivo activity of a TALEN before growing injected G0 embryos to adulthood. Typically we inject 1 cell stage embryos with different amounts (4 pg, 20 pg, 100 pg total RNA) of a 1∶1 mixture of Left TALEN and Right TALEN RNAs and measure the presence of mutations in individual 1 dpf embryos. In our experience to date, most animals that have evidence of mutations at 1 dpf will grow to become adults that transmit mutations through the germ line.
As demonstrated with the gol-ex2 TALENs, cells with two mutant alleles can be induced following injection of 1 cell embryos with TALEN RNA. Mutations are induced in a mosaic fashion and using the conditions described here, it is rare that TALEN-injected embryos exhibit strong mutant phenotypes. It may be possible to augment the activity of the TALENs to uncover mutant phenotypes in G0 animals. In addition, it is worth considering the possibility that some mutations cannot survive in the germ line in a homozygous state. As a result, it may be desirable to raise G0 animals with sub-maximal levels of mutations. As demonstrated in our studies, the frequency of TALEN-induced mutations is a function of the amount of TALENs introduced into an animal. Thus, it is possible to raise animals that carry different mutation loads.
Finally, although we have not made extensive measurements of the frequency with which unintended sequences are recognized and cleaved by TALENs in these experiments, our studies indicate off-target mutations can occur but they are sufficiently infrequent so that they are unlikely to confound analysis of targeted gene function. We found TALENs failed to cleave potential target sites that differed at about 6 positions of a 36 nt binding target, but that targets differing at only 2 positions could be effectively cleaved. The effects of off-target mutations can be minimized by studying hetero-allelic combinations of targeted mutations derived from independently mutagenized G0 founders. In sum, the current conditions for TALEN-induced mutagenesis appear sufficient for uncovering the function of almost any selected gene in the zebrafish.
HRMA sensitively detects polymorphisms. To minimize detection of polymorphisms present in the backgrounds of WT fish, we typically amplify only a small region of the genome bordering the TALEN target site and analyze that for newly induced mutations. As a result of using small amplicons, we will be unable to detect some TALEN-induced mutations that delete primer-binding sites. As we only rarely observed deletions of >30 bp in the present studies, we believe the majority of TALEN-induced mutations can be identified with the methods described here. It is also possible to detect induced mutations by HRMA using larger amplicons. Finally, although the HRMA studies presented here were performed with a LightScanner (Idaho Technology), we have obtained identical results with similar sensitivity using an Eco Real-Time PCR System (Illumina). Additional instruments initially designed for qPCR analysis are capable of performing HRMA and have been used successfully to detect TALEN-induced mutations (Tatjana Piotrowski and Steven Leach, personal communication).
All experiments were performed in accordance with, and under the supervision of, the Institutional Animal Care and Use Committee (IACUC) of the University of Utah, which is fully accredited by the AAALAC.
Wild type zebrafish Danio rerio were of the Tuebingen strain. Zebrafish were maintained under standard conditions and embryos were generated, cultured and staged as described [38], [39].
Exon sequences identified from the Zv9 Zebrafish Genome Assembly were scanned for potential TALEN target sites, which were identified using the TALEN Targeter program at https://boglab.plp.iastate.edu/node/add/talen. The following parameters were used: 1) spacer length: 14–17; 2) repeat array length of 16–21; 3) apply all additional options that restrict target choice. Preference was given to target sites: 1) close to the 5′ terminus of the gene to maximize chances of inducing premature translation stop mutations and 2) not in the first exon in case alternative promoters exist. Target sequences that are unique in the genome should be chosen following BLAST analysis to determine that highly similar Left and Right binding sites in close proximity did not exist at other sites in the genome (see Figure 6).
New final backbone vectors used to construct and express genes encoding Left and Right TALEN monomer components were generated here. The new backbone plasmids, pCS2TAL3DD and pCS2TAL3RR, were modified from pCS2-Flag-TTGZFP-FokI-DD and pCS2-HA-GAAZFP-FokI-RR plasmids [23]. First, to render the plasmids suitable for Golden Gate cloning, the single Esp3I restriction enzyme site (in the FokI nuclease domain) of each plasmid was changed from GAGACG to GCGCCG, a mutation that did not alter coding. Second, sequences encoding the ZFP domain were removed following KpnI and BamHI digestion, leaving a backbone vector with sequences 5′ to the KpnI site that provided a 5′UTR, start codon, NLS, and Flag or HA tags and sequences 3′ to the BamHI site that provided a heterodimeric FokI domain, translation termination codon, and SV40 polyA signal. Third, sequences derived from the tal1c gene and ready to accept an RVD repeat array by Golden Gate cloning were placed in frame at the KpnI and BamHI sites. The tal1c sequences were obtained from pTAL3 (sequence positions 1214–2210, www.addgene.org) using primers that added a KpnI site at the 5′ end (TAL3N153F-GTAGGATCCGGTACCGTGGATCTACGCACGCTCGG) and a BamHI site at the 3′ end (TAL3C63R-GTGGGATCCGGCAACGCGATGGGACG) of the amplified pTAL3 sequence. The amplified sequence encoded only a central portion of TAL1c, in which lacZ sequence had been substituted for the RVD repeat array. Cloning into the KpnI/BamHI sites of the backbone vector transferred sequence that provided 136 aa of TALE immediately N′-terminal to the RVD repeat array, an Esp3I restriction enzyme site, lacZ sequences, an Esp3I restriction enzyme site, and 63 aa of TALE immediately C′-terminal to the RVD repeat array. The TALE backbone truncations were designed after TALENs that previously had been shown to function well [27]. Golden Gate cloning of RVD repeat arrays into pCS2TAL3DD or pCS2TAL3RR results in replacement of the lacZ sequences with sequences encoding a designed RVD repeat array and yields a gene encoding an intact TALEN monomer. The pCS2TAL3-DD and pCS2TAL3-RR plasmids are available through Addgene (#37275 and #37276, respectively) with complete sequence information accessible at GenBank (accession numbers JX051360 and JX051361, respectively).
The TALEN Golden Gate assembly system described in Cermark et al [30] was used with modifications. RVD repeat arrays were assembled exactly as described [30]. Plasmids providing RVD repeats for Golden Gate cloning are described in [30] and are available through Addgene. Briefly, two rounds of Golden Gate cloning assembly were used to generate a TALEN gene with n RVD repeat modules. First, two arrays were generated, corresponding to repeat modules 1–10 and 11 – n-1. Resulting vectors that acquired arrays were identified as white transformants on IPTG/X-gal plates. Correct assembly was determined first by the size of repeat array inserts liberated following XbaI and AflII restriction enzyme digestion and then sequencing of plasmids with the correct insert size. Second, the two arrays and sequences encoding the nth motif were transferred into the backbone vectors. RVD repeat array sequences were cloned into pCS2TAL3DD to generate a Left TALEN gene and into pCS2TAL3RR to generate a right TALEN gene. Backbone vectors that acquired arrays were identified as white transformants on IPTG/X-gal plates. Correct assembly was determined first by the size of the insert liberated by SphI and BamHI restriction enzyme digestion and second by sequencing junction regions.
5′-capped mRNA was generated by transcription in vitro of pCS2TAL3DD and pCS2TAL3RR TALEN plasmid templates that had been linearized with NotI (mMESSAGE mMACHINE SP6 kit, Ambion/Invitrogen). Equal amounts of Left and Right TALEN mRNA were injected together into the cytoplasm of 1 cell stage zebrafish embryos.
To prepare genomic DNA from embryos, individual 1 or 2 dpf embryos were incubated in 50 ul DNA extraction buffer [10 mM Tris-HCl (pH 8.0), 1 mM EDTA, 50 mM KCl, 0.3% Tween-20, 0.3% NP-40] containing 500 ug/ml proteinase K at 55°C, 2 h. The reaction was terminated by incubation at 99°C, 5 min. The average gDNA concentration was roughly 60 ng/ul.
To detect TALEN-induced mutations by HRMA, a 90–120 bp amplicon that included the entire genomic target site was generated. Primers flanking the target site were used to amplify the genomic region in a 10 ul PCR reaction containing: 1 ul embryonic gDNA, 1X LightScanner Master Mix (containing the LC Green Plus dye, Idaho Technology), 200 uM each dNTP, and 200 nM each Forward and Reverse primers (see Table S1). Amplification/duplex formation conditions were: denaturation at 95°C, 3 min; 50 cycles {95°C, 30 s–70°C, 18 s}; denaturation at 95°C, 30 s; renaturation at 25°C, 30 s; 10°C. HRMA data was collected on a LightScanner (Idaho Technology) and analyzed using the LightScanner Call-IT Software.
|
10.1371/journal.pgen.1004193 | Mitogen-Activated Protein Kinase (MAPK) Pathway Regulates Branching by Remodeling Epithelial Cell Adhesion | Although the growth factor (GF) signaling guiding renal branching is well characterized, the intracellular cascades mediating GF functions are poorly understood. We studied mitogen-activated protein kinase (MAPK) pathway specifically in the branching epithelia of developing kidney by genetically abrogating the pathway activity in mice lacking simultaneously dual-specificity protein kinases Mek1 and Mek2. Our data show that MAPK pathway is heterogeneously activated in the subset of G1- and S-phase epithelial cells, and its tissue-specific deletion results in severe renal hypodysplasia. Consequently to the deletion of Mek1/2, the activation of ERK1/2 in the epithelium is lost and normal branching pattern in mutant kidneys is substituted with elongation-only phenotype, in which the epithelium is largely unable to form novel branches and complex three-dimensional patterns, but able to grow without primary defects in mitosis. Cellular characterization of double mutant epithelium showed increased E-cadherin at the cell surfaces with its particular accumulation at baso-lateral locations. This indicates changes in cellular adhesion, which were revealed by electron microscopic analysis demonstrating intercellular gaps and increased extracellular space in double mutant epithelium. When challenged to form monolayer cultures, the mutant epithelial cells were impaired in spreading and displayed strong focal adhesions in addition to spiky E-cadherin. Inhibition of MAPK activity reduced paxillin phosphorylation on serine 83 while remnants of phospho-paxillin, together with another focal adhesion (FA) protein vinculin, were augmented at cell surface contacts. We show that MAPK activity is required for branching morphogenesis, and propose that it promotes cell cycle progression and higher cellular motility through remodeling of cellular adhesions.
| Development of the ureter and collecting ducts of the kidney requires extensive growth and branching of an epithelial tube, the ureteric bud. While many genes that control this process are known, the intracellular signaling pathways that underlie renal morphogenesis remain poorly understood. The cellular changes that contribute to ureteric bud morphogenesis, such as adhesion and movements, are guided by intracellular signaling triggered by stimuli at the cell surface. Mitogen-activated protein kinase (MAPK) pathway is known to regulate proliferation in general, but its precise functions during different cell cycle phases are debatable. Moreover, the role of MAPK activity in control of cell adhesion has been demonstrated in cultured cells, but such regulation in vivo remains to be elucidated. Here, we examine the importance of the MAPK activity in ureteric bud branching, and find that simultaneous lack of Mek1 and Mek2 genes allows elongation of the bud but specifically arrests new branch formation. We show that lack of MAPK activity leads to changes in focal adhesion molecules and E-cadherin mediated cell adhesion and delay in cell cycle progression. Our findings may help to understand the origins of certain congenital malformations in humans.
| Receptor tyrosine kinase (RTK) signaling is a key mechanism through which extracellular stimuli guide development of the kidney and many other organs, but the specific in vivo functions of intracellular cascades activated downstream of RTKs remain poorly characterized. The kidney develops as a result of classical reciprocal inductive tissue interactions between the nephron-producing metanephric mesenchyme (MM), and the branching epithelium of the ureteric bud (UB), a structure later giving rise to the collecting duct system of the functional organ [1]. Renal differentiation begins with the formation of UB, which invades the surrounding MM, and subsequently starts its branching. UB morphogenesis is largely instructed by the MM, which secretes growth factors such as glial cell-line derived neurotrophic factor (GDNF) and members of fibroblast growth factor (FGF) family. Their RTK receptors, namely RET and FGF receptor 2, expressed in UB epithelial cells, regulate UB development [2].
Based on genetic and in vitro experiments, GDNF/RET signaling is required for early UB morphogenesis [3]–[5], while the requirement for FGFR signaling appears to arise later during normal kidney development [6], or in situations where RET signaling is absent [7]. Although the molecular basis of UB branching has been extensively studied, relatively little is known of the cellular cascades and responses regulating the formation of new branches in vivo. Binding of GDNF and FGF to their receptors activates several intracellular pathways of which phosphoinositide 3-kinase (PI3K)/AKT, mitogen-activated protein kinase (MAPK) and phospholipase Cγ (PLCγ) function during renal differentiation [8]. Inhibition of the PI3K pathway in kidney organ cultures suggests that primary UB formation depends on chemotactic cell motility induced by this pathway [9], whereas similar experiments with MEK inhibitors suggest that the MAPK pathway is also required for UB morphogenesis [10], [11]. Attempts to genetically confirm such functions are largely missing although deletion of the protein tyrosine phosphatase Pntpn11, which positively regulates MAPK, JAK/Stat and PI3K/Akt, suggests that these intracellular cascades also mediate pivotal functions during in vivo development [12], [13]. Mutations in specific RET docking sites known to activate certain intracellular pathways indicate that induction of PLCγ via Y1015 as well as simultaneous activation of PI3K and MAPK via Y1062 pathways are involved in renal differentiation [14]–[17].
Active cell proliferation occurs in UB tips [18], which are the major sites for generation of new branches formed through bifurcation of an existing buds [11]. In addition to proliferation, which appears to involve transient delamination of the cells from monolayer [19], active cell movements needing constant turnover of cellular adhesions have been implicated in UB morphogenesis [20], [21]. MAPK pathway, which is well known cell cycle regulator, functions through the RAS-RAF-MEK-ERK cascade, but its specific requirements during different cell cycle phases are highly cell type specific. The activation of RAF kinases leads to rather linear signal transduction upon phosphorylation of dual-specificity protein kinases MEK1 and −2, which in turn phosphorylate ERK1 and −2 (presently their only known substrates) [22]. ERKs have a wide variety of nuclear and cytosolic targets including cyclin D1 and focal adhesion (FA) scaffold protein paxillin, which also associates with MEK [23], [24]. Either disruption of ERK/paxillin complex or lack of ERK induced phosphorylation on serine 83 abolishes cell spreading and branching morphogenesis [24], [25]. Interestingly, paxillin and another FA protein, vinculin, are found also in adherens junctions (AJ), where they associate with β-catenin to modulate adhesion at sites of cell-cell contact [26], [27]. Vinculin stabilizes E-cadherin at AJs where it potentiates E-cadherin mechanosensory responses [28], [29].
Here we have studied the in vivo functions of MAPK pathway during renal branching by deleting Mek1 [30] specifically in UB epithelium in Mek2 -null background [31]. As previously suggested by chemical inhibition of MAPK in whole kidney cultures [10], [11], our results show definitively that loss of MAPK activity specifically in the UB prevents the generation of new branches while allowing bud elongation. The MAPK pathway appears to contribute to UB branching guidance by carrying out dual functions; it regulates G1/S-phase transition during cell cycle progression, and epithelial cell adhesion through paxillin phosphorylation affecting FA and AJ dynamics.
The pattern of MAPK pathway activity was first studied in kidneys at different developmental stages. As shown before [20], pERK1/2 localized on one side of Wolffian duct epithelium at E10.5, just before UB outgrowth. A day later when the UB had branched once to form the so-called T-bud, prominent pErk1/2 staining was detected both in the epithelium and surrounding MM (Figure 1A). During subsequent branching, MAPK activity was restricted to UB tip regions, in a pattern similar to Ret expression, and to early nephron progenitors in the MM and dispersed cells in the medulla (Figure 1B–D).
Closer examination of pERK1/2 staining revealed striking heterogeneity in MAPK activity between adjacent epithelial cells (Figures 1A–B and 2A–F). In the pseudo-stratified E10.5 Wolffian duct epithelium [20], the mitotic nuclei localize at the apical surface, while the S phase nuclei are found on the basal surface, and G1/2 nuclei within the middle epithelial zone (due to interkinetic nuclear migration). Most pERK1/2 positive cells in E10.5 Wolffian duct epithelium were found in the middle and basal zones (Figure 2A). Later, during the active UB branching phase, a similar pattern of pERK1/2 was maintained (Figure 2B–F) suggesting that MAPK pathway is activated in a subset of cells in G- and S-phases of the cell cycle. Accordingly, pulse labelling of proliferating cells with uridine analog 5-ethynyluridine (EdU) followed by simultaneous detection of pERK1/2 and EdU-positive cells showed co-localization in the UB tips (Figure 2B). Notably, while a large fraction of cells in G- or S-phases were positive for pERK1/2, none of the mitotic cells (identified by their round shape and expression of phosphorylated histone H3) were pERK1/2 positive (Figure 2C–F). This was constantly found in six distinct kidneys (E10.5–13.5) accounting approximately 100 UBs, and suggests lack of MAPK activity during mitosis in UB epithelial cells.
Abundant pERK1/2 in developing kidney (Figures 1 and 2) together with in vitro chemical inhibition studies [10], [11] suggested that MAPK pathway could be important for kidney morphogenesis in vivo. Conventional knockout of Mek1 is embryonic lethal while generation of a conditional allele allows its tissue-specific deletion [30] from Wolffian duct and UB lineages using Hoxb7CreGFP transgenic mice [6]. This resulted in normal looking embryonic kidney (Figure 3A and S1A), similarly to Mek1 deletion in epidermal keratinocytes [32]. Since ubiquitously expressed Mek2 can phosphorylate ERK1/2 and may compensate the loss of Mek1 in Hoxb7CreGFP;Mek1F/F kidneys, we reduced the gene dosages of Mek1 and −2 in UB epithelium. Conventional deletion of Mek2 alone, which results in phenotypically normal mice [31], or UB-specific double heterozygosity for Mek1 and –2, had no effect on UB branching, kidney differentiation or phosphorylation of ERK1/2 (Figure S1D–F and data not show). Normal UB branching pattern and renal differentiation were observed in vivo and in organ culture, even in the absence of three out of four Mek1 and −2 alleles regardless of allelic combinations (Figures 3B and S1A–L).
Next Mek1 was removed from UB in Mek2-/- background to examine effects on renal differentiation. Hoxb7CreGFP;Mek1F/F;Mek2-/- (henceforth called “dko” for double knock-out) mice were born in the expected Mendelian ratio (data not shown) but died within 72 h due to obvious renal defects, including severe renal hypodysplasia and sporadic hydroureters (16%) (Figure 3C, E and G). Histological examination showed disorganized medulla-cortex compartmentalization and few but well-differentiated glomeruli with associated tubuli and dilated epithelium (Figure 3D–E). Staining with the collecting duct epithelium marker calbindin and nephron segment markers Tamm-Horsefall and Na/K ATPase indicated that the cysts originate both in collecting ducts and secondarily in nephron tubules (Figure 3F–G and S1M–N).
Rudimentary kidneys in dko mice suggested that UB branching, the key process by which the kidney grows in size and acquires its typical shape could be perturbed. Time lapse imaging of in vitro cultured kidneys demonstrated a remarkable reduction in formation of novel branches in UB tips; average of 10.5 tips in controls was reduced to 3.8 in dko kidneys (Figure 4A–F, p<0.001 two-tailed T-test, n = 5). After generation of the primary UB at the correct time and with normal morphology, the subsequent epithelial morphogenesis in dko kidneys failed to start, and the UB tips usually elongated in only one direction (mean trunk lengths in controls: 142.3 µm, n = 83, three distinct kidneys, and in dko: 195.9 µm, n = 18, two distinct kidneys, two-tailed T-test, p<0.05; compare Figure 4A-C to D–F, S1O). A similar phenomenon was observed in intact dko kidneys imaged by confocal microscopy (Figure 4G–H). Typically very few if any UB tips at E13.5 had enlarged into T-bud resembling structures, which are signs of active branching. Thin UB tips were sparsely distributed in dko kidneys (average of 8.9 tips/kidney), leaving large areas of kidney devoid of UB branches, while in control kidneys UB epithelium was distributed over the entire cortical surface areas of kidneys (average of 41.9 tips/kidney, Figure 4G–I).
Formation of the primary UB at E11.5 in dko kidneys (Figure 4D) was surprising given the strong pERK1/2 staining in early kidneys (Figures 2A and 1A). This suggested that Mek1 might not yet be deleted by Hoxb7CreGFP in early UB epithelium, or that residual MAPK activity is maintained during the initiation of renal development. MEK1 was ubiquitously expressed in developing control kidneys while specifically lost from the UB of dko kidney from E11.5 (n = 5) onwards (Figure S2A–B), in contrast to pERK1/2, whose localization and staining intensities in dko kidneys were comparable to control kidneys at E10.5 (n = 5, data not shown) and E11.5 (n = 4, Figure S2C–D), and only abolished from the UBs at E12.5 onwards (Figure S2E–F and data not shown).
Application of exogenous GDNF to kidney cultures induces extra UB formation and swelling of UB tips (Figure S2H and [5]). We employed exogenous GDNF and chemical MEK-inhibition in kidney cultures to further test if the MAPK pathway is dispensable for UB outgrowth from the Wolffian duct. MEK-inhibition by UO126 dose-dependently blocked UB branching in kidney cultures (Figure S2I and K) mimicking the defects seen in dko kidneys (Figure 4D–F) and previous findings [10], [11]. Pretreatment with UO126 followed by application of GDNF in the presence of inhibitor blocked typical GDNF responses (Figure S2J), suggesting that the function of MEK1/2 cannot be overcome by RTK activation. Simultaneous UO126-inhibition and activation of RET by exogenous GDNF without UO126 pretreatment (Figure S2L) had the same effect. Thus, normal UB outgrowth in dko kidneys is likely due to delay in abolishing pERK1/2 activity at early stages of renal development.
As GDNF/RET signaling is the key RTK regulating UB morphogenesis in the normal context [7], we wanted to evaluate the linkage between RET and the MAPK pathway at the molecular level. Previous genetic engineering of Ret gene on the docking site known to mediate concurrent activation of MAPK and AKT cascades showed their importance for renal differentiation, but evidence for a specific requirement for MAPK activity downstream of RET was lacking [14], [17]. To address this, we examined if known GDNF/Ret signaling targets [33], [34] are regulated through MAPK pathway. In situ hybridization of ten GDNF/RET target genes in control and dko kidneys (Figures 5A–E and S3) revealed reduction in chemokine receptor Cxcr4 and in Spry1, a negative regulator of RTK signaling, which exerts its action at least partially by blocking MAPK pathway [35]. Downregulation of specific GDNF target genes in Mek1/2-deficient UBs suggested that MAPK pathway is an important intracellular mediator of RET signaling.
Various studies have shown that RTK signaling can promote proliferation through MEK/ERK pathway [23] and that ERK1/2 regulates G1/S transition in proliferating cells [36], while its function in G2/M transitions and M-phase remains ambiguous [37]–[39]. To reveal the cellular basis of defective UB branch formation in the absence of MAPK activity, we first examined proliferation at the onset of the morphologically distinct phenotype. Analysis of the mitotic indices at E12.5 showed that the percentages of pHH3+ UB epithelial cells were comparable in controls and dkos, but the amount of UB epithelium in dko kidneys was significantly reduced when quantified as total number of epithelial cells (Figure 6A–D). This data indicated that in the absence of Mek1 and −2, UB epithelial cells initially enter mitosis as efficiently as control cells suggesting that G2/M phase occurs independently of pERK1/2. However, at E14.5 dko UB was almost completely devoid of pHH3 (Figure S4A–B) showing a gradual decrease in mitosis. The reduced overall number of mutant UB cells indicated potential problems in cell survival or impairment in other cell cycle phases. Cleaved-caspase3+ apoptotic cells were very sporadically found in UB epithelium of both control and dko kidneys (Figure S4C–D) showing that increased cell death is not causing reduction in cell numbers of dko UB epithelium. Apoptosis was though slightly increased in renal mesenchyme, likely due to decreased UB-numbers in dko kidneys, which leave more mesenchymal cells without induction signal.
The S phase was studied by labeling the newly synthesized DNA with 5-ethynyl-2-deoxyuridine (EdU). Significantly fewer UB cells were EdU+ in the dko epithelium after 1 h pulse, revealing significant reduction of cells in S phase (Figure 6E–G). Since ERK is known to regulate the induction of cyclin D1 [40], [41], whose up-regulation is a key step for G1/S transition, and cyclin D1 at mRNA level is positively regulated by GDNF [33], it was a potential candidate for mediating the effect of the MAPK pathway on the cell cycle. In control kidneys cyclin D1 was up-regulated in early nephron progenitors and throughout the cortical UB epithelium (Figure S4E, G). In the absence of Mek1/2, only scattered, single cyclin D1 positive cells were very rarely detected in UB (Figure S4F, H), supporting the idea that MEK1/2-activated ERK functions in the G1/S transition phase during epithelial branching morphogenesis.
Greatly reduced formation of new branches in dko kidneys could involve alterations in cell adhesion properties either at the cell-to-matrix contacts made by focal adhesions, or at the cell-cell contacts formed by E-cadherin based adherens junctions. Since paxillin is a direct phosphorylation (Ser83) target of ERK in innermedullary collecting duct cells [25], we first studied the effect of chemical MEK-inhibition on pPaxillin in a ureteric bud derived cell line [42]. MAPK activity was dose-dependently inhibited by UO126, which also reduced significantly the level of pSer83 paxillin (Figure 7A–B). Immunofluorescence staining showed that inhibition of MAPK activity caused reduction in pPaxillin cytosolic pools but its plasmamembraneous localization was intensified together with E-cadherin, which appeared also stronger at the cell surfaces (Figure 7C–F, S5A–B). Simultaneously vinculin, another FA protein was also more pronounced in the cell surface (Figure 7G–H′). This preferential membranous localization was time dependent as after 2 h of UO126 such differences were not obvious (data not shown).
We next tested if genetic loss of MAPK activity could have changes in FA proteins by generating primary cell cultures from UBs isolated from control and dko kidneys. All control UBs dissected at E11.5–12.5 (n = 10) produced single cell monolayers in approximately 48 h in culture [21], [42], [43], but dko UBs isolated at E12.5 were slower in delaminating from the epithelium (7/7) and two out of seven samples failed to generate monolayers (data not shown). We thus isolated UBs from E11.5 dko kidneys (before pERK1/2 was lost and when the morphological phenotype was still comparable to controls) (Figure S2D, 4C–D), and found that monolayer formation was significantly improved (S5C–D′) but dko cells displayed thick, spiky E-cadherin at cell surfaces (Figure 7I, L). The dko cells remained tightly packed and impaired in spreading after 48 h of culture as seen by very strong and large appearance of FAs (Figure 7J–K, M–N).
Like FAs, AJs are constantly formed and disassembled during development and tissue homeostasis [44]. We next studied AJ molecule E-cadherin, which localized abundantly in the apical end of lateral cell walls of the control UB tips, being otherwise uniformly distributed along the lateral membranes (Figure 8A, C), while in medullary UB epithelium occasional staining was also observed in the basal end of lateral membranes (arrows in Figure S6A). The UB tips of dko kidneys displayed stronger overall E-cadherin staining than controls (Figure 8B, D, asterisks in S5B). Additionally, E-cadherin localization had shifted from apical to more basal sides of lateral membranes, and was also accumulating in basal membranes (Figure 8B, D) where it was rarely observed in control UB tips. Our visual observations were confirmed by quantitative measurements: E-cadherin intensities in general and the ratios of basal-to-lateral intensities were significantly higher both in cortical and medullary UB epithelium of dko than control kidneys (Figure S6C).
The cytoplasmic domain of E-cadherin is linked to the cytoskeleton through a complex of proteins including p120-, β- and α-catenins [45]. We analyzed β-catenin (Figure S6D–G″) and filamentous actin (Figure S6H and J), which appeared normal in dko kidneys. The function of p120 and related proteins ARVCF and p0071 is to cluster and stabilize E-cadherin at AJs [46], but they appeared normal in dko kidneys (data not shown). While the shift in E-cadherin subcellular localization could also reflect general problems in apical-basal polarity, the apical markers Par3 (data not shown) and ZO1 localized to apical cell surface similarly in control and dko UB epithelium (Figure S6I and K), indicating that apical cell polarity was maintained in the absence of MAPK activity. Staining of control and dko kidneys with tight junction marker claudin7 [47] revealed similar localization in wild type and dko UBs (Figure S6L–M) suggesting that MAPK activity is specifically involved in regulating E-cadherin mediated AJs.
To investigate potential functional consequences of the altered FAs and E-cadherin distribution, we analyzed UB epithelium by electron microscopy (EM). Control UB epithelium appeared as a uniform sheet of cells tightly apposed through cell-cell contacts on their lateral membranes, while the double mutant epithelium had increased extracellular space and many gaps between neighboring cells as seen by disintegration of the lateral membranes at several sites along their contact surface (Figure 8E–F, S7). However, the contact sites where adhesion was maintained appeared more electron dense.
By genetic studies focusing on MAPK pathway functions in renal development, we show that only one allele of either Mek1 or Mek2 is enough to support normal renal branching. Simultaneous deletion of both genes abolishes formation of new branches and complex 3D patterns and shows that this pathway is a necessary mediator of RTK signaling. The functional requirement for MEKs in UB morphogenesis is similar to that reported in skin [32] but different than what has been observed in placenta, where double heterozygosity results in embryonic lethality [48]. Similarly, Erk1;Erk2 gene dosage and protein levels are critical for survival and normal proliferation as shown by generation of double heterozygotes, most of whom die during gestation [49]. Such differences in functional requirements of MAPK activity imply tissue specific roles for the pathway components and call for investigations on their specific contributions to development and homeostasis.
While RET is an important RTK regulating renal branching, its exact cellular functions (e.g., proliferation, migration, extracellular matrix remodeling, changes in cell shapes and adhesive properties) and their precise contributions have remained indistinct [50]. Identification of GDNF/RET targets recently revealed several interesting candidates for mediating these functions [33], [34] but the intracellular cascades regulating the expression changes remain to be solved. Similarly to previous chemical inhibition studies [33], we saw that the expression of transcription factors Etv4 and -5 was normal whenever UBs were present and thus does not require MAPK activity. This is opposite to the requirement of tyrosine phosphatase Shp2, which appears to regulate Etv4 and -5 [12]. The negative RTK regulator, Spry1, which suppresses MAPK activity, and chemokine receptor Cxcr4, which is involved in migration in several cell types [51] were the only genes examined whose expression was reduced in dko UB epithelium. Our results suggest that different intracellular cascades regulate expression of specific target genes, and further experimentation will reveal potential genetic interactions between RET signaling and MAPK pathway.
We show that lack of MAPK activity disturbs cell proliferation by reducing the number of cells in G1/S without primarily affecting mitosis, which is in line with the finding that pERK1/2 staining was absent in the mitotic, pHH3+ cells. Reduction in markers of G1/S phase but nearly normal cell numbers in M phase suggest changes in cell cycle kinetics, so that the double mutant UB cells spend a longer time in M-phase than control cells. Such a phenomenon was demonstrated in human retinal pigment epithelia, where sustained inhibition of ERK1/2 transiently delays the cell cycle progression [37]. The fact that mutant kidneys are smaller than those in control mice, together with the reduced cell number in G1/S, supports the view that MAPK activity, through the regulation of cyclin D1 levels in UB epithelium, is required for progression from G1- to S-phase as shown previously for several cell types [52]. Consequently the defects in G1/S transition also have an effect on mitosis, which is reduced at later stages of kidney development.
In epithelial remodeling and morphogenesis, during which cells move relative to each other, cell-matrix and cell-cell contacts are constantly assembled and disassembled. We observed that UB epithelium lacking MAPK activity had significantly increased E-cadherin in general and particularly extending to more baso-lateral location than in controls, indicating problems in AJ dynamics. Rapid turnover of the E-cadherin-based homophilic AJs involves its endocytosis and recycling back to the plasma membranes [53], and defects in such processes could result in sustained cell surface localization of E-cadherin. During endocytosis, internalization of E-cadherin is initiated by tyrosine phosphorylation of E-cadherin, which induces its dissociation from p120 [54]. Normal distribution of p120 and related proteins (p0071 and ARVCF) in dko UBs suggests normal E-cadherin endocytosis, as increased localization of p120-family proteins at the cell surface would be expected if endocytosis was affected. However, normal appearance of p120 does not conclusively exclude changes in E-cadherin endocytosis or recycling.
Growing evidence indicates active crosstalk between different adhesion sites and this can be at least partially mediated by localization of certain proteins like vinculin and paxillin to both focal adhesions and AJs [26], [55]. Paxillin is a direct phosphorylation target of ERK-proteins and lack of pERK-induced phosphorylation on serine 83 of paxillin leads to functional defects in cell spreading and branching morphogenesis [24], [25]. We found accordingly that inhibition of MAPK activity in a ureteric bud-derived cell line [42] reduces general pPaxillin levels but also observed a simultaneous shift in its cellular localization to the cell-cell contacts, where concomitant increases in vinculin and E-cadherin levels were also detected. Additionally, primary cells derived from the double mutant UBs were impaired in their capacity to form monolayers, remained tightly packed and displayed stronger FAs than control cells, which all argue that such regulation is functionally significant. Both paxillin and vinculin bind to β-catenin in AJs, where at least vinculin has the potential to stabilize E-cadherin and, in certain cell types, to potentiate mechanosensing [27]–[29]. Taken together with the observation that lack of MAPK activity both in vivo and in vitro intensifies E-cadherin localization on cell surfaces, and recent evidence for cross-talk between FA and AJ, we suggest that MEK1/2-activated ERK1/2 regulates cellular adhesion by phosphorylating paxillin at Ser83, which facilitates normal FA dynamics and composition. The molecular mechanism by which lack of paxillin phosphorylation leads to increased plasmamembranous localization of itself and vinculin remains a subject of future studies, but non-receptor tyrosine kinases Src and focal adhesion kinase (FAK) as well the small GTPase Rac1 are shown to mediate shuttling between FAs and AJs in other cell types, and therefore are good candidates [55], [56].
Increased E-cadherin in the dko UB epithelium intuitively suggests stronger intercellular adhesion, but electron microscopy revealed big gaps in lateral membranes of adjacent cells. However the contact sites that were maintained in the dko epithelium appeared electron dense, suggesting that they might mediate strong contacts. Therefore it is possible that the gaps in lateral membranes are actually caused by increased adhesion at the localized sites where the contacts are maintained and then such stronger contacts generate ruptures to membranes next adhesion when cells are trying to move relative to each other. In support of this, vinculin in AJs promotes high E-cadherin based adhesion strength [57].
Our findings demonstrate the importance of Mek1/Mek2 in activation of MAPK pathway during UB branching, which is largely blocked in dko kidneys. Based on our results we suggest that MAPK activity regulates cyclin D1-mediated progression of cell cycle from G1 to S, and normal cellular adhesion through phosphorylation of paxillin in FAs. Absence of MAPK activity amplifies FA proteins vinculin and remnants of pPaxillin on cell surfaces where they participate in stabilization of E-cadherin to AJs possibly to strengthen intercellular adhesion. Taken together with heterogeneous pERK1/2 localization in UB epithelium, where certain cells show high ERK-activity and others no activity at all, this may imply that MAPK activity tunes cells for higher motility and thereby together with driving proliferation promotes novel branch formation.
Mek1-floxed, Mek2-null mice and Hoxb7CreGFP mice and their genotyping by PCR has been described [6], [31], [32]. All mice were on mixed genetic background with contributions from C57BL6/Rcc and 129/SvEv. Embryonic staging was as described earlier [58] and all experiments were approved by Finnish Animal Care and Use Committee.
NMRI or Hoxb7CreGFP mice were used for MEK-inhibition experiments. Kidneys were cultured on Trowell-type system in medium of F12:DMEM/10%FBS/Glutamax/penicillin-streptomycin, and imaged with epifluorescent microscope (see below). UO126 (Cell Signaling technologies, Inc.) dose-dependence testing was done with 5, 10, 15 and 50 µM by adding only the inhibitor, or UO126 in combination with GDNF (produced by Icosagen Ldt., see details in Supplementary information) either at the same time or sequentially. Already 5 and 10 µM concentrations of UO126 inhibited UB branching but upon replacement with GDNF resulted in some degree of response, while 15 µM mimicked closest the UB pattern in dko kidneys.
EdU (25 mg/kg) was injected intraperitoneally to pregnant females at day 12 of pregnancy. After an hour pulse the females were sacrificed and embryos collected for 4% PFA-fixation followed by standard processing for paraffin embedding (see below).
Embryos and tissues of indicated stages were collected and fixed with 4% PFA. Further processing for frozen and paraffin sections was according to standard procedures, the latter utilizing an automatic tissue processor (Leica ASP 200). The primary and secondary antibodies used are shown in Table S1. HE, whole mount and frozen section immunostaining were performed as previously described [20], [58]. Paraffin sections were gradually dehydrated followed by heat induced antigen retrieval. Sections were then blocked with 10% fetal bovine serum (Hyclone) for one hour at RT followed by o/n primary antibody incubation. Fluorescent detection was performed similarly to that with frozen sections and chromogenic detection was done with EnVision Detection System-HRP (DAB) kit (Dako). EdU incorporation detection step was done by using Click-iT EdU Alexa Fluor Imaging kit (Molecular Probes/Invitrogen) according the manufacturer's instructions after the antigen retrieval.
PFA-fixed E13 urogenital blocks were embedded in 4% low melting agarose (NuSieve GTG, Lonza) and 50 µm thick vibratome sections were cut (HM650, Microm Int.) for whole mount in situ hybridization performed with InSituPro automate (Intavis) according [59].
For EM, E11.5 and E12.5 kidneys were fixed with 2% glutaraldehyde (Fluka) followed by 1% osmiumtetroxide (EMS) post-fixation and graded series of dehydration. Transitional acetone incubation was performed before gradual embedding in Epon (TAAB Ltd.). Ultrathin sections (Leica ultracut UCT ultramicrotome, Leica) were collected on pioloform-coated single-slot copper grids and post stained with uranyl acetate and lead citrate for analyzing with transmission electron microscope (Tecnai G2 Spirit 120 kV TEM with Veleta and Quemesa CCD cameras) operated at 100 kV.
Epifluorescence images were produced with Zeiss Imager M2 Axio (Germany) equipped with Zeiss AxioCam HRm camera (Germany) and Axio Vision 4 software. Confocal imaging was done with mutiphoton Leica TCS SP5 MP confocal microscope (Germany) utilizing LAS-AF software. Chromogenic immunostaining was imaged with Olympus BX61 light microscope (Japan) equipped with Olympus Color View Soft Imaging System camera (Japan) and Cell F imaging software.
Quantification of mitotic indices from total of four control and four dko kidneys sectioned through to give 150 and 127 sections, respectively, and stained similarly as previously described [21] followed by counting of all UB epithelial cells in both groups at 40× magnification. EdU percentages were counted at 40× magnification from entire kidneys sectioned through (n = 4 for both groups). Proliferation significances were tested with Independent samples T-test (1-tailed, equal variances). Quantification of stalk lengths, as indicated in Figure S1O, was performed from 43 trunks of three distinct control kidneys, and 18 trunks from two distinct dko kidneys using Image J program. Statistical testing was done with Independent samples T-test (2-tailed, unequal variances) utilizing the SPSS software. For quantification of basal E-cadherin intensities, E13.5 UBs were imaged with multiphoton confocal microscope at 63× magnification with 251.8 nm optical section thickness. Intensity measurements were performed with Image J on 12-bit images with applied 2 pixel median filter. The average basal and lateral pixel intensities were obtained as mean grey values (MGV) from four tips/genotype so that 70 control cells and 97 dko cells in total were analyzed in the way indicated in Figure S6C′. Basal-to-lateral intensity ratios were calculated by dividing basal MGV with lateral MGV and cortical and medullar UB regions were compared with Independent samples T-test (two tailed, equal variances) employing SPSS program after removing three outlier ratios from the control and one from the dko set.
UB cells were cultured in DMEM/10%FBS/Glutamax/penicillin-streptomycin media. 1×105 cells for MEK-inhibition experiments (2 h) were plated 24 h before performing the inhibition by UO126 after which they were collected in lysis buffer. For paxillin phosphorylation, UB cells were serum-starved for 4 h and followed by fetal bovine serum (FBS) treatment of indicated times and 15 µM UO126 was used for inhibiting MAPK activity 30 min prior plus during the induction with FBS. Western blotting was done as previously described [60]. Rabbit anti-pErk1/2 (Cell Signaling, 1∶2000), anti-ERK2 (K-23, Sant Cruz, 1∶1000), p(S83) paxillin (ECM Biosciences, 1∶1000) and mouse paxillin (ECM Biosciences, 1∶1000) were used with HRP-conjugated secondary antibodies to detect proteins which were visualized by Pierce ECL Western Blotting detection system (Thermo Scientific) and Fuji film LAS1000.
For immunofluorescence staining cells were treated with 15 µM UO126 for 4 h, then fixed with 4% PFA for 10 min and stained with anti-E-cadherin (R&D Systems, 1∶300), pPaxillin, vinculin (Sigma, 1∶500) and 568-phalloidin followed by visualization of primary antibodies with corresponding Alexa-fluor secondary ABs (1∶400, Jackson Immuno Research). Cells were imaged by Zeiss Imager M2 Axio (see above).
All work on animals was conducted under PHS guidelines and approved by the relevant Institutional Animal Care and Use Committees.
|
10.1371/journal.pgen.1005703 | A Spatial Framework for Understanding Population Structure and Admixture | Geographic patterns of genetic variation within modern populations, produced by complex histories of migration, can be difficult to infer and visually summarize. A general consequence of geographically limited dispersal is that samples from nearby locations tend to be more closely related than samples from distant locations, and so genetic covariance often recapitulates geographic proximity. We use genome-wide polymorphism data to build “geogenetic maps,” which, when applied to stationary populations, produces a map of the geographic positions of the populations, but with distances distorted to reflect historical rates of gene flow. In the underlying model, allele frequency covariance is a decreasing function of geogenetic distance, and nonlocal gene flow such as admixture can be identified as anomalously strong covariance over long distances. This admixture is explicitly co-estimated and depicted as arrows, from the source of admixture to the recipient, on the geogenetic map. We demonstrate the utility of this method on a circum-Tibetan sampling of the greenish warbler (Phylloscopus trochiloides), in which we find evidence for gene flow between the adjacent, terminal populations of the ring species. We also analyze a global sampling of human populations, for which we largely recover the geography of the sampling, with support for significant histories of admixture in many samples. This new tool for understanding and visualizing patterns of population structure is implemented in a Bayesian framework in the program SpaceMix.
| In this paper, we introduce a statistical method for inferring, for a set of sequenced samples, a map in which the distances between population locations reflect genetic, rather than geographic, proximity. Two populations that are sampled at distant locations but that are genetically similar (perhaps one was recently founded by a colonization event from the other) may have inferred locations that are nearby, while two populations that are sampled close together, but that are genetically dissimilar (e.g., are separated by a barrier), may have inferred locations that are farther apart. The result is a “geogenetic” map in which the distances between populations are effective distances, indicative of the way that populations perceive the distances between themselves: the “organism’s-eye view” of the world. Added to this, “admixture” can be thought of as the outcome of unusually long-distance gene flow; it results in relatedness between populations that is anomalously high given the distance that separates them. We depict the effect of admixture using arrows, from a source of admixture to its target, on the inferred map. The inferred geogenetic map is an intuitive and information-rich visual summary of patterns of population structure.
| There are many different methods to learn how population structure and demographic processes have left their mark on patterns of genetic variation within and between populations. Model-based approaches focus on developing a detailed view of the migrational history of a small number of populations, often assuming one or a small number of large, randomly mating populations (i.e. little or no geographic structure). There has been considerable recent progress in this area, using a variety of summaries such as the allele frequency spectrum [1–3], or approximations to the coalescent applied to sequence data [4–6].
Other approaches are designed only to visualize patterns of genetic relatedness and population structure, without using a particular population genetic model. Such methods can deal with many populations or individuals as the unit of analysis. Examples of this second set of methods include clustering methods [7–9] and reduced dimensionality representations of the data (e.g. [10–12]).
A third set of methods that describe relatedness between populations by constructing a “population phylogeny” was pioneered by Cavalli-Sforza and Edwards [13], as were methods to test whether a tree is a good model of population history [14] (see [15] for a review). Tree-based approaches are appealing because trees are easy to visualize and explain, but the underlying assumptions (unstructured populations that split at discrete points in time) rarely hold true.
Recently, there has been a resurgence of interest in these tree-based methods. Some use population trees as a null model to test and quantify the signal of admixture between samples [16]. Others, such as TreeMix [17] and MixMapper [18], visualize population relationships using a directed acyclic graph; for instance, TreeMix connects branches in a population tree with additional edges to explain excess covariance between groups of populations.
There has also been renewed interest in methods for dimensionality reduction for the visualization of patterns of genetic variation [11], especially Principal Components Analysis (PCA; also pioneered by Cavalli-Sforza [19]). Examining such low-dimensional visual summaries has become an indispensable step in the analysis of modern genomic datasets of thousands of loci typed in tens or hundreds of samples. Generally, these visualizations are constructed by plotting the first few eigenvectors of the covariance matrix of normalized allele frequencies against each other.
Both PCA and tree-based methods are valuable as genetic inference and visualization tools, but both also suffer from serious limitations. Because gene flow is frequently pervasive, patterns of relatedness between samples may often be only poorly represented by a tree-based model. PCA is more flexible, as it assumes no explicit model of population-genetic processes, simply describing the axes of greatest variance in the average coalescent times between pairs of samples [20]. This allows PCA to describe more geographically continuous relationships: applied to human populations within continents, it often shows a close correspondence to geographic locations [21, 22]. However, the interpretation of PCA is more difficult, as the results can be strongly affected by the size and design of sampling, and the linearity and orthogonality requirements of the PC axes can lead to counterintuitive results [23–25].
What is desired, then, is a method for inferring and visualizing patterns of population differentiation that can recapitulate complex, non-hierarchical structures, while also admitting simple and intuitive interpretation. Since gene flow and population movements are often constrained by geography, it is natural to base such a method in a geographic framework. There is a rich history of population genetics theory for populations distributed in continuous space [26–29], as well as exciting new developments in the field [30]. The pattern of increasing genetic differentiation with geographic distance was termed “Isolation by Distance” by Wright [31], and is ubiquitous in natural populations [32]. Descriptive models of such patterns rely only on the weak assumption that an individual’s mating opportunities are spatially limited by dispersal; a large set of models, ranging from equilibrium migration-drift models to non-equilibrium models, such as recent spatial expansions of populations, give rise to the empirical pattern of isolation by distance.
In this paper, we present a statistical framework for studying the spatial distribution of genetic variation and genetic admixture based on a flexible parameterization of the relationship between genetic and geographic distances. Within this framework, the pattern of genetic relatedness between the samples is represented by a map, in which inferred distances between samples are proportional to their genetic differentiation, and long distance relatedness (in excess of that predicted by the map) is modeled as genetic admixture. These ‘geogenetic’ maps are simple, intuitive, low-dimensional summaries of population structure, and provide a natural framework for the inference and visualization of spatial patterns of genetic variation and the signature of genetic admixture. The implementation of this method, SpaceMix, is available at https://github.com/gbradburd/SpaceMix.
The genetic data we model consist of allele counts at L unlinked, bi-allelic single nucleotide polymorphisms (SNPs), sampled across K populations. After arbitrarily choosing an allele to count at each locus, denote the number of counted alleles at locus ℓ in population k as Ck,ℓ, and the total number of alleles observed as Sk,ℓ. The sample frequency at locus ℓ in population k is f ^ k , ℓ = C k , ℓ / S k , ℓ. Although we will refer to “populations”, each could consist of a single individual (Sk,ℓ = 2 for a diploid). We will depict results as coordinates on a map; however, the method does not require user-specified sampling locations.
We first compute standardized sample allele frequencies at locus ℓ in population k, by
X ^ k , ℓ = ( f ^ k , ℓ - f ¯ ℓ ) / f ¯ ℓ ( 1 - f ¯ ℓ ) , (1)
where f ^ k , ℓ is the sample allele frequency at locus ℓ in population k, and f ¯ ℓ is the average of the K sample allele frequencies, weighted by mean population size. This normalization is widely used [11, 33]; mean-centering makes the result invariant to choice of which allele to count at each locus, and dividing by f ¯ ℓ ( 1 - f ¯ ℓ ) makes each locus have roughly unit variance if the amount of drift since a common ancestor is small.
We work with the empirical covariance matrix of these standardized sample allele frequencies, calculated across loci, namely, Ω ^ = ( 1 / L ) X ^ X ^ T. Using the sample mean to mean-center X ^ has implications on their covariance structure, discussed in the Methods (“The standardized sample covariance”). For clarity, here we proceed as if f ¯ ℓ were instead an unobserved, global mean allele frequency at locus ℓ.
We wish to model the distribution of alleles among populations as the result of a spatial process, in which migration moves genes locally on an unobserved landscape. Migration homogenizes those differences between populations that arise through genetic drift; populations with higher levels of historical or ongoing migration share more of their demographic history, and so have more strongly correlated allele frequencies.
We assume that the standardized sample frequencies are generated independently at each locus by a spatial process, and so have mean zero and a covariance matrix determined by the pairwise geographic distances between samples. To build the geogenetic map, we arbitrarily choose a simple and flexible parametric form for the covariance matrix in which covariance between allele frequencies decays exponentially with a power of their distance [34–36]: the covariance between standardized population allele frequencies (i.e. X ^ values) between populations i and j is assumed to be, for i ≠ j,
F ( D i , j ) = 1 α 0 exp ( - ( α 1 D i , j ) α 2 ) , (2)
where Di,j is the geogenetic distance between populations i and j, α0 controls the within-population variance (or the covariance when distance between points is 0, known as a “sill” in the geospatial literature), α1 controls the rate of the decay of covariance per unit pairwise distance, and α2 determines the shape of that decay. Within-population variance may vary across samples due to either noise from a finite sample size or demographic history unique to that sample (e.g., bottlenecks or endogamy). To accommodate this heterogeneity we introduce population-specific variance terms, resulting in the covariance matrix for standardized sample frequencies
Ω i , j = F ( D i , j ) + δ i , j ( 1 S i ¯ + η i ) , (3)
where δi,j = 1 if i = j and is 0 otherwise, ηk is a nonnegative sample-specific variance term (nugget) to account for variance specific to population k that is not accounted for by the spatial model, and S ¯ k is the mean sample size across all loci in population k, so that 1 / S k ¯ accounts for the variance introduced by sampling within the population.
The distribution of the sample covariance matrix Ω ^ is not known in general, but the central limit theorem implies that if the number of loci is large, it will be close to Wishart. Therefore, we assume that Ω ^ is Wishart distributed with degrees of freedom equal to the number of loci (L) used and mean equal to the parametric form Ω given in Eq (3). We denote this by
P ( Ω ^ ∣ Ω ) = W ( L Ω ^ ∣ Ω , L ) . (4)
Note that if the standardized sample frequencies are Gaussian, then the sample covariance matrix is a sufficient statistic, so that calculating the likelihood of Ω ^ is the same as calculating the probability of the data up to a constant. Handily, it also means that once the sample covariance matrix has been calculated, all other computations do not scale with the number of loci, making the method scalable to genome size datasets. This modeling approach rests on the assumption that the loci in the dataset are independent, that is, not in linkage disequilibrium (LD). Linkage disequilibrium between loci included in the dataset will have the effect of decreasing the true number of degrees of freedom, effectively making this likelihood calculation a composite likelihood and artificially increasing confidence in parameter estimation. We discuss possible ways to accommodate linkage disequilibrium further in the Discussion.
Non-equilibrium processes like long distance admixture, colonization, or population expansion events will distort the relationship between covariance and distance across the range, as will barriers to dispersal on the landscape. To accommodate these heterogeneous processes we infer the locations of populations on a map that reflects genetic, rather than geographic, proximity. To generate this map, we treat populations’ locations (i.e. coordinates in the geogenetic map) as parameters that we estimate with a Bayesian inference procedure (described in the Methods). These location parameters for each population are denoted by G, and determine the matrix of pairwise geogenetic distances between populations, D(G), which together with the parameters α → and η determine the parametric covariance matrix Ω (given by Eq (3)). We acknowledge this dependence by writing Ω ( α → , D ( G ) , η ).
The prior distributions on the parameters that control the shape and scale of the decay of covariance with distance (α → and η) are given in the Methods. The priors on the geogenetic locations, G, are independent across populations; because the observed locations naturally inform the prior for populations locations, we use a very weak prior on population k’s location parameter (Gk) that is centered around the observed location. This prior on geogenetic locations also encourages the resulting inferred geogenetic map to be anchored in the observed locations and to represent (informally) the minimum distortion to geographic space necessary to satisfy the constraints placed by genetic similarities of populations. In practice, we also compare results to those produced using random locations as the “observed” locations, and can change the variance on the spatial priors to ascertain the effect of the prior on inference.
We then write the posterior probability of the parameters as
P ( G , α → , η ∣ Ω ^ , L ) ∝ P ( Ω ^ ∣ Ω ( α → , D ( G ) , η ) ) P ( α → ) P ( G ) P ( η ) , (5)
where P() denotes the various priors, and the constant of proportionality is the normalization constant.
We then use a Markov chain Monte Carlo algorithm to estimate the posterior distribution on the parameters as described in more detail in the Methods.
We first apply the method to several scenarios simulated using the coalescent simulator ms [37]. Each scenario is simulated using a stepping stone model in which populations are arranged on a grid with symmetric migration to nearest neighbors (eight neighbors, including diagonals) with 10 haploid individuals sampled from every other population at 10,000 unlinked loci (for details on all simulations, see Methods and Supplementary Materials). The basic scenario is shown in Fig 1a, which is then embellished in various ways. In the SpaceMix analysis of each simulated dataset, we treat population locations as unknown parameters to be estimated as part of the model, and center the priors on each population’s location at a random point. The resulting geogenetic maps are generated using the parameters that have maximum posterior probability. Since overall translation, rotation, and scale are nuisance parameters, we present inferred locations after a Procrustes transformation (best-fit rotation, translation, and dilation) to match the coordinates used to simulate the data. The axes of the resultant maps are presented as Northings and Eastings, as population locations in this geogenetic space no longer conform to the latitude or longitude of the original sampling locations. In S1 Fig, we show the relationship between genetic covariance, geographic distance, and inferred geogenetic distance for these simulations.
The lattice scenarios, illustrated in Figs 1 and 2, are: homogeneous migration rates across the grid; a longitudinal barrier across the center of the grid; a series of recent expansion events; and an admixture event between opposite corners of the lattice. In the simple lattice scenario with homogeneous migration rates (Fig 1a and 1d), SpaceMix recovers the lattice structure used to simulate the data (i.e., populations correctly find their nearest neighbors). After adding a longitudinal barrier to dispersal across which migration rates are reduced by a factor of 5 (Fig 1b), the two halves of the map are pushed farther away from one another, reflecting the decreased gene flow between them.
In the expansion scenario, in which all populations in the last five columns of the grid have expanded simultaneously in the immediate past from the nearest population in their row (Fig 1c), the daughter populations of the expansion event cluster with their parent populations, reflecting the higher relatedness (per unit of geographic separation) between them.
In all scenarios, populations at the corners of the lattice are pulled in somewhat because these have the least amount of data informing their relative placements, and because, without nearest-neighbor migration from farther outside the lattice, they are in fact more closely related to their neighbors.
We also examined the effects of uneven sampling on inference by subsampling a 9 × 9 grid into a variety of subsets that had successfully greater ‘uneven-ness’ of sampling, and comparing PCA and SpaceMix on these unevenly subsampled datasets. The results of these analyses are shown in S3–S9 Figs. As sampling becomes more uneven, the maps produced by plotting Principal Component axis 1 (PC1) against PC2 diverge more and more from the true geographic configuration of the samples (S10 Fig), quickly resembling the triangular shape commonly seen in PCA plots of real datasets. SpaceMix also becomes less certain about placement of some samples, but to a much smaller extent, and in all scenarios SpaceMix produces geogenetic maps that are more faithful to the true geographic configuration of the samples than those generated using PCA.
We next simulated a long-distance admixture event on the same grid, by sampling half of the alleles of each individual in the northeast corner population from the southwest corner population (Fig 2a). We then ran a SpaceMix analysis in which the locations of these populations were estimated (Fig 2b). The admixture creates excess covariance over anomalously long distances, which is clearly difficult to accommodate with a two-dimensional geogenetic map. Fig 2b shows the torturous lengths to which the method goes to fit a good geogenetic map: the admixed population 30 is between population 1, the source of its admixture, and populations 24, 25, and 29, the nearest neighbors to the location of its non-admixed portion. However, this warping of space is difficult to interpret, and would be even more so in empirical data for which a researcher does not know the true demographic history.
To incorporate recent admixture, we allow each allele sampled in population k to have a probability wk (0 ≤ wk ≤ 0.5) of being sampled from location G k *, which we refer to as population k’s source of admixture, and a probability 1 − wk of being sampled from location Gk. With no nugget, each allele would be sampled independently, but the nugget introduces correlations between the alleles sampled in each population.
With this addition, the parametric covariance matrix before given by Eq (3) becomes a function of all the pairwise spatial covariances between the locations of populations i and j and the points from which they draw admixture (illustrated in Fig 3); now, we model the covariance between X ^ i , ℓ and X ^ j , ℓ, for each ℓ, as
Ω i , j * = ( 1 - w i ) ( 1 - w j ) F ( D i , j ) + w i ( 1 - w j ) F ( D i * , j ) + w j ( 1 - w i ) F ( D i , j * ) + w i w j F ( D i * , j * ) + δ i , j ( η i + 1 / S ¯ i ) (6)
where D is the 2k × 2k matrix of pairwise distances between all inferred locations and sources of admixture, and for readability, we denote, e.g., F ( D ( G i , G j * ) ), as F(Di,j*). The spatial covariance, F(D), is as given in Eq (2), and we reintroduce the nugget, ηk, and the sample size effect, 1 / S k ¯, for each population as above in Eq (3).
We proceed in our inference procedure as before, but now with the locations of the sources of admixture and the admixture proportions to infer. The likelihood of the sample covariance matrix is exactly as before in Eq (4), except with Ω replaced by Ω*. The posterior probability of these parameters can be expressed as a function of this parametric admixed covariance, Ω*,
P ( G , G * , w , α → , η ∣ Ω ^ , L ) ∝ P ( Ω ^ ∣ Ω * ) P ( α → ) P ( G ) P ( G * ) P ( w ) P ( η ) (7)
as specified by the parameters w, G*, α →, and η, and the inferred locations, G. We place a weak spatial prior on the sources of admixture, G* around the centroid of the observed locations. The admixture proportions, w, are capped at 0.5, to ensure identifiability, and are heavily weighted towards small values to be conservative with respect to admixture inference. These priors are detailed in the Methods.
The models described above may be used in various combinations. In the simplest model, locations are not estimated for populations, nor do they draw admixture; the only parameters to be estimated are those of the spatial covariance function given in Eq (2), and the population-specific variance terms (ηi). In the most complex model, population locations, the locations of their sources of admixture, and the proportions of admixture are all estimated jointly in addition to the parameters of the spatial covariance function and the population specific variances. We discuss the utility of these different models in the Methods.
Allowing admixture gives sensible results for the scenario of Fig 2a; in the resulting map, the only population that draws substantial admixture is the one that is actually admixed, and it draws admixture (95% CI: 0.36—0.40) from the correct location (Fig 2c).
A more subtle simulated admixture scenario, with admixture proportion of 10% across a geographic barrier, is shown Fig 4a. The resulting SpaceMix map (Fig 4b), separates the east and west sides of the grid to accommodate the effect of the barrier, and the admixed population (population 23) draws admixture from very close to its true source (population 13), and in close to the correct amount (w ¯ ( 23 ) = 0 . 05; 95% CI = 0.02–0.08).
Another difficult scenario is shown in Fig 4c, where 40% admixture has occurred between two populations immediately adjacent to each other on either side of a barrier. Here, the admixed population 18 is correctly identified as admixed (Fig 4d); however, its intermediate genetic relationships are explained through an estimated location close to its true admixture source (population 13) and source of admixture (95% CI: 0.04–0.14) on the far margin of the half of the grid on its own side of the barrier. Because there is no sampled intervening population between admixed population 18 and its source of admixture 13, the model is able to explain population 18’s higher covariance with population 13 via its estimated location G(18), rather than via that of its source of admixture G ( 18 ) *. In each of these scenarios, the estimated admixture proportion is less than that used to simulate the data. This is due to the stringent prior we place against admixture. We discuss these examples further in the Methods.
To demonstrate the applications of this novel method, we analyzed population genomic data from two systems: the greenish warbler ring species complex, and a global sampling of contemporary human populations. Maps showing our sampling in these two systems are given in Fig 5, and information on the specific samples included is given in the Supplementary Materials, S1 and S2 Tables. For all analyses presented below, we centered the priors on location parameters at randomly chosen locations rather than at the observed geographical locations. Each geogenetic map shown here is the maximum a posteriori estimate (over all parameters), transformed by rotation, translation, and scaling to best fit inferred locations (G) to the observed latitude and longitudes (a full Procrustes transformation). As with the simulations described above, the axes of the geogenetic maps are presented in Eastings and Northings.
The greenish warbler (Phylloscopus trochiloides) species complex is broadly distributed in their breeding habitat around the Tibetan plateau, and exhibits gradients around the ring in a range of phenotypes including song, as well as in allele frequencies [38–40]. At the northern end of the ring in central Siberia, where the eastern and western arms of population expansion meet, there are discontinuities in call and morphology, as well as reproductive isolation and a genetic discontinuity [39, 41]. It is proposed that the species complex represents a ring species, in which selection and/or drift, acting in the populations as they spread northward on either side of the Tibetan plateau, have led to the evolution of reproductive isolation between the terminal forms.
The question of whether it fits the most strict definition of a ring species focuses on whether gene flow around the plateau has truly been continuous throughout the history of the expansion or if, alternatively, discontinuities in migration around the species complex’s range have facilitated periods of differentiation in genotype or phenotype without gene flow [42–44] (see Wake and Schneider [45] for discussion). Alcaide et. al [46] have suggested that the greenish warbler species complex constitutes a ‘broken’ ring species, in which historical discontinuities in gene flow have facilitated the evolution of reproductive isolation between adjacent forms.
To investigate this question, we applied SpaceMix to the dataset from Alcaide et. al [46], consisting of 95 individuals sampled at 22 distinct locations and sequenced at 2,334 SNPs, of which 2,247 were bi-allelic and retained for SpaceMix runs. These loci were treated as independent (i.e., un-linked). We discuss ways to accommodate linkage disequilibrium further in the Discussion.
We first ran SpaceMix on the population dataset, with no admixture. The resulting inferred map (Fig 6a) largely recapitulates the geography of the sampled populations around the ring. The Turkish population (TU, Phylloscopus trochiloides ssp. nitidus) clusters with the populations in the subspecies ludlowi, due to its recent expansion, but also has a relatively high nugget parameter (see S11 Fig panel a), reflecting the population history it does not share with its ludlowi neighbors. In the north, where the twin waves of expansion around the Tibetan Plateau are hypothesized to meet, the inferred geogenetic distance between populations from opposite sides of the ring was much greater than their observed geographic separation, reflecting the reproductive isolation between these adjacent forms (see S12 Fig).
We then ran the method allowing admixture (Fig 6b). The only population sample with appreciable admixture is the Stolby sample (ST; w = 0.19, 95% credible interval: 0.146-0.238; S13 Fig). This sample is known to be composed of an equal mixture of eastern plumbeitarsus and western viridanus individuals [46]. Multiple runs agreed well on the level of admixture of the Stolby sample (see S14 Fig). What does vary across runs is whether the Stolby sample has an estimated location by the viridanus cluster while drawing admixture from near the plumbeitarsus cluster, or vise versa; however, this is to be expected given the 50/50 nature of the sample’s makeup (S14 Fig). The somewhat intermediate position of the Stolby sample, and its non-50/50 admixture proportion, likely partially reflect the influence of the priors (S15 Fig).
We repeated these analyses (with and without admixture) with a dataset in which we treated each individual as the unit of analysis (Fig 7). No individual drew appreciable admixture (see S16 Fig for admixture proportions), and so we discuss the results with admixture (those without admixture are nearly identical, see S17–S19 Figs). As with the analysis on multi-sample populations, the results approximately mirror the geography of the individuals.
There are, however, a number of obvious departures in the individual geogenetic map from the population map. The most obvious is that the location of a pair of nitidus samples (in purple) is very far from the rest of the samples. These individuals appear to be closely related, and in the population-level analysis, this increase in shared ancestry was accounted for by a large nugget for the nitidus population (S11 Fig panel a). However, in the individual-level analysis, a nugget is estimated separately for each sample, so, the model must accommodate the much higher relatedness between this pair of individuals through estimated locations that are close to each other and far from the rest of the samples. The same phenomenon seems to be at work in determining the locations of a pair of individuals, one identified as P. t. ludlowi (Lud-MN3), one as P. t. trochiloides (Tro-LN11), as they also show an unusually low pairwise sequence divergence (see S20 Fig).
The split between viridanus and plumbeitarsus individuals (blue and red, respectively), in the north at the contact zone of the two waves of expansion, is clearer now than in the population-based analysis, as the estimated locations of individuals from the Stolby population are near their respective clusters. Although the geogenetic separation between the viridanus and plumbeitarsus individuals is greater than their geographic separation, they are still closer to each other than we would expect if all gene flow between the two was mediated by the southern populations, in which case we would expect the populations to form a line, with viridanus at one end and plumbeitarsus at the other. This horseshoe configuration, with viridanus and plumbeitarsus at its tips, is steady within and among runs of the MCMC and choice of position priors (see S18 Fig).
Is this biologically meaningful? A similar horseshoe shape appears when a principal components (PC) analysis is conducted and individuals are plotted on the first two PCs (see S21 Fig and [46]). However, as discussed by Novembre and Stephens [23], such patterns in PC analysis can arise for somewhat unintuitive reasons. If populations are simulated under a one dimensional stepping stone model, then plotting individuals on the first two PCs results in a horseshoe (e.g. see S22 Fig panel b) not because of gene flow connecting the tips, but rather because of the orthogonality requirement of PCs (see [23] for more discussion). In contrast, when SpaceMix is applied to data simulated on a one dimensional array of populations, the placement of samples is consistent with a line (see S22 Fig panels c and d). The proximity of viridanus and plumbeitarsus in geogenetic space may be due to gene flow between the tips of the horseshoe north of the Tibetan Plateau. This conclusion is in agreement with that of Alcaide et al. [46], who observed evidence of hybridization between viridanus and plumbeitarsus using assignment methods.
The SpaceMix map also diverges from the observed map in the distribution of individuals from the subspecies ludlowi (in green). These samples were taken from seven sampling locations along the southwest margin of the Tibetan Plateau, but, in the SpaceMix analysis, they partition into two main clusters, one near the trochiloides cluster, and one near the viridanus cluster. This break between samples from the same subspecies, which is concordant with the findings of Alcaide et al. [46], makes the ludlowi cluster unusual compared to the estimated spatial distributions of the other subspecies (see S23 Fig), and suggests a break in historic or current gene flow.
Human population structure is a complex product of the forces of migration and drift acting on both local and global scales, patterned by geography [21, 47], time [48, 49], admixture [50], landscape and environment [36, 51, 52], and shaped by culture [16, 53, 54]. To visualize the patterns these processes have induced, we create a geogenetic map for a worldwide sample of modern human populations. Of course, human history at these geographic scales has many aspects that are not well captured by static maps with discrete “arrows” of admixture. Nonetheless, we talk about the locations of samples and their sources of admixture as if these are fixed, even though both reflect the compounding of drift and gene flow over many historical processes. We therefore urge caution in the interpretation of our results, and view them as a simplistic but rich visualization of patterns of population structure.
We used the dataset of Hellenthal et al. [50], comprised of 1,490 individuals from 95 population samples (see Fig 5b for map of sampling), as well as the latitude and longitude attributed to each sample. In the analyses presented on human genotype data below, we have thinned the total dataset for LD in windows of 50 base-pairs, with a step-size of 5 base-pairs, and an upper limit of 0.2 on pairwise r2 [55, 56]. We then used a random subset of 10,000 SNPs to estimate the sample covariance.
We ran two sets of SpaceMix analyses: in the first, we estimated population sample locations, and in the second, we also allowed admixture. We note that few of the putative admixture events that we report have escaped the notice of previous investigators, which is unsurprising given the depth of recent attention on human admixture studies, particularly on the subset of these samples that are in the HGDP dataset [50, 57–60]. Below, when discussing a pattern we see in our analyses, we often cite other authors who have seen or suggested similar patterns. However, what is novel here is the ability to visualize these admixture events in a geographic context, and that these admixture signals stand out against a null model of migration in continuous space (rather than tree-based models).
When we only infer the location of each sample, the map roughly recapitulates the geography of the samples (Fig 8a), a result that holds nicely when we zoom in on the more heavily sampled area of Eurasia (Fig 8b). We see that samples both in the Americas and in Oceania lie close to the East Asian samples, but that they form two clusters on opposite sides. The proximity of these groups to the East Asians represents the fact that both groups share an ancestral population in the relatively recent past with East Eurasian populations, but the two expansions occurred independently. As in our simulations (Fig 1f) population expansions/bottlenecks have distorted the relationship between geographic and geogenetic distance. Geogenetic distances between samples within Africa are much greater than those between any other group (see S24 Fig), and the slope of the relationship between geographic and geogenetic distances between populations on each continent decays with distance from Africa. This pattern is consistent with a history of human colonization events characterized by serial bottlenecks [61–63] following an out-of-Africa expansion, and subsequent expansions into Western Eurasia, East Asia, the Americas, and Oceania (although see Pickrell and Reich [64] for a discussion of other models).
To investigate possible patterns of admixture further, we ran a SpaceMix analysis with admixture (results shown in Figs 9 and 10). The biggest change between the geogenetic map of human populations inferred with admixture and that without is the positioning of African samples with respect to the rest of the world. The relatively large geogenetic distances between these groups reflects the fact that Eurasian, North African, Oceanian, and American populations all share relatively large amounts of population history (and hence genetic drift) not shared with the Sub-Saharan African samples. Relative to the geogenetic map inferred without admixture, the inclusion of admixture shifts the estimated locations of admixed samples intermediate between Sub-Saharan Africa and North Africa/the Middle East toward one cluster or the other, which, in turn, pushes each of those major clusters to move relatively farther apart. The Ethiopian and Ethiopian Jewish samples have estimated locations closer to the Sub-Saharan samples than those of the North African samples, but draw substantial amounts of admixture (∼40%) from close to where the Egyptian sample has positioned itself in the the Middle East cluster, as do the Sandawe [65, 66]. The SanKhomani draw admixture from near Syria, which may reflect multiple distinct geographic sources of admixture [50, 67]. Interestingly the Bantu South African sample, though it has an estimated location near the other Bantu samples, draws admixture from close to the San populations. This is consistent with previous signals of the expansion of Bantu-speaking peoples into southern Africa [50, 66–68]. The inferred sample-specific drift parameters (the nuggets) are similar between runs with and without admixture (S25 Fig).
The majority of North African samples (Egyptian, Tunisian, Moroccan, Mozabite) join the Middle Eastern samples (positioned in rough accord with their sampling location along North Africa), and draw admixture from near the Ethiopian samples. All of the Middle Eastern samples draw admixture from close to the geogenetic location of the Ethiopian samples and where most of the North African samples draw admixture from, representing the complex history of North African–Middle Eastern gene flow [50, 69].
A number of other population samples draw admixture from Africa. The Sindhi, Makrani, and Brahui draw admixture from close to the location of the Bantu samples [50], and the Balochi and Kalash draw admixture from some distance away from African population samples. Of the European samples, the Spanish and both East and West Sicilian samples draw small amounts of admixture from close to the Ethiopian samples, presumably reflecting a North African ancestry component [54, 70].
The other significant signal of admixture is between East and West Eurasia, a signal documented by many authors [50, 57, 58, 71]. The majority of samples maintain their relative positions within each of these groups; however, there are several samples that show admixture between eastern and western Eurasia. The Uzbekistani and Hazara samples have estimated locations close to the East Asian samples and draw a substantial admixture proportion from close to the Georgian and Armenian estimated locations. Conversely, the Uygur sample has an estimated location close to the Burusho, Kalash, and Pathan samples, and draws admixture from near the Mongola and Hezhen samples. The Tu sample (with a geogenetic location in East Asia) draws a small amount of ancestry from close to the estimated location of the Uygur. The estimated location of the Chuvash sample is near the Russian and Lithuanian samples, and the Chuvash draw admixture from close to the Yakut (as do the Turkish, to a smaller extent). There are several other East-West connections: the Russian and Adygei samples have admixture from a location “north” of the East Asian samples, and the Cambodia sample draws admixture from close to the Egyptian sample [17, 50].
There are also a number of samples that draw admixture from locations that are not immediately interpretable. For example, the Hadza and Bantu Kenyan samples draw admixture from somewhat close to India, and the Xibo and Yakut from close to “northwest” of Europe. The Pathan samples draw admixture from a location far from any other samples’ locations, but close to where the India samples also draws admixture from. The Myanmar and the Burusho samples both draw admixture far from the locations estimated for other samples as well.
There are a number of possible explanations for these results. As we only allow a single admixture arrow for each sample, populations with multiple, geographically distinct sources of admixture may have estimated admixture locations that average over those sources. This may be the case for the Hadza and Bantu Keynan samples [50]. A second possibility is that the relatively steep prior on admixture proportion forces samples to draw lower proportions of admixture from locations that overshoot their true sources; this may explain the Xibo and Yakut admixture locations. A final explanation is that good proxies for the sources of admixture may not be included in our sampling, either because of of the limited geographic sampling of current day populations, or because of old admixture events from populations for which there are not other more direct modern descendant populations. The admixture into the Indian and Pathan samples (whose admixture source also clusters with the Indian Jew samples in some MCMC runs) may be an example of this; Reich et al. [16] and Moorjani et al. [72] have hypothesized that many populations from the Indian subcontinent may be descended from an admixture event involving an ancestral Southern Indian population not otherwise represented in this dataset.
In S26 and S27 Figs, we show the results of other independent MCMC analyses on these data. The broad-scale patterns and results discussed above are consistent across these runs. However, as is to be expected, there is significant heterogeneity in the exact layout of sample and admixture locations. For example, there is some play, among MCMC runs, in the internal orientation of the African locations with respect to the east-west axis within the Eurasian cluster. For some samples that draw a significant amount of admixture, such as the central Asian populations (Uygur, Hazara and Uzbekistani), the estimated location switches with that of their source of admixture (as was also seen across MCMC runs in the warbler data analysis). Similarly the Ethiopian and Ethiopian Jew samples have estimated locations, in some MCMC runs, close to the other North African samples, and draw admixture from near the Sub-Saharan samples (as do the other North African samples).
In this paper we have presented a statistical framework for modeling the geography of population structure from genomic sequencing data. We have demonstrated that the method, SpaceMix, is able to accurately present patterns of population structure in a variety of simulated scenarios, which included the effects of uneven sampling, spatially heterogeneous migration, population expansion, and population admixture. In empirical applications of SpaceMix, we have largely recovered previously estimated population relationships in a circum-Tibetan sample of greenish warblers and in a global sample of human populations, while also providing a novel way to depict these relationships. The geogenetic maps SpaceMix generates serve as simple, intuitive, and information-rich summaries of patterns of population structure. SpaceMix combines the advantages of other methods for inferring and illustrating patterns of population structure, using model-based inference to infer population relationships (like TreeMix [17], and MixMapper [18]), and producing powerful visualizations of genetic structure on a map (like PCA [11] and SPA [73]).
The patterns of genetic variation observed in modern populations are the product of a complex history of demographic processes. We choose to model those patterns as the outcome of a spatial process with geographically determined migration, and we have included statistical elements to accommodate deviations from spatial expectations. However, the true history of a sample of real individuals is vastly more complex than any low-dimensional summary, and, as with any summary of population genetic data, SpaceMix results should be interpreted with this in mind. Furthermore, our “admixture” events are shorthands for demographic relationships that occurred over possibly substantial lengths of time and regions of the globe; approximating this by a single arrow between two points on a map is certainly an oversimplification. Aspects of population history that are better described as a population phylogeny may be difficult to interpret using SpaceMix, and may be better suited to visualization with model-based clustering-based methods [7] or TreeMix/MixMapper-like methods [17, 18]. There is obviously no one best approach to studying and visualizing population structure; investigators should employ a range of appropriate methods to identify those that provide useful insight.
SpaceMix, like PCA, is well suited to describing population structure in a continuous fashion–but it also has a number of advantages over PCA. PCA is a general-purpose tool for exploratory visualization of high-dimensional data; in application to genetic data, PCA can quickly identify problematic samples and major axes of variation. Since geography is a major cause of differentiation, the first one or two PC axes often correspond to geography [23]. However, because PCs are linear functions of the genotypes, sometimes many PCs must be used to depict patterns produced by simple isolation by distance [23].
These higher order PCs can be hard to interpret in empirical data (see discussion in the warbler section). The recently introduced SPA approach [74], which also assumes allele frequencies are monotonically increasing in a given direction, may suffer from the same problem, which SpaceMix avoids (although PCA and SPA are both significantly faster than SpaceMix). Similarly, unevenness of sampling can greatly distort PC maps, as illustrated in the comparison of the uneven subsampling simulation scenarios shown in S8 and S9 Figs.
The generality of PCA is also its weakness: it displays any structure, not necessarily geographical structure. SpaceMix, since it works explicitly with local correlations on maps, is designed to visualize the relationships between samples induced by geographically limited dispersal, and so is less easily misled by other types of structure. Our explicit modeling of admixture is also helpful; in PCA, admixed individuals appear in intermediate locations in PC biplots, but are not distinguished from individuals in intermediate populations.
The application of SpaceMix to humans illustrates the utility of our approach: the first two PCs of this dataset resemble a triangle (S28 Fig), with its arms corresponding to the Africa/Non-Africa split and the spread of populations across Eurasia. In contrast, while the SpaceMix geogenetic map is dominated by the genetic drift induced by migration out of Africa, it also captures much more detail than is contained in the first two PCs (e.g., Fig 9b). SpaceMix’s explicitly geographic model avoids the tendency of PC biplots towards triangular plots, as was also seen when applied to unevenly sampled datasets (S3–S9 Figs).
An advantage of PCA is that it can explain more complex patterns of population structure by allowing up to K different axes. Although SpaceMix can easily be extended to more than two dimensions, simply by allowing Gi to describe the location of a sample in d dimensions, interpretation and visualization of these higher dimensions is more difficult, and so we have stuck to two dimensions. On the other hand, SpaceMix can describe in two dimensions patterns that PCA, due to the constraints of linearity, would need more to describe.
Our method shows the utility of representing both isolation by distance and long-distance admixture on a 2-D geogenetic map. While we generate this map using likelihood-based inference relating a parametric covariance matrix to the observed empirical matrix, it would be interesting to explore other methods of creating this geogenetic map (e.g., [30, 74–76]). Such methods may offer computational speedups and also potentially help place SpaceMix within a broader statistical framework.
One of the greatest strengths of SpaceMix is the introduction of admixture arrows. Although PCA can be interpreted in light of simple admixture events [20], and new methods can locate the recent, spatially admixed ancestry of out-of-sample individuals [73, 74], neither approach explicitly models admixture between multiple geographically distant locations, as SpaceMix does. Assignment methods are designed to deal with many admixed samples [7], but they have no null spatial model for testing admixture. We feel that an isolation by distance null model is often more appropriate for testing for admixture, especially when there is geographically dense sampling. SpaceMix offers a useful tool to understand and visualize spatial patterns of genetic relatedness when many samples are admixed.
As currently implemented, SpaceMix allows each population to have only a single source of admixture, but some modern populations draw substantial proportions of their ancestry from more than two geographically distant regions. In such cases the inferred source of admixture in a SpaceMix analysis may fall between the true locations of the parental populations. Although it is statistically and computationally feasible to allow each population to choose more than one source of admixture, we were concerned about both the identifiability and the interpretability of such a model, and have not implemented it. However, there may be empirical datasets in which such a modeling scheme is required to effectively map patterns of population structure. In addition, we have assumed that only single populations are admixed, when in fact it is likely that particular admixture events may affect multiple samples.
One concern is that the multiple admixed samples (from a single admixture event) may simply have clustered estimated locations, and not need to draw admixture from elsewhere due to the fact that their frequencies are well described by their proximity to other admixed populations. Along these lines, it is noticeable that many of our European samples draw little admixture from elsewhere (also noted by [50] using a different approach), despite evidence of substantial ancient admixture [77]. This may reflect the fact that all of the European samples are affected by the admixture events, and are relatively over-represented in our sample. However, this may also simply reflect the fact that the admixture is ancient, and that the ancient populations that took part in these events are not well represented by our extant sampling. Reassuringly, we see multiple cases where similarly admixed populations (Central Asians, Middle Eastern, and North African) populations are separately identified as admixed. This suggests that geogenetic clustering (in lieu of drawing admixture) of populations that share similar histories of admixture is not a huge concern (at least in some cases). The method could in theory be modified to allow geogenetically proximal populations to draw from the same admixture event; however, this may be difficult to make fully automated.
In this paper, we have treated the loci in the dataset as independent, and, where necessary, we have thinned empirical datasets to decrease LD between loci. One possible approach that avoids the necessity of thinning the data would be to calculate the sample covariance in large (e.g., megabase), non-overlapping windows along the genome, then average those sample covariances across all windows. Another approach is to use empirical LD between loci to estimate the effective number of independent loci in the dataset, and use this quantity as the number of degrees of freedom in the Wishart likelihood calculation. Additionally, although we have focused on the covariance among alleles at the same locus, linkage disequilibrium (covariance of alleles among loci) holds rich information about the timing and source of admixture events [50, 72, 78, 79] as well as information about isolation by distance [47]. Just as population graph approaches have been extended to incorporate information from LD [59], a spatial covariance approach could be informed by LD. A null model inspired by models of LD under isolation by distance models [80, 81] could be fitted, allowing the covariance among alleles to decay with their geographic distance and the recombination distance between the loci. In such a framework, sources and time-scales of admixture could be identified through unusually long-distance LD between geographically separated populations.
The landscape of allele frequencies on which the location of populations that were the source of population’s admixture are estimated is entirely informed by the placement of other modern samples, even though the admixture events may have occurred many generations ago. This immediately leads to the caveat that, instead of “location of the parental population,” we should refer to the “location of the closest descendants of the parental population.” The increased sequencing of ancient DNA (see Pickrell and Reich [64] for a recent review) promises an interesting way forward on that front, and it will also be exciting to learn where ancient individuals fall on modern maps, as well as how the inclusion of ancient individuals changes the configuration of those maps [49]. The inclusion of ancient DNA samples in the analyzed sample offers a way to get better representation of the ancestral populations from which the ancestors of modern samples received their admixture. However, it is also possible to model genetic drift as a spatiotemporal process, in which covariance in allele frequencies decays with distance in both space and in time. We are currently exploring using ancient DNA samples as ‘fossil calibrations’ on allele frequency landscapes at points in the past, so that modern day samples may draw admixture from coordinates estimated in spacetime.
Here we describe in more detail the algorithm we use to estimate the posterior distribution defined by Eq (7) of the population locations, G, their sources of admixture, G*, their admixture proportions, w, their independent drift parameters, η, and the parameters of the model of isolation by distance, α →. First, we give the exact form of the covariance matrix we use, and then describe the Markov chain Monte Carlo algorithm that samples parameter values from the posterior distribution.
As motivation, consider several randomly mating (Wright-Fisher) populations that all split from an ancestral population in which a neutral allele is present at frequency ϵℓ, and then subsequently exchange migrants. Since the allele is neutral, the mean change in its frequency in each population after t generations is zero, and if t is much smaller than the population size (so the frequencies remain close to ϵℓ), the variance is proportional to ϵ ℓ ( 1 - ϵ ℓ ). Conveniently, additional variance introduced by binomial sampling of alleles is also proportional to ϵ ℓ ( 1 - ϵ ℓ ). It would then be natural to consider the covariance matrix of
X k , ℓ = f ^ k , ℓ - ϵ ℓ ϵ ℓ ( 1 - ϵ ℓ ) , (8)
since these standardized allele frequencies would be independent if the loci are unlinked, and would have mean zero and variance independent of the sample sizes or allele frequencies. The central limit theorem would then imply that in the limit of a large number of loci, the sample covariance matrix XT X is Wishart with degrees of freedom equal to the number of loci and mean determined by the pattern of migration.
Although the conditions are not strictly met, these theoretical considerations indicate that such a normalization may be a reasonable thing to do, even after substituting the empirical mean allele frequency f ¯ ℓ in place of ϵℓ, which is what we do to define X ^ k , ℓ in Eq (1). Recall that the sample allele frequency at locus ℓ in population k is given by f ^ k , ℓ = C k , ℓ / S k , ℓ, where Ck,ℓ is the number of (arbitrarily chosen) counted alleles, and Sk,ℓ is the total number of sampled alleles. As sample size may vary across loci, we first calculate S ¯ k, the mean sample size in population k, as S ¯ k = 1 L ∑ ℓ = 1 L S k , ℓ. We then compute the global mean allele frequency at locus ℓ as
f ¯ ℓ = 1 ∑ K S k , ℓ ∑ K f ^ k , ℓ S k , ℓ . (9)
If sample size were constant across all loci in each population, this would be equivalent to defining the variance-normalized sample frequencies
Y ^ k , ℓ = f ^ k , ℓ f ¯ ℓ ( 1 - f ¯ ℓ ) (10)
and writing X ^ ℓ = T Y ℓ where T is the mean centering matrix whose elements are given by
T i j = δ i , j - S ¯ j ∑ k = 1 K S ¯ j , (11)
where δi,j = 1 if i = j and is 0 otherwise. If the covariance matrix of Y is Ω*, then the covariance matrix of X ^ ℓ would be TT Ω*T. Since allowing T to vary by locus would be computationally infeasible, we make one final assumption, that the covariance matrix of the standardized frequencies X ^ ℓ at each locus is given by TT Ω*T. This makes it inadvisable to include loci for which there are large differences in sample sizes across populations. This mean centering acts to to reduce the covariance among populations in X ^ ℓ compared to f ^ ℓ, and can induce negative covariance between more unrelated populations (as, across loci, they are often on opposite sides of the mean).
Additionally, the covariance matrix of the standardized frequencies has rank K − 1 rather than K, and so the corresponding Wishart distribution is singular. To circumvent this problem we compute the likelihood of a (K − 1)-dimensional projection of the data. Any projection would do; we choose a projection matrix Ψ by dropping the last column of the orthogonal matrix in the QR decomposition of T, and compute the likelihood of the empirical covariance matrix of allele frequencies Ω ^ = X ^ T X ^ as
P ( Ω ^ ∣ Ω * ) = W ( L Ψ T X T X Ψ ∣ Ψ T Ω * Ψ , L ) . (12)
The inference algorithm described here may be used to estimate the parameters with any of these held fixed, for instance: (1) population locations are fixed, and they do not draw any admixture; (2) population locations are estimated, but not admixture; (3) populations may draw admixture, but their own locations are fixed; or (4) population locations and admixture are both estimated. The free parameters for each of options are given in Table 1.
Although we anticipate most empirical researchers will be interested in the joint inference of a geogenetic map with admixture (Model 4), we have presented these models separately, as we believe each have their own utility. Model 1 and Model 3 can each be used to infer landscapes of allele frequencies, upon which genotyped individuals can be probabilistically placed (following [35]). This application may be useful to determine the geographic origin of potentially contraband biological samples (e.g., ivory), or the most likely source of museum specimens missing sampling metadata. Model 3 has the potential to improve the performance of these spatial assignment methods over Model 1, as the inclusion of admixture in the model may allow for more accurate inference of allele frequency surfaces. Model 2 directly parallels Principal Component Analysis. Informally the visual comparison of Models 2 and 4 can allow investigators to understand how ignoring long distance admixture distorts relationships among populations. Formally, the fit of these various models could be compared by cross validation, but we do not implement that here.
Below, we outline the inference procedure for the most parameter-rich model (inference on both population locations, their sources of admixture, and the proportions in which they draw admixture, in addition to inference of the parameters of the spatial covariance function). A table of all parameters, their descriptions, and their priors is given in Table 2.
We now specify in detail the Markov chain Monte Carlo algorithm we use to sample from the posterior distribution on the parameters, for Bayesian inference.
We assume that the user has specified the following data: the allelic count data, C, from K population over L variant loci, where Ck,ℓ gives the number of observations of a given allele at locus ℓ in population k; the sample size data, S, from K population over L variant loci, where Sk,ℓ gives the total number of alleles typed at locus ℓ in population k.
It is not necessary, but a user may also specify the geographic sampling locations, G(obs), from each of the K populations, where G k ( o b s ) gives the longitude and latitude of the kth sampled individual(s).
The geographic location data may be missing, or generated randomly, for some or all of the samples; if so, the spatial priors on estimated population locations, G, and their sources of admixture, G* will not be tethered to the true map.
We ran our simulations using a coalescent framework in the program ms [37]. The full command line arguments for all simulations are included in the S1 Text. Briefly, we simulated populations on a lattice, with nearest neighbor migration rate mi,j, as well as migration on the diagonal of the unit square at rate m i , j / 2. For each locus in the dataset, we used the -s option to specify a single segregating site, and then we simulated 10,000 loci independently, which were subsequently conglomerated into a single dataset for each scenario. For all simulations, except the “Populations on a line” scenario (S22 Fig), we sampled only every other population, and, from each population, we sampled 10 haplotypes (corresponding to 5 diploid individuals). In the “Populations on a line” scenario, we simulated no intervening populations, such that every population was sampled.
To simulate a barrier event, we divided the migration rate between neighbors separated by the longitudinal barrier by a factor of 5. To simulate an expansion event, we used the -ej option to move all lineages from each daughter population to its parent population at a very recent point in the past. For admixture events, we used the -es and -ej options to first (again, going backward in time) split the admixed population into itself and a new subpopulation of index k + 1, and second, to move all lineages in the (kth+1) into the source of admixture. Forward in time, this procedure corresponds to cloning the population that is the source of admixture, then merging it, in some admixture proportion, with the (now) admixed population.
A natural concern is whether all of the parameters we infer are separately identifiable, most notably whether population locations, admixture locations, and proportions can be estimated. That is, if a population has received some admixture from another population, what is to stop it from having an estimated location near that population in geogenetic space to satisfy its increased resemblance to that population, rather than drawing admixture from that location? We do not provide a formal proof, but here build and illustrate some relevant intuition.
Admixture is identifiable in our model because there are covariance relationships among populations that cannot simply be satisfied by shifting population locations around (as demonstrated by the tortured nature of Fig 2b). Consider the simple spatial admixture scenario shown in Fig 11. Populations A–D are arrayed along a line, but there is recent admixture from D into B (such that 40% of the alleles assigned to B are sampled from location D). The lines show the expected covariance under isolation by distance that each population (A, C, or D, as indicated by line color) has with a putative population at a given distance. The dots show the admixed covariance between B and the three other populations, as well as B’s variance with itself (B-B) as specified by Eq (6), with no nugget or sampling effect.
Due to its admixture from D, B has lower covariance with A than expected given its distance, somewhat higher covariance with C, and much higher covariance with D. In addition, the variance of B is lower than that of the other three populations, which each have variance 1: the value of the covariance when the distance is zero. This lower variance results from the fact that the frequencies at B represent a mixture of the frequency at D and the frequency at B before the admixture.
Now, using this example scenario, let us return to the concern posed above: that admixture location and population location are not identifiable. For the sake of simplicity, assume that we hold the locations of A, C, and D constant, as well as the decay of covariance with distance (as could be the case if A–D are part of a larger analysis). The covariance relationships of B to the other populations cannot be simply satisfied if B had an estimated location near D, as B would then have a covariance with C that is higher, and a covariance with A that is lower, than that we actually observe.
Introducing admixture into the model allows it to satisfy all of these conditions: it can draw ancestry from D but keep part of its resemblance to A, it avoids B having an estimated location too close to C, and it explains B’s low variance. Even in the absence of a sample from population C, B is better described as a linear mixture of a population close to A and D. However, there are specific scenarios in which a limited sampling scheme (both in size and location), can lead to tradeoffs in the likelihood between estimated population location and that of its source of admixture.
The analyses depicted in Fig 2c, Fig 4b and 4c, give examples of these tradeoffs. In each, the inferred admixture proportions in the admixed populations are less than those used to simulate the data, and the model is able to explain the high covariance the admixed populations have with their sources of admixture via their inferred location, rather than just via their inferred source of admixture and admixture proportion. The reason the model explains these admixed populations’ anomalous covariance with their inferred location, rather than with their admixture source, is that we place a very harsh prior against admixture inference (Table 2). The prior is designed to make inference conservative with respect to admixture, but it has the side effect of skewing the posterior probability toward lower admixture proportions.
Below, we describe the specifics of our analyses of the greenish warbler dataset and the global human dataset. The analysis procedure for each dataset is given here:
For each analysis,
Five independent chains were run for 5 × 106 MCMC iterations each in which population locations were estimated (but no admixture). Population locations were initiated at the origin (i.e. at iteration 1 of the MCMC, Gi = (0,0)), or at uniformly distributed coordinates between the minimum and maximum observed range of latitude and longitude, and all other parameters were drawn randomly from their priors at the start of each chain. The chain with the highest posterior probability at the end of the analysis was selected and identified as the “Best Short Run”. A chain was initiated from the parameter values in the last iteration of the Best Short Run. Because inference of admixture proportion and location was not allowed in the five initial runs, admixture proportions were initiated at 0 and admixture locations, G* were initiated at the origin. This chain (the “Long Run”) was run for 108 iterations, and sampled every 105 iterations for a total of 1000 draws from the posterior.
For each dataset, we ran two analyses using the observed population locations as the prior on G. Then, to assess the potential influence of the spatial prior on population locations, we ran one analysis in which the observed locations were replaced with random, uniformly distributed locations between the observed minima and maxima of latitude and longitude. For the warbler dataset, we repeated this analysis procedure, treating each sequenced individual as its own population. For clarity and ease of interpretation, we present a full Procrustes superimposition of the inferred population locations (G) and their sources of admixture (G*), using the observed latitude and longitude of the populations/individuals (G) to give a reference position and orientation. As results were generally consistent across multiple runs for each dataset regardless of the prior employed, we (unless stated otherwise) present only the results from the ‘random’ prior analyses.
Finally, we compared the SpaceMix map to a map derived from a Principal Components Analysis [11]. For this analysis, we calculated the eigendecomposition of the mean-centered allelic covariance matrix, then plotted individuals’ coordinates on the first two eigenvectors (e.g., [21]). For consistency of presentation, we show the full Procrustes superimposition of the PC coordinate space around the geographic sampling locations.
To contrast our approach to tree-based approaches to admixture we applied TreeMix [17] to our spatial simulations. We took the ms output on which we had also run Spacemix (Figs 1, 2 and 4), and converted it into TreeMix format. We also included an outgroup sequence for each dataset, which consisted of a single haploid individual who carried the 0 (ancestral allele) at every locus. We ran TreeMix without migration edges on the processed ms.treemix.file.gz file to construct the initial tree, which was rooted using the myoutgroup sequence using the following command:
treemix -i ms.treemix.file.gz -root myoutgroup -o treemix_output
We then sequentially added admixture migration edges, using the following command to add another edge to the existing tree (“prev.treemix”):
treemix -i ms.treemix.file.gz -root myoutgroup -g prev.treemix.vertices.gz prev.treemix.edges.gz -m 1 -o treemix_output
The TreeMix graphs and residual covariance matrices were visualized using the scripts provided with TreeMix.
In S30 Fig we show the tree and admixture graphs produced when TreeMix is run on a lattice stepping stone model (a smaller scale version of this exercise was previously done by Pickrell and Pritchard [17]). The tree produced by running TreeMix is rake-like, showing the lack of deep shared sub-division. However, while the tree captures some features of isolation by distance (e.g., neighboring samples are often sister to each other), the tree structure forces many unnatural splits of geographically neighboring populations (as was previously found [17]; see their S14 Fig). The admixture migration arrows act to mitigate the strongest departures from the tree, such as geographically neighboring samples that were forced into separate places on the tree, but are unable to fully accommodate the spatial relationships between samples. Different runs of TreeMix on the same dataset result in quite different trees and orders of migration events, reflecting both the high degree of symmetry in our simulated samples on a grid, and also the poor fit of the tree model to spatial data. In S31 and S32 Figs, we also present the results of TreeMix run on our expansion and barrier simulations.
In S33, S34 and S35 Figs, we show the application of TreeMix to the scenarios simulated with admixture events. For none of our scenarios was the true admixture the first migration edge added; in fact, only for the corner admixture scenario was the true admixture event in the first three edges added. This reflects the fact that TreeMix has to add migration edges to cope with the residual covariance induced by the poor fit of a tree to spatially simulated data, and so misses more subtle (but real) admixture events. The poor performance of TreeMix here is the result of the spatially simulated data not conforming to the underlying assumption of the TreeMix tree-like model.
|
10.1371/journal.pgen.1005174 | The Chromatin Remodeler CHD8 Is Required for Activation of Progesterone Receptor-Dependent Enhancers | While the importance of gene enhancers in transcriptional regulation is well established, the mechanisms and the protein factors that determine enhancers activity have only recently begun to be unravelled. Recent studies have shown that progesterone receptor (PR) binds regions that display typical features of gene enhancers. Here, we show by ChIP-seq experiments that the chromatin remodeler CHD8 mostly binds promoters under proliferation conditions. However, upon progestin stimulation, CHD8 re-localizes to PR enhancers also enriched in p300 and H3K4me1. Consistently, CHD8 depletion severely impairs progestin-dependent gene regulation. CHD8 binding is PR-dependent but independent of the pioneering factor FOXA1. The SWI/SNF chromatin-remodelling complex is required for PR-dependent gene activation. Interestingly, we show that CHD8 interacts with the SWI/SNF complex and that depletion of BRG1 and BRM, the ATPases of SWI/SNF complex, impairs CHD8 recruitment. We also show that CHD8 is not required for H3K27 acetylation, but contributes to increase accessibility of the enhancer to DNaseI. Furthermore, CHD8 was required for RNAPII recruiting to the enhancers and for transcription of enhancer-derived RNAs (eRNAs). Taken together our data demonstrate that CHD8 is involved in late stages of PR enhancers activation.
| A lot of research has been devoted during the last decades to understand the mechanisms that control gene promoters activity, however, much less is known about enhancers. Only recently, the use of genome-wide chromatin immunoprecipitation techniques has revealed the existence of more than 400,000 enhancers in the human genome. We are starting to understand the importance of these regulatory elements and how they are activated or repressed. In this work we discover that the chromatin remodeler CHD8 is recruited to Progesteron Receptor-dependent enhancers upon hormone treatment. CHD8 is required for late steps in the activation of these enhancers, including transcription of the enhancers and synthesis of eRNA (long noncoding RNAs derived form the enhancers).
| During the last decade it has become clear that regulation of gene transcription is accompanied by extensive changes in the chromatin organization of promoters [1]. More recently, efforts have concentrated on elucidating the chromatin dynamics of distal regulatory regions and particularly enhancers [2,3,4]. Enhancers were originally defined as regulatory sequences that can activate gene expression independently of their proximity or orientation with respect to their target genes [5]. Even though histone modifications signatures (e.g., high levels of histone H3 lysine 4 monomethyl (H3K4me1) and H3K27 acetyl (H3K27ac) modifications) have provided insight for the discovery of enhancer-like regions, the mechanisms and the factors that control enhancers activation are not yet well known.
Chromatin changes are normally performed by two types of enzymes: enzymes that chemically modify histones or ATP-dependent chromatin remodelers. ATP-dependent remodelling is performed by enzymes of the SNF2 family that use the energy of ATP hydrolysis to destabilize the interaction between DNA and histones [6,7]. In humans there are 26 ATPases of this family with specific roles in gene transcription and in other aspects of DNA metabolism. One of these ATPases is CHD8 which, in addition to the ATPase domain, contains two chromodomains in the amino terminus of the protein, and two BRK domains in the carboxy terminus [8]. CHD8 is able to remodel nucleosomes in vitro in an ATP-dependent reaction [9]; however, its in vivo functions are unclear. Inactivation of CHD8 by homologous recombination in mice provokes a strong growth retardation from embryonic day 5.5 and developmental arrest accompanied by massive apoptosis [10]. Ishihara et al. reported that CHD8 interacts with CTCF and plays a role in insulation activity [11]. It has also been reported that CHD8 represses beta-catenin target genes, and suppresses p53-dependent activation and apoptosis, by promoting histone H1 recruitment [9,12,13,14]. A role in repression of MLH1 gene, associated to MAFG, has been also shown [15]. In contrast to this repressive role, we have shown that CHD8 is required for E2F-dependent activation of G1/S specific promoters [16,17]. Additionally, it has been reported that CHD8 is required for estrogen-dependent induction of Cyclin E2 gene [18], and for recruitment of androgen receptor (AR) and activation of the TMPRSS2 gene [19]. These two studies indicate that CHD8 is also involved in steroid hormone-dependent transcriptional regulation although the mechanisms of this regulation are unknown.
In this work, we have investigated the role of CHD8 in progesterone-dependent transcriptional regulation. Progesterone controls transcription through a complex mechanism [20]. On the one hand, progesterone-bound progesterone receptor (PR) is able to bind specific DNA sequences in chromatin, called PRE, and to recruit histone modification enzymes and ATP-dependent chromatin remodelers, such as NURF and BAF complexes (a member of the SWI/SNF complex family) [21,22]. On the other hand, a small fraction of PR is attached to the cytoplasmic side of the cell membrane and, in the presence of hormone, interacts with tyrosine kinases provoking the activation of various kinase cascades, including ERK1/ERK2 [23,24]. Activated ERK1/ERK2 phosphorylates PR and the kinase MSK1, forming a ternary complex that binds to chromatin. Recent genome-wide studies demonstrated that most PR binding sites (PRbs) are distal regulatory regions that map at introns and intergenic regions and that display typical histone modifications of enhancer regions [25,26,27]. Here, we demonstrate that CHD8 is required for the activation of PR enhancers. In proliferating T47D breast carcinoma cells, CHD8 is mostly associated with promoters. However, upon progesterone treatment, CHD8 was quickly recruited to a subset of transcriptionally competent PR enhancers. In agreement with these data, depletion of CHD8 strongly impaired progesterone-regulated gene expression. CHD8 recruitment to the enhancers was dependent of PR but independent of the pioneering factor FOXA1. Interestingly, depletion of the SWI/SNF complex ATPases, BRG1 and BRM, impaired CHD8 recruitment. Furthermore, we observed that CHD8 interacts with the SWI/SNF complex. CHD8 was not required for H3K27 acetylation, but depletion of CHD8 impaired hormone-dependent RNAPII recruitment at enhancers and synthesis of enhancer RNAs (eRNAs), suggesting that CHD8 is required for enhancer transcription.
To identify the genome-wide distribution of CHD8 in proliferating human breast cancer cells T47D-MTVL [28], we performed chromatin immunoprecipitation of CHD8 followed by deep sequencing (ChIP-seq). We found 12655, 4900 and 2500 peaks of CHD8 by using three different confidence threshold values, respectively (ChIPseeqer threshold level 10–10, 10–15 and 10–20). In all cases, false discovery rates (FDR) were lower than 0.03 (see Materials and Methods). At the lowest threshold about 48% of the peaks (6257) were located within promoters, with a strong enrichment around TSS (Fig 1A and 1B) and about 50% of the peaks were distributed between introns (2187) and intergenic regions (3748). Interestingly, if confidence threshold for peaks identification is increased, the percentage of peaks associated with promoters raise to about 78% (Fig 1A). This is due to differences in CHD8 signal intensity. In fact, CHD8 signals at TSSs are higher than at intergenic or intronic regions (Fig 1C and 1D). Therefore, when only highly significant peaks are analyzed most of them map at promoters. Consistently, a strong overlap between CHD8 and RNA polymerase II (RNAPII) or H3K4me3 ChIP-seq signals was observed (Fig 1C and 1E). CHD8 occupancy was confirmed by ChIP-qPCR in three selected target promoters (CCND1, HDS11B2 and CCNE2) using a different anti-CHD8 antibody (S1A Fig). Furthermore, as control, we also verified that knockdown of CHD8 decreased ChIP-qPCR signal (S1B Fig). All these experiments validated our ChIP-seq results. The rest of the analysis was performed using the 12655 CHD8 sites identified at the lowest, but still very significant, threshold. We have previously reported that CHD8 binds 1965 promoters in a ChIP-on-chip analysis of proliferating cervical carcinoma C33 cells [17]. Approximately 60% of the promoters identified as CHD8 targets by ChIP-on-chip (1175) were also identified by ChIP-seq, despite the different cell lines used. CHD8 bound genes were enriched in Gene Ontology categories related to macromolecular biosynthetic processes (transcription, mRNA processing) and cell cycle (S2A Fig). Consistently with our previous data from C33 cells [17] CHD8 target promoters were strongly enriched in E2F (p-value = 3.6 × 10–135), ELK-1 (p-value = 2.2 × 10–122), AP-2 (p-value = 4.7 × 10–100), and SP1 (p-value = 4.2 × 10–80) transcription factors binding sites (S2B Fig). Taken together these results indicate that CHD8 is mostly bound to promoters in proliferating T47D-MTVL cells.
CHD8 has been suggested to be a nuclear receptor co-activator [18,19]; however, very little is known about this function of the protein. To gain insight into its role in hormone dependent transcriptional regulation we have analyzed by ChIP-seq the distribution of CHD8 in progesterone-treated T47D-MTVL cells. For that, cells were subjected to 48 h of serum deprivation and then stimulated during 5 or 45 minutes with the synthetic progestin R5020 (10 nM) or the vehicle (ethanol) as control. A very small number of peaks were found in vehicle treated cells, suggesting that serum deprivation strongly decreases the association of CHD8 to the chromatin. These data extend our previous observation about absence of CHD8 in four G1/S transition genes in quiescent cells [17]. However, 1132 and 4532 progestin-induced CHD8 peaks were found at 5 and 45 minutes, respectively, suggesting that CHD8 is quickly recruited to the chromatin after progestin treatment. Most of the sites found after 5 minutes (73%) were also identified after 45 minutes (Fig 2A). Only 18% (832) of the hormone-specific peaks were also found under proliferating conditions (Fig 2A). A large majority of the hormone specific sites were found at intronic and intergenic regions (Fig 2B) and low enrichment was found at TSS (S3 Fig). A representative example of the ChIP-seq data close to four well-known progesterone-dependent genes (HSD11B2, FKBP5, NFE2L3 and IL6ST) is shown in Fig 2C.
De novo analysis of sequence motifs in progestin-dependent CHD8 bound regions identified a highly significant enrichment (p-value = 4.7 × 10–75) for the sequence CTGTNC, which is very similar to the consensus sequence of the progesterone receptor binding site TGTYCY [25] (Fig 2D). PR binds 24436 sites in T47D-MTVL cells 60 min after progestin treatment [25]. Interestingly, 83.2% of the CHD8 progestin-dependent peaks (3770) co-localized with PR peaks, indicating that CHD8 is recruited to a subset of PRbs upon hormone induction (Fig 2E). Thus, CHD8 was significantly enriched around PRbs (Fig 2F). Ballaré et al. have recently reported that PRbs present high nucleosome occupancy and that functional PR sites, involved in transcription control, display a high nucleosome-remodelling index (NRI) [25]. NRI is the ratio between the nucleosome occupancy before and after hormone administration. Strikingly, we found strong CHD8 occupancy at PRbs with high NRI (top 10% higher NRI) and weak CHD8 enrichment at PRbs with low NRI (top 10% lower NRI), suggesting that CHD8 is associated with functional PRbs, where a strong nucleosome remodelling is occurring (Fig 2G and 2H). Similarly to CHD8 binding sites, PRbs are mostly found in introns and intergenic regions [25]. Distal regulatory regions and enhancers are enriched in the histone acetyltransferase p300 [29]. In the presence of R5020, CHD8 signal was strongly enriched around p300 binding sites (Fig 2I), indicating that CHD8 binding sites display enhancer characteristics. Another typical enhancer feature is the presence of monomethylated histone H3 lysine 4 (H3K4me1) [30]. CHD8 was moderately enriched in all regions with high H3K4me1 (S4 Fig). Most interestingly, we observed strong enrichment of CHD8 around H3K4me1 containing PRbs (Fig 2J). Taken together all these data indicate that CHD8 binds functional PR enhancers in a hormone-dependent manner.
It has been reported that CHD8 interacts and cooperates with the insulator factor CTCF (CCCTC-binding factor) [11]. Therefore, we have studied the co-localization of both factors under normal growth conditions or after progesterone stimulation. In proliferating T47D-MTVL cells about 16.5% (p-value = 6.0 x 10–130, hypergeometric distribution) of the CHD8 containing regions were also enriched in CTCF (S5A Fig). However, this was only 4.4% of the CTCF sites. CTCF mostly binds to intergenic or intronic regions [31]. Consistently, most of the CTCF-CHD8 co-occupied sites (66%) were also at intergenic and intronic regions (S5B Fig). Upon progesterone stimulation only about 5.4% of the CHD8 sites and 0.6% of the CTCF sites are co-occupied by both factors (S5C Fig), suggesting that CHD8 and CTCF do not cooperate for progesterone-mediated regulation.
To correlate CHD8 binding sites with CHD8-regulated gene expression we performed a transcriptomic analysis of T47D-MTVL cells transfected with a control siRNA or a siRNA specifically targeting CHD8 and stimulated during 6 h with progestin or vehicle (Fig 3A and 3B). We found 1170 genes differentially expressed (FDR< 0.01 and lineal change > 1.5 fold) after progestin treatment of control cells, of which 793 were up-regulated and 377 down-regulated. About 52.5% of these genes (614) were misregulated in CHD8-depleted cells with respect to control cells. Interestingly, CHD8-dependent genes presented lower induction of up-regulated genes and lower repression of down-regulated genes, indicating that CHD8 is required for progesterone-dependent regulation of a subset of genes (Fig 3C). Consistently, around 42% of the CHD8-dependent genes (257) were found to be close to CHD8 genomic locations (P = 9.96 x 10–103) (Fig 3D).
Next, we verified by RT-qPCR that CHD8 was required for normal progestin-dependent induction of several well known progesterone dependent genes that contain close hormone-dependent CHD8 binding sites, such as HSD11B2, DUSP1, FKBP5, NFE2L3 and IL6ST genes. Thus, depletion of CHD8 severely impaired accumulation of mRNA, both 45 min and 6 h after progestin treatment (Fig 3E). As a control, we confirmed that a different siRNA that targets CHD8 had similar effects on expression of progesterone dependent genes (S6 Fig). T47D-MTVL cells contain a single copy of the MMTV-Luc transgene integrated in their genome [28]. CHD8 was also strongly enriched in a progestin-dependent manner at the MMTV-Luc transgene promoter (Fig 3F). Furthermore, depletion of CHD8 severely impaired induction of MMTV-Luc (Fig 3E and S6 Fig). In summary, these data demonstrate that CHD8 is necessary for progestin-dependent regulation, at least in a subset of target genes.
Next we selected four CHD8 peaks close to HSD11B2, FKBP5, NFE2L3 and IL6ST genes that display high ChIP-seq enrichment for CHD8 and PR upon progestin stimulation (Fig 4A). These regions were also enriched for the typical enhancer factor p300 (Fig 4A). Therefore, we called these regions HSD11B2e, FKBP5e, NFE2L3e and IL6STe. All enhancers were located upstream of the corresponding genes (Fig 2C). Then, we decided to investigate how PR affects CHD8 recruitment to these enhancers and vice versa. First we analyzed PR and CHD8 recruitment to CHD8 progestin-dependent peaks in T47D-MTVL cells or in T47D-YV, a PR-negative clonal derivative cell line of T47D [32] (Fig 4B). We verified by western blotting that levels of CHD8 were identical in T47D-MTVL and in T47D-YV (Fig 4B). ChIP experiments performed 45 minutes after progestin stimulation confirmed that high levels of PR are recruited to all analyzed regions in T47D cells but, as expected, not in T47D-YV cells (Fig 4C). Interestingly, CHD8 was also strongly recruited to all analyzed regions in T47D-MTVL cells but not in T47D-YV cells, suggesting that PR is required for hormone-dependent CHD8 recruitment to PRbs (Fig 4C). A significant recruitment of CHD8 to the IL6ST regulatory region was observed in T47D-YV cells suggesting that CHD8 is recruited to this region, at least in part, independently of PR. Residual binding of progestin to other nuclear receptor might be responsible of this effect.
Next, we determined whether CHD8 is involved in PR recruitment to PRbs in T47D-MTVL cells. Fig 4D shows that CHD8 depletion (siCHD8) did not affect progestin-dependent PR recruitment to the four analyzed regulatory regions. As a control, we verified that silencing of CHD8 strongly decreased its association with chromatin, validating the ChIP signals. Taken together, these data indicate that PR is required for CHD8 recruitment to progesterone-dependent CHD8 binding sites, but PR does not require CHD8 for binding.
FOXA1 is a fork-head family transcription factor able to directly bind to DNA in the surface of a nucleosome [33]. Because of this pioneering ability, FOXA1 is able to facilitate estrogen receptor binding to the chromatin of hormone-dependent enhancers [34,35]. Since PR is able to bind directly to nucleosomes [28], it is unclear whether FOXA1 also cooperates with PR. ChIP-seq data of FOXA1 distribution in unstimulated T47D cells are available from ENCODE. Using these data we have observed that 52% (2358) of the CHD8 binding sites were also enriched for FOXA1 (Fig 5A). Furthermore, 54% (2022) of the CHD8-PR co-occupied regions were also occupied by FOXA1. Given this very significant enrichment (P = 1.17 x 10–173) we decided to study the role of FOXA1 in CHD8 and PR recruiting. First, we investigated using ChIP whether FOXA1 binds to the analyzed regulatory regions of HSD11B2, FKBP5, NFE2L3 and IL6ST genes. As shown in Fig 5B significant levels of FOXA1 were found at all studied regions. FOXA1 occupancy after progestin treatment at HSD11B2e, NFE2L3e and IL6STe enhancers was similar in T47D-MTVL and T47D-YV cells, suggesting that PR is not involved in FOXA1 recruitment at these regions (Fig 5B). However, a possible role of ligand-bound PR in FOXA1 recruitment was observed at the FKBP5e enhancer (Fig 5B).
Next, we determined the effect of FOXA1 silencing (Fig 5C) on the hormone-dependent recruitment of CHD8 and PR in T47D-MTVL cells. Surprisingly, depletion of FOXA1 significantly increased the level of occupancy of both PR and CHD8 at the four regulatory regions analysed (Fig 5D and 5E), suggesting that, at least in these progestin-dependent regulatory regions, FOXA1 does not help PR and subsequent CHD8 recruitment. Moreover, our data suggest that FOXA1 might compete with PR for binding to PR-enhancers.
PBAF and BAF are closely related chromatin remodelling complexes of the SWI/SNF family, which share multiple protein subunits. The BAF complex is required for progesterone-dependent gene activation [21,22,25]. Furthermore, BAF250 and BAF57, two of the subunits of the complex, interact with PR in a hormone-dependent manner [22]. Then, we decided to investigate whether CHD8 interacts with the SWI/SNF complexes. For that, we performed immunoprecipitation using anti-CHD8 antibodies from extracts of T47D-MTVL cells. CHD8 co-precipitated with the core subunits INI1/hSNF5/BAF47, BAF170 and BAF155 and with the ATPase BRG1 in T47D-MTVL cells (Fig 6A). All these subunits form part of both SWI/SNF complexes: BAF and PBAF. To identify the type of SWI/SNF complex interacting with CHD8, we also investigated the presence of BAF180, a PBAF-specific subunit, and BAF250, a BAF-specific subunit. As shown in Fig 6A both subunits co-precipitated with CHD8 indicating that CHD8 can interact with both complexes. The CHD8-SWI/SNF interaction was found both, in the presence and in the absence of hormone (S7A Fig). Next, we studied whether the SWI/SNF complex is required for CHD8 recruitment to PRbs. First, we verified by ChIP that BAF155, one of the core subunits of the complex, is recruited to the four analyzed CHD8-bound enhancers and to the MMTV promoter (Fig 6B). Then, we demonstrated that knockdown of BRG1 and BRM (S7B and S7C Fig) significantly impaired recruitment of CHD8 to the four analyzed PR enhancers and to the MMTV promoter (Fig 6C). Taken together, these data suggest that CHD8 interacts with the SWI/SNF complex and that this interaction contributes to recruit or to stabilize CHD8 at PRbs.
Presence of H3K27Ac distinguishes active enhancer states from those poised for activation [36,37,38]. H3K27 is acetylated by p300 [39] and we have shown that CHD8 is enriched around p300 binding sites (Fig 2I). To further characterize the role of CHD8 in enhancer activation we investigated whether CHD8 affects the level of H3K27Ac at the four selected CHD8 binding sites. At the FKBP5e and IL6STe regions H3K27 acetylation was enhanced by R5020 (Fig 7A). However, significant levels of H3K27Ac were already observed under un-stimulated conditions at the HSD11B2e and NFE2L3e enhancers, which were not further stimulated by progestin (Fig 7A). Interestingly, CHD8 depletion did not affect the level of H3K27Ac at any of the analyzed regions, suggesting that CHD8 is not involved in p300 recruiting or H3K27 acetylation.
Next we investigated whether CHD8 is required to open the chromatin of its target enhancers. For that, we performed quantitative DNase I sensitivity assays in HSD11B2e, FKBP5e, NFE2L3e and IL6STe enhancer regions in the presence of hormone or vehicle, as described in [40]. Interestingly, CHD8 knockdown decreased hormone-dependent accessibility to the enzyme of all the analyzed regions suggesting that CHD8 is involved in the hormone-dependent chromatin remodelling of these regions (Fig 7B).
Several studies have shown that many active enhancers present significant occupancy of RNAPII and regulated production of bidirectional RNA, called eRNAs [41,42,43,44,45]. We have investigated whether RNAPII is recruited to FKBP5e, NFE2L3e and IL6STe enhancer regions and the order of recruitment with respect to PR and CHD8. For that, binding of PR, CHD8 and RNAPII at 0, 2, 5, 15 and 30 min after progestin stimulation was determined by ChIP-qPCR (Fig 8A). As previously reported for the MMTV promoter [21], PR was found at the enhancers as early as 2 min after hormone addition. CHD8 was absent at 2 min but was found at the 5 min time point, remaining in the enhancers during the 15 and 30 min time points. Finally, RNAPII occupancy increased between 15 and 30 min at the FKPB5e and only at 30 min at NFE2L3e and IL6STe enhancers (Fig 8A). These data indicate that RNAPII is recruited to the enhancers after PR and CHD8 and suggest that RNAPII recruitment is a late event during the process of enhancer activation. Then, we investigated the role of CHD8 in RNAPII recruitment. Interestingly, hormone-stimulated RNAPII occupancy was strongly impaired in CHD8-depleted cells (Fig 8B).
Next, eRNA synthesis at FKBP5e, NFE2L3e and IL6STe enhancers was evaluated by RT-qPCR both 45 min and 6 h upon R5020 addition (see Materials and Methods). Progestin increased between 2 and 7 fold production of eRNAs already at 45 min and expression was maintained after 6 h (Fig 8C). As a control, we verified that no qPCR signal was observed in the absence of reverse transcriptase. Consistently with the effect of CHD8 depletion in RNAPII occupancy, silencing of CHD8 significantly impaired eRNA synthesis from the three analyzed regions (Fig 8C). These results indicate that CHD8 is required for RNAPII recruitment and enhancer transcription, at least from a subset of PR enhancers.
Steroid hormone transcriptional regulation requires binding of nuclear receptors to thousand of binding sites that mostly map in intergenic and intronic regions and that show features of transcription enhancers. The recent discovery of enhancer-associated transcripts opens the door to investigate how enhancer transcription is controlled. We, and others, have previously found that the chromatin remodeler CHD8 is found at promoters [9,13,14,16,17,19,46]. The data presented in this manuscript demonstrate that shortly after progestin stimulation of quiescent cells, CHD8 binds progesterone enhancers. Consistently, depletion of CHD8 impairs progestin-dependent transcriptional response. CHD8 recruiting requires PR, but not the pioneering factor FOXA1. We also show that CHD8 interacts with the SWI/SNF complex and that SWI/SNF is important for normal CHD8 recruitment. Furthermore, we demonstrate that CHD8 is not required for acetylation of H3K27 in enhancers, but it is necessary for DNase I accessibility, for normal recruitment of RNAPII and for the synthesis of progestin-dependent eRNAs, suggesting that CHD8 plays a role in late phases of progesterone enhancers activation.
We have recently reported a promoter ChIP-on-chip analysis demonstrating that, in cervix carcinoma C33 cells, CHD8 binds around 2000 active promoters also enriched in H3K4me2 and H3K4me3, and that it is required for expression of E2F-dependent G1/S transition genes [17]. In agreement with these results, our genome-wide ChIP-seq analysis reveals that most of CHD8 is bound to promoters also enriched in RNAPII and H3K4me3 in proliferating T47D-MTVL cells. While this paper was in revision another group has also shown that CHD8 binds thousand of active promoters in an iPSC-derived neuronal progenitor cell line [47]. All these data from different cell lines support that CHD8 is a promoter-associated factor. However, upon progestin stimulation CHD8 was rarely found at promoters; instead it was recruited to a number of PRbs mostly located in intergenic and intronic regions enriched in H3K4me1 and p300, indicating that these sites are progesterone-dependent enhancers. Interestingly, under normal proliferation conditions, CHD8 was found at the promoters of genes close to these enhancers (p < 0.001). While this fact might suggest that CHD8 dynamically move from enhancers to close promoters, further experiments are required to demonstrate this association. Our data demonstrate that CHD8 binds both, promoters and enhancers upon specific stimulation such as progesterone. CHD8 has also been involved in AR-dependent [19] and ER-dependent [18] regulation. It is, therefore, possible that CHD8 also binds AR or ER enhancers in the presence of the appropriated hormonal stimulus.
CHD7 is a paralogue of CHD8 and both proteins has been shown to interact [48]. ChIP-seq analysis of mice ES cells has demonstrated that CHD7 is mostly associated with a subset of active enhancers and promoters [49,50]. Comparison of CHD8 genomic distribution with that of CHD7 available from ENCODE showed that only 7.1% and 5.6% of the CHD8 peaks overlap with CHD7 containing regions in K562 or H1 stem cells, respectively. Most enhancers are tissue- and cell line-specific and therefore it is difficult to compare genomic positions between different cell lines. Nevertheless, these data suggest that although CHD7 and CHD8 may cooperate in some genomic locations, they also have specific independent roles.
It has been reported that CTCF interacts with the carboxy-terminus of CHD8 and that CHD8 co-occupy several CTCF binding sites [11]. Our genome-wide study demonstrates that, in the absence of progesterone, 16.5% of the CHD8 binding sites are enriched for CTCF and 4.4% of CTCF sites are occupied by CHD8. While this percentage is largely higher than expected by chance, it is clear that CHD8 might only be involved in CTCF function in a small percentage of insulators. Since CTCF sites are often located at enhancer regions [51] and enhancer looping activity has been related to CTCF and cohesins [52], it is possible that the CHD8-CTCF co-occupancy is more related to the role of CHD8 in enhancers than to specific functions in insulation. In the presence of progestin the number of sites co-occupied by CTCF and CHD8 is even lower, suggesting that their association is not related to progesterone dependent transcription.
Here we show that CHD8 is not recruited to four selected progesterone enhancers in the absence of PR, indicating that PR is essential for the recruiting. This is consistent with our time-course experiment where PR was found at the enhancers earlier than CHD8 upon hormone addition. However, we have been unable to detect direct PR-CHD8 interaction suggesting that CHD8 may be recruited through interaction with other co-regulators of PR enhancers. Despite the very significant overlapping between the pioneering factor FOXA1 and CHD8 binding sites, we show that FOXA1 silencing even increased CHD8 occupancy at the four analyzed enhancers, indicating that FOXA1 is not required for CHD8 recruitment. The chromatin remodelling BAF complex is recruited to PRbs [25] and is required for progesterone-dependent remodelling of the MMTV promoter [21,22]. Interestingly, we have observed that CHD8 interacts with both SWI/SNF complexes: BAF and PBAF. In addition, two high throughput proteomic analysis have identified co-immunoprecipitation of CHD8 and SWI/SNF subunits [53,54]. Since PR directly interacts with BAF complex it is possible that CHD8 is recruited to PRbs through interaction with the BAF complex. Consistently with this hypothesis, knocking down of BRG1 and BRM reduced CHD8 recruiting. Interestingly, it has been reported that the CHD8 paralogue, CHD7, interacts with PBAF, the other human SWI/SNF complex [55]. Whether other CHD8 paralogues, CHD6 and CHD9, also interact with SWI/SNF complexes is unknown. It is also unclear the extent to which CHD8 and SWI/SNF complexes cooperate and whether they can have common and independent targets. It is worth noting that both, SWI/SNF complexes and CHD8 are required for activation of E2F-dependent genes at the G1/S transition [17] [56].
On the other hand, we have reported that the chromodomains of CHD8 binds dimethylated and trimethylated H3K4 peptides [16]. H3K4me2 modification is typically associated with enhancers [57,58,59]. In addition, Vicent et al. reported that H3K4 methylation by the ASCOM (ASC-2 [activating signal cointegrator-2] complex) is required for the progesterone-dependent induction of the MMTV promoter [21]. Therefore, it is possible that CHD8 chromodomains interaction with methylated H3K4 also contributes to the recruitment or the stabilization of CHD8 at PR enhancers.
We have shown that CHD8 depletion does not impair PR binding. Since PR directly interacts with BAF complex it is unlikely that CHD8 is required for BAF recruitment or activity. In fact, we show that BRG1 and BRM are required for normal CHD8 occupancy of PR enhancers suggesting that CHD8 acts downstream of the BAF complex. Another common step of enhancer activation is acetylation of H3K27 by the histone acetyltransferase p300 [29,36,37,38]. Ballare et al., found that progestin-dependent PR biding sites were enriched in regions that contained p300 under un-stimulated conditions and that, in general, the level of p300 increased upon hormone treatment. We have observed that progestin stimulated H3K27 acetylation at the FKBP5e and IL6STe enhancers but not at the HSD11B2e and NFE2L3e enhancers. Depletion of CHD8 did not affect levels of H3K27Ac at any of the regions, suggesting that CHD8 was not involved in this step of enhancer activation.
Numerous reports have evidenced that enhancer activation involves RNAPII recruiting and synthesis of eRNA, monodirectional or bidirectional transcripts of 0.5 to 5 kb. Recent results demonstrate that eRNAs are involved in transcriptional activation of neighbouring genes, in enhancer-promoter looping and in directing chromatin-remodelling events at specific promoters [60,61,62,63,64]. The time-course experiment shown in Fig 8A indicates that RNAPII is recruited between 15 and 30 minutes after hormone stimulation, while PR and CHD8 reach the enhancers around 2 and 5 min after stimulation, respectively. We show that depletion of CHD8 strongly impairs recruiting of RNAPII and synthesis of eRNAs, demonstrating that CHD8 is required for these late events of progesterone enhancers activation. In Drosophila, the CHD8 orthologous Kismet is found at transcriptionally active genes in polytene chromosomes [65]. Kismet mutations reduce the level of phosphorylated elongating RNAPII but not the level of initiating RNAPII, suggesting that Kismet is involved in transcription elongation. Human CHD8 interacts with elongating RNAPII [16]. Furthermore, CHD8-depleted cells are hypersensitive to drugs that inhibit phosphorylation of serine 2 of the carboxy-terminal domain (CTD) of POLR2A, the largest subunit of RNAPII (DRB and flavopiridol), an early step of the transcription cycle [16]. These data suggest that, as Kismet, CHD8 may also be involved in elongation. Kaikkonen et al., have recently reported that eRNA synthesis is sensitive to flavopiridol [58], suggesting that eRNA synthesis also requires CTD serine 2 phosphorylation. Therefore, it is possible that CHD8 is required for transcription elongation of eRNAs at PR enhancers.
We also show that CHD8 is required to increase DNase I sensitivity at enhancers upon hormone treatment, suggesting that CHD8 may be involved in the hormone-dependent nucleosomal remodeling that occurs at a subset of PRbs [25]. It is well known that transcription increases DNase I sensitivity of gene bodies [66,67]. Therefore, it is also possible that CHD8 effect on DNase I sensitivity is caused by its role in enhancer transcription.
Vicent et al. demonstrated that BAF and NURF chromatin remodelling machines are required for PR-dependent activation of the MMTV promoter and other PR enhancers [22,25]. Now we add a third actor, CHD8, to the list of remodelers required for progesterone-dependent activation. However, CHD8 seems to be recruited only to at subset of PRbs, including the MMTV promoter. Future experiments will be required to find what determines CHD8 recruitment to some PRbs and not to others. It is worth noting that it has been reported that CHD9 interacts in vitro with nuclear receptors such as PPARA (PPARα), NR1I3 (CAR), NR3C1, ESR1 (ERα) and RXRA [68,69]. So, it is tempting to speculate that other CHD8 paralogues may also regulate, redundantly or non-redundantly with CHD8, hormone-dependent enhancers activation.
T47D-MTVL human breast cancer cells carrying one stably integrated copy of the luciferase reporter gene under the control of the MMTV promoter [28] and T47D-YV cells (PR-negative clonal derivative cell line of T47D [32,70]) were routinely grown in RPMI 1640 medium supplemented with 10% FBS, 2 mM L-glutamine, 100 U/ml penicillin and 100 mg/ml streptomycin. Cells were grown exponentially (in 10% FBS) or subjected to serum-free conditions in RPMI medium without phenol red during 48 h. After serum starvation, cells were incubated with 10 nM R5020 or vehicle (ethanol; EtOH) for the indicated times.
ChIP assays were performed as described [71] using anti-CHD8 (A301-224A) from Bethyl Laboratories or home-made rabbit anti-CHD8 [16], anti-RNAPII (N-20) (sc-899), anti-PR (H-190) (sc-7208), anti-BAF155 (R-18) (sc-9746) and anti-FOXA1 (H-120) (sc-22841) from Santa Cruz Biotechnology, and anti-H3K27Ac (ab4729) from Abcam. Chromatin was sonicated to an average fragment size of 400 to 500 bp using the Diagenode Bioruptor. Rabbit IgG (Sigma) was used as a control for non-specific interactions. Input was prepared with 10% of the chromatin material used for immunoprecipitation. Input material was diluted 1:10 before PCR amplification. Quantification of immunoprecipitated DNA was performed by real-time PCR (qPCR) with the Applied Biosystems 7500 FAST real-time PCR system, using Applied Biosystems Power SYBR green master mix. Sample quantifications by qPCR were performed in triplicate. Sequences of all oligonucleotides are available upon request. Data are the average of at least three independent experiments.
ChIP was performed as described above using anti-CHD8 (A301-224A, Bethyl Laboratories). ChIP-DNA was purified and subjected to deep sequencing using the Solexa Genome Analyzer (Illumina). The sequence reads were aligned to the human genome reference (assembly hg19). ChIP-seq peak calling, genomic annotation of peaks and comparison between ChIP-seq and ENCODE datasets were performed using ChIPseeqer (v.2.1) [72]. An empirical approach was followed to estimate the FDR, which involves using control data set as ChIP-seq data and the ChIP-seq data as the pseudo-control data and running peak detection. The FDR is defined as the ratio of the number of peaks detected in this pseudo-control analysis, to the number of peaks detected in the real ChIP-seq experiment. Motif analysis was performed using FIRE algorithm [73], included in the ChIPseeqer framework, MEME suite [74], Weeder PScan [75] and TRANSFAC database [76]. CHD8 ChIP-seq data are available from the GEO database (accession number GSE49134). RNAPII, H3K4me3 and PR ChIP-seq data in T47D were previously reported [25]. GEO accession number for FOXA1 and CTCF ChIP-seq data are GMS803409, GSM803348.
All siRNAs were transfected using Oligofectamine (Invitrogen) according to the manufacturer’s instructions, with following siRNA sequences: for siCHD8, 5′-GAGCAAGCUCAACACCAUC-3′; siFOXA1, 5′-GAGAGAAAAAAUCAACAGC-3′; siCt, 5′-CGUACGCGGAAUACUUCGA-3′; siBRG1, 5′-GCGACUCACUGACGGAGAA-3′; and siBRM, 5′-GAAAGGAGGUGCUAAGACA-3′. After transfection, medium was replaced by serum-free fresh medium without phenol red. After 48 h in serum-free conditions, cells were treated with 10 nM R5020 or vehicle (EtOH) for the indicated times. The down-regulation of CHD8, FOXA1, BRG1 and BRM was confirmed by RT-PCR and Western blotting, respectively.
T47D-MTVL cells were grown in serum-free conditions, as explained previously, during 48 h and treated with 10 nM R5020 or vehicle (EtOH) during 6 h. Total RNA was isolated in triplicate from cells using RNeasy Mini Kit (Qiagen). Purity and quality of isolated RNA were assessed by RNA 6000 Nano assay on a 2100 Bioanalyzer (Agilent Technologies, Santa 6 Clara, CA). RNA (100 ng) was used for production of end-labelled biotinylated ssDNA. Labelled ssDNA was hybridized to the GeneChip human Gene 1.0 ST array oligonucleotide microarray (Affymetrix, Santa Clara, CA) according to manufacturer’s recommendations. The arrays were scanned using the GeneChip Scanner 3000 7G (Affymetrix), and raw data were extracted from the scanned images and analyzed with the Affymetrix GeneChip Command Console Software (Affymetrix). The raw array data were pre-processed and normalized using the Robust Multichip Average (RMA) method [77]. Data were further processed using oneChannelGUI [78]. The log2 intensities for each probe were used for further analysis. Genes where considered as hormone induced when change of gene expression was >1.5 (linear fold change) and p-value < 0.01. Gene expression was considered as CHD8-dependent when [hormone-dependent change in siCHD8]/[hormone-dependent change in siCt] was higher than 1.20 or lower than 0.8. Microarray data are available from the GEO database (accession number GSE62257).
DNase I assay was performed as previously described [40]. Briefly, 2 μg of crosslinked chromatin were treated with 0.5, 1, and 2 units of DNase I (Roche) for 3 min at 37°C. Control samples were incubated in the absence of DNaseI. Reactions were stopped by adding EDTA and the crosslinking was reversed by incubating the samples at 65°C. DNA was then isolated, quantified and used as template for qPCR reactions using specific primers.
Total RNA was prepared by using the RNeasy Kit (Qiagen), as described in the manufacturer’s instructions; note that the step of DNase I digestion was included to avoid potential DNA contamination. cDNA was generated from 800 ng of total RNA (for progesterone-dependent mRNA induction analysis) using Superscript First Strand Synthesis System (Invitrogen). For progesterone-dependent eRNA induction analysis 2 μg of total RNA was used. In the case of FKBP5e and HSD11B2e eRNA determination, strand-specific oligonucleotides were used for RT, in order to avoid expression form close promoters. cDNA (2 μl) was used as a template for qPCR. Gene products were quantified by real-time PCR with the Applied Biosystems 7500 FAST real-time PCR system, using Applied Biosystems Power SYBR green master mix. Sequences of all oligonucleotides are available upon request. Values were normalized to the expression of the 28S housekeeping gene. Each experiment was performed at least in duplicate, and qPCR quantifications were performed in triplicate.
Co-immunoprecipitations were performed as described in [17] using the anti-CHD8 antibody (A301-224A) from Bethyl Laboratories. Rabbit or mouse purified IgG (Sigma-Aldrich) were used as a control. 3% Input and precipitated proteins were separated by SDS/PAGE, and visualized by Western blotting with the indicated antibodies using ECL Plus (GE Healthcare), according to the manufacturer’s instructions. Antibodies used for western blotting were: anti-BRG1 (H88, sc-10768), anti-BAF155 (R-18, sc-9746), anti-BAF170 (E-6, sc-17838), anti-PR (H-190) (sc-7208) and anti-FOXA1 (H-120) (sc-22841), and anti-hSNF5 (C20, sc-16189) from Santa Cruz Biotechnology; anti-CHD8 (A301-224A) and anti-BAF180 (A301-590A) from Bethyl; anti-BAF250 (04–080) from Millipore; α-tubulin antibody (DM1A, T9026) from Sigma Aldrich and anti-BRM (ab15597) from Abcam.
|
10.1371/journal.pgen.1006820 | Identification of a Sjögren's syndrome susceptibility locus at OAS1 that influences isoform switching, protein expression, and responsiveness to type I interferons | Sjögren’s syndrome (SS) is a common, autoimmune exocrinopathy distinguished by keratoconjunctivitis sicca and xerostomia. Patients frequently develop serious complications including lymphoma, pulmonary dysfunction, neuropathy, vasculitis, and debilitating fatigue. Dysregulation of type I interferon (IFN) pathway is a prominent feature of SS and is correlated with increased autoantibody titers and disease severity. To identify genetic determinants of IFN pathway dysregulation in SS, we performed cis-expression quantitative trait locus (eQTL) analyses focusing on differentially expressed type I IFN-inducible transcripts identified through a transcriptome profiling study. Multiple cis-eQTLs were associated with transcript levels of 2'-5'-oligoadenylate synthetase 1 (OAS1) peaking at rs10774671 (PeQTL = 6.05 × 10−14). Association of rs10774671 with SS susceptibility was identified and confirmed through meta-analysis of two independent cohorts (Pmeta = 2.59 × 10−9; odds ratio = 0.75; 95% confidence interval = 0.66–0.86). The risk allele of rs10774671 shifts splicing of OAS1 from production of the p46 isoform to multiple alternative transcripts, including p42, p48, and p44. We found that the isoforms were differentially expressed within each genotype in controls and patients with and without autoantibodies. Furthermore, our results showed that the three alternatively spliced isoforms lacked translational response to type I IFN stimulation. The p48 and p44 isoforms also had impaired protein expression governed by the 3' end of the transcripts. The SS risk allele of rs10774671 has been shown by others to be associated with reduced OAS1 enzymatic activity and ability to clear viral infections, as well as reduced responsiveness to IFN treatment. Our results establish OAS1 as a risk locus for SS and support a potential role for defective viral clearance due to altered IFN response as a genetic pathophysiological basis of this complex autoimmune disease.
| Sjögren’s syndrome (SS) is a common autoimmune condition where immune cells infiltrate moisture-producing glands, leading to dryness typically in the eyes and mouth. SS patients also manifest debilitating fatigue as well as other diseases in liver, lung, kidney, and skin. The cause of this complex disease is still not fully understood; however, an environmental trigger, such as viral infections, in individuals with genetic risk factor(s) is thought to contribute to the development of SS. Type 1 interferons (IFNs) are one of the first defenders after viral infection and induce the expression of various virus-responding genes. Perpetual elevation of type 1 IFN signaling has been observed in SS patients. Here, we first replicated previously identified RNA transcripts contributing to the abnormal type 1 IFN signaling in SS patients. We then identified a disease-associated genetic variant in an IFN-inducible gene, OAS1. This variant governs splicing of OAS1, altering the transcript into multiple isoforms that lack protein expression and responsiveness to IFNs. The results of this study may provide insight into the genetic basis of SS, as well as other autoimmune disease with similar dysregulation in the type 1 IFN system.
| Sjögren’s syndrome (SS) is a common systemic autoimmune disease with a prevalence rate (~0.7% of European Americans) second only to rheumatoid arthritis (RA) [1]. SS is distinguished by immune cell infiltration, functional destruction, and irreversible dysfunction of exocrine glands, most notably salivary and lacrimal glands [2]. Secondary manifestations of exocrine gland dysfunction may include severe dental decay and corneal scarring. Approximately one-third of patients experience extra-glandular manifestations of disease, such as debilitating fatigue, a 16-fold increased risk of developing lymphoma, neuropathies, Raynaud’s phenomenon, arthralgia, and dermatologic symptoms [2–6].
Both glandular dysfunction and extra-glandular manifestations are associated with autoantibodies, a hallmark of autoimmunity [7–9]. Approximately 70% and 40% of SS patients exhibit autoantibodies targeting ribonucleoproteins, Ro/SSA (Ro52 and Ro60) and La/SSB, respectively [10]. These autoantibodies have the capacity to bind necrotic and apoptotic material, thus creating RNA-immune complexes that can activate cells of the immune system and aggravate autoinflammation [11]. Such RNA-containing immune complexes are taken up by the Fc gamma receptor IIa on plasmacytoid dendritic cells (pDCs) [12], which activates intracellular Toll-like receptors 7 and 9 and stimulates type I interferon (IFN) responsive loci [13].
The etiology of SS is still largely unknown, though it involves a complex interplay between both genetic and environmental factors [14–16]. Viral infections, such as Epstein-Barr virus (EBV) and cytomegalovirus [17–19], may initiate prolonged inflammation in glandular lesions and formation of germinal center-like structures commonly linked to autoantibody production in SS [14, 19]. Autoantibodies can be detected up to 18–20 years prior to diagnosis in 81% of SS patients [20]. Indeed, cross-reactivity between antibodies against EBV and the Ro60 antigen has previously been reported [21], and possible subclinical reactivation of the virus has been associated with active joint involvement in SS [22]. Recently, the virus-like genomic repeat element L1 was identified as an endogenous trigger of the IFN pathway, and its expression correlates with type I IFN expression and L1 DNA demethylation [23].
Type I IFNs are key antiviral immune mediators of innate immune responses in infected cells, while at the same time enhancing antigen presentation and inducing production of pro-inflammatory cytokines and chemokines [24], thus initiating adaptive immunity [25]. Overexpression of type I IFN-inducible genes, known as “the IFN signature”, is a common feature of many autoimmune diseases [26, 27], including RA patients with poor clinical outcome [28–30] and systemic lupus erythematosus [31, 32], where the predominant IFN producing cells, pDCs, are reduced in number in the blood but are abundant in skin and lymph nodes [33]. In SS, the IFN signature is observed in both peripheral blood and salivary glands [12, 34–37], and associates with systemic manifestations, greater disease severity, and autoantibody titers [8, 9]. It has been proposed that viral infections contribute to perpetual activation of type I IFN signaling and the resulting dysregulation of innate immunity, ultimately resulting in activation of the adaptive immune response and autoantibody production in SS and other autoimmune diseases [34, 38].
Genome-wide association [GWA] studies in autoimmune diseases have identified multiple genetic risk variants involved in type I IFN signaling pathways [39–41], including associations of IRF5 and STAT4 with SS susceptibility [15, 42, 43]. Suggestive associations of FCGR2A, PRDM1 (PR domain containing 1, regulated by IRF5) [44], and IRF8 with SS have also been reported [15], along with two genes within the NF-κB pathway (TNIP1 and TNFAIP3), which regulates early phase type I IFN production during viral infection [45]. Despite the evidence indicating an important role of the type I IFN pathway in SS, no direct functional mechanisms for SS-associated variants contributing to the substantial upregulation of IFN signature transcripts have been described.
The vast majority of disease associated single-nucleotide polymorphisms (SNPs) identified in GWA studies are non-coding [46] and are not likely to impact protein function directly, thus requiring a combination of genetic studies and gene expression analyses to point towards mechanisms that link genetics with functional effects [47]. Specifically, the extensive linkage disequilibrium (LD) observed between associated polymorphisms renders it hard to identify causal variant(s) of disease. Systemic evaluation of genome-wide functional elements by the Encyclopedia of DNA Elements (ENCODE) project reveals that 80% of the human genome has at least one biochemical function, and many of the genetic variants are within cis- or trans- regulatory sites that impact gene expression [48]. Furthermore, genome-wide cis-expression quantitative trait locus (eQTL) mapping studies in different tissues have identified more than 3,000 genes associated with nearby genetic variants [49, 50].
Through combining GWA and gene expression data from SS patients, we sought to identify and characterize SS-associated variants that influence the expression of genes within the IFN signature by utilizing a genomic convergence approach (Fig 1). Through cis-eQTL analyses we identified an association of a SNP rs10774671, located within the OAS1 gene locus, with SS. Functional studies were performed to assess biological consequence of the OAS1 variants.
To select candidate genes in the IFN signature, we first evaluated dysregulated transcripts in SS through a microarray-based gene expression profiling study. Whole blood transcriptome profiles from 115 anti-Ro/SSA positive SS cases and 56 healthy controls of European ancestry were compared, as the IFN signature is enriched in SS patients seropositive for anti-Ro/SSA [34]. After quality control (QC) and normalization, 13,893 probes (in 10,966 genes) remained, for which Welch’s t-tests, false discovery rate (FDR)-adjusted P values (q values), and fold changes (FC; the difference of the mean between log2-transformed values from cases and controls) were calculated (see Methods for details). Differentially expressed genes were selected by q < 0.05 and FC > 2 or < -2. We found 73 differentially expressed genes in our dataset, among which 57 genes are regulated by type I IFNs (S1 Table). The majority of dysregulated genes (66 out of 73) were overexpressed in SS patients and formed the IFN signature in cases after unsupervised hierarchical clustering (Fig 2A). Of note, the IFN signature was observed in most but not all anti-Ro/SSA positive SS cases, in accordance with our previous work [34], and the intensity of this feature was heterogeneous among patients (Fig 2B). These results suggest that the expression of IFN signature genes might be influenced by genetic variants, which could be identified through cis-eQTL analysis.
We hypothesized that variants near the differentially expressed IFN signature genes may potentially influence disease susceptibility through cis-regulatory mechanisms. We would, however, expect to identify many cis-eQTLs in these regions regardless of whether they associate with disease susceptibility or not. Therefore, instead of performing cis-eQTL analyses for all of the 73 dysregulated IFN signature genes, we first sought to identify variants that showed a disease association of Passoc<0.05 for subsequent evaluation of their role in altering gene expression. Genetic associations with SS susceptibility for 2,163 SNPs in regions of the 73 differentially expressed genes were tested using a combined dataset (Dataset 1; Table 1) from genome-wide genotyping arrays consisting of 765 SS cases and 3,825 population controls of European ancestry. We identified suggestive associations (Passoc<1×10−4; this threshold was determined by Bonferroni correction for independent variants with r2<0.2) of genetic variants within the OAS1 region (top association at rs10774671, Passoc = 8.47×10−5), which is regulated by type I IFNs (S1 Table). Furthermore, we identified suggestive associations in ARGN (S1 Table) and observed nominal associations (1×10−4≤Passoc<0.05) with SS susceptibility in 42 additional regions (S1 Table).
To determine whether these disease-associated genetic variants (Passoc<0.05) were related to the altered expression levels of their nearby differentially expressed IFN signature genes, we performed cis- and trans-eQTL analyses for all SNPs with Passoc<0.05 (173 SNPs in 44 regions; S1 Table) using a linear model by integrating the transcriptome and genotype datasets in 178 European individuals (108 anti-Ro/SSA positive SS cases, 55 anti-Ro/SSA negative SS cases, and 15 healthy controls). Variants within and near OAS1 showed significant association with OAS1 transcript expression (Fig 3A; S1 Table). In particular, three microarray probes targeting OAS1 passed QC and were evaluated for cis-eQTLs (Fig 3B). The OAS1 transcript levels measured by all of the three probes were found to be associated with nearby genetic variants (Fig 3C). The most significant cis-eQTL for all the three probes targeting OAS1 was rs10774671 (PeQTL-Probe1 = 5.14×10−4, PeQTL-Probe2 = 2.86×10−6, and PeQTL-Probe3 = 6.05×10−14; Fig 3C). No eQTL was detected in any other differentially expressed genes (S1 Table). We also determined that none of these eQTL variants were associated with the two nearby genes, OAS2 and OAS3. Additionally, no significant trans-eQTL was detected for OAS1. Therefore, we identified a variant associated with both SS susceptibility and gene expression in the IFN signature gene OAS1.
To fine map this disease-associated region, imputation was then performed for the SS-associated OAS1 region to increase the informativeness of the genetic association and eQTL analyses results. After imputation, the most significant association with SS in the OAS1 region was at rs4767023 (Passoc = 3.82×10−5; r2 = 0.98 with the most significant genotyped SNP rs10774671; Fig 4A), whereas the top eQTL remains at rs10774671 (Fig 3A). All the variants with Passoc<1×10−4 in the OAS1 region were strongly correlated to each other (r2>0.9; Fig 4B) and could explain the association of the whole region through conditional analyses (Fig 4C and 4D). The top SS-associated variants and cis-eQTLs in the OAS1 region, including rs10774671, were in strong LD (r2>0.9; Fig 4B), thus challenging the selection of potentially functional variant(s) based on results from the association analyses. However, the top eQTL variant, rs10774671, is an A/G substitution within the consensus sequence of a splice acceptor site at the junction of the 5th intron and the 6th exon of OAS1 (Fig 3B), and is known to alter normal splicing and induce isoform switching of OAS1 [51]. In addition, all other SS-associated variants (Passoc<1×10−4) in the OAS1 locus were either intronic or outside of coding regions, lacking functional genomic elements mapped to the SNP as determined by the ENCODE project [48, 52]. Also, we performed a co-localization analysis using eCAVIAR [53] to identify the potential causal variant in the OAS1 region. We estimated colocalization posterior probability (CLPP) scores for all the tested 453 variants, and rs10774671 has the highest CLPP score among all the variants (S2 Table). Therefore, we prioritized this SS-associated cis-eQTL variant, rs10774671, for further replication and functional studies.
We replicated the genetic association of rs10774671 with SS susceptibility in an independent dataset (Dataset 2; Table 1) consisting of 514 European SS cases and 3,466 European population controls (genotyped using TaqMan assays, Prep = 5.16×10−6; odds ratio = 0.71; 95% confidence interval = 0.63–0.83). Meta-analysis was performed to combine the results between the initial genetic association study (Dataset 1) and the replication cohorts (Dataset 2) and established the association of rs10774671 with SS risk (Pmeta = 2.59×10−9; odds ratio = 0.75; 95% confidence interval = 0.66–0.86; risk allele [the A allele] frequency: case = 0.70, control = 0.64; with no heterogeneity between the two datasets as determined by I2 = 0). We also performed a stratified analysis and a permutation analysis using merged samples from Dataset 1 and Dataset 2 to determine whether the observed genetic association was restricted to anti-Ro/SSA positive or negative patients. We did not find any evidence to support the genetic effect to be specific to any sub-group of the patients (S1 Fig). In summary, we identified a potential causal variant, rs10774671, that was associated with SS susceptilibty, likely through its impact on the expression of a key IFN signature gene, OAS1.
Following establishment of the association between rs10774671 and SS susceptibility, we further determined the influence of different genotypes on the alternative splicing of OAS1. Four isoforms of OAS1 are annotated in the NCBI Reference Sequence (RefSeq; http://www.ncbi.nlm.nih.gov/refseq) database, of which we analyzed p46, p42, and p48, and p44, an un-annotated isoform previously reported in RNA-sequencing (RNA-seq) studies [54–56] (Fig 3B). The difference between these isoforms is confined to their 3' end where rs10774671 influences alternative splicing, yielding amino acid sequences of different lengths and composition. In the microarray experiments, one probe targeting OAS1 specifically recognized the 3' end of the p42 isoform (Fig 3B). The risk allele A of rs10774671 was correlated with higher expression levels of p42 (Fig 3C, right panel). However, we were not able to determine the influence of rs10774671 on the expression of other isoforms due to lack of isoform-specific probes on the microarray.
In order to determine the influence of rs10774671 on the expression of each alternatively spliced isoform of OAS1 and compare OAS1 isoform composition, we performed RNA-seq on whole blood from 57 SS cases and 27 healthy controls. After QC, the reads were aligned to the human genome using TopHat [57] without gene annotation to facilitate the detection of potentially novel isoforms of OAS1. The transcript level of each isoform was compared across samples with different genotypes of rs10774671 based on the measurement of fragments per kilobase of transcript per million mapped reads (FPKM) using Cufflinks [58]. Consistent with our microarray results, the SS risk allele A of rs10774671 was correlated with higher expression levels of p42 (P = 1.30×10−7; Fig 5A). Increased production of other alternatively spliced isoforms of OAS1, including p48 and p44, was also observed in subjects with the SS risk genotypes (GA and AA) of rs10774671 (Fig 5B and 5C). In contrast, transcript levels of the p46 isoform, was decreased in samples with the A allele (P = 3.48×10−10; Fig 5D), consistent with previous reports that interruption of the splicing consensus sequence inhibit formation of the p46 isoform [56]. These results were further confirmed by quantitative real-time PCR using primer sets targeting the specific OAS1 isoforms (S2 Fig; S3 Table). Therefore, we found that the SS-associated variant rs10774671 is a functional variant that influences alternative splicing of OAS1.
Since OAS1 is part of the IFN signature and its expression levels are correlated with the autoantibody status, we also performed a stratified eQTL analysis to investigate whether the eQTL effects are specific to any sub-group of the SS patients based on their anti-Ro/SSA positivity. We stratified the SS case samples into anti-Ro/SSA positive patients (n = 27) and anti-Ro/SSA negative patients (n = 30), and performed eQTL analyses on each of the OAS1 isoforms using linear regression while adjusting for sex. Despite reduced statistical power, we identified significant eQTL results for the p46, p42, and p48 isoforms in both subsets of samples. By using the Z-test as described in S1 Fig, we did not find any significant difference of the eQTL effects between the two sub-groups (S3 Fig).
Comparing the total OAS1 transcript level from the microarray study within each genotype revealed significantly higher gene expression in SS patients as compared to control in the GA group (Fig 6A). There is a trend towards higher total OAS1 transcripts in the AA and GG groups of SS patients as well, though it is not statistically significant. Interestingly, the highest transcript levels are seen in the anti-Ro/SSA positive cases (Fig 6B), significantly higher than both anti-Ro/SSA negative cases and healthy controls. The same results were observed in the RNA-seq data (S4 Fig), indicating that the total OAS1 transcript levels regardless of isoform are also influenced by disease status or the presence of autoantibodies.
To further dissect the functional mechanism of rs10774671 in predisposing disease risk, we utilized Western blots to evaluate the difference in protein levels of the normally spliced isoform p46 (formed by the protective allele G of rs10774671) and the alternatively spliced isoforms of OAS1 in EBV-immortalized B cells from SS patients. Consistent with the RNA-seq results, the protein expression of p46 was substantially lower in subjects carrying the A allele of rs10774671, whereas p42 was the dominant isoform in the GA and AA subjects without stimulation (Fig 7A). Interestingly, both protein and mRNA levels of the p46 isoform were upregulated after stimulation by type I IFN in the GG and GA subjects (P = 3×10−4 and P = 2×10−3, respectively; Fig 7A and 7B). However, protein expression of the p42 isoform remained unchanged upon IFN stimulation (Fig 7A), even though its transcript level significantly increased after stimulation (Fig 7B). The protein expressions of p48 and p44 were low in all of the samples, and not responsive to type I IFN stimulation (Fig 7A). We then cloned and transfected each isoform into human embryonic kidney (HEK) cell line 293T cells and observed similar protein expression results as in EBV cells: lower protein levels for p48 and p44 compared to p46, even though their transcript levels were equivalent (Fig 7C). These results suggest that the alternative isoforms of OAS1 that are associated with the disease risk variant of rs10774671 fail to generate proteins after transcription.
The catalytic OAS1 domain is located at the N terminus, though the isoforms differ in their C terminus. It has been suggested that these differences affect affinity of OAS1 protein for different viruses [59]. However, our data suggested that the alternatively spliced 3'-terminus influenced the lack of post-transcriptional expression of the p48 and p44 isoforms. To test this hypothesis, we generated several truncated forms of p48 and p44 at the 3'-terminus and transfected them into HEK 293T cells. The truncation of both p48 and p44 transcripts at the 3'-end resulted in restoration of protein expression (Fig 7D). Our results demonstrated that the alternatively spliced 3'-end between 1,047 and 1,155 bp of the p48 isoform and the 3'-end between 1,083 and 1,137 bp of the p44 isoform were responsible for the impaired protein expression (Fig 7D). In addition, we recombined the green fluorescent protein (GFP) transcript with the 3'-end from different OAS1 isoforms and expressed them in HEK 293T cells. The alternatively spliced 3'-terminus of p48 and p44 resulted in reduced expression of GFP when linked to the 3' end of GFP transcript (Fig 7E; S5 Fig). These results further confirmed the impact of the alternatively spliced 3'-end of OAS1 on protein expression. Determining mechanisms for how the 3'-terminus from the alternatively spliced OAS1 isoforms influences protein expression and type I IFN responsiveness needs further study.
Overexpression of genes in the IFN pathway is a distinctive feature of multiple autoimmune diseases, though no evident mechanism has thus far been revealed. We identified and established rs10774671 as a risk locus for SS. The A allele of rs10774671 is correlated with reduced OAS1 enzymatic activity in human peripheral blood mononuclear cells [51], and is associated with increased susceptibility to West Nile virus [60] and chronic hepatitis C virus infections [61]. OAS1 is a member of the 2'-5'-oligoadenylate synthetase family, which is upregulated by type I IFNs during innate immune responses to viral infection and activates latent RNase L, leading to viral RNA degradation and clearance [62, 63]. The SS risk allele A of rs10774671 causes alternative transcript splicing and consequently less functional isoforms are activated by type I IFNs.
Failure to clear virus might lead to subclinical, chronic infection that drives the sustained overexpression of IFN, but viral proteins may also indirectly cause IFN production through adaptive immune responses. For example, antibodies generated towards EBV nuclear antigen-1 cross-react with Ro/SSA [21], and anti-Ro/SSA antibodies may in turn form immune complexes that stimulate type I IFNs [64]. As viruses evidently play a role in SS pathophysiology [18], genetic variants affecting the antiviral properties of OAS1 might be a contributing factor. A recent study showed that while the presence of antibodies to hepatitis D virus was equal in SS patients and otherwise healthy controls, the virus itself was present in significantly more patients [65], indeed suggesting that viral clearance is restrained. Epitope spreading [66, 67], antibody cross-reaction [21], or molecular mimicry [68, 69] are likely consequences of subclinical, chronic or recurrent infection.
The basal activity of OAS1, which varies greatly among individuals, is thought to be under strong genetic control [51]. The enzyme activity in the GG genotype with predominantly the p46 isoform is higher than GA (intermediate) and AA (low) [51]. OAS1 isoforms p42 and p46 have been detected at the protein level in human cells, whereas the p44a/p44b, p48, and p52 isoforms have been detected at mRNA levels [54, 70–72]. In addition to the RNase L activation properties, the tetramer forming p48 isozyme also exhibits proapoptotic activity [73], a property partly accredited to IFN-γ [74], and is shown to interact with Bcl-2 [75]. Bcl-2 is an anti-apoptotic protein negatively regulated by Ro52 [76], and in salivary gland epithelial cells Bcl-2 is essential in regulation of IFN-γ induced apoptosis [77]. It has been postulated that p46 is a more efficient synthetase than p48, explaining the increased basal activity of p46 [51]. Although there is no evidence showing that the differences in the C-terminus alter protein function [78], our truncation experiments indicated that the alternatively spliced C-terminus governs the post-transcriptional protein expression.
Interestingly, we found lower protein expressions of p44 and p48 after type I IFN stimulation despite equivalent transcript levels compared to p46. This indicates that p44 and p48 production, which is governed by the A allele, is less responsive to IFN stimulation as compared to p46. Lack of response to IFNs has also been seen in multiple sclerosis (MS), in which patients carrying the homozygous rs10774671 GG genotype, a protective genotype in MS associated with less active disease, were more responsive to IFN-β treatment than AA and AG patients, as measured by time to first relapse [79].
We searched the Genotype-Tissue Expression (GTEx) database and confirmed the association between rs10774671 and OAS1 expression in whole blood [80]. We also searched eQTL for all our top variants in the 43 SS-associated genes (besides OAS1) as well as any variants in LD with those top variants (r2>0.8). Out of the 614 variants we checked, two variants were also eQTLs for their corresponding genes: the top SS-associated variant in ANKRD22 (rs1147601, Passoc = 2.38×10−3, PeQTL-GTEx = 6.4×10−6) and the top SS-associated variant in EPSTI1 (rs7323736, Passoc = 1.79×10−2, PeQTL-GTEx = 2.6×10−6). However, both of these variants were only nominally associated with SS susceptibility and did not pass our suggestive significance threshold for disease association (Passoc<1×10−4). Nevertheless, these variants and genes could be plausible targets for future replication studies to assess their disease associations.
The rs10774671 A/G variant is a common splice site variation, and there is a skewed distribution of genotypes in autoimmune diseases like type I diabetes (T1D) [81] and MS [79] despite ambiguous genetic association with disease: the alternative allele A renders risk to SS and MS, whereas the reference allele G increases susceptibility to T1D. We hypothesized that these opposite risk effects may be due to different functional isoform usages in different disease-relevant tissues. Through searching the GTEx database for rs10774671 eQTLs, we found rs10774671 is a significant eQTL of OAS1 in 5 tissues (S6A Fig). Interestingly, the eQTL effect in the Esophagus—Mucosa tissue is in the opposite direction compared to other tissues. In whole blood, the p46 isoform is predominant, thus the A allele caused reduced expression of OAS1 as a whole (S6B Fig); however, in Esophagus Mucosa, the isoform is p42 (S6C Fig) and results in an opposite effect of rs10774671 on the total OAS1 expression. We propose that the ambiguous genetic effects of rs10774671 on different diseases might be due to different functional isoforms in disease-relevant tissues (not necessarily Esophagus—Mucosa). While reduced expression of the functional isoform p46 in whole blood increases risk of SS and MS, it protects individuals from T1D. The downstream differences of various isoforms in protein levels, isoform expression, responsiveness to IFN, and basal activity between genotypes flag OAS1 as a highly relevant protein in autoimmune diseases, despite no direct effect on IFN expression.
OAS1 is one of several genes relevant in overall IFN response found to be disease associated in SS. Others include IL-12A [15], which can induce both type I and type II IFNs [82]; STAT4 [15], which, although not explicitly overexpressed in the IFN signature, plays an important role in the cross-talk between type I and type II IFNs [83–85]; and IRF5 [15], a transcription factor in the IFN pathway [86]. The rs10774671 is a known cis-eQTL and splicing QTL, observed in whole blood [87] as in our study, in lymphoblastoid cells [88], and in monocytes, both naïve CD14 and in cells stimulated with LPS and IFN-γ [89]; but no trans-eQTLs are known. It is possible that other variant(s) in high LD with rs10774671 could contribute additional functional impact(s), such as the rs11352835 in exon 7 seen in MS [59]. Genomic editing approaches that introduce single point mutations or deletions in the OAS1 region will further advance the dissection of the causal SS-associated variant in this haplotype. Animal models that express the risk isoforms due to the risk allele can also be used to observe whether they spontaneously develop SS-like symptoms, and whether the chances for developing such symptoms increase after exposure to viral infections. Our study also highlights the importance of utilizing genomic convergence to identify and prioritize susceptibility genes for human complex disease. The complex mechanisms underlying the IFN signature in SS cannot be explained as a single eQTL driven overexpression. However, we have in this study established OAS1 as a risk locus with functional consequences affecting isoform composition, and that may play a fundamental role in dysregulation of both viral clearance and apoptosis.
All patients in this study fulfilled the 2002 American-European Consensus Group (AECG) criteria for primary SS [7]. Seropositivity of anti-Ro/SSA autoantibodies was determined by the antibody index ≥1 using the Bio-Plex assay (Bio-Rad) following the manufacturer’s protocol. The present study was approved by the Oklahoma Medical Research Foundation Institutional Review Board (IRB#1—Biomedical), operation under Federalwide Assurance (FWA) # 00001389 and IRB # 00000114 under IORG 0000079 approved by the Office for Human Research protection (OHRP), Department of Health and Human Services (DHHS). The OMRF IRB is in compliance with local regulations and the regulations of the United States Food and Drug Administration as described in 21 CFR Parts 50, 56 and 11, the International Conference on Harmonization (ICH) E6, and the United States Department of Health and Human Services at 45 CFR 46. The current study was approved under IRB#07–12 and all patients provided written informed consent.
Quantitative levels of the differentially expressed transcripts from the microarray analysis were used as phenotypic traits in 178 European subjects described above. Variants showing nominal association with SS (Passoc<0.05) in the genetic association analysis were selected to test for cis-eQTLs, defined by variant-transcript pairs within 50kb of the target genes or trans-eQTLs for variants at least 1Mb away. Association of genotype with transcript expression was evaluated using both linear regression (adjusted for gender and disease status) and analysis of variance (ANOVA) in Matrix-eQTL [107]. FDR-adjusted P values were calculated to determine the significance of the eQTL. The results of the cis-eQTL analyses were plotted in Prism 6. We also used a tool, PEER, based on a Bayesian framework to adjust for unknown non-genetic factors in gene expression [108]. We transformed our expression values using all the genes that passed QC by running PEER for 15 factors. We then used the PEER residuals from the 44 SS-associated (Passoc<0.05) IFN signature genes as quantitative traits to determine eQTL while adjusting for other known potentially confounding factors: sex, disease status, anti-Ro/SSA status, and age.
In addition to additive genetic models, we also performed eQTL analyses using linear regression by three other models: recessive (recode genotype from 0,1,2 [where 2 equals to AA] to 0,0,1), dominant (0,1,1), and overdominant (0,1,0). We used the coefficient of determination (R2) to evaluate the goodness-of-fit in each of these models. As shown in S4 Table, both the p42 and p48 isoforms fit the additive model best (highest R2), whereas the recessive model outperformed in the p46 isoform regression (R2rec = 0.57 vs. R2add = 0.54). However, the difference of the R2 between the dominant and additive models in the p46 eQTL analysis is subtle. Also, the outperformance of the recessive model cannot be confirmed by qPCR (where the additive model has the highest R2). Therefore, we only reported the additive results in the main text. However, the alternative genetic models for the eQTL effect observed in different isoforms may reflect distinct disease mechanisms rendered by these isoforms, and thus detailed contribution of different isoforms on disease susceptibility warrant further functional study.
The co-localization analysis between genetic association and cis-eQTL results in the OAS1 region was performed using eCAVIAR [53]. We used the z-scores (calculated by β/standard error) and the LD matrix (calculated using PLINK—r) from both the genetic association and cis-eQTL results as input and assumed one causal variant to obtain colocalization posterior probability (CLPP) scores for all the tested 453 variants in the OAS1 region.
Peripheral blood mRNA transcripts from 27 anti-Ro/SSA positive SS cases, 33 anti-Ro/SSA negative SS cases, and 30 healthy controls were isolated and measured as described above. RNA-seq was performed using the Illumina HiSeq 2000 employing standard procedures. Multiplexing of 6 samples per lane was utilized. Post sequence data were processed with Illumina Pipeline software v.1.7. Quality of raw sequence data was assessed using FASTQC. We assessed the quality of each sample using AQM [92] as described above. A total of 6 samples were removed from analysis due to significantly different expression patterns revealed by PC analysis. Raw FASTQ files were aligned to the human reference genome (hg19) using TopHat [57] that aligns the reads across splicing junctions independent of gene annotations, which benefits de novo detection of alternative splicing events. The total gene transcript level was determined by normalized read counts (raw read counts divided by estimated size factor) in DESeq [109].
To determine alternative splicing events, the reference-independent construction of the transcripts was performed using Cufflinks [58] to identify transcripts >1% of the most abundant isoform in each sample. We only kept the transcripts that were detected in more than 10% of the samples for further analysis. The previously annotated isoforms (p46, p42 and p48) and an un-annotated isoform identified across multiple samples (p44) were used as reference to reconstruct the isoforms of OAS1. The novel identified isoforms of OAS1 were also checked manually in the Integrative Genomics Viewer (IGV) [110] to confirm the transcripts and cross-exon reads. The FPKM values calculated by Cufflinks were used to determine the expression levels of each isoform of OAS1.
Total RNA was extracted using TRIzol reagents (Life Technologies) from EBV-immortalized B cells pre-selected for the presence of target OAS1 isoforms based on the RNA-seq results from whole blood. Following DNase treatment (Life Technologies) and cDNA synthesis (iScript kit from Bio-Rad), full-length and truncated OAS1 transcripts were amplified from cDNA using primer sets specific for the different OAS1 isoforms and truncated forms (S3 Table). Each OAS1 isoform transcript was individually cloned into pcDNA3.1 (Invitrogen) with an Xpress epitope tag at the 5'-terminus to facilitate the detection of transfected protein using Western-blot with anti-Xpress antibody. The plasmid was transfected into the HEK 293T cells using FuGENE transfection reagents (Promega) following manufacturer’s protocols.
The protein expression of OAS1 isoforms was evaluated in EBV-immortalised B cells from SS patients, four independent samples from each genotype group GG, GA and AA, treated or not treated with type I interferon (universal type I IFN, 1500 U/mL, for 24 hours). The cells were lysed in RIPA buffer and cell lysate protein concentration determined using the Qubit Protein Assay kit (Thermo Fisher Scientific). A total of 30 μg protein from each cell extract was separated on a 10% Bis-Tris gel (10% Criterion™ XT Bis-Tris Gel, BioRad, Cat #: 3450112) following the manufacturer’s instructions, the gels cut according to the weight of the OAS1 protein, and simultaneously transferred to a single PVDF membrane, thus ensuring the comparability of Western blot bands from all gels. The OAS1 isoforms were visualized using an anti-OAS1 antibody targeting the shared epitope (Rabbit polyclonal anti-human OAS1, Abcam, Cat #: ab86343) and ECL Prime Western Blotting Detection Reagents (Amersham, Cat #: RPN2232).
|
10.1371/journal.pgen.1002755 | Polymorphisms in the Mitochondrial Ribosome Recycling Factor EF-G2mt/MEF2 Compromise Cell Respiratory Function and Increase Atorvastatin Toxicity | Mitochondrial translation, essential for synthesis of the electron transport chain complexes in the mitochondria, is governed by nuclear encoded genes. Polymorphisms within these genes are increasingly being implicated in disease and may also trigger adverse drug reactions. Statins, a class of HMG-CoA reductase inhibitors used to treat hypercholesterolemia, are among the most widely prescribed drugs in the world. However, a significant proportion of users suffer side effects of varying severity that commonly affect skeletal muscle. The mitochondria are one of the molecular targets of statins, and these drugs have been known to uncover otherwise silent mitochondrial mutations. Based on yeast genetic studies, we identify the mitochondrial translation factor MEF2 as a mediator of atorvastatin toxicity. The human ortholog of MEF2 is the Elongation Factor Gene (EF-G) 2, which has previously been shown to play a specific role in mitochondrial ribosome recycling. Using small interfering RNA (siRNA) silencing of expression in human cell lines, we demonstrate that the EF-G2mt gene is required for cell growth on galactose medium, signifying an essential role for this gene in aerobic respiration. Furthermore, EF-G2mt silenced cell lines have increased susceptibility to cell death in the presence of atorvastatin. Using yeast as a model, conserved amino acid variants, which arise from non-synonymous single nucleotide polymorphisms (SNPs) in the EF-G2mt gene, were generated in the yeast MEF2 gene. Although these mutations do not produce an obvious growth phenotype, three mutations reveal an atorvastatin-sensitive phenotype and further analysis uncovers a decreased respiratory capacity. These findings constitute the first reported phenotype associated with SNPs in the EF-G2mt gene and implicate the human EF-G2mt gene as a pharmacogenetic candidate gene for statin toxicity in humans.
| The mitochondria are responsible for producing the cell's energy. Energy production is the result of carefully orchestrated interactions between proteins encoded by the mitochondrial DNA and by nuclear DNA. Sequence variations in genes encoding these proteins have been shown to cause disease and adverse drug reactions in patients. The cholesterol-lowering drugs statins are one class of drugs that interfere with mitochondrial function. Statins are one of the most prescribed drugs in the western world, but many users suffer side effects, commonly muscle pain. In severe cases this can lead to muscle breakdown and liver failure. In this study, we discover that disruption of a mitochondrial translation gene, EF-G2mt, impedes respiration and increases cell death when exposed to statin. Using the simple unicellular organism yeast as a model, the activity of naturally occurring human EF-G2mt variants is tested. Three of these variants render yeast cells more sensitive to statin. Patients who possess these EF-G2mt variations may be more susceptible to statin side effects. Importantly, the test for statin sensitivity also led to the discovery of mutants that have a reduced energy production capacity. The decreased ability to produce energy is linked to a number of diseases, including myopathies and liver failure.
| The primary function of the mitochondria is the aerobic production of ATP, a process that is reliant on a series of protein complexes that comprise the electron transport chain. Several components of the electron transport chain are encoded in the mitochondrial genome, the translation of which is governed largely by nuclear encoded genes. Increasingly, mutations within these genes are being implicated with respiratory deficiency, an underlying factor in a number of diseases, including myopathies and liver failure [1], [2], [3], [4]. For example, pathogenic mutations in the human mitochondrial elongation factor genes, EF-G1mt and EF-Tu(mt), have been implicated with severe lactic acidosis and encephalopathy [1], [2], [3]. Recently a mutation in a novel gene, believed to be a member of the class of mitochondrial peptide release factors, was identified in patients exhibiting symptoms of Leigh syndrome [4].
In addition to disease, there is also emerging evidence that respiratory deficiencies are responsible for adverse drug reactions. Consequently, treatment with certain drugs have uncovered otherwise silent mitochondrial mutations [5], [6]. The group of cholesterol-lowering drugs, statins, are one example. The primary target of statins is 3-hydroxy-3-methylglutaryl-coenzyme A (HMG-CoA) reductase, the rate limiting enzyme of the sterol synthesis pathway, but increasingly, studies are reporting signs of statin-induced mitochondrial dysfunction [7], [8]. This is believed to be a factor in the myopathic side-effects of statins. Approximately 0.1 to 0.5 percent of statin users experience severe myopathic symptoms (defined as serum creatine kinase levels more than 10 times the upper limit of normal) and many more suffer milder musculoskeletal pain [9], [10]. Frequently such patients present symptoms that are similar to those of patients with mitochondrial myopathies [11]. To date, there have been several case studies reporting the presence of a subclinical MELAS (mitochondrial encephalopathy, lactic acidosis and stroke-like episodes) mutation within the mitochondrial DNA (mtDNA) of patients who have developed severe myopathic symptoms following statin medication [12], [13], [14]. It is expected that existing weakness in mitochondrial function can be exacerbated upon exposure to statin, leading to the uncovering of previously asymptomatic mutations in mitochondrial genes.
The yeast Saccharomyces cerevisiae has been the model of choice for studies of mitochondrial function. In addition to mitochondrial similarities with human cells, the ability of yeast to survive in the absence of mtDNA, the simplicity with which both nuclear and mtDNA can be manipulated and the extensive number of tools and resources available specifically for yeast research has greatly contributed to an understanding of potentially pathogenic mutations [15], [16], [17]. Statins were first isolated as secondary metabolites from fungi, the presumption being that the strong antifungal properties of statins provide an ecological advantage for the producer over other fungi, similar to that of antibiotics. We and others have demonstrated that upon exposure to statin, yeast, as well as having reduced cell viability, also display evidence of mitochondrial dysfunction [18], [19], [20].
In this study, we identify a nuclear gene encoding a mitochondrial translation factor as a modulator of atorvastatin toxicity in yeast (MEF2) and human cell lines (EF-G2mt). The eukaryotic mitochondrial protein synthesis system consists of four phases; initiation, elongation, termination and ribosome recycling, each carefully orchestrated by a series of nuclear encoded proteins [21], [22]. The human EF-G2mt gene, originally named a mitochondrial elongation factor based on sequence homology with bacterial EF-G, has since been shown to function as a ribosome recycling factor [23], [24]. EF-G2mt is believed to interact with the already known ribosome recycling factor (RRF1) to promote dissociation of the ribosomal subunits following termination of translation [23]. In bacteria, the dual role of translocation and ribosome recycling are shared by a single EF-G protein [25]. Eukaryotic cells harbour two EF-G proteins in their mitochondria and it appears that these have distinct functions, the EF-G1mt protein for translocation and the EF-G2mt protein for ribosome recycling [23], [26]. The human EF-G2mt protein is conserved across the majority of eukaryotic species [23]. With its yeast Mef2p counterpart, the human EF-G2mt protein shares greater than 32 percent homology and four of the five protein domains. We use the atorvastatin-sensitive phenotype of the yeast MEF2 gene to uncover naturally occurring human variants of EF-G2mt that have respiratory deficient phenotypes. These findings have ramifications for patient drug response and possibly also for disease.
In light of the emerging evidence that mitochondria are important in dictating statin toxicity, which in turn can reveal underlying respiratory defects that have important health implications, experiments were designed to discover mitochondrial lesions that affect statin sensitivity. Two published fitness profiling experiments in yeast have observed statin sensitivity in hundreds of heterozygous deletion mutants following statin exposure for 20 generations of growth [27], [28]. Using the Gene Ontology term finder available on the Saccharomyces genome database website (www.yeastgenome.org), we discovered approximately 14–17% of genes conferring statin-sensitivity are associated with the mitochondria. One of the most sensitive of these mitochondrial associated genes was the MEF2 gene, encoding a mitochondrial translation factor believed to have a role in ribosome recycling [23], [27].
The growth of deletion mutants in a competition style assay is a very sensitive method of detecting differences in growth rate. However, these assays are prone to a higher incidence of false positive and non-replicable results [29]. To confirm the MEF2 phenotype, a number of yeast deletion mutants that ranked as the most statin sensitive in the fitness profiling experiments were compared [27], [28]. Both heterozygous and haploid mutants were tested and cell viability was assessed after five days exposure to 110 µM atorvastatin. The concentration of 110 µM atorvastatin is approximate to those concentrations used in the original genome-wide fitness profiling screens (62.5 µM and 125 µM atorvastatin) [27]. Furthermore, during a previous investigation of the effects of different atorvastatin concentrations in yeast, we have shown that 110 µM atorvastatin does not inhibit cell growth but is sufficient to cause a significant decrease in intracellular ergosterol (approximately 85%), accompanied by loss of cell viability after prolonged exposure (5 days) [18]. Of the heterozygous mutants tested, only the hmg1Δ/HMG 1 strain was confirmed to be sensitive to atorvastatin (Figure 1A). However, of the haploid mutants tested, three displayed a statin hypersensitive phenotype (Figure 1B). The mef2Δ mutant emerged as the strain which exhibited the greatest loss of cell viability in the presence of atorvastatin, with an almost 20-fold reduction in cell viability compared with the wild-type. The hmg1Δ strain displayed a 5-fold reduction in cell viability and disruption of the HTZ1 gene, encoding a histone protein, resulted in a 2-fold loss of cell viability (Figure 1B).
Yeast mutants defective in mitochondrial translation undergo rapid loss of mtDNA and we have previously shown this to be the case for the mef2Δ mutant [30]. Consequently, mef2Δ is ρ0 (completely devoid of mtDNA). To determine whether atorvastatin-sensitivity is the consequence of abrogation of the MEF2 gene or simply from the absence of mtDNA, mef2Δ statin sensitivity was compared with that of ethidium bromide generated cytoplasmic ρ0 mutants. In order to ensure that there were no secondary site mutations in the mef2Δ deletion mutant that originated from the S. cerevisiae gene deletion collection, new mef2Δ haploid strains were created. Results show that although the ρ0 strains were more sensitive to atorvastatin than the respiratory positive parent, they did not display the same degree of sensitivity as mef2Δ (Table 1). Therefore, mutation of the MEF2 gene is a critical determinant of statin sensitivity through mtDNA dependent and independent functions.
The human EF-G2mt gene, ortholog of the yeast MEF2 gene, encodes a recently characterised mitochondrial ribosome recycling factor [23], but to date, no functional analysis has been performed for this gene. To determine whether the EF-G2mt gene is essential for human cell function and to ascertain whether depletion influences statin toxicity, an siRNA pool comprising of four individual EF-G2mt targeted siRNAs was used to silence EF-G2mt expression in the human rhabdomyosarcoma (RD) cell line. The RD cell line has previously been established as a skeletal muscle model for mitochondrial disorders and has also been used in studies of statin toxicity [31], [32], [33]. At 72 hours post-transfection, greater than 80 percent silencing was consistently achieved and cells remained viable. Cells were then re-transfected at this time point to enable continued depletion of EF-G2mt activity. This strategy had previously uncovered an essential role for cell viability for the first discovered ribosomal recycling factor gene, RRF1 [34]. However, at 72 hours post-re-transfection (six days after the initial transfection), RD cells remained viable even though EF-G2mt mRNA concentration had decreased by 99.9 percent. Additionally, there was no decrease in mtDNA levels upon EF-G2mt silencing as analysed using quantitative PCR (Figure S1) [35].
Although gross inhibition of mitochondrial translation in human cell lines results in loss of cell viability [34], a more subtle mitochondrial phenotype may be masked by the phenomenon of the Crabtree effect whereby many human cell lines, when grown in the presence of glucose, derive their energy almost solely by fermentative means [36]. To circumvent this effect and force cells to rely on mitochondrial respiration as their primary energy source, galactose was used to replace glucose as the carbon source [37]. Silencing of EF-G2mt led to a marked decline in the growth of RD cells at four days post-transfection, which was maintained for a period of seven days (Figure 2). This growth defect on galactose medium signals an impairment in oxidative phosphorylation (OXPHOS).
Based on the findings in yeast, it was predicted that decreased EF-G2mt activity would also enhance the effects of atorvastatin toxicity. EF-G2mt silenced RD cells were subjected to various concentrations of atorvastatin in medium containing either galactose or glucose as the carbon source for a period of 48 hours. In glucose medium, there was no difference in statin sensitivity between the EF-G2mt silenced cells and those transfected with the non-targeting control. However, based on IC50 values (defined here as a 50% loss of viability at 48 hours) in galactose medium, EF-G2mt silenced cells were over 20 percent more sensitive to atorvastatin than cells transfected with the non-targeting siRNA pool (Table 2). These results confirm a role for the human EF-G2mt gene in cell resistance to atorvastatin in a human skeletal muscle cell line. Notably, a similar increase in sensitivity (17%) was observed using the human hepatic HepG2 cell line (a model for statin-induced liver toxicity), although it should be noted that HepG2 cells are approximately 10 times more statin resistant than RD cells and this elevation in atorvastatin sensitivity was not statistically significant in these experiments (Table 2).
A global alignment of the amino acid sequence of the human EF-G2mt protein (Isoform I, AAH15712.1) with the yeast Mef2 protein (CAA59392) reveals 32.1% amino acid sequence identity (Figure 3 and Figure S2). At the commencement of this study, there were nine published non-synonymous Single Nucleotide Polymorphisms (SNPs) in the human EF-G2mt gene, of which five were either conserved or semi-conserved in the yeast MEF2 gene. Three of these variants, EF-G2mtI627T, EF-G2mtE594G and EF-G2mtK334R, are considered rare, with a heterozygosity frequency below one percent. One of the variants, EF-G2mtR744G, has a heterozygosity frequency of three percent and the EF-G2mtR774Q allele has a heterozygosity frequency greater than 20 percent. These five SNPs were selected for functional analysis using the yeast MEF2 gene as a model.
For each of the five selected human EF-G2mt variants, single nucleotide base pair substitutions were constructed directly into the chromosomal copy of the yeast MEF2 gene to replace the codon specific to the wild-type amino acid residue with a codon that corresponds to the amino acid present in the human EF-G2mt protein variants. The Mef2p variants constructed were mef2K769Q, mef2R740G, mef2I616T, mef2D578G and mef2K308R which correspond to EF-G2mtR774Q, EF-G2mtR744G, EF-G2mtI627T, EF-G2mtE594G and EF-G2mtK334R respectively. To assess for respiratory competence, mef2 mutants were grown on medium containing the non-fermentable carbon source glycerol. After 72 hours, all five mutants were proficient in the production of colonies on both glucose and glycerol medium. The number and size of colonies produced by the mutant strains on glycerol medium was equal to that of the wild-type, indicating that all mutants are respiratory competent (Figure 4A). Moreover, based on measurements of cell growth in both glucose and glycerol liquid medium, there were no growth defects exhibited by any of the mef2 mutants (Table S1). Mitochondrial DNA stability was measured periodically for up to 32 generations. The frequency of cells which spontaneously lose mtDNA amongst populations of each mef2 mutant remained equal to that of the wild-type (approximately 2 to 3%), verifying that mtDNA is stable over successive generations.
The five mef2 mutants were then assayed for atorvastatin sensitivity. Following five days of exposure to 110 µM atorvastatin, three of the mef2 mutants exhibited a statin hypersensitive phenotype. Viability of the mef2I616T, mef2D578G and mef2K308R mutants was reduced to 20.3, 22.8 and 24.2 percent respectively (Figure 4B). The two other mef2 mutants, mef2K769Q and mef2R740G, did not exhibit a statin sensitive phenotype. The unmasking of a phenotype for the mef2K308R, mef2D578G and mef2I616T variants by atorvastatin is a strong indicator that these mutations have an effect on Mef2p function. A similar effect conferred by these alleles in the human EF-G2mt protein could have vital consequences for statin users.
Although no obvious growth phenotype was observed, the statin-sensitive phenotype of three of these mutants indicates a subtle defect in mitochondrial function. Staining of mef2 mutant cells with the nucleic acid staining dye 4′,6-diamidino-2-phenylindol (DAPI) confirmed the presence of mtDNA nucleoids in cells of all five mutants and quantification of mtDNA copy number using quantitative PCR (qPCR) [38] showed that mtDNA levels were the same as that of the wild-type (Figure S3). It therefore appears that the EF-G2mt equivalent mutations do not destabilise Mef2p function so as to compromise mtDNA stability.
To investigate the possibility of a respiratory phenotype, oxygen consumption for the three statin-sensitive mef2 mutant cultures was measured (in the absence of atorvastatin) using a non-invasive oxoluminescent device [35]. All three mutants, mef2K308R, mef2D578G and mef2I616T, exhibited a significantly reduced respiration capacity, approximately one third lower than that of the wild-type (Figure 5). The two mef2 mutants that did not display an atorvastatin-sensitive phenotype had respiration rates much closer to that of the wild-type. Together, these results demonstrate a respiratory phenotype conferred by at least three of the EF-G2mt equivalent mef2 variants and this correlates with the inability of these mef2 mutants to tolerate atorvastatin toxicity.
Interestingly, the partially respiratory deficient mef2 mutants exhibit a greater statin sensitivity than the ethidium bromide generated ρ0 strains, which are completely devoid of respiratory function (Table 1). This, in support of the mef2Δ results, indicates the role of a non-respiratory function of MEF2 in the statin response. In addition to a lack of mtDNA, we have previously shown, by staining cells with MitoTracker Red CMXRos, that the mef2Δ strain has a reduced mitochondrial membrane potential (ΔΨ) and that mitochondria appear fewer, with a tendency to aggregate [30]. The MitoTracker Red CMXRos probe enters the mitochondrial matrix dependent on ΔΨ. This same method was used to test ΔΨ and mitochondrial morphology in the mef2 mutants. Cells were visualised using a laser scanning confocal microscope and, in contrast to the mef2Δ strain, all five mef2 mutants displayed mitochondria that stain brightly and are arranged in a tubular network, typical of mitochondria in wild-type yeast. Staining of the three partially respiratory deficient mutants is shown in Figure 6. These results indicate that the mef2 mutations do not disrupt the function of Mef2p in maintaining ΔΨ and so does not explain the enhanced statin sensitivity of these mutants. It is known that statin toxicity can cause loss of ΔΨ [39]. Therefore one possibility is that atorvastatin acts synergistically with a mutated Mef2p to exacerbate loss of ΔΨ and compromise cell viability. However, other mechanisms, such as modulation of the mitochondrial retrograde response, cannot be discounted [40].
The crystal structure of the human EF-G2mt protein has not yet been elucidated. Therefore, to gain insight into the molecular effects of the five chosen amino acid variations, a computational model of the EF-G2mt protein was constructed using the SWISS-MODEL server [41]. The model is based on the experimentally determined crystal structure of the Thermus. thermophilus EF-G protein [42] which shares 39 percent identity with the human EF-G2mt protein and four of the five EF-G2mt protein domains. As the N-terminal and C-terminal regions of the EF-G2mt protein share particularly low homology with the template, the initial 65 and final 10 amino acid residues could not be accurately modelled. For this reason, the location of amino acid variant EF-G2mtR774Q, positioned very close to the C-terminus, was omitted (Figure 7A). Stereochemical quality of the model was assessed by generating a Ramachandran plot using PROCHECK. Eighty seven percent of residues fall within the most favoured regions of the plot, indicating ideal stereochemistry.
Using the in silico model, we can make some hypothetical predictions about the function of the protein variants. The human EF-G2mtK334R variant, which corresponds to the respiratory compromised yeast mef2K308R variant, is located on a coil close to the surface of the GTP-binding domain (domain I) that is highly conserved in EF-G2mt homologs from all major eukaryotic species. The presence of GTP is essential for ribosome dissociation at the termination of translation, and subsequent hydrolysis of GTP is then required in order to release the EF-G2mt protein from the ribosome [23]. The EF-G2mtE594G and EF-G2mtI627T variants, equivalent to the respiratory compromised mef2D578G and mef2I616T mutants respectively, are both located in domain IV and occur close to the EF-G2mt protein surface. The EF-G2mtE594G variant is expected to diminish the largely negative electrostatic surface potential of this domain, thereby interfering in the protein's interaction with both the mitochondrial ribosome and Rrf1 [23], [25]. The EF-G2mtI627T variant is located on an alpha helix whose structural integrity is disturbed by the threonine hydroxyl group. The final two EF-G2mt protein variants, in which the corresponding mef2 mutants did not exhibit a phenotype, are located in domain V, the C-terminal region that accelerates (but is not essential for) the ribosome recycling action of the EF-G2mt protein [23].
In the last year, data from the 1000 Genomes project has expanded the number of known polymorphisms in the human genome and a further 11 non-synonymous SNPs have been discovered within the human EF-G2mt gene (dbSNP, National Centre for Biotechnology Information (NCBI)) [43]. Of these, one is fully conserved in the S. cerevisiae Mef2 protein and another five are semi-conserved (Figure 3 and Figure S2). The location of five of these variants is shown on the EF-G2mt protein model in Figure 7B. Although experimental confirmation is essential, based on the findings above it is hypothesised that the domain IV (EF-G2mtK621N and EF-G2mtF609Y) and domain I (EF-G2mtV165G) variants will affect protein activity. Therefore, these naturally occurring EF-G2mt variants may have respiratory deficient consequences.
By exploiting the genetic tractability of yeast, complemented by siRNA silencing studies in human cell lines, we have identified a mitochondrial translation factor as a mediator of atorvastatin toxicity and also made fundamental discoveries about the function of human variants within the EF-G2mt gene. The human EF-G2mt gene, ortholog of the yeast MEF2 gene, was originally identified as a mitochondrial elongation factor gene. However, a recently published comprehensive analysis of the EF-G2mt protein has shown that it functions as a ribosome recycling factor, interacting with the first discovered ribosome recycling factor protein, Rrf1, to dissociate the ribosomal subunits at the termination of translation [23]. In yeast, deletion of the MEF2 gene results in loss of mtDNA [30], a circumstance which would be lethal in higher eukaryotic cells without the supplementation of uridine and pyruvate [44]. Nevertheless, siRNA silencing experiments show that in contrast to its RRF1 counterpart, knockdown of human EF-G2mt expression does not compromise cell viability in glucose medium. Furthermore, silencing of EF-G2mt expression does not deplete cellular mtDNA content. It is possible that some compensatory mechanism enables ribosome recycling to continue to a sufficient degree to maintain cell viability in fermentative cell lines.
When galactose is used as the carbon source instead of glucose, the ATP produced via glycolysis is insufficient for cell energy requirements; therefore, there is a greater reliance on the production of ATP through the oxidative metabolism of glutamine. This more closely resembles the metabolic activity of cells in a human physiological system [37], [45]. By using galactose to force reliance on the mitochondria for cell energy production, it was shown that reduced EF-G2mt activity does indeed compromise cell proliferation in respiring cells. It is consequently expected that the EF-G2mt gene is essential for cell function in a human system. Furthermore, in support of the notion that abnormalities in mitochondrial function sensitise cells to statin toxicity, the growth defect of EF-G2mt silenced cells in galactose medium was even further exacerbated upon exposure to atorvastatin.
For cells in which mitochondrial function is challenged, exposure to mitochondrial toxicants, such as statins, places additional stress on mitochondrial function and this has the potential to trigger pathogenicity. Indeed we have shown that ρ0 strains, lacking respiratory capacity, are more sensitive to atorvastatin than those with a functioning mitochondrial genome in both yeast and human cells. Studies have shown that statins exert their mitochondrial toxicity effects by inhibiting function of the electron transport chain but there is also evidence of non-respiratory mitochondrial consequences [46], [47], [48]. These include a loss of mitochondrial membrane potential, aberrant mitochondrial morphology and apoptosis [39]. These effects, in combination with the absence of respiratory function may explain the hypersensitivity of ρ0 cells to statin.
The hypersensitivity of respiratory deficient cells to statins may have clinical ramifications for patients that have variations within mitochondrial functioning genes. Statins have been known to aggravate clinically silent disease associated mutations resulting in myopathies. In fact, mutations (which in many cases were asymptomatic) for three common myopathic diseases; carnitine palmitoyltransferase II deficiency, McArdle disease and myoadenylate deaminase deficiency (AMPD deficiency), are thought to be the underlying determinants responsible for statin-induced myopathy in up to 10 percent of patients showing adverse effects [5]. There are also reports of the statin-induced triggering of MELAS syndrome in patients whose MELAS mutations were clinically silent [13], [14]. As MELAS syndrome arises from mutations in mtDNA, it strongly implicates mitochondrial dysfunction in susceptibility to statin toxicity. This is further supported by the identification of two commonly occurring SNPs in the human COQ2 gene that are associated with an increased risk of statin intolerance [49]. The COQ2 gene is required for the synthesis of coenzyme Q10, an essential component of the mitochondrial electron transport chain [49].
Despite the presence of a number of SNPs within the EF-G2mt gene, the function of these variants for cell fitness had never been investigated. In this study, five EF-G2mt equivalent SNPs were selected for functional analysis in the yeast MEF2 gene. Unlike the mef2Δ deletant, all mef2 mutants grew proficiently on both glucose and the non-fermentable carbon source glycerol. Furthermore, the mutations had no effect on mtDNA stability, ΔΨ or mitochondrial morphology. However, exposure of the mutants to toxic concentrations of atorvastatin has uncovered a phenotype for three of the mef2 mutants; mef2K308R, mef2D578G and mef2I616T, equivalent to the EF-G2mtK334R, EF-G2mtE594G and EF-G2mtI627T alleles respectively. In silico protein homology modelling reveals these three mutations are located in either the GTP binding domain (domain I) of the EF-G2mt protein or an external domain (domain IV) necessary for ribosomal interaction.
Based on the observations that statins exacerbate clinically silent disease associated mutations, it was predicted that the three mutations compromise mitochondrial function in a subtle yet potentially significant way. The subsequent observation that oxygen consumption was significantly reduced in the statin-sensitive mef2 mutants confirmed this hypothesis and demonstrates a sub-optimal mitochondrial function for the three EF-G2mt equivalent mef2 mutants. This sub-optimal mitochondrial function is expected to contribute to the atorvastatin-sensitive phenotype of the three mef2 mutants. However, comparison of the statin sensitive phenotype of the mef2K308R, mef2D578G, mef2I616T and mef2Δ mutants with cytoplasmic ρ0 strains indicates that statin sensitivity is not fully explained by the reduced respiratory capacity of these mutants and further studies are required to completely elucidate the precise mechanism.
This study constitutes the first report of a phenotype associated with EF-G2mt, demonstrating an essential role for aerobic respiration in human cell lines and an importance for cell tolerance to atorvastatin. Atorvastatin constituted the focus of this study and is the highest selling and also one of the more potent of the statins [50]. Although the various statins have been shown to differ in their cellular toxicity effects, all of them have been implicated with mitochondrial dysfunction [48], [50], [51], [52], [53]. It would therefore be expected that a mitochondrial mediator of atorvastatin toxicity may also mediate cell response to the other statins. In support of this, preliminary experiments confirm that the atorvastatin sensitive mef2 mutants also exhibit sensitivity to lovastatin.
With an estimated 38 million people around the world undertaking statin treatment [9], the identification of novel biomarkers for statin toxicity has the potential to personalise therapy for millions of individuals. To date, only a handful of genes have been associated with statin toxicity [9], but the finding of the EF-G2mt gene as a potential pharmacogenetic candidate has strengthened the association between existing mitochondrial dysfunction and statin hypersensitivity. Importantly, the discovery of naturally occurring human polymorphisms within the EF-G2mt gene that affect respiratory function indicates that these variants, either alone or in combination with other polymorphisms, have significant pathogenic consequences. This opens avenues for further clinical investigations into a possible association between EF-G2mt variants and disease.
Atorvastatin calcium was purchased from 7 Chemicals (India). Stock solutions were prepared by dissolving atorvastatin in methanol at a concentration of 20 mg/mL and solutions were stored at −20°C. 5-fluoroorotic acid and geneticin (G418) were purchased from Sigma.
The haploid wild-type S. cerevisiae strains used were of the background BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) and BY4742 (MATα his3Δ1 leu2Δ0 lys2Δ0 ura3Δ0). The diploid strain was BY4743 (MATa/α his3Δ1/his3Δ1 leu2Δ0/leu2Δ0 LYS2/lys2Δ0 MET15/met15Δ0 ura3Δ0/ura3Δ0) [54]. The mef2Δ/MEF2, mef2Δ and the cytoplasmic ρ0 strains were previously described [30]. Prior to experiments, yeast strains were cultured in liquid YEPD (1% yeast extract, 2% peptone, 2% dextrose) medium for 24 hours, subcultured and then grown to exponential phase in Synthetic Complete (SC) medium (0.67% Difco yeast nitrogen base without amino acids, 2% dextrose, 0.79 g L−1 amino acid supplement (Sunrise Science Products, Australia)). Cells were incubated at 30°C with shaking.
Triplicate exponential phase cultures grown in SC medium were diluted to an optical density (OD595 nm) of 0.2 and 1.25 mL of this culture was added to 3.75 mL SC medium with the appropriate concentration of atorvastatin. After a maximum of five days growth, cells were diluted and plated onto solid YEPD (YEPD plus 2% agar) medium and viability counts were performed 48 hours later.
HepG2 cells and RD Cells are from the American Type Culture Collection (ATCC). Cells were cultured at 37°C and 5% CO2 in Dulbecco's Modified Eagle's Medium (DMEM) medium (Gibco) containing 4.5 g/L glucose and 10% fetal bovine serum (FBS) (Bovogen Biologicals, Australia). For experiments in which galactose was used as the carbon source, cells were grown in glucose free DMEM with 10% fetal bovine serum and 4.5 g/L galactose. To generate the ρ0 cell lines, used as respiratory deficient controls, HepG2 and RD cells were cultured for eight weeks in DMEM medium containing 100 ng/mL ethidium bromide and supplemented with 10% FBS, 50 µg/mL uridine and 100 µg/mL sodium pyruvate [31], [44]. Following ethidium bromide treatment, ρ0 cells were maintained in medium supplemented with uridine and sodium pyruvate. Depletion of mtDNA was confirmed using qPCR.
The ON-TARGETplus SMARTpool, comprising of four siRNAs targeting the EF-G2MT transcript (NM_170691) was purchased from Dharmacon (cat. # L-017534-01-0005, Dharmacon, Thermo Fisher Scientific, Lafayette, CO.). The corresponding ON-TARGETplus non-targeting SMARTpool (cat. #D-001810-10-05) and DharmaFECT transfection agents were also purchased from Dharmacon. DharmaFECT agent 2 (0.2 µL/100 µL) and DharmaFECT agent 4 (0.4 µL/100 µL) were used to transfect RD cells and HepG2 cells respectively. All siRNAs were used at a final concentration of 25 nM. To transfect cells, equal volumes of the siRNA and transfection agent were mixed, allowed to incubate for 30 minutes, and 100 µL of the solution was added to 400 µL DMEM medium containing 10% FBS. This solution was added to the attached cells which had been washed in PBS. Gene expression silencing was assessed at 72 hours post-transfection by qPCR. For experiments in which cells were re-transfected, the above procedure was performed again at 72 hours subsequent to the initial transfection.
At 24 hours post-transfection, approximately 1×104 cells/well of each culture were seeded into eight wells (one for each day of the eight-day proliferation assay) of a 96-well plate and allowed to attach overnight. Each day, the number of viable cells in one of the eight wells was assessed using the CellTiter-Glo assay (Promega) as described above. Cell medium was changed every two days.
Approximately 5×104 cells/well were seeded into wells of a 96-well plate. For siRNA transfected cells, seeding occurred at 24 hours post-transfection. Following overnight incubation to enable attachment, media was changed to DMEM containing 10% FBS plus the appropriate atorvastatin concentration. Atorvastatin concentrations ranged from 0 to 128 µM for RD cells and 0 to 1024 µM for HepG2 cells. Cell survival after 48 hours in the presence of atorvastatin was estimated using the CellTiter-Glo luminescent cell viability assay (Promega), which measures intracellular ATP concentration. The assay was performed according to the manufacturer's instructions and luminescence was quantified using a Tecan Genios microplate reader. IC50 values were calculated from dose-response curves that were generated using least-squares linear regression.
Codon modifications used to alter yeast Mef2p amino acid residues to those of the corresponding EF-G2mt variants are A2305C, A2219G, T1848C, A1734G and A924G for the yeast mef2K769Q, mef2R740G, mef2I616T, mef2D578G and mef2K308R respectively. The GenBank accession number for the MEF2 sequence used was NC_001142.9. The single base pair substitutions were created in the yeast MEF2 gene according to the double-strand break mediated delitto perfetto method [55]. The pGSKU plasmid containing the CORE-I-SceI cassette was kindly provided by Francesca Storici (Georgia Tech, Atlanta, GA). The CORE-I-SceI cassette was PCR amplified with chimeric primers (Table S2) that contain 50 bp homologous to the site of insertion and 20 bp for amplification of the cassette. Cassette amplification was performed in 50 µL PCR reactions (25 µL 2× Phusion Flash PCR Master mix (Finnzymes), 0.5 µM each primer and approximately 1–10 ng purified pGSKU plasmid) using a PTC-200 thermal cycler (MJ Research) (Initial denaturation, 98°C for 10 seconds; denature, 98°C for 1 second; anneal/extend, 72°C for 75 seconds; repeat denature and anneal/extend for 30 cycles; final extension, 72°C for 1 min). Cells were transformed with 10 µL of the concentrated PCR product as previously described [55]. Transformants were selected by plating onto synthetic medium lacking uracil (0.67% Difco yeast nitrogen base without amino acids, 2% dextrose, 0.79 g L−1 uracil dropout amino acid supplement (Sunrise Science Products, Australia) and 2% agar) and after 24 hours, G418 resistance was checked by replica plating onto solid YEPD containing 200 µg/mL G418 (Sigma). Colony PCR was performed to confirm accurate integration of the cassette. Cells containing the integrated CORE-I-SceI cassette were then transformed with 0.5 nM of each strand of a pair of complementary 80 bp oligonucleotides (Table S2) containing the desired substitution and possessing 40 bp on either side of the oligo which is targeted to the regions adjacent to the integrated marker (GeneWorks, Adelaide). To induce the I-SceI mediated double strand break at the recombination site, prior to transformation cells were cultured in 50 mL of synthetic complete medium containing 2% galactose instead of glucose and incubated at 30°C with shaking for a period of six to eight hours. Following transformation, cells were plated onto solid YEPD medium and after 24 hours, replica plated onto synthetic complete medium containing 1 g/L 5-fluoroorotic acid and 60 mg/L uracil to check for loss of the URA3 marker. Loss of the CORE-I-SceI cassette was confirmed using colony PCR (primers listed in Table S2) and the resulting PCR product was sequenced to ensure the desired mutation was present. Due to complete disruption of the MEF2 gene during insertion of the cassette, all resulting mef2 mutants were lacking mtDNA. Therefore, to reintroduce mtDNA, mutant strains were mated with the wild-type of opposite mating type (the BY4741 strain (MATa)). Diploid strains were then sporulated and dissected, resulting in 2∶2 segregation of wild-type and mef2 mutant colonies, all of which were ρ+. Sequencing was used to identify the colonies which possessed the desired mutation.
Cell respiration was measured by monitoring dissolved oxygen levels in 30 mL of exponential phase yeast cells at an optical density of 0.2 in YEPD medium. Cells were cultured in 50 mL Erlenmeyer flasks sealed with a rubber bung to minimise the exchange of oxygen with the external environment. Each flask was equipped with a PreSens PSt3 oxygen sensitive spot (NomaCorc) and the percentage of dissolved oxygen in the medium was measured using an oxoluminescent device, the NOMASense oxygen analyser system (NomaCorc) [35], every 15 minutes until percent oxygen reached below 10% for the wild-type strain.
MitoTracker Red CMXRos (Molecular Probes) was added directly to a 500 µL volume of exponential phase culture in YEPD to a final concentration of 250 nM (1×) or 2500 nM (10×). Cells were incubated at 30°C for 30 minutes, washed with fresh medium and resuspended in YEPD medium. Cells were mounted onto a glass slide and viewed immediately using a laser scanning confocal microscope (Zeiss LSM 710, Carl Zeiss Microimaging, Germany) controlled using the ZEN 2010 software (Carl Zeiss Microimaging, Germany). The excitation line used was 543 nm and the laser power was set at 2%. Cells were viewed using 630× optical magnification and 3× digital magnification. All samples were analysed using the same settings.
The protein model of the human EF-G2mt protein was based on the crystal structure of the Thermus thermophilus EF-G protein (Protein Data Bank 2bm0) [42]. This shares 39% residue identity. Selection of the T. thermophilus template and homology modelling was carried out using the SWISS-MODEL server in ‘project mode’ to enable alignment inspection prior to modelling [41]. The completed model was then submitted to the SWISS-MODEL suite of quality check programs which tests for model quality and stereochemistry using algorithms such as ANOLEA [56] and PROCHECK [57]. The model was visualised using the Visual Molecular Dynamics program (VMD), version 1.8.7 (University of Illinois) and this software was also used for the assignment of protein secondary structure and the assessment of electrostatic potential.
All sequences, both yeast and human, were obtained from the Ensembl database. There are three known human EF-G2mt protein isoforms but isoform I (AAH15712.1) was used throughout this study. Human EF-G2mt SNPs were identified using the dbSNP database available on the National Center for Biotechnology Information (NCBI) website (www.ncbi.nlm.nih.gov/SNP/index.html). The Lalign program available on the Swiss EMBnet server (www.ch.embnet.org/software/lalign_form.html) was used to generate global alignments of protein sequences, based on the BLOSUM50 matrix [58]. Graphical representations were constructed using the BioEdit version 7.0.5 sequence alignment editor.
|
10.1371/journal.pcbi.1002999 | Outlier Responses Reflect Sensitivity to Statistical Structure in the Human Brain | We constantly look for patterns in the environment that allow us to learn its key regularities. These regularities are fundamental in enabling us to make predictions about what is likely to happen next. The physiological study of regularity extraction has focused primarily on repetitive sequence-based rules within the sensory environment, or on stimulus-outcome associations in the context of reward-based decision-making. Here we ask whether we implicitly encode non-sequential stochastic regularities, and detect violations therein. We addressed this question using a novel experimental design and both behavioural and magnetoencephalographic (MEG) metrics associated with responses to pure-tone sounds with frequencies sampled from a Gaussian distribution. We observed that sounds in the tail of the distribution evoked a larger response than those that fell at the centre. This response resembled the mismatch negativity (MMN) evoked by surprising or unlikely events in traditional oddball paradigms. Crucially, responses to physically identical outliers were greater when the distribution was narrower. These results show that humans implicitly keep track of the uncertainty induced by apparently random distributions of sensory events. Source reconstruction suggested that the statistical-context-sensitive responses arose in a temporo-parietal network, areas that have been associated with attention orientation to unexpected events. Our results demonstrate a very early neurophysiological marker of the brain's ability to implicitly encode complex statistical structure in the environment. We suggest that this sensitivity provides a computational basis for our ability to make perceptual inferences in noisy environments and to make decisions in an uncertain world.
| Survival crucially depends on our ability to extract information from the environment. This ability relies on learning about regularities that enable us to make predictions about what is likely to happen next. Sensitivity to violations of these regularities is necessary for timely reactions and adaptive responses to unexpected, or odd, events. Prior work on speech acquisition and artificial grammar learning has provided important behavioural evidence that humans are able to learn statistical regularities, but it still falls considerably short of providing a biological understanding for how these processes might take place in the brain. The neurophysiological study of regularity extraction has so far been limited, to either sequence-based rules or to simple change-detection paradigms, and thus the neurobiological mechanisms that underpin statistical learning remain unknown. Here we provide both behavioural and neurophysiological evidence to show that humans keep track of the uncertainty in apparently random distributions of events. Our work demonstrates that an early neurophysiological signal underlies the fundamental human ability of learning and making inferences in an uncertain world.
| The survival of an organism often depends on its ability to form expectations about the structure of its sensory environment, and to monitor the environment for violations of these expectations so as to respond to unexpected and potentially threatening events [1]–[5]. In many instances this goal is rendered challenging by the unpredictability of even the normal environment [6].
Several studies have examined an ability that humans have to implicitly learn regularities in observed stimuli [7]. In the auditory domain many of these studies have used an oddball paradigm [8], in which participants are presented with a sequence of events that mostly obey a certain rule, but which is punctuated by occasional “oddballs” or events that violate that rule. These oddballs frequently evoke conspicuous neurophysiological activity, reflected in the so-called mismatch negativity (MMN) response. The MMN can be observed in electro- and magneto-encephalographic recordings (EEG and MEG) with a time latency of about 100 to 200 ms from violation onset [8]. In the classical, and simplest, oddball paradigm, the main sequence comprises identical tonal stimuli called “standards”. The oddballs are events that differ from the standards in some physical aspect such as frequency [9]–[12], duration [13], or amplitude [14]. The MMN response is robustly elicited in all these cases and represents a neurophysiological marker both of the internalisation of the regularity, and of the change detection.
Other experiments have observed MMN signals associated with the violation of more sophisticated rules. Examples include: a tonal sequence in which the higher the frequency of a tone, the louder its amplitude, with violation by a high-frequency soft or low-frequency loud tone [15]; a sequence of regularly descending tone pairs broken by an occasional ascending combination [16]; or a regular rhythmic pattern violated by an unexpectedly-timed event [17].
Despite the sophistication of these regularities, and the occasional randomisation of any aspects of the standard stimuli that are irrelevant to the rule [13], these studies predominantly rely on establishing a deterministic, often sequence-based pattern [18], [19], against which oddballs may be judged. A few studies have introduced some variability into the distribution of standards and observed that MMN amplitude decreased when the range of variability increased [20]–[22]. However, in these studies the “oddball” was still outside the distribution of standards, and the standards were chosen from a small set of discrete known tones, to which listeners had become accustomed. By contrast, we were interested in whether human listeners implicitly learn a statistical regularity in a stream of unknown, continuously-distributed, stochastic stimuli: a simple probabilistic pattern that could not be encoded as a deterministic sequence-based rule, a finite set of known standard stimuli, or by a categorical separation between expected events and outliers. In our study, the definition of the standards, or the context, is entirely probabilistic. As opposed to situations [20], [22] where oddballs were clearly outside the range of standard variability, here both standards and oddballs are part of the same distribution. Importantly, by its nature, such regularity does not provide for a categorical separation between expected events and outliers. Instead, as expectations are themselves uncertain, oddball events must be defined quantitatively in terms of how unlikely, or surprising they are given the probabilistic expectations formed by the majority of stimuli. Thus, our prediction was that a comparison between the neurophysiological responses evoked by very likely events to those evoked by very unlikely events should reveal an MMN-like deflection. Moreover, the strength of this MMN should depend on the relative likelihoods of the stimuli being compared. In particular, by establishing two different statistical contexts, with two different distributions of expected events, we could manipulate the probability associated with the occurrence of the very same physical stimulus in each of the two contexts.
Our key prediction here was that the stimulus should evoke a larger prediction error, or MMN response, when embedded within the context in which it was less likely. In effect, by investigating human listeners' sensitivity to statistical outliers embedded within a distribution of random stimuli, we could determine both whether listeners are able to implicitly learn and encode a statistical pattern, and whether detection of outliers from this pattern evokes an MMN signal similar to that evoked by violations of deterministic sequences.
Ten human participants reported changes in the luminance of a fixation cross while ignoring a background sound made up of tone pulses. The frequencies of most of the tone pulses in the background sounds were drawn from either a narrow or a broad distribution both of which were Gaussian in the log domain (Figure 1). These distributions provided two different contexts for the presentation of probe tones that were interspersed within the random streams. The probe tone frequencies were either equal to the centre frequency of the contextual distributions (500 Hz, likely or “standard”), or two octaves above it (2000 Hz, unlikely or “odd”). In fact, in this first experiment each probe tone coincided with a change in luminance in the fixation target, but participants were not informed of this.
Nonetheless, we found that the reaction time to the luminance change was shorter when it was accompanied by an odd probe tone rather than a standard tone regardless of the context (p = 0.002, ANOVA main effect, see Figure 2). This finding is consistent with earlier observations that auditory outliers embedded in simple deterministic patterns also facilitate visual target detection [23]. We also found that reaction times were shorter overall in the narrow as compared to the broad context (p = 0.004, ANOVA main effect) and, crucially, that responses to luminance changes paired with odd probe tones in the narrow context were faster than those to changes paired with the same odd probe tones in the broad context (p = 0.043, ANOVA interaction, p = 0.0076 post-hoc t-test).
Thus, these behavioural data show that listeners were indeed sensitive to the contextual distribution and its associated probabilities. Although the odd probes themselves were embedded at the same rate in both contexts, when tones of similar frequency were less likely to occur by chance within the background distribution, then the associated facilitation of behavioural responses was stronger.
We next asked whether sensitivity to contextual statistical properties had a neurophysiological fingerprint. To avoid possibly confounding motor or attentional signals associated with the visual task, we modified the experimental design to break the association between probe sounds and visual events. Thus, in this variant of the task, the visual and auditory streams were entirely unrelated. Eighteen naïve participants performed this modified task while we recorded neural activity with MEG. As before, participants were presented with probe sounds embedded in two contextual frequency distributions characterised by equal means and different variances, so that physically-identical odd probes were more unlikely under the narrow (low variance) than under the broad (high variance) distribution. These two distributions were presented in two separate blocks and the block order was counter-balanced across participants to avoid order effects (Figure 1).
We performed a full spatio-temporal statistical analysis, searching for significant differences between the magnetic fields evoked by odd and standard probe sounds (evoked response fields or ERFs), treating these probes as analogous to the oddballs and standard events respectively in a classical MMN paradigm. At the scalp level, we found three left-lateralised clusters (family wise error (FWE) corrected at p<0.05) peaking at about 160, 190, and 310 ms over temporal-parietal areas (Figure 3a). These corresponded in latency and shape to the traditional MMN and P3a components, as predicted. Thus, this finding suggests that in a background of tones with randomly-distributed frequencies, sounds whose frequencies lie distant from the mean are registered physiologically as outliers and treated differently to tones whose frequencies fall at the centres of the distributions. However, although the timing, form and localisation of this response are all reminiscent of such a statistically-driven MMN effect, we could not rule out a contribution due to the differing tone frequencies of the odd and standard probes.
Thus, the crucial test was our second prediction: that the size of the outlier response (odd probe response minus standard probe response) would be greater in the context of the narrow (relative to the broad) distribution. This is exactly what we found: an MMN-like response peaking at about 120 ms over bilateral temporal-parietal areas, which was stronger (more negative) in the context of the narrow distribution (Figure 3b). Here, the comparison is between responses evoked by physically identical sounds; the only difference lies in the context, with the frequency of the odd probe tone being much less likely to occur under the narrow distribution. This difference in the magnitude of the MMN-like response was mediated by context-dependent changes in the amplitude of both the responses to odd and to standard probes; responses to standard probes were smaller and those to odd probes were larger when the same two sounds were played in the narrow context (Figure S2).
We then reconstructed putative magnetic-field sources from the scalp activity (using a multiple sparse priors MSP inverse solution [24], [25]), in order to infer the cortical regions most likely to have generated the signals observed in the expected MMN time-window (between 100–200 ms from stimulus onset). As described above, the differences in fields measured at the scalp were clearly statistically significant, even when we accounted for the fact that tests were repeated for each electrode and each time point. Here, we sought to identify the most likely sources of these significant effects. As the number of cortical voxels considered was much larger than the number of electrodes, differences in reconstructed source activity generally did not appear significant after naïve correction for multiple testing – even though the significant scalp-field differences must, of course, originate from sources somewhere in the brain. Thus, we identified putative source locations using uncorrected (p<0.05) significance thresholds, and then performed a set-wise significance analysis using anatomical masks defined by prior studies. This procedure does leave open the possibility of errors in the precise localisation of activity within each defined set.
We found a main effect of surprise in bilateral visual, parietal, and sensory-motor cortices (p<0.05, uncorrected) (see Figure 4a). We also found an interaction effect (p<0.05, uncorrected) in similar areas as well as in auditory association cortex, manifest in a stronger response to odd probes under the narrow compared to the broad distribution. Although these effects were not significant when corrected for the multiple comparisons performed over the whole brain, they were included in sets that showed significant effects (p<0.05, corrected) when defined by anatomical masks derived from previous independent studies (WFU Pick Atlas [26]) for regions expected a priori, such as the effects seen in temporal [9], [27]–[30] and the parietal [28], [30] regions. Thus, these sources showed MMN-like responses that were larger when the probes were more unlikely under the contextual distribution (Figure 4b).
As opposed to what we had predicted, we did not find a main effect of surprise in the associative auditory cortex [9], [27], [29]. In the narrow context alone, there was indeed a simple main effect of surprise in the associative auditory cortex (Figure 4c). However, the same contrast did not reveal an effect under the broad distribution (even at p<0.1). This might indicate that an MMN-like response is generated in the auditory cortex only in the narrow context, where the odd probes are more unlikely, although it may also just reflect an interaction between a larger MMN effect-size in the narrow context and a noisier reconstruction of auditory cortical sources than parietal ones.
The analyses so far show that there is an MMN-like differential response to sounds that are unlikely in a statistical context, when contrasted with higher probability sounds, and that this difference is stronger when the difference in probabilities is greater. There has been some debate about the extent to which MMN-like responses to frequency deviants may be mediated by frequency- and temporally-local adaptive processes [31]–[34]. A number of different mechanistic theories for MMN generation have been discussed in great depth (see [35], [36] for reviews). One explanation rests on the fact that changes lead to release from adaptation to repeated events, or refractoriness, resulting in an enhanced response to a novel stimulus [33], [34]. While this theory is useful in the case of repeated standards, it does not explain very well why MMN is elicited by more abstract rules that do not involve change, or a break in repetition [15], [16]. Recent efforts to disentangle refractoriness and memory-comparison-based contributions to MMN have been able to demonstrate that there is more to MMN than simple adaptation [37]. However, in light of this debate, it was important to determine how far local adaptive or refractory processes might have contributed to the phenomena we observed.
Although tones with exactly the odd probe frequency never occurred by chance under either contextual distribution, tones with nearby frequencies did, and did so more often in the broad context than in the narrow. Thus a local spread of adaptive effects, whereby these contextual tones reduce the size of physiological response to tones of nearby frequencies including the odd probe tones, might contribute to the difference in response magnitudes between the two contexts. On a long-time scale spanning multiple stimuli [38], such an effect could mediate the physiological mechanism underlying sensitivity to environmental statistics. However, if the bulk of the effect were due to adaptation arising from very recent stimulation, then there would be less reason to posit a mechanism that accrues and responds to longer-term regularities.
To study the contribution of local adaptation, we grouped responses to odd probe sounds according to the number of preceding sounds, , that fell outside a frequency window of width specified in octaves and centred on the frequency of the odd probes (Figure 1). Responses to tones with larger would be expected to suffer less short-range adaptation. We used window widths ranging from one-third of an octave (roughly the equivalent rectangular bandwidth, ERB, of a psychoacoustically-defined ‘auditory filter’ [39]) to five times the ERB. Wider windows yielded too few data to be interpretable.
As there were relatively few responses contributing to each average ERF in this analysis, particularly for larger values of , we applied a spatial projection filter to find a single weighted combination of all the MEG channels, common to all participants, that maximised the signal-to-noise ratio of the filtered signal for all four types of probe sounds [40]: standard and odd in narrow and broad contexts (Figure 5a, b). Importantly this spatial filter was not designed to accentuate differences between the responses to the different probes. Indeed, depending on the geometry of the signals and noise, it might be expected to slightly suppress the magnitude of these differences, whilst simultaneously reducing noise in each signal.
Figure 5c shows the filtered components associated with odd probe sounds selected according to increasing minimum values of , using the ERB-sized window, as well as the filtered components for standard probes without selection by , for both distributional contexts. As the threshold value of increases, the peak amplitude of the response to odd probes in the broad context grows more negative, presumably reflecting a contribution from local adaptive mechanisms. However, even when the preceding 15 tones (lasting 7.5 s) fell outside the ERB window, the response to odd probes in the broad context does not reach the same level as that to odd probes in the narrow context (p<0.05 random permutation tests applied separately for each threshold). This general result held true for a range of exclusion windows (Figure 5d), with the mean ERFs remaining systematically separate, although for the larger exclusion windows there were too few stimuli with large enough values of for the effect to still reach significance. The same observation held when responses were grouped according to small ranges of , rather than by threshold values (Figure S1). Thus, the difference in response magnitude to odd probe sounds in the two distributional contexts cannot be attributed solely to local adaptation mechanisms, and depends on the distribution of stimuli well outside the auditory filter or more than 15 stimuli or 7.5 s in the past.
A striking and unexpected observation was that the response to odd probes in the narrow context seemed to be relatively unaffected by . This might well reflect a difference in the adaptive impact of other odd probe sounds when compared to tones whose frequencies are drawn from the distributional context. In the narrow context most preceding tones that fell within were themselves odd probe sounds. The fact that very little adaptation is seen even when such a tone happens to have fallen in the very recent past ( or 2, see Figure S1), raises the intriguing possibility that once isolated tones are marked by a physiological mechanism as outliers, they have only reduced, or even non-existent, adaptive impact on subsequent tones.
We provide behavioural and neurophysiological evidence that humans implicitly track statistical regularities of the sensory environment. Specifically, our findings show that stimuli that fall outside an established stochastic pattern evoke behavioural and neurophysiological responses previously associated with violations of a repetitive or deterministic sequence. Furthermore, we found that exactly the same physical stimulus, arriving at the same rate, evokes faster reaction times (Figure 2) and larger MMN (Figure 3b) when it is embedded in a statistical pattern, which makes its occurrence less likely. Furthermore we demonstrated that sensitivity to statistical context, as indexed by different MMN amplitudes, goes beyond local adaptation of afferent activity in narrow frequency bands (Figure 5), although adaptation plays an important role. These observations are consistent with the idea that observers build an internal model of the predominant stochastic distribution of stimuli, and implicitly and automatically monitor the environment for stimuli that are outliers to this distribution.
Earlier behavioural studies [23], [41] have shown that a high tone embedded in a sequence of low tones (or, indeed, a similar regularity-violating stimulus within another modality) improves the detection of a simultaneously-presented visual target. In our experiment, the visual stimulus was obvious enough that it was very rarely missed. Instead, we noted a consistent decrease in reaction time when the target onset was simultaneous with an auditory outlier. We speculate that a reaction-time effect of this sort is unlikely to depend on relatively slow executive or voluntary attentive mechanisms that exploit a learnt association; but instead reflects a low-level multimodal integration driven by the rapidity of auditory processing [42]. In particular, this suggests that the processes by which a statistical model of the environment is formed, and exceptions detected, lie at an early stage of sensory processing and can act autonomously of controlled attentionally-demanding executive processing.
This view is supported by the finding that the same probe stimuli embedded in identical random-frequency backgrounds evoke an MMN-like response (Figure 3b). There is an extensive literature showing rapid, and presumably automatic, sensitivity to changes in stimuli and to violations of deterministic [16] or sequence-based rules [18]. We add to these findings by showing that observers also implicitly learn statistical structure, and can detect outliers from a random distribution. This is evident in the MMN-like MEG signal that has similar timing to the conventional MMN evoked by sequence violations. Again, we find that the underlying physiological processes are sensitive to the overall statistics of stimuli and to the likelihood of an event conditioned upon its temporal context [22]: identical odd probes generated larger MMN responses when embedded in a narrower random-frequency context.
Source reconstruction suggested that the statistical-context-sensitive responses arose in the parietal and temporal cortices [9], [27], [30]. Intriguingly, sensitivity to sound statistics was not strong in the auditory cortex, as observed in previous studies that used conventional MMN paradigms, but rather in a region posterior to the primary auditory cortex, which agrees with prior work on statistical learning [29], [43]. These temporal-parietal areas have also been associated with stimulus-driven bottom-up saliency [28] and attention orientation to unexpected events [44]. Some behavioural studies have suggested that such involuntary redirection of attention may interfere [13], [14] with goal-directed behaviour, whereas others have shown evidence pointing towards facilitation [23], [41].
An important question is whether the sensitivity we see reflects global statistics, integrated over a wide range of frequencies and long time periods, or whether it is partly due to local adaptation. In particular, it might be that known stimulus-specific adaptive phenomena in the auditory cortex [45] underlie the physiological and behavioural changes that we observe. Three lines of argument point towards more global processing. First, although most stimuli in the auditory stream had frequencies drawn from either the narrow or the broad contextual distribution, the odd probe stimuli themselves were presented at the same rate within both contexts. Furthermore, even in the broad context only about 1% of random contextual tones fell within a semitone-wide band around the odd probe frequency. Thus the pattern of 2000-Hz tones themselves differed negligibly between the two contexts (a very different situation to that found in studies of stimulus-specific adaptation [38], [45] where it is the probabilities of the different tones themselves that are varied), and the difference in both behavioural and physiological response effects must have been due to the distribution of contextual tones at other frequencies. Second, we looked at the impact that the recency of local-frequency stimulation had on the magnitude of the MMN. We certainly noted signs of local frequency-specific adaptation in the responses to odd probe tones occurring within the broad context: there, odd probes proceeded by longer periods in which all stimuli fell outside a frequency window (i.e., large ) evoked a stronger response than those with recent in-window stimulation (Figure 5c, d). However, even when we looked only at odd probes in the broad context for which the preceding 15 or more tones (spanning 7.5 s) fell outside a window with rectangular bandwidth equivalent to a psychophysical auditory filter or wider, the size of the MMN was significantly larger than that evoked by odd probes in the narrow context. (Long time-scale adaptation has also been reported in the context of SSA [38], but may well point to more global processing even in that context). Third, we observed very little, if any, local adaptive difference in the magnitude of MMN evoked by odd probes in the narrow context. In this context most of the potentially adapting stimuli that fell within the frequency window of the analysis were themselves other odd probe tones. Thus, it seems that when presented in the narrow context, odd probe tones remain surprising even if another odd probe was presented very recently. Indeed, it seems possible that odd probes in the narrow context are heard as distinct, as though they were generated by a different process. Again, this finding points to a sensitivity to the global context [22], [31] within which specific stimuli are heard. Taken together, these results highlight the fact that the brain's ability to detect regular patterns in the environment, and register violations of these patterns, goes beyond repetitive [8] or sequence-based rules [16], [46]. Instead, the apparently implicit physiological process associated with the MMN seems to extract at least the first two moments of the distribution of stimuli in the environment, and then to drive stronger responses to tones that are more improbable within this encoded distribution.
Our results demonstrate the brain's ability to implicitly and efficiently encode the statistical regularity of events in the environment. In keeping with normative ideas of predictive coding [3], [5], [47] and of sensitivity to varying forms of uncertainty [6], we suggest that this computation might be essential for perceptual processing in a noisy environment, and for decision-making in an uncertain world.
Informed consent was obtained from each subject, after full explanation of the experiment, according to the procedures approved by the University College London Research Ethics Committee.
We recorded behavioural reaction time data from ten participants (5 females, 5 males, age range 24–32 years, and mean age 26 years), and MEG data from a separate pool of eighteen participants (8 males, 10 females, age range 19–47 years, and mean age 28 years). All participants were healthy volunteers, had normal to corrected vision and hearing, and were naive to the purpose of the study. Participants were monetarily compensated for their time.
We designed a novel paradigm in which participants passively listened to a stream of pure tone sounds, whilst performing a visual change-detection task. The experimental design is depicted in Figure 1. The frequencies of most of the tone pulses were sampled from a Gaussian distribution in log-frequency, centred at 500 Hz and with one of two standard deviations: low ( octaves) or high ( octaves). All tone pulses had an equal duration of 50 ms with smooth rise and fall periods of 10 ms each, were set to the same comfortable loudness level throughout the experiment, and were presented every 500 ms. Probe tones, whose frequencies were either equal to the mean of the distributions (standard probes: ) or two octaves above it (odd probes: ), were embedded within the random-frequency streams. Both types of probe tone were inserted into the stream pseudo-randomly, with each presented 10% of the time. The probe tones were not distinguished from those of the background distribution, and thus participants experienced a slightly distorted Gaussian distribution of frequencies which combined two point-masses of 10% probability each (the standard and odd probes) with the Gaussian contributing 80% of the frequencies. Although this renders the overall distribution not strictly Gaussian, the 10% prevalence of probes was necessary to ensure a sufficient number of trials (and SNR) to see the MMN-like response. Only the width of the Gaussian part of the distribution differed between the two conditions. Behavioural and physiological comparisons were all based on responses to the probes alone, and these were identical in both conditions. Participants were told to ignore the sounds and respond only to the visual task, which required them to press a key each time they saw a brief change in the luminance of a fixation cross. The interval between successive luminance changes was randomly chosen between 2000 ms and 5000 ms, in steps of 500 ms. In the behavioural experiment (Experiment 1), these changes were coincident with the probe sounds, both standard and odd. In the MEG experiment (Experiment 2) luminance changes were made independent of the sound stream so as to avoid introducing confounding motor or attentional signals.
Experiment 1 lasted approximately 30 minutes and was divided into 4 blocks. Both high and low variance conditions were presented in all blocks for half of the time, and in randomised order. In Experiment 2, the high and low variance conditions were presented in two separate blocks that each lasted for 13 minutes (resulting in about 160 probe tones of each type per condition per participant). The order of these blocks was counter-balanced across participants.
The stimuli and task protocols were written in MATLAB, using the Cogent 2000 toolbox (http://www.vislab.ucl.ac.uk/cogent.php).
Measurements were acquired with a CTF 275-channel whole-head MEG system, with 274 functioning second-order axial gradiometers arranged in a helmet shaped array. Three energised electrical coils were attached to the fiducials (nasion, and left and right preauricular), in order to continuously monitor the position of each participant's head with respect to the MEG sensors. Auditory stimuli were binaurally presented at a comfortable loudness level through flexible tubing connected to piezo electric transducers positioned approximately 1 m below the sensor array. Data were collected at a sampling rate of 600 Hz, and recording epochs extracted stretching from 100 ms before to 350 ms after the onset of each sound (Figure 1, inset). For the spatio-temporal and source analyses, the data were filtered with a passband between 0.5 and 30 Hz, down-sampled to 200 Hz, and baseline-corrected with reference to the pre-tone interval (−100–0 ms).
The averaged sensor data, or ERFs, were converted into 3-dimensional spatio-temporal volumes. This was achieved by interpolating and dividing the scalp data per time point into a 2-dimensional spatial 64×64 matrix. We obtained one 2D image for every time bin. These images were then stacked according to their peristimulus temporal order resulting in a 3D spatio-temporal image volume with dimensions 64×64×91 per participant.
For each subject, the 3D spatio-temporal image volumes were modelled with a mass univariate general linear model (GLM) as implemented in SPM [48]. We modelled the data with one regressor per condition: standard (s) and odd (o) probes, under high (h) and low (l) contextual variances, yielding coefficients , as well as two nuisance regressors that accounted for the block and variance factor confound. We then performed within-subject F-contrasts for (1) the main effect of surprise (regardless of contextual variance) [], and (2) the interaction or differences between odds and standards in the context of a low, as compared to a high variance distribution, []. We then carried these contrasts over to a one-sample between-subject F-test statistic and assessed the significance of the tests across the group. This approach allowed us to make inferences on all 3 data dimensions, i.e., 2-dimensional sensor space over the whole peristimulus time dimension. The same sort of statistical analysis was performed on the 3D spatial image volume obtained after the source localization step (see below). All sensor effects are reported at a threshold of p<0.05, with a family-wise error (FWE) correction for multiple-comparison for the whole volume.
We obtained source estimates on the cortical mesh by reconstructing scalp activity with a single-shell head model, and inverting a forward model with multiple sparse priors (MSP) assumptions for the variance components [25] under group constraints [24]. This allowed for inferences about the most likely cortical regions that generated the data observed in the MMN time-window [100–200 ms], pre-selected according to predictions derived from previous MMN studies and our scalp results. We obtained images from these reconstructions for each of the four conditions in every subject. These images were smoothed at FWHM 8×8×8 mm3. We then computed the main effect of surprise and interaction (surprise by contextual variance) using conventional SPM analysis [48]. Similarly to the spatio-temporal statistical tests (described above), we were able to search for significant effects over the whole brain 3-dimensional space. Effects (t-statistics) are displayed at an uncorrected threshold of p<0.05 (Figure 4). These weaker significance criteria were used for post-hoc visualisation, once the effects had been established under robust criteria at the scalp level.
The local adaptation analysis (described below) required the construction of ERFs based on relatively small numbers of responses. As single-channel ERFs for low trial counts exhibited substantial noise, we used a single-output spatial filter to combine all channels in a way that maximised the average signal-to-noise ratio across the four probe conditions [40]. We constructed two covariance matrices as follows. Let be a vector representing the multidimensional MEG measurements at sample (of in the epoch), associated with the nth repetition (of Nc for the condition), of probe tones under the condition labelled by c (either standard or odd probes, in either the high- or low-variance context); and let be the mean measurement over all Nc stimuli of that condition. Then the condition-specific signal power matrix isand the noise covariance matrix isWe constructed overall signal power S and noise covariance V matrices by averaging the four corresponding condition-specific matrices. The maximum signal-to-noise spatial filter was then the eigenvector w corresponding to the largest eigenvalue solution of the generalised eigenvalue problem:ERFs were constructed for each condition using the one-dimensional filtered signal (Figure 5c):The filter w represents a compromise between spatial patterns of greater signal and those of lesser noise (Figure 5a). However, it is possible to recover the effective signal direction either by correlating the filtered output with the multidimensional MEG signal or (equivalently, and more simply) pre-multiplying w by the noise matrix V (Figure 5b).
To examine the contribution of local adaptive mechanisms to our findings, we labelled responses to odd probe tones according to the number (Na) of preceding tones of all types that fell outside a window of full-width in log-frequency centred at the odd probe frequency (see Figure 1). We then averaged responses for which Na exceeded a threshold value (Figure 5c, d) to yield separate ERFs for each local adaptation condition. The (negative) peak value was computed for each separate response by averaging samples that fell within 30 ms of the minimum of the corresponding ERF curve. These peak values were compared between the broad and narrow context using a one-tailed sampled permutation test in which the context labels for each response were randomly permuted 2500 times, with the fraction of permutations for which the difference between broad and narrow context responses was greater than that observed for the true assignment providing the corresponding p-value. As we were interested in the hypothesis that the difference would be significant in all adaptation groups, we used an independent threshold of 0.05 for significance.
All the analyses were performed with SPM (http://www.fil.ion.ucl.ac.uk/spm/) and in-house MATLAB code.
|
10.1371/journal.pmed.1002428 | Bioequivalence between innovator and generic tacrolimus in liver and kidney transplant recipients: A randomized, crossover clinical trial | Although the generic drug approval process has a long-term successful track record, concerns remain for approval of narrow therapeutic index generic immunosuppressants, such as tacrolimus, in transplant recipients. Several professional transplant societies and publications have generated skepticism of the generic approval process. Three major areas of concern are that the pharmacokinetic properties of generic products and the innovator (that is, “brand”) product in healthy volunteers may not reflect those in transplant recipients, bioequivalence between generic and innovator may not ensure bioequivalence between generics, and high-risk patients may have specific bioequivalence concerns. Such concerns have been fueled by anecdotal observations and retrospective and uncontrolled published studies, while well-designed, controlled prospective studies testing the validity of the regulatory bioequivalence testing approach for narrow therapeutic index immunosuppressants in transplant recipients have been lacking. Thus, the present study prospectively assesses bioequivalence between innovator tacrolimus and 2 generics in individuals with a kidney or liver transplant.
From December 2013 through October 2014, a prospective, replicate dosing, partially blinded, randomized, 3-treatment, 6-period crossover bioequivalence study was conducted at the University of Cincinnati in individuals with a kidney (n = 35) or liver transplant (n = 36). Abbreviated New Drug Applications (ANDA) data that included manufacturing and healthy individual pharmacokinetic data for all generics were evaluated to select the 2 most disparate generics from innovator, and these were named Generic Hi and Generic Lo. During the 8-week study period, pharmacokinetic studies assessed the bioequivalence of Generic Hi and Generic Lo with the Innovator tacrolimus and with each other. Bioequivalence of the major tacrolimus metabolite was also assessed. All products fell within the US Food and Drug Administration (FDA) average bioequivalence (ABE) acceptance criteria of a 90% confidence interval contained within the confidence limits of 80.00% and 125.00%. Within-subject variability was similar for the area under the curve (AUC) (range 12.11–15.81) and the concentration maximum (Cmax) (range 17.96–24.72) for all products. The within-subject variability was utilized to calculate the scaled average bioequivalence (SCABE) 90% confidence interval. The calculated SCABE 90% confidence interval was 84.65%–118.13% and 80.00%–125.00% for AUC and Cmax, respectively. The more stringent SCABE acceptance criteria were met for all product comparisons for AUC and Cmax in both individuals with a kidney transplant and those with a liver transplant. European Medicines Agency (EMA) acceptance criteria for narrow therapeutic index drugs were also met, with the only exception being in the case of Brand versus Generic Lo, in which the upper limits of the 90% confidence intervals were 111.30% (kidney) and 112.12% (liver). These were only slightly above the upper EMA acceptance criteria limit for an AUC of 111.11%. SCABE criteria were also met for the major tacrolimus metabolite 13-O-desmethyl tacrolimus for AUC, but it failed the EMA criterion. No acute rejections, no differences in renal function in all individuals, and no differences in liver function were observed in individuals with a liver transplant using the Tukey honest significant difference (HSD) test for multiple comparisons. Fifty-two percent and 65% of all individuals with a kidney or liver transplant, respectively, reported an adverse event. The Exact McNemar test for paired categorical data with adjustments for multiple comparisons was used to compare adverse event rates among the products. No statistically significant differences among any pairs of products were found for any adverse event code or for adverse events overall. Limitations of this study include that the observations were made under strictly controlled conditions that did not allow for the impact of nonadherence or feeding on the possible pharmacokinetic differences. Generic Hi and Lo were selected based upon bioequivalence data in healthy volunteers because no pharmacokinetic data in recipients were available for all products. The safety data should be interpreted in light of the small number of participants and the short observation periods. Lastly, only the 1 mg tacrolimus strength was utilized in this study.
Using an innovative, controlled bioequivalence study design, we observed equivalence between tacrolimus innovator and 2 generic products as well as between 2 generic products in individuals after kidney or liver transplantation following current FDA bioequivalence metrics. These results support the position that bioequivalence for the narrow therapeutic index drug tacrolimus translates from healthy volunteers to individuals receiving a kidney or liver transplant and provides evidence that generic products that are bioequivalent with the innovator product are also bioequivalent to each other.
ClinicalTrials.gov NCT01889758.
| Consensus documents developed by professional transplantation societies worldwide have cautioned the use of generic immunosuppressants such as tacrolimus in individuals with a solid organ transplant. Reasons have included repeated switching between innovator (that is, “brand” products) and generics and among different generics, especially when not controlled by physicians.
There was uncertainty in the transplant community as to whether tacrolimus generics that are bioequivalent to the innovator are also bioequivalent to each other.
For market approval, generic drug products of the narrow therapeutic index drug tacrolimus had to be studied only in healthy individuals and not in the much more complex organ transplant population.
We performed a randomized, prospective, 3-treatment, 6-period, crossover, replicate dose study in individuals with a kidney or liver transplant.
Thirty-five individuals with a kidney transplant and 36 individuals with a liver transplant receiving tacrolimus were studied to compare the tacrolimus time concentration profiles of 3 different products in their blood: namely, Innovator (Prograf), Generic Hi (Sandoz), and Generic Lo (Dr. Reddy) 1.0 mg tacrolimus capsules.
Generic products were selected based upon pharmacokinetic data from healthy volunteer studies since bioequivalence data were not available in individuals with an organ transplant.
We observed bioequivalence based on average bioequivalence and scaled average bioequivalence criteria in individuals after kidney or liver transplant between tacrolimus innovator and the 2 generics on the US market as well as between the 2 generics.
Similar tacrolimus exposure is expected in individuals with a kidney or liver transplant when receiving Prograf, Sandoz generic, or Dr. Reddy’s generic tacrolimus.
| Most individuals receiving a solid organ transplant require lifelong immunosuppression. Switching to generic immunosuppressants may lead to significant savings and improved adherence [1,2], which is essential for long-term graft survival [3]. The current US Food and Drug Administration (FDA) generic drug approval process has performed well [4]. However, concerns persist regarding whether 2-way crossover studies in healthy individuals using conventional average bioequivalence (ABE) acceptance criteria of a 90% confidence interval contained within the confidence limits of 80.00% to 125.00% are a valid approach for generic immunosuppressant approval for use after transplantation [5,6]. This debate started when cyclosporine generics were developed over 15 years ago [7,8] and was reinvigorated when tacrolimus generics were approved. Consensus documents developed by professional societies from the US, Europe, and Canada [9–12] have cautioned against generic immunosuppressant use, citing (1) the lack of data in transplant recipients, especially “high risk” transplant recipients; (2) the need to implement stricter bioequivalence standards, as tacrolimus is a narrow therapeutic index (NTI) drug for which small changes in dose or exposure can result in therapeutic failure or toxicity; and (3) the lack of bioequivalence data between generics. Molnar et al. published a systematic review and meta-analysis to compare the clinical efficacy and bioequivalence of generic immunosuppressive drugs in individuals with a transplant and concluded that high-quality data were lacking. The authors went further to state that given the serious consequences of rejection and allograft failure, well-designed studies on the bioequivalence and safety of generic immunosuppression in individuals with a transplant are needed [13].
Differing worldwide bioequivalence regulatory standards for NTI drugs make it difficult to interpret bioequivalence study results [14–16]. For NTI drugs, the European Medicines Agency (EMA) requires a narrower 90.00%–111.11% acceptance criterion for the area under the curve (AUC, a measure of actual body exposure to a drug) but uses the usual 80.00%–125.00% acceptance criterion for the concentration maximum (Cmax) for NTI drugs [14]. Health Canada has adopted standards similar to those of EMA, with an AUC acceptance criterion of 90.00%–112.00% [15]. The FDA has classified tacrolimus as an NTI drug and recommended the scaled average bioequivalence (SCABE) approach to determine bioequivalence [16,17]. With this SCABE approach, both generic and innovator products are given twice with fully replicating measurements in each individual. An innovator pharmaceutical product is the one that was first authorized for marketing on the basis of quality, safety and efficacy. This allows for determination of within-subject variability, which is then used for scaling the bioequivalence acceptance limits based on the reference product for all products tested. This approach creates more stringent bioequivalence criteria: (1) the ABE limits for both AUC and Cmax are narrowed based on the within-subject variability of the reference product and are never wider than 80.00%–125.00%, and (2) the within-subject variabilities of all products are compared to each other.
Tacrolimus has a complex pharmacokinetic profile, as it is metabolized mainly by hepatic and intestinal cytochrome (CYP) P4503A enzymes and over 90% is eliminated as metabolites. CYP3A5 expressers (CYP3A5 *1/*1 and CYP3A5 *1/*3) are considered patients who “poorly absorb” and may exhibit higher within-subject variability of tacrolimus pharmacokinetics than nonexpressers (CYP3A5 *3/*3). These genetic differences have been associated with poorer outcomes. Tacrolimus is also a substrate of the drug efflux protein p-glycoprotein, ABCB1, thus impacting tacrolimus exposure [17–24]. Because of these complex metabolic and transport processes, stringent ABE testing is used to ensure product excipients do not impact these processes.
Given the aforementioned public concerns [9–12], we hypothesized that 2 generic tacrolimus products currently on the US market meet both FDA ABE and SCABE limits in individuals with a kidney or liver transplant when compared to innovator tacrolimus and when compared to each other in a high-quality study. All products met these bioequivalence criteria. In addition, we applied EMA NTI ABE criteria, and all products met the criteria except for one that narrowly fell above the AUC limit.
The study design was developed in collaboration with the American Society of Transplantation (Mount Laurel, New Jersey, US), the American Society of Transplant Surgeons (Arlington, Virginia, US), and the FDA. Individuals were recruited from the University of Cincinnati Medical Center and The Christ Hospital in Cincinnati. This trial adhered to the Declaration of Helsinki and was approved by local institutions’ review boards (2012–4891) and the FDA Research Involving Human Subjects Committee (13-018D). All individuals provided written informed consent. The study was monitored locally and by the FDA and registered on clinicaltrials.gov (NCT-01889758). Methodologies for tacrolimus quantification in whole blood (Tables A–F in S1 Appendix and Fig A in S1 Appendix) and genetic polymorphism testing (Table G in S1 Appendix) are described. The study protocol and statistical analysis plan are also included in the supporting information as S1 Text and S2 Text. Changes from the prespecified analysis plan included the analysis of the minimum concentration (Cmin) in lieu of C0 and C12 tacrolimus concentrations as appropriate based upon guidance documents, and dose normalization was not performed because each individual received the same dose in all treatment periods. No interim analyses were conducted prior to these data analyses.
At study initiation, 5 FDA-approved generic tacrolimus products were available in addition to the innovator product, Prograf (Astellas, Northbrook, Illinois) [25]. Abbreviated New Drug Applications (ANDAs) are submitted to the FDA for all generic products and represent the only data readily available for all products. ANDA data include, but are not limited to, pharmacokinetic data in healthy volunteers that demonstrate bioequivalence between the innovator and generic products by evaluating the pharmacokinetic parameters of AUC and Cmax. Product composition, manufacturing, and pharmacokinetic data for all approved generics were reviewed to identify the 2 most disparate generics. ANDA pharmacokinetic data from each product are provided in Table H-I in S1 Appendix. Additional manufacturing comparisons are summarized in Table J in S1 Appendix [26–31]. One tacrolimus product (Panacea; Baddi, India) was FDA-approved but not commercially available, and it was therefore excluded. Pharmacokinetic parameters of AUC and Cmax were examined for the greatest difference between the generic and the innovator product as being the most disparate and named Generic Hi and Generic Lo. Sandoz tacrolimus (Sandoz, Princeton, New Jersey, US) was identified as Generic Hi based upon higher point estimates and higher upper 90% confidence interval compared to innovator. Dr. Reddy tacrolimus (Dr. Reddy, Bachupally, India) was identified as Generic Lo based upon lower point estimates and lowest lower 90% confidence interval compared to innovator. Single tacrolimus 1 mg capsule lots (the most frequently prescribed dosage strength) of Innovator (Prograf), Generic Hi (Sandoz, Princeton, New Jersey, US), and Generic Lo (Dr. Reddy, Bachupally, India) were purchased from a pharmacy wholesaler and controlled by the University of Cincinnati Medical Center Investigational Drug Services. The University of Iowa Pharmaceuticals (Iowa City, Iowa, US), iC42 Clinical Research and Development (Aurora, Colorado, US), and the FDA independently performed dissolution, purity, and content uniformity testing according to applicable US Pharmacopeia Convention guidelines [29]. Similar results were obtained by both groups. The FDA results are reported in Tables K–O in S1 Appendix.
Individuals with a kidney or liver transplant included in the present study were at least 18 years old, with stable organ function and no evidence of rejection. Said individuals were at steady state and on stable doses of immunosuppressants including tacrolimus with no expected changes to their immunosuppressive drug regimens to eliminate confounders that occur early post-transplant or during times of rejection. Other eligibility criteria are listed in Section F of S1 Appendix. Study participants were stratified by organ type and randomized to 1 of 3 treatment sequences, each including 2 periods of Innovator (Prograf), Generic Hi (Sandoz), and Generic Lo (Dr. Reddy) (Fig 1). The replicate dosing design of administering each product twice allowed for analysis of within-subject variability by product. An independent statistician generated a randomization list using SAS (version 9.03, SAS Institute, Cary, North Carolina, US) and provided it to the investigational drug pharmacist. Eligible transplant recipients were recruited from 2 clinical sites, but all screening visits occurred at the University of Cincinnati Medical Center. The investigational drug pharmacist consecutively assigned individuals to a treatment sequence as received, independent of site. All parties were blinded to the randomization sequence allocation until after the pharmacokinetics (PK) analysis was completed. (Additional blinding information is located in Section G of S1 Appendix.)
Eligible transplant recipients were screened via telephone. Potential study participants completed a baseline visit, including written informed consent and laboratory, physical, and genetic polymorphism testing (including CYP3A5*3, CYP3A4*1B, CYP3A4*22, POR*28, and 3 ABCB1 SNPs) [18–24]. The study assessment schedule is shown in Table P in S1 Appendix. Study participants were seen 2 weeks later for baseline laboratory examination and randomization and were provided with study drug. Medications were dispensed with a Medication Event Monitoring System (MEMS, AARDEX, Palo Alto, California) bottle cap for electronic monitoring of study medication access. Pill counts were recorded at each pharmacokinetic visit. Individuals were excluded from the analysis if they were nonadherent within 48 hours of PK assessment (Section I in S1 Appendix).
Tacrolimus doses remained constant during the entire study period. After a 7-day treatment period, individuals underwent a 12-hour tacrolimus pharmacokinetic assessment with dosing and sampling times strictly controlled and monitored. The 7-day treatment period was adequate to reach steady state based upon observed half-life in individuals receiving a kidney or liver transplant. Fifteen tacrolimus samples were collected at C0 (before the morning dose) and at 20, 40, 60, 80, 100, 120, 140, and 160 minutes and 3, 4, 5, 6, 8, and 12 hours following dosing. Six 12-hour pharmacokinetic assessments were completed after 7 days of administration of each tacrolimus product. All samples were analyzed for tacrolimus and metabolites using a validated high-performance liquid chromatography-tandem mass spectrometry (LC-MS/MS) assay (Sections A and B in S1 Appendix). All individuals fasted until a standardized meal was allowed after the 4-hour and 12-hour sample collection. At each visit, study participants were assessed for safety lab tests, adverse events, medication adherence, and medication regimen changes (Table P in S1 Appendix).
Data were stored electronically using a REDCap database [32], including all laboratory, patient diary, and bioanalytical data. Data monitoring and analysis plans defined a priori were executed. After data were monitored and all queries resolved, the database was locked. Only individuals completing all 6 pharmacokinetic study periods were analyzed. Actual sample collection times were used for analysis. For each type of organ transplant, a sample size of 24 individuals was required to achieve 90% statistical power for concluding bioequivalence in crossover trials at an alpha of 0.05 using standard bioequivalence limits of 80%–125% and assuming a true difference of 0 [16,33,34].
The primary outcome was to determine whether Innovator (Prograf), Generic Hi (Sandoz), and Generic Lo (Dr. Reddy) tacrolimus products were bioequivalent with each other by comparing their AUC and Cmax using conventional ABE limits (the 90% CIs of the ratio of geometric means of the 2 products for Cmax and AUC were within the range of 80%–125%) [34] and reference SCABE limits [16]. The observed Cmin represented the minimum concentration and was analyzed in lieu of C0 or C12. Observed Tmax represented the actual time at which the maximum concentration was measured. Each patient served as his or her own control; therefore, dose normalization was not required. The investigators and the FDA independently analyzed the primary end-point data using WinNonlin software (version 6.4. Phoenix, Certara, Princeton, New Jersey, US) and SAS (version 9.3, SAS Institute, Cary, North Carolina), respectively. The investigators’ analysis results are reported.
Secondary outcomes included ABE and SCABE assessment in prespecified subgroups and assessments for safety and efficacy. Subgroups in which there were at least 6 recipients and the statistical model converged are reported. The prespecified subgroups included recipient sex; age; African-American race; diabetes; CYP3A4/5, POR 28, and ABCB1 polymorphisms; and donor CYP3A5, as these subgroups are reported to strongly influence tacrolimus concentrations [18]. The study was not powered to identify differences by subgroup. In addition, pharmacokinetic parameters of the primary tacrolimus metabolite, 13-O-desmethyl tacrolimus, were compared.
Safety was assessed at baseline and weekly throughout the study by complete metabolic panels and complete blood cell count with differential. Baseline and weekly assessments included markers of transplant function in individuals with a kidney transplant (creatinine) or liver transplant (aspartate aminotransferase [AST], alanine aminotransferase [ALT], and alkaline phosphatase). To compare kidney and liver function tests among products, a mixed effects model was run with a term designating the product received in each period and a random subject term and using the Tukey’s honest significant difference (HSD) test for multiple comparisons, where needed [35]. The study was not powered to assess differences in transplant organ function. Total daily tacrolimus dose data were summarized using means and standard deviation, and between-group differences were analyzed using the t test. Reports of adverse events were collected at each visit and coded utilizing the Common Terminology Criteria for Adverse Events (CTCAE v4.0) [36]. The Exact McNemar test for paired categorical data with adjustments for multiple comparisons was used to compare adverse event rates among the products [37]. No statistically significant differences among any pairs of products were found for any adverse event code or for adverse events overall.
ABE was assessed within each transplant organ group (i.e., kidney and liver) by using a mixed effects analysis of variance model for a 6-period crossover design with the loge-transformed pharmacokinetic parameter estimates (AUC, Cmax, and Cmin) as the response variable. Fixed effect terms in the model included sequence, period, and treatment. Random effects included subject nested within sequence. The error variance structure accounted for the repeated measures of treatments within each subject. Two-sided 90% confidence intervals using the differences in least square means and the appropriate error terms from the model were calculated for each pairwise assessment of bioequivalence. The estimates and end points of the confidence intervals were back-transformed to obtain the ratios of the parameters being assessed for bioequivalence and the corresponding 90% confidence interval for the ratios. If the entire confidence interval was contained within the range of 80% to 125%, then ABE was established.
To assess SCABE, the estimate of within-subject variability for each treatment was obtained by using a mixed effects model within each organ type and treatment group. Fixed effect terms included sequence, replication (i.e., first or second), and sequence-by-replication interaction. Random effect terms include subject nested within sequence. This model provided estimates of the within-subject variability for each treatment, and these were then used to adjust the bioequivalence end points to obtain the SCABE limits and calculate the criterion bound in accordance with FDA guidance for NTI drugs [16]. SCABE was concluded if each of the following criteria were met: (1) the 2-sided 90% confidence interval calculated for the ABE assessment must fall entirely within the SCABE limits, (2) the criterion bound must be less than 0, and (3) the upper 90% confidence limit for the ratio of the within-subject variabilities for the 2 treatments being assessed must be less than 2.5 [16].
From December 2013 through October 2014, 42 individuals with a kidney transplant and 40 individuals with a liver transplant were consented and followed as per the study protocol. Seventy-one individuals were analyzable (kidney, n = 35; liver, n = 36). The most frequent causes of noneligibility during screening were (1) a greater than 3-hour drive from the study center, (2) receiving the 0.5-mg tacrolimus dosage form, (3) renal function < 35 ml/min, (4) not receiving tacrolimus, (5) a history of multiorgan transplant (i.e., kidney and pancreas or liver and kidney), (6) documented nonadherence, or (7) a history of cancer. Moreover, individuals with a liver transplant and active hepatitis C were not eligible. A complete list of inclusion and exclusion criteria is in Section F in S1 Appendix. Consolidated Standards Of Reporting Trials (CONSORT) flow diagrams (Fig 2) and checklist (S1 CONSORT Checklist) are provided.
The demographic and baseline characteristics reported to impact the tacrolimus pharmacokinetics of the analyzed individuals are summarized in Table 1 and were similar to intent-to-treat individuals (Table Q in S2 Appendix). Most individuals received tacrolimus, mycophenolate, and corticosteroid-free immunosuppression. Immunosuppressive regimens, including tacrolimus doses, remained constant throughout the 6-week study. Individuals with a kidney or liver transplant received a median mg/day (IQR) tacrolimus dose of 5.0 (4.0–8.0) or 4.0 (3.0–6.0), respectively. During the study, no patient initiated, discontinued, or changed doses of known CYP3A inhibitors or inducers that could impact pharmacokinetic observations.
Adherence was evaluated using the MEMS system to insure the quality of the pharmacokinetic evaluations. Adherence was defined as the degree to which the number of medication doses taken each day matched the number of prescribed doses. Over 6 weeks, MEMS-based adherence was 99.75% (range: 97.67%–100%). Three individuals were excluded from the analysis due to nonadherence (kidney, n = 1; liver, n = 2) (Fig 2). MEMS-based adherence rates are reported in Tables R and S in S2 Appendix.
Tacrolimus 12-hour concentrations following chronic dosing are presented for individuals with a kidney (Fig 3A) or liver transplant (Fig 3B). Two tacrolimus concentration-time curves for each product are depicted, representing the first and second exposures to the product. Values for the mean and the standard deviation (SD) for each pharmacokinetic time point are reported in Table 2. There were no statistically significant differences observed by product between time points as assessed by Kruskal-Wallis [38].
Pharmacokinetic parameters for individuals with a kidney or liver transplant by product are summarized in Table 3. Point estimates of the geometric means with the resulting 2-sided 95% confidence intervals are presented for AUC, Cmax, Cmin, and Tmax by product. Product selection was based upon pharmacokinetic studies in healthy individuals as previously described. The Generic Hi (Sandoz) observed exposure was numerically higher than that of Innovator in both individuals with a kidney transplant and those with a liver transplant. In this study, the point estimate for Generic Lo exposure was numerically higher than that of both Innovator and Generic Hi in both individuals with a kidney transplant and those with a liver transplant. However, this can be expected to be caused by random variability since these products were shown to be bioequivalent. (Table 3)
The primary end point of bioequivalence of these pharmacokinetic parameters was tested using SCABE on log-transformed data. When comparing Innovator with Generic Hi and Generic Lo, and Generic Hi with Generic Lo, AUC, Cmax, and Cmin fell within conventional ABE limits of 80%–125%, as well as within the tighter SCABE acceptance limits. Comparing interindividual variability of systemic tacrolimus exposure (AUC), within-subject variability ranged from 12.11% to 15.81%. Similarly, Cmax within-subject variability ranged from 17.96% to 24.72% for all products. Per the FDA guidance for SCABE testing, all products exhibited similar pharmacokinetic variability since the upper limit of the 90% confidence interval for the ratio of within-subject variability was equal to or less than 2.5 (Tables 4 and 5). SCABE criteria were met for all product comparisons for AUC and Cmax in both individuals with a kidney and those with a liver transplant. In reference to the EMA bioequivalence acceptance range for AUC of 90.00%–111.11%, these limits were met, with the only exception being in the case of Innovator versus Generic Lo. Here the upper limits of the 90% -confidence intervals—111.30% (kidney) and 112.12% (liver)—were slightly above the upper EMA AUC acceptance criterion [14], whereas only Innovator versus Generic Lo in liver transplant recipients was also above the Health Canada AUC bioequivalence acceptance limits for critical dose drugs of 90%–112%[15].
Individual pharmacokinetic tacrolimus time-concentration curves for each pharmacokinetic period by product are presented (Figs D and E in S2 Appendix) [39]. Upon visual inspection of individual pharmacokinetic curves, differences can be observed upon comparison between products and between replicate administration of the same product.
The study was not powered to show a difference by any subgroup analyzed. A majority of individuals with a kidney transplant were genotyped as nonexpressers with CYP3A5 *3/*3 (n = 23), and 12 were expressers carrying *1/*3 (n = 10), and *1/*1(n = 2) variants (Figs F and G in S2 Appendix). Most individuals with a liver transplant were genotyped as nonexpressers (n = 30); however, 6 expressed the *1/*3 variant (Figs H and I in S2 Appendix). Donor samples were available for 17 individuals with a kidney transplant and for 24 with a liver transplant. Most kidney donors were genotyped as nonexpressers (n = 12); however, 5 were expressers with *1/*3 (n = 4) and *1/*1 (n = 1) variants (Figs J and K in S2 Appendix). Most liver individual donors were genotyped as CYP3A5 nonexpressers (n = 14); however, 10 were expressers with *1/*3 (n = 5) and *1/*1 (n = 5) variants (Figs L and M in S2 Appendix).
Individuals with a kidney transplant expressing CYP3A5 required significantly higher tacrolimus doses to achieve target tacrolimus trough blood levels (8.17 ± 2.5 versus 4.26 ± 1.9 mg/day [p = 0.0002]). In individuals with a liver transplant, recipient CYP3A5 genotype had no effect on the tacrolimus doses required to achieve target trough blood levels (4.67 ± 2.2 versus 4.5 ± 1.2 mg/day [p = 0.80]). When assessing the donor variant status, donor CYP3A5 expression had no impact on tacrolimus dose requirements for either individuals with a kidney transplant (6.25 ± 2.1 versus 5.40 ± 4.2 mg/day [p = 0.69]) or those with a liver transplant (4.71 ± 1.9 versus 5.70 ± 2.4 mg/day [p = 0.29]). No differences in within-subject variability were observed by CYP3A5 genotype in individuals with a kidney or liver transplant. No association between AUC and ABCB1 genotype as well as no impact on dosing requirements for individuals with a kidney (ANOVA p = 0.08) or liver transplant (ANOVA p = 0.35) was found (Figs N and O in S2 Appendix for AUCs).
ABE and SCABE limits for AUC were calculated for subgroups to assess for consistency of results (Figs PA–PC in S2 Appendix [kidney] and Figs QA–QC in S2 Appendix [liver]). In general, all FDA ABE limits were met, with most also meeting the stricter SCABE limits; however, most exceeded the EMA upper limit of 111.11%. For subgroups not meeting the ABE or SCABE criteria, the number of observations was 10 or fewer, resulting in low statistical power to conclude ABE or SCABE, except for POR*28. ABE was concluded for both POR*28 carriers and noncarriers for individuals with a kidney or liver transplant. For POR*28 carriers, SCABE was demonstrated except in the case of Generic Lo versus Innovator in individuals with a kidney transplant (n = 17) and Generic Hi versus Generic Lo in individuals with a liver transplant (n = 17). For POR*28 noncarriers, SCABE was demonstrated except for Generic Hi versus Innovator in individuals with a kidney transplant (n = 18) and individuals with a liver transplant (n = 19) and for Generic Lo versus Innovator in individuals with a liver transplant (n = 19). ABE criteria were met for most subgroups, except for individuals with a kidney transplant and a CYP3A4*1B genotype and individuals with a liver transplant and a CYP3A5 *1/*1 or *1/*3 genotype (Figs PA-C in S2 Appendix [kidney], Figs QA–C in S2 Appendix [liver]).
The blood concentrations of the major tacrolimus metabolite, 13-O-desmethyl tacrolimus, were also found to meet FDA ABE and SCABE AUC bioequivalence acceptance criteria in individuals with a kidney or liver transplant, but they failed EMA AUC acceptance limits (Tables T–U in S2 Appendix).
One serious adverse event of pyelonephritis was reported in an individual with a kidney transplant, resulting in hospitalization prior to study drug administration. This event resolved with treatment, and the individual was withdrawn. Fifty-two percent and 65% of all individuals with a kidney or liver transplant, respectively, reported an adverse event. The adverse events are sorted by formulation and by individuals with a kidney or liver transplant and CTCAE disorder classification. The Exact McNemar test for paired categorical data with adjustments for multiple comparisons was used to compare adverse event rates among the products. No statistically significant differences among any pairs of products were found for any adverse event code or for adverse events overall (Table V in S2 Appendix).
No acute rejections occurred during the study period of 6 weeks. Baseline and weekly assessments included markers of transplant function in individuals with a kidney (creatinine) or liver (AST, ALT, and alkaline phosphatase) transplant. To compare kidney and liver function tests among products, a mixed effects model was run with a term designating the product received in each period and a random subject term and using the Tukey HSD test for multiple comparisons, where needed. No statistically significant differences were found among products (Fig R–W in S2 Appendix).
Public concerns remain regarding generic tacrolimus use in individuals with a kidney or liver transplant despite the significant market penetration of generic tacrolimus in the US. Historically, concerns were generated by a lack of definitive clinical evidence with properly controlled trials in target populations [13]. Limitations of previous studies [40–45] include retrospective evaluations, case reports, poor study design (underpowered or without appropriate controls), analysis of trough concentrations only, lack of analysis of confounders such as comedications and comorbidities, incorrect pharmacokinetic analysis, and use of nonspecific immunoassays in which metabolites may interfere with tacrolimus concentration measurements, thus leading to considerable bias and limited conclusions [13].
The present randomized, prospective, 3-treatment, 6-period, crossover, replicate-dosing study in stable individuals with a kidney or liver transplant systematically addresses the aforementioned public concerns regarding generic tacrolimus. The replicate dosing study design allowed the analysis of tacrolimus products using the tighter SCABE standards required by the FDA for NTI drugs. Clinically, the present study represents a scenario in which an individual with a kidney or liver transplant is randomly switched between 3 tacrolimus products every week for 6 weeks. The pharmacokinetic parameters (AUC and Cmax) demonstrated bioequivalence by SCABE criteria, implying similar tacrolimus exposure is achieved when individuals with a kidney or liver transplant are switched between these tacrolimus products. Although not required for bioequivalence testing, Cmin also met the SCABE criteria.
These results support a previous prospective, multicenter, open-label, randomized, 2-period, crossover, pharmacokinetic study comparing twice-daily generic tacrolimus (Sandoz) versus reference tacrolimus (Prograf) in stable kidney transplant recipients [30]. In 68 kidney transplant recipients, there were no significant differences in AUC C0, Cmax, or Tmax between the generic and reference products, resulting in ratios of the geometric mean and 90% CI for AUC and Cmax that were reported as 102% (97%–108%) and 109% (101%–118%), respectively [46]. Post hoc analysis revealed the products also met SCABE acceptance criteria [47]. In contrast to the present study, the latter did not include comparisons to other generic tacrolimus products or the comparison of 2 generics, and liver transplant recipients and relevant genetic polymorphisms were not analyzed.
Pharmacogenomic profiling of individuals was performed, specifically, genotyping of CYP3A5 polymorphisms to identify the “poor absorber” [20,23]. The genetic polymorphisms most important for tacrolimus pharmacokinetics were assessed and, in general, had no effect on bioequivalence. The only exceptions were that bioequivalence of tacrolimus AUC was not found for individuals with a kidney transplant expressing CYP3A4*1B and individuals with a liver transplant expressing CYP3A5, but this analysis was underpowered for said subgroups. Moreover, an influence of POR*28 polymorphism on bioequivalence using SCABE metrics could not be excluded, although ABE criteria were met. Dosing differences by genotype were similar across products [48]. The major tacrolimus metabolite concentrations were bioequivalent for AUC. In contrast to tacrolimus, 13-O-desmethy tacrolimus is not directly administered but formed from tacrolimus, mostly by intestinal and liver cytochrome P4503A enzymes. Thus, its formation is greatly influenced by the aforementioned polymorphisms and its pharmacokinetics more variable than that of tacrolimus.
In this study, the 2 tacrolimus generic products met US FDA SCABE criteria when compared to the innovator product and with each other in individuals after a kidney or liver transplant. However, when applying the more rigid EMA criteria, the Generic Lo product failed AUC testing when compared to Innovator in individuals receiving a kidney or liver transplant. The EMA requires a narrow 90% confidence interval contained within the confidence limits of 90.00%–111.11% acceptance criterion for AUC, but not for Cmax, for which the usual 80.00%–125.00% acceptance limit applies [14]. As shown in Table 4, in general these acceptance criteria were met, with the only exception being that of Innovator versus Generic Lo, for which the upper limits of the 90% confidence intervals—111.30% (kidney) and 112.11% (liver)—were slightly above the upper EMA acceptance criterion for AUC. The conflicting approval guidelines lead to different interpretations of the bioequivalence data of the same study. In this context, it should be considered that the EMA and Health Canada bioequivalence limits for NTI drugs were set with single-dose healthy volunteer studies in mind, a population that is inherently less variable than individuals receiving a transplant after multiple doses [4]. While the FDA SCABE limits adjust based on the pharmacokinetic variability of the innovator in the study population, the EMA and Health Canada bioequivalence limits for NTI drugs are fixed, which explains the different conclusions when FDA, EMA, or Health Canada limits are employed. This study highlights the need for global harmonization of bioequivalence approval standards of NTI drugs to prevent different interpretations of bioequivalence study results.
Individual pharmacokinetic time concentration curves for all products are provided in Section C in S2 Appendix. These data visually depict the intraindividual variability that can be observed within the same product despite administering the same lot at the same dose in a controlled setting. Such observed intraindividual variability led the FDA to require repeat crossover study designs for NTI drugs, which requires that each individual receives each of the tested products twice to assess and to compare intraindividual variability between tacrolimus innovator and generic products. Upon visual inspection of individual pharmacokinetic curves, differences can be observed between products and between replicate administration of the same product.
Safety was similar across products over the observation period of 6 weeks. Although we did not evaluate the long-term impact of generic tacrolimus on acute rejection and graft survival, this study evaluates pharmacokinetic parameters as a surrogate for safety exposure.
Several design elements strengthen the findings, such as validated, specific, and sensitive high-performance liquid chromatography (HPLC)-tandem mass spectrometry analysis of tacrolimus and metabolites, quality control of the tacrolimus test batches, independent parallel analysis by the study team and the FDA, genotypic analysis of individuals donating or receiving a kidney or liver transplant, and close adherence monitoring using a combination of diaries, MEMS caps, and pill counts. Finally, this study was adequately powered to assess bioequivalence in both kidney and liver transplant recipients. However, our study also had certain limitations. A differential carryover or sequence effect cannot be fully excluded, even though we did not detect any statistically significant sequence effects in statistical modeling. When considering the half-life of immediate-release tacrolimus and the length of the treatment period, this effect is unlikely to occur. The pharmacokinetic profiling occurred in strictly controlled conditions with recipients who were highly adherent, which did not allow the evaluation about the impact of nonadherence or the impact of feeding on possible pharmacokinetic differences between products. The study design also did not allow for assessment of the potential impact of differing appearances of the 3 tacrolimus products on adherence. The safety data should be interpreted cautiously in the light of the small number of participants and short observation periods. Lastly, only the 1 mg tacrolimus dosage strength was utilized, limiting generalizability to the 0.5 mg and 5 mg dosage strengths; however, the 1 mg capsule is the most common clinically utilized dosage strength.
The present study was specifically designed to address lingering concerns in the transplant community [8–11]. While typical single-dose healthy volunteer bioequivalence studies mainly assess prescribability, our study in steady-state, stable individuals with a kidney or liver transplant mainly assessed the more important switchability between innovator and generics and between generics [5,6]. The present study suggests that tacrolimus and the tested generic products in healthy volunteers were also bioequivalent in individuals with a kidney or liver transplant. Moreover, the generics were bioequivalent to each other. Even the tighter FDA SCABE criteria were met, and there was no difference between the different tacrolimus products in terms of within-subject variability.
|
10.1371/journal.pntd.0003043 | An Integrated Lab-on-Chip for Rapid Identification and Simultaneous Differentiation of Tropical Pathogens | Tropical pathogens often cause febrile illnesses in humans and are responsible for considerable morbidity and mortality. The similarities in clinical symptoms provoked by these pathogens make diagnosis difficult. Thus, early, rapid and accurate diagnosis will be crucial in patient management and in the control of these diseases. In this study, a microfluidic lab-on-chip integrating multiplex molecular amplification and DNA microarray hybridization was developed for simultaneous detection and species differentiation of 26 globally important tropical pathogens. The analytical performance of the lab-on-chip for each pathogen ranged from 102 to 103 DNA or RNA copies. Assay performance was further verified with human whole blood spiked with Plasmodium falciparum and Chikungunya virus that yielded a range of detection from 200 to 4×105 parasites, and from 250 to 4×107 PFU respectively. This lab-on-chip was subsequently assessed and evaluated using 170 retrospective patient specimens in Singapore and Thailand. The lab-on-chip had a detection sensitivity of 83.1% and a specificity of 100% for P. falciparum; a sensitivity of 91.3% and a specificity of 99.3% for P. vivax; a positive 90.0% agreement and a specificity of 100% for Chikungunya virus; and a positive 85.0% agreement and a specificity of 100% for Dengue virus serotype 3 with reference methods conducted on the samples. Results suggested the practicality of an amplification microarray-based approach in a field setting for high-throughput detection and identification of tropical pathogens.
| Tropical diseases consist of a group of debilitating and fatal infections that occur primarily in rural and urban settings of tropical and subtropical countries. While the primary indices of an infection are mostly the presentation of clinical signs and symptoms, outcomes due to an infection with tropical pathogens are often unspecific. Accurate diagnosis is crucial for timely intervention, appropriate and adequate treatments, and patient management to prevent development of sequelae and transmission. Although, multiplex assays are available for the simultaneous detection of tropical pathogens, they are generally of low throughput. Performing parallel assays to cover the detection for a comprehensive scope of tropical infections that include protozoan, bacterial and viral infections is undoubtedly labor-intensive and time consuming. We present an integrated lab-on-chip using microfluidics technology coupled with reverse transcription (RT), PCR amplification, and microarray hybridization for the simultaneous identification and differentiation of 26 tropical pathogens that cause 14 globally important tropical diseases. Such diagnostics capacity would facilitate evidence-based management of patients, improve the specificity of treatment and, in some cases, even allow contact tracing and other disease-control measures.
| Many infectious diseases are more prevalent in the tropical and subtropical regions where ecological, geographical and socioeconomic factors facilitate their propagation. The high diversity of such tropical pathogens include bacteria, fungi, helminths, parasites, and viruses that mirrors the rich biodiversity in the tropics and sub-tropical regions [1]–[3]. Many of these pathogens are transmissible through an insect vector or an invertebrate host [4]–[7], and transmission is affected by climate that can significantly influence vector behavior and physiology [8], including the extrinsic incubation period of vector-borne pathogens [9], [10]. Furthermore, global changes such as anthropogenic climate change and climate variability, habitat encroachment by the growing human population, volume of international travel, migration, trade and pollution create new opportunities for microbial spread [11]–[13].
The world is subjected to a plethora of tropical pathogens. Table 1 provides an overview of 14 tropical diseases, stratified into protozoan, bacterial, and viral infections that are globally important. However, some of these tropical diseases are often intimately connected to paucity of local and global burden estimates, poverty, geographical isolation and lack of coordinated approaches for disease controls [14]. Firstly, there are protozoan infections: malaria, which remains one of the most devastating and difficult parasitic diseases to be controlled and further threatened by the emergence and spread of resistance to anti-malarial drugs [15]–[17]; Chagas disease which is one of the most neglected tropical disease with a lifelong infection [18]–[20]; and human African trypanosomiasis with 60 million people at risk in Africa [21]–[23]. Next are bacterial infections: leptospirosis, which has been identified as one of the most widespread zoonosis in the world, exemplified by outbreaks in rural and urban environments [24]–[27], and more recently, emerged as a disease of the adventure traveler [28]; meliodosis that has been reported with a global distribution [29], [30]; and salmonellosis, which causes enteric fever and has a high global incidence [31]. Finally, the most prevalent infections are those of viral origins: Chikungunya fever in the Indian Ocean islands, the Indian subcontinent, southeast Asia, Africa, Europe and its emergence in the Americas [32]–[37]; Dengue fever including the emergence of dengue hemorrhagic fever [38]–[43]; West Nile fever in America [44], [45] and the increasing extensive distribution through Africa, Middle East, Europe and Asia [46]; Japanese encephalitis in Australasia [47] and in Asia [48]; yellow fever in West and Central Africa [49]; high incidence rates of hand, food and mouth disease in Asia [50]–[53]; Rift valley fever which has spread to Yemen, Saudi Arabia, northern Egypt and the French island of Mayotte [54]; and Hantavirus hemorrhagic fever which can cause serious diseases in humans with mortality rates of 12% (hemorrhagic fever with renal syndrome) and 60% (Hantavirus pulmonary syndrome) in some outbreaks [4], [55]. Despite being medically important, the incidence rates of some of these diseases are grossly underestimated and this reflects the clinical index of suspicion of the diseases which could have resulted from a lack of access to rapid diagnostics [18], [25], [29].
The global spread of tropical diseases emphasizes the importance of preparedness to address them. The first goal of this preparedness is fast and accurate diagnosis of medically important diseases. Differential diagnosis is based mainly on clinical examination, taking into account which diseases are locally prevalent, potential exposure, and the relevant travel history. However, the similarity and the non-specific nature of the symptoms provoked by many tropical pathogens (Table 1) complicates correct diagnosis by classical clinical observations [25], [56]–[61]. Yet, a correct diagnosis is necessary to institute effective control measures, from timely therapeutic intervention [62], [63], to effective treatment [64] and effective clinical management in deploying appropriate community-wide control measures to improve the patients' clinical outcome, disease mapping, impact monitoring, and post-elimination surveillance. Correct diagnosis can only be determined through reliable laboratory-confirmed detection and identification of tropical pathogens in clinical specimens.
Polymerase chain reaction (PCR) has been used in the diagnosis of several infectious diseases [51], [65]–[70] as it is a highly specific and sensitive method for molecular detection [71]–[73]. Moreover, much progress has been made with molecular multiplexing [74]–[78]. With the advent of microarray technology which permits simultaneous detection of a given sequence in a sample by hybridization to thousands of defined probes [79], amplification and microarray integrated assays have been made possible [74], [80]–[82].
In this study, microfluidic technology was combined with reverse transcription (RT), PCR amplification, and microarray hybridization to develop a silicon based micro electro mechanical systems (MEMS) integrated lab-on-chip that can simultaneously detect and differentiate between 26 pathogen species (including bacteria, parasites and viruses) that cause 14 tropical diseases. The detection platform is composed of the disposable lab-on-chip, a temperature control system (TCS) for the accurate control of thermal process and an optical reader for the fluorescence microarray image acquisition. The ability of the lab-on-chip to provide a “blood-to-diagnosis” solution in the detection of known and divergent pathogens was demonstrated on retrospective patient specimens. This system allows the simultaneous identification and discrimination of a large number of candidate tropical pathogens. It is undoubtedly a potential game-changer in the field of molecular diagnostics, as it provides an effective and rapid means to establish the presence of defined potential pathogens.
The use of human samples was approved by the National Healthcare Group's Domain-Specific Ethics Review Board (DSRB reference no. B/08/026), and written informed consent was obtained from all participants. Approval was also obtained for the use of archived samples from The Oxford Tropical Research Ethics Committee (OxTREC) as part of the surveillance routine.
Plasma samples from 30 PCR-confirmed Chikungunya virus (CHIKV) patients who were admitted to the Communicable Disease Centre at Tan Tock Seng Hospital (TTSH) during the outbreak from August 1 to September 23, 2008 [83], [84] were included. Plasma samples were also collated from 10 healthy donors with informed consent (DSRB reference no. B/08/026) and used as negative controls. RNA samples were extracted using the QIAamp viral RNA mini kit (Qiagen, Hilden, Germany), according to manufacturer's instructions.
One hundred and twenty five archived nuclei acid samples extracted from specimens at the Shoklo Malaria Research Unit (SMRU) clinic on the Thai-Burmese border between 1999 and 2011 as part of two surveillance studies [16], [85] were included. DNA extracts from packed red blood cells obtained from patients (refugees and migrants) with a clear malaria diagnosis (part of the malaria burden observational study) were tested with the lab-on-chip assay. The sensitivity and specificity of the chip assay was then determined against microscopy diagnosis used by the Thailand clinics [16]. Non-malaria specimens collected from patients presenting with undifferentiated febrile illness were also evaluated with the lab-on-chip. Viral RNA extracted from acute plasma specimens that had been stored at −80°C since 2008 were used. These had previously been tested with a range of tests including Dengue RT-PCR [85]. In addition, whole blood samples from 5 native healthy volunteers were extracted using the DNeasy blood and tissue kit and QIAamp viral RNA mini kit (Qiagen) and used as negative controls.
Cultures of the 3D7 clone of the NF54 strain of Plasmodium falciparum (P. falciparum) were performed using sealable flasks with RPMI-HEPES medium at pH 7.4, supplemented with 50 mg/mL hypoxanthine, 25 mM NaHCO3, 2.5 mg/mL gentamicin, and 0.5% (weight/volume) Albumax II (Gibco, Singapore) in an atmosphere containing 5% CO2, as previously described [86], [87].
The CHIKV isolate used in this study was originally isolated from a French patient returning from Reunion Island during the 2006 CHIKF outbreak [88]. After passages in Vero-E6 cultures, virus stocks were washed, and precleared by centrifugation before storing at −80°C. All virus stocks were titered by plaque assay and quantified by quantitative RT-PCR (qRT-PCR) as previously described [89], [90].
Target gene sequences of each pathogen (Table S1 in Text S1) were first obtained from Genbank database. Sequence alignments were performed using the ClustalW algorithm [91] in MegAlign (DNAStar, Inc., Madison, WI). A consensus sequence representing clinically relevant strains (Table S1 in Text S1) was created for each pathogen. Each target oligonucleotide sequence was designed through multiple, successive steps of evaluation of candidate sequences, based on user-defined criteria, followed by analysis with Basic Local Alignment Tool (BLAST) [92] against the nucleotide sequence database (nr/nt) [93] for non-target genomes potentially present in the specimen that could cause interference. Genus-specific PCR primers were designed for all chosen target genes sequences as previously described [94], [95]. Genus-specific (for Plasmodium, Flaviviruses and Hantaviruses) and species-specific capture probes were selected to target 2 to 4 regions of the targeted gene to confirm specificity and to overcome the problem of poor hybridization within the amplicon as a result of strain-specific gene polymorphisms. Efforts to improve specificity included the design of short length capture probes of 20 to 30 nucleotides in line with other studies which have shown that shorter length probes showed higher specificity [96].
For each pathogen, a PCR product encompassing the targeted region was prepared using the consensus sequence and cloned into the T7 polymerase expression vector pGEMT-easy (Promega, Madison, WI) as described [70]. Serial diluted plasmid DNA or in vitro-transcribed RNA from respective quantified stocks was used as the DNA copy number control for DNA pathogens or RNA copy number control for RNA pathogens.
The lab-on-chip was manufactured on a silicon wafer based on MEMS and mounted on a printed circuit board (PCB) support that provides mechanical, thermal, and electrical connection [94], [95], [97] (Figure S1 in Text S1). It encompassed two silicon microreactors (12 µL) connected to a microarray chamber. The microarray chamber (3.5 mm×9.0 mm) contains 126 spots consisting of duplicate oligo-probes spotted onto the surface through a piezo-array system [95] to ensure that differential signals do not occur by chance. The enzymatic thermal cycling and hybridization reactions on the lab-on-chip are performed by the electronic TCS. Tropical pathogen detection was split into two chip versions to be subjected to two different multiplex reactions; DNA chip with a customized microarray layout specific for DNA pathogens and RNA chip specialized for RNA pathogen detection.
PCR was performed on a DNA chip in a constituted reaction of 200 nM forward and 500 nM Cy5-conjugated reverse primers in 23 µL final volume using the QuantiTect multiplex RT-PCR NoROX kit (Qiagen). Amplification was carried out with initial denaturation at 90°C for 15 min, followed by 45 cycles of 95°C for 15 sec, 60°C for 40 sec, and 72°C for 30 sec, then final extension at 72°C for 60 sec.
RT-PCR was carried out on the RNA chip using SuperScript III one-step RT-PCR system with platinum Taq (Life Technologies) in a 23 µL reaction volume containing a concentration of forward and Cy5-conjugated reverse primers in the range of 200 nM to 700 nM. After a 20-min reverse transcription step at 50°C, enzyme activation was initiated at 95°C for 120 sec, followed by denaturation at 95°C for 10 sec. Amplification was performed in a manner of touch down PCR to enhance the specificity of the initial primer-template duplex formation and hence specificity of the final PCR product [98]. The annealing temperature in the initial cycle was initiated at 60°C (5°C above the average melting temperature of the primers for RNA pathogen detection). In the subsequent 10 cycles, the annealing temperature was decreased in steps of 1°C/cycle until a temperature was reached to 50°C, and followed by extension at 72°C for 50 sec. Following these 10 cycles, 40 cycles with a temperature of 95°C for 15 sec, annealing temperature of 56°C for 40 sec, and then a final extension for 50 sec at 72°C completed the program.
Upon completion of PCR or RT-PCR, denaturation of amplicons proceeded at 95°C for 3 min, followed by hybridization at 58°C for 30 min. The lab-on-chip was washed and spin dried. The dried chip was scanned in the optical reader [95] (Veredus Laboratories) which has an excitation filter for Cy5. Accompanied software analysis was based on hybridization of amplicons to target-specific capture probes with the highest signals expected from a perfect match. Spot segmentation and intensity calculation of the microarray image was performed by overlaying a virtual grid over the microarray image using the corner features as reference points.
For positive detection of Plasmodium parasites, Flaviviruses and Hantavirus, at least 1 out of 2 genus-specific probes must give a positive signal to indicate the presence of the respective genera, and at least 50% of species-specific probes must hybridize for species differentiation (Table S1 in Text S1). For the rest of the pathogens, at least 2 out of 3 pathogen-specific probes must give a positive signal for a positive detection of the pathogen (Table S1 in Text S1).
The detection threshold and specificity of the lab-on-chip assay was evaluated by using 4 µL of quantitative standards (to cover a range of 101 to 104 copies per chip for each pathogen) and assessing the signal intensity and presence of cross hybridization at each copy number. Triplicates were run to ensure intra-experimental reproducibility. The lowest titer (DNA or RNA copies per chip) with 2 or more out of 3 chips positive for the assayed pathogen was further expanded to another 21 replicate runs to confirm the LoD which was the indicated titer that would yield more than 95% positive detection, as well as to evaluate inter-assay reproducibility.
Sorted P. falciparum parasites were serial diluted in phosphate-buffered saline (PBS) and added to whole blood to obtain spiked samples with final concentrations of 1 to 103 parasites/µL [87]. In parallel, CHIKV virus stock was serial diluted before spiking into aliquots of whole blood to cover 1 to 105 PFU/µL. Spiked experiments were repeated twice for inter-experimental reproducibility. Sensitivity of the chip assay was compared with that of nested PCR [99] or qRT-PCR [70] respectively. The volume of the isolated nuclei acid subsequently used in for all comparison assays was kept constant at 4 µL.
All statistical analyses were performed using Prism 6.03 (GraphPad Software, Inc., La Jolla, CA). Lab-on-chip outcome on previously laboratory-confirmed samples was analyzed using Fisher exact test. P values less than 0.05 were considered statistically significant.
The objective of developing a portable microfluidic integrated lab-on-chip (Figure S1 in Text S1) was to provide a seamless one-time screening test for multiple tropical pathogens that exhibit similar or non-specific symptoms. Twenty-six pathogen species that cause 14 globally important but yet neglected tropical diseases (Table 1 and Table S1 in Text S1) were considered for the panel. A typical workflow for the detection of these pathogens was defined. It comprises of a processing step (blue) that includes the sample extraction and reaction setup. This is then followed by the on-chip identification and differentiation (red) (Figure 1) to ensure accurate implementation of the assay. Microarray spots were simultaneously assessed to calculate differences in signal intensities, thereby identifying unique patterns (Figure S2 in Text S1). Hybridization to a series of target-specific probe sets provided presence/absence information for the tropical pathogen, while also revealing the species of the causative agent (Figures S2, S3 in Text S1).
The rationale of the analytical evaluation of the lab-on-chip was to define the LoD of the assay for all the pathogens. LoD of the lab-on-chip was determined as the lowest copy number which, in terms of plasmid copy for DNA or RNA transcript copy for RNA, when added to the chip, led to more than 95% positive pathogen identification outcome. Table S1 in Text S1 shows the lowest detectable dilution for each pathogen. The results revealed an individual sensitivity that ranged from 102 to 103 copies per chip (Figure 2). Target-specific hybridization signal saturation was observed at concentrations as low as 104 copies for all the pathogens (Figure 2). Notably, a highly sensitive detection range of 3 orders of magnitude between LoD and signal saturation was achieved for most of the tropical pathogens, mainly S. enterica, T. brucei and T. cruzi under the DNA pathogen category, together with RNA viruses such as West Nile virus, yellow fever virus, Enterovirus 71 and rift valley virus (Figure 2). Although a narrow detection range of 10 copies was observed for Hantaviruses with LoD at 103 copies, the rest of the pathogens stayed within the broad detection range of approximately 2 orders of magnitude. It should be noted that to date, cases of Hantavirus infections in patients yielded very low or non-detectable viral load levels [55], [100]. Probe specificity evaluation showed no significant cross reactivity (Figures S2 and S3 in Text S1).
The efficiency of a detection assay is often dependent on the efficiency of the nuclei acid extraction method from clinical specimens [101], [102]. Some methods may even interfere with the PCR reaction [103], [104]. The purpose of the investigation was to assess the efficiency of the extraction method and the sensitivity of the lab-on-chip using P. falciparum and CHIKV as targets.
The read-out for the lab-on-chip and that of nested PCR is illustrated in Table 2 and in Figure 3. The presence of P. falciparum in the extracted spiked samples was demonstrated by the presence of hybridized genus-specific and species-specific probes on the microarray for lab-on-chip, while that by nested PCR relied on the presence a PCR band on agarose gel [99]. Positive detection of P. falciparum by the lab-on-chip was observed at 100 parasites, while positive bands were detected at 5 parasites by nested PCR (Table 2 and Figure 3). Although the nested PCR method [99] is more sensitive with a difference of more than one log when compared to the lab-on-chip (Table 2 and Figure 3), it is more labor intensive.
The estimated PFU isolated from CHIKV-spiked samples (in red) compared to the viral load derived from qRT-PCR is shown in Table 3 and in Figure 4. The detection threshold for CHIKV was 50 PFU (Figure 4B, 4C). More importantly, the sensitivity of the detection range of the lab-on-chip and viral load quantification by qRT-PCR are similar, clearly demonstrating the superiority of the lab-on-chip (Figure 4).
In order to assess the clinical performance of the assay, the lab-on-chip was evaluated on retrospective clinical specimens to compare its diagnostic capability with reference methods. The screening and order of diagnostic testing of 170 samples received in Singapore and Thailand are illustrated in Figure 5. Sixty-four out of 77 P. falciparum positive samples and 21 out of 23 P. vivax positive samples were concordant with the microscopic diagnosis (Tables 4, 5). The sensitivity and the specificity for the detection of P. falciparum was 83.1% (72.9% to 90.7%) and 100% (96.1% to 100%) (Table 4, Figure 6), and that of P. vivax was 91.3% (71.9% to 98.9%) and 99.3% (96.3% to 99.9%) (Table 5, Figure 6). Fourteen P. falciparum positive samples with low levels of parasitemia did not yield a positive detection for P. falciparum, but 11 out of the 14 were tested positive for Plasmodium. Although species differentiation was not achieved with these 11 samples, the assay did provide a diagnosis for Plasmodium.
The validation also yielded a good positive 90.0% agreement (73.5% to 97.9%) and excellent specificity 100% (97.4% to 100%) for the CHIKV detection (Table 6, Figure 6). Finally, the assay showed an average positive 85% agreement (62.1% to 96.8%) (17 out of 20 DENV positive samples) and a specificity of 100% (97.5% to 100%) for DENV 3 detection (Table 7, Figure 6). The 3 CHIKV samples that were not detected positive by the lab-on-chip were that with low viral load of less than 102 viral copies/µL quantified by qRT-PCR [70]. All healthy donor samples tested were negative.
While every disease presents specific diagnostic challenges, clinical needs associated with specificity, sensitivity, total analysis time, and implementation would eventually impact the design and development of the diagnostic method. In this study, an integrated strategy for miniaturizing and simplifying complex laboratory assays for the detection of 14 globally important tropical diseases stood out favorably in terms of seamless implementation and pathogen coverage compared to conventional laboratory diagnostic methodologies.
The mainstay to detect protozoan infections such as Chagas disease, human African trypanosomiasis, and malaria infection relies in the conclusive visualization of the parasites in blood [18],[21],[105]. The reliable identification of these infections requires high quality training in specimen preparation and a competency in identifying the parasites when compared to the facile interpretation of the lab-on-chip microarray analysis.
Bacteria culture remains as one of the most effective procedures in identifying bacterial infections [106]–[108] and is also crucial in generating pools of clinical strains for pathogenesis studies. However, the process is labor and time intensive, spanning from a few days to several weeks when compared to the lab-on-chip assay that is completed within 4 hours. It is also dependent on stringent transport conditions and well-maintained equipments to maintain specimen viability.
While methods based on serological reactivity to pathogen-specific antibodies [109]–[111] have been developed to identify several viral infections and are useful in differentiating viruses within the same family or genus, cross reactivity remains a conflicting issue [100], [112], [113]. In spite of cross reactivity issues, serology is still widely used to confirm diagnosis due to limitations in the detection window of nucleic acids [83], [85], [100]. Here, the analytical performance of the lab-on-chip has highlighted its specificity with no cross reactivity observed between the 5 Plasmodium species, between DENV and the other 3 Flaviviruses, and among the 6 Hantaviruses, achieved in just one test. Future iterations of the lab-on-chip could include protein-based arrays as additional serology screens [114], [115] for some diseases that are clinically warranted as orthogonal diagnosis based on nucleic acid, protein, and other biomarkers will be where the field is heading.
Simultaneous laboratory screening of a clinical specimen from a patient with unspecific symptoms for as many tropical agents as possible would either lead to pathogen identification or narrow down the possible causes through elimination. However, combining the various assays for parallel screening of tropical diseases is not a feasible approach given the high diversity of the protocols with many limitations associated with each pathogen. Even though amplification microarray assays [80]–[82] have been developed to circumvent the need for parallel tests, detection in these assays was restricted to one virus family, despite an improvement in pathogen coverage, and thus still considered as low throughput. Moreover, simultaneous detection was achieved only after 3 separate amplification reactions for the 3 respective virus families [80]. Miniaturized total analysis systems [116] have evolved, that has led to miniaturized PCR devices being extensively studied [117]. A few reports have demonstrated rapid on-chip detection of Influenza A virus [118], [119] and human immunodeficiency virus [120], however the development of a miniaturized assay for the detection of multiple tropical diseases pathogens including the validation on patient specimens has yet to be demonstrated.
The design and process of the lab-on-chip evaluation was approached systematically. It was first evaluated using quantitative standards. The LoD of the lab-on-chip was shown to range from 102 to 103 copies and signal saturation for target-specific capture probes' hybridization was at 104 copies. This observation was crucial as the efficiency of the chip to detect the relevant pathogen in a clinical sample load on the chip containing 104 or more copies of that pathogen would be 100%. When considering the detection limit of the lab-on-chip of the pathogen in a clinical sample, the target concentration required to get the minimum amount of nuclei acids after sample extraction in the amplification reaction must be investigated. Comparison of the lab-on-chip with nested PCR using spiked P. falciparum samples and with qRT-PCR on spiked CHIKV samples has proven the efficiency of the extraction method and also emphasized a more superior trade-off between the assay's sensitivity and its utility in the systemic differentiation of P. falciparum and detection of CHIKV. The lab-on-chip assay's ability to detect CHIKV at 50 PFU/µL demonstrated high clinical relevance as it was shown that the mean CHIKV viral load in patients ranged between 126 to 241 PFU/µL [83].
One of the key objectives of the clinical validation was to investigate the lab-on-chip's performance and acceptability in field settings and the degree to which the results would determine the quality of the diagnosis for surveillance and patient management to improve health outcomes. The clinical validation of P. vivax offered a sensitivity that was equivalent to microscopy. Although there was a proportion of P. falciparium samples (14 out of 77 samples) with low parasitiamia that were not positively detected for P. falciparum on the lab-on-chip, the assay did manage to give a partial diagnosis (of the samples being Plasmodium positive) for 11 of these samples. Although the lab-on-chip did not positively differentiate samples with extremely low levels of parasitemia, the low parasite burden of these patients could represent the early stages of malaria. Taken together, the analytical performance of the lab-on-chip for P. falciparum and P. vivax in the range of 102 copies, and the demonstration of its diagnostic utility using spiked samples and clinical specimens showed the applicability of the assay for Plasmodium detection.
The clinical performance of the lab-on-chip for DENV and CHIKV was comparable to RT-PCR. For DENV, comparisons among the diagnostic tests at SMRU have demonstrated RT-PCR to have the best operating characteristics (sensitivity 89%, specificity 96%, positive predictive value 94%, negative predictive value 92%) [85]. This suggested that the chip would be potentially sufficient to function as a single assay for confirmation of Dengue infection, since it allowed for accurate confirmation. Similarly, the assay sensitivity for CHIKV was on par with that of RT-PCR, and achieved a positive 90% agreement with patients' samples.
The cost of the assay compared to that of single assays is high. Advancements in the integration of the lab-on-chip with nuclei extraction capabilities [95] and a higher density microarray with reduced chip cost would provide a more cost-effective comprehensive coverage. While the lab-on-chip assay has showed that miniaturized multiplex PCR could reach the desired clinical sensitivity, future work should attempt to recalibrate the mix of multiplex primers and modify amplification cycling conditions for improved sensitivity. One of the key milestones for lab-on-chip systems would be the direct testing of clinical specimens obtained during the acute infection phase and provide accurate diagnosis to complement clinical assessments.
|
10.1371/journal.pgen.1002516 | MNS1 Is Essential for Spermiogenesis and Motile Ciliary Functions in Mice | During spermiogenesis, haploid round spermatids undergo dramatic cell differentiation and morphogenesis to give rise to mature spermatozoa for fertilization, including nuclear elongation, chromatin remodeling, acrosome formation, and development of flagella. The molecular mechanisms underlining these fundamental processes remain poorly understood. Here, we report that MNS1, a coiled-coil protein of unknown function, is essential for spermiogenesis. We find that MNS1 is expressed in the germ cells in the testes and localizes to sperm flagella in a detergent-resistant manner, indicating that it is an integral component of flagella. MNS1–deficient males are sterile, as they exhibit a sharp reduction in sperm production and the remnant sperm are immotile with abnormal short tails. In MNS1–deficient sperm flagella, the characteristic arrangement of “9+2” microtubules and outer dense fibers are completely disrupted. In addition, MNS1–deficient mice display situs inversus and hydrocephalus. MNS1–deficient tracheal motile cilia lack some outer dynein arms in the axoneme. Moreover, MNS1 monomers interact with each other and are able to form polymers in cultured somatic cells. These results demonstrate that MNS1 is essential for spermiogenesis, the assembly of sperm flagella, and motile ciliary functions.
| Cilia are microtubule-based structures present in virtually all cells in vertebrates. Cilia have diverse functions in development, growth, signaling, and fertilization. Primary ciliary dyskinesia (PCD) affects one in 16,000 individuals. PCD is characterized by bronchiectasis and chronic sinusitis, and is often associated with situs inversus and male infertility. The genetic cause of PCD is heterogeneous. Some cases of PCD in humans and animals are caused by single genic mutations such as mutations in genes encoding microtubule-based dynein arm components. We have characterized a protein called MNS1 and found that it plays an essential role in ciliary functions in mice. MNS1 is a novel and integral component of sperm flagella. Mice lacking MNS1 exhibit male sterility as evidenced by abnormal assembly of sperm flagella. MNS1–deficient mice also display defects in left–right asymmetry patterning of internal organs and hydrocephalus. Therefore, mutations in MNS1 may contribute to male infertility and PCD in humans.
| Spermatogenesis is divided into three phases: mitotic, meiotic and haploid. During the haploid phase (spermiogenesis), spermatids undergo a complex differentiation process to develop into spermatozoa, including chromatin remodeling, nuclear elongation, cytoplasm elimination, acrosome formation, and flagellum development [1]. The sperm flagellum is a complex structure whose integrity is essential for sperm motility and fertilization of the egg [2], [3]. Structural defects in the flagella of sperm from infertile men are responsible for motility abnormalities and may underlie some cases of male infertility in humans [4].
In contrast with flagella, cilia are present in nearly all cell types in vertebrate and play diverse functions [5], [6]. For example, respiratory cilia are important for mucus clearance [7]. Ependymal cilia facilitate cerebrospinal fluid flow [8]–[10]. Nodal cilia are essential for left-right asymmetry patterning during embryogenesis [11], [12]. Flagella and some motile cilia such as respiratory and ependymal cilia consist of nine outer doublet microtubules and one pair of single microtubules in the center (9+2 axoneme). Nodal cilia lack the central pair of microtubules and thus contain a 9+0 axoneme. In the axoneme of motile cilia, dynein arms, ATP-dependent motor proteins, are attached to and projected from outer doublet microtubules. Primary ciliary dyskinesia (PCD) is a heterogeneous group of autosomal recessive disorders characterized by recurrent respiratory infections [13]. Half of PCD patients display situs inversus and PCD is sometimes associated with male infertility. Flagella and cilia are complex structures that consist of >600 proteins [5]. Genetic studies in diverse organisms have begun to uncover the role for some of these proteins in ciliary and flagellar functions [9], [10], .
Mouse MNS1 (meiosis-specific nuclear structural protein 1) was previously identified due to cross-reactivity with an anti-lamin antibody in a study of the perinuclear matrix [20]. In this original investigation, MNS1 appeared to be expressed specifically in testes. The MNS1 protein contains long coiled-coil domains but no other known functional motifs. In that report, it was concluded that MNS1 was specifically expressed in pachytene spermatocytes and thus might function in maintaining proper nuclear morphology during meiosis [20]. However, the physiological function of MNS1 remains unknown. Here we report that, in contrast with the previous study [20], MNS1 is abundantly expressed in post-meiotic spermatids and is required for spermiogenesis. In addition, MNS1 is also required for motile ciliary functions.
To study the expression and localization of MNS1, we generated specific antibodies against the N-terminal and C-terminal fragments of mouse MNS1 respectively. Western blot analysis of adult mouse tissues revealed that MNS1 is abundantly expressed as two closely migrating protein species in the testis but not in other tissues except at very low levels in lung and ovary (Figure 1A). Immunofluorescence analysis showed that the expression of MNS1 is restricted to germ cells in the testis and appears to be predominantly cytoplasmic (Figure S1). The MNS1 protein was first detected at a low level in pachytene spermatocytes at stage VIII and then was abundantly expressed in late pachytene spermatocytes, diplotene spermatocytes, and post-meiotic spermatids (Figure S1). MNS1 was not detected in early germ cells ranging from spermatogonia to mid-pachytene spermatocytes (Figure S1). Our results showed that the expression of MNS1 is not restricted to pachytene spermatocytes as had been previously reported [20].
Immunofluorescence analysis showed that MNS1 localizes to sperm flagella (Figure 1B). In contrast with the relatively continuous distribution of tubulin in the sperm tail, MNS1 localization was not uniform, appearing in a dot-on-a-string fashion (Figure 1B, inset). Our MNS1 antibody was specific, since no staining was observed with Mns1-deficient sperm (data not shown). To assess the biochemical nature of MNS1 localization in sperm tails, we treated epididymal sperm with SDS-EDTA solution to divide sperm proteins into two fractions: soluble (supernatant) and SDS-resistant (pellet) (Figure 1C) [21], [22]. While SDS-treatment solublizes the tubulin components of the axoneme, other axonemal components (such as Tektin), outer dense fibers (ODF), mitochondrial sheath (MS), and fibrous sheath (FS) remain in the SDS-resistant structures [22], [23]. Western blot analysis showed that MNS1, like ODF2, resides in the SDS-resistant flagellar structures (Figure 1C). This result was consistent with the identification of MNS1 as one of the SDS-resistant proteins in the sperm flagella in a systematic proteomic profiling study [22]. As expected, MNS1 still localizes to SDS-treated sperm flagellar remnants (Figure 1D). Collectively, these data demonstrate that MNS1 is an integral component of sperm flagella.
To determine the requirement of Mns1 for spermatogenesis, we generated Mns1 mutant mice by homologous recombination in embryonic stem (ES) cells (Figure 2A). Mouse Mns1 (encoding a protein of 491 aa) is a 10-exon gene spanning over a 20-kb genomic region on Chr. 9. In the Mns1 mutant, deletion of exons 3–8 removes aa 76–423 (Figure 2A). As expected, the internally deleted MNS1 protein (143 aa, MNS1Δ) was expressed in Mns1+/− and Mns1−/− testes (Figure 2B). Western blot analysis revealed two additional MNS1 cross-reactive bands (∼32–37 kD) in the testis (Figure 2B). These two intermediate bands were likely to be specific to MNS1, since they were not detected in Mns1−/− testes (Figure 2B). However, neither the two intermediate protein species nor MNS1Δ was detected in mutant sperm, suggesting that they are not incorporated into the flagella (Figure 2C).
Interbreeding of heterozygous mice yielded slightly fewer homozygous Mns1 mutant offspring (Mns1+/+, Mns1+/−, Mns1−/−: 73, 155, 50) (χ2 = 7.49, P = 0.024), suggesting lethality in some Mns1-deficient embryos or pups. However, Mns1−/− mice grew to adulthood with no gross abnormalities and no increased lethality was observed in Mns1−/− mice after weaning. The body weight of Mns1−/− adult mice was not different from that of wild type animals. Notably, Mns1−/− males were sterile, whereas Mns1−/− females were fertile. The weight of Mns1−/− testes (158±18 mg per pair) from 8-wk-old mice was slightly lower (13%) than that of the wild type (182±10 mg per pair) (Student's t test, P<0.019). Strikingly, the sperm count from Mns1−/− epididymides (0.77×106 per cauda) was dramatically reduced to only 8% of the wild type level (9.4×106 per cauda) (P<0.0001). The Mns1-deficient epididymides were mostly filled with cell debris and contained much fewer sperm (Figure S2). Histological analysis revealed that seminiferous tubules from adult Mns1−/− testes contained all stages of germ cells but the mutant elongated spermatids appeared to lack flagella, in comparison with wild type elongated spermatids (Figure 2D, 2E). Thus, MNS1 is required for male fertility and is essential for spermiogenesis.
We next analyzed the sperm from the cauda epididymis. The sperm flagellum is divided into the midpiece, principal piece, and the short end piece (Figure 3A). We found that the great majority (98%) of sperm from Mns1−/− mice have a very short crooked tail (less than 25% the wild type length) (Figure 3A, 3B). Rarely, sperm with a long flagellum was observed but its flagellum appeared to be abnormal (Figure 3B). In addition to abnormal flagellar morphology, freshly isolated MNS1-deficient sperm exhibited no motility upon microscopic examination. However, the head of MNS1-deficient sperm appeared normal in morphology (Figure 3D, inset) and contained protamine 1 (Figure S3), suggesting that nuclear elongation and nuclear condensation events were not disrupted during spermiogenesis in MNS1-deficient mice. Acrosomes in MNS1-deficient sperm also appeared to be normal in morphology (Figure 3D).
Morphological analysis revealed abnormal assembly of sperm flagella in Mns1−/− mice. Immunofluorescence microscopy showed that MNS1-deficient sperm flagella were positive for tubulin, demonstrating the presence of microtubules (Figure 3D). In addition to the “9+2” microtubular axoneme and outer dense fibers (ODFs) of the wild type flagella, the midpiece and principal piece consist of mitochondrial sheath and fibrous sheath respectively (Figure 3E, 3G). In the MNS1-deficient flagella, the axonemal microtubules and ODFs were completely disorganized (Figure 3F, 3H). The mitochondrial sheath did form but in uneven thickness around the axoneme in the midpiece of the mutant sperm (Figure 3F). The fibrous sheath appeared to be missing in the principal piece of MNS1-deficient sperm (Figure 3H). These analyses demonstrated that MNS1 is essential for the assembly of sperm flagella in mice.
We next tested whether MNS1 is required for somatic ciliary functions. Motile cilia such as ependymal cilia and nodal cilia, also consist of axonemes. Nodal cilia are required for left-right patterning of internal organs during embryogenesis and nodal ciliary dysfunction causes situs inversus [11], [12]. Ependymal cilia are responsible for the flow of cerebrospinal fluid and their defects lead to hydrocephalus [8]–[10]. We found that Mns1−/− mice exhibited abnormal left-right asymmetry. Based on lung lobation patterns, all 13 wild type and 34 Mns1+/− embryos exhibited normal left-right patterning (Figure 4A). However, eight of 22 Mns1−/− embryos examined (E12.5–E16.5) displayed normal situs, eight presented with situs inversus (Figure 4B), and six exhibited left isomerism, indicating that MNS1 is required for nodal ciliary function. In contrast with wild type mice (Figure 4C), post-natal day 24-old Mns1−/− mice developed hydrocephalus (Figure 4D), suggesting that MNS1 is also important for ependymal ciliary function. By histological analysis of brain from mice at post-natal days 2, 4, and 16, we found that Mns1-deficient mice exhibited no hydrocephalus at post-natal day 2 but began to develop hydrocephalus at post-natal day 4. Taken together, these data have shown that MNS1 is essential for motile ciliary functions.
Non-motile cilia are thought to be involved in sensory perception and thus are important for signaling processes such as Hedgehog and Wnt signaling pathways [24], [25]. Defects in non-motile ciliary functions cause polycystic kidney disease, polydactyly, retinal degeneration, etc [5], [6]. However, the MNS1-deficient mice lacked polycystic kidney disease and polydactyly, suggesting that MNS1 is not required for non-motile ciliary functions.
Next we performed ultrastructural analysis of tracheal motile cilia. Unlike the disorganized axonemal structure in MNS1-deficient sperm flagella (Figure 3), the characteristic “9+2” axonemal structure was present in MNS1-deficient tracheal cilia (Figure 5A, 5B). However, close examination revealed that MNS1-deficient motile cilia lack some outer dynein arms and contain only ∼4 of them (3.8±1.6, n = 24) per cross section, in contrast with the 9 outer dynein arms in the wild type (Figure 5B), suggesting that MNS1 is an axonemal protein. MNS1 was previously identified in the ciliary proteome of human bronchial epithelial cells [26]. Our Western blot analysis showed that MNS1 protein is expressed in wild type trachea but absent in Mns1−/− trachea (Figure 5C).
To elucidate a potential molecular mechanism for the function of MNS1 in spermiogenesis and motile ciliary functions, we performed a yeast two-hybrid screen of a human testis cDNA library using the full-length human MNS1 as bait. In this screen, we identified clones encoding the C-terminal half of MNS1. We confirmed the self-interaction of mouse MNS1 by yeast two-hybrid assay (data not shown). We then performed GST pulldown experiments to test if MNS1 physically binds to itself. These in vitro experiments showed that MNS1 binds to GST-MNS1N, GST-MNS1C. and GST-MNS1Δ, but not GST alone, indicating that MNS1 might form dimers or oligomers (Figure 6A). The self-interaction of MNS1 could be mediated by its coiled-coil domain, as both MNS1N and MNS1C contain coiled coil domains. We then analyzed the secondary structure of MNS1Δ (143 aa), which consists of the N-terminal 75 residues and the C-terminal 68 residues (Figure 6B). MNS1Δ was predicted to harbor a coiled coil domain in its middle region, which spans residues from both fragments. Therefore, the coiled coil domain in MNS1Δ may mediate its interaction with the full-length MNS1.
To determine whether MNS1 is able to form oligomers, we expressed MNS1 in NIH 3T3 fibroblast cells. Our data suggest that the full-length MNS1 protein forms short filaments in the cytoplasm (Figure 6B and Figure S4A). In contrast, the three MNS1 truncated proteins (MNS1N, MNS1C, or MNS1Δ) localized diffusely throughout the cells and did not form filaments (Figure 6B and Figure S4B–S4D). These data suggest that only the full-length MNS1 protein is capable of forming polymers. Although over-expression of the full length MNS1 may form non-specific aggregates rather than organized polymers in this assay, we favor the interpretation that MNS1 forms filaments. However, additional electron microscopic and biochemical studies are required to fully understand the structural basis.
Using this filament formation assay, we then tested if the three truncated MNS1 proteins would incorporate into the full-length MNS1 structures (Figure 6B). We co-expressed the full-length MNS1 and each of the three truncated MNS1-GFP fusion proteins in NIH 3T3 cells (Figure 6C). Consistent with our yeast two-hybrid interaction and GST pulldown data, MNS1C localized to the MNS1 filaments, remarkably, to one end of the filaments, suggesting the polarity of the MNS1 filaments (Figure 6C). Interestingly, MNS1N also localized to one end of MNS1 filaments. In contrast, MNS1Δ localized continuously along the MNS1 filaments (Figure 6C), while GFP-MNS1Δ alone did not from aggregates (Figure S5). These data showed that all three MNS1 mutant proteins interact with the full-length MNS1 protein. Importantly, these data suggested that MNS1 filaments possess a polarity and that MNS1 may be assembled in a head to tail fashion.
Here we report that MNS1 is essential for spermiogenesis but dispensable for meiosis. In a previous study, it was concluded that MNS1 is specifically expressed in pachytene spermatocytes [20]. However, we found that the MNS1 protein is abundantly expressed in late pachytene spermatocytes, diplotene spermatocytes, and spermatids and thus is not restricted to pachytene spermatocytes. Using different MNS1 antibodies that we generated, we always detected MNS1 as two closely migrating bands (∼60 kD) in protein extracts from testis, epididymal sperm, and trachea. Our antibodies were specific for MNS1, as evidenced by the absence of two ∼60 kD bands in MNS1-deficient testes, sperm, and trachea. However, in the previous study, the “MNS1” antibody recognized only one 60 kD band in testes and did not recognize any 60 kD protein in epididymal sperm by Western blotting analyses, raising the possibility that the antibody generated in that study might not be specific to MNS1 [20].
Several lines of evidence support that MNS1 plays an essential role in the assembly of sperm flagella. Firstly, MNS1 is an integral component of sperm flagella. The localization of MNS1 to SDS-treated sperm flagella suggests that MNS1 is an SDS-resistant component of sperm flagella. Secondly, MNS1 monomers interact with each other and are able to form fibrous polymers when ectopically expressed, implying that MNS1 might be assembled into filamentous structures in the flagella. Thirdly, in the absence of MNS1, sperm flagella are very short. Microtubules and ODFs are completely disorganized in the mutant sperm.
Inactivation of MEIG1, PACRG, or SPEF2 causes disorganization of microtubules and ODFs in sperm flagella, a phenotype similar to that in MNS1-deficient mice [27]–[30]. MEIG1 and PACRG interact with each other and both are expressed in spermatids [27], [29]. PACRG localizes to the postacrosomal region of the sperm head and the midpiece of the flagellum [29]. Disruption of SPEF2 by a LINE1 retrotransposon insertion in boars and inactivation of SPEF2 in mice led to immotile short-tail sperm defect [10], [31]. However, our yeast two-hybrid screen using MNS1 as bait did not detect any of these proteins. A proteomic screen has identified ∼50 proteins, including MNS1, from the accessory structures of mouse sperm flagellum [22]. These genetic and proteomic studies underscore the complexity in the structure and assembly of sperm flagella.
Based on the NCBI database search, MNS1 is conserved only in organisms with motile cilia. According to the ciliome database [5], MNS1 homologues were identified in a number of genomic and proteomic studies of ciliary proteins, for example, in Chlamydomonas reinhardtii [32]–[34], C. elegans [35], [36], and in humans [26], [37]. Therefore, in addition to male sterility, MNS1-deficient mice exhibit situs inversus and hydrocephalus, suggesting an essential role for MNS1 in motile cilia function and implicating MNS1 in primary ciliary dyskinesia (PCD). Interestingly, MNS1-deficient tracheal motile cilia lack some outer dynein arms, suggesting that MNS1 is an axonemal protein in the motile cilia. A number of genes are implicated in human, dog, and mouse PCD, including DNAI1 [13], DNAH11 [17], DNAH5 [18], [19], TXNDC3 [38], Spag6 [9], Pcdp1 [16], CCDC39 [15], CCDC40 [14], Spef2 [10], etc. However, MNS1-deficient mice grew to adulthood, suggesting that hydrocephalus is relatively mild in these mutant mice. Therefore, our study of MNS1 has important implications for male infertility in humans. Our findings on the function of MNS1 from mouse studies will help to select a cohort of infertile men with defined phenotypes for mutation screening in the human MNS1 gene.
Mice were maintained and used for experimentation according to the guidelines of the Institutional Animal Care and Use Committee of the University of Pennsylvania.
Two GST-mouse MNS1 (amino acids 1–233 and 392–491) fusion proteins were expressed in Escherichia coli using the pGEX4T-1 vector and affinity purified with glutathione Sepharose. Purified recombinant proteins were used to immunize rabbits at Cocalico Biologicals Inc, resulting in MNS1 antiserum UP2060 against GST-MNS1 (aa 1–233) and UP2284 against GST-MNS1 (aa 392–491). Specific antibodies were affinity purified with the immunoblot method as previously described [39].
The epididymal sperm were collected from 3 male mice (C57BL/6) by squeezing the caudal epididymides in 1xPBS solution and centrifugation at 800 g for 5 min at RT. Sperm were homogenized in 1 ml of SDS-EDTA solution (1% SDS, 75 mM NaCl, 24 mM EDTA, pH 6.0) and centrifuged at 5000 g for 30 min at RT. 100 µl of SDS-PAGE sample buffer (62.5 mM Tris, pH 6.8, 3% SDS, 10% glycerol, 5% β-mercaptoethanol, 0.02% bromophenol blue) was added to 100 µl of supernatant, while the pellet was resuspended in 200 µl of SDS-PAGE sample buffer. The samples were heated at 95°C for 10 min and 20 µl of each sample was used for western blot analysis.
To generate the Mns1 targeting construct, DNA fragments were amplified by high-fidelity PCR using an Mns1 BAC clone (RP23-349N4) as template (Figure 2A). V6.5 ES cells were electroporated with linearized targeting construct (pUP101/ClaI). Screening of ES cells was described previously and revealed a targeting frequency of 11% [40]. Two homologously targeted ES cell clones (B6 and G5) were injected into B6C3F1 (Taconic) blastocysts. The Mns1 mutant allele was transmitted through the germline in chimeric mice derived from both clones. All offspring were genotyped by PCR. Wild-type allele (500 bp) was assayed by PCR with the primers GTCAGGAAGATCTACGAGGA and CCAGAAGTCTTGTGCCCTCT. The Mns1 knockout (300 bp) allele was assayed by PCR with the primers GTCAGTTTGCTGTTGTAGAGT and CCTACCGGTGGATGTGGAATGTGTG.
A pretransformed Mate & Plate human testis cDNA library (Clontech) was screened using the full-length human MNS1 cloned into pGBKT7 as bait. GST-pulldown experiments were performed as previously described [41].
For histology, testes were fixed in Bouin's solution and brains were fixed in 4% PFA overnight. Samples were processed, sectioned, and stained with hematoxylin and eosin. For immunofluorescence analysis, testes were fixed in 4% PFA for 3 h at 4°C. Epididymal sperm and NIH 3T3 cells on slides were fixed with 4% PFA for 5 min at RT. Immunofluorescence was performed as previously described [42].
For Western blot analyses, adult tissues were homogenized in a glass homogenizer in SDS-PAGE sample buffer and heated at 95°C for 10 min. Protein concentration was measured using the Bradford assay. 30 µg of total protein for each sample was resolved on 10% SDS-PAGE gel and electro-blotted onto PVDF membrane. Primary antibodies used were MNS1 (this study), ODF2 (Santa Cruz), ACRV1 (P. P. Reddi) [43], Protamine 1 (SHAL) [44], β-tubulin (DSHB), ACTB (Sigma-Aldrich). Texas Red or FITC-conjugated secondary antibodies were used for immunofluorescence. HRP-conjugated secondary antibodies were used for Western blot analyses.
EM was performed at the Biomedical Imaging Core facility at the University of Pennsylvania. Adult testes and dissected trachea were fixed in 2.5% glutaraldehyde and 2% paraformaldehyde overnight, and postfixed in 1% osmium tetroxide for 1 hour. The specimens were further processed and sectioned at the core facility as previously described [40].
The cDNAs encoding full-length or truncated MNS1 were cloned into pcDNA3.1/V5-His TOPO TA for V5 tagging, or pEGFP-C1 for GFP fusion. Transfection of NIH 3T3 cells was performed using the Lipofectamine reagent (Invitrogen). 24–48 hours after transfection, cells were washed with PBS, fixed in PFA, stained with anti-V5 antibody (Invitrogen), incubated with secondary antibodies, mounted with DAPI-containing Vectashield mounting medium (Vector Labs) and visualized on a Zeiss Axioskop 40 microscope with a digital imaging system.
|
10.1371/journal.pbio.2002940 | The Ink4a/Arf locus operates as a regulator of the circadian clock modulating RAS activity | The mammalian circadian clock and the cell cycle are two major biological oscillators whose coupling influences cell fate decisions. In the present study, we use a model-driven experimental approach to investigate the interplay between clock and cell cycle components and the dysregulatory effects of RAS on this coupled system. In particular, we focus on the Ink4a/Arf locus as one of the bridging clock-cell cycle elements. Upon perturbations by the rat sarcoma viral oncogene (RAS), differential effects on the circadian phenotype were observed in wild-type and Ink4a/Arf knock-out mouse embryonic fibroblasts (MEFs), which could be reproduced by our modelling simulations and correlated with opposing cell cycle fate decisions. Interestingly, the observed changes can be attributed to in silico phase shifts in the expression of core-clock elements. A genome-wide analysis revealed a set of differentially expressed genes that form an intricate network with the circadian system with enriched pathways involved in opposing cell cycle phenotypes. In addition, a machine learning approach complemented by cell cycle analysis classified the observed cell cycle fate decisions as dependent on Ink4a/Arf and the oncogene RAS and highlighted a putative fine-tuning role of Bmal1 as an elicitor of such processes, ultimately resulting in increased cell proliferation in the Ink4a/Arf knock-out scenario. This indicates that the dysregulation of the core-clock might work as an enhancer of RAS-mediated regulation of the cell cycle. Our combined in silico and in vitro approach highlights the important role of the circadian clock as an Ink4a/Arf-dependent modulator of oncogene-induced cell fate decisions, reinforcing its function as a tumour-suppressor and the close interplay between the clock and the cell cycle network.
| In mammals, the circadian clock controls the punctual regulation of biological processes, which, in turn, affect physiology and behaviour, allowing for the synchronisation of internal time to environmental light-dark cycles. Malfunctions of the circadian clock are associated with pathological phenotypes including cancer. Given the range of molecular time-dependent processes, including metabolism, DNA repair, and the cell cycle, the clock is hypothesised to act as a tumour suppressor. With the help of mathematical modelling and whole-genome analysis combined with machine learning, we investigated the RAS-dependent dysregulation of the circadian clock. We find that the tumour-suppressor Ink4a/Arf acts as a key mediator of RAS oncogene-induced changes in the circadian system, thereby mediating the interplay between the clock and the cell cycle.
| Earth’s rotation within repetitive cycles of 24 hours (h) led to the evolution of an endogenous timing system across all phyla: the circadian clock, which allows organisms to anticipate and adapt to environmental changes such as light and darkness. At the molecular level, the generation of circadian rhythms in each cell is based on a complex interplay of positive and negative transcriptional/translational feedback loops. In mammals, two main feedback loops have been dissected in greater detail [1]. In the period (PER)/cryptochrome (CRY) loop, a heterodimer of the proteins circadian locomotor output cycles kaput (CLOCK) and brain and muscle aryl hydrocarbon receptor nuclear translocator-like protein (BMAL) drives the expression of genes from the Per and Cry families. The resulting proteins translocate back to the nucleus, form complexes, and bind to the CLOCK/BMAL heterodimer, thereby inhibiting their own synthesis. CLOCK/BMAL also induces the transcription of the nuclear receptor gene families reverse strand of ERBA (Rev-Erb) and RAR-related orphan receptors (Ror) within the ROR/Bmal/REV-ERB loop. REV-ERB and ROR compete for the binding to ROR response elements (ROREs) in the promoter region of Bmal1, regulating its transcription in antagonistic ways. By controlling rhythmic RNA and protein abundance, the cellular endogenous clock regulates circadian rhythms of a variety of biological processes such as rest/activity cycles, metabolism, hormone level, immune functions, and the metabolism of drugs [2,3]. Disturbances of the circadian system are associated with many pathological phenotypes including cancer [4,5], though the effect of the circadian clock in tumourigenesis is still an issue of ongoing debate. Several epidemiological studies reported increased occurrence of cancer in long-term shift workers [6,7], indicating that disrupted circadian rhythms may constitute a cancer risk factor. However, a causative connection is still disputed [8]. Particularly at the molecular level, there is a need for further investigation regarding the putative role of the circadian clock in tumourigenesis. An oncogenic-driven mouse model identified CLOCK and BMAL1 as playing a key role in regulating proliferation and differentiation pointing to a complex role of the circadian clock in cancer progression [9]. Perturbations of the circadian clock (via experimental chronic jet lag) in mice led to accelerated tumour growth, which could be counterbalanced by regular timing of food access [10]. While the disruption of core-clock genes such as Per1 and Per2 has also been associated with cancer promoting mechanisms, [11,12], their role in tumourigenesis is still debatable because a recent study showed that Per1 or Per2 deficiency does not lead to more tumour-prone phenotypes in mice [13]. In turn, cancer strongly influences circadian rhythms of biological processes, such as fatty acid and cholesterol biosynthesis, and metabolic oscillations [14,15]. Current attempts of chronobiological cancer treatment provide promising results, leading to a reduced toxic effect of drugs and increasing survival times of cancer patients [16,17].
One of the processes via which the circadian clock could potentially influence tumourigenesis, namely by triggering malignant proliferation, might be the cell-division cycle in whose regulation the clock is involved [18]. Several links between the circadian clock and the cell cycle have been reported for mammalian cells [19–22]. The clock was found to unidirectionally gate the cell cycle in mouse liver cells via a circadian expression of the cell cycle regulator Wee1 [19], whereas in mouse fibroblasts, the circadian clock was reported to be phase-shifted by mitosis, possibly via concentration changes of the PER-CRY complexes [20]. More recently, the non-POU domain containing octamer binding (NONO)-PER protein complex has been reported to couple the clock and the cell cycle by activating the rhythmic transcription of the cell cycle checkpoint gene Ink4a [22], which encodes, as part of the Cdkn2a locus, the cyclin-dependent kinase (CDK) inhibitor p16Ink4a [23]. p16Ink4a can be activated by the oncogene RAS, leading to cell cycle arrest in the Gap 1 (G1) phase, a tumour-suppressive mechanism counteracting abnormal cell proliferation [24]. A correlation between the expression level of Per2 and the expression of Ink4a mRNA has also been reported [25], indicating that PER is a positive regulator of Ink4a expression and might be responsible for its circadian expression. In addition, Cdkn2a encodes for another tumour suppressor and cell cycle regulator protein, alternate open reading frame (ARF) [26], which is mainly activated by mitogenic stimulation through the cell cycle checkpoint protein avian myelocytomatosis viral oncogene homolog (MYC) [27], but can also be induced via oncogenic RAS [28]. ARF interacts in a circadian time (CT)-dependent manner with a mouse double minute 2 homolog (MDM2), a negative regulator of the tumour suppressor gene p53 [29]. p53, in turn, directly modulates the expression of Per2 by binding to a response element in its promoter region, which is overlapping with the E-box cis-element essential for the CLOCK/BMAL binding and the transcriptional activation of Per2 [30]. In addition, the transcription factor E2F was described as a putative bridging element between the circadian clock and the cell cycle [31].
Recently, we demonstrated that RAS induces a dysregulation of the mammalian circadian clock in HaCaT keratinocyte cells [32]. We showed that the induction of RAS leads to a lengthening of the circadian period while inhibition of the RAS/mitogen-activated protein kinases (MAPK) pathway causes a period shortening. These findings were supported by using a mathematical model to analyse the rhythmic properties of the core-clock network in silico [1,32]. Other mathematical models have been developed that investigate the cell cycle-related functions of the tumour suppressor genes p53 and Arf, as well as Mdm2 [33–35]. Despite the accumulating experimental and in silico evidence pointing to a role of the circadian clock as a suppressor of aberrant cell proliferation, the underlying molecular mechanisms are not yet fully understood. Although mathematical models exist that couple elements of the cell cycle to the components of the circadian clock [36–42], none of them include the cell cycle regulatory genes Ink4a and Arf, nor do they include specific parameters for the investigation of the role of oncogenic-mediated signalling via RAS.
Using mouse embryonic fibroblasts (MEFs) as a model system, we investigated the Ink4a/Arf- and Bmal1-dependent influence of RAS on the circadian phenotype. To attain a deeper understanding concerning the molecular interactions of the circadian clock with components of the cell cycle, we constructed a single-cell semi-quantitative mathematical model. The model allows for the coupling of components of the cell division cycle with the core-clock network, enabling the interpretation of the RAS-mediated effect on the circadian clock phenotype observed in Ink4a/Arf knockout MEFs as compared to their wild-type (WT) counterparts. In particular, the model was used to attain a better understanding of the mechanism by which INK4a and ARF might influence properties of the circadian clock. Furthermore, we performed a comprehensive bioinformatics analysis to investigate the systemic effects of different experimental perturbations at the cellular level. Expression data of a set of core-clock genes, as well as in silico simulations, show that RAS overexpression influences the transcriptional expression and period of core-clock genes such as Bmal and Per differentially in the Ink4a/Arf+/+ and Ink4a/Arf-/- scenarios, which points to a regulatory role of Ink4a/Arf in the RAS-mediated effect on the circadian clock. By analysing the relevance of selected cell cycle components in mediating the RAS-induced change on the circadian period via Ink4a/Arf in silico, we show that the presence of the transcription factors E2F1 and p53 is necessary for simulating this phenotype. Genome-wide transcriptional profiling data from Ink4a/Arf-/- MEFs and their corresponding WT MEFs (Ink4a/Arf+/+) are in line with the mathematical predictions showing that Ink4a/Arf play an essential role in the RAS-mediated effect on the gene expression levels of core-clock and cell cycle-related genes. Hence, our findings resulting from a combined computational and experimental approach provide a deeper insight into the dynamics of the cross-talk between oncogene-induced perturbations of the circadian system and the cell cycle and reveal a novel role for Ink4a/Arf as an important regulator of the RAS-mediated effect on the circadian clock phenotype.
Recently, we showed that oncogenic RAS influences properties of the circadian clock by lengthening the circadian period in different cell types [32]. To attain a deeper mechanistic insight on how RAS perturbs the circadian clock and on its subsequent effect on the cell cycle and cell proliferation, we used a well-established cellular model system of MEFs. We compared the circadian phenotypes of MEFs from WT mice (Ink4a/Arf+/+) to their littermate knock-out MEFs (Ink4a/Arf-/-), which carry a targeted deletion of exons 2 and 3 of the tumour suppressor gene Cdkn2a, disrupting both Ink4a and Arf [23]. We analysed the effect of RAS on the circadian clock phenotype by bioluminescence recordings and the induction of senescent cell cycle arrest by SA-ß-Gal staining. MEFs were lentivirally transduced with a Bmal1-promoter driven luciferase construct (Bmal1:Luc) and bioluminescence was recorded for 5 days as schematically represented in Fig 1A. The bioluminescence data shows similar clock phenotypes for Ink4a/Arf-/- MEFs and their corresponding Ink4a/Arf+/+ littermates with average period values of around 24 h (T = 24.7 ± 0.3 h for Ink4a/Arf+/+ MEFs and T = 24.2 ± 0.2 h for Ink4a/Arf-/- MEFs, n = 5; mean and SEM; representative results in Fig 1B and 1C, summary of the data in Table 1), indicating that the period of the circadian clock is not influenced by the knock-out of Ink4a/Arf. The viability of the cells was not affected by the knockout (S1A Fig).
To specifically study the interaction of the RAS/MAPK pathway with components of the circadian clock, oncogenic RAS (H-RAS G12V) was overexpressed in both cell types. Interestingly, upon RAS overexpression, we observed a change of the circadian clock phenotype: the RAS overexpressing Ink4a/Arf+/+ cells (T = 26.7 ± 0.7 h, n = 5; mean and SEM) show a 2 h longer period as compared to the WT condition (T = 24.7 ± 0.3 h, n = 5; mean and SEM; p = 0.004, Student t test), whereas in the RAS-infected Ink4a/Arf-/- MEFs, a decrease of 2.5 h in the period was observed (T = 21.7 ± 0.6 h, n = 5; mean and SEM; p = 0.0003, Student t test) as compared to cells with unmodified RAS levels (T = 24.2 ± 0.2 h, n = 5; mean and SEM; representative results in Fig 1B and 1C, summary of the data in Table 1). To investigate a putative effect of RAS on core-clock components, Bmal1 was downregulated by short hairpin RNA (shRNA) prior to the stable expression of oncogenic RAS in both Ink4a/Arf-/- MEFs and their WT littermates. As expected, the downregulation of Bmal1 disrupted circadian rhythmicity (representative results in Fig 1D and 1E, summary of the data in Table 1). Bioluminescence data from several independent experiments (five WT mice and Ink4a/Arf-/- littermates, with two independent transductions per condition) are summarised in Fig 1F and Table 1 and reproduced our representative results shown in Fig 1B and 1C. Ink4a/Arf-/- MEFs proliferate faster than the Ink4a/Arf+/+ MEFs, independent of the downregulation of Bmal1 (Fig 1G).
To investigate the effect of RAS overexpression and Bmal1 downregulation on cellular senescence, we performed SA-ß-Gal staining using Ink4a/Arf-/- MEFS and WT littermates (n = 3 WT mice and Ink4a/Arf-/- littermates) and tested for senescence-associated galactosidase activity. While the Ink4a/Arf+/+ MEF population has a low amount of senescence cells (7 ± 1.44%, n = 3; mean and SEM), RAS overexpression leads to a strong increase (94.5 ± 1.15%, n = 3; mean and SEM; Fig 1H and 1I). The results are in agreement with published data showing that RAS overexpression increases the percentage of senescent cells in WT MEFs [24]. In comparison, the population of Ink4a/Arf-/- MEFs shows the lowest number of senescent cells (0.83 ± 0.44%, n = 3; mean and SEM) independent of RAS overexpression (1.5 ± 0.5%, n = 3; mean and SEM; Fig 1H and 1I). The downregulation of Bmal1 shows no significant effect on the senescence phenotype (Fig 1H and 1I).
To explore the observed Ink4a/Arf-dependent effect of RAS overexpression on the clock—an increased circadian period in WT MEFs but a decreased period in Ink4a/Arf-/- cells (Fig 1B–1E)—we developed a novel semi-quantitative mathematical model coupling the mammalian cell cycle and the circadian clock using ordinary differential equations (ODEs).
The model contains all elements of our previously published single-cell model of the mammalian circadian core-clock [1], the cell cycle elements INK4a and ARF, and a minimal selection of their respective interaction partners, which allow for the connection to the core-clock. These include the cell cycle checkpoint regulators Myc (G1/S) and Wee1 (G2/M) and components of two INK4a- and ARF-dependent signalling pathways: the INK4a/ retinoblastoma-associated protein 1 (RB1)/E2F1 pathway (module 1) and the ARF/MDM2/p53 pathway (module 2). The resulting regulatory network includes nine additional elements (MYC, WEE1, ARF, MDM2, INK4, CDK/Cyc, p53, RB1, E2F1) and their corresponding transcriptional and translational components (Fig 2A, S2 Fig). In total, the model comprises 46 variables and 170 parameters.
In the model, the CLOCK/BMAL complex activates the transcription of Wee1 and represses Myc transcription as reported in the literature [19,43,44]. The transcriptional activation of Ink4a by the PER-NONO complex in a circadian manner [45,46] is modelled as a positive interaction between the PER/CRY complex and Ink4a transcription. By forming a complex with D-type CDKs CDK4 and CDK6, INK4a prevents the interaction with the cell cycle checkpoint regulator Cyclin D (CycD), thereby inhibiting the subsequent phosphorylation of the RB1, a key regulator of the E2F family of transcription factors (E2F1, E2F2, and E2F3a) [47–51]. The dephosphorylated active form of RB1 inhibits the dissociation of the RB1/E2F complex leading to the inactivation of E2F-mediated transcription of cell cycle genes. The cell cycle/core-clock loop in module 1 is completed by a predicted transcriptional activation of Bmal by E2F because E2F potentially binds to the promoter region of Bmal1 (as reported by MotifMap, the genome-wide map of candidate regulatory motif sites for humans [52]). The prediction of E2F as an important binding element between the clock and the cell cycle is further supported by data from the unicellular red alga Cyanidioschyzon merolae in which time-dependent phosphorylation of E2F promotes the G1/S transition and a mutation of the E2F phosphorylation sites results in an uncoupling of cell cycle progression from the circadian clock [31]. Altogether, these results point to a putative connection between E2F and the clock which—given the existing data—can be assumed to happen via Bmal1. Module 2 contains the ARF/MDM2/p53 pathway and represents an indirect feedback from ARF to the core-clock. ARF is encoded by the same gene locus as INK4a and can be activated by oncogenic MYC or oncogenic RAS [47]. Accumulated ARF associates with the p53 inhibitor MDM2 and leads to its rapid proteasomal degradation. This decreases MDM2-mediated ubiquitination of the tumour suppressor p53 and promotes its stabilisation, which in turn activates the transcription of Mdm2 [53,54]. A recent study describes a p53 response element located in the promoter region of Per2, which overlaps with the E-box cis-elements crucial for CLOCK/BMAL-mediated Per2 transcription [30]. Thus, the binding of p53 strongly represses the transcription of Per2 by competing with the CLOCK/BMAL binding to its promoter [30]. Apart from the strong negative influence of p53 on Per2, PER2 is also known to transcriptionally modulate p53 [55], and this regulation is thought to be positive [39]. For simplicity, we have merged the mutual influence into one negative interaction from p53 to Per2. In addition, p53 inhibits the phosphorylation of RB1 via the p21/CDK/CycE/RB1 pathway [56,57]. Although we cannot exclude the possibility that other elements may also be involved in connecting the clock and the cell cycle elements INK4a and ARF, the chosen modules represent a minimal functional set well-supported by published data [44,58–60] that enables us to investigate the properties of the system in silico. The model development and the complete network scheme is described in greater detail in the supplementary information (S1 Text, S2 Fig). The resulting model is robust to perturbations in the range of ± 10% as shown by the control coefficient analysis over all parameters (S1 Text). The period of the model system was adjusted to 23.65 h and the phase of Bmal mRNA expression was set to CT 21 h.
To examine whether simulations of the model are in agreement with biological phenotypes of the clock, we compared the peak phases of the in silico mRNA expression patterns of core-clock genes and the PER/CRY protein complexes with experimental data retrieved from the literature [1]. The peak phases for all core-clock genes are within the range of published experimental peak phases of core-clock mRNAs (Fig 2B, Table 2). Furthermore, the model successfully reproduces the correct phase relations among the core-clock components (Fig 2C).
Data from our previous work points to a role of RAS as a regulator of the circadian clock period [32] as was also reported by other studies [61]. With our novel mathematical model, we investigated whether the observed RAS-induced change in the circadian clock period (Figs 1B and 2C) can be simulated both in the Ink4a/Arf WT and in the knock-out condition (Fig 2D). For the simulation, we adapted a method from our previous work on RAS-mediated dysregulation of the circadian clock in cancer, in which a factor ktt was introduced to the activation/inhibition rates describing CLOCK/BMAL-mediated transcription: ktt = 1 describes a normal RAS expression level whereas ktt < 1 indicates a reduction in the transcriptional activity of CLOCK/BMAL caused by RAS overexpression [32].
The double knock-out of Ink4a/Arf was achieved by setting the initial conditions of INK4a and ARF mRNAs, cytoplasmic and nuclear proteins, as well as their rate of change to 0 (S1 Text, equations 1, 2, 8, 10, 14, 19 = 0). We measured the period in our model for a transient region, defined as the mean of the time between the first four peaks (three periods) after introducing the perturbation of RAS (represented by ktt < 1) to the system. In this transient region, there are still fluctuations of the modelled system that can represent the observed biological noise of retrovirus-mediated RAS overexpression. More information concerning the model analysis can be found in S1 Text.
The model predicts a slightly longer circadian period for the Ink4a/Arf-/- system (23.68 h compared to the adjusted period of 23.65 h in the WT) when RAS is expressed at WT levels (ktt = 1; Fig 2D), which results in a phase shift over time (S3A Fig). As expected from the experimental data shown in Fig 1 and our previously published results [32], there is a lengthening of the period upon RAS overexpression in Ink4a/Arf+/+ MEFs (Fig 2D). Interestingly, the opposite effect on the period is predicted by simulations in the Ink4a/Arf-/- system. Thus, the in silico period changes are in agreement with the experimentally measured phenotypes in Ink4a/Arf-/- MEFs and their WT littermates (Fig 1B and 1C) and show the same tendency of an increase/a decrease of the period length in response to RAS overexpression. The simulations were not fitted to exactly reproduce the values of the experimentally measured periods, which were obtained from MEFs originating from different mice and therefore show a certain biological variation. The simulated Ink4a/Arf-/- system shows a nonmonotonic dependency of the period length on the strength of the RAS overexpression. With increasing RAS, the period length first decreases, reaching its minimum for ktt = 0.7, only to increase afterwards in response to even higher simulated levels of RAS (0.4 < ktt < 0.7; Fig 2D). After introducing the perturbation of RAS to the system by varying the parameter ktt, there are nonmonotonic phase shifts of Bmal expression that depend on the strength of RAS overexpression (S1 Text). The change of the ktt value causes Bmal oscillations to peak at different times: The system with the strongest RAS overexpression and lowest simulated ktt value (ktt = 0.4) peaks the earliest for the first iteration but is then superseded by the system with ktt = 0.7 for the following iterations. The WT system always peaks last. These phase shifts are especially prominent for the Ink4a/Arf-/- system, which represents one possible explanation for the observed nonmonotonic period changes (Fig 2D). Yet the effect of RAS overexpression on the length of the period is also dependent on the time when the perturbation is introduced to the model system (S3B Fig). Interestingly, the model predicts a longer period for an inhibition of RAS (ktt > 1) in the Ink4a/Arf-/- system (S1 Text). We were able to confirm this prediction experimentally by inhibiting RAS in the Ink4a/Arf-/- MEFs by using a MEK inhibitor that blocks the downstream chain in the RAS signalling pathway. Contrary to the shortened period observed when overexpressing RAS, we now observed an increase of the period (T = 23.76 ± 0.1 h for Ink4a/Arf-/- MEFs and T = 25.24 ± 0.1 h for Ink4a/Arf-/- -RAS MEFs; n = 3; mean and SEM; S1C–S1E Fig). Furthermore, we simulated the overexpression of RAS in the Ink4a/Arf+/+ condition (ktt = 0.6). The model predicts that RAS overexpression leads to an increase of the expression level of Ink4a as compared to the normal RAS scenario (Fig 2E). This is in line with published data showing that RAS activates Ink4a, leading to cell cycle arrest [24,27] and correlates with the RAS-induced increased number of senescent cells shown in Fig 1I. Moreover, we downregulated RAS in the Ink4a/Arf-/- MEFs and observed a long period phenotype, as opposed to the period change observed for RAS overexpression (S1C–S1E Fig). Furthermore, we used an RAS-inducible construct to investigate the effects of different levels of RAS induction. Our data indicates that both longer and shorter periods can be observed, in a RAS-dependent manner, in agreement with the simulations from our mathematical model (S1F–S1J Fig).
Taken together, the results demonstrate that the model reproduces important circadian properties, which are in agreement with experimental data, and that it can be used to further elucidate the mechanism of RAS-induced and Ink4a/Arf-dependent changes of the circadian period.
In order to investigate the relative influence of the INK4a/RB1/E2F1 (module 1) and ARF/MDM2/p53 (module 2) pathways in mediating the RAS-induced effect on the core-clock in silico, we tested whether the presence of key elements in the network and the oscillations of both modules are necessary to reproduce the experimentally determined clock period phenotypes of Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs.
As shown in Fig 3A, module 1 represents the connection of INK4a to the clock via the INK4a-dependent inhibition of E2F, a transcription factor that we predict to regulate the transcription of Bmal in the model. Still, we cannot exclude the possibility that such a regulation may happen via additional elements that were not investigated within the scope of this study (S3 Text). However, additional elements would not change the validity of the conclusions derived from the model as long as the delays in the expression values between the defined core-clock elements remain. These delays or phase differences, which were retrieved from published experimental data, were used as constraints in our model (S1 Text). Module 2 (Fig 3B) represents the ARF-mediated activation of the transcription factor p53, which is known to repress the CLOCK/BMAL-mediated transcription of Per2. In both modules, the nuclear protein elements show oscillations in their simulated expression patterns according to the circadian rhythm of the core-clock system (Figs 3C and 1D).
The connection of module 1 to the core-clock was disrupted by setting the concentration of E2FN to 0 (S1 Text, equations 7, 12, 22 = 0; initial concentration of E2FN = 0), thereby removing its predicted regulation of Bmal in the model. Ink4a/Arf knockout was modelled as described above. Upon the disconnection of module 1, both the Ink4a/Arf+/+ and the Ink4a/Arf-/- system acquire shorter periods when RAS is overexpressed (ktt = 0.7) as compared to normal RAS conditions (ktt = 1; Fig 3E). These results are not in line with our previous simulations for the full network (Fig 2D) and the experimental observations (Fig 1A and 1B), which show an Ink4a/Arf-dependent effect on the period upon RAS overexpression. Thus, it seems that in our modelling scenario the predicted connection between E2F and Bmal is indeed necessary to reproduce the observed period changes. When comparing the expression and the period length of Bmal oscillations before and after the perturbation by RAS, it becomes evident that the knockout of module 1 results in a lower period value of 22.86 h for both the Ink4a/Arf+/+ and the Ink4a/Arf-/- system. This results in a phase shift of Bmal oscillations when compared to the oscillations in the WT, causing the differing effect on the Bmal period length upon the perturbation by RAS (S1 Text). We further investigated the relevance of the oscillations in module 1 for the RAS-mediated effect on the circadian period. The connection between E2FN and Bmal was maintained, but the expression of E2FN was clamped to the constitutive concentration of its mean value (S1 Text, E2FN = 5.7, equations 7, 12, 22 = 0). The constant expression of E2F leads to similar period lengths of 23.63 h for the Ink4a/Arf+/+ and the Ink4a/Arf-/- systems (Fig 3F). In this case, RAS overexpression causes a nonmonotonic change of the period length depending on the value of ktt, but independent of the Ink4a/Arf status resulting in slightly shorter periods for 0.7 ≤ ktt < 1 and longer periods for 0.4 ≤ ktt < 0.7. Again, this does neither reproduce the results of the previous simulation nor the experimental observations. Bmal oscillation profiles of both systems shows that there is only a very small phase shift between the Ink4a/Arf+/+ and the Ink4a/Arf-/- systems, potentially explaining why both systems react similar upon perturbation by RAS (S1 Text). These results indicate that in the model, both the connection of the INK4a/RB1/E2F1 module to the core-clock via E2F regulation of Bmal and the low-amplitude oscillations of E2F are crucial to simulate the contrary effect of RAS overexpression on the circadian period.
Next, we tested whether the connection of module 2 to the core-clock is also necessary to reproduce the experimentally observed phenotype by setting the concentration of the component p53N to 0 (S1 Text, equation 16 = 0). The expression profile of Bmal shows that the perturbation is introduced at a similar oscillation phase for both systems, which differs from that of the WT Bmal oscillations (S1 Text). As before, both the Ink4a/Arf+/+ and the Ink4a/Arf-/- system show a nonmonotonic change of the period dependent on the strength of the simulated RAS overexpression: for 0.7 ≤ ktt < 1, there is a decrease in period length, followed by a slight increase for 0.4 ≤ ktt < 0.7. This does not reproduce the experimentally observed RAS-induced period changes indicating that the connection of the ARF/MDM2/p53 pathway to the core-clock is crucial as well. To simulate a constitutive connection of module 2 to the core-clock, we clamped the oscillatory expression of p53N in module 2 to its average expression value (S1 Text, equation 16 = 0.6). Interestingly, by using a constitutive concentration of p53N, we were able to simulate circadian phenotypes similar to the simulations of the whole network and the experimental observations (Fig 3H). Although p53 is no longer oscillating, there is a similar phase shift between the Ink4a/Arf+/+ and the Ink4a/Arf-/- system as in the original (S1 Text). This suggests that the oscillation of the ARF/MDM2/p53 pathway plays only a minor role in regulating the period in response to RAS overexpression and can instead be substituted by a constitutive expression.
Taken together, these results indicate that upon RAS overexpression, the connections of both the ARF/MDM2/p53 and the INK4a/RB1/E2F1 pathway to the core-clock are necessary to produce the observed period phenotypes. The simulations from the mathematical model predict that a circadian expression of the INK4a/RB1/E2F1 pathway is crucial for reproducing the Ink4a/Arf-dependent change in rhythmicity in response to RAS overexpression. This again points to dynamical effects such as phase shifting in gene expression, which are hard to identify experimentally because even small phase shifts can cause large effects in a feedback loop.
Furthermore, we simulated the expression of representative components of the core-clock (Per), module 1 (E2f), and module 2 (p53N) under four different conditions (Ink4a/Arf+/+, Ink4a/Arf+/++RAS, Ink4a/Arf-/-, and Ink4a/Arf-/-+RAS) and compared them to experimental 24-h time-course measurements of Per2, E2f, and p53 under the same conditions. The knockout of INK4a and ARF and the overexpression of RAS (ktt = 0.6) were modelled as described above. Both the in silico simulations and the experimental measurements show that Per is oscillating with a circadian rhythm in all conditions (Fig 3I). As expected, there is a phase shift in the simulated Per expression upon the knockout of Ink4a and Arf, with the knockout peaking slightly later than the WT, which is in line with the phase shift observed in the simulations of Bmal expression (S3A Fig). The same tendency can be observed in the experimental time-course measurements of Per2. RAS overexpression reduces the amplitude of Per oscillations both in the WT and the Ink4a/Arf-/- system (Fig 3I). E2f and p53 exhibit only low expression changes. The simulated expression patterns of E2f show oscillations with similar amplitudes in all four conditions while in the experimental measurements, the WT has a higher fold change than the perturbed conditions (Fig 3J). The in silico expression of p53N shows low amplitude oscillations in the WT conditions (both with and without RAS overexpression), which are out-of-phase with Per oscillations (S1 Text) in agreement with a recent work that modelled the spatiotemporal regulation of p53 by Per2 [39]. In the knockout condition, its expression is constitutive and does not change upon overexpression of RAS (Fig 3K), which is reflected in the low fold change observed for p53. This supports our prediction that the constitutive, but not the oscillatory expression of the ARF/MDM2/p53 pathway is necessary to reproduce the Ink4a/Arf-dependent effect on the circadian period upon RAS overexpression as shown in the modular analysis.
To further investigate the hypothesis that Ink4a/Arf plays a crucial role in the RAS-mediated effect on the core-clock with propagating effects on clock-regulated genes, we conducted a genome-wide screening of gene expression upon RAS overexpression and Bmal1 downregulation using microarrays of Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs under the eight different conditions summarised in Fig 1A. The experimental setup was validated by reverse transcription quantitative PCR (RT-qPCR) for Bmal1, Ink4a/Arf, and the oncogene HRas (S1B Fig).
We carried out a principal component analysis (PCA) based on genome-wide expression values (Fig 4A, S2 Fig). The eight conditions were separated into four groups along the first three principal components. Notably, the conditions with downregulated Bmal1 are grouped with the corresponding WT counterparts, indicating a limited effect of the Bmal1 knockdown at the whole genome level. Upon RAS overexpression, Ink4a/Arf+/+ MEFs and Ink4a/Arf-/- MEFs are visibly separated, reinforcing our hypothesis that Ink4a/Arf influences the RAS-mediated effect on the system.
To explore possible correlations between components of the circadian clock and the genes most affected by RAS overexpression and/or knockdown of Bmal1 in Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs, we used molecular interaction data from the IntAct database [62] to create a network of the two gene sets. The first gene set consists of the top 50 differentially expressed genes whose expression levels exhibit the highest variance across the four groups determined by the PCA (Fig 4B). The list of the topmost 50 differentially expressed genes is provided in S1 Table. Interestingly, 32 out of the top 50 differentially expressed genes have been shown to oscillate in the suprachiasmatic nucleus (SCN) or other peripheral tissues, such as liver, kidney, or lung, according to the database of mammalian circadian gene expression, CircaDB [63] (S1 Table). The second gene set is comprised of 166 clock and clock-controlled genes that constitute our previously published network of circadian regulated genes (NCRG) [64].
For the network, we only included elements of the NCRG connected to the set of differentially expressed genes via a direct interaction or via one connecting element and further reduced the network to one connected component. The resulting network consists of three gene/protein sets that form an undirected graph of 331 nodes, 36 of which are part of the original differentially expressed genes and 122 are part of the NCRG (S4 and S5 Figs). The remaining 173 nodes are connecting elements of the two gene sets. The genes from the mathematical model are present in all three subsets of the network (Fig 4C). Components of modules 1 and 2 can be found in the set of differentially expressed genes (Ink4a/Arf) and the set of connecting genes (Cdk4, Cdk6, and Cyclin D1 of module 1 and p53 and Mdm2 of module 2) in addition to Myc. Five core-clock genes are part of the reduced NCRG (Bmal1, Cry1, Cry2, Rorγ, and Rev-Erbβ). The significance of the number of direct interactions between the differentially expressed genes and the circadian clock was tested by comparing the network properties to those of 100 randomly generated networks based on the set of differentially expressed genes and a set of 122 randomly chosen genes. This resulted in a considerably lower average of 3 ± 2 interactions (mean and SD) between the set of differentially expressed genes and the random gene sets as compared to 34 interactions with the NCRG based on data by iRefIndex [65].
To analyse the biological significance of the network, we used consensuspath.db to determine enriched Wikipathway terms (Fig 4D and 4F). As expected, circadian rhythm-related genes are the topmost enriched pathway. In addition, there is an overrepresentation of pathways related to various types of cancer such as pancreatic cancer, breast cancer, bladder cancer, glioblastoma, and retinoblastoma. Furthermore, there are several enriched pathways that are related to carcinogenesis, including the Wnt signalling pathway (Gsk3β, Apc), the ErbB pathway (Erbb2, Gsk3, Myc), the TNFα pathway (Nfkb, Mapk3), the Notch signalling pathway (Ncor1,2), and the TGF-β signalling pathway (Smad4).
Additionally, several cell cycle-related pathways are among the topmost enriched terms of the whole network including contrasting cell cycle-fate phenotypes, such as apoptosis signalling, G1 to S cell cycle control, and senescence. The enrichment analysis for the set of connecting elements yields similar results as the analysis for the whole network, but includes also the MAPK signalling pathway and DNA replication among its overrepresented pathways (Fig 4F).
The high number of direct interconnections and connections via one intermediate element between the two gene sets indicates a strong interplay between the mammalian circadian core-clock and the genes that were differentially expressed upon RAS overexpression and/or knockdown of Bmal1 in Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs. The presence of various cell cycle-related genes in the set of connecting components in addition to the results of the enrichment analysis of overrepresented pathways hints at an involvement of the cell cycle in this connection. The enrichment analysis places the network connections into a potential cancer context that involves genes affected by disturbances of the circadian system and RAS oncogenic signalling, in addition to cell cycle-related genes.
To specifically investigate the connection between components of the circadian clock and the cell cycle and the effects of RAS overexpression and Bmal1 downregulation in the system, we analysed the expression levels of 23 genes included in our mathematical model (Fig 4E, S2 Table) and validated the results for selected genes with RT-qPCR (Fig 4G). The clustering based on the gene set mimics the overall separation on the genome-wide level as determined by the PCA (Fig 4A) and is specific for the selected genes when compared to random sets of the same size (p = 4.54e-07). Notably, the changes in expression of clock and cell cycle-related genes upon RAS overexpression and downregulation of Bmal1 differ between Ink4a/Arf+/+ MEFs and their Ink4a/Arf-/- littermates.
As expected, Ink4a/Arf itself is up-regulated upon RAS overexpression as predicted by the model (Fig 2E), which is in agreement with published data [24]. This increase is higher when combined with the downregulation of Bmal1. When knocking out Ink4a/Arf, the connection between the clock and cell cycle via the activation of Ink4a by the NONO-PER protein complex is severed. This leads to an increase in the expression levels of several core-clock genes such as Cry2 and Per2, while the expression level of Bmal1 does not change significantly (Fig 4G). The knockout of Ink4a/Arf differentially influences a number of cell cycle genes. Despite being inhibited by Arf, expression of the oncogene Mdm2 decreases upon Ink4a/Arf knockout, which in turn, leads to an increase of the expression levels of the tumour-suppressor gene p53 (Fig 4G). The consistent up-regulation of p53 in all knock-out conditions leads to a decrease in the expression levels of the tumour suppressors p21 and Rb1, which might ultimately result in the proliferation effect observed for the Ink4a/Arf-/- MEFs (Fig 1G).
Perturbation of the circadian clock by knockdown of the core-clock gene Bmal1 leads to an increase in Cry2 and Per2 expression levels. However, the distinct up-regulation of Per2 levels in Ink4a/Arf+/+ MEFs is not found in Ink4a/Arf-/- MEFs. Interestingly, the tumour suppressor genes p53 and Rb1 are up-regulated upon Bmal1 knockdown, which points to a possible compensation of the tumour-suppressor role of the circadian system.
RAS overexpression in the WT MEFs decreases the average expression levels of most core-clock genes, such as Cry2 and Per2 with the exception of Bmal1, which shows an increased expression (Fig 4E and 4G). This fits with our previously published data in which we showed that overexpression of RAS decreases the activity of BMAL1 [32], which should lead to a lower transcriptional activation of elements of the Per and Cry families. In the Ink4a/Arf-/- MEFs, we observed a stronger RAS-induced downregulation of Cry2 and Per2, indicating that the influence of RAS on the circadian clock is indeed mediated by Ink4a/Arf. RAS overexpression concurrent with Bmal1 downregulation leads to an increase in the expression levels of the clock genes Cry2 and Per2 and the cell cycle genes Mdm2 and Rb1 in Ink4a/Arf+/+ MEFs, whereas upon knockdown of Ink4a/Arf, the opposite effect can be observed for Cry2, Per2, and Rb1 (Fig 4G). This shows that the effects of RAS overexpression and Bmal1 downregulation of the expression levels of clock and cell cycle-related genes can differ between Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs, particularly when both perturbations are combined.
Given the strong influence of the Ink4a/Arf knockout on several clock and cell cycle genes, we further analysed the expression across the different experimental conditions of 32 manually curated genes associated with senescence [66–69] (S2 Table, Fig 5A).
In order to predict whether the Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs exhibit normal, proliferating,or senescent behaviour upon RAS overexpression and/or downregulation of Bmal1, a machine learning analysis with a support vector machine (SVM) was performed. The SVM algorithm creates a model based on information from a set of training conditions to predictively classify the remaining conditions based on the expression of the 32 senescence-associated genes. For the training set, we selected conditions with strong cell cycle-fate phenotypes as determined by the experimental data: Ink4a/Arf+/+ (normal), Ink4a/Arf+/++RAS (senescence), and Ink4a/Arf-/- (proliferation; Fig 1G and 1I). A two-dimensional representation of the resulting classification based on the expression of the genes Rb1 and Suv39h1, a histone methyltransferase known to be essential for senescence [70], is shown in Fig 5B.
Ink4a/Arf+/+ shBmal1 MEFs were classified as normal, whereas senescent behaviour was predicted for Ink4a/Arf+/+ shBmal1+RAS MEFs. This is in line with the experimental results of the SA-β-gal staining, which showed a high percentage of senescent cells in Ink4a/Arf+/+ shBmal1+RAS MEFs, but not for Ink4a/Arf+/+ shBmal1 MEFs (Fig 1I) and similar cell growth for Ink4a/Arf+/+ shBmal1 MEFs and their WT littermates (Fig 1G). The Ink4a/Arf-/- conditions were all predicted to be proliferating. Experimentally, both Ink4a/Arf-/- and Ink4a/Arf-/- shBmal1 MEFs have been shown to proliferate (Fig 1G). For the remaining conditions, we conducted a fluorescence-activated cell sorting (FACS) analysis to determine the percentage of cells in each cell cycle phase and compared them to those of the three training conditions (Ink4a/Arf+/+, Ink4a/Arf+/++RAS, Ink4a/Arf-/-; Fig 5C).
FACS analysis of the WT MEFs yields a distribution of 44.1 ± 2.7% (n = 3, mean and SEM) cells in G1 phase, 10.8 ± 2.5% in S phase, and 45.1 ± 2.7% cells in G2/M phase. The Ink4a/Arf+/++RAS MEFs show a higher percentage of cells in G1 phase (63.6 ± 0.3%) as opposed to a lesser number of cells in G2/M phase (25.0 ± 0.9%) as expected for senescent cells arrested in G1 phase. In contrast, the Ink4a/Arf-/- MEFs that show a proliferating phenotype (Fig 1G) have a larger number of cells in S phase (19.7 ± 2.5%) than the WT. Similarly, Ink4a/Arf-/- shBmal+RAS MEFs have a higher percentage of cells in S phase (24.2 ± 10.2%) than the WT, indicating a proliferating behaviour, as do Ink4a/Arf-/-+RAS MEFs, albeit to a smaller extent (13.3 ± 0.8%) (Fig 5C). This shows that the expression data of the 32 senescence-associated genes suffices to correctly predict the senescence/proliferation state of Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs upon different perturbations.
In order to compare our observations regarding the interplay between the clock and the cell cycle in the MEFs system to human experimental models, we additionally analysed available microarray gene expression data from different human cells (see Materials and methods). These include the human primary fibroblast cell line IMR-90 and the human colorectal cancer (CRC) cell lines RKO, LIM1215, Caco2, HCT116, SW403, HT29, COLO205, SW480, and SW620.
H-Ras expressing IMR-90 cells undergo oncogene-induced senescence [71] in agreement with our observations in the MEFs system. In the published array set (GEO-GSE33613), the downregulation of ERK2 in H-Ras expressing IMR-90 cells causes the cells to bypass senescence and to proliferate instead. We analysed this dataset and found that both Bmal1 and Ink4a/Arf were downregulated in the proliferating cells when compared to the H-Ras expressing IMR-90 cells as observed in the MEF cells (S6A Fig).
We further analysed available array data of human CRC cells from our own group (GEO-GSE46549) [2] and analysed the expression of all the genes, which are represented in our mathematical model. The comparison of the resulting heat map (S6F Fig) with the one of Fig 4E shows large differences between the cancer cells and the primary MEF cells analysed. Moreover, the genetic profile is very different in between the CRC cells.
In the metastatic SW620 cells, we measured an up-regulation of Ink4a/Arf and a downregulation of HRas and Bmal1 as compared to the primary SW480 cells (S6B Fig). Furthermore, we carried out cell cycle measurements by FACS for both cell lines, which show an increase in the number of cells in S-phase and a corresponding decrease in the percentage of cells in G1-phase for SW620 (S6E Fig). This points to an increase of proliferation associated to a reduction of Bmal1 despite an increased level of Ink4a/Arf in these cells. We further downregulated Bmal1 in both SW480 and SW620 and analysed the subsequent effects on the cell cycle (S6C and S6D Fig). We measured an increase of approximately 7.5% of cells in the S-phase and a similar decrease of cells in the G1-phase for the SW480 cells (S6E Fig). No significant effect was observed in the SW620 cells in which Bmal1 was already at low levels before its downregulation (S6B Fig). It seems that the decreased expression levels of Bmal1 are associated to more proliferative scenarios in these cell lines.
Altogether, our results reinforce the hypothesis that Ink4a/Arf—being a part of the cellular response upon RAS overexpression—acts as a regulator for the oncogene-induced effect on the clock phenotype and cell cycle-fate decision.
In this study, we explore a potential mechanism via which the oncogene RAS is able to dysregulate the circadian clock and the interplay of this interaction with the cell-division cycle. We build up upon previously published data, in which we showed that the overexpression of RAS in cancer cell lines leads to a lengthening of the clock period while its downregulation causes a shortening [32]. RAS is also known as an elicitor of cell cycle decisions depending on the presence of the tumour suppressor genes Ink4a and Arf [23]. Moreover, recent data points to an additional connection between the circadian clock and cell proliferation via the NONO-PER protein complex that activates the rhythmic transcription of Ink4a [46]. Hence, we hypothesise that the RAS-dependent dysregulation of the circadian clock might be achieved via elements involved in the cell cycle that are interlocked with the circadian system. We explored this idea in silico, by means of a novel mathematical model which connects the core-clock and the cell cycle, and experimentally, using an Ink4a/Arf knock-out mouse model system.
We analysed the circadian clock phenotype of WT MEFs and their littermates carrying a knockout of the tumour suppressors Ink4a and Arf, under different conditions. While the clock period is similar for the WT and the knock-out MEFs, both cell types show an RAS-dependent response of the clock phenotype with an opposite effect on the period length. As expected from our previous results [32], the WT Ink4a/Arf+/+ MEFs exhibit a longer period upon RAS overexpression. Interestingly, in the Ink4a/Arf-/- MEFs, RAS overexpression causes a shortening of the period. We hypothesise that this differential effect might be due to a phase shift in the expression of circadian elements that is induced by the knockout of Ink4a/Arf. The cell cycle checkpoint and tumour suppressor gene Ink4a/Arf seems to influence the RAS-mediated effect on the clock phenotype, pointing to a cross-talk between the core-clock and the cell cycle in an oncogenic-dependent manner.
A number of cell cycle regulators such as Myc, Wee1, and Ink4a/Arf are known to be clock-controlled [18,19], but the complex molecular mechanisms that couple these two biological oscillators are not entirely understood. Deeper insights into the underlying mechanisms might be of great advantage for understanding how a disruption of our internal timing system might induce malignant proliferation of cells in cancer.
Thus, to investigate the interplay between the circadian clock and cell cycle elements, we developed a novel mathematical model of the mammalian core-clock that includes the tumour suppressors Ink4a/Arf, as well as core-clock elements and cell cycle checkpoint genes, some of which have been reported to be directly regulated by the circadian clock as described above.
The model predicts RAS-induced Ink4a/Arf-dependent changes in the length of the clock period, as well as an RAS-induced increase of Ink4a transcription levels, which correlates with the experimentally observed cell cycle arrest phenotype. For the model construction, we assumed that Bmal1 transcription can be regulated by E2F as an additional coupling element between the clock and the cell cycle based on candidate regulatory motif sites and published data. Indeed, if we increase the strength of the E2F activation on Bmal in the model (see control coefficient analysis, S1 Text) we obtain a longer period. This is consistent with the increase of Bmal1 transcription in the WT MEFs after RAS induction and the longer period phenotype observed. The postulated connection between E2F activators and Bmal1 is an interesting topic that needs further investigation.
Our in silico data show that the Ink4a/Arf-dependent changes in circadian phenotypes upon RAS overexpression might be due to phase shifts in the oscillations of the core-clock and subsequently also in cell cycle components. A small phase shift between the WT and the knock-out conditions observed in the modelling simulations could also be detected in time-course measurements of the core-clock gene Per2.
In addition, an in silico analysis that introduced RAS overexpression at different time points showed that the time of the perturbation might be essential for the differing effects on the period phenotype—a finding that deserves further experimental investigation and could have consequences regarding chronotherapy research [16], but goes beyond the scope of this study. By independently removing the ARF/MDM2/p53 pathway and the INK4a/RB1/E2F pathway from the model, we show that the presence of both pathways is necessary to reproduce the experimentally observed RAS-mediated dysregulation of the circadian phenotype. Although components of both modules oscillate in the WT simulations, the modular analysis showed that circadian oscillations of the INK4a/RB1/E2F module are sufficient to simulate the observed RAS-dependent changes in the period phenotypes. This indicates that the INK4a/RB1/E2F1 pathway is necessary for maintaining the rhythmicity of the system and that its connection to the clock might be rate limiting to control cell proliferation.
Our mathematical model describes a free-running clock, which is consistent with the experimental approach in which MEFs are placed under constant conditions. As a potential future perspective, it would be interesting to investigate how our conclusions would vary when cells are subjected to synchronising cues such as metabolic signals linked to the feeding/fasting cycles, e.g., by integrating our clock-cell cycle model with a recent model of the mammalian liver clock and metabolic sensors [72].
The tumour suppressor role of INK4a has been reported in different studies, e.g., Ink4a-deficient mice develop spontaneous melanomas and are more susceptible to carcinogens than WT mice [73]. Moreover, the tumour suppressor p53 is known to modulate Per2—a clock gene that interacts with NONO and Ink4a—therefore, it is likely that these components are relevant for controlling clock-dependent cell proliferation [30,39]. To get a deeper insight into the consequences of disturbing the coupling between the cell cycle and the core-clock, we performed a whole-genome analysis using MEFs from Ink4a/Arf-/- mice and their WT littermates.
The genome-wide PCA of the expression data yields a separation of the Ink4a/Arf+/+ MEFs and Ink4a/Arf-/- MEFs into four main groups. The knockout of Ink4a/Arf has the strongest effect, followed by the perturbation by RAS. The clustering of a smaller gene set derived from genes of the mathematical model yields a similar separation of Ink4a/Arf+/+ MEFs and Ink4a/Arf-/- MEFs as does the genome-wide PCA. This indicates that the genome-wide differences caused by RAS overexpression and Bmal1 downregulation are well mimicked in our interacting subset of clock and cell cycle genes, and reinforces the postulated interconnection between the core-clock and the selected cell cycle elements. To explore the putative effect of these correlations on the circadian system, we determined the interconnections between the 50 topmost differentially expressed genes and a set of circadian-regulated genes, recently published by our group [64]. The resulting network reveals a strong interplay of the two sets via protein–protein and protein–nucleic acid interactions, which are complemented by an additional third set of newly retrieved connecting elements. Remarkably, components of our mathematical model are distributed among the set of connecting elements, the topmost differentially expressed genes and the set of circadian-regulated genes, emphasising the strong coupling between the circadian system and cell cycle components. This synergy is further reinforced by the pathway enrichment analysis of the whole network, which highlights several cell cycle-related terms as well as cancer-related effector pathways.
In our MEF model system, the overexpression of the oncogene RAS leads to an increased expression of the core-clock gene Bmal1. Even in the Bmal1 knock-down condition, the Bmal1 levels increase upon RAS overexpression, though they are still lower than in the WT, indicating a direct influence of RAS on clock components. Here, it is important to notice that Bmal1 may influence biological processes in different ways: by disrupting circadian rhythms, and/or by influencing the expression of genes, which are direct targets of CLOCK/BMAL1 as a result of other potential, noncircadian functions of Bmal1. Therefore, the output results of Bmal1 knock-down phenotypes must be seen in light of this wider perspective.
We further observed Ink4a/Arf-dependent changes in the expression levels of core-clock and cell cycle-related genes: RAS overexpression in combination with shBmal1 leads to an increase of Cry2 and Per2 when Ink4a/Arf is present, whereas it results in its downregulation in the knock-out MEFs. The concurrent perturbations (overexpression of RAS and simultaneous knockdown of Bmal1) lead to an additive effect on the expression levels of genes of the Per family. This indicates that Ink4a/Arf may be involved in regulating the gene expression of core-clock genes upon RAS overexpression, including but not limited to members of the Per family. This correlates with changes in the expression levels of the cell cycle-related gene Rb1 that is also up-regulated in the Ink4a/Arf+/+ shBmal1+RAS MEFs and downregulated in the corresponding knock-out condition. We speculate that upon RAS-induced dysregulation of the core-clock, the expression level changes of the clock gene Per2 might lead to the differential regulation of Rb1. In addition, the predicted loss of oscillations of p53 upon Ink4a/Arf knockout could account for the downregulation of Rb1 in the knock-out MEFs and be the cause for the observed experimental variations in Rb1 gene expression upon perturbations by RAS and shBmal1. Ultimately, this could result in abnormal cell fate decision phenotypes, including cell-cycle arrest, which would also be in agreement with published data [74] and points to a correlation between the core-clock gene Bmal1 and cellular senescence.
Based on a curated list of senescence-related genes derived from the literature and on our results from β-Gal-staining and growth measurements, we used a machine learning approach to classify the cell cycle fate decision of our experimental model system upon perturbations by RAS overexpression and dysregulation of the circadian system.
While in the Ink4a/Arf+/+ MEFs overexpression of RAS leads to oncogene-induced senescence, the Ink4a/Arf-/- MEFs are classified as proliferating despite the up-regulation of the oncogene RAS. We validated the conditions that were predicted to proliferate by cell-cycle analysis and compared them to the WT and the senescent conditions. The downregulation of Bmal1 shows an enhancing effect of the Ink4/Arf knockout-induced proliferation phenotype, resulting in the highest percentage of cells in the S phase and lowest in G1 phase.
Altogether, our experimental and modelling analyses strongly point to a role for Ink4a/Arf in mediating the clock connection to important cell cycle checkpoints and controlling its response upon RAS oncogenic signalling. This led us to propose a model for the interplay between the cell cycle and the circadian clock that highlights Ink4a/Arf as an important regulator for the period-changing effect of RAS on the circadian clock as depicted in Fig 6. The Ink4a/Arf locus seems to act as a switch for the effect of RAS. If the Ink4a/Arf genes are functional while the system is perturbed by the oncogenic signalling, the period increases and the cells receive a senescence signal (Fig 6A). If on the contrary, Ink4a/Arf genes are impaired, then the period of the system decreases upon RAS perturbation, and the cells receive a proliferation signal (Fig 6B). Whereas some studies have questioned the role of the clock as a tumour suppressor [13], our data is in line with several publications including a recent work in which enhancing circadian clock function seems to control cancer progression [75], and points to clock dysregulation as a tumourigenesis-promoting factor. It is conceivable that the tumourigenesis promoting/inhibiting properties of the circadian clock are context dependent, and to unravel the full complexity of such interactions will require a tumour-specific circadian analysis.
Our combined theoretical and experimental approach using a primary cell model system interrelates high-throughput data with mathematical modelling, which together provide compelling evidence for the control of both cellular oscillators, the cell division cycle, and the circadian clock via the oncogene RAS. Furthermore, our data points to the existence of a fine-tuning system of circadian regulation via Ink4a/Arf, as a response to RAS overexpression with severe consequences on cell fate decisions.
To see whether our results hold true for other systems as well, we analysed the expression of Bmal1 and Ink4a/Arf in several publicly available human cells, e.g., fibroblasts and CRC cell lines with RAS overexpression or Bmal1 downregulation. We found that RAS overexpression leads to increased levels of Ink4a/Arf and Bmal1 in the senescent human fibroblast cell line IMR-90 as compared to their proliferative counterparts. In the CRC cell line SW480 (derived from the primary tumour), Bmal1 downregulation by shRNA lead to an increase of cells in S-phase and a decrease of cells in G1-phase resembling the cell-cycle phenotype observed in the metastasis-derived cell line SW620, in which Bmal1 was already downregulated independent of the shRNA.
Based on our results, it seems to us that the clock is likely to act as a tumour suppressor, and that it is of advantage for cancer cells to circumvent circadian control. One cannot stop wondering whether disrupted circadian timing should be included as a next potential hallmark of cancer.
The semiquantitative mathematical model was derived by using the same approach as in our previously published model of the mammalian circadian core-clock [1,32]. It contains 46 variables and 170 parameters. A detailed representation of the underlying network is shown in S2 Fig. The model comprises a set of 45 ODEs, which were implemented using Matlab R2015a (MathWorks, Natick, MA), with ODE45, a built-in solver for non-stiff differential equations by using a Runge-Kutta method. Both the relative error tolerance and the components of the absolute error tolerance vector were set to 10−9. An integration step of 0.01 was used. The system of equations was assembled by using Hill-type kinetics, mass action kinetics, and Michaelis-Menten kinetics. The model was parameterised following extensive literature research. The remaining free parameters were either determined analytically or fine-tuned to fit known core-clock phase relationships (delays) for the WT scenario. Refer to S1 Text for detailed lists of model variables, parameters, and equations.
HEK293T cells (human, kidney, ATCC Number: CRL-11268) were seeded in a 75 cm2 cell culture flask and transfected with 6 μg of psPAX packaging plasmid, 3.6 μg pMD2G envelope plasmid, and 8.4 μg of either Bmal1-promoter-driven luciferase expression plasmid, pLKO.1 shBmal1 or nonsilencing shBmal1 control (Dharmacon Inc., Lafayette, CO) using the CalPhos-mammalian transfection kit (BD Biosciences, San Jose, CA). Virus particles were harvested and spun at 4000 x g for 15 minutes to remove cell debris. Supernatant was passed through a 0.45 μm filter (Sarstedt Group, Newton, NC) and used for lentiviral transduction. WT and KO primary MEFs were isolated from embryos as previously described [76] and cultured in DMEM (Gibco Laboratories, Gaithersburg, MD) containing 10% fetal calf serum (Sigma-Aldrich, St. Louis, MO) and 1% penicillin-streptomycin (Merck Millipore, Burlington, MA). Cells were incubated at 37°C at 3% O2, 5% CO2. MEFs were transduced with 1.5 ml virus filtrate plus 8 μg/μl protamine sulfate (Sigma-Alrich) in 35 mm dishes. After 1 day, the medium was replaced with selection medium (Bmal1:Luc hygromycin 100 μg/ml, shBmal1 puromycin 10 μg/ml).
SW480 (human, colon, ATCC Number: CCL-228) and SW620 (human, colon, ATCC Number: CCL-227) cell lines were maintained in DMEM low glucose (Lonza Group, Basel, Switzerland) culture medium supplemented with 10% FBS (Life Technologies, Durham, NC), 1% penicillin-streptomycin (Life Technologies), 2 mM Ultraglutamine (Lonza Group), and 1% HEPES (Life Technologies). Cells were incubated at 37°C in a humidified atmosphere with 5% CO2.
A TRC lentiviral shRNA glycerol set (Dharmacon) specifically for Bmal1 was used consisting of five individual shRNAs. The construct that gave best knock-down efficiency was determined by gene expression analysis and used for further experiments. The transduction was carried out as indicated above for the MEFs.
IMR90 human diploid fibroblasts (HDFs) (human, lung, ATCC Number: CCL-186) were cultivated in DMEM medium, 10% FBS, 100 U/ml penicillin-streptomycin, and transduced with retroviral construct pBabe-Ras-BSD or pBabe-empty-BSD as control. After blasticidine selection, cells were harvested at different days and cell pellets were frozen in −80°C for RNA extraction.
Low-passage Phoenix cells were grown in a 10-cm petri dish to a maximal density of 70%. 20 μg MSCV-Ras-BSD plasmid (MSCV Retroviral Expression System; Clonetech Laboratories, Mountain View, CA, and human hRas cDNA, Dharmacon), 15 μg helper plasmid, and 62.5 μl CaCl2 were mixed and adjusted with sterile water to 500 μl. Subsequently, 10 ml DMEM medium containing 25 μM chloroquine and the precipitate was added. After 12 h of incubation, virus particles were collected. MEF cells were seeded at subconfluent density 12 h after Phoenix cell transfection. The first virus supernatant was harvested within 12 h after transfection by aspiration and filtered through a 0.45 μm filter. The virus supernatant including 4 μg/ml polybrene was added to the MEFs after medium change. After 12 h of incubation, the second virus supernatant was harvested as described above, supplemented with polybrene and added to the MEFs. After spinoculation of the plates (1,500 rpm, 10 min, 32°C), cells were grown until the next round of transduction. The third and fourth virus supernatants were collected 12 h and 24 h later according to the same procedure. Following the last transduction, selection medium containing 10 μg/ml blasticidin was added and cells were selected. The protocol underlying the animal work, in compliance with Federation of European Laboratory Animal Sciences Association’s (FELASA) guidelines, was reviewed and approved by the Landesamt für Gesundheit und Soziales (LAGeSo), Berlin, with the identification number G105/10.
For RAS-inhibition assays, 2 × 105 cells were plated in 35 mm dishes and cultured overnight. On the next day, the cells were synchronised by a single pulse of 1 μM dexamethasone (Sigma-Aldrich) for 45 min after which an MEK inhibitor (UO126 final concentrations 10 μM; Promega Corporation, Durham, NC) was added to the culture together with 2 ml of reporter medium containing luciferin (250 μM final concentration; PJK GmbH, Kleinblittersdorf, Germany). Luciferase activity was monitored for five days using a photomultiplier tube (PMT)-based device (LumiCycle; Actimetrics Inc., Evanston, IL). The Chronostar software was used for data analysis [77].
For RAS-overexpression assays, 2 × 105 cells were transduced with a construct encoding H-RASG12V as described previously [41] were plated in 35 mm dishes and cultured in medium containing 4-hydroxytamoxifen (4OHT) (Sigma-Aldrich) was used at 1 nM, 10 nM, and 100 nM for 5 days. On the sixth day, cells were synchronised by a single pulse of 1 μM dexamethasone (Sigma-Aldrich) for 45 min. Individual dishes were washed with phosphate-buffered saline, and 2 ml of reporter medium containing luciferin (250 μM final concentration, PJK GmbH) was added per dish, after which 4OHT was added again to the culture in the concentrations as previously described. Ethanol was used as a solvent control.
Ink4a/Arf+/+ and Ink4a/Arf-/- MEFs (from five WT mice and KO littermates) harbouring the Bmal1-promoter-driven luciferase reporter construct (two independent transductions per condition) were synchronised with dexamethasone (1 μM final concentration, Sigma-Aldrich) for 45 min. Individual dishes were washed with phosphate-buffered saline, and 2 ml of reporter medium containing luciferin (250 μM final concentration, PJK GmbH) was added per dish. Luciferase activity was monitored for five days using a PMT-based device (LumiCycle, Actimetrics). The Chronostar software was used for data analysis [77].
MEF cells were seeded in 6- or 12-well plates containing 5 mm round glass cover-slips and incubated under standard conditions overnight to let the cells attach. Afterwards, medium was removed and cells were fixed in freshly prepared fixation solution (0.25% Glutaraldehyde, 2% Paraformaldehyde in PBS containing 1 mM MgCl2) at room temperature. After 15 min incubation, fixation solution was removed and cells were washed twice in phosphate-buffered saline. Staining solution was added and plates were transferred and incubated in a humidified atmosphere at 37°C for 16 h. Cells were gently washed with PBS and mounted afterwards. On each slide 200 cells were analysed, and the blue-stained cells were counted as positive.
ChIP-qPCR assays were performed using the iDeal ChIP-seq kit (Diagenode s.a., Liège, Belgium) with a primer pair specific for the promoters of Bmal1. Primer sequences are available upon request.
Total RNA from MEFs was isolated with the RNeasy Mini kit (Qiagen, Hilden, Germany) following the manufacturer instructions. Microarray hybridisation was carried out by the Labor für funktionelle Genomforschung (LFGC, Charité—Universitätsmedizin Berlin) using Affymetrix Mouse Exon 1.0 ST arrays. For RT-qPCR analysis, the extracted RNA was reverse transcribed into cDNA by using random hexamers (Eurofins MWG Operon, Huntsville, AL) and Reverse Transcriptase (Life Technologies). RT-qPCR was performed using mouse QuantiTect Primer assays (Qiagen) and SYBR Green fluorescence assays (Life Technologies). For the detection of human Ras expression, a human Quantitect Primer assay (Qiagen) was used. RNA was analysed with a real-time PCR System (Applied Biosystems, Foster City, CA). For selected genes, we harvested MEFs around one circadian cycle in three-h intervals and performed a time-course RT-qPCR analysis. Gene expression levels were normalised to mouse Gapdh mRNA (Eurofins MWG Operon, fwd: ACGGGAAGCTCACTGGCATGGCCTT rev: CATGAGGTCCACCACCCTGTTGCTG). The log2-fold change in expression of the target genes in the different conditions (Ink4a/Arf+/+/Ink4a/Arf-/- MEFs with and without Ras and Ink4a/Arf+/+/Ink4a/Arf-/- MEFs with and without shBmal1) in relation to the control (Ink4a/Arf+/+) was calculated as ΔΔCт. For the time-course analysis, the log2-fold change was calculated in comparison to the expression at 0 h. The cosine function fitted to the time-course data was calculated with the R package HarmonicRegression [78].
For the human cell lines SW480, SW620, and IMR-90 RT-qPCR was performed by using human QuantiTect Primer assays (Qiagen) and SYBR Green (Bio-Rad Laboratories, Hercules, CA) in 96-well plates. Tbp or Gapdh were used as housekeeping genes. The qPCR reaction and the subsequent melting curve were performed using a real-time PCR Detection System (Bio-Rad). Cт values were determined by using the regression method. The log2-fold change in gene expression in relation to the respective control was calculated as ΔΔCт.
MEF cells (from three mice and KO littermates), SW480, and SW620 cells used for cell cycle evaluations were labelled with 10 μM of BrdU/1x106 cells for 60 min, then collected and washed with 1 x PBS and fixed with ice-cold 80% ethanol. After DNA denaturation with 2 N HCl/Triton x-100, samples were neutralised with 0.1 M Na2B4O7, washed with PBS, and stained with 3 μL of anti-BrdU-FITC (BD Biosciences Clone B44), in a 1 x PBS solution containing 0.5% Tween20, 1% BSA, and 10 mg/mL of RNase (AppliChem GmbH, Darmstadt, Germany) for 1 h at room temperature. Supernatant was removed, cells resuspended in 200 μL of 1xPBS containing 5 μg/mL of PI (Sigma-Aldrich), and read in FACSCabilur (Becton Dickinson, Franklin Lakes, NJ). The cells of interest were gated based on forward scatter/side scatter (FSC versus SSC) values (S4 Text). The cell cycle analysis was conducted by fitting a univariate cell cycle model to the previously gated population using the Watson pragmatic algorithm [79] as implemented in FlowJo v10.2 (FlowJo, LLC, Ashland, OR).
The microarray data was analysed in R version 3.2.3 using the oligo package [80]. Expression levels of genes were calculated using the Robust Multi-Array Average (RMA) preprocessing procedure [81]. Mouse exon arrays were annotated by using the moex10sttranscriptcluster.db package. Heatmaps of clustered gene expression were generated by using the ComplexHeatmap package [82]. Hierarchical clustering with Pearson correlation was used to build the heatmaps. The R packages arrayQualityMetrics [83], affycoretools, and ReportingTools [84] were used for quality control and statistical testing of the arrays (S2 Text).
The microarray dataset has been deposited in the ArrayExpress database at EMBL-EBI (www.ebi.ac.uk/arrayexpress) under the accession number E-MTAB-5943. The calculation of the principal components was performed with built-in functions of R and visualised with the scatterplot3d package. Differentially expressed genes were determined with the limma package [85]. Analysis of the statistical significance of the resulting clusters was performed with the SigClust package [86].
We also tested the expression of our genes of interest (Ink4a/Arf, Bmal1, HRas) in publicly available microarrays of different human cell lines: human primary fibroblasts (IMR-90) with stably expressing H-RasV12 and shRNA against ERK2 or a nontargeting shRNA (GEO-GSE33613), and the CRC cell lines RKO, LIM1215, Caco2, HCT116, SW403, HT29, COLO205, SW480, and SW620 (GEO-GSE46549).
Gene and protein regulatory networks were generated using Cytoscape version 3.4 [87]. Additional information for the network was retrieved from the IntAct database of the EBI-EMBL [62] by mapping the NCBI Gene ID to Ensembl Gene IDs and Uniprot IDs. Pathway enrichment was performed with consensuspath.db. Random network analysis was performed with the iRefR package for R and iRefIndex version 14.0 with the IntAct database.
In order to predict the senescence/proliferation behaviour of the different experimental conditions, a machine learning analysis was performed on a subset of the microarray data with the SVM algorithm of the e1071 R package. The dataset consists of 32 manually curated genes associated with senescence [66–69]. The conditions used as the training set were Ink4a/Arf+/+, Ink4a/Arf+/++RAS, and Ink4a/Arf-/-+RAS.
|
10.1371/journal.pntd.0002945 | Exploring the Relationship between Access to Water, Sanitation and Hygiene and Soil-Transmitted Helminth Infection: A Demonstration of Two Recursive Partitioning Tools | Soil-transmitted helminths (STH) – a class of parasites that affect billions of people – can be mitigated using mass drug administration, though reinfection following treatment occurs within a few months. Improvements to water, sanitation and hygiene (WASH) likely provide sustained benefit, but few rigorous studies have evaluated the specific WASH components most influential in reducing infection. There is a need for alternative analytic approaches to help identify, characterize and further refine the WASH components that are most important to STH reinfection. Traditional epidemiological approaches are not well-suited for assessing the complex and highly correlated relationships commonly seen in WASH.
We introduce two recursive partitioning approaches: classification and regression trees (C&RT) and conditional inference trees (CIT), which can be used to identify complex interactions between WASH indicators and identify sub-populations that may be susceptible to STH reinfection. We illustrate the advantages and disadvantages of these approaches utilizing school- and household-level WASH indicators gathered as part of a school-based randomized control trial in Kenya that measured STH reinfection of pupils 10 months following deworming treatment.
C&RT and CIT analyses resulted in strikingly different decision trees. C&RT may be the preferred approach if interest lies in using WASH indicators to classify individuals or communities as STH infected or uninfected, whereas CIT is most appropriate for identifying WASH indicators that may be causally associated with STH infection. Both tools are well-suited for identifying complex interactions among WASH indicators.
C&RT and CIT are two analytic approaches that may offer valuable insight regarding the identification, selection and refinement of WASH indicators and their interactions with regards to STH control programs; however, they represent solutions to two distinct research questions and careful consideration should be made before deciding which approach is most appropriate.
| Soil-transmitted helminths (STH) are pervasive enteric parasites that lead to cognitive, nutritional and educational sequelae. Mass drug administration is employed to reduce morbidity, but reinfection occurs rapidly in the absence of changes to other environmental conditions, such as improvements to water, sanitation and hygiene (WASH). Since WASH behaviors and conditions are highly interrelated, typical epidemiological methods are limited. Few rigorous studies have assessed the impact of WASH components as they complement deworming and even fewer have sought to prioritize among the available indicators or identify complex interactions. In this paper we introduce two recursive partitioning approaches: classification and regression trees (C&RT) and conditional inference trees (CIT). We demonstrate these two tools using data from a school-based cluster-randomized trial conducted in Kenya. We discuss the advantages and disadvantages of each tool and give examples of how they may be used to improve STH control programs.
| Infection with soil-transmitted helminths (STH), intestinal nematodes, is classified by the World Health Organization (WHO) as a neglected tropical disease (NTD). More than 1 billion people are infected and up to 5.3 billion are at risk of infection with at least one species of STH, including roundworm (Ascaris lumbricoides), whipworm (Trichuris trichiura), or hookworm (Necator americanus or Ancylostoma duodenale) [1]–[3]. STH infection occurs through fecal exposure, either through the skin in contaminated soil (in the case of hookworm) or ingestion of fecal material, typically in soil, on food, or on fingers [4]. Morbidity is most acute in school-age children, though high levels of hookworm infection can persist into adulthood [4]. It is estimated that between 5 and 39 million disability adjusted life years are lost due to STH infection [5], [6].
Though a recent review found limited evidence [7], STH infections have been found to impact on growth and nutrition of children [8] and reduce pupil absence in some studies [9], [10]. Control of STH is a priority for the WHO [11] and several countries, including Kenya, are scaling up mass drug administration in school-age children to reduce STH-related morbidity [12], [13]. These infections can be treated safely and effectively with the anthelminthic drugs albendazole or mebendazole [4], [14]. However, in the absence of improved access to water, sanitation, and hygiene (WASH), reinfection occurs and the prevalence and intensity of infection can reach pre-treatment levels in as few as six months, with 94% reinfection after 12 months [15]. Access to WASH includes hardware – such as toilet facilities that separate human feces, protected water supply, and soap – as well as behaviors, such as hand washing at key times and toilet use. The UNICEF and WHO Joint Monitoring Program (JMP) is the most widely cited source of data on what is considered “improved” water supply and sanitation [16], but the JMP does not provide guidance on hand washing, nor are its definitions specific to STH control. Even in countries with moderate access to improved water and sanitation in sub-Saharan Africa there is considerable geographic inequity [17]; these same marginalized populations without access are the ones with high risk of STH [18].
WASH components thought to be most critical for control of STH are the use of a clean toilet facility and the presence of water and soap for hand washing; however, few randomized trials have been conducted to assess the relationship between WASH and STH infection. Three randomized controlled trials have found evidence that improved hand washing with soap can lead to lower STH infection [19]–[21]. Nonetheless, in a study by Dumba et al., researchers did not find any impact of a participatory hygiene and sanitation transformation (PHAST) intervention compared with a control group that received deworming alone [22].
A recent meta-analysis of 36, mostly observational, studies suggested that access to and use of sanitation facilities is associated with significant reductions in the prevalence of STH infection, with an odds ratio [OR] of 0.54 (95% CI: 0.43–0.69) for A. lumbricoides, 0.58 (95% CI: 0.45–0.75) for T. trichiura, and 0.60 (95% CI: 0.48–0.75) for hookworm [23]. In a separate meta-analysis, soap use (OR: 0.53, 95% CI: 0.29–0.98), wearing shoes (OR: 0.38, 95% CI: 0.18–0.81) and drinking treated water (OR: 0.45, 95% CI: 0.36–0.58) were associated with lower STH infection [24]. Access to piped water was associated with lower infection with A. lumbricoides (OR:0.39, 95% CI: 0.39–0.41) and T. trichiura (OR: 0.57, 95% CI: 0.45–0.72). However, because nearly all studies in these meta-analyses were observational, it was not possible to disentangle the impacts of individual WASH components or the relationship between WASH and socio-economic status, potentially biasing many of these results.
WHO has set the goal of elimination of STH as a public health problem by 2020, which is provisionally defined as a prevalence of moderate- and high-intensity STH infection of <1% (WHO, 2012). To achieve this goal, and to sustain the gains made possible through mass drug administration, WASH improvements and intersectoral collaboration will be critical [11], [25]. However, identifying and characterizing those WASH components that are most effective at reducing or preventing STH infection is non-trivial, in part because of the ethical challenges of conducting randomized control trials which are necessary for establishing causal relationships [24], and yet will be essential for developing evidence on the success of STH control programs [18].
One challenge is that access to the different components of WASH in both the public and private sphere is highly interrelated, and little is known about the relative contributions of each independent WASH component in mitigating infection with STH. Furthermore, readily measurable WASH components relevant for STH control have not been identified or validated. Indeed, current WHO guidelines for STH control refer to WASH in general terms [11], [26]. The vast majority of studies examining the association between WASH components and STH infection have considered the main effects; however, because of the inherent connectedness of WASH components – e.g. water must be present for hand washing to occur – it is also critical to consider interactions. The number of potentially measurable WASH components is quite large, and when one also considers all the potential first, second-, and higher-order interaction terms, most datasets would not have sufficient power to detect all important associations using standard analytic approaches. A need exists to identify alternative analytic approaches to help identify, characterize, and further refine those WASH components that are most important to STH infection.
The goal of this analysis is to introduce two analytic approaches that are relatively new to the NTD and WASH communities: classification and regression trees (C&RT) and conditional inference trees (CIT). Both C&RT and CIT are a type of recursive partitioning, a nonparametric analytic approach well-suited for handling datasets with large numbers of predictor variables, identifying complex interactions, and selecting independent variables that are most predictive of or associated with the outcome [27]. These approaches are particularly useful for hypothesis generation and as a precursor to other model building approaches. We demonstrate how both methods can be applied to a dataset measuring household- and school-level WASH components and STH infection in Kenyan school children. This is a secondary analysis of the data; the primary results from this study have been reported elsewhere [9]. We discuss the relative merits and weaknesses of each approach and make recommendations for their uses.
Data collection for this study was approved by the Institutional Review Board at Emory University and the Ethics Committee at Great Lakes University of Kisumu (Kenya). We obtained a loco parentis from the head teacher at each school. Children provided oral consent to participate in this study, which was documented on the electronic data collection form. The ethics committees approved both a waiver of parental consent and the use of oral consent for study participants.
This study utilized data from a cluster-randomized trial to assess the impact of improved school and household WASH access on STH infection in Nyanza Province, Kenya from 2007–2009 [28]. Data for this analysis were collected in February 2009 – the final survey round of the trial – from 1,106 students in 39 public primary schools (Checklist S1: STROBE Checklist). Twenty of these schools had been randomly selected to receive a school-based WASH intervention that included construction of ventilated-improved pit latrine facilities at the school, hand washing and drinking water storage containers, teacher training on hygiene behavior change, and a one-year supply of dilute sodium-hypochlorite used for treatment of drinking water at the point of use. Pupils in all schools – both intervention and control – were dewormed at baseline (May, 2007) and midterm (April, 2008) using 400 mg of albendazole.
Pupils from grades 3 to 5 who were between the ages of 7 and 13 and had been dewormed during the previous round of data collection were randomly selected and enrolled into the original trial. The mean number of pupils was 302 and 275 in the intervention and control schools, respectively. This age group was selected because they experience the greatest burden of A. lumbricoides and T. trichiura, though peak morbidity for hookworm occurs later [29]–[31]. Systematic random selection of 30 pupils was conducted using a list of pupils from the school records, though some pupils were absent the day of the study. Only one child per household was enrolled to avoid the need to adjust for intra-household correlations. Of the 1106 students included in the study, 1095 provided a single analyzable stool and had valid Kato-Katz results. The original sample included pupils from 40 schools (20 intervention and 20 control). However, one control school was dropped from the analysis after children were treated with an additional round of deworming drugs.
Stool samples were collected and transported to the laboratory in cool boxes and examined microscopically within one hour of preparation using the Kato-Katz method [32]. Each stool sample was processed on two separate slides and read by different laboratory technicians to ascertain the eggs per gram of each STH species. Presence of infection was defined as detection of one or more eggs on either slide. Because all individuals were dewormed 10 months prior to the study, any infection observed was interpreted as incident infection. This analysis includes data on pupils from both intervention and control arms.
Data on individual demographics, household WASH conditions, and school WASH conditions were collected using structured observations and questionnaires. Pupils were interviewed to determine their age, sex, shoe wearing, comfort using the latrine at home and school, knowledge on hand washing and water supply treatment, opinion about latrine conditions, access to hand washing and drinking water at school, and their soil eating behavior (known as pica or geophagy), a common practice in western Kenya [33]. As a complement to the direct observations made at each school, pupil responses regarding school-based access to drinking water, hand washing water, soap, and latrines were aggregated at the school-level as an estimate of school WASH access.
One caregiver – typically the maternal head of household – for each pupil enrolled in the study was interviewed in his or her home to determine if one or both parents was alive and, if alive, the highest level of education achieved, socio-economic status through an asset index; access to an improved drinking water source, as defined by UNICEF and WHO [34]; if treated water is used; and the presence and condition of a household latrine. School head teachers were interviewed about the school's access to an improved water source during the dry season, pupil to latrine ratio, and latrine conditions. Enrollment data were taken from official school records.
In order to calculate socio-economic status, we used a principal component analysis (PCA) using assets observed at the household [35]. These assets included household construction materials, ownership of goods such as a TV and radio, and connection to electricity [36]. PCA was also used to construct an index of sanitation conditions at both the household and school, which included odor, presence of flies, presence of feces, wall material, condition of the slab, and presence of a functioning door. These components were put on a scale from 1–4 and the resulting value was a relativistic score of the average conditions for all latrines at the school or for the latrine at home. Two scores – latrine cleanliness and latrine structure – were derived based on factor loading. We did not include the score for latrine conditions at the home in our tree analysis, since that would limit the analysis to children with latrines at home. Acceptable latrines were classified as those for which no parameter scored in the lowest two values for each of the five sanitation categories.
This study compares two different recursive partitioning approaches: C&RT and CIT. Recursive partitioning is a nonparametric regression approach; it is a form of hierarchical clustering in which the data are sequentially split into dichotomous groups such that each resulting group contains increasingly similar responses for the outcome [37], [38]. Recursive partitioning has several advantages over traditional logistic regression. C&RT and CIT are supervised clustering approaches; they create partitions based on an outcome variable, as opposed to other clustering approaches such as k-means and PCA, which do not involve the outcome [39], . As nonparametric approaches, C&RT and CIT make no assumption of a monotonic or parametric relationship with the outcome, can be used to identify complex interactions among the independent variables without a priori specification of interaction terms, and can handle datasets where the number of independent variables is high relative to the number of observations. This final feature is particularly attractive to studies such as this, where a goal is to identify a few best predictors from many.
Both C&RT and CIT result in the formation of a decision tree with three levels consisting of a root node, internal nodes, and terminal nodes. Every tree starts with a “root node” that contains the sample of data from which the tree will be grown (e.g. the study population). The data are then partitioned into two “child nodes” based on the value the independent variable (IV) that best meets some partitioning criterion. The resulting child nodes each contain a subset of the original data. Each child node may be further partitioned, again based on the value of an IV. This process continues until no further partitions remain or some set of partitioning criteria are no longer met, resulting in terminal nodes. Terminal nodes, by definition, cannot have offspring. C&RT and CIT differ in the partitioning criteria used to select the IVs.
Under C&RT the data are partitioned according to the IV that results in the greatest improvement in the distribution homogeneity of the outcome [41], also referred to as reducing node impurity. Put another way, the data are split according to the IV that best improves predictive accuracy in the child nodes. The predictive accuracy of each potential binary split is considered independently and the split offering the greatest improvement is chosen to partition the data.
The initial tree generated by the recursive partitioning process of C&RT tends to be large (i.e. contain many splits of the data) and runs the risk of over-fitting the data. This motivates a second stage of tree construction called “pruning”, which can be viewed as analogous to backwards selection in linear regression. Through pruning, partitions of the data that are deemed to be the most superfluous are removed from the bottom-up. Cross-validation is then used to select the optimal sub-tree from the initial tree. In this study, C&RT analysis was performed using the ‘rpart’ package in R, version 2.13.2, available at http://cran.r-project.org/web/packages/rpart/index.html. For a more detailed description of this method see Therneau et al [42].
With CIT, the partitioning criterion is based on statistical significance and, unlike C&RT, accounts for conditional relationships between IVs. In the first step of the algorithm, the global null hypothesis of independence between all the IVs and the outcome is tested; if the null cannot be rejected, partitioning stops. If the global null hypothesis is rejected, then the IV that is the most significant in the model, conditional on the other covariates, is selected. When the selected IV is dichotomous, the choice of the best binary split is trivial; for non-dichotomous variables, the algorithm identifies the best binary split from all possible splits. Because CITs are based on statistical inference, pruning is not necessary. In this study, CIT analysis was performed using the ctree function in the ‘party’ package in R, version 2.13.2, available at http://cran.r-project.org/web/packages/party/index.html. See Hothorn et al for more information on this method [43].
All of the demographic and WASH indicators listed in Tables 1–3 were included in the analysis as IVs to be selected to partition the trees. The outcome of interest was any STH infection, coded dichotomously, with a “1” indicating the presence of at least one infection by A. lumbricoides, T. trichiura, or hookworm. Both C&RT and CIT trees were grown with the restriction that each node must have a minimum of 20 observations.
The C&RT results were validated using 10-fold cross-validation and the optimal tree was selected by pruning to the smallest tree within one standard error of the minimum cross-validated error tree [44]. Two different CITs were generated. In the primary analysis a minimum p-value of 0.05 was used for the partitioning criterion; in a secondary analysis, p-values were adjusted for multiple comparisons, using the Bonferroni correction.
Table 1 contains the prevalence of key demographic and WASH variables measured at the pupil level. Over half (575; 52%) of the pupils surveyed were boys with a mean age of 10.4 years. Approximately a third (383; 35%) of pupils were observed without shoes at school and 138 (13%) reported some form of soil eating, known as geophagy. Pupils had a mixed impression of their school latrines, with 55% and 64% reporting the latrines to be dirty and have a strong odor, respectively; while 66% reported the latrines to be “comfortable.”
Table 2 contains information regarding the household demographic and WASH characteristics. The prevalence of orphanhood was high in the households surveyed, with 12% of mothers and 33% of fathers deceased. Of the households with living mothers, only 67 (6%) had completed at least secondary education and 423 (39%) had no formal education. Nearly half of the households had an improved water source in the dry season (537; 49%). The presence of a latrine was high (63%), though few hand washing stations were observed (39%).
Information regarding school-level WASH characteristics is in Table 3. Available water for drinking and hand washing was observed at the time of interview in nearly 60% of schools. Only half of the schools had an improved water source in the dry season, while approximately 70% had an improved source in the rainy season. The frequency with which pupils reported constant availability of water for drinking and hand washing at school varied widely between schools. Most students reported that soap was not always available at school.
Of the 1095 pupils tested by Kato Katz, 18% tested positive for at least one worm infection, with 81 pupils testing positive for A. lumbricoides, 75 for hookworm and 74 for T. trichiura (Table 4). Thirty-three children (3%) tested positive for two STH species. No samples were positive for all three worm types. More detailed information on the worm burden and main effects in this data have been previously published in Freeman et al [9].
Initial attempts to generate a classification tree failed to result in anything other than the root node (the starting dataset with no partitions) after pruning. This means that after cross-validation the tree generated showed no significant improvement in predictive accuracy over the starting dataset. A second C&RT analysis was performed, this time with sensitivity weighted more heavily than specificity under the assumption that in a situation of STH control, the identification of true STH infections is likely to be prioritized over true no infections. This was achieved by setting a misclassification cost of 2∶1 for STH positive vs. STH negative infections. Typically the C&RT algorithm tries to minimize the proportion of misclassified cases, where misclassification costs are taken to be equal for every case (e.g. those individuals with both positive and negative stool examinations for STH). With the 2∶1 weighting employed, misclassified positive individuals count twice as much as misclassified negative individuals.
The 2∶1 misclassification weighting resulted in a pruned classification tree with five partitions and six terminal nodes for predicting the incidence of infection by any of the three helminths. The following independent variables appeared in the final pruned tree: “latrine structure”, “latrine cleanliness”, “% pupils reporting drinking water always available at school”, and “father status” (Figure 1). The first split of the tree was the PCA factor for school latrine cleanliness, indicating that this variable was best at classifying STH infection status in the data. For both PCA-derived school sanitation variables appearing in the final tree (latrine cleanliness and latrine structure), the C&RT algorithm found an optimal partition of the PCA scores; however, because the numeric values of these scores have no external generalizability we relabeled the dichotomous groups as having “better” or “worse” sanitation conditions.
There were six terminal nodes in the final C&RT tree. Terminal node T5 had the greatest proportion of positive cases with 19 of the 28 children positive for one or more species of helminth; this terminal node corresponds to schools with good latrine cleanliness and structure but low reported drinking water availability. The C&RT algorithm labeled terminal nodes T4 and T5 as predictive of a “positive” STH infection, while the remaining four terminal nodes were predictive of no STH infection. Note that terminal node T4 is predictive of “positive” infection despite only 41% of observations being positive because of the 2∶1 weighting favoring sensitivity over specificity.
Table 5 shows the distribution of STH infection at each pair of child nodes emanating from the internal nodes in the C&RT analysis. Based on the classification tree, pupils with “worse” latrine cleanliness scores were twice as likely to be infected with STH (30%), compared to those with “better” latrine cleanliness scores (15%). Among pupils in schools with “better” latrine cleanliness scores, those with “better” latrine structure had twice the rate of infection (24%), compared to those with “worse” latrine structural conditions (12%). For schools with better latrine structure, greater values for “% pupils reporting drinking water always available at school” was predictive of lower levels of infection. Among those schools with “worse” latrine cleanliness and greater drinking water availability at school, the pupil-level variable for father's education status was identified as the best classifier of STH infection. The optimal partition of this pupil-level ordinal variable, which distinguished deceased fathers from living fathers with various levels of education (see Table 2), occurred between deceased fathers and living fathers regardless of education status.
The primary CIT tree, generated without adjustment for multiple comparisons, had 12 terminal nodes (Figure 2). The IV most significantly associated with STH infection, conditional on all other variables in the model, was “District”. Of the 11 IVs appearing in the tree, 5 were measured at the schoollevel, 5 at the household level and 1 (“age”) at the pupil level. Five of the IVs in the final tree were demographic measures while the remaining six were WASH indicators. Terminal node T7 had the greatest proportion of positive cases with 44% (n = 12) of individuals testing STH positive. This node corresponds to pupils from Kisumu East or Rachuonyo Districts with low family SES who attend a school with high enrollment rates and little to no soap reportedly available.
Table 6 shows the distribution of STH infection for each pair of child nodes emanating from the internal nodes in the CIT analysis. In Nyando District, greater reported availability of water for hand washing in the schools was associated with a greater incidence of STH infection (34% vs. 16%, from terminal nodes T1 and T2). In the Kisumu East and Rachuonyo Districts, among those schools with high enrollment rates (>80th percentile), pupils whose household was in the lowest SES quintile had twice the incidence of STH infection (32% vs. 16%; Table 6). The IV “% pupils reporting soap always available at school” appeared twice in the tree in Figure 2 (internal nodes 8 & 10) but with opposite directions of association; at internal node 8, little to no soap availability was associated with increased STH infection (44% vs. 24%), whereas at internal node 10 lack of soap was associated with decreased STH infection (11% vs. 27%).
The second conditional inference tree, grown with p-values adjusted for multiple comparisons, is shown in Figure 3. This tree is a sub-tree of the tree in Figure 2 and represents a more conservative approach. The tree in Figure 3 has four terminal nodes, with the greatest STH incidence seen in the terminal node for pupils in Nyando District attending schools where more than 63% of the students reported hand washing water available (34% of pupils in this node were positive for STH).
The branching of the IVs in both the classification and conditional inference trees can be used to identify potential interactions between WASH indicators that may be important predictors of STH infection. The classification tree illustrates that when latrine cleanliness is better, latrine structure is an important determinant of STH infection. In this instance, worse latrine structure predicts lower STH infection (Figure 1). When both latrine cleanliness and structure are good, the pupil-reported presence of water at school is important, with poor water availability (<24% of students reporting constant water availability) associated with higher STH infection (T5, Figure 1) and more consistent water availability predicting less STH infection (T6, Figure 1). The conditional inference tree suggests that the interaction of living in the Nyando District and the reported availability of hand washing water at school are together associated with STH infection (Figure 2). Similarly, for those living in the Kisumu East and Rachuonyo Districts the CIT analysis identified an interaction between pupil enrollment, low SES and age (among other interactions present in the tree).
In the past several years classification and regression trees, first introduced by Breiman et. al. in 1984 [45], have started to gain recognition as a statistical tool in NTD research. C&RT is most commonly used in the NTD research community as a tool for disease prediction and classification [46]–[48], as well as to identify the hierarchical importance of predictor variables [49], [50]. To our knowledge there have been no studies that have used C&RT or CIT to examine WASH data in relationship to enteric disease prevalence or incidence, specifically STH. Given the policy and programmatic interest in quantifying the impact of WASH on STH, these recursive partitioning approaches may prove to be important analytic tools.
In this study we show how C&RT and CIT can be used as tools to better understand household- and school-level WASH indicators and their relationship with STH infection in children. While both C&RT and CIT result in the generation of a decision tree, the dissimilarities between Figures 1 and 2 make it clear that these two approaches are not identical. Instead C&RT and CIT represent solutions to two distinct research questions and careful consideration should be made before deciding which tool is most appropriate.
A C&RT analysis is well-suited when the goal is prediction or classification. Independent variables are chosen to partition the data according to the one that results in the biggest improvement to predictive accuracy, and not according to association with the outcome, as in the CIT analysis. The difference in predictive performance between these two types of trees is apparent when comparing the sensitivity and specificity of the classification tree (Figure 1) with the conditional inference tree (Figure 2), where observations in each terminal node of the CIT are classified according to the majority of observations in that node (e.g. if ≥50% of pupils in the terminal node are STH positive, then all individuals in that node will be classified as “positive”). In Figure 1, of the 197 children testing positive for STH infection, 62 were correctly classified by the tree, for a sensitivity of 31%; of 898 children who tested negative, 828 were correctly classified, resulting in a specificity of 92%. By contrast, in Figure 2, all of the terminal nodes had fewer than 50% positive cases, resulting in a sensitivity of 0% and a specificity of 100%. It is important to keep in mind that the C&RT results were generated with a preference towards sensitivity (using a 2∶1 misclassification cost), which is why terminal node T4 is classified as “positive” despite only 41% of cases being positive. This weighting was necessary to obtain a final pruned tree that was more than just the root node. If we ignore this weighting and classify node T4 as “negative” according to the majority of observations, then the sensitivity of the C&RT tree falls to 10% and the specificity grows to 99%.
While one might expect CIT results to have poorer predictive accuracy, the poor predictive accuracy of the C&RT results is surprising and may have more to do with our data than the method itself. Firstly, the dataset used had relatively few positive cases (18% positive overall), making classification more difficult. Secondly, although infection was measured at the pupil level, nearly half of the IVs (12 of 27) were school level. While the presence of school-level IVs in the final tree could mean that school-level factors may play a greater role in driving STH infection than household factors, it also leads to some major drawbacks. In Figure 1, terminal nodes T1, T2, T5 and T6 are not classifying STH infection beyond the school level; a predicted STH status of “positive” or “negative” is assigned to all students in the same school because there are no individual-level WASH factors to further differentiate pupils within the same school. Only among the subgroup of pupils in Figure 1 for whom “father status” is predictive of STH infection are we able to classify infection at the pupil level. As a result, the sensitivity of the STH tree is limited by the sensitivity obtained at the school level. Taken together, the scarcity of positive cases and use of school-level variables likely explain the failure of the initial C&RT analysis to identify anything beyond the root node. A further limitation of using data clustered at the school level is that to our knowledge there is no way to account for the design effect (decrease in the sample variance due to clustered sampling) in either the C&RT or CIT analysis packages. This may affect the validity of the CIT results if the design effect varies by IV.
C&RT may be the preferred approach if interest lies in using WASH indicators to classify individuals, or communities, as STH infected or uninfected. Under this framework, a representative dataset would be used to grow a classification tree (as was done in this analysis), which could then be used to predict the STH status of future data collected from individuals or communities. One potential application of this approach is the development of a rapid screening tool for classifying communities as “likely endemic”, “likely sub-endemic” or “likely non-endemic” for STH based on the values of the set of IVs in the classification tree. Such an approach might lessen the initial need for specimen collection and conserve resources.
When the goal is to identify the IVs most associated with the outcome, in order to estimate causal effects, conditional inference trees are the more useful tool. At each branch of a conditional inference tree the IV that is most significantly associated with the outcome is chosen to partition the data, resulting in a tree built on statistical significance. An advantage of CIT over traditional regression is that, as a nonparametric approach, it can identify non-linear associations in the data. A CIT analysis can be used as a form of variable selection, identifying the few IVs that result in the most significant main and joint effects. Such an approach would be useful when trying to determine which of the many WASH indicators to include in an intervention package in order to see the greatest impact on STH incidence.
Both C&RT and CIT are useful as exploratory analytic tools for identifying complex interactions in the data. Using standard analytic approaches to assess interaction often requires including the product of two or more predictors in the same model and assessing whether the resulting coefficient differs significantly from zero. As the number of potential predictors increases, the number of potential second-, third- and higher-order interactions grows, becoming too large to include in any one model without substantial a priori knowledge about complex interactions [51]. A C&RT or CIT analysis allows one to move beyond the commonly reported main effects, to identify complex interactions (e.g. non-linear, multi-order interactions) in the data. Often the interactions identified by the branching of the tree are ones that might not have been identified a priori by the researcher (e.g. the interaction of district, pupil enrollment, low SES and age found in Figure 2) and can be used to generate hypotheses leading to further research. These interactions can also be used to identify potentially vulnerable sub-populations. For example, the tree in Figure 2 suggests that children in Nyando District who attend schools with greater reported prevalence of hand washing water have a greater risk of STH infection – a surprising and somewhat paradoxical finding. It is important to interpret any interactions in the context in which they were generated, namely for prediction (C&RT) or association (CIT).
As with any analysis, it is important to consider the data and potential for biases when interpreting results. One challenge is that many WASH interventions are highly correlated. Research in other fields has shown that when two IVs are highly correlated, the effect estimate of the better measured variable will capture some of the effect of the less-well measured variable, making the overall effect estimates less accurate [52], [53]. This is of potential concern to WASH analyses where the potential for measurement error is high, due to the sensitivity of the topics and the way in which some indicators are measured (e.g. reported vs. observed measures). Examples of highly correlated variables in our analysis are “% pupils reporting soap always available at school”, “% pupils reporting hand washing water always available at school” and “intervention”. Because the intervention involved making soap and water available in some schools, these three variables are highly correlated (rho = 0.71, p<0.0001) in our data and the presence of any one of these in the final tree is likely capturing, in part, the effect of the others. One advantage of a CIT analysis is that IVs are selected conditional on all other variables in the model. That is, each partition represents the variable most strongly associated with STH infection controlling for the other IVs (i.e., WASH indicators) in the model. By simultaneously controlling for the other IVs, a CIT analysis alleviates many of the concerns with correlated data that are particularly pertinent to WASH analyses, though it does not resolve concerns about measurement error.
Another advantage of CIT over C&RT is that the former can handle independent variables with different types of classification (e.g. continuous, categorical, ordinal, etc.) without bias [43]. Studies have shown that C&RT favors continuous IVs over dichotomous or ordinal IVs [54], [55]. It may be that the appearance of “latrine cleanliness” and “latrine structure” in the first splits of Figure 1 has more to do with the fact they are among the few continuous IVs eligible to be selected for splitting and were unduly favored by C&RT. That neither of these two best classifiers in the C&RT analysis appeared in the CIT analysis is somewhat surprising (if they are truly good classifiers one would also expect them to be strongly associated with STH infection) and may instead be indicative of this bias towards favoring IVs with many possible splits. The optimal splits for “latrine cleanliness” and “latrine structure” could have been selected because the C&RT algorithm was able to identify a partition that most closely approximated some unmeasured risk factor for STH.
This is not to say that continuous variables should not be used in C&RT. Another way that C&RT is being used by the NTD community is to identify the best dichotomous cutoffs for continuous IVs. Martinez et. al used C&RT to find optimal cutoff for chromosome counts used in leprosy diagnosis while Levecke et. al utilized C&RT to find cutoffs for study design factors for the fecal egg count reduction test to monitor STH drug efficacy [56], [57]. It should be noted that CIT can also be used to identify optimal cutoffs associated with the outcome. A main effects analysis, using logistic regression, found neither of the principal components for latrine structure and cleanliness to be strongly associated with overall STH infection—which explains why they did not appear in the CIT analysis—but when dichotomized according to the cutoffs identified by C&RT (from Figures 1) the resulting indicator variables were highly significant in the regression analysis (results not shown).
While the goal of this study was primarily methodological, it is important to discuss some of the findings, particularly those that were surprising and at times counter-intuitive. In the first split of the C&RT tree, children in schools with worse latrine cleanliness had twice the incidence of infection compared to those with better latrine cleanliness (30% vs. 15%), which is in-line with our expectations. One example of a counterintuitive finding in the C&RT analysis is that among schools with better latrine cleanliness, better latrine structure is associated with greater infection. It is possible that latrines with better structure are more likely to be used, and even good cleanliness was insufficient to prevent infection. Another counter-intuitive finding from the C&RT results was that among schools with poor latrine cleanliness, poorer water availability is associated with less STH infection (Table 5, node 2). One reason for this surprising finding could be that the causal direction of association is not well-understood. Is the latrine cleanliness poor because the latrines are often used [58], or does poor cleanliness prevent the latrines from being used? Another surprising finding is that having ones' mother alive is predictive of less STH infection in the CIT analysis (Figure 2, terminal nodes T11 vs T12) while having ones' father alive is predictive of more STH infection in the C&RT analysis (Figure 1, terminal nodes T3 vs T4). While it is possible that there is a causal explanation behind this, it is also important to remember that the subsets of data used to make these split selections are likely to be quite different, based on the previous partitions leading to the nodes. The variable “% pupils reporting soap always available at school” appears twice in Figure 2 but suggests opposite directions of association. This could be due in part to the complex interactions leading up to these two terminal nodes. For example, by looking at the previous branching of the tree in Figure 2, one can see that node 10 contains schools that did not receive the intervention whereas the population in node 8 contains both intervention and non-intervention schools. It is possible that students in node 8 attending an intervention school that should have received soap, but report that it did not, are somehow more susceptible to STH infection. These, and other unexpected findings, further support our suggestion that these two approaches be used for hypothesis generating and that follow-up analyses be conducted to better understand these interactions identified in the data and determine if they are generalizable to other populations.
Another way of analyzing the trees in Figures 1 and 2 is by assessing the WASH characteristics of the IVs included at each split. In the classification tree, two of the IVs selected to partition the tree are related to water, two to sanitation and one is demographic; no indicators for hygiene appear in the C&RT tree. By contrast, 3 of the 11 independent variables in the conditional inference tree are indicators of hygiene, 1 of sanitation and 1 of water (and 1, “intervention” encompassing all three WASH components); the remaining 5 IVs selected are demographic indicators. These differences between the types of WASH indicators present in the two trees again serve to highlight the differences in the two approaches of growing trees: prediction vs. statistical significance. The trees are also suggestive of the complex relationship between demographic, water, sanitation and hygiene variables and STH infection; both trees involve many of these different variables with no single variable or class of variable emerging as the most related to STH infection.
The differences in the Figure 1 and 2 results also emphasize one of the common criticisms of recursive partitioning algorithms, that any single tree is often highly unstable [37]. With only 1095 observations in the entire dataset, tree stability is a definite concern in this analysis. Increasing sample size or incorporating multiple datasets to grow and test the trees can help to improve tree stability. Random forests is an alternative approach in which an ensemble of classification or conditional inference trees are grown and the results combined to get more stable measures of variable importance [38]; however a downside is that random forests do not result in a single tree diagram and cannot be readily used to identify complex interactions.
In addition to tree stability, the small sample size will also limit the power and generalizability of any one tree. With small datasets one runs the risk of generating trees that over fit the data, and thus any substantive findings may have limited applicability beyond the data. Cross-validation and splits based on statistical inference help to limit this concern in C&RT and CIT, respectively. Nonetheless, we caution the reader against over-interpreting the substantive findings in this paper. Moving forward, as issues related to WASH take on greater priority, we expect larger datasets to become available that will increase both the relevance and utility of these analytic approaches. A good example is the Global Trachoma Mapping Project, which, in addition to measuring trachoma prevalence, has collected data on water and sanitation indices on over half a million people [59].
A limitation of the data used in this analysis is that two of the independent variables that appeared in the final classification tree are data-specific; the indicators for latrine cleanliness and structure were created from two principal components. Because these variables are derived from the data they may not be generalizable to other studies and their numerical value has little direct interpretation or value as a metric; however, our findings regarding the direction of effect and relative importance of these characteristics remain valid. This analysis also highlights the need to arrive at methods for measuring these and other water and sanitation indicators that are feasible, replicable and generalizable.
C&RT and CIT are two analytic tools that may be of use to the NTD and WASH communities, depending on the research objective. When prediction of the outcome is the goal, C&RT is likely to be the most favorable tool, whereas CIT is good for identifying the IVs most significantly associated with the outcome. Both methods can be used to identify complex interactions in the data; however, these interactions should be interpreted in the context of the tool (e.g. prediction vs. association). These interactions can then be incorporated into subsequent parametric analyses or used to generate hypotheses for future research. This study supports the WHO's goal for STH elimination by contributing to the research on the impacts of WASH in mitigating STH infection. With the help of this and future research it will hopefully be possible to identify the WASH indicators of greatest relevance for STH control.
|
10.1371/journal.ppat.1001233 | Structural and Functional Studies of Nonstructural Protein 2 of the Hepatitis C Virus Reveal Its Key Role as Organizer of Virion Assembly | Non-structural protein 2 (NS2) plays an important role in hepatitis C virus (HCV) assembly, but neither the exact contribution of this protein to the assembly process nor its complete structure are known. In this study we used a combination of genetic, biochemical and structural methods to decipher the role of NS2 in infectious virus particle formation. A large panel of NS2 mutations targeting the N-terminal membrane binding region was generated. They were selected based on a membrane topology model that we established by determining the NMR structures of N-terminal NS2 transmembrane segments. Mutants affected in virion assembly, but not RNA replication, were selected for pseudoreversion in cell culture. Rescue mutations restoring virus assembly to various degrees emerged in E2, p7, NS3 and NS2 itself arguing for an interaction between these proteins. To confirm this assumption we developed a fully functional JFH1 genome expressing an N-terminally tagged NS2 demonstrating efficient pull-down of NS2 with p7, E2 and NS3 and, to a lower extent, NS5A. Several of the mutations blocking virus assembly disrupted some of these interactions that were restored to various degrees by those pseudoreversions that also restored assembly. Immunofluorescence analyses revealed a time-dependent NS2 colocalization with E2 at sites close to lipid droplets (LDs) together with NS3 and NS5A. Importantly, NS2 of a mutant defective in assembly abrogates NS2 colocalization around LDs with E2 and NS3, which is restored by a pseudoreversion in p7, whereas NS5A is recruited to LDs in an NS2-independent manner. In conclusion, our results suggest that NS2 orchestrates HCV particle formation by participation in multiple protein-protein interactions required for their recruitment to assembly sites in close proximity of LDs.
| Formation of infectious virus particles (assembly) is a complex process by which structural proteins and the viral genome must be transferred to the same subcellular sites to allow their direct or indirect interaction. In case of the hepatitis C virus (HCV), this process appears to take place in close proximity of lipid droplets (LDs) and requires in addition to the structural proteins core, envelope glycoprotein 1 (E1) and E2 two auxiliary factors, designated p7 and nonstructural protein 2 (NS2), contributing to virion formation by unknown mechanisms. In this study we used a combination of structural, genetic and biochemical assays to study the role of NS2 in HCV assembly. By using nuclear magnetic resonance spectroscopy of NS2 peptides we established a membrane topology model of the amino-terminal membrane binding domain of NS2. We found that this protein participates in multiple interactions with E2, p7, NS3 and NS5A that appear to recruit the viral proteins to sites in close proximity of LDs. In this respect, NS2 is a key organizer of the assembly of infectious HCV particles.
| Chronic infection with the hepatitis C virus (HCV) is amongst the most frequent causes of liver cirrhosis and hepatocellular carcinoma [1]. About 3% of the world population is persistently infected with this virus and inspite of significant decline of new infections, owing to the long incubation period, a profound rise in the frequency of long-term complications such as steatosis, cirrhosis and liver cancer is expected [2].
HCV is the predominant member of the genus Hepacivirus in the family Flaviviridae. These viruses are enveloped and possess a single strand RNA of positive polarity. In case of HCV the genome has a length of ∼9.6 kb and it encodes a single polyprotein that is cleaved co- and post-translationally by cellular and viral proteases into 10 different products [3], [4]: core, envelope protein 1 (E1), E2, p7, nonstructural protein 2 (NS2), NS3, NS4A, NS4B, NS5A and NS5B. Core, E1 and E2 are the main viral constituents of the HCV particle. P7 and NS2 are essential ‘co-factors’ for virus assembly [5], [6], but dispensable for RNA replication [7]. This process is catalyzed by the concerted action of NS3 to NS5B proteins forming –together with cellular proteins- a membrane-associated replicase complex [8].
Studies of HCV assembly and release have become possible with the identification of the genotype 2a isolate JFH1 that efficiently replicates in the human hepatoma cell line Huh-7 and supports production of infectious virus particles [9]. This culture system has been improved by the identification of virus titer-enhancing mutations increasing infectivity yields by up to 1,000-fold [10]–[12] and the construction of JFH1 chimeras in which the region encoding core to NS2 has been replaced by analogous genome fragments from other HCV isolates [13], [14].
With the advent of these cell culture systems, first insights into HCV assembly and the roles of p7 and NS2 in this process could be gained. P7 is a small hydrophobic protein composed of two transmembrane segments (TMS) [15], [16]. It is capable to form hexa- or heptameric complexes that can act as a viroporin [17]–[19]. P7 is dispensable for RNA replication [7], but crucial for infectivity in vivo [20] likely because of its critical role in virus particle assembly [5], [21]. Whether p7 is a component of the virion is discussed controversially [21], [22].
NS2 is a 217 amino acids (aa) long cysteine-protease composed of a highly hydrophobic N-terminal membrane binding domain (MBD) and a C-terminal globular and cytosolic protease subdomain. The latter is capable to form dimers creating a composite active site [23]. This protease is not directly required for RNA replication, but has to be cleaved off the N-terminus of NS3 to allow formation of an active replicase [24]. Recently it was shown that NS2 is essential for HCV assembly [5], [6]. Interestingly, protease activity is not required for particle formation, but rather global integrity of both NS2 subdomains [25], [26]. Although the precise mode-of-action of NS2 during HCV assembly is not known, a recent study suggests that this protein acts at a late stage of infectious particle formation [27].
The exact membrane topology and architecture of the N-terminal MBD of NS2 is not known. It might be composed of three trans-membrane segments (TMS) [28], but alternative models are possible. We have recently shown that TMS1 (aa 1–23) adopts an overall helical fold interrupted by flexible glycine residues at position 10 and 11 [6]. While TMS-1 is clearly predicted as a single membrane-spanning trans-membrane helix, this is less clear for TMS2 and TMS3 that may span the membrane bilayer or reside on its cytosolic surface in a helix-loop-helix configuration.
By using combinations of reverse genetic and biochemical approaches, convincing evidence has been obtained that also factors of the viral replicase are essential for particle formation, most notably NS3 and NS5A [25], [29]–[32]. The latter is a highly phosphorylated RNA binding protein composed of an N-terminal amphipathic alpha-helix serving as a membrane anchor and contributing to targeting of the protein to lipid droplets (LDs) [33], [34], and three domains [35]. Domain I forms a dimer and is essential for RNA replication [36]. Most of domain II is dispensable for replication [30] whereas the C-terminal domain III is essential for virus production, most likely via interaction with the core protein [32]. This interaction appears to be regulated by casein kinase II-mediated phosphorylation of NS5A [29].
Assembly of HCV particles is tightly linked to lipid metabolism, LDs and the machinery required for production and secretion of very-low-density lipoproteins (VLDL) [31], [37]–[39]. Several models of HCV assembly have been put forward, but the precise details are unknown (reviewed in [40]). While these models can explain the early steps of nucleocapsid formation, it is unclear how these nucleocapsids acquire the membranous viral envelope and the envelope glycoproteins and how this process is linked to VLDL formation and secretion. NS2 may play a central role in these reactions, but the precise mechanisms are not known [25], [27].
In this study we undertook a detailed structural and functional characterization of the N-terminal MBD of NS2. We solved the NMR-structures of TMS2 and TMS3 and propose a model of NS2 membrane topology. In addition, we performed a structure-activity study of the MBD and established an interaction map of NS2. The data reveal that NS2 serves as a key organizer participating in multiple protein-protein interactions that are required for the assembly of infectious HCV particles.
We reported recently that a transmembrane segment denoted TMS1 was almost invariably predicted in the very N-terminal region (aa 1–23) of NS2, irrespective of the analyzed genotypes and subtypes [6]. TMS in the 23–102 region ([6] and references therein) yielded inconsistent results that depended both on the genotype examined and the method used (data not shown). By using secondary structure predictions and the algorithm developed by Wimley and White to calculate the propensity of an aa sequence to interact with membranes (Figure S1, A and B) we could deduce that the consensus segments 17–45 and 72–96 exhibit a clear propensity to partition into the membrane bilayer and likely include transmembrane helical passages (Figure 1A and supplementary Figure S1). In contrast, the aa segment 49–71 is predicted not to show such properties. Based on these results, the NS2 MBD sequence was divided into the three segments: 1–27, 27–59, and 60–99, each containing a putative transmembrane helix (Figure 1A).
To determine the capacity of these segments to associate with membranes, we analyzed proteins comprising full length NS2 or putative NS2 TMS that were C-terminally fused to green fluorescent protein (GFP) by fluorescence microscopy. The NS2-GFP fusion protein showed a fluorescence pattern that included the nuclear membrane, was strongest in the perinuclear region, and extended in a reticular pattern throughout the cytoplasm (Figure 1B). This pattern corresponds to the endoplasmic reticulum (ER), as corroborated by the colocalization with protein disulfide isomerase (PDI). Each of the predicted TMS of NS2 showed a very similar subcellular localization whereas GFP expressed individually was diffusely distributed throughout the cell including the nucleus. These observations indicate a clear propensity of each of the three segments to associate with membranes.
To gain insight into the structure and membranotropic properties of NS2 segments 27–59 and 60–99, the corresponding peptides of the Con1 strain (1b) designated NS2[27]-[59] and NS2[60-99] were chemically synthesized, purified to homogeneity, and their structures were analyzed by circular dichroism and nuclear magnetic resonance in membrane mimetic environments (for details see supplementary Figure S1 and materials and methods S1). The 3D model structures obtained for both peptides identified one α-helical segment in case of NS2[27]-[59] and three well defined helical segments in case of NS2[60-99] (Figure 1C). Based on physicochemical considerations, a transmembrane orientation of the amphipathic α-helix in TMS2 could only be achieved upon interaction with another complementary TMS neutralizing the polar and charged residues located in the hydrophobic core of the membrane. In case of NS2[60-99], the 35 aa segment including all three helices would be too long for a single transmembrane passage given the average length of transmembrane helices (16 to 25 aa). As the most hydrophobic stretch extends between aa 82–93, and considering that both edges of this stretch include large hydrophobic residues, we assume that segment ∼77–97 forms a TMS. The first small helix (64–69), which includes the short hydrophobic stretch VILL might be located in the membrane interface, possibly in-plane of the membrane.
These considerations together with the available structural data for NS2 TMS1 [6] allow us to propose a model for the membrane association and topology of NS2 MBD (Figure 1D). It would contain three transmembrane, mainly helical segments (TMS1: 4–23; TMS2: 27–49; and TMS3: 72–94), connected by a small cytosolic loop (aa 24–26) and by a luminal segment (aa 50–71) containing a short helix supposed to interact with the membrane interface. Although in this model the three TMS and protease ectodomain are represented as separated entities, given the dimeric structure of the protease domain [23] we expect a packed overall NS2 structure and eventually higher-order complexes of NS2 dimers mediated by intermolecular interactions between MBDs.
To correlate the structure of NS2 MBD with function, we conducted an extensive mutagenesis of this domain based on the NMR secondary structures in order to identify residues and structural determinants that are most critical for HCV assembly without affecting RNA replication. Mutants with a very low, but still detectable assembly competence were then used to select for pseudoreversions capable to rescue the assembly defect, which was achieved by serial passage of virus in Huh7.5 cells. We anticipated that these pseudoreversions would reside either within NS2, which might be used to refine the structure model, or in other viral proteins that thus would be candidates for interaction with NS2.
For the reverse genetic studies we used the JFH1 derivative JFH1mut4-6 containing three mutations (V2153A and V2440L in NS5A and V2941M in NS5B) elevating virus titers close to the level of the highly efficient chimera Jc1 (∼106 TCID50/ml) without affecting RNA replication [10] (Figure 2A). We chose JFH1mut4-6 for several reasons: first, assembly efficiency of the parental JFH1 genome is very low (∼103 TCID50/ml), which precludes its use to select for pseudoreversions; second, the titer enhancing mutations reside in the replicase, thus avoiding possible effects on the structural proteins and the assembly factors p7 and NS2; third, the Jc1 chimera has a cross-over site of two different HCV genomes within NS2 [14], which may confound phenotypes caused by mutations within NS2.
In the first set of experiments we generated a panel of NS2 mutants in which individual α-helices of the different TMS of JFH1mut4-6 were replaced by those of the genotype 1b isolate Con1. These ‘helix-swap’ mutations should not affect the secondary structure and thus preserve overall structure, folding and topology of NS2, but disrupt genotype-specific protein-protein interactions. Importantly, since each of the exchanged α-helices of JFH1 and Con1 differ by several aa residues the risk to select for revertants rather than pseudorevertants was very low.
Based on the NMR structures reported earlier [6] and in this study, we constructed 7 helix-swap mutants (aa sequences of affected helices are boxed in Figure 1A): JFH1-CT1.h and CT2.h in which we exchanged the α-helix of TMS1 (aa 11–22 of NS2) or TMS2 (aa 34–46), respectively; JFH1-CT3.h1, JFH1-CT3.h2 and JFH1- CT3.h3 in which the individual short α-helices of TMS3 were exchanged (aa 62–69; 75–85 and 89–97, respectively). In addition, we generated constructs JFH1-CT3.h12 and CT3.h23 in which two short α-helices of TMS3 were exchanged at the same time (aa 62–85 and 75–97, respectively).
As shown in Figure 2B, save for the mutant in which α-helix1 in TMS3 was exchanged (mutant JFH1-CT3.h1), all mutants were profoundly impaired in virus production and infectivity titers were reduced up to 1,000 fold at 72 h after transfection of Huh7.5 cells. This impairment correlated with the degree of sequence conservation between the exchanged helices. Aa sequence alignments revealed that Con1 – JFH1 sequence similarities were in the range of 58–81%, but in case of helix1 of TMS3 similarity was only 37%. Since this helix swap mutant was unaffected in assembly, this region most likely is required for interactions with the membrane and thus genotype independent.
Analysis of intra- and extracellular infectivity revealed that in all cases reduced titers were due to impaired assembly rather than virus release (Figure 2C). With the exception of mutant CT3.h23 producing lower amounts of NS2 assembly defects could not be ascribed to gross alterations of NS2 abundance (Figure 2D). The protein of higher molecular weight detected with mutant JFH1-CT2.h corresponded to uncleaved p7-NS2 arguing for a processing defect of this mutant, which was not the case for all the other mutants. In addition, a distinct product of smaller size (about 16 kDa) was detected on longer exposure (not shown) and this protein most likely corresponds to an N-terminal cleavage product of NS2 designated tNS2 [6] (see below). In summary, these results show that the integrity of the N-terminal MBD of NS2 is important for HCV assembly and that all 3 TMS are required.
To further narrow down the regions within individual TMS of NS2 that are most crucial for assembly, we generated a panel of single aa substitutions that were designed on the basis of the degree of conservation across the different genotypes, aa size, charge, polarity and hydrophobicity (Figure 1A). Large and hydrophobic aa were replaced by smaller and less hydrophobic ones (Y26P, Y39A, W51A, F77A or LL83-84AA); small aa probably serving as flexible linkers between individual α-helices were replaced by larger aa assumed to reduce flexibility of the N-terminal MBD (G10A/S/T/P, P53I, P73I and G88L); charged aa potentially involved in electrostatic interactions were replaced by aa with opposite charge thus possibly introducing repulsive forces (K27E, E45R, R58E, R61E and K81E); finally, the polar threonine residue at position 80 was replaced by the larger and hydrophobic aa leucine. All mutants were tested for protein expression, kinetic of infectivity release as well as amounts of intra- and extracellular infectivity. The results shown in Figure 3A and B demonstrate that both infectivity release and intracellular infectivity levels were reduced with all mutants, albeit to very different degrees. In agreement with our earlier results demonstrating the important role of the glycine residue at position 10 for virion assembly [6], we found that mutants G10P and G10T did not support virus production and even less drastic alanine or serine substitutions reduced infectivity titers up to 300-fold. Substitutions residing in loop1 that connects α-helix 1 and 2 strongly reduced or completely abolished production of infectious HCV particles (Y26P and K27E, respectively). In contrast, aa substitutions in loop2 (W51A, P53I, R58E and R61E) slowed down the kinetic of infectivity release, which was best detected at 24 h post transfection, whereas intra- and extracellular infectivity titers were reduced only moderately at later time points as compared to the parental construct JFH1mut4-6 (wt).
Alanine substitutions of aromatic aa residues in TMS2 (Y39) or TMS3 (F77) reduced infectivity titers up to 1,000-fold whereas mutations introducing electrostatic repulsion in TMS2 or TMS3 (E45R and K81E, respectively) blocked virus production almost completely. Substitutions targeting the flexible region between helix one and two or helix two and three in TMS3 (P73I and G88L, respectively) strongly reduced infectivity titers (1,000-fold at 72 h p.e.). Surprisingly substitutions affecting the highly conserved polar aa residue at position 80 (T80L) or the two leucine residues at positions 83 and 84 (LL83-84AA) did not give rise to a detectable phenotype. Interestingly, as shown in Figure 1D, residues 80, 83 and 84 are located on one helix side, suggesting that it is not important for protein – protein interactions. In contrast, residues 77 and 81 are located on the opposite side, arguing that this helix side might be involved in interactions.
Western blot analysis of NS2 proteins expressed in cells after transfection with each of the mutants or the parental construct revealed no gross difference in the abundance of this protein and the other HCV proteins arguing for similar replication levels and protein stabilities (Figure 3C). Nevertheless, some variations in abundance of individual HCV proteins were detected such as lower amounts of NS2 in case of K81E and reduced amounts of core protein in case of LL83AA. However, these rather subtle differences are very unlikely to account for the often drastic impairment of HCV assembly. In agreement with an earlier report, ‘truncated NS2’ [6] was detected to variable levels, but its abundance did not correlate with assembly phenotypes.
Given the important role of tryptophan in protein-protein interaction within a membrane [41] and their preferred location at the membrane interface [42], [43], we also analyzed a panel of mutations affecting the two fully conserved W35 and W36 residues. A striking correlation was found between reduction of aromaticity as well as size of residues at these sites and reduction of virus production (data not shown) arguing that the aromatic side chains of W35 and W36 are involved in essential interactions such as membrane tethering of TMS2 via aromatic ring stacking.
Having generated a panel of NS2 mutants with a selective assembly defect, and –in some cases- an additional virus release defect (e.g. Y26A; Figure 3B) we wanted to establish a genetic interaction map of NS2. For this purpose we used a cell culture adaptation strategy, which was possible, because the parental construct JFH1mut4-6 that was used for mutagenesis already supports high level virus production and therefore, selection for pseudoreversion would not give rise to undesired mutations enhancing assembly in general [10]. Culture supernatants collected from cells 72 h after transfection with a given NS2 mutant were concentrated and used to infect naive Huh7.5 cells that were passaged 6 times. After 4 additional passages of culture supernatants, they were used to inoculate naïve Huh7.5 cells and virus titers produced therefrom were determined by TCID50 assay. In case of mutants with elevated virus titers, cell lysates were used to prepare total RNA, HCV genomes were amplified by RT-PCR and amplicons spanning most of the 5′ NTR up to the middle of NS3 were either directly sequenced or cloned prior to sequence analysis. In the latter case at least two independent clones were analyzed and only mutations conserved between the two cDNA clones were considered in order to discriminate against mutations that might have been introduced by PCR. Pseudoreversions outside the analyzed region, including NS5A, were not considered because we used the JFH1mut4-6 genome that already contained titer-enhancing mutations in NS5A to allow adaptation. Mutations identified by this approach were inserted into the corresponding parental NS2 mutant and replication as well as assembly properties were analyzed by Western blot and TCID50 assays. A summary of all pseudoreversions identified in this way along with their degree of titer enhancement is given in Table 1.
Several assembly deficient mutants (G10T, G10P, K81D, G88L and JFH1-CT3.h23) could not be adapted, because the virus was rapidly lost during cell passages suggesting that the genetic barrier was too high and assembly impairment too strong. Nevertheless, for most mutants we could select for pseudoreversions with the exception of G10A and K81E where reversion to wild type occured, which was achieved by just one nucleotide substitution. Since we had inserted in addition a silent nucleotide exchange in the subsequent codon that was retained in the selected virus, we could rule out a contamination with wild type virus. All of the other adapted mutants contained pseudoreversions. They resided primarily within NS2 and two ‘classes’ of pseudoreversions could be discriminated: first, those residing at the same position as the primary mutation, but with a different substituting aa residue; second, pseudoreversions at a different site than the primary mutation. Pseudoreversions belonging to the first class (R45G and I73S) enhanced virus production 88,000- and 3,100-fold and thus back to the level of the parental genome JFH1mut4-6 (Figure 4A and Table 1). Pseudoreversions belonging to the second class (double mutants G10S-T23N, W36F-Q32R and Y39A-G25R) increased infectivity titers ∼220-, 450- and 10-fold, respectively, with G10S-T23N also reaching wild type levels (Figure 4A and Table 1). For this reason the T23N substitution was also combined with mutants G10T and G10P, but titer increase was very moderate (Figure 4A).
In addition to pseudoreversions in NS2, one particular mutation was identified in NS3 (Q221L) that rescued infectivity titers of the W35F and the W36L mutants by 1,400- and 38,000-fold (Figure 4B). This NS3 mutation has previously been described as a general titer enhancing substitution [25], [27]. For this reason we inserted the Q221L mutation into mutants W35A, W35L and W36A that are completely defective in assembly and that could not be adapted with our approach. As shown in Figure 4B, infectivity titers were enhanced in all cases corroborating the more general assembly enhancing phenotype exerted by this particular NS3 mutation. Moreover, the same mutation was also capable to rescue virus production of several NS2 mutants even in trans (supplementary Figure S2). Although in this case rescue efficiency was lower as compared to direct insertion of the mutation into the NS2 mutant genome, this observation suggests that a genetic separation of the replication and assembly function of NS3 is possible.
In case of the F77A mutation residing in TMS3.2 of NS2 two mutations located in E2 could be selected: Y215S located in the ectodomain and V341A in the N-terminus of the TMS of E2 (Table 1). When we tested these substitions in the context of the parental NS2 mutant we found that Y215S completely abrogated infectious virus production concomitant with a strong reduction in intracellular NS2 amounts (Figure 4A). However, the V341A substitution enhanced virus titer ∼100-fold arguing for a (direct or indirect) interaction between these two proteins for efficient particle production.
A more complex pattern of mutations was found upon selection for pseudoreversions in case of helix-swap mutants JFH1-CT1.h, JFH1-CT2.h, JFH1-CT3.h2 and JFH1-CT3.h3 (Table 1 and Figure 4C). For CT1.h, we found E3D residing in the N-terminal helix of p7 [16], F14L in TMS1 of NS2 and I17M in the membrane-binding amphipathic helix α0 of the NS3 protease domain. When inserted into the parental virus, strongest rescue of assembly was found with the p7 mutation and infectivity titers were further increased 10-fold at 24 h or 2-fold at 72 h p.t. when combined with the NS2 pseudoreversion (CT1.h - p7NS2). In contrast, the NS3 mutation had a slightly negative effect (CT1.h-NS3-I17M). For the JFH1-CT2.h mutant we detected two pseudoreversions residing in the turn connecting the N-terminal helix and the TM1 helix of p7 (N15D) [16] and the loop connecting TMS1 and TMS2 of NS2 (G25R). In this case, the mutation in NS2 rescued HCV assembly strongest (∼265-fold) whereas the p7 mutation had a very moderate effect and in combination with the NS2 mutation did not enhance virus titers further. In case of pseudoreversions detected with helix-swap mutants JFH1-CT3.h2 and JFH1-CT3.h3, the only titer enhancing mutations were found in NS2 (K172R, T21A) with T21A arguing for an interaction between TMS1 and TMS3 of NS2. The double mutation identified in E1 of JFH1-CT3.h2 construct (E151D/I181S) had no effect.
For all tested single aa mutants and helix-swap mutants and their corresponding pseudorevertants, the enhancement of infectivity titers in cell culture supernatants correlated with increased amounts of intracellular infectivity showing that the pseudoreversions rescued primarily assembly rather than virus release (supplementary Figure S3). Moreover, with the exception of the F77A-E2-Y215S double mutant, amounts of NS2 as well as NS5A, NS3 and core protein were not grossly altered (Figure 4 and supplementary Figure S4, respectively) suggesting that overall protein stabilities were not profoundly affected by the mutations. We note however that for mutant F77A already producing somewhat lower amounts of NS2 as compared to the wildtype, NS2 abundance was reduced much more by the Y215S substitution in E2. Moreover, in case of helix-swap mutant CT2.h and the corresponding pseudorevertants cleavage between p7 and NS2 was impaired (Figure 4C). Interestingly, even in case of rescue mutant CT2.h-NS2, the substitution in NS2 enhancing virus production about 250-fold does not affect the amounts of this uncleaved precursor arguing that assembly competence of this mutant is restored in a manner that still allows delayed p7-NS2 cleavage, thus compensating e.g. an impaired p7-NS2 interaction (see below).
Although in most cases, selection for pseudoreversion resulted in compensatory mutations within NS2 itself, we also identified pseudoreversions in E2, p7 and NS3. In case of helix-swap mutant CT1.h, the E3D substitution in p7 restored almost wild type infectivity titers (Table 1). Likewise, the assembly defect of mutant W35F in NS2 was compensated by Q221L in NS3 and NS2 mutant F77A was compensated by the V341A substitution in E2. These results suggested that NS2 might interact (directly or indirectly) with each of these proteins. To support this assumption by pull-down assays, we first generated a fully functional JFH1-derivative with a tagged NS2 protein suitable for efficient immunoprecipitation and allowing capture of NS2 independent from any mutation that might affect recognition with the NS2-specific antibody. To this end we constructed a series of mutants in which NS2 was fused N- or C-terminally with several tags such as the FLAG-, hexa-histidine (His)- or hemagglutinin (HA)-tag. All genomes with C-terminal fusions no longer supported HCV particle production (data not shown). Moreover, when we tried to select for titer enhancing pseudorevertants of assembly deficient NS2-tagged mutants, in all cases the tag was partially or completely deleted (not shown). In contrast, viable mutants were obtained with N-terminally tagged NS2 versions in which the first 5 codons of NS2 were duplicated upstream of the heterologous sequence that was composed of a single copy of the tag and a linker sequence encoding for Gly-Ser-Gly preceeding NS2. In addition, variants were generated with a second insertion of the tag sequence to increase efficiency of immunoprecipitation (Figure 5A). Analysis of the kinetics of virus production revealed that both the single Flag-tagged (F-NS2) and the HA-Flag-double tagged variant (HAF-NS2) produced amounts of intra- and extracellular virions that were comparable to the parental genome JFH1mut4-6 (data not shown and Figure 5B). In contrast, the variants with the tandem Flag-tag (FF-NS2) or the His-Flag-tag combination (HisF-NS2) produced lower amounts of extra- and intracellular infectious particles arguing for an assembly defect. Western blot analysis revealed comparable replication of all constructs and no defect of polyprotein processing was detected (Figure 5C). The sizes of the various NS2 proteins and their immunoreactivities confirmed that the tag(s) remained fused to mature (fully processed) NS2.
Taking advantage of these assembly-competent tagged NS2 constructs, we selected JFH1mut4-6HAF-NS2 to determine NS2 interactants. Huh7.5 cells were transfected with this construct and NS2-containing immunocomplexes captured from lysates that were prepared 72 h p.t. were analyzed by Western-blot for coprecipitation of core, E2, p7, NS3 and NS5A. Specificity of immunoprecipitation and Western blot analysis was determined by using the parental JFHmut4-6 construct that lacked the N-terminal NS2-tag. The results in Figure 5D show that the tagged NS2 protein co-precipitated with E2, p7, NS3 and NS5A, but not with core. In contrast, no signal was found in case of the non-tagged genome inspite of comparable amounts of viral proteins in cell lysates, demonstrating specificity of these co-precipitations.
To analyze whether the pseudoreversions in E2, p7 and NS3 affect the NS2 interaction pattern, we chose those NS2 mutants for which infectivity titers were enhanced by a pseudoreversion outside of NS2: CT1.h and CT1.h-p7-E3D; W35F and W35F-NS3-Q221L; F77A and F77A-E2-V341A. Lysates of all samples harvested 72 h after electroporation together with positive and negative controls were subjected to HA-specific pull-down and immunocomplexes were analyzed by Western blot (Figure 6A). Pull-down efficiencies were quantified by densitometry scanning and normalized to protein amounts detected in the corresponding cell lysate (Figure 6B); based on this quantification fold enhancement of coimmunoprecipitation achieved by the pseudoreversion was determined (Figure 6C).
For all mutants, NS2 interaction with the other viral proteins was reduced, but to very different extents. Most pronounced was the impairment of NS2 interaction with E2, p7 and NS3, whereas interaction with NS5A was less affected. Importantly, the E3D pseudoreversion in p7 introduced into CT1.h enhanced interaction with E2 and NS3 back to wild type levels correlating well with the rescue of assembly competence of this helix-swap mutant. Unfortunately this pseudoreversion disrupted the epitope recognized by the p7-specific antibody and therefore, the degree of coprecipitation of this p7 with NS2 could not be determined. Interaction of NS2 with NS5A was elevated even above the wild type level (Figure 6B, C). As expected, the Q221L pseudoreversion in NS3 introduced into the W35F NS2 mutant increased NS2 – NS3 interaction, but surprisingly had little or no effect on NS2 interaction with E2 or p7, respectively, and even a negative effect on interaction with NS5A (Figure 6B, C). The pseudoreversion in the TMS of E2 (V341A) introduced into the F77A mutant moderately enhanced interaction of NS2 with E2 and NS3, but no significant enhancement of interaction with p7 and NS5A was detected. Interestingly, NS2 containing this F77A substitution coprecipitated with both phospho-variants of NS5A to the same extent, whereas all other NS2 proteins tested preferentially interacted with the basal phosphorylated form p56 (Figure 6A). This phenotype of the F77A mutant was not altered by the adaptive mutation residing in E2.
To support and extend the interaction patterns described above with an alternative assay we performed colocalization studies of NS2 with structural and other nonstructural proteins. In the initial set of experiments, we determined the subcellular localization of NS2 (Figure 7A) and observed a profound change from a reticular ER staining pattern 36 h post transfection (NS2 colocalization with the ER marker PDI is not shown) to a strong punctate NS2 stain accumulating in close proximity of LDs 72 h post transfection. By counting ∼200 cells we defined two phenotypes, based on the number of LDs with NS2 accumulation: phenotype 1 with less than 10 NS2-positive LD structures per cell and phenotype 2 with more than 10. A time-dependent increase of phenotype 2 was also observed although the overall percentage was lower, which was probably due to lower replication as compared to RNA transfection (supplementary Figure S5).
The functional relevance of these two phenotypes is supported by the analysis of the NS2 mutants and their respective pseudorevertants (Figure 7B). We found that NS2 decorated LDs were much less frequent in an assembly deficient mutant and even 72 h after transfection the majority of NS2 was localized at the ER. Importantly, upon insertion of the corresponding pseudoreversion a shift back to phenotype 2 representing higher abundance of NS2 ‘positive’ LDs 72 h p.t. was detected (Figure 7B).
To determine whether other viral proteins might be recruited to LDs in an NS2-dependent manner we performed colocalization studies. As shown in Figure 8A, at each analyzed time point we found a striking colocalization of NS2 and E2 in case of the wild type, consistent with the coimmunoprecipitation results. In addition, we detected a strong accumulation of both proteins around LDs 72 h p.t. (Figure 8A). A similar pattern, but less colocalization as determined by Pearson's correlation coefficient, was found for NS2 with NS3 (Figure 8B). Interestingly, a lower degree of colocalization of NS2 and NS5A predominated 36 h p.t. and NS5A localized in close proximity of LDs independent of NS2. Accumulation of NS2 around LDs at the later time point coincided with increased NS2-NS5A colocalization at these sites. No significant colocalization was detected between NS2 and core protein at LDs. However, a small amount of core colocalized with NS2 in a reticular, presumably ER-derived compartment (supplementary Figure S6). Attempts to detect p7 by immunofluorescence were not successful with the available antibodies and insertion of tags into p7 very much impaired assembly (not shown). Therefore, p7 – NS2 colocalization could not be studied.
Given the most pronounced loss of NS2 accumulation around LDs (i.e. low frequency of phenotype 1) with mutant CT1.h we determined for this construct and the corresponding pseudorevertant NS2 colocalization with E2, NS3 and NS5A as well as HCV protein accumulation at LDs. For the parental NS2 mutant we found that E2 no longer localized to LDs and localization of NS3 to these sites was strongly impaired (Figure 9A). However, recruitment of these viral proteins to LDs and strong colocalization at LDs was restored by the pseudoreversion in p7 (E3D; Figure 9A, lower panel). In contrast, NS5A was recruited to LDs independent from NS2 and NS2 – NS5A colocalization was also restored by this pseudoreversion.
When analyzing a larger panel of NS2 mutants and their corresponding pseudorevertants for colocalization of these HCV proteins in a quantitative manner (Figure 9B) we found for CT1.h a slight reduction of NS2 colocalization with E2, NS3 and NS5A that was partially restored by the pseudoreversion residing in p7 (CT1.h-p7). In case of NS2 mutants W35F and F77A only NS2 colocalization with NS5A was impaired, but restored by the corresponding pseudoreversion in NS3 or E2 (W35F-NS3-Q221L or F77A-E2-V341A, respectively). In contrast, colocalization of NS2 with E2 or NS3 was unaffected by these NS2 mutations (Figure 9B). This result could be explained by the fact that the mutations in NS2 might impair accumulation around LDs and thus would lead to accumulation of NS2 at ER membranes where also the majority of E2 and NS3 reside. Therefore, colocalization of these NS2 proteins with E2 and NS3 (at the ER membrane) might be strong. In contrast, NS5A is recruited to LDs independent of NS2 and therefore, NS2 mutations that no longer are recruited to LDs might have lower colocalization rate that would be restored by the pseudoreversion that rescues ‘LD targeting’ of NS2.
The important role of NS2 for HCV assembly has been shown by several earlier reports [5], [6], [25]–[27]. However, the way this protein contributes to virion formation remains enigmatic. We addressed this question by using several complementary approaches including the structural analysis of the MBD, reverse and forward genetic analyses and a combination of coimmunoprecipitation and subcellular localization studies. Our results provide evidence that NS2 recruits the envelope glycoproteins (presumably in conjunction with p7) and probably also NS3 to LDs and serves as a key organizer of HCV assembly by participating in multiple protein - protein interactions required for virion formation. The implications of these results are discussed in the following.
By using NMR of synthetic peptides we solved the secondary structures of TMS2 and 3 and propose a membrane topology model of the overall N-terminal MBD (Figure 1D). This model supports and very much extends an earlier report [28] and proposes 3 transmembrane α-helices, connected by flexible loop regions. While TMS1 and 2 consist of one α-helix, TMS3 is composed of three. Each of TMS1 - 3 is capable to mediate membrane association on its own. To determine whether individual helices within TMS3 are sufficient for membrane targeting we analyzed subcellular localization patterns of NS2-GFP fusion proteins comprising NS2 aa residues 60–88 or 74–99. However, these proteins displayed a predominantly diffuse fluorescence signal, arguing that all 3 helices of TMS3 are required for membrane targeting (J.G. and D.M., unpublished).
The model of three TMS is consistent with homo-intramolecular TMS interactions revealed by the pseudoreversions indicating that TMS1 interacts with TMS2 and TMS3. The model is also consistent with hetero-intermolecular interactions by which TMS1 and TMS2 could interact with p7 whereas TMS3 could interact with the TMS of E1 and E2 (Figure 10). Moreover, the fact that point mutations in the long and variable connecting loop between TMS2 and TMS3 had no effect on virus production is in keeping with its ER luminal location. Conversely, the sensitivity to mutation of the small loop between TMS1 and TMS2 is consistent with its cytosolic localization However, it should be stressed that helices observed in TMS2 and TMS3 are not classical membrane anchoring TM helices, since they contain polar and charged residues. In addition, the TMS2 helix exhibits an amphipathic character suggesting that it could associate with the membrane interface, at least transiently. Based on physicochemical considerations, a transmembrane orientation of this helix is expected to be achieved only upon interaction with another complementary transmembrane segment neutralizing the polar and charged residues located in the hydrophobic core of the membrane. In this context, it is possible that the transmembrane association of TMS2 and TMS3 occurs in the translocon during NS2 biosynthesis. Alternatively, these TMS might be first released into the cytosol where they could interact at the membrane interface and then associate with the membrane to adopt their final transmembrane topology. Interestingly, the length of the connecting loop between TMS2 and TMS3 and the absence of an interaction between these TMS suggest that TMS3 might be an independent entity possibly interacting with distant partners. The fact that chimeric genomes with high assembly competence can be obtained when using a cross-over site right after TMS1 of NS2 indicates that TMS1 is functionally separated from the remainder of the NS2 MBD [14]. Overall, the MBD of NS2 appears to be composed of a series of structural elements with own functional properties, but with the capacity to acquire new functions upon intra- and intermolecular interactions. This structural plasticity is likely essential to ensure the multiple interactions mediated by the NS2 MBD.
Almost all helix-swap mutations reduced assembly competence arguing for genotype specific incompatibilities between individual TMS of either NS2 or other viral proteins, such as p7 and E2. The only exception was mutant CT3.h1 affecting helix1 of TMS3 that acts most likely as an adaptable linker between TMS2 and TMS3. This helix is the least conserved sequence of the NS2 MBD suggesting that it mediates interactions with the membrane in a genotype-independent manner.
Selection of assembly-impaired NS2 mutants for titer enhancing mutations compensating the assembly defect to the most part lead to pseudoreversions within NS2. This was the case for all helix-swap mutants and several single aa exchanges. Six out of 9 pseudoreversions within NS2 were found in the loop region connecting TMS1 and 2. This loop resides on the cytosolic side of the ER membrane and by interaction with membrane phospholipids it may stabilize membrane association of NS2 or is involved in intra- or intermolecular protein-protein interactions. Overall, these pseudoreversions most likely restore structural alterations induced by the primary mutation.
Based on our model of the NS2 MBD (Figure 1D), at least some of these mutations could be explained. One example is the Y39A pseudoreversion compensating the assembly defect caused by the G25R mutation suggesting that aa residues 25 and 39, which are located on TMS1 and TMS2, respectively, might be in contact (Figure 10). We assume that the “hole” created in TMS2 by the Y39A substitution is compensated by a bulky aa in the interacting TMS1 counterpart, thus ‘filling up’ the hole in the mutated TMS2. This interaction likely occurs at or close to the membrane interface, where the charge of Arg is well tolerated. Importantly, this assumption is corroborated by the G25R pseudoreversion that was selected with helix-swap mutant JFH1-CT2.h, which has a histidine residue at position 25. We therefore conclude that the beginning of the loop between TMS1 and TMS2 likely interacts with the helix in TMS2. Another example are the pseudoreversions K172R and T21A in NS2 that were selected with helix-swap mutants JFH1-CT3.h2 and JFH1-CT3.h3, respectively, suggesting interactions between TMS1 and TMS3 (Figure 10). While this can be easily explained for position 21, the aa at position 172 is more remote from the membrane surface. Nevertheless, this residue is at the junction between the two subdomains of the NS2 protease domain and thus still suitable to contact TMS3.
We tried to integrate all these informations into our NMR-based structure model of NS2 MBD, but these attempts were confounded by the fact that NS2 is a dimer, which most likely forms higher-order oligomeric complexes. Therefore, we do not know whether a given mutation restores intra- or intermolecular interactions. Nevertheless, the tight correlation between structural integrity of NS2 and its role in assembly is underlined by the fact that titer-enhaning mutations within NS2 have also been found by us and others when using JFH1 wild type or various virus chimeras with low assembly competence [10], [44]–[49].
Apart from pseudoreversions within NS2, we also identified two in p7. Importantly, in case of the helix-swap mutant affecting TMS1, the pseudoreversion in p7 (E3D) enhanced virus production almost back to wild type level. This result argues for an interaction between TMS1 of NS2 and p7 (Figure 10). Unfortunately, this assumption could not be tested directly, because this mutation destroyed the epitope recognized by the p7-specific antibody. However, we have earlier described that for most virus chimeras the best junction for fusion of the genome segments resides after TMS1 of NS2, whereas an intergenotypic fusion right after p7 was severely impaired in assembly [14]. Thus, genotypic compatibility between TMS1 of NS2 and the structural proteins as well as p7 appears to be required for efficient assembly.
A direct interaction between NS2 and envelope glycoproteins might be suggested by the pseudoreversions V341A residing in the N-terminus of the TMS of E2 and mutation I181S in E1 (Figure 10). We note that V341 in E2 and the primary NS2 mutation F77 responsible for the assembly defect are both most likely located in the membrane hydrophobic core, close to the ER membrane interface (Figure 1D for NS2 and Figure 1 in [50] for E2). Moreover, I181S in E1 selected as pseudoreversion with the helix-swap mutant CT3.h2 resides in the center of the TMS of E1 and thus could also directly form a stable in-membrane interaction with NS2 (Figure 10).
W35F and W36L independently adapted via the Q221L pseudoreversion in NS3 that has also been found in two earlier reports [25], [51]. This reversion is highly potent and restores viral infectivity up to ∼38,000-fold. Interestingly, this NS3 mutation also rescues assembly in trans showing that the replication and assembly function of NS3 can be separated genetically. While the mechanism by which Q221L enhances assembly is not known, we note that this residue resides on the helicase NTPase subdomain surface in a basic patch and is well accessible. This positively charged surface area might interact with the membrane surface by electrostatic interactions. In this way the aa residue at position 221 could contact the NS2 MBD at the membrane interface, at least transiently (Figure 10). According to this hypothesis, the replacement of the polar residue (Q) by a large hydrophobic aa (Leu) might reinforce membrane binding.
Co-immunoprecipitation studies revealed stable interactions of NS2 with NS3, p7 and E2 whereas interaction with NS5A was rather weak. Importantly, none of the tested conditions revealed NS2 interaction with core. These results were well supported by immunofluorescence studies demonstrating a profound and rapid colocalization of NS2 with E2 and NS3 at the ER or an ER-derived membrane compartment prior to accumulation around LDs. Several lines of evidence suggest that NS2 recruits E2 –and thus most likely also E1 that forms a very stable E1/E2 heterodimer [52]– and eventually also p7 to assembly sites in close proximity of LDs. First, we detected a profound colocalization of NS2 and E2 for each time point after infection or transfection; second, in NS2 assembly-defective viruses E2 localized primarily to the ER; third, upon insertion of the corresponding pseudoreversion E2 and NS2 colocalized again to LDs. The NS2-independent LD localization of NS5A and its weak interaction with NS2 is in agreement with previous data showing that NS5A expressed on its own is targeted to LDs, for which the N-terminal amphipathic helix appears to be most critical [33].
No significant colocalization of NS2 and core at LDs was detected. However, a small fraction of core protein presumably residing at the ER colocalized with NS2 both at the early and the late time points after transfection. Although LDs have been described as sites of HCV assembly [31] the weak NS2 – core colocalization is not in contradiction to this observation. In fact, it is speculated that the early steps of HCV assembly (nucleocapsid formation) might take place at LDs whereas the envelopement is thought to occur at the ER or an ER-derived compartment. Since NS2 probably acts at a late step of assembly [27] and might be involved in envelopment of the nucleocapsid, the colocalization of core and NS2 at the ER in close proximity of LDs rather than directly on LDs would support such a model. Moreover, given the complex membrane topology of NS2 this protein most likely can not move onto the surface of LDs that is formed by a membrane monolayer.
The results described in this study together with earlier reports [25], [27] invite speculation how NS2 might contribute to assembly. It is assumed that the early steps (nucleocapsid formation) occur in close proximity of LDs that may serve as assembly platforms [31]. By interaction between core and (RNA-containing) NS5A, capsid formation might be triggered [32]. How the envelope is acquired is not known, but we assume that NS2 plays a central role in this step. Since the TMS of E1 and E2 lack a cytosolic domain that could interact with the core protein, an adaptor protein such as NS2 that in turn efficiently binds p7 and E2 (and the latter forming heterodimers with E1), might be required to ‘deliver’ the envelope proteins to assembly sites in close proximity of LDs. This process could be facilitated by a particular membrane lipid environment supporting recruitment of the NS2 complex as well as the (lipid-binding) nucleocapsid. Alternatively, one or several host cell factors such as CIDE-B, described as an NS2 interactant [53] and required for lipid homeostasis [54], might be recruited by NS2 and aid in assembly. Moreover, NS2 efficiently binds to NS3 arguing that NS2 can form an additional complex with the replicase. How such a complex would contribute to assembly is unclear, but it may ‘tether’ the replicase to the assembly sites thus facilitating core – NS5A interaction. Alternatively, NS2 may form just one higher-order protein complex including in addition to E1/E2 and p7 NS3. This is probably facilitated by the N-terminal MBD that might form ‘clusters’ within the membrane. Finally, it is possible that the strong NS2 - NS3 interaction affects cleavage at the NS2-3 site, in this case contributing to assembly in a rather indirect manner.
In conclusion, our results point to a central role of NS2 in HCV assembly by formation of (a) multiprotein complex(es) with structural and eventually also nonstructural proteins and recruiting them to assembly sites in close proximity of LDs. In this respect, NS2 acts as a central organizer of HCV virion formation.
Sequence analyses were performed by using Network Protein Sequence Analysis (NPSA) (http://npsa-pbil.ibcp.fr [55]) and European HCV Database (http://euhcvdb.ibcp.fr [56]). Multiple-sequence alignments were performed with CLUSTAL W [57], by using the default options. Protein secondary structures were deduced from a large set of prediction methods available at the NPSA website, including HNNC, SIMPA96, MLRC, SOPM, PHD, and Predator (http://npsa-pbil.ibcp.fr/NPSA and references therein). Octanol hydrophobicity plots were generated with MPEx (http://blanco.biomol.uci.edu/mpex/) by using the scale developed by Wimley and White [58].
Monolayers of the highly permissive cell lines Huh7-Lunet [59] and Huh7.5 [60] were grown in Dulbecco's modified minimal essential medium (DMEM; Life Technologies, Karlsruhe, Germany) supplemented with 2 mM L-glutamine, nonessential amino acids, 100 U/ml of penicillin, 100 µg/ml of streptomycin, and 10% fetal calf serum. Owing to highest permissiveness for JFH-1, Huh7.5 cells were used for virus production and infection assays whereas Huh7-Lunet cells and derivatives thereof were used for immunofluorescence analyses because of their superior morphology as compared to Huh7.5 cells.
NS2-GFP fusion constructs were derived from pFK1-9605Con1 ([7]; HCV Con1 strain). First, a BamHI restriction site was eliminated by introducing a silent mutation replacing the cytidine at nucleotide position 2920 by an adenosine. Sequences encoding NS2 fragments from codon 1–27, or 27–59, or 60–99, or 1–99 or the complete NS2 coding region were fused to EcoRI and BamHI recognition sequences by PCR and amplified fragments were inserted via these two restriction sites into pCMV-KEB-GFP [61], yielding constructs pCMVNS21-27-GFP, pCMVNS227-59-GFP, pCMVNS260-99-GFP, pCMVNS21-99-GFP and pCMVNS2-GFP.
Unless otherwise stated all mutations were introduced into JFH1mut4-6 [10] corresponding to the JFH1 genome [9], but containing three virus titer enhancing mutations that do not affect RNA replication (V2153A, V2440L and V2941M). All nucleotide and aa numbers refer to the JFH1 genome (GenBank accession no. AB047639). Single aa substitutions and helix-swap mutations were introduced by PCR-based site-directed mutagenesis or overlap-PCR, respectively, using standard procedures. In case of the helix-swap mutations the following nucleotide sequences of JFH1 were replaced by the corresponding sequences of Con1: nucleotides 2811 - 2839 in case of pFK-JFH1-CT1.h; nucleotides 2879 - 2911 for pFK-JFH1-CT2.h; nucleotides 2967 - 2986 for pFK-JFH1-CT3.h1; nucleotides 3002 - 3025 in case of pFK-JFH1-CT3.h2; nucleotides 3047 - 3070 with pFK-JFH1-CT3.h3. To generate the JFH1 genome encoding a tagged NS2 protein a sequence encoding the peptide YDAPVSGDYKDDDDKGSG (corresponding to the first 5 aa of NS2, a 2 aa flexible linker (SG) containig a BspEI site, a Flag tag and a flexible GSG linker) was inserted by overlap PCR between nucleotide 2779 and 2780 of the JFH1 genome. A silent G to A mutation at position 2794 was introduced to create a BsrGI restriction site whereas the natural BsrGI site at position 7786 was destroyed by a silent A to T mutation. In addition, a silent A to C nucleotide substitution at position 1741 was introduced to create a dam methylation site affecting the BspEI cleavage site at this position. To generate genomes with double tagged NS2, oligonucleotides encoding the Flag-, or hexahistidine- or HA-tag fused to the GSG linker were inserted in-frame into the BspEI site.
The experimental procedures used to generate in vitro transcripts from cloned HCV sequences and transfection of Huh-7 cells by electroporation have been described in detail recently [6]. For trans-complementation assays a mixture of 7.5 µg NS2 mutant and 5 µg helper replicon RNA was used. After electroporation, cells were immediately transferred to complete DMEM and seeded as required for the assay.
Virus titers were determined as described elsewhere with slight modifications [13]. In brief, Huh7.5 cells were seeded into 96-well plates and fixed 3 - 4 days after infection. For immunohistochemistry we used an antibody specific for the JFH1 NS3 helicase (2E3, generated in cooperation with H. Tang, Florida State University, USA) at a dilution of 1∶100 or the 9E10 monoclonal antibody specific against NS5A protein in dilution 1∶2,000 (NS5A-9E10; kindly provided by C.M. Rice, New York, USA). Bound antibody was detected with a peroxidase-conjugated secondary antibody specific to murine IgG (Sigma-Aldrich) diluted 1∶200 in PBS. Virus titers (50% tissue culture infective dose per ml; [TCID50/ml]) were calculated as described recently [62].
Huh7.5 cells were electroporated with 10 µg in vitro transcript and culture supernatant harvested 72 h later was concentrated by ultrafiltration using an Amicon Ultra Centrifugal Filter Column (Milipore). Naïve Huh7.5 cells were inoculated with this concentrate and continuously passaged up to 6 times. Thereafter, culture supernatants were passaged 4 times on naïve Huh7.5 cells and virus titers were determined by TCID50 or immunofluorescence assay. Details of the adaptation method have been described elsewhere [63].
HCV RNA present in Huh7.5 cells was amplified and cloned as described previously [10], [63]. In brief, total RNA was isolated from a confluent 10 cm-diameter dish of Huh7.5 cells infected with the adapted virus population by using the Nucleo Spin RNAII Kit (Macherey-Nagel, Düren, Germany) as recommended by the manufacturer. One µg total RNA and 50 pmol of primer A9482 (5′-GGA ACA GTT AGC TAT GGA GTG TAC C-3′) were applied for cDNA synthesis by using the Expand-RT system (Roche, Mannheim, Germany) as recommended by the manufacturer. Two to four microliters of the reaction mixture were used to amplify the 5′ half of the HCV genome with the Expand Long Template PCR kit (Roche) according to the instructions of the manufacturer with primers S59-EcoRI (5′-TGT CTT CAC GCA GAA AGC GCC TAG-3′) and A4614 (5′-CTG AGC TGG TAT TAT GGA GAC GTC C-3′). PCR products were directly sequenced (mutants G10S, E45R, P73I and F77A) or inserted into pFK-I389Luc-EI/NS3-3′/JFH1-dg after restriction with EcoRI and SpeI. Sequence analysis of two independent plasmid clones was performed with an appropriate set of primers.
The following antisera were used in this study: rabbit polyclonal antibody specific for NS2 (NS2-1519; [6]); rabbit polyclonal antibody specific for the core protein (C-830; [64]); rabbit polyclonal antibody specific for NS3 of JFH1 (NS3-4949; [30]); mouse monoclonal antibody specific for JFH1 NS3 (NS3-2E3; generated in co-operation with H. Tang, Florida State University, USA); rabbit polyclonal antibody specific for JFH1 NS5A (NS5A-52; [6]); mouse monoclonal antibody specific for NS5A (NS5A-9E10, kindly provided by C.M. Rice, New York, USA); rabbit polyclonal anti p7-2716 and anti p7-2717 (kindly provided by M. Harris and S. Griffin, Leeds, UK); rabbit polyclonal antibody specific for J6 E2 [6]; mouse monoclonal antibody specific for E2 protein (AP33, kindly provided by Arvin Patel, Glasgow, U.K.); mouse monoclonal anti-Flag antibody, mouse monoclonal anti-HA antibody and mouse monoclonal anti-β-actin, all from Sigma-Aldrich (Munich, Germany). Monoclonal antibody (mAb) 1D3 against protein disulfide isomerase (PDI) was purchased from StressGen (Victoria, BC, Canada).
Western blot analysis was performed as described previously [6]. Samples harvested 48 h after transfection were heated for 20 min at 95°C in sample buffer (125 mM Tris/HCl, 2% (w/v) SDS, 5% (v/v) 2-mercaptoethanol, 10% (v/v) glycerol, 0.001% (w/v) bromophenol blue, pH 6.8) and separated by SDS polyacrylamide gel electrophoresis. Proteins were electro-transferred to a polyvinylidene fluoride (PVDF) membrane (PerkinElmer Life Sciences) for 1 h. Membrane was blocked overnight in PBS supplemented with 0.5% Tween (PBS-T) and 5% dried milk (PBS-M) at 4°C prior to 1 h incubation with primary antibody diluted in 2% milk in PBS-T. Membrane was washed 3 times with PBS-T and incubated for 1 h with horseradish-peroxidase conjugated secondary antibody. Bound antibodies were detected after 3 times washing with the ECL Plus Western Blotting Detection System (GE Healthcare Europe, Freiburg, Germany).
Huh7.5 cells were mock treated or transfected with HCV RNA, and samples were harvested 72 h later by scraping into IP buffer (0.5% n-dodecyl-β-D-maltoside, 100 mM NaCl, 20 mM Tris pH 7,5). After 60 min incubation on ice, cell debris was removed by 15 min centrifugation at 20,000xg. Samples were incubated with HA-specific antibody beads (Sigma Aldrich) over night at 4°C. After three times washing with IP buffer, samples were eluted into sample buffer and separated by electrophoresis into a 11% Tris-Tricine gel as described elsewhere [65]. Proteins were transferred onto PVDF membrane and HCV proteins were detected by Western blot as described above.
U-2 OS human osteosarcoma cells grown on glass coverslips were transfected with GFP fusion constructs, fixed 24 to 48 h post transfection with 2% paraformaldehyde, and mounted in SlowFade (Molecular Probes, Eugene, OR). Immunofluorescence staining was performed as described previously [66]. Bound primary antibody was revealed with Alexa-488-conjugated goat anti-mouse antibody (Molecular Probes). Mounted coverslips were examined using a Leica SP5 AOBS confocal laser scanning microscope. Immunofluorescence detection of HCV proteins in Huh7-Lunet cells was conducted in the analogous way with some modifications. Cells were transfected with HCV RNA and fixed 24, 36, 48 and 72 h post-transfection with 4% paraformaldehyde. Bound primary antibodies were detected with a Alexa-488-conjugated goat anti-rabbit antibody or a Alexa-568-conjugated goat anti-mouse antibody (Molecular Probes). LDs were stained with HCS LipidTOX™ Deep Red neutral lipid stain (Molecular Probes). Coverslips were mounted in Fluoromount-G mounting medium (Electron Microscopy Sciences, Ft. Washington, USA) and examined with a Perkin Elmer spinning disk confocal ERS 6Line microscope. Images were deconvolved with the Huygens Essential 3.5 software using a theoretical point spread function. 3D reconstructed images were created using the Volocity 5.3. software package.
The NS2[27]-[59] and NS2[60-99] peptides representing aa segments 27–59 and 60–99 of NS2 of the Con1 strain (AC number AJ238799) were synthetized by Clonestar Biotech and purified by RP-HPLC (purity >98%).
CD, NMR spectroscopy, NMR-derived constraints and structure calculation, and molecular modeling and structure representation were performed by standard approaches as described in materials and methods S1.
The atomic coordinates for the NMR structures of peptides NS2[27]-[59] and NS2[60-99] and the NMR restraints in 50% TFE are available in the Research Collaboratory for Structural Bioinformatics (RCSB) Protein Data Bank under accession number 2KWT and 2KWZ respectively. The chemical shifts of all NS2[27]-[59] and NS2[60-99] residues have been deposited in the BioMagResBank (BMRB) under the accession number 16886 and 16892, respectively.
|
10.1371/journal.pntd.0003296 | The Relationship between Water, Sanitation and Schistosomiasis: A Systematic Review and Meta-analysis | Access to “safe” water and “adequate” sanitation are emphasized as important measures for schistosomiasis control. Indeed, the schistosomes' lifecycles suggest that their transmission may be reduced through safe water and adequate sanitation. However, the evidence has not previously been compiled in a systematic review.
We carried out a systematic review and meta-analysis of studies reporting schistosome infection rates in people who do or do not have access to safe water and adequate sanitation. PubMed, Web of Science, Embase, and the Cochrane Library were searched from inception to 31 December 2013, without restrictions on year of publication or language. Studies' titles and abstracts were screened by two independent assessors. Papers deemed of interest were read in full and appropriate studies included in the meta-analysis. Publication bias was assessed through the visual inspection of funnel plots and through Egger's test. Heterogeneity of datasets within the meta-analysis was quantified using Higgins' I2.
Safe water supplies were associated with significantly lower odds of schistosomiasis (odds ratio (OR) = 0.53, 95% confidence interval (CI): 0.47–0.61). Adequate sanitation was associated with lower odds of Schistosoma mansoni, (OR = 0.59, 95% CI: 0.47–0.73) and Schistosoma haematobium (OR = 0.69, 95% CI: 0.57–0.84). Included studies were mainly cross-sectional and quality was largely poor.
Our systematic review and meta-analysis suggests that increasing access to safe water and adequate sanitation are important measures to reduce the odds of schistosome infection. However, most of the studies were observational and quality was poor. Hence, there is a pressing need for adequately powered cluster randomized trials comparing schistosome infection risk with access to safe water and adequate sanitation, more studies which rigorously define water and sanitation, and new research on the relationships between water, sanitation, hygiene, human behavior, and schistosome transmission.
| Schistosomiasis is a serious disease in many developing countries, and the control of schistosomiasis relies on the large-scale administration of praziquantel. However, this strategy fails to address the root causes of schistosomiasis, which people acquire during contact with freshwater bodies that contain infected snails. It is suggested that improving access to clean water and sanitation reduces the risk of schistosomiasis transmission. Moreover, the use of soap, detergent, and endod (a berry sometimes used as a substitute for soap) might kill snails and the parasite larvae they excrete. We systematically reviewed the literature and performed a meta-analysis to study the association between people's access to clean water, sanitation, and good hygiene and the risk of schistosomiasis. People with access to clean water and adequate sanitation were at lower risks of schistosomiasis. No studies were found to explore the relationship between hygiene and risk of schistosomiasis. The difference in infection rates between people with and without access to clean water and sanitation varies widely between studies, suggesting that the impact of water and sanitation on schistosomiasis transmission is mediated by many other social and environmental factors. Further research is needed on the impact of water, sanitation and hygiene interventions for schistosomiasis control.
| More than 200 million people are estimated to be infected with schistosomes, among about 800 million at risk of schistosomiasis [1]. Three species of schistosome comprise the majority of these infections. Intestinal schistosomiasis is mostly caused by Schistosoma mansoni and Schistosoma japonicum, and the parasite eggs are released in the feces. In urogenital schistosomiasis, caused by Schistosoma haematobium, the eggs are released in the urine [2], [3]. Chronic intestinal schistosomiasis is manifested by debilitating symptoms, such as hepatosplenomegaly (enlargement of the liver and spleen) [2], [3], while S. haematobium is associated with an increased risk of developing bladder cancer [4], and thought to exacerbate the transmission of HIV and its progression to AIDS [5]. Both intestinal and urogenital schistosomiasis can cause anemia and malnutrition [6], and occasionally the eggs enter the central nervous system, causing symptoms such as seizures and focal neurological deficits [2], [3], [7].
Access to safe water and adequate sanitation are considered important components of schistosomiasis control, which at present largely relies on preventive chemotherapy with a single drug, praziquantel [8]. Adult schistosomes live within humans and, particularly in the case of S. japonicum, other mammals (e.g., water buffaloes) [9]. Aquatic snails (in the case of S. mansoni and S. haematobium) or amphibious snails (S. japonicum) act as intermediate hosts and release cercariae. People become infected during contact with infested water, when these cercariae penetrate through the skin. In turn, snails are infected by miracidia, which are released from eggs in the definitive host's urine or feces [2], [3]. Humans avoiding water contact and preventing urine and feces from entering freshwater bodies should therefore halt schistosome transmission. Furthermore, soap and endod (a natural soap substitute) are toxic to cercariae, miracidia, and snails, suggesting that their use may protect from schistosome infection, and thus implying a possible role for hygiene in schistosomiasis control [10], [11]. However, water, sanitation, and hygiene (WASH) are inadequate in large parts of low- and middle-income countries, where schistosomiasis is endemic [2], [3], [12]. Over the past 20 years, the need for multisectoral and integrated approaches to the control of schistosomiasis and other neglected tropical diseases (NTDs) has been emphasized [13]–[26]. Investigation of such approaches are particularly crucial as countries aim for elimination of schistosomiasis in line with World Health Assembly (WHA) resolution 65.19, put forward in May 2012.
Inadequate WASH is estimated to be responsible for 4.0% of deaths and 5.7% of disease burden worldwide, primarily driven by its role in the transmission of diarrheal disease and helminthiases [27]. The evidence for the impact of integrated control of NTDs is accumulating. In a 1978 study in St. Lucia, S. mansoni infection rates fell following the provision of safe water supply [28]. In the People's Republic of China, Wang et al. (2009) found that the integration of improved water and sanitation provision significantly reduced infections with the soil-transmitted helminths (STHs) Ascaris lumbricoides and Trichuris trichiura in addition to S. japonicum [29]. In Ethiopia, King et al. (2013) documented declines in S. mansoni and STH prevalences during a trachoma control program, which increased the use of improved water sources and latrines [30]. Asaolu and Ofoezie (2003) found sanitation and health education to be important interventions for the control of schistosomiasis and other helminthiases [31]. In Kenya, Freeman et al. (2013) quantified how a school WASH intervention reduced A. lumbricoides infection above provision of mass drug administration alone [32].
Relatively little evidence has been systematically collated and analyzed to inform policy-relevant discussions about the importance of including WASH as a part of schistosomiasis control. A previous review, conducted more than 20 years ago, identified four rigorous studies comparing schistosome infection rates with access to clean water, with a median reduction in schistosome morbidity for people with access to improved water supplies of 77% [33]. Many more relevant studies have been published since, providing the motivation for the current systematic review and meta-analysis.
We carried out a systematic review and meta-analysis of studies comparing Schistosoma infection rates in people with and without access (defined as the availability or use of) to safe water, adequate sanitation, and good hygiene, according to the ‘Meta-analysis Of Observational Studies in Epidemiology’ (MOOSE) guidelines [34], and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses' (PRISMA) statement [35]. Our protocol is available in Text S1, our MOOSE checklist in Text S2, and our PRISMA checklist in Text S3.
We systematically searched PubMed, Web of Science, Embase, and the Cochrane Library from inception to 31 December 2013. Two sets of search terms were developed: one for the diseases, and one for WASH. The standardized ‘improved’ water and sanitation definitions, developed by the World Health Organization (WHO) and UNICEF Joint Monitoring Programme (JMP) [12], were rarely used in the literature. Furthermore, studies seldom distinguished reliably between availability and use of WASH. Therefore, the categories of WASH ‘availability’ and ‘use’ were combined to form the category of ‘access to’. We considered all types of water in Box 1 as ‘safe’, and all types of sanitation in Box 1 as ‘adequate’. We considered ‘well’ to be a safe water source, except in Brazil, where ‘wells’ often consisted of pond-like water bodies, in contrast to the hand-dug wells in sub-Saharan Africa that are unlikely to contain snails or allow for water contact [36]. We considered use of soap during water contact as ‘good’ hygiene practice.
Search terms were combined as follows, such that each WASH term would be searched in conjunction with each disease term: (schistosomiasis or schistosome or schistosoma or bilharzia or bilharziasis or snail fever) and (water or borehole or standpipe or rainwater or sanitation or sanitary engineering or latrine or toilet or pit or open defecation or open urination or shower or laundry or hygiene or detergent or soap or risk factor). We did not use Medical Subject Headings (MeSH) terms since some WASH MeSH terms had been introduced only recently, and hence, relevant literature might have been missed during our search.
We also scanned the bibliographies of previous reviews pertaining to WASH and other NTDs [33], [37]–[39]. Additionally, when any study under consideration cited another which appeared to provide relevant evidence, the second study was eligible for inclusion. If a study demonstrated that eligible data had been collected but not reported, the authors were contacted and kindly asked to provide the data for further analysis.
Odds ratios (ORs) with 95% confidence intervals (CIs) for prevalence of Schistosoma infection according to availability of WASH were used as summary measures in all meta-analyses. Any paper reporting these directly, or providing data from which an OR with a 95% CI could be calculated (for instance 2×2 tables of numbers of people infected and not infected amongst those with and without access to safe water, adequate sanitation, or good hygiene, or sensitivities, specificities, and positive predictive values of these as diagnostics of Schistosoma infection), was eligible for inclusion. The searches were carried out without restrictions on language or year of publication.
Studies returned by the searches were screened independently by two assessors (JETG and DC), and disagreements were discussed until consensus was reached. First, the duplicates were removed. Next, titles, and then abstracts (of the remaining papers; if available) were reviewed in order to exclude papers whose titles or abstracts revealed that they were definitely not about WASH, not about human schistosomiasis, or did not contain data that would qualify for inclusion in the meta-analysis. Papers without abstracts or where abstracts were not available were reviewed in full.
The full texts of the remaining papers were sought from Imperial College London, the Swiss Tropical and Public Health Institute, the London School of Hygiene and Tropical Medicine, and the Wellcome and British Libraries. Those obtained were read by JETG and DC, and papers not reporting prevalence according to availability of water and/or sanitation were excluded. Papers in French, Portuguese, and Chinese were discussed with fluent speakers.
Data (2×2 tables where available, or ORs and corresponding 95% CIs) were extracted independently by JETG and DC from the papers, or (when supplied), from authors' correspondence. Crude or unadjusted ORs from bivariate analyses were taken where available, to minimize the risk of water supplies' impact being reported as due to the water contact they prevent, rather than due to the water supplies per se. Discrepancies were discussed and, if needed, a third person (JU) was consulted until consensus was reached. Where studies reported datasets from different settings, all datasets were eligible for inclusion. Where they reported different ORs for different forms of water or sanitation in the same setting, all ORs were included in the meta-analysis (double-counting some participants was felt to be preferable to the bias that would be induced by choosing one of the ORs). Where a 2×2 table contained one or more zeros, a Woolf-Haldane continuity correction was applied and 0.5 was added to all four of that table's elements [40].
Study quality was assessed using a checklist based on the GRADE approach [41] and other recent and similar systematic reviews [37]–[39]. Study assessment considered diagnostics (with sedimentation for intestinal schistosomiasis being rewarded due to its higher sensitivity compared with a single Kato-Katz thick smear reading) [42], method of assessment of WASH, correction for confounders, response rates, and other strengths and weaknesses (see Tables S2, S3, S4).
S. haematobium may be less susceptible than S. mansoni and S. japonicum to control with sanitation, since urination into water bodies is generally thought to be less easily controlled than open defecation [43], [44]. On the other hand, all human schistosomes infect people during contact with infested water, so we might expect water supplies to have a similar effect on infection with any schistosome species. We therefore pooled different species in the water meta-analysis, but carried out species-specific analyses for sanitation. The effect of species was subsequently investigated in the water sub-analyses. No studies reported data eligible for an analysis of hygiene and schistosomiasis.
The impact of WASH on schistosomiasis is likely to be mediated by a number of other factors, including behavioral and environmental ones, and aspects related to socioeconomic status (SES), which may vary between study settings. It is therefore reasonable to expect some variability in the true effect size between studies. Hence, random effects models [45] in StatsDirect version 2.8.0 (StatsDirect Ltd, Altrincham, United Kingdom) were employed in all the meta-analyses. These models weighted datasets' effect sizes by their inverse variances.
Publication bias was assessed through the visual inspection of funnel plots and through Egger's test [46]. Higgins' I2 was used to assess heterogeneity between studies [47]. Where heterogeneity was high (I2>75%) and a meta-analysis included at least one study of a different age group (adults, children, or mixed, with children defined as those below 18 years of age, or attending school), from a different continent, with a different schistosome species, with water in a different location, or with a different kind of sanitation, sub-analyses divided the datasets according to these factors to see if this reduced heterogeneity.
Sensitivity analyses were used to check for the impact of the largest studies on the three meta-analyses. All datasets from the study contributing the greatest weight to each meta-analysis was removed, and the effect on the results was investigated.
The searches and bibliographies of previous reviews returned 9,114 studies, 5,404 of which were unique (Figure 1). Finally, 44 relevant studies containing 90 datasets were identified. These 90 datasets consist of 54 datasets comparing safe water with schistosomiasis (35 on S. mansoni, 17 on S. haematobium, and two on S. japonicum), 24 comparing adequate sanitation with S. mansoni, and 12 comparing adequate sanitation with S. haematobium. No eligible studies compared sanitation with S. japonicum, or hygienic practices with Schistosoma infection rates, so meta-analyses were not conducted for these associations. A number of studies discussed related topics such as the survival of free-living schistosome stages, or the relationship between water supplies and water contact. However, these did not meet the inclusion criteria and were therefore excluded from the current review. The full list of excluded papers (along with reasons for exclusion) is found in Table S1.
Safe water sources were most commonly described as ‘tap’ or ‘piped’ (24 datasets), followed by ‘borehole’, ‘well’, or ‘standpipe’ (18 datasets), and ‘not using environmental water bodies such as rivers and lakes’ (four datasets), then by ‘adequate source of drinking water’ (three datasets), not using ‘unsafe’ water, and ‘domestic drinking water’ (two datasets each). The remaining dataset referred to ‘clean household water’. In the sanitation and S. mansoni analysis, adequate sanitation was mostly described as a ‘latrine’ (12 datasets), followed by ‘latrine or flush toilet’ (six datasets). Two datasets referred to each of ‘septic tank’ or ‘cesspool’, ‘sewer connection’ or ‘latrine’, and ‘sewerage’. In the sanitation and S. haematobium analysis, adequate sanitation was most commonly described as a ‘latrine’ (eight datasets), followed by ‘latrine’ or ‘flush toilet’ (two datasets), and finally by ‘septic tank’ or ‘cesspool’, then ‘toilet’ (one dataset each).
Studies most commonly included children and adults (21 studies). Another 19 studies included only children (i.e. individuals below the age of 18 years, or in school), while four studies were of adults only. The included studies were most commonly from Africa (21 studies). Another 17 studies were from Brazil. The remaining six studies were from Asia (four in Yemen and two in the People's Republic of China). The most common language was English (40 studies), and the remaining four studies were in Portuguese. Three studies had case-control designs, while the remaining 41 contained descriptive cross-sectional data. Study quality was generally low, with water and sanitation rarely being defined in a uniform way, or assessed through inspections. Furthermore, very few studies provided data split according to confounders such SES. Of the 21 studies whose authors were contacted, data were only provided for the studies by Knopp et al. (2013) [48], [49], Arndt et al. (2013) [50], Fürst et al. (2013) [51], and Sady et al. (2013) [52].
This is the first systematic review of the association between WASH and Schistosoma infection. Access to safe water supplies were found to be associated with significantly less infection with S. haematobium, S. mansoni, and S. japonicum, while adequate sanitation was found to be associated with significantly less infection with both S. mansoni and S. haematobium. No observational studies were found assessing the association between good hygiene, defined as the use of soap during water contact, and Schistosoma infection.
Since schistosome cercariae are susceptible to water treatment and even to water storage [92]–[95], it is reasonable to assume that piped water should not pose a risk of transmission. Thus the ability of safe water sources to prevent Schistosoma infection would depend on how well they prevent dermal contact with schistosome-infested environmental water bodies. Jordan et al. (1975) found that provision of piped water to the household was much more effective than centralized community access in preventing water contact and reducing schistosomiasis transmission [96]. However, we found similar ORs for household access and community access (OR = 0.57, 95% CI: 0.28–1.17 for household access rather than use of environmental water bodies, and OR = 0.60, 95% CI: 0.47–0.76 for community access rather than use of environmental water bodies). We identified two observational studies comparing schistosome infection rates in people household access and community access [97], [98], but again, neither study reported significantly lower infection rates in people with household rather than community water supplies. These studies were not included in the meta-analyses since both household and community water sources were ‘safe’, and thus these studies did not meet our inclusion criteria.
Schistosome eggs are released in the urine and the feces of human hosts, but to sustain transmission an egg must enter freshwater and hatch to release a miracidium, which then infects an intermediate host snail [3]. This infected host snail will later release cercariae, which may infect people coming into contact with the water. Thus sanitation's impact upon schistosome transmission is dependent upon its ability to reduce fecal or urinary contamination of freshwater containing intermediate host snails, rather than contamination of the environment in general. Furthermore, owing to exponential reproduction of the parasite within the intermediate host snail, even small numbers of schistosome eggs entering freshwater may give rise to a disproportionately large risk of infection in people coming into contact with that water [95].
The high heterogeneity throughout the meta-analyses could not be attributed to differences in any one of: the schistosome species, ages of study participants, type of sanitation, location of water source, or geography of study (stratified by continent). Perhaps such heterogeneity could be due to a combination of many setting-specific community, ecological, and occupational factors such as the above, presence of intermediate snail hosts, and reasons for water contact, and the input of miracidia into the water. A recent geographical analysis of national survey and demographic health survey data found absence of piped water to be associated with significantly increased infection with S. haematobium, but absence of a toilet facility to be associated with a significantly lower odds of S. mansoni infection [99]. These findings perhaps reflect that the aforementioned factors are much stronger predictors of infection than WASH, and for example people without adequate WASH may remain uninfected due to a lack of snail intermediate hosts in the locality. Similarly, some studies in this meta-analysis may have included people with inadequate WASH who were nevertheless not exposed to schistosomiasis due to a lack of intermediate host snails nearby, or people with adequate WASH but nevertheless exposed to schistosomes, for example during activities such as fishing.
The lack of observational studies comparing Schistosoma infection in people who do and do not use soap during water contact, perhaps reflects the fact that hygienic practices can be more temporary than access to water or sanitation infrastructure. However, in Ethiopia, Erko et al. (2002) found after distribution of soap bars containing endod, the prevalence of schistosome infection in women dropped significantly [100].
In our view, the biggest limitation of the current meta-analysis is the possibility of socioeconomic confounding. Farooq et al. (1966) found that people without latrines showed a higher prevalence of schistosome infection, but that this difference was no longer apparent if the analysis was carried out separately for the sub-populations living in houses made of mud, or bricks, respectively [53]. The authors concluded that the higher infection rates were due to lower SES, which could be measured by house construction or by access to sanitation, rather than any reduction of schistosomiasis transmission arising due to sanitation. Similarly, in Brazil, Gazzinelli et al. (2006) found significantly higher infection rates in households without either a motorcycle or a car, another indicator of low SES [101]. Safe water supplies are also more prevalent amongst those of higher SES, meaning that possible confounding by SES potentially runs through all the meta-analyses presented here. On the other hand, WASH can depend on environmental and other factors, in addition to SES. An example is provided by Barbosa et al. (2013) [102], who compared two rural Brazilian communities and found better sanitation in the community of lower SES. Unfortunately, very few studies reported data that were stratified according to, or adjusted for, SES.
Most of the studies containing data used in the meta-analysis were multivariable analyses, which analyzed the importance of various risk factors (including absence of water and sanitation) for Schistosoma infection. As such, they were not focussed on WASH and often did not precisely define the water and sanitation available to, or used by, participants, or indeed distinguish between availability and use of safe and adequate WASH. Regarding sanitation, it was rare for a study to define where latrines or toilets drained to, and we may have therefore included some studies where the adequate sanitation drained directly into lakes or rivers, facilitating schistosome transmission. Very few studies carried out quality control of the schistosomiasis diagnosis (e.g., reading a random sample of 10% of Kato-Katz thick smears by a senior laboratory technician), and none carried out quality control of the WASH data collection (e.g., spot checks whether reported data on availability and use of sanitation are correct).
WASH not being the focus of most studies also raises the possibility of a weak publication bias (the funnel plots and Egger's tests suggested that this was unlikely, but not impossible), and it has also led to imprecisely defined WASH (particularly in the case of those lacking the safe water or adequate sanitation of interest). WASH was always assessed through questionnaires rather than direct inspection. Furthermore, the included studies always compared WASH directly with schistosome infection rates. New research of the relationship between WASH, human exposure through water contact, human contamination of freshwater, cercarial, miracidial and snail populations, and infection rates is needed, in order to provide a deeper understanding of the relationship between WASH and the transmission likelihood of schistosomes. Very few studies reported WASH in a way that allowed for comparison with the JMP definitions [12]. This observation is explained by the fact that the JMP definitions were first put forward only in 2000 [103], and have been further developed subsequently. Many of our included studies were conducted before this. Furthermore, people's use of different water supplies and sanitation may vary with activities and season [91], and therefore the dichotomisation of water supplies into ‘safe’ and ‘unsafe’, and of sanitation into ‘adequate’ and ‘inadequate’ risks oversimplifying access to WASH. (Box 2).
Data on infection with different intestinal parasites was often aggregated, with WASH variables presented as risk factors for infection with any parasite. In many cases E-mail addresses were not available, or we received no replies. We were therefore unable to include these studies, despite the fact that the authors had collected data that would qualify for inclusion.
Water contact and thus schistosome transmission, typically takes place outside the home (public exposure), not within the household (domestic exposure) [104], [105]. The individual is exposed to cercariae released by snails infected not just by him- or herself but also by his or her neighbors. With this in mind, one may expect the associations between water, and particularly sanitation, to be most strongly associated with schistosome infection at the community- rather than the household-level, as has been suggested for other diseases [106]. However, very few such analyses have compared schistosome infection rates between communities with different levels of WASH. Yang et al. (2009) did adopt such an approach and found S. japonicum infection rates to be significantly lower in communities where more than 50% of people used ‘hygienic lavatories’ [63].
A meta-analysis of observational studies found both safe water supplies and adequate sanitation to be associated with significantly lower odds of Schistosoma infection. This meta-analysis lends support to more consideration of environmental factors and living conditions in schistosomiasis control, and adds to the growing body of evidence about the relationship between WASH and NTDs. Previous meta-analyses have found significant associations between sanitation and STH infection [37], WASH and STH infection [39], and WASH and trachoma [38]. However, the possible confounding caused by factors such as SES shows that adequately powered cluster randomized controlled trials assessing the impact of WASH on human behavior and schistosome infections, and cercarial, miracidial and snail populations, must play an integral role in informing future policy-making. Such studies are needed to inform the potentially crucial role that WASH could play in the elimination of schistosomiasis, in line with World Health Assembly resolution 65.19.
|
10.1371/journal.pbio.1002011 | Signal Peptide-Binding Drug as a Selective Inhibitor of Co-Translational Protein Translocation | In eukaryotic cells, surface expression of most type I transmembrane proteins requires translation and simultaneous insertion of the precursor protein into the endoplasmic reticulum (ER) membrane for subsequent routing to the cell surface. This co-translational translocation pathway is initiated when a hydrophobic N-terminal signal peptide (SP) on the nascent protein emerges from the ribosome, binds the cytosolic signal recognition particle (SRP), and targets the ribosome-nascent chain complex to the Sec61 translocon, a universally conserved protein-conducting channel in the ER-membrane. Despite their common function in Sec61 targeting and ER translocation, SPs have diverse but unique primary sequences. Thus, drugs that recognise SPs could be exploited to inhibit translocation of specific proteins into the ER. Here, through flow cytometric analysis the small-molecule macrocycle cyclotriazadisulfonamide (CADA) is identified as a highly selective human CD4 (hCD4) down-modulator. We show that CADA inhibits CD4 biogenesis and that this is due to its ability to inhibit co-translational translocation of CD4 into the lumen of the ER, both in cells as in a cell-free in vitro translation/translocation system. The activity of CADA maps to the cleavable N-terminal SP of hCD4. Moreover, through surface plasmon resonance analysis we were able to show direct binding of CADA to the SP of hCD4 and identify this SP as the target of our drug. Furthermore, CADA locks the SP in the translocon during a post-targeting step, possibly in a folded state, and prevents the translocation of the associated protein into the ER lumen. Instead, the precursor protein is routed to the cytosol for degradation. These findings demonstrate that a synthetic, cell-permeable small-molecule can be developed as a SP-binding drug to selectively inhibit protein translocation and to reversibly regulate the expression of specific target proteins.
| All cells are highly crowded with proteins that, once synthesized, have to reach their proper subcellular location in order to maintain the cellular homeostasis. Approximately 30% of the proteome needs to be sorted from the cytosol and inserted into, or transported through, biological membranes. For proteins sorted via the secretory pathway, an important step is the translocation into a cellular compartment called the endoplasmic reticulum (ER). The cell uses an elegant way to discriminate proteins that need to be translocated into the ER from those that have to reside in the cytosol by scanning for the presence of an N-terminal ER-entry tag. Although these tags, called signal peptides, have a common structure, they each contain a unique hydrophobic peptide sequence. In this work, we describe how a small chemical drug, CADA, can bind to one specific signal peptide present in the human CD4 pre-protein. We show that by influencing the signal peptide orientation in the translocation channel located in the ER membrane, CADA prevents CD4 translocation into the ER lumen. As a consequence, the CD4 protein is not properly synthesized and routed to the cell surface, resulting in a clear reduction in the amount of surface CD4, a membrane protein found on immune cells, and implicated in HIV-infection and other diseases. We believe that other drugs can be designed to selectively regulate, in a similar way, ER translocation of specific target proteins.
| CD4 is a type I integral membrane glycoprotein that is expressed on the surface of thymocytes, T-helper lymphocytes, and cells of the macrophage/monocyte lineage [1]. It plays a central role in immune responses but also represents an obligatory component of the cellular receptor complex for HIV [2],[3]. Several reports demonstrate that down-modulation of surface CD4 protects cells from HIV infection [4]–[8]. In addition, natural CD4 down-modulation by memory CD4+ T cells in vivo protects African green monkeys from developing AIDS after infection with simian immunodeficiency virus (SIV), while maintaining the immunological functions normally attributed to CD4+ T cells [9]. Reduction in surface CD4 can be elicited by several factors that interfere with its translation or intracellular trafficking (reviewed in [10]). Phorbol esters are known to induce CD4 endocytosis through serine phosphorylation of the cytoplasmic tail of CD4 [11]. The concerted action of the three HIV-1 proteins Nef, Env, and Vpu results in a complete removal of CD4 from the surface of HIV infected cells through (i) enhanced routing of CD4 to the endoplasmic reticulum (ER) degradation pathway [12],[13] and (ii) activated endocytosis and lysosomal degradation [14],[15].
Surface expression of type I transmembrane (TM) proteins, such as CD4 receptors, requires translation of precursor proteins and their insertion into the ER membrane for subsequent routing to the cell surface. This co-translational translocation pathway begins when a hydrophobic N-terminal signal peptide (SP) on the nascent protein emerges from the ribosome and is recognized by the signal recognition particle (SRP). This complex of ribosome, nascent chain, and SRP is then targeted to the ER membrane via the interaction between SRP and its membrane receptor. Subsequently, the ribosome tightly binds to the Sec61 complex in the ER-membrane, a protein-conducting channel composed of the membrane proteins Sec61α, Sec61β, and Sec61γ. Finally, the ribosome continues the translation and the elongating polypeptide chain moves directly from the ribosome exit tunnel into the associated membrane channel. When the TM domain within the nascent polypeptide chain reaches the Sec61 complex, the channel opens laterally and the membrane anchor is released into the lipid bilayer (reviewed in [16],[17]). Simultaneously with the translocation of the polypeptide chain, cleavage of the signal sequence occurs at the luminal side of the ER together with other possible modifications such as N-glycosylation and proper folding of the polypeptide.
A screen for anti-HIV drugs led to the identification of CADA, a cyclotriazadisulfonamide with broad spectrum antiviral activity against laboratory strains and clinical isolates of HIV-1, as well as HIV-2 and SIV [18],[19]. The anti-HIV activity of CADA and its analogues correlated with their ability to down-modulate cell surface CD4 expression [8]. In the present study, we focused on the mechanism of action and molecular target of CADA. We demonstrate that CADA inhibits CD4 biogenesis during the early translational steps. More specifically, we show that (i) CADA selectively binds to the SP of human CD4 (hCD4), (ii) CADA prevents the growing CD4 polypeptide from entering the lumen of the ER, (iii) the SP of hCD4 is first inserted into the translocon channel with its N-terminus facing the lumen of the ER (Nlum/Ccyt) before an obligate inversion into a Ncyt/Clum topology takes place, and (iv) CADA locks the SP of hCD4 in an intermediate position during inversion and prevents further translocation of the polypeptide chain into the ER lumen.
To evaluate the effect of CADA on CD4 expression, different cells were treated with the compound under various conditions (Figures 1 and S1). CADA induced a dose-dependent down-modulation of hCD4 regardless of the cell background in which it was expressed, i.e., in primary T-cells and T-cell lines that express the receptor naturally, as well as in transfected cells stably expressing CD4 (Figure 1B–1D). This down-modulating effect appeared to be reversible (Figure 1G), and selective for CD4: in a set of 14 different surface receptors CADA inhibited only CD4 (Figure 1E and 1F). The sensitivity of CD4 for CADA was species-specific: the compound did not affect the expression of mouse CD4, whereas primary T-cells from macaques responded to the compound in a similar way to human T-cells (Figure 1D). Thus, CADA selectively and reversibly down-modulates CD4 of primate origin in a dose-dependent manner.
To elucidate how CADA interferes with CD4 protein expression, we analyzed the impact of CADA on the life cycle of CD4. A kinetic study with CADA and the phorbol ester PMA, a drug known to induce rapid endocytosis and degradation of surface CD4 [11], suggested an effect of CADA on CD4 translation or transport to the cell surface. As shown in Figure 1H, the appearance of newly synthesized CD4 molecules at the surface was prevented by CADA. To test directly whether CADA interferes with the de novo synthesis of CD4 we performed pulse-chase experiments followed by immunoprecipitation for CD4. CADA profoundly inhibited CD4 synthesis in CHO.CD4+ cells as indicated by the absence of this protein band in CADA-treated samples (Figure 2A and 2B). Although CADA strongly affected CD4, analysis of the total cellular protein extract revealed no general inhibition of protein synthesis (Figure 2A, lanes 3 and 4). We further analyzed this CD4-specificity of CADA by examining protein synthesis in different cellular compartments. Both cytosolic and membrane fractions were investigated in CD4-negative and CD4-positive CHO cells. The expression of cytosolic proteins was not altered by CADA-treatment, both in the absence or presence of CD4 (Figure 2C and 2D). In contrast, addition of the translation elongation inhibitor cycloheximide (CHX) resulted in an almost complete protein shut-down. Interestingly, CADA seemed not to affect other membrane proteins in the CD4-negative cells. Also, a similar membrane protein expression pattern was observed in DMSO and CADA-treated CD4+ CHO cells, except for one protein band migrating around 80 kDa. As this protein band could not be detected in the CD4-negative cells, we concluded that the affected protein was most likely CD4-YFP, an 80 kDa fusion protein that is stably and highly expressed in these CHO cells. The significant reduction in synthesis of membrane-associated proteins in CADA-treated CD4+ CHO cells (Figure 2D) can be ascribed to the complete block of CD4-YFP production. To isolate the glycosylated surface proteins from the membrane fraction that contains proteins not only from the surface membrane but also from different intracellular organelles and export pathways, we used Concanavalin A (ConA) agarose beads. Again, CD4-YFP was strongly inhibited by CADA, whereas the expression of other glycosylated membrane proteins was not affected (Figure 2E). In addition, CADA did not inhibit the secretion of proteins into the culture medium (Figure 2E). Similar observations were made in other CADA-treated cells, such as U87 and SupT1 (Figure S2). Finally, down-modulation of CD4 occurred post-transcriptionally, because CD4 messenger RNA levels were not altered by the compound (Figure S1F). From these data, we concluded that CADA has a high selectivity for the protein synthesis of hCD4.
Next, we established which domain of CD4 is required for drug sensitivity by investigating C-terminal deletion mutants (Figure 3). Deletion of the cytosolic tail of CD4, which is involved in signal transduction and endocytosis of the receptor [11],[20], did not affect its sensitivity to CADA (Figure 3A, mutant hCD4-426). Exchanging the extracellular D3, D4, and transmembrane (TM) domains of CD4 with a related type I TM protein such as the alpha chain of CD8, expression of which is not affected by CADA (Figure 1F), failed to prevent CADA-induced down-modulation (Figure 3A, mutant hCD4-CD8). Furthermore, human/mouse chimaeric fusion constructs excluded a potential role for the immunoglobulin-like domains D1 and D2 of hCD4 in CADA-sensitivity. In agreement with primary murine T-cells (Figure 1D), wild-type mouse CD4 (mCD4 WT) did not respond to the down-modulating activity of CADA, whereas hCD4 containing either mouse D1 or mouse D2 did (Figure 3B). These data narrowed down the CADA-sensitive region of hCD4 to the N-terminal SP and the first seven N-terminal residues of the mature protein, as this is the only remaining fragment of hCD4 common to all CADA-susceptible constructs (Figures 3 and S1E).
Previous experiments demonstrated an inhibitory effect of CADA on the expression of membrane-anchored proteins that contained the SP of hCD4. However, removal of the TM domain of CD4 did not diminish the sensitivity of the protein to CADA (Figure 4B), showing that CADA-susceptibility is not membrane anchor-dependent. Therefore, in our further study we also included smaller secreted proteins (Figure 4A).
In general, precursors of type I TM proteins (e.g., CD4) contain an amino-terminal SP that is involved in the early steps of biogenesis such as targeting of the nascent polypeptide to the ER membrane for co-translational translocation [21],[22]. To determine how CADA interferes with CD4 synthesis, translation and translocation of CD4 were analyzed in vitro using cell-free rabbit reticulocyte lysate with or without pancreatic rough microsomes (RMs). In the absence of RMs, translation of full length and truncated hCD4 was unaffected by CADA (Figure 4C, lanes 1 and 2, open arrowhead). However, in the presence of RM, translocation of hCD4 into the RM lumen was markedly inhibited by CADA as indicated by a reduction in glycosylated WT hCD4 (Figure 4C, lanes 4 and 7, star) and SP-cleaved truncated hCD4 (Figure 4C, lanes 4 and 7, solid arrowhead) that were resistant to degradation by added proteases (Figure 4C, lanes 5 and 8). In agreement with the in cellulo results of Figure 4B, the in vitro translocation of full length (hCD4 WT) and truncated secreted (hCD4-201) CD4 was dose-dependently inhibited by CADA (Figure 4D). In contrast, the translocation of two control molecules, mCD4-186 and bovine pre-prolactin (pPL), was not affected by CADA at any concentration tested (Figure 4D and 4E).
We next prepared two chimaeric constructs (Figure 4A) in which the mature domain of pPL was fused either directly to the SP of hCD4 ([hCD4]-pPL) or to the seventh residue of mature hCD4 ([hCD4]-(7)-pPL). Translocation of [hCD4]-pPL was profoundly inhibited by CADA (Figure 4D and 4E). While this mutant was slightly less sensitive to the compound than WT CD4, this confirms that sensitivity to CADA is determined primarily by the hCD4 SP. However, the presence of seven additional amino acids of the N-terminal CD4 mature domain enhanced CADA sensitivity, resulting in dose-dependent translocation inhibitory levels similar to WT CD4 (Figure 4E, construct [hCD4]-(7)-pPL). Furthermore, fusing these N-terminal 32 residues of hCD4 to a non-SP containing yellow fluorescent protein (YFP) resulted in translocation of a non-CD4 related cytosolic protein into the RM lumen and full susceptibility to CADA (Figure S3A–S3C). These data show that inhibition of protein translocation by CADA is specific to the SP and the first seven N-terminal residues of mature hCD4.
To investigate if CADA inhibits CD4 translocation into the ER in vivo, we performed experiments to rescue cytosolic CD4 in HEK293T cells by inhibiting proteosomal degradation with MG132. In order to detect all precursor forms of CD4, we generated an hCD4 construct that contained the simian virus 5 (V5) epitope at the N-terminal end of the mature protein (outlined in Figure S3E). Only when the cells were treated with the combination of CADA and the proteasome inhibitor MG132, a higher molecular form of CD4 could be detected that corresponded to the precursor form of the V5-tagged CD4 (Figure 4F). Accordingly, for [hCD4]-YFP proteasome inhibition with MG132 also rescued the precursor form in CADA-treated samples (Figure S3D). These in vivo results indicate that CADA diverts CD4 synthesis towards cytosolic proteosomal degradation.
We then looked for direct interaction between CADA and the hCD4 SP. Chemically synthesized SPs of hCD4 and mCD4 were captured on a streptavidin sensor chip and examined by surface plasmon resonance (SPR). Selective and dose-dependent binding of CADA to hCD4 SP was observed, and almost no binding to mCD4 SP (Figures 5A and S4A). Also, CADA did not show interaction with the SP of bovine pPL (Figure S4C). As a control, SRP interaction with the SPs was evaluated and revealed equal binding profiles of SRP to human and mouse SP, thus excluding non-functionality of mSP or insufficient peptide-coupling to the chip (Figures 5B and S4B). In addition, the lack of hCD4 SP binding by the structural analog MFS105 (Figure 5C), a CADA-derivative with no CD4 receptor down-modulating activity (Figure S4E) [8], further strengthens the selectivity of CADA for hCD4 SP.
SPs display a general structure consisting of three regions: a (mostly) positively charged N-terminus (N-region), a central hydrophobic alpha helical region (H-region), and a more polar C-terminal part (C-region) that includes the SP cleavage site [23]. On the basis of the unresponsiveness of mouse CD4 to CADA (Figures 1D, 3B, 4D, and 4E), we generated human/mouse chimaeric constructs with exchanged SP subregions in order to analyze the contribution of each subregion of hCD4 SP to CADA-susceptibility (Figure 5D). Exchanging the N-region of the SP (construct [mhh]-hCD4) slightly decreased the sensitivity of CD4 to CADA, as analyzed by flow cytometry (Figure 5E). A stronger reduction in susceptibility was observed when the C-region of mSP and the first seven N-terminal residues of mature mCD4 were inserted in hCD4 (Figure 5E, construct [hhm]-hCD4). Swapping the hydrophobic H-region had a major negative impact on the sensitivity to CADA (Figure 5E, construct [hmh]-hCD4). Reversibly, insertion of the hydrophobic alpha helix of hCD4 SP into CADA-insensitive SPs, such as those of murine CD4 or bovine pPL, significantly enhanced their responsiveness to CADA (Figure S5C and S5D). In addition, the different inhibitory levels of CADA on surface expression of the human-murine chimaeras could be linked to different degrees in ER translocation inhibition (Figure S5A and S5B). Although a more than 10-fold reduction in CADA-activity was noted in the in vitro assay as compared to the in cellulo data (Figure 5E versus S5B), the chimaeras responded to CADA in the same relative order. These data suggest contributions from all three hCD4 SP subregions to CADA-sensitivity, but show a crucial role for the hydrophobic H-region.
To identify the step at which CADA interferes with the translocation of hCD4, we analyzed the targeting and translocation of [CD4]-pPL nascent chains. Transcripts lacking a stop codon were truncated at sequential sites in the mature domain (Figure 6A) and translated in vitro. Via an intact peptidyl-tRNA bond the nascent chains remain attached to the ribosomes and are synchronized at a defined length. Addition of pancreatic RMs will result in docking of the different ribosome-nascent chain complexes (RNCs) to the ER membranes, giving translocation intermediates that represent static snapshots of the movement of the SP within the translocon channel. Subsequently, the nascent chains are released from membrane-bound ribosomes by puromycin, allowing for a controlled translocation into the ER lumen.
Nascent chains of 80 residues (80-mers) were translated in the absence of RM (Figure 6B, left panel, first lane). These chains are bound to the ribosome as peptidyl-tRNAs with about 30 residues buried inside the ribosome. Protease treatment of these RNCs will degrade the N-terminal part of the peptide that is exposed to the cytosolic compartment (about 50 residues), and will generate a faster migrating protein band on the gel (Figure 6B, lane 6). Addition of RM to the nascent chains will allow for targeting of the RNCs to the membrane and insertion of the SP into the translocon channel. If well-targeted, nascent chains will be shielded from exogenous protease because of a tight interaction between the ribosome and the translocon after transfer from SRP [24],[25], and appear as intact RNCs after proteinase K (PK) treatment. Equal protease-protected peptidyl-tRNA bands were observed in control and CADA samples, ruling out an inhibitory activity of CADA on targeting and transfer of the nascent chains to the translocon (Figures 6B, lanes 7 and 8, arrow). Addition of puromycin to the targeted chains induced the release of the nascent chains from the ribosome, with subsequent translocation of the peptides into the PK-protected RM lumen and cleavage of the SP (Figure 6B, lanes 4 and 9, solid arrowhead). However, in the presence of CADA, very few cleaved species were observed (Figure 6B, left panel, lane 5, solid arrowhead), indicating a profound inhibition by CADA on the co-translational translocation of pPL species that contained the SP of hCD4 but not of those with the SP of WT pPL (Figure 6B, left and right panel, respectively). This inhibition was also dose-dependent (Figure 6C). Translocation inhibition by CADA was observed at all chain lengths, but was most effective on nascent chains up to 80 residues (Figure S7).
Remarkably, we could delay the administration of CADA to the membranes until targeting was completed. A similar inhibitory effect on the translocation of 80-mers was recorded when the compound was applied to the RM before targeting or to the RNC/RM mixture 15 minutes after initiation of targeting (Figure 6D). Taken together, these results indicate that CADA acts at a step after the targeting and transfer of the nascent chains to the translocon, but before the growing peptide chain has reached the luminal side of the membrane.
Although N-terminal signal sequences are generally considered to insert in a tail-first, hairpin-looped topology [26],[27], hydrophobic sequences can also enter the translocon in a head-first configuration or reorient within the translocon channel [28]–[32]. To determine if CADA might have a role in changing the topology of the SP during translocation, we introduced a diagnostic N-glycosylation site at the N-terminus of the SP (Figure 7A). Head-first N-terminal translocation of the SP will give rise to glycosylated species, whereas tail-first C-terminal orientation will result in SP-cleaved forms.
The hCD4 SP was extended with 17 residues based on a construct used in other studies [33],[34], and either contained or lacked (control) an N-linked-glycosylation site (Figure 7A). Full length proteins with the extended hCD4 SP translocated well, irrespective of the presence of the glycosylation site, and responded to CADA in a dose-dependent way (Figure S8A). Through analysis of nascent chains we investigated if and when head-first translocation occurred, and determined the stepwise movement of the extended SP within the translocon (Figure 7B). At shorter truncations (i.e., 17+58-m), a substantial fraction of the nascent chains were glycosylated, indicating that targeting may begin with the N-terminus of the SP in the ribosome/translocon complex facing the RM lumen. At this chain length, the binding of the RNC to the translocon was already highly stable as evidenced by its high-salt resistance (Figure S8E). Elongation of the mature part of the polypeptide with four residues (17+62-m) resulted in a profound increase in glycosylated RNCs (Figure 7B, second lane of third panel, star). At longer truncations, the glycosylation efficiency decreased gradually and was almost undetectable for the 17+88-mer. Interestingly, administration of CADA to the RNCs had very little effect on the N-terminal glycosylation of the shorter chains (17+58-m) and the high-salt resistant binding of the RNC to the translocon (Figure 7B, third lane of second panel, and Figure S8E). However, the inhibitory activity of CADA increased significantly when the C-terminus of the polypeptide was extended, and resulted in non-detectable glycosylation levels for chains with a minimum length of (17+71) residues (Figure 7B). Notably, administration of CADA to the 17+71-mers resulted in a loss of the high-salt resistant binding of the RNC to the translocon (Figure S8E), suggesting that with the compound the positioning of the nascent chain in the channel was altered so that the peptide tether could allow the ribosome to dissociate from the channel. Analysis of the puromycin-treated samples revealed a similar N-terminal glycosylation pattern for the released chains as for the intact RNCs (Figure 7C and 7D). Again, the inhibitory activity of CADA on N-terminal glycosylation gradually increased with growing chain length for chains up to (17+71) residues (Figure 7D). Breaking the peptidyl-tRNA bond with puromycin also allowed for translocation of the chains into the RM lumen with subsequent SP cleavage. SP-cleaved species were hardly detectable for the shorter chains, irrespective of the presence of a glycosylation site (Figure 7C and 7E, blue line). However, translocated SP-cleaved chains that were PK-protected first appeared for the 17+80-mers, i.e., a chain length at which a profound decrease in N-terminal glycosylation was noted (Figure 7C–7E). This suggests that at this specific chain length the SP is mainly positioned in the hairpin-looped topology with the C-terminus facing the RM lumen. In accordance with the data from the WT SP-containing chains (Figure S7B), maximum C-terminal translocation and SP-cleavage were noted for the 17+80-mers, thus when the distance from the SP cleavage site to the ribosome peptidyltransferase centre (PTC) had reached a chain length of about 55 amino acids. Further extension of the polypeptide chain diminished the C-terminal translocation efficiency that was consistently blocked by CADA (Figure 7E). Notably, for the 17+80-mers CADA strongly inhibited both the (weak) N-terminal glycosylation and the C-terminal translocation and SP-cleavage with equal efficiency, in a dose-dependent manner (Figure S8C and S8D).
These results together show that the nascent chains need a minimum length in order for CADA to exert its inhibitory effect on glycosylation, whereas at all lengths where SP-cleavage can be observed translocation with SP-cleavage is inhibited. Thus, on one hand, CADA hinders vertical positioning of the polypeptide with the N-terminus faced to the ER lumen for efficient glycosylation and, on the other hand, disturbs the completion of SP inversion for a hairpin-looped structure that can be SP-cleaved (Figure 7F). This suggests that in the presence of CADA the SP is held in a folded conformation in the translocon channel, so that the polypeptide is prevented from reaching the luminal side of the ER.
In this study we have characterized CADA, a small-molecule HIV entry inhibitor as, to our knowledge, the first SP-binding drug that selectively inhibits hCD4 protein translocation into the ER in a SP-dependent way.
CD4 is a type I membrane protein expressed on the surface of a subset of immune cells [1]. Sorting of this protein to the plasma membrane requires that the CD4 pre-protein contains a cleavable SP, an obligatory component for protein targeting to the ER through co-translational translocation. This synthetic pathway begins when a hydrophobic N-terminal SP emerges from the ribosome, is recognized by the SRP, and targets the whole RNC to the ER membrane so that the emerging polypeptide is inserted into the Sec61 translocon channel [16],[17]. Simultaneously with the translocation of the polypeptide chain, cleavage of the SP occurs at the luminal side of the ER. Here, we revealed that targeting of hCD4 to the ER is initiated with a head-first N-terminal insertion (Nlum/Ccyt) of the SP in the translocon, before it inverts into an obligate Ncyt/Clum topology necessary for SP cleavage during C-terminal translocation. Thus, initial insertion of hCD4 SP in the channel occurs differently from the more widely accepted hairpin positioning (Ncyt/Clum) of cleavable N-terminal signals [26],[27],[35]. Head-first topology of non-cleavable N-terminal sequences has been proposed by the Spiess group for signal-anchor (SA) sequences that anchor the polypeptide in the bilayer [29]–[31],[36]. Recently, a detailed study with an N-terminal type II SA sequence confirmed the head-first insertion of the SA into Sec61α before inversion from type I to type II topology takes place [32]. It is plausible that cleavable N-terminal SPs insert and invert in a similar way as type II SA sequences. Moreover, for a relatively short nascent polypeptide that is C-terminally attached to the ribosome PTC, the first possible way for the N-terminal SP to interact with the translocon is most likely a head-first (Nlum/Ccyt) orientation. Initial positioning of signal sequences may also be directed by interaction with cytosolic chaperones other than SRP, and translocon-associated proteins. Recently, Sec62 has been found to mediate the orientation of SA proteins depending on the hydrophobicity of the SA sequence [37]. It would be interesting to explore if there are other type I membrane proteins with initial Nlum/Ccyt insertion and to investigate if the factors that determine this early positioning of cleavable SP are different from those reported for N-terminal SA sequences.
Reorientation of SA sequences is driven by flanking charges according to the positive-inside rule [38],[39] and inhibited by increased hydrophobicity of the SA sequence [40]. For hCD4 the two adjacent lysine residues immediately down-stream of the SP cleavage site might overrule the more dispersed positive charge of the N-domain (two arginine residues R3 and R8) and initially orient the SP according to this positive-inside rule with its C-terminus to the (negatively charged) cytosolic side of the membrane. In line with the canonical SA derived from the first TM segment of aquaporin 4 [32], hCD4 SP initiates ER targeting at a nascent chain length of 58 residues, when the size of the polypeptide between the SP C-terminus and the PTC is about 33 residues. Inversion of hCD4 SP and translocation of its C-terminal segment with SP-cleavage were observed upon lengthening the polypeptide chain (at its C-terminus) with 22 amino acids, a stretch theoretically long enough to span the bilayer membrane and thus to allow for a hairpin looped positioning of the SP. For our study we extended the N-terminal hydrophilic domain of hCD4 SP with 17 residues as this was the minimum length required to obtain optimal glycosylation of the N-terminus (Figure S9A–S9C), because of the distance between the active site of the oligosaccharyl transferase complex and the inner bilayer of the ER membrane [41]. Extension of the N-domain of SA sequences with more than 20 residues can change their signal insertion behaviour, but in favour of C-terminal insertion [36]. Our extended hCD4 SP did not shift to a preferentially C-terminal translocation with SP-cleavage (Figure S9D), suggesting that the extended hCD4 SP preserved the same insertion behaviour as WT SP.
Our data indicated equal targeting efficiencies for control and CADA-treated nascent chains, making interference of CADA with SRP-binding less likely. Furthermore, the equal mRNA levels that were detected in CADA-treated and control cells indicate that the compound does not induce an mRNA degradation response as has been observed with defective signal sequence recognition by SRP [42]. Also, early insertion of the SP into the translocon channel was not altered by our drug, ruling out inhibition of channel gating by CADA. The post-targeting translocation inhibition by CADA is probably related to the obligate inversion of hCD4 SP inside the translocon. One can expect that the reorientation of the hCD4 SP from an Nlum/Ccyt into an Ncyt/Clum topology requires a high degree of flexibility of the SP and the translocation channel to undertake such a gymnastics, which may be compromised after CADA-binding. Based on the CADA-sensitive chimaeras that contained the hydrophobic alpha-helix of hCD4 SP, we could attribute a crucial role to the core hydrophobic domain of the SP and hypothesize compound binding to this region. In fact, minor changes in this alpha-helix, such as removing the helix terminator Pro20, could already reduce CADA-dependency, but introducing this proline residue at the C-terminal part of the murine ortholog was not sufficient to render sensitivity to the drug (see Figure S6). However, not only the central hydrophobic α-helical H-region of hCD4 SP, but also its C-terminus together with the first residues of the mature protein was shown to be important for full CADA-sensitivity. Interaction of CADA with both regions might re-position the SP in the translocon and profoundly reduce flexibility or alter the balance between hydrophobic and electrostatic interactions of the SP with the translocon and the lipid environment [43]–[45]. Also, post-targeted folding of a zinc-finger modified pPL has recently been shown to inhibit co-translational translocation into ER microsomes, indicating that folding events within the ribosome-Sec61 translocon complex (RTC) can occur and induce a mechanical block within the RTC that diverts the nascent chain into the cytosol for degradation [46]. Accordingly, in the presence of CADA the non-translocated CD4 precursor forms are routed to the cytosol for proteasomal degradation, which also indicates that translocation and not translation of CD4 is inhibited by the compound.
In our experiments, C-terminal translocation of nascent chains was determined by SP-cleavage. Inhibition of the signal peptidase activity could also be interpreted as an inhibition of translocation, however, CADA did not interfere with the enzymatic activity of RM-extracted signal peptidase (Figure S9E). Furthermore, both inhibition of N-glycosylation and SP-cleavage were observed with the N-terminally extended 80-mers, making it very unlikely that CADA would inhibit both luminal enzymes that act at distinct parts of the SP. Therefore, we propose that in the presence of CADA the SP is held in a folded conformation in the translocon (presumably at the lateral gate), so that the polypeptide is prevented from moving through the channel.
The focus of this work was to unravel the mechanism through which CADA selectively regulates CD4 expression. We show that CADA selectively inhibits the biogenesis of CD4 in cells without affecting other membrane or secretory proteins, and that this is due to its ability to inhibit specifically the co-translational translocation of CD4 into the lumen of the ER. The activity of CADA maps to the cleavable N-terminal SP of hCD4. Moreover, through SPR analysis we were able to show direct binding of CADA to the SP of hCD4 and identify this SP as the main target of our drug. In view of the very high selectivity of CADA for hCD4, it is most plausible to hypothesize a unique SP as its target. However, we cannot completely exclude a secondary interaction of CADA with another target of the translocation machinery. This could be in line with the suggested bimolecular binding model for CADA based on data from unsymmetrical CADA analogs and the 3-D quantitative structure-activity relationship (3D-QSAR) study [47],[48]. If CADA interacts with Sec61 (or closely related factor) in a more unspecific and non-discriminating manner, the final decision for a substrate to become translocation defective would then depend on the degree of reduced SP-flexibility by CADA binding, suggesting that CADA-dependency would mostly rely on SP-recognition and binding.
Hydrophobic signal sequences, such as SPs, perform several functions in the biogenesis of secretory and membrane-bound proteins [26]. Although these hydrophobic sequences may represent interesting targets for drug design in order to regulate the expression level of proteins, CADA is the first small-molecule drug, to our knowledge, known to selectively bind to a SP and inhibit translocation of a specific protein in a SP-dependent way. There is precedence for the cyclic heptadepsipeptide HUN-7293 and its derivatives CAM741 and cotransin to down-modulate vascular cell adhesion molecule 1 (VCAM1) in a signal-sequence-discriminatory way [49],[50]. However, these compounds were reported to bind to the Sec61 translocon channel and to inhibit the expression of a subset of secretory and membrane proteins, including VCAM1 [34],[51]–[53]. In contrast, the expression levels of a range of cell surface receptors including VCAM1 (Figure 1F) and CD4 from a different (non-primate) species (Figure 1D), were not decreased by CADA, which is explained by the interaction of the drug with a SP of a specific cell surface receptor instead of a common Sec61 channel. Recently, the cyclic dodecadepsipeptide valinomycin has been described to selectively destabilize the SP of hamster prion protein [54]; however, this might be related to the down-regulation of the luminal chaperone protein BiP involved in protein translocation, explaining the general apoptosis-inducing effect of this drug [55]. The toxic, small-molecule eeyarestatin 1 (ESI) has been reported to inhibit co-translational translocation of a wide-range of receptors, but this compound prevents the transfer of the nascent polypeptide chain from the membrane-bound SRP delivery complex to the ER translocon, thus acting more upstream of CADA and the cyclodepsipeptides and less substrate-selective [56]. Mycolactone, the virulence factor of Mycobacterium ulcerans has recently been described to block protein translocation into the ER, and thus, to prevent secretion of innate cytokines and expression of important immune-related membrane receptors [57]. Also, apratoxin A, a natural product from a marine cyanobacterium prevents co-translational translocation of a wide-spectrum of cellular proteins, which may also explain its cytotoxic nature [58]. Cellular cytotoxicity is a major concern when using small synthetic drugs; however, cellular viability was not compromised by CADA and long-term (∼1 year) CADA-treatment of T-cells was achieved, with preservation of CD4 re-expression following workout.
In conclusion, our findings demonstrate that with cell-permeable small synthetic compounds selective SP-binding and subsequent translocation inhibition of the associated protein is feasible. This opens a new field, not only for CD4 related immunomodulation and antiviral intervention, but also for a wide range of cell surface proteins for which this principle of selective SP-targeted translocation inhibition should be applicable.
CADA.HCl and MFS105.HCl were synthesized as described previously [47]. The compounds were dissolved in DMSO and stored at room temperature in the dark.
The pcDNA3 expression vector (Invitrogen) encoding WT hCD4 was kindly provided by O. Schwartz (Institut Pasteur, Paris, France). The A2.01/hCD4.426 and A2.01/hCD4-CD8 cell clones were from C. Devaux (Montpellier, France) and have been described elsewhere [59]. The pReceiver-M16 vector encoding mouse CD4-YFP was purchased from Imagenes and the human-mouse chimera constructs [60] were kindly provided by A. Trkola (Zurich, Switzerland). Bovine pPL in the pGEM4 vector has been described previously (plasmid pGEMBP1 [61]). Human and mouse CD4 were cloned into the pGEM4 vector for in vitro translation experiments. Fusion constructs were generated by PCR and cloned into the pGEM4 vector. The human/mouse chimaeric constructs were generated by incorporating the mouse sequence into the coding human sequence using PCR overlap extension and site-directed mutagenesis (Stratagene). The same strategy was used for the murine/hCD4 and pPL/CD4 chimaeras. For the V5-tagged CD4 construct [hCD4]-V5sCD4, the simian virus 5 (V5) epitope (GKPIPNPLLGLD) was inserted in the pcDNA3.1-hCD4D1D2 plasmid, which contains the coding sequence of the two N-terminal domains of the hCD4 protein. The V5 sequence was incorporated at the N-terminus of mature CD4 between residue 32 and 33 as schematically presented in Figure S3E. Note that the first two residues of V5 (residue G and K) were already present in CD4 at positions 31 and 32.
The CD4+ T-cell lines MT-4 and SupT1 were cultured in RPMI-1640 medium, and Hela, U87, HEK293T, and CHO cells were cultured in DMEM, supplemented with 10% FCS and penicillin/streptomycin. Human PBMCs were isolated by density gradient centrifugation as described previously [18]. For the monkey PBMCs, blood was collected from cynomolgus macaques (Macaca fascicularis) as described elsewhere [62]. Murine PBMCs were isolated from blood collected from Balbc mice. Stable transfections of U87 cells were performed with FuGENE 6 Transfection Reagent (Roche Diagnostics), and transient transfections of HEK293T cells were performed with PolyJet in vitro transfection reagent (Tebu-Bio), in accordance with the manufacturer's instructions. Western blot analysis of cell lysates was performed in accordance with standard protocols. Mouse monoclonal antibodies anti-hCD4 (clone SK3, BD Biosciences), anti-clathrin heavy chain (clone 23, BD Biosciences), anti-THE SV5-tag mAb (Genscript), anti-actin (Abcam ab3280), and a HRP-conjugated goat anti-mouse antibody (Dako) were used for detection.
For the cytosolic CD4 precursor rescue experiments, HEK293T cells were transfected with a mixture of 500 ng pcDNA3.1-V5-hCD4D1D2 plasmid DNA and 2 µl lipofectamine 2,000 (Invitrogen). Six hours after transfection, growth medium was removed and replaced with fresh medium containing 10 or 2 µM CADA and/or 200 nM MG132 (Sigma). After a 22 hour incubation in the presence of the compounds, cells were washed with ice-cold PBS and lysed in ice-cold lysis buffer (25 mM Tris, 150 mM NaCl, 1 mM EDTA, 1% NP-40, 5% glycerol, pH 7.4), supplemented with 0.4 µM PMSF (Fluka) and Complete protease inhibitor cocktail (Roche).
Flow cytometric analysis of surface receptor expression was performed as described previously [18]. All FITC-, PE-, PerCp-, or APC-labelled mAbs were from BD Biosciences. Data were acquired with a FACSCalibur flow cytometer (BD Biosciences) and the CellQuest software (BD Biosciences). Data were analyzed with the FLOWJO software (Tree Star). Down-modulation of CD4 was evaluated by the decrease in fluorescence intensity on CADA-treated cells relative to matched, untreated cells stained for CD4. To calculate the efficiency of CADA on CD4 down-modulation, the median fluorescence intensity (MFI) for CD4 labelling for each sample was expressed as a percentage of the MFI of control cells (after subtracting the MFI of the unstained control cells).
Pulse-labelling of CD4 was performed on CD4+.CHO cells preincubated for 45 min at 37°C with CADA (16 µM) in medium lacking methionine, cysteine, and serum. After being pulsed with 0.75 mCi ml−1 [35S]methionine/cysteine for 30 min, cells were chased by replacing the radiolabel with pre-warmed medium containing 1 mM cysteine and 1 mM methionine. At the end of the chase this medium was replaced by ice-cold DMEM. Cells were kept on ice and lysed in NP-40 buffer (2% NP-40, 20 mM Tris [pH 7.8], 150 mM NaCl, 2 mM MgCl2), and cell lysates (separated from nuclei and debris) were immunoprecipitated for CD4 as described [63]. Proteins were analyzed by SDS-PAGE and autoradiography.
Cells were pretreated with CADA (5 µM) or DMSO for 1 h before starvation in Met/Cys free medium with CADA, DMSO, or 50 µg/ml CHX (Sigma). Cells were pulsed with 0.022 mCi ml−1 [35S]methionine/cysteine EasyTag Protein Labeling mix (Perkin Elmer) for 30 min, washed twice with PBS, and incubated in fresh medium without serum for 90 min. After collection of the supernatant, cells were washed in ice-cold PBS and then either lysed in ice-cold lysis buffer (25 mM Tris, 150 mM NaCl, 1 mM EDTA, 1% NP-40, 5% glycerol, pH 7.4) supplemented with 0.4 µM PMSF and protease inhibitor cocktail, or separated into cytosolic and membrane fractions. To collect the cytosolic proteins, cells were permeabilized with digitonin buffer (20 mM Tris, 150 mM NaCl, 2 mM MgCl2, 0.03% digitonin, pH 7.8, supplemented with 0.4 µM PMSF and protease inhibitor cocktail). Permeabilized cells were then washed in digitonin buffer and further lysed in ice-cold lysis buffer (25 mM Tris, 150 mM NaCl, 1 mM EDTA, 1% NP-40, 5% glycerol, pH 7.4, supplemented with 0.4 µM PMSF and protease inhibitor cocktail). To isolate glycosylated proteins, total cell lysates or digitonin-resistant membrane fractions were further incubated with Concanavalin A (Vector Laboratories) agarose beads overnight at 4°C by gentle rotation.
Full length cDNAs and truncated CD4-pPL nascent chains generated by PCR were transcribed in vitro using T7 RNA polymerase, and translated in rabbit reticulocyte lysate (Promega) in the presence of [35S]methionine (Perkin Elmer). Translations were performed at 25°C for 45 min (full length) or for 25 min (truncated) in the presence or absence of mammalian pancreatic microsomes (RMs). Digestion with Proteinase K (Roche) was performed on ice for 30 min and stopped by the addition of phenylmethylsulfonyl fluoride (PMSF; Thermo Fisher Scientific). Release of the nascent chains (NCs) from the targeted ribosome was induced by treatment with 2 mM puromycin for 10 min at 25°C. NCs were isolated by sedimentation at 4°C and pellets were dissolved in SDS sample buffer for analysis by SDS-PAGE. For the quantitative analysis of translocation experiments we used a Cyclone Plus phosphorimager (Perkin Elmer) with accompanying software.
For the time-of-addition experiments, [hCD4]-(7)-pPL nascent chains of 80 residues were synthesized in rabbit reticulocyte lysate by translating the mRNA for 20 min at 25°C. Translation was done in the absence of RM, and the translation mixture was left untreated (C, control) or was treated with CADA (15 µM) from the beginning of the translation reaction (R, ribosomes). In the mean-time, some RMs were pretreated with CADA (15 µM). After 20 min of translation, the untreated control sample was split in three aliquots. Next, one aliquot of control sample and the CADA-sample were given untreated RMs. The CADA-pretreated RMs were administered to the second control sample (M, microsomes). The third control sample received untreated RMs. All samples were then incubated for 10 min on ice and 5 min at 25°C. Finally, CADA was administered to the third control sample (P, post-targeting), and all samples were further incubated for 5 min at 25°C, before treatment with puromycin. NCs were isolated by sedimentation at 4°C and pellets were dissolved in SDS sample buffer for analysis by SDS-PAGE.
Chemically synthesized SPs (PEPperPRINT) were captured on a streptavidin-coupled C1 sensor chip (GE Healthcare). The amino acid sequence of the SP was: MNRGVPFRHLLLVLQLALLPAATQGKKVVLGKK-PEG11-K-biotin for hCD4, MCRAISLRRLLLLLLQLSQLLAVTQGKTLVLGKE-PEG11-K-biotin for mCD4, and MDSKGSSQKGSRLLLLLVVSNLLLCQGVVSTPVCPNGP-PEG11-K-biotin for pPL. The chip density was between 84 and 215 resonance units (RUs). A reference flow cell was used as a control for non-specific binding and refractive index changes. All interaction studies were performed at 25°C on a Biacore T200 instrument (GE Healthcare). The compound CADA was diluted in HBS-P (10 mM HEPES, 150 mM NaCl, and 0.05% surfactant P20 [pH 7.4]) supplemented with 5% DMSO. Signal recognition particle (tRNA Probes) was included as a positive control. Samples were injected for 2 min at a flow rate of 30 µl/min and the dissociation was followed for 2 min.
All data are presented as means with standard deviations (SD), unless otherwise stated. Two-tailed Student's t test was used to determine statistical significance as calculated with GraphPad software. For the flow cytometric data, statistical analysis was done on the background-corrected MFI values between compound treated and control (without compound) samples.
|
10.1371/journal.pcbi.1005089 | Noncommutative Biology: Sequential Regulation of Complex Networks | Single-cell variability in gene expression is important for generating distinct cell types, but it is unclear how cells use the same set of regulatory molecules to specifically control similarly regulated genes. While combinatorial binding of transcription factors at promoters has been proposed as a solution for cell-type specific gene expression, we found that such models resulted in substantial information bottlenecks. We sought to understand the consequences of adopting sequential logic wherein the time-ordering of factors informs the final outcome. We showed that with noncommutative control, it is possible to independently control targets that would otherwise be activated simultaneously using combinatorial logic. Consequently, sequential logic overcomes the information bottleneck inherent in complex networks. We derived scaling laws for two noncommutative models of regulation, motivated by phosphorylation/neural networks and chromosome folding, respectively, and showed that they scale super-exponentially in the number of regulators. We also showed that specificity in control is robust to the loss of a regulator. Lastly, we connected these theoretical results to real biological networks that demonstrate specificity in the context of promiscuity. These results show that achieving a desired outcome often necessitates roundabout steps.
| DNA is the blueprint of life. Yet the order in which a cell follows these instructions makes it capable of generating thousands of different fates. How this information is extracted from underlying gene regulatory networks is unclear, especially given that biological networks are highly interconnected, and that the number of signaling pathways is relatively small (approximately 5–10). The conventional approach for increasing the information capacity of a limited set of regulators is to use them in combination. Surprisingly, combinatorial logic does not increase the diversity of target configurations or cell fates, but instead causes information bottlenecks. A different approach, called sequential logic, uses noncommutative sequences of a small set of regulators to drive networks to a large number of novel configurations. If certain targets are first protected, then even promiscuous regulators can activate specific subsets of lineage-specific targets. In this paper we show how sequential logic outperforms combinatorial logic, and argue that noncommutative sequences underlie a number of cases of biological regulation, e.g. how a small number of signaling pathways generates a large diversity of cell types in development. In addition to explaining biological networks, sequential logic may be a general experimental design strategy in synthetic and single-cell biology.
| A fundamental question in systems biology is how a small number of signaling inputs specifies a large number of cell fates through the coordinated expression of thousands of genes. This problem is especially challenging given that gene regulatory and other types of networks in biology tend to be highly interconnected and their regulators promiscuous, with regulators affecting multiple targets and targets being affected by multiple regulators. Examples of this architecture include: transcription factor binding networks in bacteria [1], yeast [2, 3], plants [4], and animals [5, 6]; cellular signalling pathways involved in growth and differentiation [7–9]; the interactome of protein kinases and phosphatases [10, 11]; and synaptic connections between different layers of the brain [12]. Furthermore, because the targets and regulators are often well-mixed and mutually accessible in the cell, most actions are likely to have nonspecific and undesired effects.
At the same time, regulatory molecules drive networks to a large number of highly specific outcomes or cell fates. Although there are approximately four hundred canonical cell types in the adult human [13], recent single-cell RNA expression profiling experiments in the developing embryo [14, 15], brain [16], hematopoietic system [17, 18], and other organs [19, 20], have indicated that there may be thousands more.
Given there are only a few signaling pathways used in metazoan development [21, 22], understanding how cells reach their final outcomes when there are fewer regulators than fates and/or targets is an unsolved problem. One extensively studied solution for the control of promiscuous gene networks is combinatorial binding of DNA-binding transcription factors (TFs) at the promoter [23–31]. At the level of individual promoters, combinatorial binding ensures that individual genes are ON only when specific combinations of TFs are present (Fig 1A). However, on the genome level, combinatorial regulation restricts which sets of genes may be ON at the same time. For example, using AND logic, gene H in Fig 1A is only ON in the case that the three TFs K1, K2, and K3 are concurrent at the H promoter; but these stringent requirements mean that H can never be transcribed independently of the less highly-regulated genes A-G. (A similar conclusion holds for OR logic.)
In fact, using combinatorial control, there is a one-to-one correspondence between configurations of the targets and configurations of the regulators. As shown in Fig 1A, the ON/OFF states of 3 TFs uniquely define the binding combinations at 23 = 8 promoters. A similar conclusion holds when the regulators are expressed in a graded fashion.
This one-to-one correspondence is the fundamental limitation of combinatorial regulation: it requires an equal number of regulators and independently controlled targets and/or cell fates. Applied to embryonic development, combinatorial control requires that hundreds or thousands of cell-type specific TF combinations be generated in a spatially precise manner at the start. However, the combinatorial scheme does not explain how the TF states are regulated in the first place, and thus it offers no new insight into how cell fate is specified.
The limitations of combinatorial logic can also be understood from an information theoretic point of view. In particular, it is impossible to specify arbitrary cell fates if the regulatory layer bottlenecks the capacity of the targets to receive messages from extracellular signals. It is known that some ten to twenty types of signals [21, 22] converge onto membrane-bound regulators in many different combinations, permitting messages to be passed to the downstream targets. Much of this information stands to be lost, however, if the network relies on combinatorial logic alone: the regulatory layer simply cannot transmit messages in their entirety if there are more signals than regulators. Thus, combinatorial logic strongly circumscribes what fates are ultimately reachable. Cell fate information is lost not only if the signals are more numerous than the regulators, but also if the connections between signals and regulators are promiscuous (Fig 1B). When different signals activate the same regulators (Fig 1Bi), certain signaling inputs become redundant. On the other hand, when same signal activates different regulators (Fig 1Bii), some of the regulators become redundant. One may determine by direct enumeration exactly how redundancy decreases the number of configurations available to the targets (Materials and Methods Sections 1 and 2). These preliminary conclusions are at odds with the observation that signaling molecules are deployed over time in a complex code [32]. How do these messages in the signal space reach the targets if the regulatory layer imposes a bottleneck on information flow?
In addition, feedback regulation—a common feature of regulatory networks—exacerbates information bottlenecks when coupled with combinatorial logic. Stated another way, feedback merely widens the basin of attraction of certain promoter configurations at the expense of the number of distinct configurations. In Fig 1C, constitutive expression of K1 by C means that C is never ON independently of the targets regulated by K1. Thus, the number of accessible configurations decreases from 8 to 6 without allowing new target configurations to be explored.
We need an alternative to combinatorial logic in cell fate specification that overcomes information bottlenecks. Here, we considered time-ordered control schemes, which we refer to as sequential logic. In this scheme, regulators can be applied in a stepwise manner; the entire sequence matters, so the final configurations can differ if the same regulators are permuted in time. In order for different temporal sequences to carry distinct information, the actions of the regulators must be noncommutative. This is the case, for example, when a regulator protects its targets from the action of another regulator, as when loci recruited to repressive chromatin compartments are protected from further modification [33, 34].
While it is not surprising that noncommutative sequences like this result in different outcomes at the single promoter level, these simple mechanisms may have nontrivial implications for regulation at the genome level. In particular, noncommutativity permits the same regulators to be used at different times with distinct effects. This is seen in development when ubiquitous signaling molecules like FGF family members exert different effects depending on the time and context of their expression [35–38]. Reuse of factors could greatly expand the information capacity of the major signaling pathways.
A number of examples show that noncommutativity may be a general strategy in other areas of biology. In hematopoietic stem cells, activation of GATA2 and C/EBPα in different orders results in different cell fates [39]. In neurobiology, different temporal orderings of the same inputs lead to distinct firing patterns [40–42]. In the field of synthetic biology, a DNA switch was developed that could detect the order in which invertase enzymes were applied [43]. And in evolutionary biology, the order in which mutations arise was recently implicated in determining a genotype’s fitness [44–47]. There is also accumulating evidence for sequential logic in transcriptional control: signaling molecules and TFs in mammalian cells, including ERK [48], NF-κB [49, 50], p53 [51], as well as in yeast [52–54] have been observed to pulse, suggesting that TF timing may be used to control the transcriptional state of the cell.
By applying sequential logic, we show that, even in complex and promiscuously regulated networks, specific target configurations can be reached using a temporal sequence of regulators. In particular, we consider two models inspired by (i) kinase/neural networks and (ii) chromosome folding and show analytically that both scale super-exponentially. We further show that noncommutative networks are robust to the loss of regulators, suggesting a mechanism for regulator evolution. We also show that regulators induce different orbits in expression space, which is related to the number of networks that can be controlled in parallel. We conclude by discussing how these models apply to interconnected networks in and outside biology and by providing possible experimental tests of the theoretical concepts. Theorems and proofs are given in the Materials and Methods.
To consider how time-ordered sequences of regulators can specifically control groups of targets, we begin by analyzing a generic two-layer network that is an extension of combinatorial logic (Fig 2, Materials and Methods Sections 1 and 2). In this model, each regulator controls multiple targets, and each target is accessible to any of its regulators. The model is meant to be analogous to the cellular environment wherein regulators and targets are well-mixed. For example, targets could be substrate proteins capable of multi-site phosphorylation [55, 56], and regulators the kinases and phosphatases. Targets could also be neurons and regulators their upstream excitatory and inhibitory inputs [12]. We denote by K the set of activators (i.e. kinases) and P the set of deactivators (i.e. phosphatases). Each target has a ladder of (integer-valued) states, and together the states of the targets are a configuration of the network. (This distinction is in contrast to the common usage of “state” as a gene expression vector.) An additional parameter, the threshold T, determines the number of rungs on the ladder. Regulators ratchet the targets through their states, and only targets that have reached threshold will be ON at the end of a sequence of regulators. If each target in the group can be controlled by a unique combination of K’s and P’s, what ON/OFF configurations are possible?
In this model, termed the ratchet network (Fig 2A), each of n K’s and m P’s control N = ( n l n ) ( m l m ) unique targets, with the connectivity parameters ln and lm specifying the number of regulators to which each target connects. Consider the sequence K1 K2 P1 acting on the targets A, B, C, and D (Fig 2B). In the final configuration, B and D are ON together even though no single K connects to both, and A and C are OFF, even though both share and activator with B and D. Therefore, this simple model illustrates the important point that similarly regulated targets can be in independently controlled using sequential logic.
With threshold T = 1, not all configurations are reachable. Observe that there is no way to specifically activate A and D while leaving B and C OFF. This result is surprising given that A and D share no regulators: specificity depends on the network as a whole, not just individual targets. By going to T = 2, the forbidden configuration becomes accessible (Fig 2C), along with all ON/OFF states (below).
The model described above can be formalized as a combinatorial object that we refer to as the connectivity matrix A. This formulation is useful because it is amenable to studying scaling, and it permits a direct comparison between noncommutative ratchet networks and standard combinatorial logic. For the interested reader, the models considered in this paper have a universal formulation as noncommutative matrix operators on the vector space of configurations (Materials and Methods Section 9).
Typically, the state of N targets is represented as an N-dimensional vector. If each target is controlled by a unique (Ki, Pj) pair (i.e. ln = lm = 1), the N = nm-dimensional vector can be re-formulated as an n × m matrix
A = P 1 ⋯ P m K 1 ⋮ K n ( A 1 , 1 ⋯ A 1 , m ⋮ ⋱ ⋮ A n , 1 ⋯ A n , m ) (1)
where each entry Ai,j ∈ {0, 1, …, T} is the state of the target regulated by Ki and Pj. For example, the connectivity matrix for the network in Fig 2 is
A = P 1 P 2 K 1 K 2 ( A B C D ) . (2)
In general, a regulator may connect to multiple targets (i.e. ln, lm > 1, see below), so that each entry of A may be thought of as an M-dimensional vector (M determined in Materials and Methods Section 1). It turns out that this is an unnecessary complication; we instead let each Ai,j = 1 if at least one of the M targets regulated by Ki and Pj is ON, and Ai, j = 0 only if all M targets are OFF.
In this formulation Ki and Pj are raising and lowering operators that map n × m matrices to n × m matrices via the rules
K i A i , j = A i , j + 1 if A i , j < T A i , j if A i , j = T P i A i , j = A i , j - 1 if A i , j > 0 0 if A i , j = 0 . (3)
From Eq (3), any sequence Ki1 Ki2⋯Kik of all K’s is commutative, because any target controlled by t ≤ k of the K’s will be in state t ≤ T at the end of the sequence, regardless of the order. A similar argument holds for the P’s. However, sequences consisting of both K’s and P’s are in general noncommutative. This is due to edge effects when Ai,j = 0 or T. If Ai,j = T, for example, then Ki Pj results in Ai,j = T − 1, whereas Pj Ki gives Ai,j = T. Therefore, A gives insight into both the configuration of the targets and the noncommutativity of the regulators.
The problem of determining the number of accessible configurations in a network is reduced to finding the number of matrices satisfying certain patterns. For example, combinatorial logic with T = 1 corresponds to the special case in which the only sequences are the 2n combinations of the n K’s. In an n × 1 connectivity matrix, activating Ki corresponds to turning all 0’s in row i into 1’s. There are 2n matrices generated by this procedure. More complicated cases of combinatorial logic can be studied this way (Materials and Methods Section 2), but it turns out that the total number of network configurations is always less than 2n + m, with n + m the total number of regulators. This is important because noncommutative models can bypass the exponential limit.
We used the connectivity matrix representation of the ratchet network to determine the scaling as function of the number of regulators n and m, with each target connected to a unique (K, P) pair (i.e. ln = lm = 1) and the threshold T = 1. Ki turns 0’s to 1’s in row i and Pj turns 1’s to 0’s in column j. The rules are consistent with the one-pot reaction model in which all substrates receptive to Ki are promoted when Ki is active. For example, the sequence K1 K2 P1 in Fig 2B can be recast as
0 0 0 0 → K 1 1 1 0 0 → K 2 1 1 1 1 → P 1 0 1 0 1 . (4)
The main result is that A must avoid the patterns ( 1 0 0 1 ) and ( 0 1 1 0 ) in any 2 × 2 sub-block (Materials and Methods Section 3). Brewbaker [57] enumerated the n × m binary matrices avoiding these patterns and showed that they scale as the poly-Bernoulli numbers [58]
B m − n = B n − m = ∑ j = 0 m ( − 1 ) ( n + j ) j ! ( j + 1 ) n { n j } = ∑ j = 0 min ( n , m ) ( j ! ) 2 { m + 1 j + 1 } { n + 1 j + 1 } , (5)
where { nj} is a Stirling number of the second kind, defined combinatorially as the number of ways to put j labelled balls into n unlabelled boxes such that no box is empty [59]. These numbers scale not quite as fast as 2N = 2nm, but much faster than 2n + m, the maximum number of states in the equivalent combinatorial network (Fig 2D). Thus, a simple time-sequence model is able to generate super-exponential scaling.
Are more configurations accessible if multiple activation events are needed before reaching threshold? For example, neurons require the summation of multiple excitatory inputs to reach action potential, and proteins need to be phosphorylated at multiple sites before they are activated [55, 56]. We found that by increasing the threshold to T = 2, all 2N ON/OFF configurations of the N targets become reachable. In the connectivity matrix formulation, ( 1 0 0 1 ) and ( 0 1 1 0 ) are no longer forbidden, which we show with an inductive proof (Materials and Methods Section 4). This scaling law (Fig 2D), achieves the maximum of reachability and specificity; it far exceeds the scaling 2n + m of the combinatorial model.
Being able to reach the entire ON/OFF space of N targets is overkill for most biological networks, which only display a relatively small number of stable configurations. The major implication of this result is that multiple levels of activity permit more targets to be controlled independently.
As sequential logic allows a large number of configurations to be reached in a complex network, we asked whether increasing the connectivity of the network (ln and lm) can maintain the specificity of the network while making it robust to the loss of a regulator. This is potentially relevant to evolution of biological networks, because redundant connections allow the network to repurpose regulators for new functions without severely impairing existing ones [60].
In the ratchet model, an increase in the connectivity parameters to ln = 2 K’s and lm = 2 P’s permits multiple targets to share a common (K, P) pair (Fig 3A). The connectivity matrix incorporating the extra links in the network in Fig 3A is
P1 P2 P3A=K1K2K3(ABDEACDFCDEFABGHACGIBCHIDEFGDFGIEFHI). (6)
Now that each entry of A is a group of M > 1 targets, it makes sense to track the state of the group as a whole with a single number Ai,j. Even though a target appears in multiple entries of A, the rules prevent a regulator from altering the state of groups at remote locations (e.g. K1 cannot change the state of the group at A2, 2).
We prove in the Materials and Methods that all sequences using at least n − ln + 1 K’s and m − lm P’s are redundant with shorter sequences (Fig 3B and 3C, Materials and Methods Section 5). For example, the sequences K1 K2 K3 is required to turn ON all targets in the case ln = lm = 1, but if ln = lm = 2, the shorter sequences K1 K2, K1 K3, and K2 K3 have the same effect. We derived a recursive formula that eliminates the redundant sequences in each (n, m, ln, lm) instance to derive the number of sequences in (n, m, ln + 1, lm) and (n, m, ln, lm + 1) (Fig 3D and S2 Fig). The formula agreed exactly with an algorithm designed to find all minimal length sequences (Materials and Methods 5). Notably, increasing ln, lm reduced the number of configurations. We observed a similar effect in combinatorial logic (S1 Fig).
To investigate the robustness of sequential logic networks, we studied the effect of deleting regulators in increasingly connected networks on the number of reachable configurations (Fig 4A). We hypothesized that sequences that activate similar subsets of targets should be able to recoup permanently lost configurations. To test this, we computed the normalized correlation coefficient between configurations in the network using all K’s (the full network) and configurations in the network without K1 (the impaired network), subject to the constraint that those configurations were reached using longer sequences (Fig 4B). To focus on the recoverable fraction, we deleted all configurations that had an exact match. Highly similar configurations (yellow) clustered to the right of the plot, indicating that longer sequences can be used to recover lost configurations.
How similar are the recouped configurations? As connectivity increased, the maximum similarity became increasingly concentrated above 0.8 (Fig 4C). There is generally a tradeoff between reachability and the size of the fraction above 0.8 (Fig 4D). The tradeoff is nonlinear, however: using ln = 2 gave the greatest increase recoverability for the smallest loss of configurations, showing that an intermediate level of redundancy can buffer the network to loss of regulators. The above analyses demonstrate that specificity of control is not compromised when regulators are lost or repurposed in heavily interconnected networks.
In the ratchet model, all targets are accessible to their regulators at all times. However, in some cases targets may be shielded from regulators: for example, genes can be silenced by sequestration in various nuclear compartments [61, 62]. This was seen in a landmark study by Filion et al [63], who used a DNAse accessibility assay to show that genes associate with different regulators depending on their chromatin “color” or accessibility status.
To study the effect of accessibility and silencing on activating specific subsets of genes, we constructed the following sequestration model. In addition to the OFF state 0 and the ON state 1, each target/gene is endowed with additional orthogonal states 2 to n (allowing for a total of 2n − 1 − 1 genes). If RNA polymerase (RNAP) is associated with K1, what genes can be independently activated? In this model (Fig 5) a regulator Ki promotes targets in the 0 state to state i, and Pi returns targets in state i to 0. Any target in state i is protected from regulators other than Pi. As an example of gene regulation on a three-dimensional chromosome (Fig 5A), the sequence K3 K4 K1 P3 P4 first clusters all genes having a 3 in a repressive compartment, and then the remaining genes having a 4 in another repressive compartment. The net effect is that RNAP can only act on the gene represented by {1, 2}. We represent this abstractly as a configuration vector of k-armed targets (Fig 5B), where each entry corresponds to the state {0, 1, …, n} of a gene able to access k ≤ n of the states (see below for a mathematical description of the model). Therefore, protected states in the sequestration model allow genes to be transcribed specifically in a well-mixed environment.
We derived (see below) that the number of reachable configurations scales with the number of regulator pairs n as
f n = 2 2 n - 1 - 1 - ∑ m = 2 n - 1 n - 1 m 2 ∑ k = 3 m m k - 1 - 1 2 ∑ k = 3 m n - 1 k - 1 - m k - 1 . (7)
For n = 1, 2, 3, 4, 5, 6, this formula gives f(n) = 1, 2, 7, 89, 16897, 780304385 (Fig 6). We also relaxed the constraint that all genes have a 1 state (allowing for a total of 2n − 1 genes) and found that the number of configurations scales as cn = 2, 7, 94, 37701 with n = 1, 2, 3, 4. The full model does not have an analytical solution, but it does have upper and lower bounds related to Eq (7) (Materials and Methods Section 7, S3 Fig).
Combinatorial scaling laws of this sort are not uncommon [44, 64, 65]. Edwards and Glass [64] saw an explosion in the number of states when studying trajectories on n-cubes, and Green and Rees [65] saw a super-exponential jump when enumerating certain types of nonrepeating sequences on n letters. Furthermore, a similar small number (four) of factors are necessary and sufficient to reprogram fibroblasts to stem cells [66]. Together, these results indicate that sequences can far exceed the 2n limit set by combinatorial regulation, and that only a few regulators are necessary to make large changes in the configuration of a cell.
The sequestration network with n regulator pairs (referred to as the n-network) is described using the 1 × 2n − 1 configuration vector x. This is a simpler description than the connectivity matrix because a target affected by Ki is necessarily affected by Pi. The entries of x are the states of each target g able to be controlled by k ≤ n of the regulator pairs. Each target g is is a list {0, i1, …, ik} of the k regulators to which it responds. Because of their radial appearance, such targets are said to have k arms (see Fig 5B).
The regulators act on x according to the rules
K i x g = x g + i if i ∈ g and x g = 0 x g else P i x g = 0 if i ∈ g and x g = i x g else. (8) Eq (8) guarantees that the regulators are orthogonal in the sense that a target in state j is protected from Ki and Pi if i ≠ j; and also idempotent in that K i 2 = K i. Furthermore, sequences of regulators are noncommutative unless the only actions are P’s. This is a consequence of the fact that P’s put all affected targets into the 0 state. Although these rules are different from the ratchet model, a formulation exists that generalizes the K’s and P’s to matrix operators consistent with both models (Materials and Methods Section 9).
If x is restricted to the 2n − 1 − 1 targets all able to be regulated by K1 and at least one other K, the network is said to be reduced; otherwise we say x is full. This distinction was used in Fig 5.
A one-coloring is a configuration of x that uses only one of the states and 0. For example, the configuration x = (1, 0, 0, 1, 1, 0, 0) in the full n = 3-network is a one-coloring of 1; so is the reduced network formed by (x4, x5, x7) = (1, 1, 0). This concept is easily extended to k > 1-colorings. One-colorings are particularly important because they resemble the ON/OFF configurations of genes in an RNA-seq experiment, and we would like to know how many such configurations can be reached.
As in the ratchet model, finding the accessible states of the sequestration network amounts finding restricted patterns in x. We determined that the restricted one-colorings are those that violate a property referred to as connectivity (Materials and Methods Section 7). A configuration of x is said to be connected if all k > 3-arm targets g i ( k ) = { 0 , i 1 , … , i k } match the state of at least one of k of the 2-arm targets {0, i1, i2}, …, {0, ik − 1, ik} sharing the indices i. If the network is reduced, no k-arm target may be in the 1 state when all of 2-arm targets with which it overlaps (i.e. shares an index other than 1) are in the 0 state. This restricts the one-colorings and suggests a method to determine the scaling law for the model in Fig 5.
As an example, in the n = 4 network on the reduced set of 23 − 1 targets illustrated in Fig 5, {0, 1, 3} and {0, 1, 4} both being 0 constrains {0, 1, 3, 4} to be 0 as well. Furthermore, even though {0, 1, 2} is in the 1 state, {0, 1, 2, 4} and {0, 1, 2, 3, 4} may be 0. It is only the two-arm targets that constrain the possible configurations: for example, the longer sequence K2 K4 P2 K3 K2 P4 K1 P3 K4 P1 K3 P4 K1 P2 P3 obtains the state x = (0, 0, 1, 0, 0, 0, 1) in which only the targets {1, 4} and {0, 1, 2, 3, 4} are ON, showing that {0, 1, 2, 3, 4} need not be in the same state as {0, 1, 2, 3}, {0, 1, 2, 4}, or {0, 1, 3, 4}. In Fig 6A and 6B we show the allowed states and the sequences that generate them for n = 4; there are 90 out of a possible 224 − 1 − 1 = 128 configurations.
There are 22n − 1 − 1 one-colorings on 2n − 1 − 1 targets. How many of these violate the connectivity rule? Suppose there are m 0’s among the 2-arm targets. If m = 1, then ( m k - 1 ) = ( 1 k - 1 ) = 0 of the k ≥ 3-arm targets are constrained to be 0, as there is always another 2-arm target (in the 1 state) that each k-arm target can match. If m > 1 and m − 1 < k, however, then ( m k - 1 ) > 0, so ( m k - 1 ) k-arm targets whose states {i1, …, ik − 1} are completely contained within the set of 2-arm targets {0, 1, j1}, …, {0, 1, jm} must be 0. Hence in any violation of the connectivity rules at least one of ∑ k = 3 m ( m k - 1 ) k-arm targets will be in the 1 state and the remaining ∑ k = 3 m ( n - 1 k - 1 ) - ( m k - 1 ) k-arm targets will be 0 or 1. Furthermore, there are ( n - 1 m ) ways of specifying m 0’s, so the total number of violations is
∑ m = 2 n - 1 n - 1 m 2 ∑ k = 3 m m k - 1 - 1 2 ∑ k = 3 m n - 1 k - 1 - m k - 1 . (9)
Subtraction from 2n − 1 − 1 gives Eq (7).
Until now we have considered the reachable space of a single group of targets each starting in 0. An ensemble of networks could each start with their targets in some arbitrary state, and when a sequence is applied to the ensemble the different networks will in general span different configurations. Determining the number of orbits (defined precisely in Materials and Methods Section 8) within the set of possible configurations tells us how many networks can be controlled in parallel.
Enumerating the reachable space for both the ratchet and sequestration networks involved finding configurations that violated at least one rule. If two configurations have distinct violations, then there is no way they can communicate using the regulators. Therefore, the different orbits are the groups of configurations having the same forbidden patterns. It is possible that a violation could be alleviated by an action that changes the state of an offending target, so we require that each orbit be immune to a subset of the regulators. This could be achieved in biological networks by locking targets in protective chromatin states or by shutting down certain cellular receptors.
We determined a recursive formula for the number of orbits in the ratchet network for an arbitrary n, m (Materials and Methods Section 8). In Fig 7A we plot the orbits for the n = 4, m = 2 case. There is one large component of size B n - m and several smaller orbits of size B i - j with i ≤ n and j ≤ m. There are only a handful of singleton orbits in Fig 7A, but the number of isolated states dominates the space as n, m increase.
We were unable to find a similar solution for the sequestration network because we lack a general solution for the number of states in the main orbit. However, Fig 7B shows the computationally discovered orbits for the full network on 2n − 1 targets. A nontrivial feature is that there are orbits which use all pairs of regulators, but which do not communicate with the main orbit. For example, the sequence K2 K3 from x = (1, 0, 0, 0, 0, 0, 1) reaches the same configuration as the sequence K1 starting from x = (0, 2, 3, 2, 2, 3, 0); these configurations are part of the same orbit because both violate the connectivity rule between x7 = {0, 1, 2, 3} and the 2-arm targets x4, x5, and x6.
Another observation is that some pathways cannot be reversed by a legal action in the ratchet network orbits (indicated by a directed arrow in Fig 7), whereas there always exists a reversible path between configurations in the sequestration network orbits (no arrowheads). It can be proved that this is true in general for the sequestration network (Materials and Methods Section 8). This feature permits orbits to be found computationally by looking for reversible one-step paths in the entire configuration space.
The orbits are one explanation for the phenomenon the same signal can cause cells to behave differently [38]. More generally, the orbits demonstrate an intriguing symmetry between the targets responding to a restricted subset of the regulators on one hand, and the orbits restricted to the same subset on the other.
In this paper we first show how noncommutative, sequential logic can relieve information bottlenecks in multilayer networks. Bottlenecks in combinatorial logic may occur whenever a downstream layer has fewer elements than the layer upstream, which poses the problem of how networks process complex signals without loss of information. Noncommutative solutions such as the ratchet and sequestration models, in which the number of configurations scales super-exponentially in the number of regulators (Eqs (5) and (7)), permit longer, more complex messages to reach the targets via information “pulses.” These pulses encode a large diversity of signals into configurations of the targets that would otherwise be lost using combinatorial logic.
Noncommutativity has long been recognized as a central concept in control theory, because it allows systems with few controllers to explore a broader configuration space. For example, one generates z rotations in 3D by R − x Ry Rx, so control over z is generated by a pulse sequence of rotations in x and y, as in airplane control where roll and pitch generate yaw [67]. Infinitesimal motions in the form of generating matrices are translated into flows in a vector space by exponentiation. Because matrix multiplication is noncommutative, composition of flows is not simply the addition of generators, but rather a higher order polynomial of commutators of the generators given by the Baker-Campbell-Hausdorff formula [68]. Noncommutativity also appears in experimental physical chemistry where pulse sequences can prepare spin systems in nontrivial population configurations [69]. A formal description of these phenomena is based on the Heisenberg picture of quantum mechanics, wherein evolution of a system of many variables is given by a differential equation involving the commutator of a Hamiltonian operator.
The significance of noncommutative control for systems biology is that it becomes possible to independently control targets that would otherwise be activated by the same promiscuous regulator. In this paper, we argue that noncommutative sequences permit control over new directions in gene expression space, allowing more specific sets of targets to be controlled. Several studies have shown that TFs that can bind genes in one tissue type are in fact precluded from binding the same genes in another [70, 71]. The C. elegans TF LIN35 fails to bind targets in the germline that it binds in the intestine [71], and the SMARCA4 complex in mouse binds enhancer elements in heart, limb, and brain tissue in a tissue-specific manner [70]. One hypothetical explanation for these observations, based on the sequestration model, is that cell-type specific gene expression is the result of noncommutative sequences like K1 K2 and K2 K1 that silence certain promoters. The three-dimensional structure of the genome is a likely setting for this type of regulation.
Gene regulation is known to take place in three-dimensions, as observations of DNA looping [72], nonrandom chromosome packing [73], and clustered transcription factories [74] have shown. However, the factors that affect chromosome structure are non-specific. One such factor is the ubiquitous zinc finger protein CCCTC binding factor (CTCF) [75], which functions as both an activator of transcription by bringing enhancers and promoters together [76, 77] and as a repressor by insulating genes [78, 79]. Epigenetic modifications, such as histone methylation and acetylation [80–82], also affect three-dimensional structure. In addition, DNA looping was observed in the context of allelic exclusion during B- and T-cell lineage specification where individual alleles were recruited to heterochromatic regions while the other underwent recombination [33, 34]. Consequently, the sequestration model predicts that temporal permutations of a small set of chromatin modifying factors could specify a large number of potential chromosomal conformations and lead to different expression states and corresponding cell fate decisions.
New technologies such RNA-seq and ChIP-seq can be used to test the predictions of the noncommutativity hypothesis at the genome level. Epigenetic drugs such as azacytidine and trichostatin A inhibit DNA methylation [83] and histone deacetylation [84], respectively, and have been shown to cause global changes in gene expression alone and in combination [83, 85]. The sequestration hypothesis predicts that perturbations to the three-dimensional structure of the chromosome are noncommutative, so distinct gene expression states may be reached by permuting the order in which epigenetic drugs are applied. While the sequestration model may underlie chromosome folding, the ratchet model could form the basis of phosphorylation networks. For example, mass spectrometry studies have revealed complex phosphorylation patterns [86, 87], though the number of kinases and phosphatases is comparatively small and the networks are highly interconnected [10, 11]. As phosphoproteins are the mediator of extracellular signals, ordered disruption of signaling pathways could also lead to distinct gene expression configurations.
Analogously, the ratchet model may aid in the specification of distinct neural activity patterns, owing to the fact that connections between the different hippocampal layers overlap [12, 88]. While superficial neurons can be activated in response to spatial cues, deeper layers can be selectively activated by time sequences of inputs [40, 41, 89]. These results suggest the hypothesis that neural networks may be noncommutative. In particular, experimental support exists for the role of the dentate gyrus in pattern separation and orthogonalization by way of ensuring that even quite similar memory representations use distinct subsets of neurons [90, 91]. The ratchet model, by ordering inputs in time, is one way of reaching these specific subsets if the number of input neurons is smaller than the number of targets neurons. Memories share many common elements, including shape, color, smell, and sound, which poses problems for recall. We hypothesize that older, “fuzzier” memories could be those relegated to very long ratchet sequences. According to this hypothesis, memories are not forgotten, but are instead increasingly difficult to access, and memories that are not consolidated are those that never formed a unique ratchet sequence.
Beyond resolving bottlenecks and generating specificity, noncommutative actions offer a new interpretation of how cell fate decisions and other stepwise processes occur on abstract regulatory landscapes. The classical Waddington landscape view of development holds that cells decay to attractor configurations representing terminal outcomes [92]; this is consistent with a boolean network with many variables X converging to a fixed point [93]. In a static landscape, the final outcome is determined a priori by the nearest energy minimum. What then determines the initial configuration? In organisms such as Drosophila, maternal patterning of the embryo may account for this initial bias [94]; but in other organisms that employ mechanisms like multilineage priming [82, 95], it is not clear that every cell fate decision is made at the beginning.
Sequential logic allows cells to reach their final fate on a dynamic landscape. In the system of Fig 8A (top), for example, it is not possible for cells in the blue configuration to transition to the red fate by increasing X2, because this involves an uphill climb. However, the regulators of genetic networks may also affect the landscape directly. This is seen in Fig 8A (bottom) where the sequence K1 K2 P1 changes the landscape in such a way that the overall cost of reaching the same endpoint is much lower than the direct path (Fig 8A, top). This can be understood as the effect of regulators acting on additional variables V, which modulates the landscape in X space. For example, TFs can recruit chromatin regulators that modify global three-dimensional chromosome structure and future TF accessibility [74, 76, 96, 97], or kinases can sequester substrates in the nucleus to prevent their subsequent activation [53, 54]. Because sequential logic acts on the V’s as well as the X’s, changes that appear to be small in one dimension (Fig 8B, left) actually involve large excursions in the full space (Fig 8B, right). As a consequence, in noncommutative regulation, the landscape changes and cells can take on fates that were not accessible at the beginning.
Previous theoretical models have explored dynamic regulatory landscapes in the form of bifurcations [98, 99]. In these models, a set of kinetic parameters determines the positions of minima and maxima in the landscape. However, the noncommutative model advanced here is fundamentally different, in that using the regulators to move through X changes the landscape directly. This could happen, for example, if acting on X1 with K1 hides it from the effect of K2. Uncoupling of targets in this way may underlie the distinct effects of signals like FGF at different stages of development [35–38]. It will be interesting to explore time series data for hints that some genes pulse ON and OFF in order to protect their promoters from the actions of promiscuous regulators.
Multistep processes other than development can benefit from the type of noncommutative regulation highlighted in Fig 8. What seems like an intractable problem at the start becomes much more feasible if one realizes that the effects of actions change with time and context. This intuition is why thinking in terms of commutators [A, B] = AB − BA can make complex problems more soluble: the desired effect is often what is leftover after performing and undoing a sequence of actions. Several examples illustrate this concept.
With its increased capacity for generating diversity, sequential logic is likely to be used in evolution. A recent theoretical example in social bacteria demonstrated that in evolving a new quorum sensing receptor-ligand pair, adding new receptors prior to ligands is preferred over the opposite path [45]. An analysis of the stability and catalytic activity of a family of bacterial β-lactamase mutants showed that the ability to evolve new substrate specificity is contingent on mutations that first stabilize the protein active site [46, 100]. Finally, biological networks evolve the same functions in different orders, but the order in which these functions arise dictates which other genotypes can be reached by neutral mutations [44]. These results suggest that permuted sequences of mutation events may have different fitness costs. With extensive artificial evolution experiments underway in protein engineering [100] and bacterial mutation accumulation [47], coupled with progress in sequencing technologies, it will be possible to test this hypothesis by permuting the conditions that promote mutation.
Sequential logic can also be applied in synthetic biology to build circuits with memory [43, 101–103]. In general, the toolkit that permits up- and downregulation of genes is small, with a few staples like Lac, Tet, and Ara [104]. Significant effort has been put into generating logic gate (AND/OR) promoters [30]. To further expand the toolkit, it has been proposed that more orthogonal regulators be developed [105]. We suggest that sequential logic may be a more promising strategy to scale up the number of targets that can be independently controlled by permuting in time a small number of controllers.
More broadly, sequential logic can be used to accomplish experimental goals not possible in single-step approaches. For example, in multiplexing mRNA detection in single cells, we previously used a sequential hybridization scheme that permits the number of barcodes to exponentially [106], whereas combinatorial schemes can only specify approximately 30 barcodes. We expect many single-cell experiments to benefit from a sequential strategy in which detours facilitate achievement of the main goal with high efficiency.
Finally, our results connect outside of biology to strategic planning in social, political, and economic arenas. Anyone familiar with negotiating knows about the limitations inherent in trying to make interconnected groups of people move in specific directions, especially when the actions affect all participants at once. Multiparty negotiations and tournaments may benefit from time-ordered strategies in which enemies temporarily team up, or fringe interest groups are transiently pacified. Indeed, a conclusion from the sequestration model is that the most highly regulated targets need to be protected prior to satisfying the ones with fewer connections. Determining whether this prediction is borne out in congressional and international negotiations, for example, is an interesting question for political science. Evidence for noncommutative effects in games exists in that the initial seeding in a tournament can bias its outcome [107], and that long-term goals change players’ strategies in in the repeated prisoner’s dilemma [108]. In conclusion, the direct path to an outcome in a networks with many interacting parts may have many unintended and prohibitively expensive consequences. A multi-step strategy may achieve the same outcome with minimal cost and side effects.
In this section we determine how many targets are controlled by the same regulators in the connectivity matrix A. Then we extend A to more than 2 dimensions.
If ln = lm = 1 it is clear that each Ai,j corresponds to a single target and that each target appears only once. In general, however, a target can appear in multiple entries of A (cf. Eq (6)). To see this, consider the bipartite graph formed by all the targets and all the K’s, but none of the P’s. The handshaking lemma from graph theory [59] says that the total number of edges is one half the sum of the degrees of each vertex, which is either ln for a target or some number pn for a K regulator. There are Nln total edges, so we find 1 2 ( N l n + n p n ) = N l n or p n = N n l n for the number of links coming from each K. Similarly, the number of links emanating from each P is p m = N m l m. In terms of the connectivity matrix, pn and pm correspond to the number of unique targets in each row and column, respectively.
Because K1 connects to a fraction p n N of the targets, it follows that K1 and P1 together connect to a fraction p n p m N 2 of the targets. Therefore, the total number of targets connecting to K1 and P1 is M = N ( p n p m N 2 ) = p n p m N. Another way to see this is to consider one target in the intersection of K1 and P1. This one target uses up one of each of the regulators and one unit of connectivity, leaving a total of M = ( n - 1 l n - 1 ) ( m - 1 l m - 1 ) ways to connect other targets to the same pair of regulators. It is easily verified that these two formulations for the number of targets per matrix entry M are equivalent. This illustrates that there is not simply a one-to-one correspondence between the entries of A and the targets.
There was nothing special about the labels K and P in the above paragraphs. Thus, the connectivity matrix can easily be extended to a u-dimensional connectivity tensor where u is the number of pools of regulators. Each pool has ni regulators connecting to lni targets, and each target connects to p n i = N n i l n i regulators of pool i, ∀i ∈ {1, …, u}. The total number of targets and the total number of targets per entry are extensions of the u = 2 case, giving
N = ∏ i = 1 u n i l n i (10)
distinct targets and
M = ∏ i = 1 u p n i N u - 1 = ∏ i = 1 u n i - 1 l n i - 1 (11)
targets controlled by one factor from each of the u pools. S1A Fig shows an example network with u = 3 pools.
The number of configurations in combinatorial logic is the number of ways that N targets can each be bound by exactly u regulators, where each regulator comes from a different pool. In the main text we analyzed the case u = 1 and ln = 1. Here we extend those results to arbitrary u and ln.
First consider the case u = 2, corresponding to a pool of K’s and a pool of P’s. Whereas in the ratchet model, Ki and Pj acted separately on the entries of A, in combinatorial logic the pair (Ki, Pj) is needed to switch Ai, j from 0 to 1. Many such pairs may be active at any one time. We write this formally as
K , P A i , j = 1 if K i ∈ K and P j ∈ P 0 else, (12)
where {K} denotes a subset of the K’s. The notation (⋅, ⋅) means that a combination of factors acts on the target, instead of just a single factor.
If ln = lm = 1 there are (2n − 1)(2m − 1) + 1 ways to pick at least one of n K’s and one of m P’s, plus one way to pick nothing. If lm = 1 and ln > 1, then for a certain number α ≤ n of the K’s, any subset containing α or more K’s has the same effect as activating all n K’s at once. For example, in Eq (6), the action of ({K1, K2}, {P1, P2}) is sufficient to activate all targets in the n = m = 3, ln = lm = 2 network. To determine α, recall that there are M targets in each entry of the connectivity matrix A. Choosing i K’s means that the total number of targets is M × i, but a single column of A only contains pm unique targets. Each target is connected to ln K’s, so for a target in the intersection of i K’s and a single P, there are ln − i spots left over to choose n − i K’s and lm − 1 spots left over to choose m − 1 P’s, or ( n - i l n - i ) ( m - 1 l m - 1 ) ways total. Using the principle of inclusion-exclusion [59] this means that α is the smallest i such that
M × i - ∑ i ′ = 2 min i , l n - 1 i ′ i i ′ n - i ′ l n - i ′ m - 1 l m - 1 ≥ p m . (13)
By choosing α K’s, the number of unique targets in a column of A that can be turned ON is exactly the number represented in that column. Because all subsets with α, α + 1, …, n − 1 K’s are redundant, here are only ( 2 n - 1 ) - ∑ i = α n - 1 ( n i ) subsets of K’s that contribute to unique configurations, leaving a total of [ ( 2 n - 1 ) - ∑ i = α n - 1 ( n i ) ] ( 2 m - 1 ) + 1 unique configurations.
If the P’s also have redundant connections, the result generalizes to
Theorem 1 The number of configurations in combinatorial logic with parameters n, m, ln, lm, and u = 2 is 2 n - 1 2 m - 1 + 1 - ∑ i = α n - 1 n i 2 m - 1 - 2 n - 1 ∑ i = β m - 1 m i + ∑ i = α n - 1 n i ∑ i = β m - 1 m i , (14) where α (resp. β) is the smallest number of K’s (resp. P’s) having the same effect as all K’s (resp. P’s) at once.
This result is obtained by counting all pairings of K’s and P’s, then subtracting those pairings that have a redundant effect. For example, any combination using K3 is redundant in the connectivity matrix of Eq (6). Finally, those pairings that were excluded twice are added back in.
This result generalizes to all u with slight modifications. Because one factor from each of u pools is now required, the combinatorial equation determining state of a target is
K 1 , K 2 , … , K u A i , j , … , k = 1 if K 1 i ∈ K 1 , K 2 j ∈ K 2 , … , K u k ∈ K u 0 else. (15)
Here the double subscript Kik indicates the kth factor in the ith pool. Determining αi for each pool i of regulators requires finding the pool j ≠ i which maximizes the number Ni of targets controlled in two dimensions. If we choose αi or more regulators in the ith pool, then there is a redundancy in the jth dimension, whereas any choice of fewer than αi regulators activates fewer than Ni targets. Write N i = max j ≠ i { ( n i l n i ) ( n j l n j ) } the total number of targets and p n j = N i n j l n j the number of targets in any column of the the equivalent ni × nj connectivity matrix regulated by pools i and j. It is easy to see that these parameters reduce to their previous definitions for u = 2. Now define M i = ( n i - 1 l n i - 1 ) ( n j - 1 l n j - 1 ) as the number of targets in each entry of the equivalent ni × nj connectivity matrix. As above, αi is now the smallest r such that
M i × r - ∑ r ′ = 2 min r , l n i - 1 r ′ n i - r ′ l n i - r ′ n j - 1 l n j - 1 ≥ p n j . (16)
Once αi is determined for each pool i, the inclusion-exclusion sum can be extended using standard arguments [59]. Define by
S k = ∑ σ ∈ u k ∏ i ∈ σ ∑ j = α i n i - 1 n i j ∏ i ∉ σ 2 n i - 1 , (17)
where σ denotes all k-subsets of {1, …, u}. Then we have the final result
Theorem 2 The total number of configurations in combinatorial logic with u pools and parameters ni, lni, i ∈ {1, …, u} is S = 1 + ∑ k = 0 u - 1 k S k . (18)
This result reduces to Theorem 1 when there are only u = 2 pools. At most there are ∏ i = 1 u ( 2 n i - 1 ) ways to specify at least one target, corresponding to the 0th-order term in Eq (18). Increasing the connectivity through the lni can only reduce the number of configurations. This behavior is shown in S1B Fig for the symmetric case that all the ni and lni are equal. As u is increased the number of configurations increases dramatically, but the scaling is actually subexponential, i.e. less than 2N. Increasing connectivity through lni shifts the curves to the right.
To establish the correspondence between the reachable configurations of ratchet network (ln = lm = 1, T = 1) and the lonesum matrices, we must show (i) that A avoids the patterns ( 1 0 0 1 ) and ( 0 1 1 0 ) in any 2 × 2 sub-block, and (ii) that any lonesum matrix can be constructed from K and P actions. First observe that the value 1 in Ai, j indicates the last K affecting that index must have followed a P, whereas 0 indicates the last P must have followed a K. For the first restriction we have ( 1 0 0 1 ) implies ( P 1 … K 1 K 1 … P 2 K 2 … P 1 P 2 … K 2 ). This means P2 follows K1 follows P1 follows K2 follows P2, which is a contradiction, showing that this 2 × 2 block is unreachable. The other five unique 2 × 2 blocks are all reachable with elementary sequences. This establishes point (i) that the reachable configurations are a subset of the lonesum matrices.
To establish point (ii) that the lonesum matrices are a subset of the reachable configurations, we use an equivalent formulation of the lonesum matrices as staircase matrices composed of the rows aj = (1, …, 1, 0, …, 0) with the last 1 appearing at position ij subject to the constraint that ij ≤ ij − 1 for all ∀j ∈ {2, …, n} [109]. It is easy to see that the pattern of ones resembles an inverted staircase. We show via induction that any staircase matrix can be constructed from K and P actions. The nth row is obtained by the sequence Kn Pin + 1⋯Pm which leaves 1’s at the first in indices and 0’s at the remainder. Now assume that the kth row is obtained by the sequence Kk Pik + 1⋯Pm without affecting any of the rows n, n − 1, …, k + 1. Then the sequence Kk − 1 Pik − 1 + 1⋯Pm puts 1’s at the first ik − 1 indices of row k − 1. Because ik − 1 ≥ ik ≥ ⋯ ≥ in, none of the Pik − 1 + 1, …, Pm turn a 1 to a 0 in rows n, n − 1, …, k + 1, k. This proves the induction hypothesis and shows that the staircases matrices are a subset of the reachable configurations.
Together with the fact that the reachable configurations are a subset of the staircase matrices, this implies that the reachable configurations and the lonesum matrices are in fact the same set, and we have
Theorem 3 The number of reachable configurations in the (n, m) ratchet network with ln = lm = 1 and threshold 1 scales as the poly-Bernoulli numbers B m - n = B n - m.
With T = 2, only targets in state 2 are ON. Once a 0-1 configuration of A is obtained, however, it is a simple matter to convert it into an ON/OFF configuration by applying all the K’s. Here we use the fact that 1’s can be reached from above and below to prove the
Theorem 4 In the ratchet network represented by the matrix A with ln = lm = 1 and threshold T = 2, all binary 0-1 configurations are reachable.
Proof. We use an induction argument analogous to the proof of Theorem 3. Suppose that in row n a set of r ≤ m indices {nj} = {nj1, …, njr} should be ON. First prepare every target in row n in the 1 state using Kn, then use the sequence Kn Pjr + 1⋯Pjm to obtain 2’s at {nj1, …, njr} and 1’s at {njr + 1, …, njm}. Now assume that we can prepare rows n, n − 1, …, k + 1 in a similar 1-2 configuration with the rest of the matrix 0. We want to show that we can add row k to this set without affecting any of the previous rows. Assuming that a set of s ≤ m indices {kj1, …, kjs} should be ON, apply the sequence P j 1 … P j s K k 2 P j s + 1 … P j m to obtain 2’s at {kj1, …, kjs} and 1’s at {kjs + 1, …, kjm}. Now, because {Pj1, …, Pjs}∪{Pjs + 1, …, Pjm} = {P1, …, Pm}, all 2’s and 1’s in rows n, n − 1, …, k + 1 are now 1’s and 0’s, respectively. Applying the sequence Kn Kn − 1⋯Kk + 1 reestablishes the 1-2 configuration we had prior to fixing row k and leaves 0’s at rows 1, …, k − 1. Now that row k is also in the proper 1-2 configuration, we have proved the induction hypothesis. Once all rows in the proper 1-2 configuration, the sequence P1⋯Pm obtains the matrix in the 0-1 configuration. Since this procedure can be repeated for any collection of indices {{1j}, …, {nj}}, it follows that all binary 0-1 matrices are reachable.
When the connectivity parameters ln and lm exceed 1, certain sequences in the threshold 1 ratchet network become redundant. Our goals in this section are to (i) to characterize the redundant sequences by the number of K’s and P’s, and (ii) count the non-redundant sequences. This will obtain an upper bound on the number of configurations.
We want the shortest sequences that can activate or (deactivate) all targets; any sequences longer than this are redundant. To see why this is so, we need the concept of a cycle. We say that a target has gone through a cycle if has traversed the states 0, 1, 0 at some subsequent time points. We have the following lemma.
Lemma 5 Any sequence that takes all targets through a cycle is redundant.
Proof. The final configuration of any sequence is represented by the positions of the 1’s and 0’s of the connectivity matrix. Recall that Ai,j = 0 if an only if all targets represented by Ai,j are OFF in the final configuration. Permute the rows and columns of A until it is in staircase form with r ≤ min(n, m) steps, where a step is a group of adjacent rows or columns having the same number of 1’s and 0’s. The steps partition the rows and columns of A into subsets of indices {i1, i2, …, ir} and {j1, j2, …, jr} where the kth step is defined by 1’s at rows ik to ik + 1 − 1 and 0’s at columns jk to jk + 1 − 1. Then the sequence ∏ k = 1 r K i k … K i k - 1 P j k … P j k - 1 obtains the desired configuration of 1’s and 0’s. Being able to write a staircase matrix for the final configuration means that every target ON in the final configuration occurs only where there are 1’s in the matrix. These targets are never affected by a P in this procedure; they do not go through a cycle. Because any allowed configuration can be reached from this procedure, it follows that any sequence that uses a cycle is redundant.
Knowing that the non-redundant sequences must avoid cycles, it suffices to find the longest sequences that can be written before cycles appear.
Lemma 6 For each value of ln (lm), the maximum number of K’s (P’s) that can be used before all targets are activated (deactivated) is n − ln + 1 (m − lm).
Proof. A sequence that activates all targets has no intervening P’s. Recall that a single K activates at most N n l n targets. Then, prior to the last K being used, the number of activated targets is N - N n l n = N n ( n - l n ) ≤ N n l n ( n - l n ). This means there are at most n − ln groups of targets controlled by different K’s. Thus, at most n − ln K’s are used before the last K is used, and n − ln + 1 K’s must be sufficient to activate the complete set. The maximum number of P’s that can be used is only m − lm because we can think of every sequence starting in the zero configuration as having been preceded by a single P; this modification puts the P’s on equal footing with the K’s.
With this characterization of the non-redundant sequences our goal is to recursively eliminate sequences that use n − ln + 1 K’s and m − lm P’s. We first find the number of sequences that use up to m − lm P’s, which forms the top row in each (n, m) block in S2 Fig. Then we use these values to recursively find the number of sequences using up to n − ln + 1 K’s. The strategy is to subtract from the total number of sequences at a given (ln, lm) all those sequences using the forbidden number of regulators in order to get the new total.
Denote by a n m the number of sequences using m P’s when the total number of K’s is n. If m = 1, then all B 1 - n = 2 n sequences (except for the empty sequence) use a K and none use a P. If m = 2, the maximum number of P’s that can be used is m − lm = 1. Discarding the 2n sequences with no P, the number of sequences using a single P is
a n 1 = B 2 - n - 2 n 2 . (19)
Division by m = 2 is required to account for the fact that there are ( m 1 ) = m different ways of starting each sequence with a P, and we consider both of these equivalent. Having determined a n m, it is straightforward to determine a n m + 1. Because there are m + 1 P’s to choose from, there are ( m + 1 m ) a n m ways to write sequences with m P’s, ( m + 1 m - 1 ) a n m - 1 ways to write sequences with m − 1 P’s, …, ( m + 1 0 ) 1 ways to write sequences with 0 P’s, the only remaining sequences are those with m + 1 P’s. Knowing that the total number of sequences is B m - n, this leaves
a n m + 1 = B m - n - 2 n - ∑ j = 0 m m + 1 j a n j m + 1 (20)
total sequences using m + 1 P’s when the total number of K’s is n. Having determined this number, we can sum up all the sequences using m − lm P’s to get the first row of the (n, m) block in S2 Fig. Denote by c n m ( l n , l m ) the lmth column and lnth row of the (n, m) block. The column headers c n m ( 1 , l m ) are given by
c n m 1 , l m = 2 n + ∑ j = 0 m - l m m j a n j . (21)
We can determine the row entries for ln > 1 in the same way that we determined the column headers, the only difference being that the total number of sequences is c n m ( 1 , l m ), not B m - n unless lm = 1. Denote by b m n ( l m ) the number of sequences using n K’s when the total number of P’s is m and the P connectivity is lm. For fixed m, lm and n = 1, there are
b m 1 l m = 2 m - ∑ j = 0 l m - 1 m j , (22)
sequences, as all but the empty sequence use a single K. In complete analogy to Eq (20) we find there are
b m n + 1 l m = c n m 1 , l m - ∑ j = 0 n - l n + 1 n + 1 j b m j l m (23)
sequences using n + 1 K’s when the total number of P’s is m. Unlike in the equation for a n m, there is no division by n + 1 because all sequences starting with a different K are different. Finally, we can sum up all the sequences using n − ln + 1 K’s to get the
Theorem 7 The number of minimal length sequences in the (n, m, ln, lm) ratchet network with threshold T = 1 using no more than n − ln + 1 K’s and m − lm P’s is c n m l n , l m = ∑ j = 0 n - l n + 1 n j b m j l m . (24)
We used this formula to compute each entry in S2 Fig. Because of the complexity of this procedure, we checked it against a computer algorithm operating with the following steps. In step 1 find all B m - n sequences in the ln = lm = 1 case. In step 2 increase the connectivity (ln or lm) and find all sequences of a given length; group them by the configuration they generate. Some of these sequences will not appear in the list generated by step 1: for example, both K1 K2 and K2 K1 will be found in step 2. We are interested in index permutation e.g. 1 → 3, not letter permutation, so in step 3 delete all sequences in each length group not appearing in step 1. Repeat steps 1–3 with this new list of sequences until ln = n − 1. This code, implemented in Matlab Version 2015b, gave exact agreement with Theorem 7.
We now show that rules restrict the reachable configurations of the sequestration model in the main text to the connected one-colorings of the reduced n-network.
Theorem 8 There is a one-to-one correspondence between the reachable configurations of the reduced n-network and the connected one-colorings.
Proof. The converse direction, reachable implies connected, is easier to prove and will be discussed first. Assume that all configurations in the reduced n-network so far reached are connected. The next configuration will be reached by turning all 0’s to i’s or all j’s to 0’s by application of Ki or Pj, respectively. The k-arm targets sharing state i with the 2-arm target {0, 1, i} are either in the same state as some other 2-arm target {0, 1, i′} or are in the 0 state. So application of Ki cannot change the connectivity of the configuration. Furthermore, a k-arm target can be in the j state only if the target {0, 1, j} is in the j state, so these targets will still be matched after application of Pj. Thus, any configurations reachable from a reachable configuration must be connected.
The forward direction, connected implies reachable, is less trivial. In order to prove that all connected one-colorings in the n-network are reachable, we will use the strong form of mathematical induction. Assume the theorem holds for all networks up to n − 1. Embedded within the full n-network of 2n − 1 targets is the reduced n-network on 2n − 1 targets. Within the reduced n-network is a set of 2n − 2 targets able to access {0, 1, 2} and all subsets (including Ø) of the integers {3, …, n}. Thus, we can substitute 2 → 1 as the ON state in this embedded network and all connected one-colorings (of 2) will be reachable. The same holds in general for all 2n − k targets able to access {0, 1, k} and all subsets of the integers {k + 1, …, n}. In each of these embedded networks the substitution k → 1 as the ON state will enable us create any connected one-coloring.
Pick any connected one-coloring (of 1) in the n-network. Its opposite configuration is formed by the transformation at each target g of 1 → 0 and 0 → kmin, where kmin = min{k∈g|xpos({0, j, k}) = 0} is the smallest index that g shares with a corresponding 2-arm target at position pos({0, j, k}) of x (possibly in the full network) currently in the 0 state. The opposite of a connected one-coloring is clearly connected, because all the connected 1’s are now 0, and all the 0’s are in the same state as the 2-arm target {0, j, kmin}. If it is possible to reach the opposite configuration, then application of the sequence K1 P2…Pn yields the desired one-coloring of the n-network.
To show that the opposite configuration of the chosen one-coloring is indeed reachable, isolate the embedded networks one-by-one by application of the sequence Kk K1 Pk for k = 2, …, n, so that the targets in the n − k + 1-network are the only targets in the 0 state. By hypothesis, the connected one-colorings are reachable in all embedded networks which have at most n − k states besides 0, 1, and k. The opposite configuration in the n-network is composed of connected one-colorings (of k) in each embedded network; these are are reachable. Therefore, the one-coloring of the n-network is reachable via K1 P2…Pn. This procedure holds for any one-coloring.
How many configurations are reachable in the full n-network? Let this number be cn. The following theorems derive lower and upper bounds for cn in terms of the number of one-colorings.
Theorem 9 The formula f(n + 1) for the number of connected one-colorings in the reduced n + 1-network is a lower bound for cn.
Proof. The full n + 1-network can be partitioned into a set of 2n targets having a 1 and all subsets of {2, …, n + 1}, and 2n − 1 targets that lack 1 but have all nonempty subsets of {2, …, n + 1}. The latter set of targets is an embedded full n-network, while the former is the reduced n + 1-network. All 2(n + 1) letters are needed to form the one-colorings in the reduced n + 1-network. Every one-coloring is finally obtained by applying some permutation of K1, P2, …, Pn + 1 to a configuration that uses (at most) the states 2, …, n + 1 and 0, i.e. the full n-network. Because K1 and P1 do not affect the targets of the the embedded full n-network, there must be (at least) one sequence using only {K2, …, Kn + 1} and {P2, …, Pn + 1} that prepares the embedded full n-network in the aforementioned configuration, which means we may associate a one-coloring with (at least) one of the cn sequences in the embedded full n-network. Therefore, multiple configurations in the full n-network may map to the same one-coloring in the reduced n + 1-network. Conversely, if two one-colorings are different, they are distinguishable by their configurations immediately preceding the final K1, P2, …, Pn + 1 sequence, and must therefore map to different configurations in the full n-network. Together, these statements imply that the map from configurations in the full n-network to one-colorings in the reduced n + 1-network is many-to-one, but the map from one-colorings to configurations in the full n-network is one-to-one. Therefore, f(n + 1) ≤ cn.
Theorem 10 An upper bound on cn is n f n + n n - 1 f n - 1 f n - 1 + ⋯ + n ! f n - 1 ⋯ f 2 - 1 f 1 + 1 = ∑ k = 1 n n k ∏ j = n - k + 2 n f j - 1 f n - k + 1 + 1 . (25) where (n)k = n(n − 1)⋯(n − k + 1) is the falling factorial.
Proof. There are nf(n) one-colorings in the full n-network, plus one origin. Each one of the one-colorings can be thought of as the origin of an n − 1-network, which in turn generate (n − 1)f(n − 1) one-colorings in an embedded n − 1-network, for a total of
n f n n - 1 f n - 1
configurations using 1, 2, and perhaps 0, hence termed two-colorings. However, one of the f(n) one-colorings is the 0 state of the n-network, so it does not generate any two-colorings. Thus, there are at most 1 + nf(n) + n(n − 1)(f(n) − 1)f(n − 1) zero-, one-, and two-colorings. Now assume that the number of k-colorings is
n n - 1 ⋯ n - k + 1 f n - 1 f n - 1 - 1 ⋯ f n - k + 2 - 1 f n - k + 1 .
Of these,
n n - 1 ⋯ n - k + 1 f n - 1 f n - 1 - 1 ⋯ f n - k + 2 - 1
are origins of an n − k-network, meaning they are actually k − 1-colorings; they cannot generate any k + 1-colorings. The remaining
n n - 1 ⋯ n - k + 1 f n - 1 f n - 1 - 1 ⋯ f n - k + 2 - 1 f n - k + 1 - 1
are genuine k-colorings which can generate f(n − k) one-colorings in the n − k-network, or equivalently, k + 1-colorings. Thus, the total number of zero-, one-, two-, …, k + 1-colorings is no more than
n n - 1 ⋯ n - k + 1 f n - 1 f n - 1 - 1 ⋯ f n - k + 2 - 1 f n - k + 1 - 1 .
This induction argument proves the statement.
First we define what it means to be an origin and an orbit in the threshold-1 ratchet network and determine the number of orbits as a function of n and m. Then we prove that the configurations in the sequestration network are defined by reversible paths.
A forbidden configuration in the ratchet network contains some row or column permutation of the pattern ( 1 0 0 1 ) on any 2 × 2 sub-block of the connectivity matrix A. This is the minimum violation, but larger blocks may violate this pattern as well, for example ( 1 0 0 1 1 0 ) has 2 violations. Furthermore, application of any of the K’s or P’s in this sub-block will relieve at least one of these violations. Therefore, we define an i, j-orbit in the ratchet network as the locus of configurations having a forbidden configuration on an i × j sub-block that does not use the corresponding set of i K’s and j P’s. The origin of any i, j-orbit is the configuration having all remaining nm − ij entries of A equal to 0 (or all 1 to make the case of having only P actions symmetric with having only K’s). A matrix X having the same forbidden i × j sub-block as an origin Y is not considered to be in the orbit of Y if (i) there is no sequence of actions that transforms Y to X, or (ii) if the sequence involves one of the forbidden K’s or P’s. With these restrictions, the number of origins is equal to the number of orbits.
Denote by c i j the number of orbits in a ratchet network of size n × m with violations involving i ≤ n K’s and j ≤ m P’s. If i = j = 2 there are 2 i j - B 2 - 2 = 2 forbidden configurations that turn into origins for the remaining n − i K’s and n − j P’s. There are more orbits in these smaller networks. For every i′, j′ ≥ 2 there are ( i i ′ ) ( j j ′ ) c i ′ j ′ B i - i ′ - ( j - j ′ ) configurations reached by orbits using i′ K’s and j′ P’s. Only configurations not reached by these orbits are available as new origins when the number of K’s and P’s not to be used is i and j, respectively. Finally, there are ( n i ) ( m j ) ways to specify i ≤ n K’s and j ≤ m P’s. Then we have the
Theorem 11 For a given set of i ≤ n K’s and j ≤ m P’s, the number of i, j-orbits is cij=(2ij−Bi−j)−∑i′,j′≥2i′+j′≤i+j−1i,j(ii′)(jj′)ci′j′B′i−i′−(j−j′), (26) and the the total number of i, j-orbits in the n × m ratchet network is C i j n , m = n i m j c i j , (27) where B ′ i - i ′ - j - j ′ = B i - i ′ - j - j ′ i f i - i ′ > 0 and j - j ′ > 0 2 i - i ′ i f j - j ′ = 0 2 j - j ′ i f i - i ′ = 0 . (28)
The modification B′ ensures that an orbit lacking allowable P’s (K’s) can still use K’s (P’s). A table of values of Eq (27) is given in S4 Fig.
We noted in the main text that configuration in the sequestration network can be joined by reversible paths. A path Ki Pj or Pj Ki is reversible if a configuration reached by the sequence of actions w is also reached by the either the sequence wKi Pj or wPj Ki, but not wKi or wPj, respectively. Thus we can also prove the
Theorem 12 There always exists a reversible path between any two configurations in an orbit of the sequestration network.
Proof. Let x be a configuration in an orbit using m ≤ n of the actions, and let P denote the locus of configurations reached from x. We now need to show that P must be reversibly reached from the origin. Denote by P ¯ the complement of P, so that any y ∈ P ¯ is reversibly reached from the origin. In order for there to be no reversible path between x ∈ P and y ∈ P ¯, there must always be a state i such that Ki increases the number of targets {⋅, i} in the i state and Pi increases the number of targets {⋅, i} in the zero state. Now assume there is a configuration z ∈ P using all m allowed states. z must have at least one target in the 0 state, but this is un-allowed, because then z would violate the connection rule. Therefore, there is a maximum number m′ < m of states used by any x ∈ P. Now assume there is a configuration z′ ∈ P using all m′ allowed states. But this implies that there is a single-arm target {0, j} that must be in the zero state. Then the action Kj takes z′ to a configuration y ∈ P ¯ and Pj takes y to z. This path must be reversible, and z′ is reached reversibly from the origin. By induction we conclude that m′ = 0 and that P = Ø. Finally, because any two configurations are reached reversibly from the origin, there is a reversible path between them.
Theorem 12 defines the orbits of the sequestration network as those configurations connected by reversible paths.
In this section we show how to write the K and P regulators as matrix operators in a manner consistent with both models considered in the paper. First we define the vector space V of configurations of the N targets, then we derive the operators that transform V.
Let x ∈ V. For a network with N targets we require that ∑i xi = N. This means that x has at least N entries, and in general dim x ≥ N. Therefore, we cannot use the standard state space of N-dimensional vectors, because the operators will not conserve the number of targets. Each target has a 0 state. The number D of independent directions accessible from 0 is called the dimension of the network, and the number T of steps one can move along each dimension is called the threshold. In the ratchet model, each target has a single ladder of states with variable threshold, so D = 1 and T is allowed to vary; in the sequestration model D = n but the threshold is T = 1.
Denote by Adi the fraction of the targets of type A in state i ∈ {0, 1, …, T} along dimension d. For a subset of the targets a K-type action causes population transfer between states (d, j) and (d, i) with i = j + 1, and a P-type action the reverse. If a K regulator acts for a short time we can write the “reaction rate” equation as
x ˙ A j = - g A x A j x ˙ A i = + g A x A j (29)
where gA > 0 is a proportionality constant. This defines a matrix differential equation
x ˙ = G dj · x (30)
with x ∈ R N ( D T + 1 ) × 1 the vector of populations of the DT + 1 states of the N targets and G dj ∈ R N ( D T + 1 ) × N ( D T + 1 ) the block diagonal matrix of rate constants between the j and j + 1 population states along dimension d. Eq (29) can be rewritten
x ˙ A j x ˙ A i = - g A 0 g A 0 · x A j x A i . (31)
Because Gdj is block diagonal, Eq (30) can be solved by exponentiation on each block:
x A j t x A i t = exp - g A 0 g A 0 t · x A j 0 x A i 0 = e - g A t 0 1 - e - g A t 1 · x A j 0 x A i 0 . (32)
The restriction of the model from a continuous range of population states xAi ∈ [0, 1] to the boolean values {0, 1} formally emerges by considering the “reaction” K catalyzes on its targets to have gone to completion. We do this by taking the the limit t → ∞ in Eq (32) to get
x A j t x A i t = 0 0 1 1 · x A j 0 x A i 0 , (33)
so that the matrix Kdj defined by
K dj = lim t → ∞ exp G dj t (34)
is the block diagonal matrix having 1’s at (row, column) positions (1 + (d − 1)T + i, 1 + (d − 1)T + j) of each block that responds to K in dimension d and admits population transfer between from j to i.
Because K acts on all targets at once, it is insensitive to the initial state j. Thus the matrix corresponding to the action of K is
K d = ∏ j K dj , (35)
which is the block diagonal matrix having 1’s at (row, column) positions
1 + d - 1 T + 1 , 1 + d - 1 T + 0 , … , 1 + d - 1 T + T , 1 + d - 1 T + T - 1 and 1 + d - 1 T + T , 1 + d - 1 T + T
of each block that responds to K in dimension d.
This derivation can be repeated in the case that population goes in the opposite direction from at state j to a state i < j using a different set of rate matrices Hdj corresponding to the reverse of Eq (31). We obtain the block diagonal matrix Pd corresponding to the action of P in dimension d having 1’s at (row, column) positions
1 + d - 1 T + 0 , 1 + d - 1 T + 1 , … , 1 + d - 1 T + T - 1 , 1 + d - 1 T + T and 1 + d - 1 T + 0 , 1 + d - 1 T + 0
of each block that responds to P in dimension d. Whereas Kd is sub-diagonal, Pd is super-diagonal.
The Baker-Campbell-Hausdorf expansion shows that Kd in Eq (35) and in general any product of matrices Kd and Pd are generated by matrix exponentiation of commutators of the generators Gdj, Hdj. This is the origin of noncommutativity in both the ratchet and sequestration models.
An example in the sequestration network illustrates population transfer between states. In the n = 2 network on the targets A, B, and C the initial configuration of the network is represented by ( A 0 A 11 A 21 B 0 B 11 B 21 C 0 C 11 C 21 ) T = ( 1 0 0 1 0 0 1 0 0 ) T. Only targets A and C can access dimension 1, and only targets B and C can access dimension 2. Therefore the t → ∞ action of K1 on the network is given by
e - g A t 0 0 0 0 0 0 0 0 1 - e - g A t 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 e - g C t 0 0 0 0 0 0 0 0 1 - e - g C t 1 0 0 0 0 0 0 0 0 0 1 1 0 0 1 0 0 1 0 0 → t → ∞ 0 1 0 1 0 0 0 1 0 . (36)
Only A and C advance to state 1 and the number of targets (3) is conserved.
|
10.1371/journal.pcbi.1004595 | Network-Based Integration of Disparate Omic Data To Identify "Silent Players" in Cancer | Development of high-throughput monitoring technologies enables interrogation of cancer samples at various levels of cellular activity. Capitalizing on these developments, various public efforts such as The Cancer Genome Atlas (TCGA) generate disparate omic data for large patient cohorts. As demonstrated by recent studies, these heterogeneous data sources provide the opportunity to gain insights into the molecular changes that drive cancer pathogenesis and progression. However, these insights are limited by the vast search space and as a result low statistical power to make new discoveries. In this paper, we propose methods for integrating disparate omic data using molecular interaction networks, with a view to gaining mechanistic insights into the relationship between molecular changes at different levels of cellular activity. Namely, we hypothesize that genes that play a role in cancer development and progression may be implicated by neither frequent mutation nor differential expression, and that network-based integration of mutation and differential expression data can reveal these “silent players”. For this purpose, we utilize network-propagation algorithms to simulate the information flow in the cell at a sample-specific resolution. We then use the propagated mutation and expression signals to identify genes that are not necessarily mutated or differentially expressed genes, but have an essential role in tumor development and patient outcome. We test the proposed method on breast cancer and glioblastoma multiforme data obtained from TCGA. Our results show that the proposed method can identify important proteins that are not readily revealed by molecular data, providing insights beyond what can be gleaned by analyzing different types of molecular data in isolation.
| Identification of cancer-related genes is an important task, made more difficult by heterogeneity between samples and even within individual patients. Methods for identifying disease-related genes typically focus on individual data sets such as mutational and differential expression data, and therefore are limited to genes that are implicated by each data set in isolation. In this work we propose a method that uses protein interaction network information to integrate mutational and differential expression data on a sample-specific level, and combine this information across samples in ways that respect the commonalities and differences between distinct mutation and differential expression profiles. We use this information to identify genes that are associated with cancer but not readily identifiable by mutations or differential expression alone. Our method highlights the features that significantly predict a gene’s association with cancer, shows improved predictive power in recovering cancer-related genes in known pathways, and identifies genes that are neither frequently mutated nor differentially expressed but show significant association with survival.
| The sequencing revolution of the last decade is producing vast amounts of data with clinical relevance. However, translating these data to biomedical understanding remains a formidable challenge due to the typically low statistical power associated with sequencing studies, disease heterogeneity, experimental limitations and more. A promising strategy to circumvent some of these problems is the integration of sequence data with other types of “omic” data [1]. In the context of cancer, comprehensive data generation efforts such as The Cancer Genome Atlas (TCGA) and and the COSMIC cancer gene census [2] provide excellent opportunities in this regard, since they interrogate large sets of samples for multiple types of omic data.
An important and well-studied problem in this field is the prioritization of genes for specific diseases. State-of-the-art methods for tackling this problem rely on the observation that proteins causing similar diseases tend to lie close to one another in a protein-protein interaction network. We have previously devised prioritization methods that start from known causal proteins and propagate their signal in the network to predict novel causal proteins [3, 4]. Here, we aim to harness the network propagation methodology to the integration of multiple omic data types in the context of cancer, with a view to gaining mechanistic insights into the relationship between molecular changes at different levels of cellular activity.
In recent years, there have been substantial efforts in integrating multiple omic data types that provide information on cancer pathogenesis and progression, with a view to predicting patient outcome, identifying drug targets, and understanding the functional relationships among key players in cancer. In the context of predicting patient outcome, Hofree et al. [5] used a network propagation based strategy to incorporate the functional relationships among mutated genes into the clustering of patients. They showed that the resulting clustering correlates with patient outcomes better than the clustering of patients according to mutation data alone. Similarly, several groups demonstrated that integration of transcriptomic data with protein-protein interaction networks leads to the identification of protein subnetworks that serve as reliable markers for the prediction of survival in such cancers as glioblastoma multiforme [6] and ovarian cancer [7].
In the context of understanding the functional relationships among key players in cancer, enrichment-based approaches aimed at identifying significantly mutated pathways provide insights into how different mutations influences similar biological processes [8]. Analysis of mutually exclusive mutations further elucidate the functional relationships among mutated genes by interpreting mutual exclusivity among mutations in the context of networks, thereby recovering key functional modules that provide systems-level insights into the mechanisms of pathogenesis [9]. Integration of sequence data with gene expression data based on eQTL analysis is also shown to be effective in the identification of cancer-related pathways [10]. These studies establish that the addition of network information can enhance predictive power in many applications, but most of these methods focus on a single data type in addition to network relationships. Though previous studies combine mutational or differential expression data with protein interaction networks, few use network information to integrate mutational and expression data. In particular, Nibbe et al. [11] propose a method that integrates protein expression data with mRNA expression data, with the purpose of extending the scale of of proteomic data that has limited coverage of the proteome. In Nibbe et al.’s study proteomic and transcriptomic data from different patients is used to integrate mRNA-level gene expression and protein expression data. However, efforts like TCGA make it possible to extract multiple types of omic data (mutation, mRNA expression, microRNA expression etc.). In this study, we aim to develop an algorithmic framework for the integration of these multi-omic data at the level of individual samples.
We stipulate that during pathogenesis of cancer, mutations in up-stream proteins may lead to transcriptional dysregulation of down-stream genes. Similarly, transcriptional dysregulation of some processes may lead to conservation of certain mutations during neoplastic evolution. The dynamics of the interplay between genomic mutations and transcriptional dysregulation likely involves signaling proteins (e.g., kinases, phosphatases, transcription factors) that mediate the relationship between mutated genes and dysregulated gene products. However, due to limitations in proteomic and phosphoproteomic screening [12], the changes in those mediator proteins may not be readily detectable from genomic and transcriptomic data alone. We propose that such “silent” proteins can be detected by integrating mutation and differential expression data in a network context, since these proteins are likely to be in close proximity to both mutated and differentially expressed proteins in the network of protein-protein interactions (PPIs).
Based on our hypothesis, we develop an algorithmic workflow aimed at quantifying the proximity of all proteins in the human proteome to the products of mutated and differentially expressed genes in each sample. The proposed workflow is illustrated in Fig 1. Here, our emphasis is on utilizing sample-specificity to be able to deal with molecular heterogeneity of pathogenesis at the population level. In order to utilize sample-specific data, we use network propagation to separately score proteins based on their network proximity to 1) mutated and 2) differentially expressed genes in each sample. This procedure provides us with two vectors in the space of samples for each protein: a “propagated mutation profile” indicating proximity to genes mutated in each sample and a “propagated differential expression profile” indicating proximity to genes differentially expressed in each sample. We then use these vectors to extract descriptive features for each protein, to be used for predicting its involvement in the disease being studied.
We apply the proposed method to breast cancer (BRCA) and glioblastoma multiforme (GBM) data obtained from The Cancer Genome Atlas (TCGA) project. First, we assess the power of mutation data, expression data, and network-based integration of these two in unsupervised prediction of genes known to play a role in each cancer. We show that one can gain significant predictive power by propagating mutation or expression data over a PPI network, as compared to using raw mutation or differential expression data (area under ROC curve (AUC) gains of 0.16–0.18 for BRCA and 0.17–0.27 for GBM). We then combine the two signals to derive several features and used these features to train a supervised predictor with further improved AUC of 0.836 for BRCA and 0.933 for GBM. Importantly, by using this predictor we are able to recover important proteins that are not readily revealed by molecular data. These genes are supported by the literature and by an independent cancer gene resource. This observation suggests that incorporation of network data can provide insights beyond what can be gleaned from sequence or expression data in isolation. Seven of those novel predictions are further found be significantly predictive of patient outcome. Our results also suggest important features that contribute significantly to the prediction of causal genes in breast cancer and glioblastoma multiforme, which provide insights into how the crosstalk among mutated and differentially expressed proteins contributes to pathogenesis.
In this section, we first describe the datasets we use. We then explain how we use network propagation for each sample to generate “propagated mutation” and “propagated differential expression” profiles for each gene. Finally, we describe the features we extract from these propagated mutation and differential expression profiles and how we use those features to develop a model to predict causal genes in cancer.
The input to our method consists of BRCA (breast cancer invasive carcinoma) and GBM (glioblastoma multiforme) data obtained from TCGA [13]. We use two categories of data: somatic mutations obtained from whole-exome sequencing and microarray gene expression data. We also obtain differential expression status for TCGA samples from the COSMIC cancer gene census [2]. We collect this data into a binary mutation matrix M, and a binary differential gene expression matrix D, with samples as rows and genes as columns. We use C(A) to denote the set of column labels of matrix A, so that e.g. C(M) is the set of genes that appear in the TCGA somatic mutation data. Similarly, we define R(A) as the set of row labels of matrix A, corresponding to the distinct samples present in each data set.
The mutation matrices M are defined as
M [ i , j ] = 1 if gene j is mutated in sample i , 0 otherwise (1)
The differential expression matrices D are defined similarly, using differential expression status instead of somatic mutation status for each gene. BRCA data includes somatic mutations in 15189 genes across 974 samples, and differential expression in 18018 for 973 samples. GBM data likewise includes 9507 genes and 591 samples, with differential expression measurements in 17660 genes across the same 591 samples.
We use the HIPPIE protein-protein interaction network [14] (version released 2014-09-05), which contains confidence scores for 160215 interactions over 14680 proteins. All samples present in the gene expression data also appear in the mutation data. 12042 genes are contained in both the mutation and expression data, out of which 9303 are present in the HIPPIE network.
We use the network propagation method described in Vanunu et al. [4]. Given a network G = (V, E, w) with V as the set of proteins, E as the set of their interactions, w(u, v) representing the reliability of an interaction uv ∈ E, and a prior knowledge vector Y: V → [0, 1], we seek to compute a function F(v) ∀v ∈ V that is both smooth over the network and accounts for the prior knowledge about each node. In the context of our problem, the prior knowledge about each node is the mutation or differential expression status of the respective gene in a sample.
As described by Vanunu et al. [4], we use Laplacian normalization to produce the normalized network edge weight w′. Briefly, we construct a |V| × |V| matrix W from the edge weights w, and construct a diagonal matrix Δ with Δ[i, i] = ∑j W[i, j]. The normalized weight matrix is computed as W′ = Δ−1/2 WΔ−1/2. Our W′ is a 14680 × 14680 sparse matrix with each row and column corresponding to a node in the HIPPIE network, and each nonzero entry signifying an interaction between two proteins.
With the normalized weight matrix W′, we use the iterative procedure described by Zhou et al. [15] to compute F. Namely, starting with F(0) = Y, we update F at iteration t as follows:
F ( t ) = α W ′ F ( t - 1 ) + ( 1 - α ) Y (2)
This procedure is repeated iteratively until convergence; namely we stop the iterations when ‖F(t) − F(t−1)‖2 < 10−6.
We use network propagation on a sample-specific basis to compute propagated mutation and differential expression vectors for each sample. Namely, we produce new “propagated” matrices MP and DP, by separately using each row of matrices M and D as the prior knowledge vector Y in Eq 2. This is illustrated in Fig 1.
Given the data matrix A (either M or D) and each protein in the network v ∈ V, we construct the vector Y i ( A ) for sample i as follows:
Y i ( A ) [ v ] = A [ i , v ] if v ∈ C ( A ) ∩ V , 0 otherwise (3)
That is, the prior knowledge about a protein is 1 if and only if the protein is part of the HIPPIE network and the corresponding gene is mutated in sample i or differentially expressed in it. For each sample i ∈ R(A), we denote the prior information vectors by Y i ( M ) and Y i ( D ). Subsequently, using each of these prior information vectors, we use the iterative procedure described above to compute propagated mutation and expression vectors, denoted respectively as F i ( M ) and F i ( D ) for sample i.
Next, we collect each propagated vector F i ( A ) into the rows of a “propagated” matrix AP, where R(AP) = R(A) and C(AP) = V. Intuitively, the propagated matrices MP and DP contain the per-sample binary vectors of M and D smoothed over the network. In biological terms, each row of these matrices represents the network proximity of each gene product to mutated and differentially expressed genes in that sample. Consequently, as illustrated in Fig 1, the columns of these matrices provide propagated mutation and differential expression profiles for each gene product across all samples, indicating the proximity of the respective gene product to the products of mutated or differentially expressed genes in the respective sample.
We seek to use the propagated mutation and differential gene expression matrices MP and DP (with sample set S = R(MP) = R(DP)) to predict causal genes based on network proximity to mutated and differentially expressed genes in BRCA. To this end, we define several features that express the mean, variance and cross-correlation of the columns of those matrices across the n = |S| samples:
μ M [ g ] = 1 n ∑ i n M [ i , g ]: mutation frequency of gene g across samples. μ M P [ g ] = 1 n ∑ i n M P [ i , g ]: mean of propagated mutation scores across samples. μMP[g] quantifies the mean proximity of gene g to mutated genes across all samples. σ M P 2 [ g ] = Var M P [ · , g ]: variance of propagated mutation scores across samples. σ M P 2 [ g ] quantifies how inconsistently the gene products in the neighborhood of gene g are mutated across different samples. μ D [ g ] = 1 n ∑ i n D [ i , g ]: differential expression frequency across the n samples. μ D P [ g ] = 1 n ∑ i n D P [ i , g ]: mean of propagated differential expression scores across the n samples. μD[g] quantifies the mean proximity of gene g to differentially expressed genes across all samples. σ D P 2 [ g ] = Var D P [ · , g ]: variance of propagated differential expression scores across samples. σ D P 2 [ g ] quantifies how inconsistently the gene products in the neighborhood of gene g are differentially expressed across different samples. ρ[g] = Spearman correlation between MP[·, g] and DP [·, g]. ρ[g] quantifies whether samples that harbor mutations in the neighborhood of gene g also harbor differentially expressed genes in the neighborhood of gene g and vice versa. δ [ g ] = ∑ i n M P [ i , g ] · D P [ i , g ]: dot product between MP[⋅, g] and DP[⋅, g]. δ[g] can be interpreted similarly as ρ[g]. However, unlike correlation, this is a non-normalized measure of the consistency of proximity to mutated and differentially expressed genes. As such, δ[g] includes information about the magnitude of values in columns MP[⋅, g] and DP[⋅, g] as well as the agreement between those columns. χmax[g] and χmean[g]: For a gene g, high χ[g] scores denote a gene that is in close proximity to other genes that are frequently mutated or frequently differentially expressed.
χmax[g] = maxi ∈ S(max{MP[i, g], DP[i, g]}). A high χmax[g] denotes a gene that is close to mutations or differential expression in any patient. χ mean [ g ] = 1 n ∑ i n max { M P [ i , g ] , D P [ i , g ] }. χmean[g] represents the gene’s mean distance to mutations or differential expression across all samples. νmax[g], νmean[g]: A high ν[g] score denotes a gene that is in close proximity to other genes that are frequently mutated and frequently differentially expressed.
νmax[g] = maxi ∈ S(min{MP[i, g], DP[i, g]}). A high νmax[g] denotes a gene that is close to mutations and differential expression in any sample. ν mean [ g ] = 1 n ∑ i n min { M P [ i , g ] , D P [ i , g ] }. νmean[g] quantifies the gene g’s mean distance to mutations and differential expression across all samples. γ[g]: Network centrality of gene g, as quantified using eigenvector centrality. Propagation of mutation and differential expression data across the network may bias results in favor of nodes that are central to the network or have high degree [3]. Our propagation method uses node degrees to normalize edge weights, offering some correction for nodecentrality [4]. However, to explicitly account for node centrality without unfairly penalizing hub nodes, and to gain insights into the effect of network centrality, we include network centrality as a feature in the model.
An example of the νmean feature in a simulated data set is shown in Fig 2. We see that genes which score highly via propagated mutation and differential expression frequency are scored highly with νmean, conversely, genes that are proximal to only mutations or differential expression may be scored highly in each individual data set but need not be scored highly in this combined feature.
The features described above are used as input to a standard logistic regression model to predict the causal status of gene g. To train this model, we use prior knowledge of whether each gene is known to be associated with breast cancer based on the integrated breast cancer pathway (Table A in S1 Text), or in glioblastoma based on the GBM KEGG pathway (Table B in S1 Text). The logistic regression model represents the probability p that a gene is associated with the cancer of interest as
log p 1 - p = β 0 + β 1 x 1 + … + β n x n . (4)
Here, β0 represents the background probability that a gene is related to the disease, each xi represents one of the features described above, and each βi represents the magnitude to which xi influences p. In addition to estimating the magnitude of a feature’s effect on p, logistic regression models also allow for the investigation of whether a feature is statistically significant in the model fit. This framework therefore allows us to examine the relationship between the role of a gene in cancer and its mutational frequency, differential expression frequency, network distance to mutations or differential expression, and the relationship between these distances.
Using the genes labeled based on prior knowledge of the molecular basis of each cancer, we fit this model using the features described above, perform step-down via AIC (Akaike Information Criterion [16]), and use the probabilistic output of the stepped-down model as prediction scores for further analysis. We perform experiments to investigate whether this model can effectively recover cancer-related genes even though they are not frequently mutated or differentially expressed in available samples. We also evaluate the model’s performance on an independently curated set of genes known to be implicated in cancer. Finally, we investigate which features significantly contribute to the model fit, in order to gain insights into the factors that have important roles in pathogenesis.
In this section, we apply the logistic regression model we have trained to predict genes associated with breast cancer and glioblastoma and evaluate its performance and the contribution of the different features to its success. Subsequently, we examine in detail the novel predictions made by our model. We identify several predictions that are supported by the literature and find that our predictions significantly overlap with an independent resource on cancer genes. Finally, we test the clinical relevance of the predicted genes, identifying several promising candidates with significant predictive power with respect to patient survival.
We evaluate the predictive ability of our model using ROC curves, using the integrated breast cancer pathway from the NCBI BioSystems database [17] and the glioblastoma KEGG pathway [18]. We label a gene as positive if and only if it is contained in the respective pathway, and use these positive/negative labels to evaluate various prediction schemes. Better scoring systems naturally induce a higher area under the ROC curve (AUC).
We first examine the ability of naïve scoring methods in recovering known BRCA and GBM genes. Namely, we investigate how each of mutation frequency, differential expression frequency, and the network propagated mutation and differential expression, i.e., respectively the column-wise means of matrices M, DG, MP and DP described in “Consolidation of Mutation and Expression Data”, can predict known BRCA and GBM genes. The results of this analysis are shown in Fig 3 and Tables 1 and 2. We see that both mutation and differential expression frequency are slightly informative (AUC 0.581 and 0.625, respectively) in choosing genes that are part of the integrated BRCA pathway. In other words, frequency of mutation or differential expression in TCGA breast cancer samples provides some information on whether a gene is involved in the BRCA pathway, but this information is quite modest.
We see that the propagated signals (with propagation parameter α = 0.8) show much more discriminative power: the mutational AUC increases to 0.757 after network propagation, and likewise the differential expression AUC increases to 0.781. We see similar gains in predictive power in GBM: raw mutational and differential expression AUC are informative (AUC 0.679 and 0.511, respectively), and the application of network propagation to these signals boots the AUC values to 0.854 and 0.782.
Though the increase in predictive power through network propagation is considerable, we seek to improve the AUC values further through a more sophisticated integration of the propagated mutation and differential expression signals. For this purpose, we evaluate the regression model described in subsection “Consolidation of Mutation and Expression Data.”
We first fit the logistic regression model described in the aforementioned section to the full data sets, and perform a step-down procedure to remove features that do not significantly contribute to the model fit. We use the standard AIC (Akaike information criterion) measure [16] to determine whether a model term should be preserved. At each iteration of the step-down procedure, the AIC is computed for the full model and for reduced models with each single term removed. The term whose removal most improves AIC is removed from the model. The step-down procedure terminates when no term removal improves AIC. Fig 4 shows ROC curves resulting from this analysis; Fig 4a and 4c respectively show performance in recovering genes in the BRCA and GBM pathways. Fig 4b shows the accuracy in predicting genes’ membership in the COSMIC database using the BRCA model, and likewise Fig 4d shows performance in predicting COSMIC membership using the model trained from GBM data. We see that the stepped-down models improve ROC AUC when compared to the single features shown in Fig 3, and perform well when selecting genes contained in the COSMIC set.
The final model coefficients and P-values for each disease are shown in Tables 3 and 4. For BRCA, we see that νmean and δ are highly significant predictors of a gene’s membership in the integrated BRCA pathway, with a positive coefficient for νmean and a negative coefficient for δ. We also see a large negative coefficient for feature χmean. We interpret this result by noting that for some sample i and gene j, the value max{MP[i, j], DP[i, j]} is high if gene j is close to either mutations or differential expression, and genes that score highly in only one of these measures are likely to simply be frequently mutated or differentially expressed. Conversely, the ν signals measure the degree to which a gene is close to both mutations and differential expression. We indeed see that νmean is significant (P < 2 × 10−16) with positive coefficient 91.7. We see similar trends in GBM: again νmean is the most significant individual feature, with positive coefficient, and δ is also significant with a negative coefficient. Unlike BRCA, in GBM the χ features which select for proximity to mutations or differential expression are not preserved after AIC step-down. It is also notable that δ is preserved in both diseases but ρ is not. This result is not entirely surprising since ρ only represents agreement between propagated differential expression and mutation signals, and δ also quantifies a gene’s total proximity to mutations and differential expression.
We also evaluate the predictive power added by our combined features in comparison to models fit with purely mutational and differential expression data. These results are shown in Fig 5. These results show ROC AUC values for six models in each disease: one fit with all available mutational features, one fit with all available differential expression features, the full model with all features, and stepped-down versions of the three aforementioend models. We see that in both BRCA (Fig 5a) and GBM (Fig 5b), the combined models improve on performance of those fit with only mutational or differential expression features.
Additionally, we evaluate the distribution and univariate predictive power of each individual feature included in the predictive models shown above. Fig 6 shows the AUC values for each feature defined in “Consolidation of Mutation and Expression Data” in comparison with the AUC value of the fitted model that combines individual features. Fig 6a shows the AUC values for recovering BRCA genes; Fig 6b shows GBM. In both cases we see that νmean is the most informative individual feature, which favors genes that are close to both mutations and differential expression. In both BRCA and GBM we see that the predictive model improves upon the AUC values of each individual predictor.
We observe that the mean propagated mutation feature (μMP) provides better predictive performance than mutation frequency (μM) for BRCA. However, this feature is dropped from the stepped down model while mutation frequency is preserved. This observation applies to several other features for both BRCA and GBM as well. This observation demonstrates the benefit of using logistic regression, in that features that are themselves significant may be almost colinear and not all of them need to be preserved if there is overlap in the information provided by multiple features. In particular, the specific observation stated above suggests that the smoothed mutational signal in μMP is subsumed by the combined features, whereas mutation frequency provides information in addition to the information provided by other selected features. It is also interesting to note that the coefficient of mutation frequency is negative in the stepped down model. It is likely that this reflects a correction for passenger mutations (mutated genes that are not functionally related to tumorigenesis), since the information provided by driver mutations (mutated genes that play a role in tumorigenesis) is incorporated by another feature (combined propagated mutation and differential expression signals) in the model.
Fig 7 shows the CDFs of each individual feature, with separate curves for genes that are contained in each respective pathway. Fig 7a showsBRCA; Fig 7b shows GBM. These figures indicate significant difference between cancer genes and other genes in terms of the distribution of some individual features, and reveal bimodality in δ (dot product) in GBM and νmax (minimum between MP and DP, maximum across samples) in both diseases.
We also fit models with multiple values of the propagation parameter α, ranging from 0.01 to 0.99. The results are shown in Fig 8, and we see that the performance of stepped-down predictive models does not significantly depend on the propagation parameter α.
In order to evaluate the utility of our method in predicting new causal genes, we investigate the high-scoring genes that are not already known to be implicated with breast cancer and glioblastoma. The cumulative distributions of genes’ prediction scores (outputs of the stepped-down logistic regression models) are shown in Fig 9. We see that the distributions of scores are skewed toward 0, and for demonstration purposes we consider a gene to be high-scoring if its prediction score is ≥ 0.2. The highest-scoring such genes are shown along the horizontal axis of Fig 10; (Fig 10a) shows BRCA and (Fig 10b) shows GBM. Several interesting genes appear; PIK3R1 is known to be implicated in human immunodeficiency [19] and the PI3K kinase has been shown to regulate insulin-induced cell proliferation in the MCF-7 breast cancer cell line [20]. GRB2 interacts with BCAR1 as part of the CIN85 complex [21], and CBL is a known oncogene in myeloid malignancies [22].
Since our goal is the identification of potential “silent players” that cannot be selected by each data set in isolation, we identify genes scored highly (prediction score ≥ 0.2) by the combined model (Tables 3 and 4) that are not scored highly by the models shown in Fig 5. Genes for BRCA are shown in Table 5 and genes for GBM are shown in Table 6. Many of these genes are known to be implicated in diseases, but few have been previously reported as associated with cancer. GATA3 controls differentiation of luminal cells in mammary glands [23]. HRAS mutations have been reported to cause altered glucose metabolism in mammary carcinogenesis [24] and to promote epithelial-mesenchymal transition in mammary epithelial cells [25]. NOTCH1 [26] has previously been associated with head and neck squamous cell carcinoma [27], acute lymphoblastic leukemia [28], and chronic lymphocytic leukemia [29]. SHC1 interacts with the atypical kinase PEAK1, which is involved in a basal breast cancer signaling pathway [30]. Alterations in methylation of ANK1 are common in Alzheimer’s disease [31, 32]. Overexpression of ERBB2 (also known as HER2) has been shown in several cancers, including non-small cell lung [33] and endometrial cancers [34]. Mutations in the tyrosine phosphatase PTPN11 have been shown to cause a predisposition for leukemia and some solid tumors [35].
As an independent evaluation of our method, we also examine our scoring system’s ability to select genes that are included in the COSMIC cancer gene census [2]. As with our original set of BRCA interesting genes, we treat membership in the COSMIC data as a positive label for a gene, and evaluate our ability to rank these genes higher than others. Fig 4b and 4d show ROC curves for this gene selection using the models shown in Tables 3 and 4, with AUC values of 0.7833 for BRCA and 0.7701 for GBM. We evaluate the statistical significance of selection of genes in the COSMIC database among those not contained in the respective pathways for BRCA and GBM using hypergeometric tests. In BRCA, 321 genes remain in the COSMIC set after removing those that are included in the integrated BRCA pathway. 8 of the 36 genes with prediction scores ≥ 0.2 overlap with the COSMIC dataset; choosing at least 8 of 321 in 36 trials from the remaining 14562 genes yields P = 5.09 × 10−8. In GBM, 250 genes remain in the COSMIC set after removing those that are included in the respective KEGG pathway. 10 of the 40 genes with prediction scores ≥ 0.2 overlap with the COSMIC dataset; choosing at least 10 of 250 in 40 trials from the remaining 14562 genes yields P = 2.06 × 10−9.
We also examine our method’s ability to recover genes for which mutation or differential expression status is predictive of patient outcome (survival). While the main objective of this study is not to identify markers for predicting patient outcome, these results are presented as an additional validation of the silent players we identify. As such, for both BRCA and GBM, we identify the 25 top-scoring genes that are not contained in the respective pathway, and use the mutational and differential expression status of these genes to repeatedly separate the sample set into two groups. We then use the logrank test to estimate the significance of the difference in survival between those groups; P-values are shown in Fig 11. BRCA samples are shown in Fig 11a, and we see nominal statistical significance from somatic mutations in FLNB and SHC1. FLNB is involved in vascular repair and has not been shown to be associated with cancer, but SHC1 interacts with a kinase signaling pathway that has been implicated in breast cancer [36, 37]. Differential expression status in GRB2, FYN, and HTT also show utility in predicting differences in survival between groups. In GBM, we see that differential expression status of ESR2 is also nominally significant in stratifying patient survival.
Molecular data is a gold-mine for studying human disease, but current methods do not seem to exploit its full potential due to computational problems and lack of statistical power to examine all genomic markers or combinations of those. Network-based analyses provide an appealing bypass as they greatly narrow the search space. Here we have shown the power of network propagation in exploiting weak signals, from either sequence or expression studies, to predict disease causing genes. An application of our approach to breast cancer and GBM data revealed novel genes with literature support and significant association to disease outcome.
Our preliminary results can be extended in several ways. While our analysis focused on breast cancer, the methodology is general and could be applied to any multi-factorial disease for which there are available gene expression and/or sequence data. Furthermore, the method is extensible to other types of omics data such as protein expression and DNA methylation. Finally, it is interesting to study how the method can benefit from prior knowledge on disease causing genes, potentially better guiding the propagation process.
|
10.1371/journal.pcbi.1004662 | Model-Based Evaluation of Spontaneous Tumor Regression in Pilocytic Astrocytoma | Pilocytic astrocytoma (PA) is the most common brain tumor in children. This tumor is usually benign and has a good prognosis. Total resection is the treatment of choice and will cure the majority of patients. However, often only partial resection is possible due to the location of the tumor. In that case, spontaneous regression, regrowth, or progression to a more aggressive form have been observed. The dependency between the residual tumor size and spontaneous regression is not understood yet. Therefore, the prognosis is largely unpredictable and there is controversy regarding the management of patients for whom complete resection cannot be achieved. Strategies span from pure observation (wait and see) to combinations of surgery, adjuvant chemotherapy, and radiotherapy. Here, we introduce a mathematical model to investigate the growth and progression behavior of PA. In particular, we propose a Markov chain model incorporating cell proliferation and death as well as mutations. Our model analysis shows that the tumor behavior after partial resection is essentially determined by a risk coefficient γ, which can be deduced from epidemiological data about PA. Our results quantitatively predict the regression probability of a partially resected benign PA given the residual tumor size and lead to the hypothesis that this dependency is linear, implying that removing any amount of tumor mass will improve prognosis. This finding stands in contrast to diffuse malignant glioma where an extent of resection threshold has been experimentally shown, below which no benefit for survival is expected. These results have important implications for future therapeutic studies in PA that should include residual tumor volume as a prognostic factor.
| The most common brain tumor in children and young adults is pilocytic astrocytoma (PA). This tumor is usually benign and often follows an indolent course. The treatment of choice is resection and the prognosis is very favorable if total excision can be achieved. However, due to the location of the tumor, only partial resection is possible in many cases. Partially resected PA could spontaneously regress, regrow or even progress to a more aggressive type of PA. We develop a mathematical model which describes the growth, progression and regression of PA. We are able to quantitatively predict the chance for regression in dependency of the remaining tumor size. This prediction has the potential to provide decision support to clinicians after partial resection of benign PA. Furthermore, our results imply that there is no resection threshold for PA below which no survival advantage is provided. This finding stands in contrast to malignant brain tumors where such a threshold has been experimentally shown.
| Pilocytic astrocytoma (PA) is the most common pediatric brain tumor and the second most frequent tumor in childhood [1]. Three of four cases are diagnosed up to an age of 20 years with the highest age incidence between 5 and 15 years. PA is usually benign, often follows an indolent course and is mostly slow-growing [2]. In children, PA most frequently occurs in the cerebellum but can develop in the entire neuroaxis. Surgery is the treatment of choice [3]. If total excision is achieved, the prognosis is favorable with more than 90% of patients being cured [4]. However, in many cases tumor location in critical or deep areas (such as brain stem, optic pathway, or hypothalamus) restricts resection options and alternative management options are required [5, 6]. Patients with only partial resection have a worse and highly unpredictable prognosis [4, 5]. Tumors can regrow or even progress to a more aggressive tumor [3, 7–11] but spontaneous tumor regression of PA has also been observed [4, 12–15] and is a common phenomenon. A recent review in [14] estimates a fraction of 14% of all residual cerebellar astrocytoma that regress spontaneously. Other studies claim an even higher portion [16]. While regression of PA after partial resection is reported in many case series [12–16], the influence of the residual tumor size has not been evaluated yet. Moreover, the management for patients in whom complete resection cannot be achieved is still unclear. Due to the chance of regression and the indolent nature of PA, some authors propose a wait and see strategy in order to avoid potential risks induced by further therapies [4, 7, 14]. Other authors favor an aggressive surgical resection in combination with additional treatment strategies, like radiation and chemotherapy to control tumor growth [15, 17, 18].
On the molecular level, it has been shown that activation of the mitogen-activated protein kinase (MAPK) pathway is sufficient to induce the development of PA. This leads to the hypothesis that PA is a single-pathway disease [19, 20]. Furthermore, PA usually harbor only one alteration within the MAPK pathway. The majority of mutations are activating changes in the BRAF gene, the most common is the KIAA1549-BRAF fusion, but also other activating mutations have been described. A more aggressive behavior of PA is observed if additional genetic alterations occur, e.g. loss of tumor suppressor gene CDKN2A [10, 21]. Furthermore, alterations in the PI3K/AKT pathway [22] have been associated with aggressive forms of PA [9]. One proposed mechanism for the often observed slow growth of the tumors is oncogene-induced senescence, which is a mechanism limiting neoplastic growth by inducing cellular senescence. The MAPK activation might initially promote growth as well as induce senescence. Oncogene-induced senescence has also been observed in melanocytic nevi and melanoma [10].
Several mechanisms for tumor regression have been suggested, e.g. immunologic mechanisms, hormonal factors, induction of differentiation or apoptosis [13]. However, the reason why regression in PA occurs is not understood yet [4].
We formulate a mathematical model for growth, progression and regression of PA based on the above described clinical and molecular biological observations. We study the effects of competition between tumor and wild-type cells on the chance for regression. We distinguish two types of PA. Benign cases are classified as PA-I tumors and assumed to be caused by alteration of a single pathway. Tumors in which an additional alteration occurs are categorized as PA- II tumors, representing the more aggressive subset of PA. We introduce a stochastic tumor growth and progression model, namely a Moran model [23] with mutations. We chose a Moran model in this juvenile tumor, since astrocyte proliferation and diversification mainly happen during late embryogenesis and the first three weeks after birth. These processes are largely complete by early postnatal stages, while early and late postnatal development is mainly characterized by maturation processes (like continuing elaboration of astrocyte processes and building of synaptic/vascular connections) [24, 25]. Since PA are usually diagnosed between 5 and 15 years, the normal astrocyte population is not proliferating at this time anymore. Therefore, it is reasonable to assume an approximately homeostatic tissue. In such a tissue, Moran dynamics provide a natural and established framework for modeling competition between tumor and wild-type cells.
In our model, we derive the PA-regression-function describing the probability for regression in dependency of the residual tumor size after partial resection of benign PA. The accumulation of mutations in a tissue has been modeled and investigated by several authors by using a Moran model. Work by Iwasa, Michor, Komarova and Nowak [26, 27] has been extended by Schweinsberg [28] and durrett, Schmidt and Schweinsberg [29] to the case of m mutations. These models analyze tumor growth and progression [30–34] with a focus on theoretical results regarding the waiting time until a cell has accumulated a certain number of mutations. Our approach is motivated by a concrete clinical question which is the regression probability of a benign PA tumor in dependency of the residual tumor size. We modify the model introduced in [29]. In particular, we consider Moran dynamics with two mutations but two absorbing states and investigate the precise relation of the two absorption probabilities which allows the incorporation of epidemiological data to calibrate the model. From the mathematical point of view, the relation of the two absorption probabilities can be connected to the portion of stochastic tunneling events in the model presented in [29].
The behavior of the TGP process depends on its three parameters, the critical tumor size N, the mutation probability from wild-type cells to type-I cells u and the mutation probability from type-I cells to type-II cells v. The parameter regime for the analysis of the TGP model is chosen such that
u ≪ 1 N (1)
and
( N v ) 2 = : γ > 0 . (2)
In the following we explain this choice. We call the parameter γ risk coefficient.
We are interested in the regression probability of a partially resected PA-I tumor in dependency of the remaining tumor size and assume that regression of a residual tumor is achieved if no tumor cells are present anymore. All suggested mechanisms of tumor regression influence the ratio of tumor and wild-type cell birth and death rates. Therefore, we assume that competition between tumor and wild-type cells leads to tumor regression which is incorporated by Moran dynamics with relevant cell number equal to N again, see also Fig 3. Furthermore, we assume that the partial resection reduces the residual number of PA-I cells below the critical tumor size N. Hence, the regression function is defined as the extinction probability of tumor cells, i.e. the probability to reach state 0 when starting the TGP process in some state k with 1 ≤ k ≤ N − 1. For v = 0, our TGP process simplifies to a neutral two-type Moran process in which the extinction probability is an established result and equals 1 - k N [32]. Here, we derive this extinction probability for our TGP process with three cell types. For the mathematical analysis, it is convenient to express this function in terms of ρ = k N. The fraction ρ describes the ratio between the residual number of PA-I cells after partial resection k, 1 ≤ k ≤ N − 1, and the critical tumor size N.
Formally, these considerations lead to the regression function β γ N ( ρ ) defined as
β γ N ( ρ ) : = I P ( X t = 0 for some t ≥ 0 | X 0 = N ρ ) , ρ ∈ [ 0 , 1 ] . (4)
Fig 3 provides a graphical representation of regression in the TGP model.
A diffusion approximation of (Xt)t ≥ 0 leads to the Wright-Fisher diffusion process that can be utilized to approximate the term of Eq (4). This approach was introduced in [29] and leads finally to a series representation as approximation of β γ N ( ρ ). In S1 Fig it is shown that this series can be expressed by Bessel functions I n , n ∈ I N , [35] and that the regression function of the TGP model is given by
β γ ( ρ ) = 1 - ρ I 1 ( 2 γ ( 1 - ρ ) ) I 1 ( 2 γ ) (5)
for 0 ≤ ρ ≤ 1. The graph of βγ is plotted in Fig 4C for different values of the risk coefficient γ.
The regression function Eq (5) depends on the parameters of the TGP model via the risk coefficient γ, see Eq (2). This parameter is estimated such that the clinically observed fraction of PA-I tumors, denoted by p ^, equals the theoretically obtained fraction α(γ) of absorption in state N in the TGP model. Subsequently, the derived risk coefficient is substituted into the regression function given by Eq (5) in order to obtain the specific PA-regression-function. Fig 4 summarizes the overall strategy of this approach.
We estimate the clinically observed fraction of PA-I tumors on the basis of data reported in [10]. The authors analyzed 66 PAs with respect to their genetic profile and classified 57 cases as benign PA-I tumors and 9 cases as more aggressive PA-II tumors. This leads to
p ^ = 57 66 = 0 . 8636 .
In the TGP model, this clinically observed fraction corresponds to the absorption probability in state N, given by Eq (3). Therefore, we set
α ( γ ^ ) = p ^ = 0 . 8636 .
This equation allows to calculate the risk coefficient γ ^ which yields
γ ^ = 0 . 152 .
Substituting γ ^ = 0 . 152 into the regression function given by Eq (5) allows to derive the PA-regression-function given by
β 0 . 152 ( ρ ) = 2 . 3795 1 - ρ I 1 ( 0 . 7797 1 - ρ ) , 0 ≤ ρ ≤ 1 . (7)
A plot of this function is provided in Fig 4C. This figure shows that the regression function is very robust to small alterations with respect to p ^.
Note that the actual risk coefficient may be smaller than the estimated value γ ^ = 0 . 152 due to the following considerations. The parameter N in our model represents a critical tumor size above which tumor regression cannot be expected anymore. However, the number of mutated cells in a diagnosed PA-I tumor may be larger than N because tumors could grow beyond this critical size without symptoms or due to a diagnostic gap between first symptoms and diagnosis. Therefore, a PA-I tumor can consist of more than N type-I cells and should have been more susceptible for progression to PA-II than accounted for in our TGP model. Hence, the risk of progression in our TGP model and therefore γ ^ might be overestimated. However, this would not change the linear dependency between residual tumor size and regression probability which is discussed in the following.
We can show that the PA-regression-function Eq (7) is approximately linear by utilizing a Taylor expansion using Eq (6). Substituting the estimated risk coefficient of the PA-regression-function γ ^ = 0 . 152 into Eq (6) leads to
T 1 ( ρ ) = 0 . 9817 - ρ , ρ ∈ [ 0 , 1 ] . (8)
This is a very good approximation since the remainder term can be estimated by
| R 1 ( ρ ) | ≤ γ 8 = 0 . 152 8 = 0 . 0185 (9)
for ρ ∈ [0, 1]. Hence, the deviation of the PA-regression-function from the linear function T1(ρ) is very small. Moreover, if the risk coefficient was overestimated, an even smaller deviation would be observed as Eq (9) implies.
In order to provide a quantitative prediction of the regression probability given the absolute residual tumor size, we estimate the critical tumor size N in our model. Since the total cell number corresponds to the the tumor volume, we can interpret N also as minimum absolute tumor volume above which tumor regression cannot be expected anymore. The existence of this critical tumor size and its estimate of a cell number corresponding to a volume of 9 cm3 is justified in the following way. First, an extensive literature research indicated that tumor regression for residual cerebellar PA over 9 cm3 has not been reported yet, see S1 Table. Second, the prediction for patients with 78 cerebellar astrocytoma, including 62 PAs, has been investigated in [15]. Fig. 6 in [15] implies that the theoretical proportion of progression-free patients based on a Cox regression analysis with a residual tumor of 9 cm3 is estimated to be zero in the long-term. Finally, in [18], the role of the extent of resection in the long-term outcome of low-grade gliomas is investigated including 93 PAs. It is stated that “‘the predicted outcome for patients is negatively influenced by even residual tumor volumes on the order of 10 cm3”’.
Incorporating the estimation for the critical tumor size of 9 cm3 into the PA-regression-function Eq (7) allows to quantify our predictions, indicating that any volume reduction of one cm3 below the critical size will add 10% to the chance for regression (see also Fig 5 and Table 1).
In malignant brain tumors it has been shown that there is an EOR threshold below which no survival advantage is provided, e.g. in glioblastoma this threshold is 78% [37]. The existence of different tumor zones which basically reflect tumor heterogeneity is one proposed reason for such a threshold in malignant brain tumors [38]. In contrast, our results suggest the non-existence of such a threshold in PA. This is an immediate consequence of the linear dependency between residual tumor size and regression probability. If the residual tumor is smaller than the critical tumor size N, which marks the volume for which regression cannot be expected anymore, any reduction of the tumor volume will contribute to the regression probability. Importantly, this behavior stands in contrast to a non-linear dependency which would have been obtained in our model for a higher estimated risk coefficient γ, see Fig 5.
In order to gain insights into the regression behavior after partial resection of benign PA, we introduced a stochastic TGP model based on recent molecular findings, functional, and clinical data. We derived a regression function that depends on the risk coefficient γ and quantifies the probability of regression in dependency of the residual tumor size. By incorporating epidemiological data on the clinically observed fractions of PA-I and PA-II cases, we estimated γ and derived the specific PA-regression-function, given by Eq (7). The estimated PA-regression-function implies an approximately linear dependency between the residual critical tumor fraction and the regression probability as illustrated in Fig 4C. This linear dependency is supported by a Taylor approximation and an estimation of the remainder term, given by Eqs (8) and (9), respectively. Furthermore, we quantitatively predicted the chance for tumor regression for benign PA by estimating the critical tumor size N, see Table 1.
Our TGP model incorporates assumptions based on clinical observations. It is observed in the clinics that PA-I tumors grow slowly, arrest in growth, or even regress. Hence, type-I cells in our model proliferate without fitness advantage. Furthermore, we assume that the first type-II cell that occurs leads to an aggressive form of PA, corresponding to malignant progression in PA. Alternatively, one could assign a success probability s to an emerging type-II cell, which represents the probability that a single type-II cell leads to a PA-II tumor. However, it has been shown in [36] that this is equivalent to considering an analog process with type-II mutation probability sv instead of v. This alternative process would lead to the same estimated risk coefficient γ ^. Therefore, the estimated PA-regression-function would not change since this function is determined only by γ ^. Further, we use asymptotic results for N → ∞ in order to calculate the theoretical portion of PA-I and PA-II cases in the TGP model. This is justified by the fact that a tumor consists of billions of cells. Simulation results given in S2 Table support these asymptotic results. They show that excellent accordance with formulas for finite N is reached even for small values of N. Moreover, we could show that the model is robust against small changes in the proportion of PA-I versus PA-II tumors as shown in Fig 4C. This robustness is an important property of the model since the proportion of PA-I can vary between different studies, especially since the sample size is often very small [39, 40].
To our knowledge, the proposed TGP model is the first theoretical attempt to predict the regression behavior of PA. In particular, we analyzed PA regression based on the population dynamics of tumor and wild-type cells. The ratio of tumor cell birth and death rates is influenced by immunologic mechanisms, hormonal factors, induction of differentiation, or apoptosis, which could all contribute to tumor regression [13]. Since PA-I tumors grow very slowly, we assumed identical birth and death rates of type-I cells in our model.
Our findings have clinically relevant implications. There is still controversy about the best treatment strategy for PA. Since PA is a slowly growing tumor and might even spontaneously regress, a wait and see strategy is an option besides more aggressive treatment strategies like radiation and chemotherapy. The decision for a more radical therapy would depend on the risk for recurrence (or even progression) and the chance of regression. However, long-term follow-up data about the probability of regression or progression after partial resection of PA is restricted and only retrospective studies with small case numbers are available [39–42]. The linear dependency between residual tumor size and regression probability in our model implies that every resected percentage point of a PA-I tumor contributes equally to the regression probability. Hence, there is no EOR threshold, but any small reduction in tumor mass provides an improvement in prognosis by increasing the probability for tumor regression. This prediction suggests a fundamentally different treatment strategy for PA compared to glioblastoma for which such a threshold has been determined [37]. Therefore, our results indicate that resection of a tumor should be aimed at even if a complete resection may not be possible. This is supported by studies showing that in patients with PA outcome depends on the extent of resection, although these studies only differentiate between biopsy, partial, subtotal, and gross/total resection and do not measure tumor volumes [3, 15, 16]. Moreover, if complete resection cannot be achieved, our results predict that the outcome linearly depends on the residual tumor volume. If there is a reasonable chance for regression of the residual tumor, it might be less justified to accept side effects by further therapies like radiation. This is an important result since the role of additional radiation therapy in treating children with tumors is highly controversial [8]. Unfortunately, as far as we know, there are no clinical studies on treatment of PA that take into account the influence of the residual tumor volume on patient outcome. We suggest that the residual tumor volume is an important prognostic marker and that a lack of sufficient volumetric data could be a reason for different results in clinical studies on additional treatment in PA.
The results of this work should be further supported by future clinical studies that include volumetric data, which will improve the quantitative prediction of our model and form a statistical basis for clinical decision rules.
|
10.1371/journal.pcbi.1003753 | Cost-Effective Control of Plant Disease When Epidemiological Knowledge Is Incomplete: Modelling Bahia Bark Scaling of Citrus | A spatially-explicit, stochastic model is developed for Bahia bark scaling, a threat to citrus production in north-eastern Brazil, and is used to assess epidemiological principles underlying the cost-effectiveness of disease control strategies. The model is fitted via Markov chain Monte Carlo with data augmentation to snapshots of disease spread derived from a previously-reported multi-year experiment. Goodness-of-fit tests strongly supported the fit of the model, even though the detailed etiology of the disease is unknown and was not explicitly included in the model. Key epidemiological parameters including the infection rate, incubation period and scale of dispersal are estimated from the spread data. This allows us to scale-up the experimental results to predict the effect of the level of initial inoculum on disease progression in a typically-sized citrus grove. The efficacies of two cultural control measures are assessed: altering the spacing of host plants, and roguing symptomatic trees. Reducing planting density can slow disease spread significantly if the distance between hosts is sufficiently large. However, low density groves have fewer plants per hectare. The optimum density of productive plants is therefore recovered at an intermediate host spacing. Roguing, even when detection of symptomatic plants is imperfect, can lead to very effective control. However, scouting for disease symptoms incurs a cost. We use the model to balance the cost of scouting against the number of plants lost to disease, and show how to determine a roguing schedule that optimises profit. The trade-offs underlying the two optima we identify—the optimal host spacing and the optimal roguing schedule—are applicable to many pathosystems. Our work demonstrates how a carefully parameterised mathematical model can be used to find these optima. It also illustrates how mathematical models can be used in even this most challenging of situations in which the underlying epidemiology is ill-understood.
| We consider how mathematical models can be used to inform the control of plant disease, even when the identity and biology of the pathogen are not well understood. This is often the case: control of emerging epidemics is most likely to have a significant effect when epidemics remain small, but little may then be known. We analyse data from an experimental plot concerning spread of Bahia bark scaling of citrus, an economically-important disease in north-eastern Brazil, by fitting a mathematical model, which also accounts for uncertainty, to disease spread. Our model captures the epidemiological features of the disease, revealing that transmission is localised and that disease spreads relatively slowly. We use the model to investigate fundamental trade-offs underlying cultural disease control at scales relevant to citrus production. We show how optimal planting densities can be defined, which balance slower spread of disease against the profit that would be lost by growing fewer plants. We also show how the cost of looking for and removing symptomatically diseased plants can be balanced against the reduced disease it leads to. Ours is the first study to consider how a parameterised mathematical model can be used to design optimised cultural controls of plant disease.
| Mathematical models of plant disease can be used to screen and assess control strategies [1]–[10]. Although work on plants is not subject to the ethical concerns that hamper experimentation targeting pathogens of animal or human hosts, mathematical modelling nevertheless becomes particularly compelling for plant diseases when logistic constraints mean that experimentation would be costly or difficult. This situation is exemplified by diseases caused by pathogens with epidemiology necessitating long experiments to yield useful data [11]–[13], pathogens causing symptoms that are difficult to detect [14], [15], pathogens with epidemiology that is ill-understood [16], [17], and/or pathogens that would require experimental trials in the vicinity of susceptible commercial growing operations [18], [19].
Here we develop a model of Bahia bark scaling of citrus (BBSC) on grapefruit, a pathosystem subject to each of these logistical constraints. BSSC has been endemic to north-eastern Brazil since the 1960s [20], but its etiology remains unknown [21]. We use Markov chain Monte Carlo with data augmentation [22] to fit a spatially-explicit, stochastic, epidemiological model to a data-set charting the spread of BBSC through a small experimental grove. We go on to alter the host topology and parameters in this model to use it to assess the efficiency and cost-effectiveness of control at the scale of a typical grove as used in citrus production in Brazil. As little is known of the putative BBSC pathogen, and even less about any potential vector, it is difficult to reliably estimate the efficacy of any chemical [23] or biological [24] control. We therefore concentrate on cultural strategies [25], and focus on the effectiveness of reducing the density of planting [26] and of roguing [10] (i.e. searching for and removing infected plants).
The spread of plant pathogens is typically localised, and so it is intuitive that the progression of disease through a host population will be affected by planting density. Direct as well as indirect effects of host density on disease incidence have been proposed [26], and lower host densities are almost always associated with lower levels of disease [27]. Indeed the “dilution effect” caused by increased distances between pairs of hosts has been suggested to underlie the success of crop mixtures [28] and intercropping [29], although other more complex mechanisms are thought to be involved in both cases [30]–[33]. However, there are very few models specfically targeting the effect of host density on disease spread. Despite work concentrating on how percolation thresholds can be related to the distance between pairs of nearest neighbours [34], [35], tests of that theory have largely been restricted to small-scale model systems [36], and application to real pathosystems remains in its infancy [37], [38]. Percolation is also only strictly relevant to systems where spread is restricted to nearest neighbour transmission, although this does map well to the soil-borne pathogens that are the focus of that work. Other work has concentrated on how host density affects invasion thresholds [39], [40], but does not provide a clear prescription for how to optimise host densities when disease is able to invade. While there have also been studies showing how the landscape-scale dynamics of disease are conditioned on the configuration and availability of patches of suitable habitat [41], or fields planted with susceptible varieties [42], that work offers little at scales relevant to individual farmers or growers.
Roguing is commonly used for systemic diseases of high-value or perennial crops [43], particularly when labour is cheap compared with the cost of chemicals [44], or for pathogens which cannot be effectively controlled by chemical means [18], [19]. Viral pathogens for which roguing is practised include cassava mosaic [45], bunchy top of banana [46], cocao swollen shoot [47], [48], citrus tristeza [49], plum pox [50] and sweet potato chlorotic stunt [51], although roguing is also used for bacterial pathogens (e.g. almond leaf scorch, caused by Xylella fastidiosa [52]), and for fungal diseases (e.g. lettuce drop, caused by Sclerotinia minor [53]). The only constraint is that pathogens must cause symptoms that can be detected, either by visual inspection or by diagnostic testing. Roguing has been included in non-spatial mathematical models for a number of years [1], [2], [54], [55], and more recent work has embedded control by roguing in spatial models of pathogen spread [10], [56]–[58], although realistic parameterisation of pathogen dispersal is less common [6]–[8]. Typically these later models have also considered culling, in which all hosts within a particular distance of a symptomatic focal host are removed at the time of control. Some of these models [57], [58] have explicitly included economics, although the focus has been the cost of treatment (i.e. the cost of removal of diseased host plants). For perennial hosts that are not replanted, however, the cost of detection may, in fact, be more important, since an individual host can be removed at most once, but may be examined for symptoms any number of times. The only model to include detection costs used optimal control theory to show rigorously how to balance the costs of detection and control within a fixed budget [59], but the mathematical complexity of this procedure necessarily restricted attention to a non-spatial, deterministic model. There are no examples of a model-based approach that optimises the economic aspects of roguing including the cost of detection via a model parameterised to spread data.
We have taken advantage of the availability of experimental data for model fitting to frame our analyses specifically in terms of the dynamics and control of BBSC. However, the controls we examine are widely used, and the techniques we use in our modelling and fitting are applicable to a large number of pathosystems. We therefore prefer to think of the BBSC system as a data-driven case study that provides an opportunity to address the following more general questions.
BBSC affects most citrus species and varieties, but is especially severe on grapefruits [60]. Symptoms appear similar to Citrus Psorosis A, and include darkening and thickening of the bark leading to scaling lesions on the trunk and branches, dieback of young branches, and significant gum extrusion. However leaf symptoms on inoculated indicator plants, together with histopathological and molecular studies, indicate BBSC is a distinct disease. The study of Laranjeira et al. [20] resulted in the only published data focusing on BBSC spread (see Text S1). It demonstrated that the disease is polyetic and naturally transmitted. The speed of disease spread and the pattern of dispersal appear consistent with an insect vector of limited dispersion ability. However the identity of this putative vector is unclear, as is the identity of the pathogen itself [21].
BBSC currently remains restricted to two states in the Brazilian north east, Bahia and Sergipe [21], [61]. Since dispersal is thought to be localised, the principal risk of an epidemic arising elsewhere in Brazil is likely to occur by transplanting infectious plants. Introduction of BBSC by inadvertent transplantation is certainly possible: Santos et al. [62] have described BBSC symptoms in plants used for budwood in Bahia, Brazil. There is therefore a need to understand whether and how a spatially-isolated epidemic could be effectively controlled. This must be done even though our biological understanding of the epidemiology of BBSC remains limited.
We use a spatially-explicit, stochastic, compartmental SEIR model [4] to represent BBSC dynamics at the scale of a grapefruit grove. Individual host plants are categorised by disease status: (S)usceptible hosts are uninfected; (E)xposed hosts are latently infected, and so are neither symptomatic nor infectious; (I)nfected hosts are both infectious and symptomatic; and (R)emoved hosts have been removed by control (Figure 1(a)). The E to I transition occurs at rate , corresponding to average latent period (see also Table S1). Since infectious hosts are always symptomatic in our model, the average incubation period is also . Infected hosts do not appear to suffer increased mortality due to BBSC infection [20], and so in the absence of control the rate of transition from the I to R compartment is fixed at zero. However, if control by roguing is included in the model, the removal rate is set by how frequently and efficiently infected plants are detected and removed, with rounds of detection and removal according to a schedule that is fixed in advance. Since we work over a twenty year timescale, similar to the typical productive lifespan of an individual citrus host [63], [64], we do not attempt to model natural death. We also do not consider replanting of any plants removed by roguing, since this is not common in the Brazilian citrus industry, perhaps due to growers' perception that replanting removed hosts would lead to a heterogeneous grove that would be more difficult to cultivate [63].
The rate of infection of susceptible hosts depends on the disease status of all other hosts in the system. In particular, if host is susceptible at time , then it becomes latently infected (i.e. transitions to the E compartment) at rate , where(1)
The summation runs over the set of all (I)nfectious hosts, , and denotes the distance from infectious host to susceptible host . The parameter sets the rate of infection. Spatial dependency in spread is controlled by the dispersal kernel, . Here, noting the constant velocity of the epidemic front in the experimental grove [65], and following exploratory analyses that strongly supported the choice, we used the exponential kernel, normalized in two dimensions(2)where is the area of susceptible tissue presented by an individual host. The factor of is included since, strictly-speaking, the underlying normalised dispersal kernel is a probability density function, with dimensions of inverse area, meaning the observed rate of infection must be calculated by integration over the area of the recipient plant. Assuming the kernel is constant over this area reduces the integration to a simple multiplication, and so leads to Equation (2) above [66], [67]. Since the infection rate then depends entirely on the product in Equation (1), we rescale the area of a single host into the infection rate, setting
(3)(4)(5)
Our model fitting then estimates the value of directly, since it is this product which sets the observed rate of spread of disease in our model. The mean distance of dispersal is [68]. Since we model a grove that initially contains immature plants, and guided by the temporal pattern of disease spread in the experimental grove, we include a delay, , to allow young plants to reach epidemiological maturity [20], [21]. This delay prevents the disease from spreading for the first units of time, but otherwise does not affect the dynamics of infection in the model. Including this delay is therefore equivalent to considering two age classes of tree in the model: juveniles of age less than , that cannot become infected or transmit infection, and adult trees of age greater than , that are epidemiologically competent. The inclusion of this extra parameter was strongly supported by our model fitting (see Results and Text S2).
Data from the experiment of Laranjeira et al. [20] were used to fit the model. These data consist of successive snapshots over time, tracking the disease status of each host in a small experimental grove. This grove contained 240 grapefruit (Citrus paradisi Macf.) plants in 16 rows of 15. Immature plants were planted at regular 2 m2 m spacing at the start of the experiment, at a closest distance of 5 m from twenty-five BBSC symptomatic adult grapefruit plants arranged in a rectangular lattice at separation 6 m4 m (see Figure 2a). Disease progress was assessed by detailed visual inspection at three monthly intervals for the first five years of the experiment, followed by additional more irregular surveys for two years thereafter. The data consist of the visible disease status of each grapefruit plant in the experimental grove at each survey time; i.e. a series of maps showing which hosts were susceptible and which were (visibly) infected on each survey. However, since surveys were separated by at least three months, and because the transition is not visible, exact transition times of individual plants are unknown. We therefore fitted the model in Equations 4 and 5 using Markov chain Monte Carlo with data augmentation to estimate the model parameters of interest (i.e. and ) [22], [69], treating the unobserved times as additional nuisance parameters to be estimated. Posterior distributions for the epidemiological parameters could then be obtained post hoc by marginalization. Further details of the fitting methodology and expressions for likelihood functions are given in the Text S2.
One thousand independent simulations of the model were performed to assess how BBSC would spread in a typical grove (i.e. 1680 plants at 6 m4 m spacing) when disease control is not attempted. We (arbitrarily) took , and simulated progression over 20 years, a notional productive lifespan of a citrus grove [63], [64]. Parameter values used in each simulation were drawn randomly from the joint posterior distribution for and as obtained in estimation. The model was simulated using the Gillespie algorithm [70] (see Text S3 for details).
The number of plants in the central grove that are susceptible at time is , and the number of plants in the exposed compartment is . We define the number of asymptomatic plants at time as . This corresponds to the number of productive (i.e. fruit-bearing) plants at any time. We consider the final number of asymptomatic plants after twenty years, , as a simple composite measure of disease spread, corresponding to the productive trees that remain after accounting for the final size of the epidemic over a 20 year period, and we examined the response of this to values of ranging from 0.06% to 2%, i.e. from 1 to 34 initially exposed trees within the central grove. We again used 1000 independent simulations for each initial condition we considered, as we did for each set of parameters in each of the scenarios described below.
To test the effect of host density on disease dynamics, the within-row and between-row spacing of trees were altered, while constraining the total number of trees in the central grove to remain fixed at 1680. The ratio of horizontal to vertical separation was held fixed at throughout. Again we focused on the final number of asymptomatic plants () in a grove with , and considered planting densities from 50 to 500 plants per hectare.
While this approach illustrates the effect of inter-host distance on disease spread, it is an oversimplification, since fixing the number of trees at different planting densities corresponds to groves with different areas. To examine the trade-off between disease prevention and productivity we therefore considered the density of asymptomatic trees at years in the central grove as a function of host density, again for .
We modelled a programme of scouting for disease symptoms and roguing detected infected plants. This was included in the model by simulating the examination of every surviving plant in the central grove every units of time, and independently detecting symptomatic (i.e. class I) plants with probability . Any detected plants were immediately removed. We considered roguing intervals, , between 7 days and 2 years, and took the probability of detection on a round of scouting to be , supported by data from Belasque et al. [71]. Again we assessed the efficacy of control by examining the value of , the number of productive trees in the central grove after twenty years.
We considered the responses of to the roguing interval () with fixed , and to with fixed months. We also considered the response of the median value of and of the probability of eradicating the pathogen within twenty years as both and were varied simultaneously. Since the default detection probability is an estimate, we also considered the sensitivity of our results to this choice, by considering the response of the median value of as and were simultaneously varied.
Although the disease initially spreads rather slowly, almost all plants within a typical grove are expected to become symptomatic within 20 years when the initial level of infection (Figure 1(b)). On average 50% of plants become symptomatic within approximately the first 10 years. Spatial snapshots from an arbitrarily chosen run of the model (Figure 1(c)) indicate that disease spread is very localised, with infection apparently being transmitted largely (but not exclusively) between neighbouring pairs of plants. It also appears to be rather difficult for the pathogen to escape the central grove and to infect plants in the surrounding groves, although this does happen occasionally. Snapshots from other runs indicate that these aspects of BBSC dynamics are general for ; spread is localised with separate foci of infection that grow and coalesce over time, and spread is also largely restricted to the central grove, at least for the first to years. Varying the initial level of infection indicates the final number of productive (i.e. asymptomatic) plants at , , is highly dependent on (Figure 4(a)), at least for low values of . However, since decreases sharply with the amount of inoculum that is initially present, effectively the whole of the central grove becomes infected by for .
The value of depends strongly on the planting density (Figure 4(b)), with low host density leading to very little spread and so high values of (again with ). However at more realistic planting densities the spread is much more devastating. On average only of plants escape (visible) disease by years at the density of the typical grove plants per hectare).
This behaviour leads to a disease-driven trade-off in the number of productive plants per hectare. Low planting density can give excellent disease control, with very high values of , but of course also implies fewer plants per hectare. The optimum density of productive plants is therefore recovered at an intermediate host spacing: for , this was at a planting density of around 200 plants per hectare, with per hectare (Figure 4(c)). This qualitative result is robust to the initial level of infection, and there was an optimum planting density for all values of we considered. However both the optimal planting density, and per hectare at this planting density, decreased as the initial level of infection was increased (Figure 4(d)), although these responses begin to flatten off for .
We also considered the response of the yield (cf. Equation 6) to the planting density. Again for a given level of initial infection, a planting density that leads to an optimum yield per hectare can be defined (Figure 4(e)), although the density that optimises yield when ( plants per hectare) is larger than that required to maximise the value of ( plants per hectare, as described above). The response was also differently shaped, with the yield per hectare remaining at a non-zero value for even very large planting densities (compare 4(c) with 4(e)). This is because even at high densities the epidemic does not infect the entire central grove within the first few years of the epidemic, and so the yield is then non-zero (see also the inset to Figure 4(e), which shows the yield before normalisation of to fixed grove area). However, the response of the optimum planting density required to optimise yield per hectare for different values of the initial level of infection, and the response of the optimum yield per hectare itself at optimum planting density to the initial level of infection both follow a similar pattern to the responses for (compare Figure 4(d) and Figure 4(f)).
Even at relatively high initial levels of infection, , roguing can lead to excellent disease control (Figure 5(a)). At (a level at which every plant within the central grove would become infected without control within 20 years), even the rather long roguing interval would save approximately 20% of plants from visible symptoms at . As is shortened, of course increases. Values of months lead to high levels of disease control (e.g. %), and even gives . This response is comparatively robust to the initial level of infection (Figure 5(b)): although does decrease as is increased (for fixed ), it does so only relatively slowly.
The value of in fact always depends on and in this broad fashion (Figure 5(c)), decreasing as either parameter is increased. For short roguing intervals, however, was relatively irresponsive to , and indeed there was a large set of pairs for which excellent control was achieved. This was despite the more restricted range of pairs of these parameters for which the pathogen was reliably eradicated from both the central and the surrounding groves (Figure 5(d)).
We also examined the response of the median value of to changes in the roguing interval, , and the probabilty of detection, (Figure 5(e)). Unsurprisingly, the impact of the epidemic is increased as is increased or is decreased. In fact the shape of the contours of constant can be explained by a simple calculation. If the other epidemiological parameters are fixed, the efficacy of roguing is set by the effective infectious period of the average host. This is the time for which the host is infectious, i.e. the time between the emergence of infectivity after the latent period has passed and later removal of the host by roguing. If the probability of detection is , then the number of surveys required to detect a host after the emergence of symptoms upon it is a geometric random variable, with average . A particular symptomatic plant could have become infectious at any time between the final round of surveying when it was asymptomatic/uninfectious and subsequent round by which time it was symptomatic. If we assume the time of the transition between states and in our model is uniformly distributed between surveys (i.e. if we ignore any knock on effect from the slight increase in the rate of infection between rounds of detection that would occur because the number of infected plants increases between surveys), then the average effective infectious period can be approximated by(10)
For the default parameters and , the average infectious period is ; all pairs with this effective infectious period are shown by the black curve in Figure 5(e).
We used Markov chain Monte Carlo with data augmentation to fit a spatially-explicit, stochastic, epidemiological model to the spread of BBSC, and have estimated a number of key epidemiological parameters. Dispersal was exponential, with median approximately 5 m (similar to the distance between neighbouring pairs of plants in a typical citrus grove in Brazil). Laranjeira et al. [20] suggest that the BBSC pathogen may be transmitted by an air-borne vector of limited dispersion ability, and our results are consistent with that possibility. Our estimate of the dispersal scale, together with a careful review of the dispersion ability of arthropods detected in the Bahia region, may help to narrow the set of candidate vectors. Certainly a number of mites and scale insects are known to transmit viral diseases, both in citrus [74] and other perennials [75], and similar species would be an obvious place to begin such a search. Our parameter estimates are also consistent with an association between a bark wounding insect and a splash dispersed fungus.
To obtain an adequate fit to the experimental data we included a delay for plants to reach epidemiological maturity before being able to spread and/or show symptoms of the disease in our model. While it is of course rather difficult to give a mechanistic interpretation of this delay because of the uncertainities surrounding BBSC etiology, it could, for example, correspond to a need for mature tissues for symptom expression, or a bark borer insect vector that only feeds on mature bark. Irrespective of its mechanistic basis, our estimate of the delay is approximately 24 months. Laranjeira et al. [20] took the long delay before disease began to spread in the experiment as indicative of the incubation period for the pathogen that causes BBSC, which we instead estimated to be approximately 6 months. Given the very good statistical support for our model fitting, we contend that our new interpretation of the experimental results is more plausible, especially since a two year incubation period is rather long for a vectored disease.
In a grove at planting density typical of citrus production in Brazil, we predict that BBSC would spread slowly. This was unsurprising given the relatively slow rate of disease spread in the original experiment, in which the density of host plants was approximately six times higher than found in citrus production. Nevertheless, and slow spread notwithstanding, we predict BBSC would easily spread throughout an entire grove within 20 years, even for modest levels of initial infection (). In turn this indicates that careful sanitation of new plantings for BBSC symptoms is important. Despite the official programs to foster propagative plants under screenhouses in Bahia, symptomatic “mother” plants are still found [62], and most nurseries are not kept under screenhouses [76]. This clearly presents a risk, particularly since there is no diagnostic test to identify asymptomatic infected plants. This compelled us to investigate other types of control apart from sanitation. We note that, although high BBSC severity and incidence can be routinely detected in mature commercial groves in Bahia, the incidence of disease is usually quite low at the time of first detection (HP Santos-Filho, personal communication). The particular range we used was therefore intended to account for the full range of values that may occur in practice, given groves at different distances from sources of inoculum and/or with different levels of sanitation before planting. The influence of the initial level of infection on the optima we identify indicates that, for practical implementation, it would be advantageous to perform further experimentation and/or further data-collection to enable to be more precisely quantified.
We therefore used our model to examine the effect of host spacing on disease spread. As the density of hosts was increased, so did the level of disease, which of course was expected [26]. However this is particularly unfortunate given recent trends toward higher planting densities in commercial citrus production in Brazil [77]. We therefore examined the trade-off between host density and productivity in the presence of disease by considering the density of plants that escape infection over a 20 year timescale as the host spacing was altered. We found an optimum planting density, at which the reduction in productivity due to planting fewer hosts per hectare was offset by the reduced losses to disease (cf. Figure 3(c)). Although the exact nature of this optimum depended on the initial level of infection, optimal densities were typically sufficiently low that there would be enough space for an intercrop to be established. This approach is already used in Brazil, where growers sometimes plant passion fruit or pineapple between rows of citrus. However, since the intercrop would undoubtedly have its own effect(s) on pathogen dispersal [32], [33], investigating the epidemiological consequences of intercropping requires more data.
According to our simulation results, roguing, even when detection is imperfect, can control disease successfully (cf. Figure 4). Control can be achieved for relatively long roguing intervals, even for high levels of initial infection. Indeed in our scans showing the effect of roguing interval on control efficacy we used a default value of (rather than as used in assessing the effect of host density) in order to obtain a more meaningful response as the parameters of interest were changed. This good level of control was possible because of the slow rate of BBSC spread and its limited dispersal ability. Control by roguing is also aided by the absence of cryptic infection (i.e. hosts that are able to infect without showing symptoms). This contrasts with a number of other pathogens of citrus, for example Xanthomonas axonopodis, the bacterium that causes citrus canker, for which there is both significant long-range dispersal [78] and cryptic infection [6]. Indeed the recent attempt to eradicate citrus canker from Florida involved removing any host plant within 579.1 m (1900 ft) of a detected symptomatic focal plant, irrespective of apparent disease status [79]. However, the epidemiology of BBSC indicates that a similar approach is not required here, and initial tests of this type of control strategy indicated that it did not noticably outperform simple roguing (data not shown).
Control was possible even though roguing only occurred within the central grove. It did not require the pathogen to be entirely eradicated from the system, and indeed for high values of , the pathogen was eradicated only rarely (cf. Figure 5(d)), presumably because there was at least one escape of the pathogen from the central grove before it was effectively controlled there. This surprisingly high level of control despite an ever-increasing external reservoir reflects the low probability of the pathogen returning to the central grove once it has escaped (cf. Figure 1(c)), and on the occasions it does return, frequent roguing limits its impact. Ultimatately this derives again from the limited disperal ability of the pathogen that causes BBSC. For pathogens capable of faster and/or long-distance dispersal, synchronisation in control is acknowledged to be extremely important, since otherwise the pathogen is able to persist, bulk-up and repeatedly cause devastating reinvasion from uncontrolled areas that act as refugia [10]. Following common practice in the Brazilian citrus industry, removed plants were not replaced in our model, which again facilitated control. Replanting removed trees results in a constant supply of new susceptible hosts to areas with infection, which necessarily makes control more difficult.
The efficacy of roguing was characterised by considering , the average effective infectious period (Equation 10), and this quantity was an excellent predictor of the number of plants that escape disease (cf. Figure 5(e)). Investigating how this result generalises to pathogens that are harder to control would be an interesting extension, particularly because the approximation used in the calculation of is most accurate for pathogens that spread slowly. We note that, although simple, the principle underlying the calculation of has been reported incorrectly in previous studies that used non-spatial, compartmental models. Parameter values given in Table 2 of Jeger et al. [3] (see also Madden et al. [80]) indicate that if roguing is performed monthly then the equivalent removal rate would be . This assumes that symptoms and infectivity are developed immediately after rounds of surveys, and so that the average infectious period is . Given the more accurate estimate of 15 days, the rate of removal for monthly surveys with perfect detection should in fact be .
By introducing a simple measure of the profitability of a grove, we demonstrated the trade-off between the cost of detection and the benefits of control (cf. Figure 6). An optimum roguing frequency can be determined, balancing the increased cost of roguing more frequently against the improved control it leads to, although this optimum is conditioned on the initial level of disease (cf. Figure 6(d)) and the cost of examining a plant for disease symptoms relative to the difference between the sale price of the fruit from a single year's harvest and the yearly cost of cultivating a tree.
For simplicity and ease of presentation, our definition of the cost of control focused exclusively on the cost of detection and did not include the cost of removal. However, because an individual plant would potentially be surveyed many times, but can be removed at most once, we believe this is a reasonable simplification. While our methodology could readily be extended to include more complex economics (e.g. removal costs, cost of initial grove establishment, increased yield from older plants), or to allow for growers potentially ceasing cultivation if the net profit from a particular grove fell below zero despite the yield that would subsequently accrue, the broad result would certainly be robust to these changes. A more interesting extension would be control strategies that change over time. An example of this is a roguing interval that depends on the current (observed) prevalence of infection, and so that could cause surveying to slow down or even stop once the disease was judged to be under control. This differs from the implementation considered here, in which the cost of detection for low levels of initial infection and short roguing intervals may be overstated: any grower who surveyed weekly but did not find disease for a number of years would doubtless reduce the frequency of surveying or even stop entirely. Investigating this type of adaptive strategy, together with the consequential risk of failure that derives from having to predict whether the disease has actually been eradicated or has merely not been found recently, will form the basis of our future work in this area.
A number of previous models have used deterministic mean field representations of cultural control [1], [2], [54], [55], [81]. More recently stochastic, spatially-explicit models have predominated [10], [56]–[58], although typically these models are not fitted to data (a series of studies of the failed eradication of citrus canker in Florida are the exception [6]–[8]). What previous models lack, however, is a treatment of the economic aspects of control, and the trade-offs and optima to which this can lead. While significant progress in examining this type of trade-off has been made using optimal control theory [59], [82], [83], the complexity of the associated mathematics has necessarily reverted attention to deterministic, non-spatial models. Using a spatial, stochastic model parameterised with real data to balance the benefits of effective disease control against its costs is the novel aspect of our work. In addition to the additional insight into BBSC epidemiology obtained by our model fitting, providing a “real world” example showing how a mathematical model can be used to optimise and test both the epidemiological and economic aspects of control strategies for a plant disease is therefore the key contribution of this paper.
|
10.1371/journal.pcbi.1002358 | Metabolic Reconstruction for Metagenomic Data and Its Application to the Human Microbiome | Microbial communities carry out the majority of the biochemical activity on the planet, and they play integral roles in processes including metabolism and immune homeostasis in the human microbiome. Shotgun sequencing of such communities' metagenomes provides information complementary to organismal abundances from taxonomic markers, but the resulting data typically comprise short reads from hundreds of different organisms and are at best challenging to assemble comparably to single-organism genomes. Here, we describe an alternative approach to infer the functional and metabolic potential of a microbial community metagenome. We determined the gene families and pathways present or absent within a community, as well as their relative abundances, directly from short sequence reads. We validated this methodology using a collection of synthetic metagenomes, recovering the presence and abundance both of large pathways and of small functional modules with high accuracy. We subsequently applied this method, HUMAnN, to the microbial communities of 649 metagenomes drawn from seven primary body sites on 102 individuals as part of the Human Microbiome Project (HMP). This provided a means to compare functional diversity and organismal ecology in the human microbiome, and we determined a core of 24 ubiquitously present modules. Core pathways were often implemented by different enzyme families within different body sites, and 168 functional modules and 196 metabolic pathways varied in metagenomic abundance specifically to one or more niches within the microbiome. These included glycosaminoglycan degradation in the gut, as well as phosphate and amino acid transport linked to host phenotype (vaginal pH) in the posterior fornix. An implementation of our methodology is available at http://huttenhower.sph.harvard.edu/humann. This provides a means to accurately and efficiently characterize microbial metabolic pathways and functional modules directly from high-throughput sequencing reads, enabling the determination of community roles in the HMP cohort and in future metagenomic studies.
| The human body is inhabited by trillions of bacteria and other microbes, which have recently been studied in many different habitats (including gut, mouth, skin, and urogenital) by the Human Microbiome Project (HMP). These microbial communities were assayed using high-throughput DNA sequencing, but it can be challenging to determine their biological functions based solely on the resulting short sequences. To reconstruct the metabolic activities of such communities, we have developed HUMAnN, a method to accurately infer community function directly from short DNA reads. The method's accuracy was validated using a collection of synthetic microbial communities. Applying HUMAnN to data from the HMP, we showed that, unlike individual microbial species, many metabolic processes were present among all body habitats. However, the frequencies of these processes varied dramatically, and some were highly enriched within individual habitats to provide niche specialization (e.g. in the gut, which is abundant in food matter but low in oxygen). Other community functions were linked specifically to properties of the human host, such as biochemical processes only present in vaginal habitats with particularly high or low pH. Studying additional environmental or disease-associated communities using HUMAnN will further improve our understanding of how the microbial organisms in a community are linked to the biological processes they carry out.
| Human-associated microbial communities interact directly with their hosts by means of metabolic products and immune modulation, and environmental communities are further responsible for a wide range of biochemical activities [1]. Metagenomic sequencing provides a culture-independent means of studying these diverse microbiota within different ecological niches, including sites in the human body that differ strikingly in microbial composition and subsequent impacts on health [2], [3], [4]. The gut microbiota in particular have been shown to play an important role in host metabolism [5], [6] and immune response [4], and mechanisms of commensal microbial contribution to disease have been established e.g. in the vaginal [7] and skin [8] communities as well. These studies have demonstrated the importance of assaying microbial pathways, metabolism, and individual gene products by means of metagenomic sequencing to determine their roles in community-wide interactions and phenotypes. A functional interpretation of metagenomic sequences is thus key to connecting the metabolic and functional potential of a microbial community with its organismal population structure and with its influence on the surrounding environment or human host.
The functions of human-associated microbes are of particular interest, and the Human Microbiome Project (HMP [9]) has thus performed a comprehensive study of microbial communities from many different body sites in a reference population of disease-free adult subjects. The study's metagenomic data comprise over 3.5 Tbp of shotgun DNA sequences drawn from seven body habitats (including oral, gut, urogenital, nasal, and skin) from over 100 individuals. Using these data, we sought to address questions pertaining specifically to microbiome function: what metabolic and broader biomolecular functions are present within the human microbiome, how do they provide specialization within the microbial niches of distinct body sites, and how do they vary across the human host population? To address these in a high-throughput manner, we have developed a scalable methodology to reconstruct the functional potential of microbial communities from metagenomic sequences, the HMP Unified Metabolic Analysis Network (HUMAnN). To avoid the need for assembly of metagenomic reads, HUMAnN (Figure 1) allows direct profiling of the metabolic potential of microbial communities as represented by orthologous gene family and pathway abundances. The computational methodology incorporates a series of gene- and pathway-level quantification, noise reduction, and smoothing steps in order A) to identify which pathways are present or absent within a metagenomically sequenced community and B) to determine their relative abundances. HUMAnN's predictive accuracy was validated quantitatively using data from four synthetic communities, and it was subsequently used to characterize metabolic function throughout the human microbiome.
Most analytical methods for microbial community assays focus on organismal membership and population structure, i.e. “who's there” in a community [5], [10]. However, functional characterization of community metagenomes is additionally necessary to determine what metabolism and other biological activity may be occurring [11], [12]. This presents a distinct set of challenges, since inter-dependent organisms within a community may share many functional components in addition to playing individually specialized roles. Current metagenomic approaches for characterizing microbial community function include IMG/M [13], MG-RAST [14], and the recently expanded MEGAN tool [15]. Each of these relies on a “best-BLAST-hit” approach, in which individual short reads from a sequenced community (or open reading frames from assembled DNA) are searched against a characterized reference database using translated BLAST. This approach has been used to show the importance of specific community metabolic processes in a range of environmental ecologies, including ocean water and the human gut [16]. Gianoulis et al [17] in particular found that in a collection of 37 ocean communities, metabolic differences correlated specifically with environmental features such as temperature, depth, and salinity. In a particularly large dataset of 124 gut metagenomes, the MetaHIT consortium [11] qualitatively identified metabolic pathways using genes predicted from assembled sequences, which were subsequently suggested to be associated with host phenotypes including obesity [12]. Several additional studies have shown the importance of testing for pathways differentially abundant among communities of interest, e.g. among ocean environments [18] or within the infant gut [19], [20]. However, although each of these results demonstrates the importance of community metabolism and function, no one method has yet been quantitatively evaluated as a means of reconstructing microbial pathway abundances from metagenomic data.
In order to determine the distribution of microbial function within the human microbiome, we thus first validated HUMAnN's ability to quantify metabolic pathway abundances in four synthetic metagenomes containing up to 100 organisms. These were recovered with correlations over 0.9, consistently outperforming best-BLAST-hit approaches. We proceeded to scale our analysis to perform metabolic reconstruction on 649 human microbiome samples drawn from the buccal mucosa, supragingival plaque, and tongue dorsum (oral sites), anterior nares (nasal), retroauricular crease (skin), and stool (gut) communities from 102 individuals. We identified 196 metabolic pathways and 168 small modules that were differentially abundant among body sites, and we highlight here associations with environmental pH and enrichment for glycosaminoglycan degradation as examples from the vaginal and gut communities, respectively. Metabolic module abundances were substantially more variable among body sites and among individuals than was module coverage, indicating a connection between selective pressures in each microbial niche and the pathways carried by members of the community. Finally, as HUMAnN simultaneously reconstructs large pathways, specific metabolic modules, and individual enzymatic gene families, we discuss an example of glutamate metabolism in the gut community as it interacts with specific carbohydrate active enzymes (CAZys [21]). An implementation of HUMAnN is publicly available at http://huttenhower.sph.harvard.edu/humann, and the methodology can be equivalently applied to metatranscriptomic or metaproteomic data using any gene or pathway catalog of interest for future studies.
Here, we describe the methodology employed in this study in two parts: first, the computational pipeline for metagenomic metabolic reconstruction implemented in HUMAnN, and second its application to the 741 microbial community samples of the Human Microbiome Project. HUMAnN inputs metagenomic DNA sequences and infers community-wide gene and pathway abundances through a process of seven steps (Figure 1):
HUMAnN has additionally adapted ecological diversity metrics in order to provide functional diversity and richness profiles for each sample, and we validated its gene- and pathway-level accuracy using four synthetic communities of varying complexity.
To assess microbial community function and metabolism in the human microbiome, we applied this process to the metagenomic data generated by the HMP [9], comprising >3.5 Tbp of microbial DNA from 7 body sites spanning 102 individuals (Table 1). We identified modules over- or under-represented in individual body sites using the LEfSe [23] biomarker detection system, as well as associating the resulting gene and module abundances with subject clinical metadata and with external data including CAZy [21] abundances using standard nonparametric Spearman correlation.
The filtering criteria applied to HMP short reads are representative of HUMAnN's sequence preprocessing requirements. As fully described elsewhere by the HMP [24], 100 bp paired-end Illumina shotgun metagenomic reads were screened for duplicate reads and for residual human sequences. BWA [25] trimming was then applied at q = 2, followed by low-complexity filtering, and sequences resulting in less than 60 remaining valid bases were discarded. In practice, any steps removing non-microbial DNA and low-quality reads should be sufficient for HUMAnN, as retained uncharacterizable reads will be removed during the subsequent BLAST search.
Sequences passing preprocessing criteria are then searched against a characterized protein sequence database. For the HMP, we employed MBLASTX (MulticoreWare, St. Louis, MO), an accelerated translated BLAST implementation, with default parameters against a functional sequence database including the KEGG Orthology v54 [26]. All 741 HMP samples were searched in less than 13,000 CPU-hours (with 32 GB memory required on average), resulting in an average of 36% of reads mapped to at least one orthologous family, and up to the 20 most significant hits at E<1 were retained and used for further processing. The HUMAnN software additionally includes support for NCBI BLASTX, USEARCH [27], and MAPX (Real Time Genomics, San Francisco, CA) and has been tested with other sequence databases including MetaCyc [28] and CAZy [21].
HUMAnN next summarizes these BLAST results as the number of reads that matched each protein family, weighted by the quality of the matches. We used KEGG Orthology gene families (KOs) as defined by KEGG [26], a catalog of organism-independent identifiers corresponding to groups of gene sequences carrying out comparable biochemical functions. For our analysis, each KO i consists of a set of one or more specific gene sequences Gi = {gi,1, gi,2, …} from individual organisms annotated in KEGG v54. Orthologous family abundances wi were calculated independently within each metagenome for KO i and read j as:where |g| is the nucleotide length of gene sequence g in KO i, |Gi| is the number of such sequences, and pi,j is the p-value of the MBLASTX hit of read j to sequence g (or 1 if no such hit occurred), calculated from the E-value as p = 1-e−E. That is, the relative abundance of KO i in a metagenome is the number of reads j that map to a gene sequence in the family, weighted by the inverse p-value of each mapping and normalized by the average length of all gene sequences in the orthologous family. Comparisons among alternative weighting schemes (bit score, inverse E-value, and sigmoid-transformed E-value) suggested that the specific method by which multiple BLAST hits were combined had little effect on outcome (Supplemental Figure S3). Although it was surprisingly unnecessary to normalize by average gene family length in order to recover accurate pathway abundances (Supplemental Figure S1), this step was critical in inferring accurate gene family abundances (Supplemental Figures S3–4).
For each sample, the process above assigns each KO family a relative abundance; KOs are then consolidated into one or more pathways (or modules) using MinPath [22]. MinPath defines each pathway as an unstructured gene set and selects the fewest pathways that can explain the genes observed within each community. More specifically, HUMAnN associates each KO family i with a vector of relative abundances w = [wi1, wi2, …] in each metagenome. For the HMP, KOs were then assigned to zero or more pathways and modules (both as defined in KEGG) using MinPath v1.2 [22]. KOs assigned to two or more pathways/modules are effectively duplicated and their abundance included in each; this results in two independent vectors of abundance tuples of the form (KO, pathway ID) and (KO, module ID) for each metagenome. More formally, each sample is at this stage represented as a vector wp = [wj1,p1, wi1,p2, wi2,p1, wi2,p2, …], where wi,p = wj for all pathways p, and an analogous wm for modules.
For evaluation purposes, best-BLAST-hit pathway and module assignments in Supplemental Figures S1–3 were performed using only best BLAST hits, without weighting by quality of hit, normalization by gene sequence length, maximum parsimony assignments by MinPath, or the additional HUMAnN steps described below. Each best-BLAST-hit was counted once and duplicated, as for HUMAnN, into all pathways or modules within which the targeted gene occurred. For additional evaluations with best-BLAST-hit in combination with other HUMAnN processing steps, see Supplemental Figure S4.
We found an additional module/pathway filter step to be useful in removing false positive pathways selected by MinPath. Specifically, by retaining a very approximate organismal abundance profile of gene families hit during the initial BLAST process, HUMAnN is able to remove pathways in gross disagreement with observed taxa in an unsupervised manner. This is performed leniently in order to be minimally disruptive in e.g. microbial communities rich in uncharacterized organisms, and often results in depletion of false positive metazoan pathways. Specifically, taxonomic limitation is performed by removing only (KO, ID) tuples for which the same KO was assigned to multiple pathways or modules. For each sample, approximate abundances for each organism o in KEGG were calculated as a sum over all weighted, normalized BLAST hits to sequences from that organism:Each pathway/module was then assigned an approximate expected relative abundance by summing wo values over all organisms' genomes in which it was annotated. Finally, any (KO, ID) pair with two or more IDs and corresponding to a pathway/module with observed relative abundance below the average expected abundance for that ID was removed. That is, for δo,p = 1 if pathway p was annotated to organism o in KEGG and 0 otherwise, a pathway's expected abundance was:and all wi,p1 such that wi,p2>wi,p1>0 and were set to zero. Median and inter-quartile range cutoffs were also evaluated for this limitation process and the settings described here were retained due to optimal performance on synthetic data (see below and Supplemental Figure S1).
When performing this step, HUMAnN can additionally use the data on approximate taxonomic composition to divide each gene's abundance by its expected copy number in the detected organisms, providing a degree of additional normalization (as gene family copy number should not influence the abundance of pathways in which they're carried). This contrast is reflected in Supplemental Figure S1 as “Tax” and “TaxC,” respectively. As shown by our evaluation, the taxonomic limitation process both with and without copy number normalization substantially reduced false positive pathway detections caused by gene families that participate in multiple processes.
Taxonomic limitation was used by HUMAnN to reduce false positive pathways, and we found a small degree of replacement or “gap filling” of certain missing genes to likewise reduce false negatives. A small number of low abundance genes within otherwise abundant pathways often occurred due to noise or poor BLAST hits. Biological gap filling was added to increase the effective contribution of unobserved members of otherwise abundant pathways, although this had a minimal impact on overall performance in most communities. Within each retained pathway/module ID, KOs with relative abundance 1.5 interquartile ranges below the pathway median were boosted to an effective abundance equal to median for purposes of subsequent calculations. That is, for all pathways p such that there existed some wi,p>0, let be the lower inner fence of wi,p over all i∈p, and each wi,p for i∈p was set to max(wi,p, ). Add-one and Witten-Bell smoothing [29] were also evaluated as alternative methods for gap filling independent of prior biological knowledge; add-one replaces missing genes in abundant pathways with a constant value, and Witten-Bell replaces missing genes sample-wide with a small probability mass estimated from abundant genes. However, neither was retained due to a lack of improvement on synthetic data (Supplemental Figure S1).
The final outputs for each sample were thus coverage (presence/absence) and abundance values for KEGG modules and pathways. These two types of entities are quantified somewhat differently by HUMAnN, but with equivalent semantics. Pathways are defined as unordered sets of orthologous gene families; modules are defined by KEGG as combinations of required, optional, or complementary genes in notation resembling conjunctive normal form. In both cases, coverage is calculated to indicate the likelihood that all genes needed to operate the pathway or module are present; abundance is calculated as the average copy number of the pathway or module's operational subset. The definition of “operation” changes since pathways, as unordered sets, are assumed to include redundant genes (which are not explicitly indicated), whereas alternative means of accomplishing a specific metabolic module can be explicitly taken into account.
Thus each pathway's coverage and abundance were calculated fairly simply. Given the vector wp, coverage for each pathway p in a sample was calculated as the fraction of KOs in the pathway that were confidently present, specifically with abundance greater than the overall sample median. That is:Pathway abundance was calculated as the average of the upper half of its individual gene abundances, in order to be robust to low-abundance alternative enzymes; that is:for [p/2] the most abundant half of wi,p.
Modules were determined to be covered only if each gene in at least one path satisfying the module was confidently present. Specifically, given the vector wm, coverage for each module m in a sample was calculated as the harmonic mean of the Χ2 CDF with degrees of freedom evaluated at wi,m for each required i∈m, maximizing over optional genes i and alternative submodules. That is, the probability of each gene family i being present in pathway p by chance was assigned based on the sample-wide median abundance (thus adjusting for sequencing depth). This has the effect of strongly penalizing low-abundance genes, i.e. a module could not be present without all its constituent required gene abundances being confidently nonzero. Module abundances were calculated more simply as the harmonic mean of the sample gene family abundances wi, m, replacing pathways' arithmetic means since alternative enzymes are explicitly known and taken into account. These choices of parameterization both for pathways and for modules were again validated using multiple synthetic communities (Supplemental Figure S1).
Three classes of statistical tests were used to assess metabolic variability across the human microbiome. First, pathways and modules differentially abundant in at least one of the seven analyzed body sites were determined by the LEfSe system for metagenomic biomarker discovery [23]. These differences were summarized into overall patterns of variation using principal component analysis on a matrix of average module abundances per body site, Winsorized at 20% (a robust arithmetic mean [30]), filtered at a minimum of 0.01% in at least one site, and normalized to z-scores. Since LEfSe is not appropriate for HUMAnN's binary pathway coverage scores, we determined site-enriched or underenriched pathways and modules as follows: a module was in aggregate present at a site if it occurred with coverage ≥0.9 in ≥90% of the site's samples; absent if it occurred with coverage ≤0.1 in ≥90% of samples; and differential if it was present in at least one site and absent in at least one other. Pathways were analyzed identically using a ≥0.5 coverage criterion, since no large pathways consistently had coverage ≥0.9.
The third test described here associated pathway and module abundance not with human microbiome body sites, but with one or more of the subject clinical metadata variables described by the HMP [9]. These included continuous descriptors of each sample (e.g. subject age, body mass index, vaginal introitus and posterior fornix pH for women, etc.) as well as categorical variables (e.g. gender or location, see Supplemental Table S1). Pathway and module abundances were associated with these metadata first by stratifying by body site. Within each body site, each pathway/metadata pair present above 0.01% in at least 10% of samples was independently tested using Spearman's ρ for continuous metadata and the Kruskal-Wallis nonparametric ANOVA for categorical, after removing any outliers outside of the upper or lower inner fences. The resulting p-values were corrected using the Benjamini-Hochberg method within each body site and thresholded at a minimum FDR q<0.1.
Four in silico synthetic communities were constructed to validate parameter choices and to determine HUMAnN's predictive accuracy. Inspired by Mavromatis et al [31], we generated four communities, two of low complexity (LC, 20 organisms) and two of high (HC, 100 organisms). One HC and one LC had even distributions with all organisms at equal abundance, and the remaining two had log-normally distributed random abundances (see Supplemental Table S2). Organisms for the LC communities were manually selected from KEGG v54 curated reference genomes associated with the human microbiome, and HC communities were randomly generated from all manually curated bacterial genomes. A MAQ [6] error model was constructed using one lane of Illumina reads and quality scores; 106 synthetic reads were generated from this error model per organism. These were randomly mixed in proportion to organismal abundances to a total of 106 100 bp reads per community. These synthetic reads were BLASTed as above, with any hits at >90% identity discarded so as to prevent overestimates of accuracy based solely on well-characterized genomes.
Finally, gold standards of pathway/module coverage and abundance were constructed for each community by listing A) the pathways/modules annotated to at least one organism in the community and B) multiplying these by the organisms' abundances, respectively. Inferred pathway and module coverages and abundances were also calculated by applying HUMAnN to these synthetic reads as described above for the HMP samples. All software implementing these processes and the specific error model, synthetic communities, and data used for HUMAnN in the HMP are available at http://huttenhower.sph.harvard.edu/humann.
All metabolic reconstructions generated by this study are publicly available at http://hmpdacc.org/HMMRC. Taxonomic abundances derived from shotgun data are provided at http://hmpdacc.org/HMSCP, and input Illumina reads at http://hmpdacc.org/HMIWGS. The open source HUMAnN software can be obtained at http://huttenhower.sph.harvard.edu/humann.
We first validated HUMAnN's ability to accurately quantify microbial community function using a collection of four synthetic metagenomes. We proceeded to reconstruct metabolic pathways and modules for 741 human microbiome samples comprising a total of 3.5 Tbp of sequence from 18 body habitats from 102 subjects assayed metagenomically by the HMP. 686 of these samples passed quality control (see [9]), and after grouping bilateral habitats (left and right retroauricular creases), seven habitats included at least 25 samples: buccal mucosa, supragingival plaque, and tongue dorsum in the oropharynx; anterior nares and retroauricular crease representing airways and skin; stool samples representing the gut; and the vaginal posterior fornix. These habitats together comprised the 649 total microbial community samples analyzed here (Table 1). In the following sections, we report on their metabolic reconstructions using modules and pathways, discuss instances of inter- and intra-habitat functional variation, and present examples relating community function to microbial abundances and to host phenotype.
To assess the accuracy of HUMAnN's metabolic reconstructions, we constructed four synthetic metagenomic datasets with known functional profiles. Patterned on the study of Mavromatis et al [31], these included two low-complexity communities with 20 organisms each and two high-complexity with 100 organisms each. The former were manually chosen from representatives of the human microbiota, and the latter were randomly selected from KEGG high-quality bacterial genomes (see Methods, Supplemental Tables S2–3). Likewise, two of the communities contained organisms with equally distributed abundances, and two possessed lognormally distributed abundances to mimic physiological communities. A gold standard of functional modules and pathways as defined by KEGG [26] was assembled, comprising 13,980 gene families, 370 total small metabolic modules, and 309 large pathways. 251 and 303 of the latter, respectively, included at least four genes and were used here. Their presence in these communities was determined from KEGG genome annotations for the chosen organisms; these were used to evaluate a naive best-BLAST-hit metabolic reconstruction as compared with HUMAnN's inferences. In all cases, BLAST hits with >90% identity were discarded, preventing overconfident evaluations due to the gold standard's use of only well-characterized genomes and forcing a conservative estimate of HUMAnN's expected performance.
In all four communities, HUMAnN recovered metabolic module abundances with a correlation above 0.88 and a partial AUC at 10% false positives (pAUC10) above 0.73 (Figure 2, Supplemental Figure S2). Performance was generally comparable for large pathways (ave. ρ = 0.90 sd. 0.02, pAUC10 = 0.85 sd. 0.05), and in all cases HUMAnN outperformed reconstructions based on the best BLAST hit alone (Supplemental Figures S1–2). Although HUMAnN was not optimized to recover individual gene family abundances, it performed comparably to best-BLAST-hit at a correlation of 0.93 sd. 0.01 among the four communities. These synthetic communities were further used to refine the inclusion of computational steps within the HUMAnN pipeline and to assess the robustness of their parameter settings. For example, smoothing of low-abundance gene family frequency estimates proved to have surprisingly little overall impact, whereas MinPath was particularly critical for accurate pathway coverage determination (Supplemental Figure S1).
Within these four synthetic communities, only a few classes of gene families, modules, and pathways were recovered incompletely by HUMAnN. In the most complex staggered community, erroneous gene family calls included just 29 proteins missed as false negatives (of 5,640 total, 1.3% at abundance >10−4), typically short proteins <150AAs such as K13771 (the Rrf2 family transcriptional repressor) or K10533 (limonene-1,2-epoxide hydrolase). Likewise only 7 false positive proteins were detected based on closely related orthologs or strongly conserved domains (0.027% at >10−4), including K08721/OprJ (confused with K07796/cusC, blastp e = 3·10−82) and K11187/peroxiredoxin 5 (confused with peroxiredoxin 2, blastp e = 2·10−19). False positive modules (8, 3.3% at a 10−4 cutoff) were near-uniformly small gene sets overlapping with modules truly present (e.g. the urea cycle M00029, with five total genes and four present in the community). Conversely, false negatives (9, 3.7%) were most often small pathways present only in very low-abundance organisms, e.g. bicarbonate transport (M00321, four genes present at 0.022% relative abundance) or mannopine transport (M00301, four genes at 0.0014%). We specifically tuned HUMAnN to prefer false negatives to false positives based on these communities (see Supplemental Figure S2), and Figure 2 demonstrates the minimal impact of this choice even at low recall in the most complex community.
Fortunately, 89% of high-complexity (staggered) and 93% of low-complexity (even) modules were correctly called present or absent, and their inferred abundances were consistently well-correlated with the gold standard (Figure 2). These unsurprisingly included large, well-conserved pathways such as the ribosome (M00178) and polymerase (M00183), but also a variety of smaller specialized modules such as sugar transport (M00207, four genes, present in staggered/high-complexity at 0.25% and detected at 0.28%) and biotin biosynthesis (M00123, four genes, present at 1.7% and detected at 1.7%). Almost all large pathways were quantified with high accuracy, again due to their larger metagenomic footprint and detectability; examples included the TCA cycle (ko00020, 53 genes, present at 1.1% and detected at 1.2%) and base excision repair (ko00240, 152 genes, present at 1.1% and detected at 1.1%). In contrast, false positive rates for the best-BLAST-hit approach exceeded 23% of modules in the low-complexity community. Overall, as summarized in Figure 2 and in Supplemental Figures S1–3, this evaluation on synthetic metagenomes established that HUMAnN can accurately reconstruct community metabolic pathways and modules directly from short reads.
We employed this optimized system to study microbial metabolism at seven body sites spanning the human microbiomes of 102 subjects. HUMAnN was applied to these data as described above, yielding the relative abundances and coverages both of functional modules and of full pathways, as well as the abundances of individual orthologous gene families. We first focused on analysis of small metabolic modules (ave. 11.2 sd. 9.2 genes), and 232 such modules were detected in at least one of the 649 samples; larger, more broadly defined pathways were also reconstructed from the HMP data and are described below. Both modules and pathways were reconstructed by HUMAnN as coverages (presence/absence calls on a zero-to-one scale) and as relative abundances for each sample (Supplemental Tables S4–5). The resulting metabolic reconstructions were complementary to the organismal compositions of the communities (see [24]) and provided a link between microbial environment and metagenomically prevalent pathways and metabolic potential.
In these data, we observed a core of 16 metabolic modules present at >90% coverage in >90% of samples (Supplemental Table S6), in contrast to essentially no specific microbes found to be core in this population [1]. However, in agreement with a gene family core from the gut microbiomes of an independent cohort [11], these modules comprise the functionality essential for microbial life: transcription (M00183, M00049-52), translation (M00178, M00360), transport (M00207, M00222, M00239), central carbon metabolism (M00001-2, M00006), and energy production (M00120, M00125, M00157, M00164). By relaxing the coverage threshold to 30% (expected to introduce few, if any, false positives; see 4.4), only 8 additional modules were included, demonstrating robustness to this threshold. Two of these extended the categories listed above; the remainder comprised sn-Glycerol 3-phosphate transport (M00198, a membrane lipid precursor [32]), the mannose and trehalose phosphotransferase systems (M00270 and M00276), spermidine/putrescine transport (M00299), early terpenoid biosynthesis (M00364), and threonine biosynthesis (M00018), the only amino acid module to meet this prevalence threshold. It should be noted that these 24 core modules are not the most abundant; for example, sn-Glycerol 3-phosphate transport reaches only a mean relative abundance of 0.001 sd. 0.002 across all samples, compared with the ribosome (M00178) at 0.03 sd. 0.008. Nor were they evenly abundant among habitats, as examples including phosphate and sugar transport (M00197 and M00222) are highly enriched in the posterior fornix as discussed below. The organisms performing the more specialized processes in this list also vary among habitats. Spermidine and putrescine are metabolized by the abundant Streptococcus spp. in the oral community, for example, processes that can play a role in halitosis [33]. However, this metabolic module is not present in reference genomes for the skin community's abundant Corynebacterium and Propionibacterium spp. [8], [26], and its abundance is instead correlated with that of the Staphylococcus reference genomes [24] (Spearman r = 0.87, n = 26, p = 1.8·10−6). Although this stringent core is itself moderately small, most other modules were consistently present or absent across body sites, with only 24 showing strongly differential coverage among habitats (Figure 3).
To investigate microbial functions over- or under-enriched within specific niches of the human microbiome, we determined modules differentially abundant in at least one body site using the LEfSe biomarker discovery suite [23] (Figure 3). Over two thirds (168, 67%) of detected metabolic modules varied significantly in abundance in at least one habitat, demonstrating the uniqueness of each body habitat's microbial environment (Supplemental Table S7). In addition to the detailed examples below, these included such diverse processes as arginine transport (M00229) and methionine biosynthesis (M00017) enriched in all three oral habitats, an enrichment for fungal transcription (M00181) and translation (M00177) in the skin and airways, and a strong depletion of pyruvate (M00307) and second carbon oxidation (M00011) in the anaerobic gut and vaginal sites. Several overall patterns of co-variation are shown in Figure 4, where the first principal component captured primarily eukaryotic modules found only on the skin and often nares (including, intriguingly, vitamin D biosynthesis, M00102). Interestingly, it also included metabolism abundant throughout the digestive tract (i.e. oropharynx and gut), such as putrescine (M00300) and sulfate transport (M00185) [23]. The second principal component described functionality depleted in the low-complexity vaginal habitat, and the third comprised processes enriched in the gut, both discussed below. The three oral habitats are often functionally similar (apparent in components 1–3), and the fourth emphasized modules unique to the tooth surface, the only microbially colonized hard surface assayed here, in contrast to the mucosal and tongue soft tissues [23]. While the skin and nares were likewise often similar, the final principal component shown in Figure 4 differentiates the two. These summaries show that while a wide range of microbial metabolism is present throughout the human microbiome, specific subsets of this functionality are selected for by the unique combinations of nutrients, immune pressures, and environmental exposures present at each body site.
Ecological and phylometagenomic studies of organismal abundance often employ summary statistics including species richness, evenness, or diversity to characterize and compare communities [34]. These quantify how many distinct types of organisms occupy a community, the uniformity of their relative abundances, or both, respectively. These measures were historically adopted from macroecology into microbial ecology, and while the former has included assessments of functional diversity [35], [36], it has been proposed for microbial communities [37], [38] but not yet widely adopted in metagenomic studies. Functional diversity measures are thus calculated by HUMAnN as a novel means of profiling microbial community structure, with some differences from their applications in macroecology as described in the Discussion. A simple measure of richness is calculated by summing module coverage scores within each sample [34], and Pielou's evenness [39] and the Shannon and inverse Simpson [34] diversity measures are calculated from module abundances. As observed qualitatively by Turnbaugh et al [5] and analyzed quantitatively by the HMP [1], microbial metabolic function differs significantly less among subjects than does organismal diversity within each habitat. Ecological summary statistics of community function may thus represent a unique descriptor of metagenomic data complementary to standard organismal diversity measures.
Modules performing glycosaminoglycan (GAG) degradation in the gut were among those most differentially abundant among habitats. These included chondroitin sulfate degradation, dermatan sulfate degradation, and keratan sulfate degradation, as well as the related uronic acid metabolism. All four of these modules are involved in animal proteoglycan degradation for microbial carbohydrate utilization [40], and they were present in high abundance in the gut (coverage >0.9 in 136 stool samples, 100%) and rare or completely absent in other body sites (coverage <0.1 in 131 samples, 26%; abundance <0.001 in 512 samples, 99%). This degree of specificity was unusual in the HMP dataset - as mentioned above, relatively few modules were present or absent in only one such body site. However, it provides a readily identifiable example of community metabolism associated with a specific clade, as glycosaminoglycan degradation is known to be enriched among the Bacteroides species [5], [41]. The model B. thetaiotaomicron alone, for example, carries nearly 90 different polysaccharide utilization loci, many targeting dietary starches, but at least 16 specific to host mucin O-glycan degradation [42]. As the Bacteroides are one of the predominant genera in the gut microbiome [43], are largely characteristic of only that body site, and are abundant in the HMP gut samples [1], they are highly likely to be responsible for this niche-specific metabolism.
HUMAnN's reconstruction further allowed examination of the individual orthologous gene families participating in these interrelated pathways. Chondroitin and dermatan sulfate degradation share the greatest overlap with 3 KO families, including beta-glucuronidase (K01195). Beta-glucuronidase is prevalent in the gut, averaging 0.03% sd. 0.01% in our data, in contrast to all other nonzero KOs (ave. 0.007% sd. 0.03%). It is one of the primary enzymes involved in metabolism both of food matter and of pharmaceuticals [44], as well as mediating effects ranging from dietary cancer risk [45] to antibiotic activity [46]. This enzyme also links uronic acid metabolism with the rest of pentose and glucuronate processing, the former being the most abundant module of these gut-specific examples. Uronic acid, like the GAG sulfates, is a component of dietary fiber and glycoproteins degraded by intestinal bacteria [42]. In contrast, heparan sulfate degradation, an additional module included in glycoprotein degradation as defined by KEGG, is present at only low abundance among all body sites, in spite of sharing nearly half of its enzymes with the other three modules (ave. 7·10−4% sd. 0.003% in the gut, <10−5% elsewhere). This is due exclusively to the absence of the module's input enzyme, heparanase (K07964-5), a typically eukaryotic gene family implicated in tumor metastasis [47]; the healthy commensal microbiota may thus lack this activity in order to avoid undesirable inflammatory and immune response in the gut [48]. Conversely, the abundances of the four gut-specific GAG degration modules were not driven by one specific gene family (Figure 5), ranging from the most abundant beta-hexosaminidase (K12373, ave. 0.3% sd. 0.1%) to the low-abundance outlier L-iduronidase (K01217, ave. 2·10−5% sd. 3·10−7%). Given this ubiquity, it should be noted that GAGs possess wide-ranging activities including anti-cancer [49] and antimicrobial [50] properties; the abundance of degrading enzymes in the gut of any particular individual thus has the potential to affect the efficacy of GAG drugs [41], [51]. The specificity and prevalence of these four modules in the gut microbiota provides one example of HUMAnN's ability to identify community metabolism across hundreds of samples and link it to individual microbial gene families.
We next examined modules differentially abundant in the 53 HMP vaginal posterior fornix samples (Supplemental Table S7). Enriched functions included phosphate, glutamate/aspartate, and phosphotransferase transport systems (M00222, M00230, M00276, M00277, M00287); other functions such as arginine and amino acid transport (M00229, M00237) were significantly depleted. The posterior fornix community's metabolic pattern was perhaps the most distinct of the habitats examined here, in agreement with the high degree of biochemical specialization (and thus intercellular transport) observed in the low-diversity vaginal community [7]. Perhaps most interestingly, the misleadingly labeled “nitrogen fixation” module (M00175) was overabundant in the posterior fornix, as well as in buccal mucosa and stool. This module can be driven by any one of four gene families: K00531, K02588, K00536, or the complex K02586+K02591. The low number of genes involved would render its detection prone to noise; however, as described below, this module is detected very consistently only in women with vaginal pH≥4.0, and then only due to either K00536 (Enzyme Class EC 1.19.6.1, nitrogenase - flavodoxin) or, in a minority of cases, K02588 (EC 1.18.6.1, nitrogenase). These two gene families were mutually exclusive when detected, and the most common family K00536 is only included in the current reference genomes for Bacillus cereus, B. thuringiensis, and the archaeon Archaeoglobus fulgidus, none of which were detected in any vaginal samples [9]. While it seems unlikely that these enzymes are contributing to canonical nitrogen fixation per se, flavodoxin has been observed to be involved in the catalysis of nitric oxide production in the gut microbiota [52], [53], and disrupted amine production in conjunction with elevated vaginal pH is a standard symptom of bacterial vaginosis [7]. In combination with the vaginal pH-associated modules below, this provides one example of microbial functionality linked to host phenotype that could not be recovered based on reference genomes, community structure, or phylometagenomic assays alone.
Given their lower diversity, the functional profiles of these vaginal microbiomes proved to be particularly informative when associated with host phenotype and microbial membership. Taxonomic profiles detailing the abundances of microorganisms in these communities were available from the HMP's 16S rRNA gene surveys [9], from mapping of metagenomic sequences to reference genomes [24], and from previous studies of the vaginal flora [7]. Specifically, Ravel et al [7] showed the presence of five microbiome types among the vaginal communities of reproductive age women, with four clusters dominated by different Lactobacillus species and one with low levels of all lactobacilli. Communities dominated by L. crispatus corresponded to lower pH (<4), which was also the case in the HMP posterior fornix data [1]. We observed a very tight co-clustering of the metabolic repertoires of these communities with the abundances of these five groups' characteristic species (Supplemental Figure S5), suggesting that in this specialized microbial niche, there is a particularly close association of community structure and function. To expand on this, we performed a comprehensive test of all 232 modules against the HMP's clinical metadata (phs000228.v3.p1 [54]) to uncover associations between microbial metabolic profiles and host phenotype. Available metadata included gender, age, BMI, geographical location, and others, as well as the pH of the posterior fornix and vaginal introitus in women. The latter were again strongly linked not only to community structure, but also with the abundances of several metabolic modules. Specifically, pH correlated with the abundances of N-acetylgalactosamine II phosphotransferase (M00277), proline biosynthesis (M00015), and again “nitrogen fixation” (M00175), while phosphate transport (M00222), peptide/nickel transport (M00239), and lysine biosynthesis (M00016) were associated with lower pH. Species within the Lactobacilli are known to specialize in exactly these areas of nutrient uptake and carbohydrate metabolism [55], [56], and many of these differences can be observed directly within finished genomes (e.g. lysine biosynthesis is present in L. crispatus and lacking in L. gasseri [26]). It is currently near-impossible to obtain complete genomes for all organisms in complex communities, however, again emphasizing the utility of metagenomic functional reconstruction for direct association of community function with habitat and host phenotype.
A surprising feature of these data, however, was the observation of few other robust correlations between microbial metabolic abundances and host phenotype outside of the vaginal community, in contrast to recent studies of the gut microbiota [12] (Supplemental Table S1). Of all phenotypic associations with metabolic modules tested in the HMP stool microbiomes, only two reached significance (Spearman FDR q<0.1), links between diastolic blood pressure and nickel transport (M00246) and between pulse rate and methionine degradation (M00035). Although these are within the range of false positives expected at this level of stringency, recent observations of the gut microbiota in atherosclerosis [57] have also detected borderline significant shifts in organismal composition. Among other body sites, the only additional associations were correlations with BMI among the skin community: lysine biosynthesis (M00016) and sulfate transport (M00185). Although the retroauricular crease represents our smallest sample size (n = 26), these were significantly stronger (p<10−4) than any metabolic association with BMI in the stool community (n = 126, lowest p<10−3), with the metagenomic abundance of lysine biosynthesis consistently increasing and that of sulfate transport decreasing at greater BMI. Additional associations beyond this false discovery rate are included in Supplemental Table S1 and include several modules detected only at higher read counts, emphasizing the need for sufficient sequencing coverage when rare metabolic functions (as opposed to abundant taxa) are of interest. As has been shown for genotype [58] and gene expression [59] data, phenotype can be difficult to reproducibly associate with high-dimensional genomic or metagenomic features, particularly in large cohorts with complex population structure or when using multivariate models [60]. It is also critical to note that the HMP population was strictly screened for disease-free individuals [9], increasing phenotypic homogeneity and precluding the detection of microbial function perturbed in dysbioses. Further studies targeted to specific phenotypes and microbial communities of interest will be necessary to better understand the relationship between human microbiome membership, metabolism, and host phenotype.
The analyses described up to this point, and the primary outputs of HUMAnN for metagenomic samples, focus on the coverages and abundances of small functional modules and individual gene families. Such modules are typically defined to carry out a specific metabolic step, such as production of a single amino acid from its immediate precursor. They contain an average of only ∼10 genes and are structured to include both “and” relationships (multiple gene products that must function together in a complex or sequential pathway) and “or” relationships (alternative enzymes that can catalyze the same reaction). However, KEGG and other functional catalogs also define large pathways with up to several hundred genes that lack the complex combinations of relationships used to define small modules. HUMAnN can additionally recover coverages and abundances for any such pathways represented by unstructured sets of gene family identifiers. To this end, we analyzed 297 KEGG pathways present in the seven HMP body sites, containing 52 sd. 48 genes on average. While HUMAnN successfully recovered coverage and abundance information for these larger pathways in the HMP data (Supplemental Tables S8–9), we often found them to be too broadly defined for adequate analysis of mixed microbial communities, arguing for a focus on smaller functional modules, biosynthetic clusters [61], and orthologous gene families.
Demonstrating these larger pathways' lack of specificity, 190 (64%) were at most 50% covered among all 649 samples; that is, although portions of the pathways were detected, at least half of the associated gene families were missing in every assayed microbial community. Of the remaining 107 pathways, 96 (90%) had a coefficient of variation below one, indicating variation in coverage lower than their mean across all samples. In other words, only small portions of most KEGG pathways are present in the human microbiome. Those pathways that were consistently present tended to be so broadly defined (e.g. ko02060, phosphotransferases, or ko00030, the entirety of pentose phosphate) that they provided too coarse a view with which to observe metabolic variation and niche specialization among body sites. Nevertheless, applying the same significance criteria as described above, 196 pathways (66%) were differentially abundant in at least one body site using LEfSe. These are detailed in Supplemental Table S7 and are for the most part analogous to the more specific small modules described above.
In contrast to this variation, however, we were again able to recover a subset of “core” pathways with moderate coverage and low variability across all sites of the human microbiome. The pathways with the lowest coefficient of variation across all samples in our data represented surprisingly diverse biochemistry, including terpenoid biosynthesis (ko00900), RNA degradation (ko03018), pyruvate metabolism (ko00620), and one-carbon metabolism (ko00670). These pathways were all present at an average of at least 0.5% relative abundance over all samples, with average coverages from 41% (of 75 genes in ko03018) to 77% (of 73 genes in ko00620). Each of these pathways followed a very characteristic pattern: their overall pathway abundance remained near-constant across body site niches, while the modules implementing each pathway varied significantly. Terpenoid biosynthesis, for example, can be performed either by the mevalonate (module M00095) or by the non-mevalonate (M00096) module; both were present in the oropharynx and skin, only the former was present in the posterior fornix, and only the latter in the gut (Supplemental Table S5). Likewise, pyruvate metabolism includes portions of the citric acid cycle (M00173), absent from the posterior fornix and rare in the gut and skin; pyruvate oxidation (M00307), absent from the posterior fornix and rare in the gut; and pyruvate ferredoxin oxidoreductase (M00310), present only in the nares and gut. This trend complements the relationship between organismal and functional diversity described above and by the HMP overall [1], in which the communities at each body site of the human microbiome vary extensively, but the pathways needed for microbial life within these niches remain relatively stable. However, this result shows that the specific metabolic modules implementing stable, broad pathways tend to be specialized within each body site's microbial environment.
This pattern extended to individual gene families as well; while the pathways above demonstrated low variability across all body sites, other pathways were enriched at specific sites but possessed low inter-subject variability in the HMP population (Supplemental Table S9). These include, for example, glutamate metabolism in the gut (ko00250), which has been previously observed to be enriched in the stool microbiota [62]. Here, it showed low inter-individual variability in the gut, although as above, its product glutamate serves as input for several modules that were themselves highly variable among subjects. Among others, these included proline biosynthesis (M00015); glutathione biosynthesis (M00118); glutathione transport (M00348); and portions of the TCA cycle (M00009–M00011 and M00311). We determined individual carbohydrate active enzyme (CAZy) gene families correlated with these modules by comparing them with CAZy abundances derived independently from the HMP metagenomic data [21]. Intriguingly, each of these modules in the gut correlated significantly (Spearman FDR q<0.05) with multiple CAZy families, almost none of which included enzymes within the modules themselves (Supplemental Table S10). In particular, six CAZy-module pairs had significant positive or negative correlations in all seven body sites: proline biosynthesis with GH5 (cellulase) and GH18 (chitinase); 2-oxoglutarate ferredoxin oxidoreductase with GH28 (galacturonases) and GH97 (α-glucosidase and α-galactosidase); and glutathione transport with GH3 (β-glucosidase) and GH18.
Summarizing these three results, glutamate metabolism falls within a class of 57 pathways present in multiple niches within the human microbiomes sampled here but enriched in specific environments (e.g. the gut). 101 total large pathways varied little in site-specific abundance among individuals, but smaller modules within them (such as proline biosynthesis) showed greater inter-subject variability. Finally, these module-specific changes in abundance almost always (99% of all modules) correlated with one or more individual CAZy gene families detected within the same metagenome independently of the pathway's constituent genes. Together, these data suggest that while the basics of microbial metabolism remain stable among human microbiome body sites and individuals, the modules and enzymes operating as specific metabolic producers and consumers within pathways vary among environments and subjects to adapt to changing nutrient and metabolite availabilities [10].
Culture-independent metagenomic sequencing of microbial communities provides a wealth of data regarding their potential biological functions, particularly as studied for the human body by the Human Microbiome Project. Here, we have described the development of the HUMAnN methodology for high-throughput metagenomic functional reconstruction and its application to 649 communities from 7 body habitats sequenced as part of the HMP. Validation of HUMAnN's accuracy using four additional synthetic communities of increasing complexity demonstrated its ability to quantify both pathway presence and relative abundance, with correlations to true abundances >0.9. When analyzing the human microbiome, relatively few modules were specifically present or absent in any one body habitat, but over two thirds varied in abundance by habitat and 24 were core to all hosts and habitats. Less variation was evident among hosts, although we demonstrated one example in which nutrient transport and mechanisms of central carbon metabolism were strongly associated with vaginal pH. HUMAnN's functional reconstructions include the abundances of large, general pathways, smaller and more specific metabolic modules, and individual orthologous gene families; each data type proved to show distinct patterns of variation among body sites and to provide a different perspective on underlying microbial community function.
Characterization of microbial communities by large-scale shotgun metagenomic sequencing is a relatively recent advance, and computational methods for assessing these data in terms of biological function are under active development. Several previous studies have analyzed individual reads or assembled contigs using direct annotation by BLAST to orthologous gene families [5], [11] or to proxy genes [63]. Other computational pipelines such as MG-RAST [14] and MEGAN [15] do include full pathway reconstruction, generally using approaches that rely on the single best BLAST hit for each metagenomic read. Here, HUMAnN scaled easily to provide module and pathway reconstructions for >3.5 Tbp of HMP metagenomic data by avoiding metagenomic assembly and employing an accelerated translated BLAST implementation, requiring a total of approximately 13,000 CPU-hours for sequence search and 175 for metabolic reconstruction (750 and 55K sequences/second, respectively). HUMAnN is not dependent on any particular BLAST implementation, however, and provides default support for NCBI BLAST, MBLASTX as employed here, MAPX (Real Time Genomics, San Francisco, CA), and USEARCH [27]. In each case, the approximations used to accelerate search against a functionally characterized orthologous sequence database are mitigated by considering all read-to-sequence hits in a weighted manner. This leaves overall gene family abundance recovery essentially unchanged while improving full module/pathway recovery, as ambiguous BLAST hits can be resolved later in the reconstruction process when more information is available (Supplemental Figure S2). Further, each of HUMAnN's processing modules incorporates one or more types of additional knowledge, e.g. pathway parsimony by means of MinPath [22] and subsequent automatic taxonomic limitation based on BLAST organismal abundance profiles. These steps are not guaranteed to be optimal in all situations - taxonomic limitation, for example, might degrade performance in environments rich in novel or rapidly evolving organisms - but they are heuristics designed to improve reconstruction in most cases. They thus generally take advantage of the compositional information that can be leveraged when combining multiple gene family sequences into a single module or pathway, decreasing the noise potentially arising from examining single best BLAST hits.
It should be noted that when analyzing metagenomic data as described here, HUMAnN reconstructs a profile of a microbial community's metabolic potential, not its metabolic activity per se. The abundances of gene families and pathways inferred by the system describe only the enzymes encoded by one or more microbial genomes, and their relationship to realized transcriptional or protein activity may not be straightforward in the absence of additional metatranscriptomic, metaproteomic, or metametabolomic data [64]. However, these metagenomic gene family and module abundances are appropriate as inputs into more sophisticated metabolic network and systems biology models, which have recently begun to incorporate features such as predicted compartmentalization, small molecule transport, and multi-organism interactions in microbial communities [65], [66]. HUMAnN as described here was designed to infer a permissive superset of community function that does not yet include realized transcriptional activity or organismal compartmentalization, and we hope to incorporate these features during future work. It should be emphasized that HUMAnN as currently implemented is appropriate for analysis of metatranscriptomic data from short sequence reads as well, from which it will reconstruct the abundance of actively transcribed gene families or pathways within a microbiome.
The results produced by HUMAnN and analyzed above for the human microbiome include the application of several community diversity measures to microbial function. Such measures are typically applied instead to organismal abundances, where α-diversity summarizes complexity and types of different organisms within a community and β-diversity the similarities (or differences) between multiple communities' structures [34]. Such organismal diversity measures have been very successful in describing properties of the human microbiome in large populations, such as the greater similarity of children's and parents' microbiomes [5] or reduced microbial diversity in conditions such as Crohn's disease [67]. Conversely, ecological functional diversity has been developed primarily in macroecology, specifically as applied to phenotypic traits [35], [68]. To our knowledge, however, this represents the first application of α-diversity measures to molecular function within microbial communities and specifically to the human microbiome. The HMP consortium has contrasted the functional diversities reported above with comparable organismal diversity measures at the genus, species, and strain levels throughout the human microbiome [1]. Their results suggest that functional diversity is lower than phylogenetic diversity both within and between communities throughout the human microbiome; that is, the microbes within this human population vary more than do the biological processes carried by their metagenomes. It must be noted that this conclusion speaks so far only to the disease-free HMP population, however, and only to the subset of characterized orthologous gene families currently analyzable by HUMAnN. Further variability in the functional potential of the microbiota may certainly remain to be found in its substantial carriage of uncharacterized gene families (estimated as high as 80% [11]) and during disruptions of host health.
A key consideration during our development of the HUMAnN pipeline was versatility; the software implementation can easily be extended to assess any functional catalog, characterized sequences, or metagenomic sequences (e.g. 454 reads). In other analyses by the HMP, additional protein databases including MetaCyc [28], CAZy [21], virulence related proteins [69], and antibiotic resistance genes [70] were all processed using HUMAnN. MetaCyc, for example, includes both characterized sequences and metabolic modules, for which HUMAnN reconstructed coverages and abundances; other databases included no explicit pathway groupings, and gene families were used directly to examine differential abundance. While these smaller databases are less appropriate for quantitative evaluations or broad metabolic reconstruction, they can be used with HUMAnN to provide focused coverage of specific biological areas. As detailed above, in addition to its primary abundance and coverage outputs, HUMAnN by default calculates a number of basic ecological summary statistics as applied to community functional profiles; it also produces detailed gene-level outputs for each community that can be directly imported into the JCVI Metagenomics Reports (METAREP) [71] software. All components in the pipeline, including taxonomic limitation, are entirely data-driven; the methodology can therefore be used for functional reconstruction on any genomic data from microbial or eukaryotic organisms, although in a single-organism setting, there are not clear benefits over standard genome annotation pipelines. However, individual modules (such as gap filling or the inclusion of multiple BLAST hits) can be manually activated or deactivated by the user for particular datasets. Importantly for microbial communities, HUMAnN can also be used on other data types, including metaproteomic or metatranscriptomic sequences; we anticipate HUMAnN being useful in the reconstruction of pathway activities in transcriptomic sequences from different environmental communities, for example.
In closing, we would like to emphasize that HUMAnN's current approach to microbial community functional reconstruction is explicitly independent of the organismal membership of these communities. It was designed to complement taxonomic classifications of community structure, and integration of community function with membership is an area of further ongoing work [15]. Particularly in the human microbiome, full genome sequences are available for many reference strains isolated from multiple body sites, which has already allowed community membership to be analyzed simultaneously in metagenomic and 16S taxonomic marker sequences [1], [24]. By combining membership with functional reconstruction, specialized processes in specific habitats or hosts, for example, can be correlated with the organisms providing or dependent on these aspects of community function. While the general applicability of Beijerinck's 1913 hypothesis [72] that, “Everything is everywhere, and the environment selects,” is still unclear, we speculate that it may prove to be more broadly accurate for microbial function than for microbial organisms. That is, there may be a moderately stable pool of core microbial pathways, present in all communities but implemented by different organisms and gene families, with relative abundance (and activity) determined by the local selective pressures of each microbial habitat. This appears to be at least somewhat the case in the human microbiome, and further investigation will determine whether this pattern holds for the functional profiles of broader classes of microbial communities.
|
10.1371/journal.ppat.1001177 | HTLV-1 Evades Type I Interferon Antiviral Signaling by Inducing the Suppressor of Cytokine Signaling 1 (SOCS1) | Human T cell leukemia virus type 1 (HTLV-1) is the etiologic agent of Adult T cell Leukemia (ATL) and the neurological disorder HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP). Although the majority of HTLV-1–infected individuals remain asymptomatic carriers (AC) during their lifetime, 2–5% will develop either ATL or HAM/TSP, but never both. To better understand the gene expression changes in HTLV-1-associated diseases, we examined the mRNA profiles of CD4+ T cells isolated from 7 ATL, 12 HAM/TSP, 11 AC and 8 non-infected controls. Using genomic approaches followed by bioinformatic analysis, we identified gene expression pattern characteristic of HTLV-1 infected individuals and particular disease states. Of particular interest, the suppressor of cytokine signaling 1—SOCS1—was upregulated in HAM/TSP and AC patients but not in ATL. Moreover, SOCS1 was positively correlated with the expression of HTLV-1 mRNA in HAM/TSP patient samples. In primary PBMCs transfected with a HTLV-1 proviral clone and in HTLV-1-transformed MT-2 cells, HTLV-1 replication correlated with induction of SOCS1 and inhibition of IFN-α/β and IFN-stimulated gene expression. Targeting SOCS1 with siRNA restored type I IFN production and reduced HTLV-1 replication in MT-2 cells. Conversely, exogenous expression of SOCS1 resulted in enhanced HTLV-1 mRNA synthesis. In addition to inhibiting signaling downstream of the IFN receptor, SOCS1 inhibited IFN-β production by targeting IRF3 for ubiquitination and proteasomal degradation. These observations identify a novel SOCS1 driven mechanism of evasion of the type I IFN antiviral response against HTLV-1.
| Infection with HTLV-1 leads to the development of Adult T cell Leukemia (ATL) or the neurological disorder HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP). Although the majority of HTLV-1–infected individuals remain asymptomatic carriers (AC) during their lifetime, 2–5% will develop either ATL or HAM/TSP. Using gene expression profiling of CD4+ T lymphocytes from HTLV-1 infected patients, we identified Suppressor of cytokine signaling 1 (SOCS1) as being highly expressed in HAM/TSP and AC patients. SOCS1 expression positively correlated with the high HTLV-1 mRNA load that is characteristic of HAM/TSP patients. SOCS1 inhibited cellular antiviral signaling during HTLV-1 infection by degrading IRF3, an essential transcription factor in the interferon pathway. Our study reveals a novel evasion mechanism utilized by HTLV-1 that leads to increased retroviral replication, without triggering the innate immune response.
| Infection with the Human T cell Leukemia Virus type I (HTLV-I) can result in a number of disorders, including the aggressive T cell malignancy Adult T cell Leukemia (ATL) and the chronic, progressive neurologic disorder termed HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP) [1], [2], [3]. In endemic areas including Southern Japan, the Caribbean basin, Western Africa and Central/South America - where infection rates range from 2 to 30%- these diseases are major causes of mortality and morbidity [4]. The majority of HTLV-1–infected individuals remain asymptomatic (AC) during their lifetime and only ∼2–5% of AC will develop either ATL or HAM/TSP [5], [6]. Although the factors determining progression from AC to ATL or HAM/TSP remain unknown, it is well established that the risk of ATL vs. HAM/TSP development varies dramatically with the geographical distribution of HTLV-1-infected populations.
Clinically, acute ATL is characterized by abnormally elevated T cell counts, accompanied by readily observed ‘flower cells’ – multi-lobed, leukemic cells with highly condensed chromatin - hypercalcemia, prominent skin lesions, hepatosplenomegaly and suffer from serious bacterial, viral, fungal and protozoan infections. Most patients present at this final acute stage, often unaware of their HTLV-1 positive status and given a poor prognosis, with a survival estimate of 6–10 months [7]. Transformation of CD4+ T lymphocytes by HTLV-1 and the development of ATL leukemogenesis generally occur in two stages [8], [9]. After infection with the blood borne pathogen, HTLV-1 induces IL-2-dependent, CD4+ T cell proliferation, that over a period of decades in vivo, progresses with the emergence of an IL-2-independent malignant clone that has accumulated multiple secondary genetic changes in growth regulatory and tumor suppressor genes [9], [10]. HTLV-1 encodes the 40-kDa nuclear oncoprotein Tax that promotes cellular transformation through dysregulation of mitotic checkpoints, activation of cellular signaling pathways and inactivation of tumor suppressors (reviewed in [11], [12]).
HAM/TSP is a systemic immune-mediated inflammatory disease characterized by demyelination of motor neurons in the spinal cord, although other tissues can also be damaged [13]. HAM/TSP attacks in the prime of life (median age of onset: 35 years) and is associated with a clinical history that includes neurological symptoms in 80% of cases – gradual onset of leg weakness, paresthesis, and impairment of urinary or bowel function. Central nervous system (CNS) white matter lesions of the spinal cord harbor activated CD4+ and CD8+ T cells during early stages of disease, with a predominance of CD8+ T cells later in disease. HTLV-1 viral RNA has been found associated with CD4+ T cells and astrocytes in CNS lesions, suggesting that virus-infected cells migrate through the blood-brain barrier and infect CNS resident cells [14], [15]. While the mechanisms resulting in HAM/TSP development remain unresolved, it has been suggested that Tax expression in CNS cells triggers a strong virus-specific CD8+ (as well as CD4+) T cell response leading to inflammation, myelin loss, and axonal damage [16], [17]. Elevated levels of proinflammatory cytokines (IL-6, IFN-γ, IL-15, IL-1β, TNF-α and IL-12) have been detected in the serum and cerebrospinal fluid (CSF) of patients with HAM/TSP, corroborating the link between HAM/STP development and dysregulated inflammation [18], [19].
It is widely accepted that type I interferon (IFN-α/β) has a negative impact on HIV-1 replication [20], [21], and although few reports have documented the IFN antiviral effects during HTLV-1 infection, type I IFN constitutes a potent anti-retroviral mechanism that affects HTLV-1 replication [22], [23]. In return, HTLV-1 infection of pDCs results in impaired IFN-α production, and correlates with elevated HTLV-1 proviral load in infected individuals [24]. Central to the establishment of an antiviral state is the activation of diverse IFN-stimulated genes (ISGs) which restrict viral replication [25]. Interferon regulatory factors IRF3 and IRF7 play essential roles in the early phase of IFN gene activation [26]. IRF3 is constitutively expressed and is activated by C-terminal phosphorylation by IKKε and TBK1, which promotes transactivation of downstream genes such as IFN-β and IFN-α [27], [28]. In contrast, IRF7 protein is synthesized de novo upon IFN stimulation and contributes to the amplification of the IFN response, via expression of multiple IFN-α subtypes [29]. IRF-driven IFN secretion acts in a paracrine fashion to induce the expression of hundreds of genes through engagement of the IFN receptors and activation of the JAK/STAT signaling pathway, which leads to the development of an antiviral state (reviewed in [30], [31]).
IFN-induced JAK/STAT signaling is negatively regulated at different levels by several cellular factors to control the extent of the antiviral response and limit tissue damage [32], [33]. Suppressor of cytokine signaling 1 (SOCS1) belongs to the SOCS protein family and is induced after virus infection [34]. SOCS1 suppresses IFN signaling by direct binding to phosphorylated type I IFN receptor and active JAK kinase, abrogating phosphorylation of STAT1 [35]. Through its SOCS-Box domain, SOCS1 targets various proteins such as JAK, MAL, p65, Steel, Vav for proteasomal degradation [36], [37], [38]. The SOCS-Box serves as a recruiting platform for the formation of a E3 ligase complex composed of elongin B/C-Cullin 2 and Rbx2 [39], [40]. Thus, SOCS1 initiates and orchestrates the events leading to proteasomal degradation of target proteins [34]. Recently, virus-induced upregulation of SOCS1 protein has emerged as a novel mechanism employed by several viruses to evade the antiviral response [41], [42], [43]. In the present study, global gene expression profiles in CD4+ T lymphocytes were examined in a unique cohort of 30 HTLV-1 infected individuals from the Caribbean basin including ATL, HAM/TSP and asymptomatic carriers (AC) patients. Interestingly, among the many genes dysregulated in HTLV-1 infected patients, SOCS1 was highly expressed in CD4+ T cells from HAM/TSP and AC patients, but not in ATL. Subsequent biochemical analysis demonstrated that HTLV-1-induced SOCS1 expression played a positive role in viral replication through inhibition of the IFN response. SOCS1 directly interacted with IRF3 and promoted its proteasomal degradation in a SOCS-Box dependent manner, thus identifying a novel mechanism of HTLV-1 mediated evasion of the IFN response.
To analyze gene expression profiles of CD4+ T cells isolated from HTLV-1 infected patients, we gathered a unique cohort of 30 HTLV-1 infected individuals from the Caribbean basin, including 11 AC, 7 ATL, 12 HAM/TSP and 8 healthy, non-infected donors (NI) (Table S1). Microarray experiments were performed using the human ImmuneArray cDNA array (UHN Microarray Center, University of Toronto), followed by higher order analysis. About three thousand genes were analyzed with Future Selection Subset/ANOVA on log-transformed data, followed by unsupervised hierarchical clustering on 1039 genes selected by Anova analysis (p<0.01) (Figure 1A). These genes displayed differential expression patterns depending on the type of HTLV-1-associated disease. Unsupervised clustering based on the 1039 genes signature accurately discriminated between NI, HAM/TSP and ATL patients. AC samples however did not separate as an individual cluster, but rather distributed amongst HAM/TSP and ATL samples. Also, two of the HAM/TSP patients and two AC clustered with the NI group, suggesting that the profile of their circulating CD4+ T lymphocytes had not undergone significant variation compared to healthy donors.
Pair-wise correspondance analysis (PCA) was performed on the top 500 genes modulated in HTLV-1-infected versus non-infected samples (p value <0.01, false discovery rate (FDR) = 0.17%) (Figure 1B). PCA identified prevalent expression profiles among the three clinical groups, and confirmed significant class discrimination between non-HTLV-1-infected donors (NI) versus each of the HTLV-1-associated diseases when plotted in two dimensions (Figure 1B). Gene clusters common to each HTLV-1-infected clinical group, and shared within pair-wise comparisons (AC-HAM/TSP, AC-ATL and ATL-HAM/TSP), could be identified and are presented in the adjoining Tables of Figure 1B. For each grouping, genes with a high differential expression are located with quantitative spacing from the center comparator (gene expression of NI group). Specifically, SOCS1 (green square) was identified as a strongly upregulated gene in both HAM/TSP and AC patients, in agreement with the prior observation by Nishiura et al. [44]. Since SOCS1 is known to counter-regulate the anti-viral response, it was selected as a gene of interest for further study.
Efficient HTLV-1 spread must overcome cellular antiviral programs [45]; yet how HTLV-1 evades the host innate immune response is poorly understood. SOCS1 stood out among the many genes identified as having the potential to counteract the innate immune response against HTLV-1. HAM/TSP and AC patients exhibited a greater than two fold increase in SOCS1 gene expression compared to NI individuals (Fisher's test, p value = 0.054 and <0.05, respectively) (Figure 2A). However, no significant difference in mRNA levels of SOCS1 was found between ATL and NI patients, suggesting that SOCS1 expression was upregulated in AC and HAM/TSP. This increase was specific for SOCS1, as SOCS3 mRNA was unchanged in HTLV-1 infected samples compared to control samples (fold change <2) (Figure 2B). Using a separate cohort of patient samples (Table S2), we demonstrated that SOCS1 expression was strongly and positively correlated with HTLV-1 mRNA load in CD4+ T cells of HAM/TSP patients (Pearson's p<0.0001) (Figure 2C).
Since high proviral load is a hallmark of HAM/TSP pathology [46], we investigated the relationship between HTLV-1 replication and SOCS1 gene expression. Initially, the level of SOCS1 mRNA was examined in HTLV-1-carrying T cell lines (MT-2, C8166, MT-4, RMP), control T cell lines (Jurkat and CEM, Figure 2D and Figure S2), as well as PBMCs infected with the HTLV-1 infectious molecular clone pX1M-TM (Figure 2D and 2E). In non-leukemic MT-2 cells that carry an integrated replication-competent provirus and produce infectious HTLV-1 viral particles, a ∼50-fold increase in SOCS1 mRNA was detected, as compared to non-infected CEM and Jurkat cells (<5-fold). In leukemic MT-4 and C8166 cells, which carry a defective provirus, lower levels of SOCS1 mRNA were detected, suggesting that SOCS1 induction required an intact proviral genome (Figure 2D). The RMP cell line derived from an ATL patient which express low amount of HTLV-1 mRNA also displayed lower SOCS1 level (∼10-fold). In order to determine whether de novo HTLV-1 infection induced SOCS1 expression, PBMCs expressing the HTLV-1 infectious molecular clone pX1M-TM (Figure 2E) were analyzed for the level of SOCS1 and HTLV-1 mRNA at different times post-transfection. HTLV-1 RNA expression was determined by amplifying the pX region (tax/rex) of the HTLV-1 proviral genome; HTLV-1 RNA expression was modest between 24 and 72 h (<10-fold), but the viral mRNA load increased sharply at 96 h (50-fold), concomitent with a dramatic increase in SOCS1 gene expression (∼24-fold).
The initial observation that SOCS1 was induced upon HTLV-1 infection prompted us to examine whether SOCS1 also influenced viral replication. To do so, the effect of SOCS1 expression on HTLV-1 provirus replication was examined in the CEM T cells, co-expressing a SOCS1 expression vector together with the HTLV-1 provirus. The level of HTLV-1 mRNA was consistently higher in SOCS1 expressing cells compared to CEM cells expressing HTLV-1 provirus alone (e.g. 55-fold vs. 10-fold at 24 h) (Figure 3A). As a complementary strategy, SOCS1 expression was silenced in MT-2 cells (Figure 3B); siRNAs targeting SOCS1 (siSOCS1(1) siSOCS1(2) and siSOCS1(1)+siSOCS1(2)) inhibited SOCS1 levels by 50, 75 and 90%, respectively. Knock-down of SOCS1 protein expression was confirmed by immunoblot assay (Figure 3B, bottom panel). Real time PCR analysis of the HTLV-1 pX region demonstrated a significant reduction of HTLV-1 mRNA that directly correlated with the decrease in the observed SOCS1 levels (∼27, 56 and 80% decrease, respectively). These data indicate that SOCS1 induction during HTLV-1 infection leads to enhanced HTLV-1 replication.
Since SOCS1 has been shown to negatively regulate type I IFN signaling [34], [42], we sought to investigate the relationship between HTLV-1 infection, type I IFN response and SOCS1 gene expression. First, the profile of type I IFN (IFN-β and IFN-α2) and IFN-stimulated gene expression (IRF7 and CXCL10) was examined in PBMCs expressing the HTLV-1 provirus pX1M-TM (Figure 4A). IFN-β, IFN-α2 and CXCL10 mRNAs were induced (30, 3.5 and 35-fold, respectively) at 24 h post-HTLV-1 transfection, and IRF7 mRNA expression (∼11-fold) was maximal at 36 h. However, mRNA transcripts for all these genes decreased substantially (below 50% of maximal levels) by 48–72 h. At 96 h, when HTLV-1 and SOCS1 gene expression were maximal, no reactivation of antiviral gene transcription was detected (Fig 2D and 4A). We interpret this result as indicating that early after infection, transient stimulation of the antiviral response occurs and restricts de novo HTLV-1 RNA expression; at 72–96 h after infection induction of SOCS1 results in the shutdown of the type I IFN response, thus promoting high HTLV-1 mRNA expression.
IFN-α signaling is initiated by binding to the heterodimeric IFN-α receptor, followed by activation of JAK1 and TYK2 protein kinases, resulting in the phosphorylation of STAT1 and STAT2 [31]. To investigate whether HTLV-1 expression interfered with the type I IFN response, primary PBMCs expressing the HTLV-1 provirus pX1M-TM were treated with IFN-α for 10–120 min to focus on early IFN-triggered phosphorylation events. In control PBMCs, STAT1 and JAK1 phosphorylation was detected at 10 and 20 min post-IFN-α treatment, as determined by immunoblotting with specific antibodies (Figure 4B). However, in PBMCs expressing the proviral clone pX1M-TM, IFN-α-induced phosphorylation of JAK1 and STAT1 was reduced >90 and 70%, respectively (Figure 4B), while total protein levels of JAK1 and STAT1 remained unchanged in control and HTLV-1 expressing PBMCs.
To further characterize the effect of HTLV-1 on antiviral response, PBMCs expressing the HTLV-1 provirus were infected with Sendai virus (SeV) - a strong inducer of the antiviral response - and kinetics of expression of IFN genes was assessed by Q-PCR (Figure 4C, D, E). At 24 h post-transfection, PBMCs had significant HTLV-1 proviral load (∼700 fold higher than control, Figure 4C); thus at this time, PBMCs were infected with SeV (20 HAU/mL) to compare the levels IFN-α2 and IFN-β mRNA in the presence or absence of HTLV-1 provirus (Figure 4D and E). Induction of IFN-β and IFN-α2 mRNA was detected in all PBMCs as early as 3 h post-SeV infection and was sustained up to 12 h (Figure 4D and E). However, in HTLV-1 expressing PBMCs, induction of IFN-β and IFN-α2 mRNA was reduced >60%, relative to the level observed in the absence of HTLV-1 provirus. Decreased levels of IFN-β and IFN-α2 in cells expressing the HTLV-1 provirus were not due to inhibition of SeV replication, as demonstrated by immunoblot for SeV proteins (Figure 4F).
Similarly, knockdown of SOCS1 in MT-2 cells reversed the inhibition of antiviral gene expression imposed by HTLV-1 (Figure 5). Pooled siSOCS1 resulted in increased IFN-β (7.5-fold), ISG56 (2.5-fold), IFN-γ (4-fold) and CXCL10 (∼10-fold) gene expression compared to control siRNA. These results demonstrate that SOCS1 contributes to the inhibition of antiviral responses during HTLV-1 infection.
Many pathogenic viruses strategically antagonize the early innate antiviral defenses in order to maintain viral replication, often inactivating IFN signaling components as part of their immune evasion strategy (reviewed in [45]). Because IRF3 is essential for IFN gene activation, we assessed IRF3 dimerization (as a measure of activation) in PBMCs expressing the HTLV-1 provirus (Figure 6A). In control PBMCs, SeV infection induced IRF3 dimer formation at 3–12 h post-infection, whereas IRF3 dimer formation was not detected in PBMCs expressing the HTLV-1 provirus. Furthermore, IRF3 monomer levels decreased sharply during the course of HTLV-1 replication (Figure 6A). Immunoblot analysis for total IRF3 confirmed that IRF3 levels decreased in a time dependent manner in PBMCs and Jurkat cells expressing the HTLV-1 provirus, a phenomenon not observed in control PBMCs infected with SeV (Figure 6C). IRF-3 was degraded via the proteasomal pathway, as the use of the proteasome inhibitor MG132 prevented HTLV-1 mediated reduction of IRF3 protein level (Figure 6B). This observation demonstrates for the first time that HTLV-1 does not activate IRF3 in PBMCs, but rather prevents the initial steps of type I IFN production by targeting IRF3 for proteasomal degradation. Additionally, IRF3 silencing in Jurkat cells expressing the HTLV-1 provirus resulted in increased HTLV-1 mRNA expression (Figure 6D) – indicating that the degradation of IRF3 by SOCS1 enhances viral mRNA load.
Given that SOCS1 upregulation during HTLV-1 infection inhibits the expression of IFN and ISGs, we sought to investigate the role of SOCS1 in HTLV-1-mediated degradation of IRF3. In HEK293T cells expressing increasing amounts of SOCS1 together with a constant amount of IRF3, SOCS1 expression induced IRF3 degradation in a dose-dependent manner (Figure 7A). RT-PCR analysis with specific IRF3 primers showed that the level of IRF3 mRNA remained unchanged, indicating that SOCS1 had no effect on IRF3 gene expression (Figure 7A). Moreover, SOCS1 silencing in HTLV-1 infected MT-2 cells restored endogenous IRF3 expression (Figure S3). Interestingly, the addition of the proteasome inhibitors lactacystin (Figure 7B) or MG132 (data not shown) prevented IRF3 degradation in the presence of SOCS1. Furthemore, co-immunoprecipitation experiments demonstrated that SOCS1 physically interacted with IRF3 (Figure 7C), indicaing that IRF3 degradation was triggered by physical association with SOCS1.
SOCS1 induces degradation of target proteins by recruiting Elongin B/C to its SOCS-Box domain, leading to the formation of an E3 ubiquitin ligase complex able to modify substrate proteins with K48-linked ubiquitin chains. To confirm that SOCS1-mediated IRF3 degradation required E3 ligase complex activity, increasing amounts of a SOCS1 deletion mutant lacking the SOCS-Box - SOCS1-ΔB/C-Box - was expressed together with a constant amount of IRF3 (Figure 7D); SOCS1-ΔBC did not induce IRF3 degradation at any concentration (compare Figures 7D and 7A).
Proteasome-mediated degradation requires the addition of K48-polyubiquitin chain to the target protein; exogenous addition of ubiquitin mutated in its ability to link K48-polyubiquitin chains (ubiquitin-K48R, which contains a single K48R point mutation, or Ubi-KO, which contains no lysines) prevented IRF-3degradation, while HEK293 cells expressing exogenous SOCS1 readily degraded IRF3 (Figure 7E). In addition, IRF3 turnover was completely reversed in the presence of a 10-fold excess of HA-Ub K48R or HA-Ub KO (Figure 7E), thus confirming that proteosome-mediated IRF3 degradation by SOCS1 requires recruitment of Elongin B/C E3 ligase machinery and is dependent on K48-polyubiquitin chain formation.
The complexity of gene expression dysregulation in ATL or HAM/TSP diseases has been highlighted in a number of gene expression profiling [47], [48] and protein profiling studies [49]. The present study however represents the first comparative genome-wide array analysis to establish gene expression profiles for HTLV-1-associated disease states. With a unique cohort of 30 HTLV-1-infected individuals from the Caribbean basin and a custom ImmuneArray [50], we identified ∼1039 significant immune-related genes that were differentially regulated in CD4+ T cells from 11 AC, 7 ATL, 12 HAM/TSP patients, compared with CD4+ T cells from 8 NI donors from the same geographical region. Clear clinical discrimination was observed between the ATL, HAM/TSP and NI patients, both by unsupervised hierarchical cluster and principal component analysis. Our analysis revealed that the gene expression profile in ATL cells was clearly distinct from healthy CD4+ T cells, although similarities in gene expression patterns were observed between HAM/TSP samples and NI controls. This difference between HAM/TSP and ATL CD4+ T cells likely reflects numerous alterations in gene expression that occur during ATL transformation [12]. In contrast, evolution to HAM/TSP does not involve cellular transformation, but rather is characterized by a high HTLV-1 proviral load and the establishment of a pro-inflammatory microenvironment due to cytokine/chemokine production of infected and bystander immune cells. It is possible that T lymphocytes derived from early-stage HAM/TSP patients have a profile similar to healthy cells and that gene expression changes are observed only at later stages of the disease, an interesting hypothesis that needs to be investigated further. Interestingly, AC patients did not cluster as an individual group but rather distributed amongst NI, ATL and HAM/TSP patients, suggesting that extensive analysis of the genes modulated in NI, HAM/TSP and/or ATL groups may help to identify candidate genes important for early diagnosis of HTLV-1 diseases. Here, the major cellular pathways identified involved cell adhesion (CXCR4, CD2, CD63), antimicrobial defense (KLRB1, SPN, SELPLG), innate immune signaling (SOCS1, TRAF3, AIM2, TLR2, IKBKG, STAT3), antigen presentation (TRA alpha locus), and chemotaxis (CCL14, SPN, CCL13) thus supporting the idea of a global disruption of the immune system during HTLV-1 infection.
Among the many genes modulated during HTLV-1 infection, the suppressor of the interferon signaling - SOCS1 - was upregulated in HAM and AC patients but not in ATL. This observation is in agreement with a previous report published by Nishiura et al. demonstrating that SOCS1 mRNA levels were increased in HAM/TSP patients compared to NI [44]. We now demonstrate that CD4+ T cells from HAM/TSP and AC patients express increased levels of SOCS1 which strongly correlates with HTLV-1 mRNA load. Since HAM/TSP patients are characterized by a very high proviral load, we hypothesized that SOCS1 upregulation in HAM/TSP may represent an immune evasion strategy used by HTLV-1 to dampen the early IFN antiviral response. Indeed, in PBMCs expressing a HTLV-1 infectious molecular clone, and in cell lines harboring an intact HTLV-1 provirus, high levels of SOCS1 gene expression correlated with high levels of HTLV-1 transcription. Increasing HTLV-1 proviral expression blocked expression of type I IFN genes such as IFN-β, IRF7, IFN-α2, as well as the IFN-γ stimulated chemokine gene CXCL10, with maximal inhibition observed when HTLV-1 and SOCS1 gene expression levels were coordinately elevated. Furthermore, depletion of SOCS1 using siRNA decreased HTLV-1 replication and restored the type I IFN response.
IFN-α/β is known to have a negative impact on retrovirus replication. Although few studies have reported its effect on HTLV-1, type I IFN constitutes a potent anti-retroviral mechanism that limits HTLV-1 replication [22], [23]. Moreover, clinical studies using IFN-β therapy in HAM/TSP patients have demonstrated benefits in reducing HTLV-1 mRNA load and the number of pathogenic CD8+ T cells, as well as minimizing disease progression during therapy [51]. Accumulating evidence indicates that HTLV-1 possesses evasion mechanisms to counteract type I IFN signaling: for example, HTLV-1 down-regulates JAK-STAT activation by reducing phosphorylation of Tyk2 and STAT2, possibly through a Gag- or Pr-mediated mechanism [52]; and Tax further negatively modulates IFN-α-induced JAK/STAT signaling by competing with STAT2 for CBP/p300 coactivators [53].
SOCS1 is a cytokine-inducible intracellular negative regulator that inhibits type I and II IFN signaling by triggering the degradation of various components of the JAK-STAT cascade (reviewed in [32], [54]). SOCS1 can also be induced during virus infection and plays a positive role in viral replication [55], [56], [57]. SOCS1 is induced during virus infection and binds directly to the type I IFN and/or II IFN receptors to suppress IFN signaling, thereby preventing chronic inflammation. However, SOCS1 could be subverted to enhance viral replication via untimely inhibition of the IFN response. SOCS1 induction may be a direct result of viral protein activity. Bioinformatics analysis of the SOCS1 promoter region reveal the presence of CRE, AP-1 and NF-κB binding regions, suggesting the possible involvement of HTLV-1 Tax in the induction of SOCS1 expression (data not shown). Another possibility is that SOCS1 transcriptional activation is not directly regulated by viral proteins, but rather by recognition of viral RNA and downstream signaling events. For instance, Potlichet et al. reported that Influenza A virus suppresses the antiviral response by inducing SOCS1 and SOCS3 via TLR3-independent but RIG-I/IFNAR dependent pathways [43]. Moreover, IFN-γ gene expression in CD4+ T cells from HAM/TSP patients is elevated as compared to ATL or AC patients. Constitutive induction of IFN-γ may also augment SOCS1 expression, and thus increase HTLV-1 replication.
SOCS proteins exert their negative effect by promoting the ubiquitination and proteosomal degradation of key proteins involved in cytokine signaling pathways: MAL in Toll like receptor 4 signaling (TLR4), JAK2 in IFN-γ mediated signaling and NF-κB p65/RelA are all known targets of SOCS1 [38], [58], [59]. Here, we identified IRF3 as an important target for SOCS1-induced proteasomal degradation that impacts the early type I IFN antiviral response. IRF3 is ubiquitously expressed in the cytoplasm and is activated in response to viral infection, triggering IFN-β and other early ISGs expression, thus initiating the antiviral response. To counter type I IFN, many viruses have evolved strategies to interfere with IRF3 activation as an efficient means to limit IFN-β production [26], [60]. Interference of IRF3 activation also dampens the second wave of IFN signaling, including production of IFN-α. The mechanisms of IRF3 antagonism vary, and include inhibition of IRF3 phosphorylation, nuclear translocation, or transcription complex assembly as well as down-regulation of IRF3 by ubiquitin-mediated degradation. In this context, bovine herpes virus 1 infected cell protein 0 (bICP0) has been shown to act as an E3 ligase and promote IRF3 degradation in a proteasome-dependent manner, thus inhibiting the IFN response [61].
The interaction between SOCS1 and IRF3 during HTLV-1 infection promotes proteasome-mediated degradation of IRF3 and thus abrogates early IFN antiviral signaling. SOCS1-dependent IRF3 degradation required the elongin B and C binding sites within SOCS1 and K48-linked polyubiquitination of IRF3. Indeed, the SOCS box-mediated function of SOCS1 is chiefly exerted via its ubiquitin ligase activity [62] and biochemical binding studies have shown that the SOCS box interacts with the elongin B/C complex, a component of the ubiquitin/proteasome pathway that forms an E3 ligase with Cul2 (or Cul5) and Rbx-1 [40], [58]. Thus, SOCS1 serves as an adaptor to bring target proteins to the elongin B/C-Cullin E3 ligase complex for ubiquitination. Although we show from our current experiments that SOCS1 directly mediates K48-linked ubiquitination of IRF3, further studies are required to elucidate the details of SOCS1-mediated IRF3 ubiquitination, as well as the mechanisms of regulation of SOCS1 during HTLV-1 infection.
The present study reveals a novel mechanism of viral evasion of the IFN response in HTLV-1 infected T lymphocytes – the consequence of which can be directly related to the efficiency of HTLV-1 replication in patients suffering from HAM/TSP. Future studies are required to elucidate putative alternate consequences of SOCS1 upregulation in T cells [63], as well as the effect of HTLV-1 induced SOCS1 expression in other relevant viral reservoirs such as dendritic cells and astrocytes [64], [65]. Collectively, SOCS1-mediated degradation of IRF3 during HTLV-1 infection has substantial implications in the framework of known HTLV-1 pathobiology and as such opens new avenues of exploration for designing effective therapeutic strategies.
Blood samples from HTLV-1 infected patients and non-infected (NI) donors were obtained from the Centre Hospitalier Universitaire de Fort-de-France in Martinique and Institut Pasteur de Cayenne in French Guyana. Patients suffering from ATL, HAM/TSP or HTLV-1 asymptomatic carriers were recruited according to World Health Organization (WHO) criteria. According to the French Bioethics laws, the collection of samples from HAM/TSP, ATL, AC and NI has been declared to the French Ministry of Research and the study was reviewed and approved by the CPP (Comité de Protection des Personnes) Sud-Ouest/Outre-Mer III, as well as the ARH (Agence Régionale de l'Hospitalisation) from Martinique. Because the protocol is non-interventional (e.g. blood samples collected for routine health care with no additional samplings or specific procedures for subjects), no informed consent was provided by the patient, as stated by the French Public Health code and therefore the study was conducted anonymously. Clinical collection of samples for research purpose are stored at the Centre de Ressources Biologiques de Martinique (CeRBiM). The CeRBiM database has been approved by the CNIL (Commission nationale de l'informatique et des libertés). Leukophoresis from healthy donors were also obtained at the Royal Victoria Hospital, Montreal, Quebec, Canada. Informed consent were written and provided by study participants in accordance with the Declaration of Helsinki. The study was reviewed and approved by the Royal Victoria Hospital, the Jewish General Hospital, and McGill University Research Ethics Committee (REC) board of the SMBD-Jewish General Hospital.
In total, we selected for study 12 HAM/TSP, 11 asymptomatics (AC), 7 ATL and 8 not infected individuals (NI). The diagnosis of the 7 ATL cases included in patient cohort number 1 respected the international consensus recently published by Tsukasaki et al. [7]. Diagnostic criteria for ATL included serologic evidence of HTLV-1 infection, and cytologically or histologically proven T cell malignancy. Six ATL cases were classified as acute leukemia type on the basis of leukemic manifestations, with >5% typical ATL cells in the peripheral blood, and immunologically confirmed mature CD4+ T cell phenotype. One case (HISS0023) was a lymphoma type, with <5% circulating abnormal cells, the ATL cell phenotype and clonal integration of HTLV-1 being confirmed on lymph node tissue. Diagnosis of HAM/TSP was in accordance with WHO criteria [66], which comprise (1) slowly progressive spastic paraparesis with symmetrical pyramidal signs, (2) disturbance of bladder function, (3) no radiologic evidence of significant spinal cord compression, and (4) intra-thecal synthesis of anti−HTLV-1 antibodies. The asymptomatic HTLV-1 carriers did not display any neurological symptoms (Tables S1 and S2). PBMCs were isolated by centrifugation (400 g at 20°C for 25 min) on a Ficoll-Hypaque gradient (GE Healthcare Bio-Sciences Inc., Oakville, Canada). CD4+ T lymphocytes were isolated using a negative selection CD4 enrichment cocktail with the high-speed autoMACS system (Miltenyi Biotec) according to the manufacturer's instructions. In all cases, the purity of CD4+ T lymphocytes was between 90 and 95% as determined by flow cytometry. Cells were pellet and kept at −80°C until all samples were ready for RNA extraction.
The HTLV-1-carrying T cell lines MT-2, MT-4, C8166, RMP and the HTLV-1-negative T cell lines CEM and Jurkat were used for experiments. MT-2, MT-4 and C8166 cells are derived from umbilical cord blood lymphocytes after cocultivation with leukemic cells from ATL patients [67]. MT-2 cells are reported to have integrated at least fifteen copies/cell, including defective types, of HTLV-1 proviral DNA whereas C8166 cells have only one copy of proviral DNA integrated in the genome [68], [69]. The interleukin (IL-2)-independent RMP cell line is derived from CD4+ T cell of a patient with acute ATL. All cell lines were maintained in RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum, 100 U/ml penicillinG, and 100 µg/ml streptomycin. HEK293T cells were used for transient transfection and were maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% heat-inactivated fetal bovine serum, 100U/ml penicillin G, and 100 µg/ml streptomycin.
For proteasome inhibitor treatment, MG132 (Sigma-Aldrich) or Lactacystin (Boston Biochem) were used at 5 and 10 µM, respectively. Sendai virus CANTELL strain (SeV) was obtained from Charles River Laboratory (North Franklin, CT). Cells were infected with SeV at 20 hemagglutinating units (HAU) per 106 cells in serum-free medium supplemented with 10% heat-inactivated fetal bovine serum 2 h postinfection and harvested for whole cell extracts or RNA extraction at indicated times.
The HTLV-1 proviral clone pX1M-TM was kind gift from Dr David Derse (National Cancer Institute-Frederick, Frederick, USA). Myc-tagged SOCS1 full length and the deletion mutant SOCS1-ΔBCBox (amino acids 174–183) were kind gifts from Dr. Ferbeyre Gerardo (Departement de Biochimie, Universite de Montreal, Canada). Ha-Ub-K48R and Ha-Ub-KO were kind gifts from Dr. Zhijan Chen (Department of Molecular Biology, University of Texas Southwestern Medical Center, Dallas, Texas). Plasmid encoding for Flag-tagged IRF3 full length was decribed previously [27].
Total RNA was extracted using Trizol Reagent (Invitrogen) or RNeasy kit (Qiagen) according to the manufacturer's instructions. The RNA integrity and purity was assessed with the Agilent 2100 Bioanalyzer (Agilent Technologies). Total RNA was amplified using the MessageAmp II mRNA kit (Ambion, Austin, USA). Sample and universal human RNA probes (Stratagene) for microarray hybridization were prepared by labeling the amplified RNA with Cy5 or Cy3, respectively, by reverse transcription, and hybridizing the labeled cDNA on the CANVAC (http://www.canvac.ca/) human Immunoarray version 2 manufactured by the Microarray Center (UHN, Toronto, ON, Canada) containing 7256 duplicate spots representing 3628 expressed sequence tags (ESTs). Details of the labeling and hybridization procedures can be obtained at http://transnet.uhnres.utoronto.ca.
Microarrays were scanned using Scanarray Express Scanner (Packard Biosciences) or the Axon 4000B scanner at 10-µm resolution. Array images were inspected visually for poor quality spots and flagged for omission. Quantified raw data was acquired with QuantArray version 3 and saved as quantarray text files. The quantified raw data were managed and pre-processed in GeneTraffic (Iobion Informatic). Following background correction and removal of genes where both channels were less than 100 or represented by less than 90% of the samples and polished data was generated by normalization by Lowess sub-grid. The final data array was analyzed using JExpress Pro software (http://www.molmine). To establish differentially expressed genes, multi-class analysis was performed by one-way ANOVA on Log2 fold change (Log2Fc) data for ATL, HAM/TSP, AC and NI groups. Genes with a p value ≤0.01 were selected as significant (1039 total). Visualization was produced by unsupervised clustering of the 1039 genes using Pearson correlation parameters. Pair wise correspondance analysis (PCA) was performed on the first 500 genes by Future Subset Selection (FSS) t-test. Genes were selected based on false discovery rate (FDR) according to the Benjamini/Hochberg (BH) methods. Gene annotations were gathered using manual searches in NCBI as well as the ontology tools DAVID (http://david.abcc.ncifcrf.gov/) and BioRag (Bioresource for array genes, http://www.biorag.org). Fold change (Fc) for each gene was calculated as 2(Log2X-Log2NI), where Log2 X represents the Log2 (Fc) for either ATL, AC or HAM and Log2 NI represents the Log2 (Fc) for NI. Microarray data have been deposited in the NCBI Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/). The results of the microarray experiment were confirmed by quantitative PCR (Q-PCR) on 47 genes chosen on the basis of a fold change of at least 2-fold, with RNA from 3 patients per group used for validation. A strong correlation between the average fold-change determined by microarray and the average of the qPCR results was observed, with 25/47 genes having a Pearson correlation value of at least 0.6 (Figure S1A, B, C) and 18/47 genes with a value of at least 0.9 (Figure S1A).
Leukophoresis from healthy donors were obtained at the Royal Victoria Hospital, Montreal, Quebec, Canada. PBMCs were isolated by Ficoll-Hypaque gradient (GE Healthcare Bio-Sciences Inc., Oakville, Ontario, Canada) and activated for 4 days with 2 µg/ml of phytohemagglutinin-P (PHA-P) (Sigma Aldrich) and 50 U of interleukin 2 per ml (IL-2) (PBL Biomedical Laboratories). 5 µg of pX1M-TM was pulsed into 10×106 cells PBMCs in a 0.4-cm cuvette using a Gene Pulser II (Bio-Rad Laboratories) set at 0.25 kV and 0.95 µF. Cells were plated in six-well plates in complete medium and collected at indicated times for whole cell extracts or RNA extraction.
Validation of selected target genes was performed by relative quantification PCR (RQ-PCR) in 9 samples (3 NI, 3 ATL, 3 HAM). A total of 2 µg of amplified RNA from uninfected and HTLV-1-infected samples was converted to cDNA using the High Capacity cDNA Archive Kit (Applied Biosystems, Foster City, CA) according to the manufacturer's protocol. cDNA was amplified using SyBR Green I PCR master mix (Roche Applied Science, Germany) or TaqMan Universal PCR Master Mix (Applied Biosystems, Foster City, CA, USA). Real-time PCR primers were designed using the primer3 website (primer3_www.cgiv. 0.2) and listed in Supporting information (Table S3). Some predesigned primers and probe sets from TaqMan (Applied Biosystems) were also used and listed in Table S3. Data were then collected using the AB 7500 Real-Time PCR System (Applied Biosystems, Foster City, CA) and analyzed by comparative CT method using the SDS v1.3.1 Relative Quantification (RQ) Software where ddCT = dCT(Sample) – dCT (non-infected), dCT (Sample) = CT (Sample) - CT (GAPDH) and dCT (non-infected) = CT (non-infected) - CT (GAPDH).
Cells destined for immunoblotting were washed with PBS and lysed in lysis buffer (0.05% NP-40, 1% glycerol, 30 mM NaF, 40 mM β-glycerophosphate, 10 mM Na3VO4, 10 ng/ml of protease inhibitors cocktail (Sigma Aldrich, Oakville, Ontario, Canada). The protein concentration was determined by using the Bradford assay (Bio-Rad, Mississauga, Canada). Whole-cell extracts (30 µg) were resolved by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) in a 10%-acrylamide gel and transferred to a nitrocellulose membrane (Bio-Rad, Mississauga, Canada). Membranes were blocked in 5% nonfat dried milk in Tris-buffered saline (TBS) plus 0.1% Tween 20 for 1 h at room temperature. Membranes were then probed overnight with antibodies against Stat 1 phosphorylated (Tyr701) (1∶1000; Cell Signaling) and non-phosphorylated forms (p91) (1 µg/ml; Santa Cruz); phosphorylated Jak1 (Tyr 1022/1023) (1∶1000; Cell Signaling); total Jak 1 (1 µg/ml Santa Cruz); SeV (1∶10,000); SOCS1 (1 µg/ml; Zymed laboratories), IRF3 (1∶10,000; IBL, Japan) in 5% bovine serum albumin and PBS at 4°C. Incubation mixtures were washed in TBS-0.05% Tween 20 five times for a total of 25 min. Following washes, the membrane was incubated with peroxidase-conjugated goat anti-rabbit or anti-mouse antibody (KPL, Gaithersburg, MD) at a dilution of 1∶5,000 for 1 h at room temperature. Following the incubation with the secondary antibody, membranes were washed again (5 times, 5 min each) and then visualized with an enhanced chemiluminescence (detection system as recommended by the manufacturer (ECL; GE Healthcare Bio-Sciences Inc., Oakville, Ontario, Canada).
Native-PAGE was conducted as described [70]. Briefly, 10 g WCE in native sample buffer (62.5 mMTris-HCl, pH 6.8, 15% glycerol, and bromophenol blue) were resolved by electrophoresis on a 7.5% acrylamide gel (without SDS) pre-runned for 30 min at 40 mA using 25 mMTris and 192 mM glycine, pH 8.4, with and without 1% deoxycholate in the cathode and anode chamber, respectively. After transferred into nitrocellulose membrane, IRF3 monomers and dimers were detected by immunoblot using an IRF3 anti-NES antibody (1∶10, 000, IBL, Japan).
HEK293 cells (1×106 cells/60-mm dish) were transiently transfected with equal amounts (5 µg) of IRF3 and MYC-tagged SOCS1 expression plasmids by using calcium phosphate precipitation method. Cells were harvested 24 h post-transfection, washed with 1 X phosphate-buffered saline (PBS), and lysed in a 1% Triton X-100 lysis buffer (20 mM Tris-HCl, pH 7.5, 200 mM NaCl, 1% Triton X-100, 10% glycerol, 40 mM β-glycerophosphate, 0.1% protease inhibitor cocktail, 1 mM phenylmethylsulfonyl fluoride, 1 mM Na3VO4, 5 mM NaF, 1 mM dithiothreitol). Immunoprecipitations were performed by incubating WCE (300 µg) with 1 µg of anti-MYC (9E10; Sigma-Aldrich, St. Louis, MO) or 1 µg of antiserum directed against IRF3 (rabbit polyclonal antibody, IBL, Japan) coupled to 50 µl of A/G Plus-agarose beads (Santa Cruz Biotechnology, Santa Cruz, CA) overnight at 4°C with constant agitation. Immunocomplexes were washed at least 3 times in lysis buffer eluted by boiling beads in 5 volumes SDS-PAGE sample buffer. The proteins were fractioned on 10% SDS-PAGE, transferred to nitrocellulose membrane and analyzed by immunoblot assay using anti-MYC (Sigma-Aldrich) or anti-IRF3 (IBL, Japan) antibodies.
Control and SOCS1-specific RNA interference sequences were described previously [71], [72]. SOCS1 protein was knocked down using siSOCS (1), siSOCS (2) or a pool of the two siRNAs (siSOCS (1) + siSOCS (2)). siRNAs were pulsed into MT-2 cells in a 0.4-cm cuvette using a Gene Pulser II (Bio-Rad Laboratories) set at 0.25 kV and 0.95 µF. Cells were plated in six-well plates in complete medium, washed 4 and 12 h later and collected at 72 h post-transfection. RNA extinction efficiency was demonstrated by real time PCR and immunoblot assay.
Data are presented as the mean ± standard error of the mean (SEM). Statistical significance for comparison of gene expression was assessed by an unpaired Student's t test, with the expection of Figure 4, panels D and E where a two-way ANOVA with Bonferroni post-test was used. Analyses were performed using Prism 5 software (GraphPad). Statistical significance was evaluated using the following p values: p<0.05 (*), p<0.01 (**) or p<0.001 (***).
A list of accession numbers for genes and proteins mention in this study are listed in Table S4.
|
10.1371/journal.ppat.1006508 | microRNA dependent and independent deregulation of long non-coding RNAs by an oncogenic herpesvirus | Kaposi’s sarcoma (KS) is a highly prevalent cancer in AIDS patients, especially in sub-Saharan Africa. Kaposi’s sarcoma-associated herpesvirus (KSHV) is the etiological agent of KS and other cancers like Primary Effusion Lymphoma (PEL). In KS and PEL, all tumors harbor latent KSHV episomes and express latency-associated viral proteins and microRNAs (miRNAs). The exact molecular mechanisms by which latent KSHV drives tumorigenesis are not completely understood. Recent developments have highlighted the importance of aberrant long non-coding RNA (lncRNA) expression in cancer. Deregulation of lncRNAs by miRNAs is a newly described phenomenon. We hypothesized that KSHV-encoded miRNAs deregulate human lncRNAs to drive tumorigenesis. We performed lncRNA expression profiling of endothelial cells infected with wt and miRNA-deleted KSHV and identified 126 lncRNAs as putative viral miRNA targets. Here we show that KSHV deregulates host lncRNAs in both a miRNA-dependent fashion by direct interaction and in a miRNA-independent fashion through latency-associated proteins. Several lncRNAs that were previously implicated in cancer, including MEG3, ANRIL and UCA1, are deregulated by KSHV. Our results also demonstrate that KSHV-mediated UCA1 deregulation contributes to increased proliferation and migration of endothelial cells.
| KS is the most prevalent cancer associated with AIDS in sub-Saharan Africa, and is also common in males not affected by AIDS. KSHV manipulates human cells by targeting protein-coding genes and cell signaling. Here we show that KSHV alters the expression of hundreds of human lncRNAs, a broad class of regulatory molecules involved in a variety of cellular pathways including cell cycle and apoptosis. KSHV uses both latency proteins and miRNAs to target lncRNAs. miRNA-mediated targeting of lncRNAs is a novel regulatory mechanism of gene expression. Given that most herpesviruses encode miRNAs, this mechanism might be a common theme during herpesvirus infections. Understanding lncRNA deregulation by KSHV will help decipher the important molecular mechanisms underlying viral pathogenesis and tumorigenesis.
| Kaposi’s sarcoma-associated herpesvirus (KSHV) is an opportunistic human oncovirus, which causes Kaposi’s sarcoma (KS), Primary Effusion lymphoma (PEL) and Multicentric Castleman’s disease (MCD) in immunocompromised individuals, primarily AIDS patients and organ-transplant recipients [1]. KSHV uses the lytic mode of replication for spread of infection, and latency for persistence in the host. All tumor cells isolated from KS patients test positive for latent viral episomes [1]. Latent KSHV expresses only 10% of its 140-kb dsDNA genome, encoding primarily four latency proteins (Kaposin, vFLIP, vCyclin and LANA) and 25 mature miRNAs [1]. miRNAs are 21–23 nt long non-coding RNAs that recognize target mRNAs using 7 bp ‘seed sequences’ and silence them (see [2] for review). To identify the means by which KSHV causes tumors, KSHV latency proteins and miRNAs have been studied extensively [1]. Ribonomics approaches to identify targets of KSHV miRNAs have focused exclusively on mRNAs [3, 4].
Recently, lncRNAs have emerged as important regulatory molecules in cancer [5]. LncRNAs play a variety of regulatory roles in both the cytoplasm and nucleus [6, 7]. This group includes all RNA molecules longer than 200 nt with no apparent coding potential, and they have diverse functions ranging from acting as a scaffold, sponge/decoy or guide aiding in cell-signaling [6, 8]. Owing to their diversity, over 95% of the lncRNAs remain uncharacterized. Disease association is a starting point for identifying and characterizing lncRNAs with important regulatory roles. Using this approach with different cancer types, oncogenic lncRNAs such as MALAT-1, ANRIL, UCA1, and tumor suppressor lncRNAs like Gas-5 and MEG3 have been functionally characterized [5]. Another important group of disease-relevant lncRNAs includes those involved in the innate immune response following viral or bacterial infections [9]. A few studies have addressed the roles of host lncRNAs during viral infections, for example HULC (Hepatitis-B) and NRON (HIV) [10]. However, the question of whether viruses manipulate specific host lncRNAs to their advantage remains largely unexplored. Understanding deregulation of specific host lncRNAs, especially cancer-related lncRNAs by persistent oncoviruses, such as the γ-herpesviruses, would shed light on how these viruses drive oncogenesis.
Regulatory cross-talk is known to occur between miRNAs and lncRNAs, at multiple levels. LncRNAs like BIC1 and H19 act as precursors for miRNAs [11, 12] and lncRNAs such as HULC and CDR1-AS act as sponges for miRNAs [13, 14]. Conversely, human miRNA miR-9 represses the expression levels of the lncRNA MALAT1 [15]. Work from the Steitz laboratory demonstrated that the viral lncRNAs HSUR1 and HSUR2, encoded by Herpesvirus Saimiri, act as sponges for cellular miR-16, miR-142-3p and miR-27 and thereby silence some of these miRNAs in T-lymphocytes, suggesting that γ-herpesviruses can utilize virus lncRNAs to target host miRNAs[16]. Conversely, whether herpesvirus miRNAs can target and downregulate host lncRNAs remains an open question.
In this study, we demonstrate that latent KSHV infection of endothelial cells alters the host lncRNA profile. We provide evidence that KSHV deregulates hundreds of host lncRNAs including many cancer-associated lncRNAs such as UCA1, ANRIL and MEG3 in both a miRNA dependent and independent manner. Furthermore, KSHV appears to manipulate the host lncRNAs to favor proliferation and migration of latently infected endothelial cells.
Previously, we identified the mRNA targetome of viral miRNAs in PEL cells by High Throughput Sequencing-Crosslinking Immuno Precipitation (HITS-CLIP) analysis of the Ago protein [3]. The PEL cell lines we studied were BC-3 and BCBL-1, which are KSHV positive B-cell lines. We reanalyzed the HITS-CLIP data for enriched lncRNAs and compared our results with a similar reinvestigation of Ago PAR-CLIP data from lymphoblastoid cell lines infected with Epstein-Barr Virus (EBV) [17], a related γ-herpesvirus that causes cancer. We found that approximately 357 and 750 lncRNAs were a part of the KSHV and EBV miRNA targetome, respectively, and 64 lncRNAs were potentially targeted by miRNAs from both viruses (S1 Table).
We aimed to determine the effect of latent KSHV infection on the lncRNA expression profile of endothelial cells and specifically question whether KSHV encoded miRNAs targeted endothelial lncRNAs. To address these questions, we used Telomerase Immortalized Vein Endothelial (TIVE) cells, an in vitro model system to study KS [18]. We performed lncRNA expression profiling on latently infected TIVE cells harboring either the wt-KSHV or Δcluster-KSHV [19, 20], in which a region containing 10 of the 12 miRNA genes is deleted, and used the lncRNA profile of mock-infected TIVE cells as reference. The KSHV latency-associated region of the wt and mutant bacmid backbones used for this experiment is shown in Fig 1A. The profiling analysis revealed that wt-KSHV and Δcluster-KSHV infections deregulate 858 and 2372 host lncRNAs, respectively (Table 1), indicating that latent KSHV infection globally affects lncRNA expression. The higher count of deregulated lncRNAs in Δcluster-KSHV infection is likely due to increased spontaneous reactivation rate in the absence of viral miRNAs [20, 21]. The differentially expressed lncRNAs are listed in S2 Table. We grouped the deregulated lncRNAs into three categories based on a cut-off of fold change ≥ 2.0: upregulated, downregulated and rescued. We defined rescued genes as those that were downregulated in wt-KSHV-infected cells compared to mock, and were upregulated in Δcluster-KSHV-infected cells compared to wt-infected cells. We validated using qRT-PCR two downregulated lncRNAs, two upregulated lncRNAs, and three rescued lncRNAs that were identified from the microarray analysis (S1 Fig). We identified 126 candidates in the rescued category, which are putative direct targets of viral miRNAs (Fig 1B).
Based on lncRNA localization data from HUVEC cells [22], at least 9 of the 126 putative lncRNA targets of viral miRNAs we identified are exclusively nuclear localized, and 32 of them are partially nuclear localized. It is important to note that the localization information was available for only 72 out of the 126 rescued lncRNAs. Similarly, the 357 lncRNAs identified from Ago HITS-CLIP of PEL cells include nuclear resident lncRNAs such as ANRIL (CDKN2B-AS1) and MALAT-1. Moreover, several of the uncharacterized candidates of these 357 lncRNAs may be nuclear localized. miRNA-mediated regulation of nuclear localized lncRNAs seemed paradoxical at the outset, as mature miRNAs and RISCs including the Ago family proteins are believed to reside and function in the cytoplasm. Recently, several groups showed that Ago-2 complexes can be present in the nuclei of different cell types [23, 24]. Moreover, studies in Hodgkin’s lymphoma lines identified that several lncRNAs co-isolate with Ago protein [25]. To determine whether KSHV miRNAs could regulate nuclear lncRNAs, we investigated the nuclear/cytoplasmic distribution of viral miRNAs and Ago-2 in PEL cells.
We fractionated BCBL-1 cells into nucleus and cytoplasm and analyzed the distribution of KSHV miRNAs using stem-loop RT-qPCR, which amplifies mature miRNAs but not their precursors. Mature KSHV miRNAs were found in both the cytoplasmic and nuclear fraction (Fig 2A). It is important to note that a cellular miRNA hsa-miR-16 is also distributed between the nucleus and the cytoplasm (Fig 2A), and such partial nuclear localization of mature miRNAs has been previously reported in other cell lines [26, 27]. We probed the fractions for Ago-2 using western blotting (Fig 2B). Calnexin, an ER resident, was used as a control to ensure that the nuclear preparations were free of endoplasmic reticulum. A significant fraction of Ago-2 was localized in the nucleus of BCBL-1 cells. This observation is consistent with a study by Gagnon et al., which reported comparable amounts of Ago2 in the nucleus and cytoplasm of multiple cell lines including HeLa, T47D, A549 and fibroblasts [23]. These results were confirmed using immunofluorescence analysis (IFA) of Ago-2 in isolated BCBL-1 nuclei by confocal microscopy and 3D-reconstruction. The images in Fig 2C show Ago-2 in all planes of view (XY, YZ and ZX) with and without DAPI, and it is evident that Ago-2 is present inside the BCBL-1 nuclei. We observed similar results with IFA performed on KSHV-infected TIVE cells (Movie S1). Thus, we concluded that Ago2 and viral miRNAs are present in the nuclei of infected cells, and miRNAs could potentially interact via Ago2 with nuclear lncRNAs.
Of the 126 rescued lncRNAs identified based on transcriptional profiling, 98 contained seed sequence matches for at least one KSHV miRNA. Repeated sampling of 126 sequences from randomly generated DNA sequences, controlling for lncRNA length, revealed that the presence of KSHV miRNA seed matches in 98 out of 126 lncRNAs is statistically significant (p-value = 5.79 x 10−8, one-sided t-test). These data provide genetic evidence for miRNA-dependent deregulation of host lncRNAs during KSHV latency.
In order to validate that KSHV miRNAs can target host lncRNAs in the absence of KSHV infection, we chose four lncRNAs from the 98 containing seed sequences, and transfected pools of corresponding miRNA mimics into uninfected TIVE cells. The pools of mimics transfected were specific to the seed matches that those lncRNAs contained. Their respective mimic pools when compared to control mimic significantly knocked down all four lncRNAs tested, demonstrating that the viral miRNAs target lncRNAs in the absence of KSHV infection (Fig 3A).
The miRNA-dependent downregulation of lncRNAs could result from direct targeting of lncRNAs by miRNAs, or from an indirect secondary effect (e.g., through miRNA-mediated downregulation of transcription factors). To investigate direct interaction between KSHV miRNAs and lncRNAs, we performed miRNA pull-down experiments in TIVE-Ex-LTC cells. TIVE-Ex-LTC cells were derived from TIVE cells (see Materials and Methods), but grow much faster compared to TIVE cells. KSHV negative TIVE-Ex-LTC cells were transfected with biotinylated miRNA mimics for either miR-K12-6-5p, miR-K12-11* or siGLO (lacks biotin) and pull-down experiments were performed 24 h post-transfection. It is important to note that the mimics are dsRNAs that require loading into the RISC in order to bind their targets. Loc541472 has one binding site for miR-K12-6-5p but none for miR-K12-11*, and CD27-AS1 has one binding site for miR-K12-11* but none for miR-K12-6-5p. Biotinylated miR-K12-6-5p mimic pulled down 43.7% of Loc541472 and none of CD27-AS1, and miR-K12-11* mimic pulled down 12.9% of CD27-AS1, but no Loc541472, thus confirming direct miRNA-lncRNA interaction (Fig 3B). The fact that we identified putative lncRNA targets of viral miRNAs in PEL and endothelial cells by Ago HITS-CLIP and viral genetics, together with biochemical evidence for direct miRNA-lncRNA interaction, demonstrated that KSHV deregulates a subset of host lncRNAs in a miRNA-dependent fashion.
To date a very small percentage of all lncRNAs are functionally annotated, making interpretation of lncRNA expression data challenging. As a starting point, we analyzed lncRNAs that were deregulated (upregulated, downregulated and rescued) in response to latent KSHV infection for known or proposed functions in disease processes. Comparison of our dataset to two public databases [28, 29] identified 54 lncRNAs that were previously shown to be aberrantly expressed in various human cancers (S3 Table). These include HOTTIP, DLEU2, HOTAIRM1, ANRIL, MEG3 and UCA1. Ten of the 54 lncRNAs are listed in Table 2, and include oncogenic and tumor suppressor lncRNAs. HOTTIP is upregulated in hepatocellular carcinoma, osteosarcoma, lung, prostate and other cancers [30]; DLEU2 is deleted in lymphocytic leukemia and epigenetically silenced in myeloid leukemia [31, 32]. Knockdown of HOTARM1 has been shown to promote proliferation in promyelocytic leukemia cells [33]. ANRIL is an oncogenic lncRNA that promotes proliferation in numerous cancers including basal cell carcinoma (BCC), glioma, prostate and ovarian cancers [34]. UCA1 is upregulated in multiple cancers, including bladder, endometrial and pancreatic cancer and acts as an oncogenic lncRNA [35]. Loss of MEG3 expression has been reported in a wide spectrum of malignancies ranging from gliomas to colon and liver cancers [36]. To understand the mechanisms by which cancer-related lncRNAs are deregulated by KSHV, and their contribution to pathogenesis, we chose to initially study UCA1, ANRIL and MEG3.
MEG3 is a tumor suppressor lncRNA which is proposed to act by enhancing transcription from p53-dependent promoters [36]. Studies in HCT116 and U2OS cell lines have identified that MEG3 is a nuclear localized lncRNA [41], which was also confirmed in GM12878 cells by the GENCODE project [42]. According to the microarray data (S2 Table), MEG3 was slightly upregulated during latent KSHV infection. However, when validating MEG3 expression by qRT-PCR, it behaved in a rescued pattern, being suppressed in wt-KSHV infection and restored in Δcluster-KSHV-infected cells, suggesting regulation by KSHV miRNAs (Fig 4A). MEG3 contained seed sequence matches for miR-K12-3, K12-5, K12-6-5p, K12-8* and K12-9*. Uninfected TIVE cells were transfected with a pool of three KSHV miRNA mimics (miR-K12-5, K12-6-5p and K12-8*). MEG3 expression was reduced by almost 80% (Fig 4B). Furthermore, miRNA pull-down assays using biotinylated miR-K12-6-5p mimic pulled-down 24.5% of MEG3 (Fig 4C). miR-K12-11* mimic did not pull down MEG3 lncRNA. These data are consistent with viral miRNAs directly binding to and downregulating MEG3.
ANRIL is a nuclear localized oncogenic lncRNA that drives proliferation by silencing the INK4 tumor suppressor gene by recruiting PRC2 complexes [34]. The fact that ANRIL was downregulated in KSHV-infected cells from the microarray data suggested that ANRIL does not have a direct role in proliferation; however, ANRIL has recently also been implicated in innate immune responses, albeit in the context of bacterial infection [43]. Analysis of ANRIL expression by qRT-PCR showed a very strong 100-fold downregulation in wt-KSHV-infected cells, and a slightly reduced inhibition in the Δcluster-KSHV-infected TIVE cells (Fig 5A). Such strong repression is not typical of miRNAs, however, the cDNA of ANRIL had a total of 17 6-mer seed matches for 12 of 25 mature KSHV miRNAs. To investigate whether the large number of KSHV miRNA seed sequence matches in ANRIL are targeted by KSHV miRNAs, we ectopically overexpressed the shortest isoform (transcript variant 12) of ANRIL from a CMV promoter-driven vector in wt-KSHV-infected and uninfected TIVE-Ex-LTC cells. Since TIVE cells are highly resistant to plasmid transfection, we used TIVE-Ex-LTC cells for this experiment. As shown in Fig 5B, the ANRIL expression levels achieved in wt-KSHV-infected cells were 80% less compared to uninfected cells. We note that this expression difference was not due to differences in transfection efficiencies, since a control gene (LSD-1), expressed from the same vector, was expressed at similar levels in both cell lines (Fig 5B). Hence, the reduced ANRIL expression levels in infected cells compared to control cells strongly suggested post-transcriptional miRNA-dependent regulation of ANRIL. To test this, we transfected a pool of four miRNA mimics (miR-K12-1*, K12-6-5p, K12-2* and K12-11*) which led to a strong knock-down of ANRIL expression in uninfected TIVE cells compared to the control mimic (Fig 5C). Additionally, pull-down experiments in TIVE cells using biotinylated miR-K12-6-5p and miR-K12-11* mimics, for which ANRIL contains two seed matches each, significantly pulled-down 12.7% and 22.7% of ANRIL transcripts, respectively (Fig 5D).
Together these data show that ANRIL is targeted by multiple viral miRNAs. Since ANRIL also contained miRNA seed sequence matches for miR-K12-10 and K12-12, which are still present in Δcluster-KSHV (Fig 1A), we wanted to test ANRIL expression in the absence of all viral miRNAs. To this end we analyzed ANRIL expression in TIVE cells by infecting with a virus lacking all 12 miRNA genes (Δall-KSHV). Surprisingly we did not observe significantly altered ANRIL expression compared to wt-KSHV-infected cells (Fig 5E). These data suggested that ANRIL may also be negatively regulated by latency associated proteins. To directly address this question we ectopically expressed the major latency associated proteins of KSHV (LANA, vCyclin, vFLIP and Kaposin) and monitored ANRIL expression by qRT-PCR. Since TIVE-Ex-LTC cells do not express detectable levels of ANRIL, this experiment was performed in HeLa cells, which are known to robustly express ANRIL [44]. vFLIP and vCyclin downregulated ANRIL expression by almost 75% and 53%, respectively (Fig 5F). LANA and Kaposin did not have significant effects. The observation that ANRIL is negatively regulated by both miRNAs and latency associated proteins is in congruence with other host genes that are targeted by multiple viral mechanisms [45].
Urothelial Cancer Associated 1 (UCA1) is a lncRNA which was identified as highly upregulated in bladder cancer and has since been implicated in other cancers like colorectal, ovarian and renal carcinomas [35]. UCA1 is partially localized in both the nucleus and the cytoplasm and plays distinct roles in different sub-cellular compartments [46, 47]. Recently, it was shown that UCA1 transcription is induced by HIF-1α, to enhance hypoxic proliferation, migration and invasion of bladder cancer cells [35]. UCA1 was upregulated by approximately 90-fold during wtKSHV infection and approx. 300-fold during Δcluster-KSHV infection (Fig 6A). Since UCA1 was upregulated under both infection conditions and its cDNA sequence contained no seed matches for any KSHV miRNAs, UCA1 is presumably not regulated by a miRNA-dependent mechanism.
To determine which of the four major latency-associated proteins (LANA, vCyclin, vFLIP and Kaposin) upregulates UCA1, we transfected TIVE-Ex-LTC cells with expression vectors either alone or in combination. Ectopic expression of vCyclin and Kaposin led to a 3.9 and 5.7-fold upregulation of UCA1 as monitored by qRT-PCR, respectively. Furthermore, co-transfection of vCyclin and Kaposin increased UCA1 to almost 15-fold compared to empty vector suggesting synergy (Fig 6B). LANA and vFLIP had no effect. The fact that the upregulation observed in transfected cells is much less than in the context of infection could be a consequence of either an altered stoichiometry or absolute expression levels of latency proteins, or mean that other viral genes might contribute to UCA1 upregulation.
To address whether UCA1 directly contributes to KS-associated phenotypes, we knocked-down UCA1 expression using siRNAs in KSHV-infected TIVE cells. At 24, 48, 72 and 96 h post-transfection we observed 60–85% knockdown of UCA1 expression (Fig 6C). First, we assayed for proliferation using the MTS assay. We measured proliferation at 24, 48, 72 and 96 h post-transfection and observed a statistically significant and dose-dependent decrease in proliferation of cells treated with siUCA1 as compared to scrambled control (Scr). Upon treatment with 10 nM siUCA1, the proliferation rate dropped to 72% by day 1 and then progressively to 52% by day 4 (Fig 6D). We observed a similar decrease in proliferation of uninfected TIVE cells transfected with siUCA1 (S2A Fig) suggesting that UCA1 contributes to endothelial cell proliferation in general. To test whether latent KSHV upregulates UCA1 in all infected cells, we measured UCA1 expression levels in uninfected and KSHV-infected iSLK cells. UCA1 was upregulated by almost 5-fold in KSHV-infected iSLK cells (S3A Fig). Knockdown of UCA1 in uninfected and KSHV-infected iSLK cells led to a mild reduction in proliferation of these cells (S3B and S3C Fig). The magnitude of effect observed in iSLK cells was much lower than that in TIVE cells, presumably because iSLK cells are transformed and unlike TIVE cells form tumors in nude mice [18]. Next, we assayed the effect of UCA1 knockdown on migration of KSHV-infected TIVE cells. The migration assay (wound healing) involves introduction of a scratch in a monolayer of cells and measuring the percentage of the clear area that gets covered by migration at 12 hours post introduction of the scratch under serum-free conditions (Fig 6E). siUCA1-treated cells were consistently slower in migration from day 1 through day 4, as they recovered only between 12–15% of the scratch area, while Scr-treated cells recovered between 26–35% of the area (Fig 6F). A similar reduction in migration was observed on days 1 and 2 when UCA1 was knocked down in uninfected TIVE cells, however, no difference was evident after day 3 (S2B Fig). This suggests that high UCA1 levels in KSHV-infected endothelial cells contribute to increased migration of these cells. These data demonstrate that the induction of UCA1 by the KSHV latency-associated proteins Kaposin and vCyclin promotes proliferation and migration, and likely contributes to KSHV pathogenesis and tumorigenesis.
Here we show that latent KSHV infection significantly alters the lncRNA expression profile of endothelial cells. Deregulation of lncRNAs has implications in diseases such as diabetes, neurological disorders, viral infections and cancer [48, 49]. Our study establishes that KSHV employs its latency proteins and miRNAs, either alone or in combination, to target specific lncRNAs and potentially contribute to sarcomagenesis.
Post-transcriptional regulation of lncRNA expression by miRNAs is a newly described phenomenon. Yoon et al showed let-7 loaded RISCs targeted lincRNA-p21 in a HuR-dependent manner in cervical carcinoma cells, eventually destabilizing and degrading lincRNA-p21 [50]. In bladder cancer, UCA1 and miR-1 expressions were inversely correlated, and overexpression of miR-1 phenocopied the knockdown of UCA1 [51]. Further, MALAT-1, a nuclear lncRNA, was reported to be targeted by miR-9 in an Ago-2-dependent manner in the nuclei of Hodgkin’s lymphoma and glioblastoma cell lines [15]. We identified 126 lncRNAs as potential targets of viral miRNAs in endothelial cells, and we verified direct miRNA/lncRNA interactions by pulldown experiments with biotinylated KSHV miRNA mimics targeting Loc541472, CD27-AS1, ANRIL and MEG3. Results from the Ago HITS-CLIP experiment further suggest that this regulation proceeds in an Ago and hence RISC-dependent manner. As per our current understanding, RISC-mediated silencing of mRNAs proceeds via translation repression and induction of mRNA turnover [52, 53]. RNA destabilization followed by degradation is perhaps the mechanism relevant to silencing of lncRNAs. However, the details of the mechanism, especially for lncRNAs lacking a cap and/or a poly-A tail, remain to be uncovered.
An alternative and not mutually exclusive mechanism that involves direct engagement of miRNAs and lncRNAs is miRNA sponging by lncRNAs [54]. LincRNA-RoR sponges miR-145-5p thereby increasing the expression of pluripotent stem cell factors Oct4, Nanog and Sox2 [55]. The Steitz lab showed that lncRNAs encoded by Herpesvirus Saimiri, called HSURs, sequester host miRNAs in infected T-lymphocytes [16]. It is plausible that some host lncRNAs could sponge KSHV miRNAs, thereby derepressing downstream targets instead of being targeted by miRNAs themselves.
We demonstrated that viral latency proteins vCyclin and Kaposin synergistically upregulate UCA1 while vFLIP and vCyclin downregulate ANRIL. Thus, aside from miRNAs, the latency proteins play a pronounced role in perturbing lncRNA expression. This is not surprising given we identified 858 differentially expressed lncRNAs during wt-KSHV infection and only 126 were potential miRNA targets. vCyclin, an ortholog of cellular Cyclin D, upregulates expression of cell cycle regulatory genes [56]. Moreover, Kaposin stabilizes cytokine mRNAs thereby increases their turnover time [57]. vCyclin and Kaposin may act cooperatively by augmenting transcription and simultaneously preventing turnover of UCA1. We also showed that ectopically expressed vFLIP strongly downregulates ANRIL. STAT1-mediated activation of the ANRIL locus in vascular endothelial cells has been reported based on GWAS studies [58]. Studies using a mutant virus that lacks vFLIP in HUVEC cells showed activation of STAT1 in a NF-κB-dependent manner, suggesting that vFLIP probably inhibits STAT1 to downregulate ANRIL expression [59]. A recent study in endothelial cells demonstrated that ANRIL expression is induced by pro-inflammatory molecules, especially NF-κB and TNF-α, and silencing of ANRIL expression led to a reduction in IL6/IL8 response [60]. This further underlines the role of ANRIL in immunity and supports the notion that KSHV may downregulate ANRIL to evade innate immune responses.
KSHV drives latently infected cells towards proliferation by a variety of mechanisms such as encoding orthologs for cell cycle proteins like vCyclin, or interfering with the p53 pathway through LANA [61], encoding miR-K12-11, an ortholog of oncomir-155 [62], and the induction of the oncogenic host miRNA cluster miR-17/92 [45]. Here we demonstrate that KSHV also upregulates UCA1 to drive proliferation and migration in endothelial cells. UCA1 has also been shown to promote the Warburg effect [63], an effect that has been shown to be required for maintenance of latent KSHV in endothelial cells [64]. We found that 53 additional lncRNAs previously shown to be aberrantly expressed in various malignancies are deregulated by KSHV, suggesting that UCA1 exemplifies how KSHV could similarly exploit lncRNAs that contribute to phenotypes such as proliferation and migration in the context of tumorigenesis. Given that the majority of lncRNAs we catalogued in this study remain uncharacterized, the repertoire of cancer-relevant lncRNAs regulated by KSHV may be much larger. Although cancer is the pathological consequence of KSHV infection, KSHV could target lncRNAs of biological significance in other cellular processes, for example, lncRNAs involved in inflammation and innate immunity [9]. KSHV continually evades the innate immune response using several approaches, like suppressing TGF-β signaling [45], activation of NF-κB response genes [65] and encoding trace amounts of v-IL6, a truncated version of human IL-6, during latent infection [66]. Loc541472, which we show here is targeted directly by KSHV miRNAs, is antisense to the hIL-6 promoter, suggesting that targeting of this lncRNA contributes to regulation of IL-6 expression. Indeed, preliminary experiments suggest a correlation between Loc541472 and hIL-6 expression and mechanistic studies are currently ongoing.
We identified a novel paradigm by which KSHV, an oncogenic herpesvirus, regulates cellular gene expression by targeting host lncRNAs with viral miRNAs and latency proteins. Studying lncRNAs deregulated by KSHV may yield novel mechanisms by which viruses evade the host immune response and in the case of EBV and KSHV contribute to tumorigenesis, as exemplified by our data on UCA1 which modulates migration and proliferation. Finally, studies on aberrantly expressed lncRNAs in KSHV-infected cancer cells may aid the functional characterization of cellular lncRNAs and at the same time identify novel virus-specific therapeutic targets for KS.
The viruses used in this study, wt-KSHV, Δcluster-KSHV and Δall-KSHV, have the viral genome cloned into a Bac-16 backbone, as described in Brulois et al. [19] and Jain et al. [20]. Transcript variant 12 (RefSeq ID: NR_047542.1) of ANRIL was expressed from a pcDNA3.1 vector [67]. LANA, vCyclin, vFLIP and Kaposin were expressed from pcDNA3.2 vectors [68].
Telomerase immortalized vein endothelial cells (TIVE) and long-term cultured KSHV infected cells (TIVE-LTC) were generated by immortalizing passage 2 HUVEC cells (kindly provided by Dr. Keith McCrae, Case Western Reserve University) in our laboratory as described [18]. All uninfected and infected TIVE cells were grown in complete Medium-199 (1% Pen-Strep, 20% FBS), supplemented with Endothelial cell growth supplement (Sigma). TIVE-Ex-LTC cells were obtained by culturing TIVE-LTC cells as single cell dilutions without antibiotic selection, and have lost all copies of viral episomes. TIVE-Ex-LTC cells grow faster and are more transfectable compared to TIVE cells. All uninfected and infected TIVE-Ex-LTC cells were grown in complete DMEM (1% Pen-Strep, 10% FBS). Latently infected TIVE and TIVE-Ex-LTC cells were maintained under hygromycin (10 μg/mL) to prevent episome loss. Body-cavity-based lymphoma (BCBL-1) cell line was derived from KSHV positive primary effusion lymphoma (PEL) and was kindly provided by Dr. Don Ganem at UCSF [69]. BCBL-1 cells were grown in complete RPMI (2% Pen-Strep, 10% FBS). HeLa cells and iSLK cells were grown in complete DMEM (1% Pen-Strep, 10% FBS).
The method for isolating nuclear and cytoplasmic fractions was adapted from [71]. Briefly, 1 x 107 BCBL-1 cells were pelleted and washed twice with ice cold PBS. Cells were resuspended smoothly by gentle pipetting in Sucrose buffer I (SB-I: 0.32 M Sucrose, 3 mM CaCl2, 2 mM Mg(Ac)2, 0.1 mM EDTA, 10 mM Tris-HCl (pH 8), 1 mM DTT, 0.5 mM PMSF and 0.5% NP-40) using 100 μL buffer per 1 x 107 cells. Lysis was at room temperature for 60–90 s. The nuclei were pelleted at 800 x g, 4 ˚C for 5 min and the supernatant (cytoplasmic fraction) was frozen immediately and stored at -80C. The pellet was resuspended smoothly by gentle pipetting in 50 μL of SB-I and allowed to sit for 30 s at RT. The nuclei were pelleted again at 800 x g, 4 ˚C for 5 minutes. The supernatant was discarded and the pellet (now whiter) was washed twice in 1 mL ice cold PBS. The resuspension was smooth and easy indicating no nuclear rupture. 10 μL of the 1 mL suspension from the second wash was trypan blue stained and checked by microscopy to verify the purity and integrity of the isolated nuclei. The nuclear fraction was frozen immediately and stored at -80 ˚C.
TIVE cells were grown overnight on coverslips at a dilution of 1 x 104 cells per well in a 6-well plate. Nuclei isolated from PEL cells were prepared as described [72], and fixed with a 1:1 ratio of methanol and acetone for 10 min in a humid chamber at 4 ˚C. The samples were blocked in PBS with 3% BSA for 1 h at room temperature, and then incubated overnight at 4 ˚C with either primary anti-Ago2 antibody or blocking solution (control). After washing, the samples were incubated with Alexa-468 anti-rat secondary antibody for 1 hour at room temperature. The slides were then stored at -20 ˚C and imaged using a LEICA TCS SP2 AOBS Spectral Confocal microscope. The images were analyzed and figures were generated using the freeware Vaa3D [73]. The movie was generated using Volocity® 6.3.
SDS-PAGE and Western blotting were performed using whole cell lysates, or cytoplasmic or nuclear fractions prepared from 100,000 cells/well. The following antibodies were used to probe the membrane: Ago2 (11A9, [74]), β-Tubulin (Millipore, CP06-100UG), Sm antigen (Dr. Joan Steitz’s lab, Yale University), Lamin A/C (Active Motif, 39288), Calnexin (ENZO Lifesciences, ADI-SPA-865-D).
Total RNA was isolated with RNA-Bee (Tel-Test Inc.) using the protocol provided by the manufacturer. Total RNA (5–10 μg) was treated with DNase I (NEB) according to the manufacturer’s instructions and ethanol precipitated overnight. Genome-wide lncRNA microarray analysis was performed with ArrayStar using Human LncRNA Array v3.0 (8 x 60K, Arraystar). A fold change cut-off of 2.0 was applied to filter lncRNAs into different categories (upregulated, downregulated and rescued) for further analyses. Three technical replicates for each of the three samples were analyzed.
Total RNA preparations from PEL cell fractions were reverse transcribed using the TaqMan MicroRNA Reverse Transcription Kit (ThermoScientific). Stem-loop qPCR was performed using the TaqMan Gene Expression Master Mix and appropriate miRNA assays from Applied Biosystems.
Total RNA (2 μg) was reverse transcribed using SuperScript III (Life Technologies) using random hexamers according to the manufacturer’s instructions. cDNA corresponding to 50–100 ng RNA was used per 10 μL of qPCR reaction. Instruments used for real-time PCR included ABI StepOne Plus (Applied Biosystems) and LightCycler96 (Roche). qPCR primer sequences are listed in S4 Table.
TIVE cells were seeded in 48-well plates (50,000 cells/well) and transfected with pools of miRNA mimics (in equimolar ratios and a final concentration of 5 nM) purchased from Qiagen. At 48 h post transfection, the lncRNA expression levels were measured using the Power SYBR Green Cells-to-CT Kit (ThermoFisher). In the cases of ANRIL and MEG3, 10 cm plates were seeded to 70% confluency and qRT-PCR analysis was performed using the conventional approach described above.
Biotinylated miRNA mimics (miR-K12-6-5p and miR-K12-11*) were purchased from Exiqon. Pulldown was performed from TIVE and TIVE-Ex-LTC cells according to the previously published protocol [75] with minor changes. Each replicate started with 6x106 cells for TIVE-Ex-LTCs (instead of 4x106) and 8x106 cells for TIVE cells. Input RNAs saved for analysis were 5% and 20% for TIVE-Ex-LTC and TIVE cells, respectively.
TIVE-Ex-LTC cells were reverse transfected in 6-well plates (300,000 cells/ well) with 2 μg of plasmid DNA using FuGENE HD according to the manufacturer’s protocol. HeLa cells were seeded in 6-well plates (150,000 cells/ well) and were transfected 24 h later with 2 μg plasmid DNA using Lipofectamine 3000 according to the manufacturer’s protocol. DMEM (10% FBS) was used for transfection of both cell types. Comparable transfection efficiencies were ensured by co-transfection of pmaxGFP. Total RNA was harvested from transfected cells at 72 h post-transfection.
wtKSHV-infected or uninfected TIVE cells were plated in 96-well plates (20,000 cells/well for MTS assay) and 48-well plates (250,000 cells/well for wound healing assay). Uninfected and wtKSHV-infected iSLK cells were plated in 96-well plates (8000 cells/well for MTS assay). siRNAs (5nM or 10 nM) against UCA1 (Qiagen) were transfected using Lipofectamine RNAiMAX reagent (ThermoFisher) according to the manufacturer’s protocol. ON-TARGETplus Non-targeting Control siRNA (Dharmacon) was used as the scrambled negative control. At 4 h post-transfection, the serum free medium was replaced by complete Medium-199 (TIVE) or DMEM (iSLK). Comparable transfection efficiencies were ensured by co-transfection of siGLO (Dharmacon).
Statistical analyses on experimental measurements were done using two-tailed student’s t-test assuming unequal variances for all experiments reported.
Raw data files from the microarray experiment were deposited to the Gene Expression Omnibus under the accession number GSE89114.
|
10.1371/journal.pcbi.1000170 | Insights into Protein–DNA Interactions through Structure Network Analysis | Protein–DNA interactions are crucial for many cellular processes. Now with the increased availability of structures of protein–DNA complexes, gaining deeper insights into the nature of protein–DNA interactions has become possible. Earlier, investigations have characterized the interface properties by considering pairwise interactions. However, the information communicated along the interfaces is rarely a pairwise phenomenon, and we feel that a global picture can be obtained by considering a protein–DNA complex as a network of noncovalently interacting systems. Furthermore, most of the earlier investigations have been carried out from the protein point of view (protein-centric), and the present network approach aims to combine both the protein-centric and the DNA-centric points of view. Part of the study involves the development of methodology to investigate protein–DNA graphs/networks with the development of key parameters. A network representation provides a holistic view of the interacting surface and has been reported here for the first time. The second part of the study involves the analyses of these graphs in terms of clusters of interacting residues and the identification of highly connected residues (hubs) along the protein–DNA interface. A predominance of deoxyribose–amino acid clusters in β-sheet proteins, distinction of the interface clusters in helix–turn–helix, and the zipper-type proteins would not have been possible by conventional pairwise interaction analysis. Additionally, we propose a potential classification scheme for a set of protein–DNA complexes on the basis of the protein–DNA interface clusters. This provides a general idea of how the proteins interact with the different components of DNA in different complexes. Thus, we believe that the present graph-based method provides a deeper insight into the analysis of the protein–DNA recognition mechanisms by throwing more light on the nature and the specificity of these interactions.
| The interaction of proteins with DNA is crucial for several cellular processes. Some insights into the mode of interaction can be obtained from the analysis of the complexed structures. Conventional analyses are based on the identification of pairwise interactions. However, a collective representation of the network of interactions and the analyses of such networks provide valuable information, which is not easy to obtain from pairwise analyses. Although the protein structure networks have been described in the literature, this is the first time that a network representation of protein–DNA is described. Construction and analysis of such networks have given valuable information on protein–DNA interactions in terms of network parameters, such as clusters of interacting residues and hubs, which are highly connected residues. Furthermore, the results also represent both the protein- and the DNA-centric viewpoints, because the analysis is carried out on combined networks. The methodology developed here can lead to predictions, such as important residues responsible for stabilizing protein–DNA interactions, and will be of interest to experimentalists.
| A network of interactions among the macromolecules drives the cell. The protein–DNA interactions orchestrate the high fidelity processes like DNA recombination, DNA replication, and transcription. With the increasing number of high-resolution structures of macromolecular complexes, it is now possible to obtain insights into the atomic details of interactions governing their structural and functional integrity. In the present study, we focus on protein–DNA interactions, which can either be specific or non-specific depending on the functional requirement. Insights into the mechanism of protein–DNA binding and recognition have come from extensive analysis of protein–DNA interfaces [1]–[14]. Some of these investigations have been carried out at the level of pairwise interactions between the atoms/residues of the interacting partners. However, the information communicated along the interfaces is rarely a pairwise phenomenon. New insights can be gained by investigating the interactions holistically, extending beyond the pairwise analysis of atomic/residue interactions. This can be achieved through the use of efficient methods, which capture the topological features from the structures of these complexes. The concept of representing protein structures as graphs exists in the literature [15]–[21]. In these studies the amino acids in proteins are considered as nodes and the interaction between these nodes have been considered as edges for constructing different types of graphs. These protein structure graphs (PSG) have been successfully used in the analysis of protein structure, stability and function [22],[23]. PSGs have also been analyzed in protein–DNA complexes to identify significant interactions as clusters of interacting amino acids at the protein–DNA interfaces [4]. However in such studies, the interacting nucleotides of the DNA were not considered as part of the graphs, since the parameters required for representing DNA as graphs were not available at that time.
As a conceptual turning point, it has been pointed out that most of the information on protein–DNA complexes have been obtained from a “protein-centric” view and new insights are likely to emerge if protein–DNA complexes are investigated from a “protein–DNA-centric” viewpoint [24]. Recently, protein–DNA complexes have been classified on the basis of structural descriptors that highlights the significance of the protein induced distortions of the DNA [3]. In the current article, the interactions between the protein and the DNA in protein–DNA complexes have been evaluated in a collective fashion by considering the complexes as bipartite graphs. We have developed the parameters to evaluate the strength of interaction between the amino acids and the nucleotides, based on the atomic contacts. Such a combined graph based analysis of protein–DNA interactions has been presented for the first time in this study.
The protein–DNA graph is of special interest, since we are dealing with two different types of biopolymers with unique structural and chemical properties. In the case of proteins, the two amino acids are linked by a rigid peptide bond and each amino acid could be unambiguously considered as a node in the protein graph. However, in the case of DNA, the linkage between two nucleotides is through a flexible phosphodiester bond and the nodes can be defined at various levels. For example, a nucleotide as a whole or its individual chemical components such as the phosphate group, deoxyribose and the bases (A, T, G, and C) can be represented as nodes. Such a different representation of nodes has distinct advantages of their own, in interpreting the nature of interaction between the protein and DNA [13], as we will show in this study. An important component of the analysis is to quantify the strength of interaction (on the basis of the number of atomic contacts) between these chemically different molecules to capture the essence of the intermolecular interactions at the protein–DNA interface.
The present graph based analysis of protein–DNA complexes focuses on the following points. Primarily, the interface interactions of protein–DNA complexes have been investigated at a network level. This is achieved by constructing protein–DNA graphs (PDGs) on the basis of the strength of interaction between the nodes and also by performing extensive calibrations to choose the optimal strength of interactions to gain structural insights. Secondly, the clusters and hubs of such interacting amino acids and the nucleotides at the interfaces have been analyzed in a set of protein–DNA complexes.
Significant results that are inaccessible by conventional pairwise analysis of the structure or by sequence analysis have been obtained from the present work. These include the identification of spatial networks of interacting residues that are sequentially far apart, the evaluation of a scale of interaction strength along which we can compare and analyze the interaction networks of protein–DNA complexes and the identification of groups of optimally interacting residues which stabilize the structural architecture. Furthermore, we have been able to revisit the classification scheme of DNA binding proteins. Our classification schema, which is based on the concepts of graphs/network, interaction strength, and the type of interaction, is distinct from the classification schemes proposed earlier with a protein-centric point of view. We have compared our results with the other classification schemes [3],[25].
A protein–DNA graph (PDG) is a bipartite graph constructed to represent the interaction between the amino acids of the protein and the nucleotides of the DNA in a protein–DNA complex. A bipartite graph deals with two different node sets and edges are defined across the two node sets. A contact in the bipartite PDG is defined when a side chain of an amino acid interacts with the nucleotide. The interactions of the amino acid with the nucleotide can be considered at different levels: with the phosphate (p), deoxyribose sugar (S) or base (B) components individually, or with the nucleotide as a complete entity. The edges are defined upon quantification of the interaction between the amino acids and the nucleotides with the “Interaction Strength,” Iij (It is to be noted that the interaction strength mentioned here is based on the number of atom-atom contacts and in a way reflects only the local packing density.) The details of the construction of PDG are presented in the Materials and Methods section. A bipartite PDG representation of a protein–DNA complex is illustrated in Figure 1.
The nodes in a PDG are connected if the Iij evaluated between the nodes is greater than or equal to a user-defined Iij. The user-defined Iij is termed as Minimal Effective Connection (MEC). We have constructed PDGs from protein–DNA complexes using a range of MEC from 0% to 15%. A graph generated with a high MEC is sparse with strongly connected nodes, and the graph generated with a low MEC is dense with weakly connected nodes (Figure S1). Hence, the choice of the window of MEC for the analysis of the graphs becomes important, which ensures that we neither include insignificant (weak) interactions nor miss significant (strong) ones. As we have partitioned the nucleotide of the DNA into its phosphate, deoxyribose and base components, an optimal range of MEC (OMEC) is chosen for constructing protein-phosphate (P-p), protein-deoxyribose (P-S) and protein-base (P-B) graphs. The OMEC is selected to balance a trade-off between the strength of interaction and the cluster size. Thus, the MEC is further classified into weak (WMEC), optimal (OMEC) and strong (SMEC) on the basis of Iij and the ranges of MEC values are listed in Table 1. All further discussions are based on the analyses with the OMEC ranges, unless otherwise specified. We obtain P-p, P-S, and P-B clusters and hubs from the adjacency matrices of PDGs (see Materials and Methods section). The detailed analysis of protein–DNA complexes through these parameters is presented in the following section.
The propensities of amino acids to form P-p, P-S, and P-B graphs were calculated from DS1 (see Materials and Methods section). The results are presented in Figure 2. In general, we observe a higher propensity of basic residues (Arg and Lys) to occur in PDGs. Arg is more preferred in P-B graphs whereas Lys is more preferred in P-p graphs. All polar amino acids occur significantly in all the three component clusters, however the preferences vary. For instance, Ser, the smallest polar amino acid has a higher propensity to occur in P-p graphs, and Asn/Gln, which contain the planar conjugated amide group has higher propensity of occurrence in P-B graphs. The interactions of Val, Ile, Leu, Phe, Trp are higher in the P-S graphs indicating that the deoxyribose is involved in hydrophobic and van der Waals interactions. The acidic residues (Asp and Glu) are not excluded from the interface interactions, in spite of the net negative charge of DNA. However, the occurrence of these amino acids near the phosphate backbone (P-p graphs) is minimal and significant near the bases (P-B graphs). Also, Glu interacts with bases more frequently than Asp. Thus, the analysis confirms some of the expected trends of interaction between amino acids and DNA, such as the dominance of basic and polar residues, and lesser preference of hydrophobic and acidic amino acids in PDGs. Additionally, the preference of interactions with individual chemical components such as the phosphate backbone, deoxyribose sugar and bases have been elucidated in detail. Such a level of analysis is useful in understanding the subtleties involved in protein–DNA interactions, and for interpreting the nature of component graphs as will be discussed later in the context of classification of protein–DNA complexes.
The protein–DNA complexes have been classified into different groups based on the structural similarity of the proteins bound to the DNA. Luscombe et al have provided a comprehensive classification of the protein–DNA complexes based on the secondary structural motifs of proteins interacting with the DNA [25]. The classification results in eight groups of complexes: β-sheet group, β-hairpin group, helix turn helix (HTH), zipper type (ZT), zinc coordinating group, other α-helices, enzymes and others [25]. We have cured it further to remove structures with single-stranded DNA (see Table S1 for details regarding the members of the dataset). We have generated PDGs for all groups of protein–DNA complexes. The PDGs are further analyzed to investigate the properties such as the preference of proteins to interact with the DNA, the components of the DNA to which the protein binds, the dominance of a particular type of cluster (P-p, P-S, or P-B). We have also tried to find a generic pattern (if any) of clusters that could be identified amongst the groups of protein–DNA complexes. The important results for each group are discussed in the following sections.
The importance of classifying the DNA binding proteins from the protein–DNA point of view, rather than the protein-centric view is being recognized for developing better protein–DNA recognition code [24]. Since our method considers the spatial relationships between the amino acids and the nucleotides at the interface, we have attempted to classify the DNA binding proteins based on these spatial relationships.
The interface P-p, P-S, and P-B clusters are examined in the protein–DNA complexes. The analysis of these clusters shows that the complexes can either have exclusive P-p, P-S, or P-B clusters or they can contain a mixture of these types of clusters. In cases where more than one type of cluster is observed, we define overlapping (amino acids sharing the same nucleotide) or non-overlapping (amino acids making contact with different nucleotides) clusters. Based on the types of clusters observed, the complexes are classified into seven groups. The complexes containing exclusive P-p, P-S, and P-B clusters are denoted as class 1, 2, and 3, respectively. Mixtures of (P-p+P-S), (P-S+P-B), and (P-p+P-B) clusters are denoted as class 4, 5, and 6, respectively. The complexes containing all the three component clusters (P-p, P-S, and P-B) are considered as class 7. A sub-classification based on the presence of overlapping or non-overlapping (P-p+P-S) or (P-S+P-B) or (P-p+P-B) components is made for cases 4 to 7. The details of the classification of the protein–DNA complexes are presented in Table 4. Among the protein–DNA complexes of the dataset only ten complexes exhibit exclusive P-p, P-S, or P-B clusters at the protein–DNA interface and thereby fall into distinct classes 1, 2, or 3 (Table 4). Majority of the other complexes however, seem to employ concerted interactions and interact with different chemical components of the nucleotide by forming two or more types of clusters (either overlapping or non-overlapping) at the interface (see Figure S2 for illustrative figures).
It should be noted that our classification scheme based on the interaction patterns of amino acids with nucleotide components in PDG does not directly deal with the type of interaction involved (like electrostatics, van der Waals, H-bonding, etc). However, indirectly, the P-p cluster is dominated by electrostatic interaction and the P-S clusters are composed of van der Waals interactions along with stacking of aromatic residues with the deoxyribose ring. The P-B graphs are dominated by stacking of amino acids (mostly the planar side chain of Arg) with the bases, H-bonding and also charge mediated interactions.
A comparison of the present classification with the structural motif based classification by Luscombe et al [25] shows distinct differences. A major difference is that the proteins from the same group (motif based classification) fall under different classes of interface clusters. In other words, even though the proteins have the same secondary structure motif (e.g., HTH motif), their mode of interaction may vary significantly depending on factors like the sequence of DNA (cognate/non-cognate DNA binding) and the component (p, S, or B) of the nucleotide to which it binds. However, we see a few salient features, which are common to both the classification schemes. For example, most members of the Zinc Coordinating group belong to non-overlapping category of class 6 (P-p and P-B), and non-overlapping class 4 (P-p and P-S). A few exceptions are ZIF-268 DNA complex (1zaa), Glucocorticoid mutant–DNA complex (1lat), and Reverba Orphan Nuclear Receptor/DNA complex (1a6y). Similarly, in the Helix Turn Helix members (with the exception of Matα2 Homeodomain Operator Complex (1apl), lambda repressor mutant protein/ DNA complex (1lli), and Engineered Cro monomer/DNA complex (3orc)) belong to either class 4 or class 7, both of which contain P-p and P-S clusters (overlapping as well as non-overlapping). The β-sheet group, which bind to the TATA box, belong to the classes 5 and 7, both of which contain P-S and P-B clusters. These protein–DNA complexes (with the exception of TFIIB/TBP/TATA element ternary complex (1vol)) belong to the overlapping type, indicating the continuous involvement of the deoxyribose and the base moieties of the nucleotides. As we had seen earlier, the P-S clusters in these complexes are extensive in most of the cases. Finally, a majority of the enzymes belong to the classes 4 and 7 (both overlapping and non-overlapping), which have P-p and P-S clusters (Table 4).
Here we have shown that the DBPs classification based on the protein–DNA interface interaction at the molecular level differs significantly from the protein motif based classification, although a little consensus is observed. DBPs have also been classified based on other criteria. For instance, Prabakaran et al [3] have classified DBPs based on the structural descriptors (Structural Descriptor Based Classification-SDBC), involving both protein and DNA. And Siggers et al [7] have developed a score (IAS) based on the interface geometry to align interfaces, which has been used to classify protein–DNA complexes. We have also investigated the interface clusters of the complexes from the dataset used by Prabakaran et al and found only a marginal correspondence in classification. For example, Class 1 of SDBC has prominent overlapping P-p and P-B clusters when subjected to our method of classification. This class is characterized by major groove binding proteins and interact mostly with the bases of the DNA. Also Class 2 of SDBC, which has high number of both major and minor groove contacts, and structurally deformed DNA, has overlapping P-p and P-S clusters. Class 5 of SDBC shows mostly overlapping P-S and P-B clusters.
The fact that there is only a marginal overlap between different classification schemas underscores the versatilities in protein–DNA recognition mechanism. It may be valuable to use different approaches to obtain complementary information to understand the protein–DNA recognition mechanisms in detail.
The present study aims to represent a protein–DNA interface as an undirected bipartite graph based on non-covalent interactions. A quantitative method has been developed to represent the interactions between both DNA and protein as a single, combined graph. Such a representation has facilitated the study of the spatial relationships between the amino acids and the nucleotides at the protein–DNA interface in a holistic way. Thus, protein–DNA interfaces across the spectrum of complexes could be compared at a uniform level, irrespective of the structural and functional differences.
In general, we have provided a method of quantifying the interactions of proteins with the components of nucleotides (phosphate, deoxyribose and base). It is now clear that the combined representation of protein and DNA as PDGs could highlight the intricacies involved in protein–DNA recognition of some families of proteins. For instance, the predominance of protein-deoxyribose (P-S) clusters and hubs has brought out the specificity of the interaction in β-sheet proteins. Such analysis and the group specific features of protein–DNA recognition could be used as a starting point in predicting the DNA binding sites on these proteins. We have also proposed a scheme for classifying the structures based on the nature of the network connectivity present at the protein–DNA interface. Based on comparative analysis, we conclude that different classification schemes could provide complementary information on the nature of protein–DNA interactions.
Thus, the analyses performed on a dataset of protein–DNA complexes have highlighted the nature of the clusters and hubs present at the recognition site. These clusters and hubs may not only prove to be valuable in understanding the residues contributing to the stability of the protein–DNA interfaces, but also could be identified as features characteristic for a given group of proteins. The knowledge gained from the study could also provide a platform for further docking and prediction experiments.
The protein–DNA complexes with resolution better than 2.5 Å and with protein identity less than 25% were taken from PDB (Version 3.1) [39] and were further cured so that the proteins (size>40 amino acids) are bound to at least one complete turn of double-stranded DNA. This resulted in a dataset (DS1) of 118 protein–DNA complexes (Table S5), which was used for evaluating the amino acid propensities for various component graphs and for recalculating the normalization values. The interface clusters and hubs were identified on two datasets, DS2 from Luscombe et al [25] and DS3 from Prabakaran et al [3], which were used for direct comparison of their classification schemes with our graph based classification scheme. DS2 was further cured for removing identical protein chains and the complexes containing single stranded DNA (Table S1).
The interaction between the amino acids and the nucleotides at the protein–DNA interface is represented as undirected bipartite Protein–DNA Graphs (PDGs). Here, the amino acids comprise one node set and the nucleotides constitute the other node set of the bipartite-PDGs as shown in Figure 1. As the focus of this work is on the protein–DNA interface, we have adopted this bipartite graph, in which the edges are made only across the amino acids and the nucleotides. The non-covalent interactions between the amino acid side chains and the nucleotides form the basis for linking the nodes. These non-covalent interactions are evaluated from the atomic contacts of the nodes. Any two atoms from nodes i and j, are considered to make a contact if they are within a distance of 4.5 Å and the total number of such contacts (nij) is evaluated between a pair of nodes i and j. The strength of interaction (Iij) between these nodes i and j is evaluated in a manner similar to that adopted in the case of protein structure graphs [4],[23],[40], as given in Equation 1.(1)where Iij is the strength of interaction between the amino acid side chain (i) and the nucleotide (j) of the protein–DNA complex. As the graphs are undirected Iij = Iji. nij is the number of interactions existing between nodes i and j (contacts within distance of 4.5 Å). Ni and Nj are the normalization values of the corresponding nodes evaluated as described in the next subsection.
Here the evaluation of the strength of interaction is restricted only to atom-atom contact and does not explicitly take into account the details such as hydrogen bond, salt bridge interactions etc. Indirectly, this amounts to a measure of packing density at the selected region.
Iij is evaluated for all the amino acid-nucleotide (interface) interactions of a given protein–DNA complex. A threshold of Iij is used to connect two nodes in a graph. The threshold Iij representing the minimal atomic connectivity between interacting amino acids and the nucleotides is called as the Minimal Effective Connection (MEC). Thus, an edge between nodes i, j is defined if the Iij evaluated is greater than the user-defined MEC. For instance, a MEC of 6% specified by the user results in the generation of a PDG with nodes connected by the interaction strength Iij≥6%. Here the evaluation of Iij requires the normalization values (maximum number of contacts made by the unit) for corresponding nucleotide and amino acid units. The details of the evaluation of the normalization values are discussed below.
The normalization values are the estimates of the maximum non-covalent contacts an amino acid or a nucleotide can have in protein–DNA complexes. The method of obtaining normalization values for amino acids in proteins was previously given [40] and a similar method is used to obtain protein–DNA normalization values. These values are evaluated from a non-redundant dataset of protein–DNA complexes [41], for all the 20 amino acids (Na) and the A, G, T, C nucleotides (Nn) as shown below,(2)where Max ak is the maximum number of non-covalent contacts made by the amino acid, “a,” with other amino acids and nucleotides in the protein–DNA complex “k.” Similarly, Max nk is evaluated for the nucleotide “n” in the structure “k.” “m” is the total number of protein–DNA complexes from the non-redundant dataset.
Here it should be noted that, while calculating the amino acid - amino acid contacts, the contacts made by an amino acid side chain with its sequence neighbors (i±1) are ignored. However, in the case of nucleotides, the sequential base contacts (stacking) of a nucleotide (i±1) are taken into account, ignoring only the covalent phosphate and the sugar contacts of the sequential residues. We wish to point out that the normalization values of amino acids obtained here from protein–DNA complexes are not significantly different from those obtained only from protein structures [40].
Protein-nucleic acid recognition mechanisms are often mediated by amino acids through a specific or nonspecific recognition of a nucleotide backbone or base at the protein–DNA interface. Quite often the electrostatic interactions of proteins with the phosphate groups are considered as non-specific and the stacking interactions and hydrogen bonding with bases are considered as specific. Furthermore, there is substantial conformational flexibility in the phosphodiester bond and in the conformation of the deoxyribose ring. Therefore, the nucleotides have been dissected into their chemical components such as the phosphate backbone (p), deoxyribose sugar (S), and the base (B). The interactions of an amino acid with all these individual components are characterized by constructing separate interaction graphs of amino acids with all the p, S, and B components of the DNA.
The normalization values (NP, NS, and NB) for these dissected components of the nucleotide are also calculated using Equation 3 where, in place of the whole nucleotide, the dissected component of the nucleotide is considered as,(3)where Max pk is the maximum number of non-covalent contacts, made by the phosphate with other amino acid and nucleotide residues in a protein–DNA complex “k.” Similarly, Max Sk and Max Bk are evaluated for the deoxyribose sugar and the base, respectively, from the complex “k.” “m” is the total number of protein–DNA complexes from the non-redundant dataset.
These individual phosphate, sugar backbone, and base-specific normalization values are useful to obtain finer details regarding the molecular connectivity existing at the protein–DNA interface. The normalization values of the individual chemical components of the nucleotide thus obtained are given in Table 5.
The PDGs are constructed as specified above and represented as binary adjacency matrices at a given MEC. Clusters of interacting nodes are identified from the adjacency matrix using Depth First Search (DFS) algorithm [42]. This is a method of graph traversal in which we obtain all the nodes that are either directly or indirectly linked to a node Vx from which the search starts. The backtracking of Vx starts only when all such connected nodes are explored. The next start point Vy is the next unvisited node in the graph. Thus, a cluster of strongly interacting nodes (higher MEC) captures the significant interactions that exist between the amino acids and the nucleotides at the protein–DNA interface. In this study, we have focused on the interaction of protein with the components of nucleotides. Thus, we have identified the component clusters, P-p, P-S, and P-B, to capture the interaction of amino acids with the phosphate, deoxyribose and the base of the nucleotides, respectively. In many cases, amino acids interact with more than one component of a nucleotide (for example, an amino acid may interact with the phosphate atoms as well as the deoxyribose of the nucleotide). In such cases, we have defined the clusters as “overlapping clusters” (which consists of both phosphate and deoxyribose of the same nucleotide). Thus overlapping clusters constitute the interactions of amino acids captured with two or more components of a nucleotide. The details of the interface clusters for DS2 are given in Table S1 and Table S2.
Hubs are highly interacting nodes in a graph. In protein structure graphs, a node was declared as a hub if it was connected to a minimum of four nodes [23]. The same definition is being used here for the protein-sugar (P-S) and protein-base (P-B) hubs in PDGs. In the case of a protein-phosphate (P-p) graph, a phosphate residue connected to a minimum of three residues is considered as a hub. The purpose of hub analysis in this study is to identify the nucleic acid component, highly connected to the amino acid residues and vice versa. Thus, a (P-B) hub for example, may constitute an amino acid connected to four or more bases or a base being connected to four or more amino acid residues. The interface component hubs for the protein–DNA complexes of DS2 are presented in Table S1 and Table S3.
|
10.1371/journal.pgen.1000383 | HECTD2 Is Associated with Susceptibility to Mouse and Human Prion Disease | Prion diseases are fatal transmissible neurodegenerative disorders, which include Scrapie, Bovine Spongiform Encephalopathy (BSE), Creutzfeldt-Jakob Disease (CJD), and kuru. They are characterised by a prolonged clinically silent incubation period, variation in which is determined by many factors, including genetic background. We have used a heterogeneous stock of mice to identify Hectd2, an E3 ubiquitin ligase, as a quantitative trait gene for prion disease incubation time in mice. Further, we report an association between HECTD2 haplotypes and susceptibility to the acquired human prion diseases, vCJD and kuru. We report a genotype-associated differential expression of Hectd2 mRNA in mouse brains and human lymphocytes and a significant up-regulation of transcript in mice at the terminal stage of prion disease. Although the substrate of HECTD2 is unknown, these data highlight the importance of proteosome-directed protein degradation in neurodegeneration. This is the first demonstration of a mouse quantitative trait gene that also influences susceptibility to human prion diseases. Characterisation of such genes is key to understanding human risk and the molecular basis of incubation periods.
| Prion diseases are fatal transmissible neurodegenerative diseases of animals and humans for which there is no treatment. They include Bovine Spongiform Encephalopathy (BSE), and its human equivalent, variant Creutzfeldt-Jakob Disease (vCJD). Prion diseases are characterised by a long, silent incubation period before the disease emerges, and this time interval varies greatly between individuals. Differences in our genetic makeup are a key factor in this variability. We already know that natural variation within one key gene, the prion protein gene, has a major influence on incubation time, but it is now clear that a number of other genes are also important. Using a mouse model, we have identified one of these genes, Hectd2, which is thought to be involved in the process that removes unwanted proteins from the cell. We also show that HECTD2 is associated with an increased risk of two human prion diseases—vCJD in the United Kingdom and kuru in Papua New Guinea. These data will give us a better understanding of the fundamental processes involved in these diseases and go some way to explaining why some individuals exposed to BSE have developed vCJD and others have not.
| Prion diseases are fatal transmissible neurodegenerative disorders of animals and humans. These include the agriculturally and economically important diseases of scrapie and Bovine Spongiform Encephalopathy (BSE) and the human diseases sporadic Creutzfeldt-Jakob disease (CJD), variant (vCJD) and kuru. Sporadic CJD has no known aetiology and vCJD is thought to have arisen following exposure to BSE prions [1]. Kuru is a prion disease that reached epidemic proportions in the 1950s in the Fore linguistic region of Papua New Guinea and is thought to have been transmitted through endocannibalism by participation in mortuary feasts [2]. Following the cessation of this practice in the late 1950's, the incidence of disease has declined, however, it remains our only experience of a large epidemic of acquired human prion disease and provides a useful model for vCJD [3].
Although there was widespread population exposure in the UK and some other countries to BSE only around 200 have developed clinical vCJD to date, although the number infected remains unknown. This represents an on-going public health concern with a risk of iatrogenic transmission through blood and surgical instruments. vCJD has not been associated with any unusual pattern of dietary or occupational exposure to BSE prions and a significant genetic component to risk seems probable therefore the identification of susceptibility factors is key to estimating individual risk [1].
All prion diseases have prolonged clinically silent incubation periods which in humans span over 50 years [2]. Marked variation in incubation period occurs between inbred lines of mice and this is determined by multiple genetic loci in addition to the prion protein gene [4],[5]
Previous studies have identified several quantitative trait loci for prion disease incubation time in mice. However, the resulting regions of interest spanned many megabases and were consequently too large for individual candidate gene analysis [6]–[10]. Several different strategies are available for fine mapping [11]–[13] and we chose to use a heterogeneous stock of mice. These are produced to model an out-bred population of mice, however they have the advantage of starting with a defined number of parental alleles. Heterogeneous stocks of mice have been shown to be a useful mapping tool because they provide a high level of recombination and the development of specific mapping software allows for convenient multipoint linkage analysis [14],[15]. This approach led to the identification of Hectd2, an E3 ubiquitin ligase, as a quantitative trait gene for prion disease incubation time. Mouse models are extremely useful for studying human prion diseases as they faithfully recapitulate many key features of the disease and indeed rodents are naturally susceptible to prion diseases. It is expected that susceptibility genes and pathways identified in mice will also be relevant to human prion diseases. To test this hypothesis we carried out an association study with HECTD2 markers and samples from different human prion diseases and successfully found a significant association with two acquired forms of prion disease: vCJD and kuru.
To fine map regions thought to contain quantitative trait loci for prion disease incubation time we utilised the Northport heterogeneous stock [16] (HS) of mice (gift of Robert Hitzemann), which was produced by semi-randomly mating eight inbred lines of mice (A/J, AKR/J, BALB/cJ, C3H/HeJ, C57BL/6J, CBA/J, DBA/2J, LP/J). Approximately 1000 mice were inoculated intracerebrally with mouse-adapted scrapie prions (Chandler/RML) and incubation times (in days) were determined as previously described [6],[17]. Regions of interest for fine mapping include those identified in previous crosses [6]–[10] and also from other studies. In this study we focus on a region of Mmu19 as a result of an interest in candidate genes on human chromosome 10 (unpublished data).
Nine microsatellite markers from chromosome 19 (D19Mit86-D19Mit112 see Table S1) at approximately 1 cM intervals were genotyped in approximately 400 animals which represent the extreme 20% of both sides of the incubation time distribution. Multipoint linkage analysis was carried out using HAPPY (http://www.well.ox.ac.uk/happy) [18]. A peak of linkage (−logP = 5.88) was seen between D19Mit63 and D19Mit65, a region of approximately 2.9 Mb (Figure S1A). Significant linkage was taken as −logP>3 as defined by a permutation test (n = 1000) carried out by HAPPY. This interval explains 6.9% of the observed variance therefore as predicted by other QTL mapping studies [6]–[10], other loci are expected to contribute to prion disease incubation time. Trait estimates for each strain are shown in Table S2.
Twenty seven RefSeq genes were identified within this region (NCBI build 37), 22 of which were sequenced (Table S3) in the parental strains of the HS in order to identify polymorphisms. Sequencing was not exhaustive and focused primarily on the exons including 5′ and 3′UTRs, intron/exon boundaries and potential promoters as defined by the literature for each gene or PROSCAN (http://www-bimas.cit.nih.gov/molbio/proscan). 177 polymorphisms were identified across the region which included single nucleotide polymorphisms (SNPs), simple repeats and insert/deletion polymorphisms. Most of the variation was observed in non-coding regions, however, several non-synonymous changes were seen (for detail see Table S4). All variants were assessed using an additional function of HAPPY which assigns a probability that any polymorphism is a quantitative trait nucleotide (QTN) [19]. This predicts which strain distribution pattern (SDP) most closely fits the pattern identified by the microsatellites in the HS animals (Figure S1B, Table S4). The main candidates to emerge from this analysis are Hectd2, Exoc6, Cyp26c1, Cyp26a1, Plce1 and Lgi1 (Table 1). Some SDPs are broadly conserved across the whole gene (e.g. −logP = 6.12 Hectd2, −logP = 6.74 Cyp26c1 and Plce1) whereas others represent single polymorphisms (e.g. −logP = 6.74 Cyp26a1 and Lgi1).
The −logP values assigned by HAPPY are predictions. We therefore tested representative polymorphisms from either each gene, or each strain distribution pattern, by genotyping the HS (Table 2). The only highly significant SNPs were seen in Hectd2 (P = 0.0008, 0.0013 and 0.0022, ANOVA) suggesting that Hectd2 is the most promising candidate in this region. However, these analyses are not exhaustive and it is not possible to exclude the possibility that variation in other genes or intergenic regions also contribute to prion disease incubation time.
Seven SNPs were identified in the 3′UTR of Hectd2 (Table S4), however, it is not clear whether they would affect regulation. Five polymorphisms occur within the predicted promoter (−226 to +25) one of which affects a potentially functional site. Sequence for the C57BL6/J allele from −216 to −210 is TGGGCGG and the insertion of 6 Gs in the alternative allele gives TGGGGGGGGGCGG. Both variants contain the consensus sequence for a Sp1 binding site (shown in bold) however the insertion also generates an overlapping large T antigen binding site (underlined). It is unclear whether additional mouse proteins could bind to this sequence or whether Sp1 binding would be affected. The significant SDPs were spread across the whole of Hectd2 therefore we cannot exclude any of these closely linked polymorphisms either individually or collectively from a contribution to the phenotype.
To determine whether the polymorphisms detected in Hectd2 have an effect on expression, RNA was extracted from whole brains of 8 week old males from the parental strains of the HS (except LP). Samples were analysed by real time RT-PCR. To examine genotype-related differential expression, strains were grouped according to the major strain distribution pattern seen in Hectd2 (Group A = A, AKR, BALB; Group B = C3H, C57, CBA, DBA). Expression was ×2.4 greater in group A than group B (P = 2.85×10−9, unpaired t-test) (Figure 1A). Where incubation time data are available, the increase in Hectd2 expression is associated with a shorter incubation time (R2 = 0.61) [20]–[22] (See also Figure S2). A potential role for Hectd2 in prion disease pathogenesis was explored by comparing the mRNA expression levels between normal mice and those at the end stage of disease following infection with Chandler/RML prions. For C57BL/6, expression was ×5.0 greater in the prion infected mice (P = 2.66×10−8, unpaired t-test) (Figure 1B).
Our data indicate that Hectd2 influences prion disease incubation time in mice. We therefore analysed HECTD2 in a hypothesis-driven association study of human prion disease. We analysed 834 samples from patients with prion disease or strong resistance to prion disease and 1162 relevant control population samples. We tested whether genetic variation at HECTD2 was associated with a phenotype of variant and sporadic CJD. In Papua New Guinea (PNG) we genotyped patients who died from the epidemic prion disease kuru, transmitted by endocannibalism, and compared these data with elderly women known to have had multiple exposures to kuru at mortuary feasts prior to the cessation of endocannibalism in the late 1950's, but who are long-term survivors [2],[23]. See methods for details of the patient data, populations, phenotype ascertainment and population stratification data.
We initially tested a single SNP, rs12249854(A/T), located in a HECTD2 intron, and showed that the minor allele (A) was significantly over-represented in vCJD (n = 117, 8.1%) compared to controls (n = 601, 3.9%), P = 0.0049, (OR 2.11, 95% CI 1.19–3.77, trend test 1 d.f.), and between sporadic CJD (n = 452, 6.3%) and controls, P = 0.012, (OR 1.65, 95% CI 1.11–2.46, trend test 1 d.f.). Given that sample sizes are necessarily small in both sporadic and variant CJD, these data are consistent with the association of rs12249854 with risk in both prion disease categories and a large effect size. We went on to test whether the risk rs12249854 allele modified the phenotype of human prion disease. Although the age of onset of sporadic CJD with rs12249854AA was younger than other genotypes (53.5 years Vs 68.8 years for rs12249854AT and 69.0 years for rs12249854TT, P = 0.048 t-test), this genotype was rare (n = 3) and the finding therefore was not robust. There was no association between vCJD year of presentation, or age of onset with rs12249854 genotype. Insufficient data were available to look at any association with duration of illness.
We went on to analyse a further seven SNPs in HECTD2 selected to capture global genetic diversity based on Hapmap (http://hapmap.org) [24] data (Table S6). In the United Kingdom (UK), we found strong linkage disequilibrium (LD) and a simple haplotype structure across the entire gene (Table 3). Three haplotypes were >1% frequency, the most common two haplotypes (1 and 2) differed at all SNPs, a third haplotype (3) was distinguished from the most common haplotype (1) by a single SNP upstream of HECTD2. Increased risk of vCJD was associated with haplotype 2, possessing rs12249854A, but the extensive LD prevented us from identifying the functional SNP. In PNG, however, we found considerably more diversity with four common haplotypes, 1, 2 and two novel haplotypes 4 and 5 (see methods for haplotype inference). Haplotype 2, most significantly associated with vCJD (haplotype association test, P = 0.006), showed no significance between kuru and the elderly female survivors of mortuary feasts. Rather, in PNG we found that a population specific haplotype (designated 4) was strongly associated with kuru (P = 0.0009). Haplotype 4 differs from haplotype 2 at a single SNP, rs12247672, which itself is significant in vCJD (P = 0.0039) but not at all in kuru (P = 0.6138). Our data suggest that there is evidence for HECTD2 association in both vCJD and kuru however the functional polymorphisms are likely to be different. This is not necessarily surprising given the distinct evolutionary history and consequent genetic differences that exist between the UK and PNG populations. It should also be noted that although vCJD and kuru are both acquired human prion diseases that share many characteristics they are also derived from different sources and caused by distinct prion strains [25],[26] therefore the mechanism of HECTD2 involvement may also be different.
We sequenced the ORF and promoter of HECTD2 in 16 vCJD, multi kuru-exposure survivors, and both UK and PNG controls. Three polymorphisms were found, of which only one is potentially functional (Table 4). rs7081363 occurs in the promoter (−247) and the minor allele is predicted to remove an Sp1 binding site (GGCG/AGG). rs7081363 was genotyped in our samples and shown to be in complete LD with rs12249854 in the UK population (vCJD P = 0.0012; sporadic CJD P = 0.0065). We were unable to genotype the kuru samples due to poor DNA quality, however, analysis of all other samples suggest that rs7081363 is unlikely to be significant in PNG.
To determine whether the susceptibility alleles in the UK population are associated with differential mRNA expression, HECTD2 expression levels in blood lymphocytes (n = 140, UK blood donors) were quantified by real-time RT-PCR. Samples were grouped according to rs12249854 genotype, however, due to the low frequency of the minor allele (A), no homozygotes (AA) were seen. The mean expression level was ×2.3 greater in the heterozygotes than for the major allele homozygotes (TT) (P = 0.0008 Mann-Whitney test, Figure 1C). This suggests that a higher level of HECTD2 mRNA expression may be linked with vCJD in the UK population.
Our data show that HECTD2 is linked to prion disease incubation time in mouse and is associated with sporadic and variant CJD and kuru in humans and an increase in expression is associated with a susceptibility genotype and disease pathogenesis. In mouse, we cannot exclude the possibility of other nearby genes or intergenic regions also being implicated as our sequencing studies were not exhaustive. However, in human, the LD block, based on HapMap [24] data, includes only HECTD2 and does not extend into the neighbouring genes suggesting that the association observed stems from HECTD2 and not any other gene in the area.
In mouse, the promoter, 3′UTR polymorphisms and the associated differential expression suggest a mechanism by which Hectd2 may influence the incubation time phenotype. Similarly, in the UK population a promoter polymorphism is also associated with a susceptibility phenotype and a resulting increase in expression level. This suggests that the mode of HECTD2 action in prion disease may be independent of host and prion strain. Due to lack of available material it has not been possible to replicate these experiments in our kuru samples, however, our haplotype study suggest that a different polymorphism is likely to be functional in the PNG population. This does not rule out the possibility that differential expression is also important in PNG, through an alternative polymorphism, although this may be difficult to determine. Our expression analysis in terminally sick mice suggest that HECTD2 is upregulated during the course of infection therefore we can speculate that a higher base line of expression reduces the time taken to reach a threshold level thereby reducing the incubation time.
The ubiquitin-proteosome system has been implicated in the pathogenesis of several neurodegenerative diseases which show an accumulation of an abnormally folded protein including prion disease, Parkinson's disease and Alzheimer's disease [27]–[29]. By homology to other family members, HECTD2 is an E3 ubiquitin ligase suggesting that common pathways are involved in the neurodegenerative processes of these different diseases. Specifically, the mouse mahoganoid coat-colour mutation is found in the gene Mahogunin which is an E3 ubiquitin ligase [30]. A null mutation of Mahogunin causes an age-related progressive neurodegenerative phenotype characterised by spongiform degeneration, neuronal loss and astrocytosis. The phenotype resembles that of prion disease however there is no PrPSc accumulation. Mutations in the E3 ubiquitin ligase parkin are associated with autosomal recessive juvenile parkinsonism and loss of ubiquitin-protein ligase activity in patients has been shown to be associated with protein accumulation [31]. E3 ubiquitin ligases have also been implicated in the pathogenesis of polyglutamine diseases in particular it has been shown that mutations in the E6-AP ubiquitin ligase reduces the frequency of nuclear inclusions in mice expressing mutant ataxin-1 while accelerating the Purkinje cell pathology [32]. Further, HECTD2 maps to a region of human chromosome 10q previously linked with Alzheimer's disease [33] suggesting that HECTD2 may also be a susceptibility factor for Alzheimer's disease and other neurodegenerative disorders.
Group sizes for vCJD, kuru and elderly female survivors of mortuary feasts are of necessity small, however we believe that the combined weight of data from the mouse genetic studies, expression analyses and our association study of independent human prion diseases from different populations provide sufficient evidence to support a role for HECTD2 in prion disease. This supports a significant role for the ubiquitin-proteasome system in prion pathogenesis [27],[34],[35] and will contribute to modelling and understanding genetic risk of developing prion disease following BSE and secondary human prion exposure.
The clinical and laboratory studies were approved by the local research ethics committee of University College London Institute of Neurology and National Hospital for Neurology and Neurosurgery and by the Medical Research Advisory Committee of the Government of PNG. Full participation of the PNG communities involved was established and maintained through discussions with village leaders, communities, families and individuals.
28 pairs of Northport HS mice were obtained from R. Hitzemann (Portland, Oregon, USA) at generation 35. Offspring from these pairs were randomly mated to produce a total of 49 pairs. 1000 offspring (generation 37) were used for inoculation. All other inbred lines were obtained from Harlan, UK. Mice were identified by individual transponder tags (Trovan) and tail biopsies were obtained for DNA extraction. Mice were anaesthetized with isofluorane/O2 and inoculated intra-cerebrally into the right parietal lobe with 30 µl Chandler/RML prions as previously described [6]. Incubation time was calculated retrospectively after a definite diagnosis of scrapie had been made and defined as the number of days from inoculation to the onset of clinical signs [17]. All procedures were conducted in accordance with UK regulations (Local ethics approval and Home Office regulation) and international standards on animal welfare.
|
10.1371/journal.pntd.0000301 | Needles in the EST Haystack: Large-Scale Identification and Analysis of Excretory-Secretory (ES) Proteins in Parasitic Nematodes Using Expressed Sequence Tags (ESTs) | Parasitic nematodes of humans, other animals and plants continue to impose a significant public health and economic burden worldwide, due to the diseases they cause. Promising antiparasitic drug and vaccine candidates have been discovered from excreted or secreted (ES) proteins released from the parasite and exposed to the immune system of the host. Mining the entire expressed sequence tag (EST) data available from parasitic nematodes represents an approach to discover such ES targets.
In this study, we predicted, using EST2Secretome, a novel, high-throughput, computational workflow system, 4,710 ES proteins from 452,134 ESTs derived from 39 different species of nematodes, parasitic in animals (including humans) or plants. In total, 2,632, 786, and 1,292 ES proteins were predicted for animal-, human-, and plant-parasitic nematodes. Subsequently, we systematically analysed ES proteins using computational methods. Of these 4,710 proteins, 2,490 (52.8%) had orthologues in Caenorhabditis elegans, whereas 621 (13.8%) appeared to be novel, currently having no significant match to any molecule available in public databases. Of the C. elegans homologues, 267 had strong “loss-of-function” phenotypes by RNA interference (RNAi) in this nematode. We could functionally classify 1,948 (41.3%) sequences using the Gene Ontology (GO) terms, establish pathway associations for 573 (12.2%) sequences using Kyoto Encyclopaedia of Genes and Genomes (KEGG), and identify protein interaction partners for 1,774 (37.6%) molecules. We also mapped 758 (16.1%) proteins to protein domains including the nematode-specific protein family “transthyretin-like” and “chromadorea ALT,” considered as vaccine candidates against filariasis in humans.
We report the large-scale analysis of ES proteins inferred from EST data for a range of parasitic nematodes. This set of ES proteins provides an inventory of known and novel members of ES proteins as a foundation for studies focused on understanding the biology of parasitic nematodes and their interactions with their hosts, as well as for the development of novel drugs or vaccines for parasite intervention and control.
| Excretory-secretory (ES) proteins are an important class of proteins in many organisms, spanning from bacteria to human beings, and are potential drug targets for several diseases. In this study, we first developed a software platform, EST2Secretome, comprised of carefully selected computational tools to identify and analyse ES proteins from expressed sequence tags (ESTs). By employing EST2Secretome, we analysed 4,710 ES proteins derived from 0.5 million ESTs for 39 economically important and disease-causing parasites from the phylum Nematoda. Several known and novel ES proteins that were either parasite- or nematode-specific were discovered, focussing on those that are either absent from or very divergent from similar molecules in their animal or plant hosts. In addition, we found many nematode-specific protein families of domains “transthyretin-like” and “chromadorea ALT,” considered vaccine candidates for filariasis in humans. We report numerous C. elegans homologues with loss-of-function RNAi phenotypes essential for parasite survival and therefore potential targets for parasite intervention. Overall, by developing freely available software to analyse large-scale EST data, we enabled researchers working on parasites for neglected tropical diseases to select specific genes and/or proteins to carry out directed functional assays for demystifying the molecular complexities of host–parasite interactions in a cell.
| Molecules secreted by a cell, often referred to excretory/secretory (ES) products, play pivotal biological roles across a diverse range of taxa, ranging from bacteria to mammals [1]. ES proteins can represent 8±20% of the proteome of an organism [1],[2]. ES proteins include functionally diverse classes of molecules, such as cytokines, chemokines, hormones, digestive enzymes, antibodies, extracellular proteinases, morphogens, toxins and antimicrobial peptides. Some of these proteins are known to be involved in vital biological processes, including cell adhesion, cell migration, cell-cell communication, differentiation, proliferation, morphogenesis and the regulation of immune responses [3]. ES proteins can circulate throughout the body of an organism (in the extracellular space), are localized to or released from the cell surface, making them readily accessible to drugs and/or the immune system. These characteristics make them attractive as targets for novel therapeutics, which are currently the focus of major drug discovery research programmes [4]. For example, knowledge of the molecular basis of secretory pathways in bacteria has facilitated the rational design of heterologous protein production pathways in biotechnology and in the development of novel antibiotics. From a more fundamental perspective, proteins secreted by pathogens are of particular interest in relation to the pathogen-host interactions, because they are present or active at the interface between the parasite and host cells, and can regulate the host response and/or cause disease [5],[6].
ES proteins have long been the focus of biochemical and immunological studies of parasitic helminths, as such worms secrete biologically active mediators which can modify or customize their niche within the host, in order to evade immune attack or to regulate or stimulate a particular host response [7],[8],[9],[10]. Parasitic nematodes are responsible for a range of neglected tropical diseases, such as ancylostomatosis, necatoriasis, lymphatic filariasis, onchocerciasis, ascariasis and strongyloidiasis in humans [11],[12], and others can cause massive production or economic losses to farmers as well as to animal and plant industries [13].
There have been efforts to identify and characterize ES proteins in different parasitic nematodes in various studies. For instance, Robinson et al. [14] used a proteomic approach to identify ES glycoproteins in Trichinella spiralis, an enoplid nematode (or trichina) of musculature. In another effort, Yatsuda et al. [9] undertook an analysis of ES products from Haemonchus contortus (barber's pole worm), a parasite of small ruminants; these authors identified several novel and known proteins but were only able (based on comparative analysis) to investigate known proteins, such as serine, metallo- and aspartyl- proteases and the microsomal peptidase H11, a vaccine candidate, previously recognised as a “hidden antigen” [15]. The precise role of ES proteins from parasitic nematodes in mediating cellular processes is largely unknown due to the difficulty in experimentally assigning function to individual proteins [14]. In this context, computational approaches applied to identify and annotate ES proteins have significantly complemented experimental studies of different cells, tissues, organs and organisms. For example, in an early study, Grimmond et al. [16] developed a computational strategy to identify and functionally classify secreted proteins in the mouse, based on the presence of a cleavable signal peptide (required for its entry into the secretory pathway), along with the lack of any transmembrane (TM) domain or intracellular localization signals, in full-length molecules. This study was followed by the computational reconstruction of the secretome in human skeletal muscle from protein sequence data by Bortoluzzi et al. [17]. Also, Martinez et al. [18] identified and annotated the secreted proteins involved in the early development of the kidney in the mouse from microarray ‘expression’ profiling, using computational strategies.
While expressed sequence tag (EST) data have been mined for many interesting functional molecules [19],[20], predicting ES proteins from ESTs has been relatively uncommon. For example, Vanholme et al. [21] identified putative secreted proteins from EST data sets for the plant parasitic nematode, Heterodera schachtii. Harcus et al. [22] investigated the signal sequences inferred from the EST data for the parasitic nematode Nippostrongylus brasiliensis, and related them to “accelerated evolution” of secreted proteins in this parasite, compared with host or non-parasitic organisms. Ranganathan et al. [23] identified ES proteins from EST data for the bovine lungworm, Dictyocaulus viviparus, whereas Nagaraj et al. [24] identified and classified putative secreted proteins from Trichostrongylus vitrinus, a parasitic nematode of ruminants and suggested some molecules as candidates for developing novel anthelmintics or vaccines. One of the suggested molecules, Tv-stp1, was investigated further and functionality established [25].
While single EST or protein data sets have been examined for the presence of secretory or ES proteins, large-scale analysis has not been conducted to date, due to the lack of effective high-throughput, computational pipelines for analysis [16]. Recently, we designed a high-throughput EST analysis pipeline, ESTExplorer [26] to provide comprehensive DNA and protein-level annotations. Based on earlier work [23],[24], ESTExplorer has been adapted to predict ES proteins with high confidence, and then provide extensive annotation, including Gene Ontologies (GO), pathway mapping, protein domain identification and predict protein-protein interactions. Our new pipeline, EST2Secretome, is a freely available web server that can directly process vast amounts of EST data or entire proteomes.
In the present study, approximately 500,000 ESTs, representing 39 economically important and disease-causing parasitic nematodes of humans, other animals and plants, were subjected to a comprehensive analysis and detailed annotation of inferred ES proteins using EST2Secretome, with specific reference to candidate molecules already being assessed as intervention targets. We compared the predicted ES proteins with those inferred from the free-living nematode C. elegans, to establish whether these proteins could be nematode-specific and propose their functionality. Also, we examined whether the ES proteins had homologues in their respective hosts (animal, human or plant), as such proteins and their genes are less likely to be useful as intervention targets. Pathway, interactome and literature-based ES protein analyses have assisted in gleaning sets of candidate molecules for future experimental studies. The present results lay a foundation for understanding the functional complexity of ES proteins from parasitic nematodes and their interactions with other proteins (within the nematodes) and/or with host proteomes.
EST2Secretome (http://EST2secretome.biolinfo.org/) is a comprehensive workflow system comprising carefully selected computational tools to identify and annotate ES proteins inferred from ESTs. EST2Secretome provides a user-friendly interface and detailed online help to assist researchers in the analysis of EST data sets for ES proteins. The workflow can be divided into three phases, with Phase I dedicated to pre-processing, assembly and conceptual translation, similar to that of ESTExplorer (details described in Nagaraj et al. [26]). In Phase II, putative ES proteins are identified based on the presence of signal sequences and the absence of transmembrane helices. Phase III contains a comprehensive annotation layer, comprising a suite of bioinformatic tools to annotate the ES proteins inferred in Phase II. ESTs can be submitted to Phase I for EST pre-processing, assembly and conceptual translation, followed by the identification of putative ES proteins in Phase II and annotation in Phase III. Alternatively, instead of EST data, protein sequences may be submitted directly to Phase II to identify putative ES proteins and functionally annotate them in Phase III.
Phase I of EST2Secretome shares SeqClean, RepeatMasker and CAP3 (contig assembly program) programs with ESTExplorer [26], based on the analysis presented elsewhere [20]. The contig and singleton sequences generated by CAP3 are transferred to the program ESTScan [27] for conceptual translation into proteins, using the genetic code from the nearest organism. EST2Secretome currently implements the genetic codes for 15 organisms, covering the most studied organisms, including human, mouse, rat, pig, dog, chicken, rice, wheat, thale cress (Arabidopsis thaliana), zebrafish, fly, yeast and a free-living roundworm (Caenorhabditis elegans) (Figure 1).
In Phase II, putative ES proteins are identified from the protein sequences generated in Phase I, using the two programs SignalP [28] and TMHMM [29] (Figure 1). SignalP first checks whether a signal sequence [30] is predicted both the artificial neural network and the hidden Markov model probability scores (SignalPNN and SignalP-HMM), using default parameters that can be modified by experienced users. Subsequently, all proteins with signal sequences are passed on to TMHMM [29], a hidden Markov model-based transmembrane helix prediction program, to “filter out” of transmembrane proteins. The subset lacking transmembrane helices is selected as ES proteins for further annotation.
Phase III is the annotation layer, comprising a suite of six computational tools for the functional annotation of ES proteins, of which the first three (Gene Ontology using BLAST2GO, InterProScan and pathway mapping using KOBAS) are also implemented in ESTExplorer and described elsewhere [26]. The other three components are unique to EST2Secretome and incorporate protein BLAST searches against three different data sets derived from Wormpep [31] for locating nematode homologues, IntAct [32] for protein-protein interaction data and a non-redundant known secreted protein database (SecProtSearch) derived from the literature, the secreted protein database, SPD [33] and the manually curated signal peptide database, SPdb [34]. Mapping to Wormpep gives a list of homologous proteins in C. elegans, linked to WormBase [31]. Homologues from the IntAct database are determined using the concept of interlogs (evolutionarily conserved interactions identified by conservation among homologous proteins in different species) and are linked to all molecular interaction partners of homologous proteins. EST2Secretome provides a link to the relevant interlog page at IntAct, containing all interaction partners. The interaction data culled from these interlogs can be extrapolated to predict protein interactions of the query sequence, for validation by complementary double-stranded RNA interference (RNAi), gene deletion or fluorescence-based interaction studies. The final module compares the query sequence to a specialised data set of known secreted proteins (SecProtSearch), in order to identify orthologous secreted proteins, which would provide a second level of validation for the ES protein dataset. Phase III (Figure 1) thus allows extensive characterization and validation of ES proteins predicted by EST2Secretome.
Once an EST (or a protein dataset) has been submitted to EST2Secretome, a status page is accessible, for the monitoring of the progress of the analysis, at the program level. As each selected program is completed, the status page is updated and the output from that program becomes available. The outcome from each run is summarized, with links to output files from each selected program being listed. When a large dataset is analysed using a workflow, it is challenging to collate the results of the analysis from multiple steps. To address this issue, EST2Secretome provides a summary file for each ES protein, comprising the assembled contig/singleton sequence, the peptide sequence and all the annotations (such as homologous proteins, protein domains, pathways and interaction partners).
The details of the EST2Secretome workflow, including the software and hardware used, are provided on the website. A detailed tutorial, frequently asked questions (FAQ) and sample EST and protein datasets are available online for the effective use of EST2Secretome.
452,134 ESTs (as at 18 December 2007) from 39 parasitic nematodes (7 from human, 18 from other animals and 14 from plants, Table 1) were downloaded from dbEST [19]. ESTs from each organism were submitted to Phase I of EST2Secretome, where they were pre-processed (SeqClean and RepeatMasker), aligned/clustered using CAP3 [35], with a minimum sequence overlap length “cut-off” of 30 bases and an identity threshold of 90%, for the removal of flanking vector and adapter sequences, followed by assembly. These high quality contigs and singletons were conceptually translated using ESTScan [27], based on a “smat” matrix, generated from available mRNA data for each organism. When the smat file for a specific organism is not available, the nearest well-studied organism has to be selected as a reference, based on taxonomy, and its smat file is used instead. We used data (25,481 cDNA sequences) from C. elegans (as it is the best studied nematode) for the generation of the smat file. The conceptually translated peptide data were transferred to Phase II of EST2Secretome, for the prediction of ES proteins, by sequentially running the SignalP [28] and TMHMM [29] programs. For SignalP, the threshold values for the D-score and the Signal peptide probability were both set to 0.5, based on a validation carried out for 1946 sequences of experimentally verified signal peptides from the recently updated SPdb [34], with an accuracy of prediction of 98.1%. Any protein that simultaneously fulfilled the threshold set for both the D-score and the Signal peptide probability score, was classified as a secretory-excretory (ES) protein. Inferred ES proteins were then tested for the presence of transmembrane domains using the transmembrane helix and membrane topology prediction program, TMHMM [29] and sequences containing predicted transmembrane regions were eliminated to yield only those proteins that were predicted as destined for secretion.
Inferred ES proteins were annotated by selecting all of the programs in Phase III of the EST2Secretome. Gene Ontology (GO) [36] terms were assigned using BLAST2GO (v 1.6.2) [37]. Sequences were then mapped to biological pathways employing the KEGG Orthology-Based Annotation System (KOBAS) [38], with C. elegans data selected for the construction of background pathway maps. The query sequences were then compared using BLASTP against Wormpep v183 (e-value threshold of 1e-05). For each predicted ES sequence, the protein domain/family/motif was mapped using InterProScan [39], including 13 member databases, and the results were tabulated in decreasing order of abundance. Inferred ES protein sequence data were queried against the IntAct database (version 1.7.0) [32] to retrieve all interaction partners (e-value threshold of 1e-05). A comparison of homologues, based on BLAST scores from three different datasets, can be efficiently compared and presented visually using the program SimiTri [40]. In the case of parasitic nematodes, we generated BLAST-indexed datasets for the host organisms (human, other mammals or plant), C. elegans as the primary reference organism for nematodes and parasitic nematodes, based on NCBI protein datasets (defined by keyword), followed by local processing to add or remove selected organisms.
EST2Secretome made possible the large-scale analysis and annotation of all publicly available EST data for nematodes that are parasitic in humans, other animals and plants. In total, 452,134 ESTs from 39 parasitic nematodes were downloaded from dbEST [19]. The organisms were broadly categorised on the basis of the host(s) they infect (Table 1) with seven, 18 and 14 nematodes parasitic in humans, other animals and plants, respectively, being selected for secretome analysis. Putative ES proteins were identified in the first two phases of EST2Secretome (see Figure 2). Phase I pre-processing and assembly resulted in a total of 152,702 representative ESTs (rESTs) comprising 53,377 contigs and 99,326 singletons, with 152,702 rESTs being conceptually translated into 101,514 peptide sequences. In Phase II, these conceptually translated peptide sequences were first analysed for the presence of N-terminal signal peptide, followed by the absence of transmembrane helices. We thus identified a total of 4,710 putative soluble ES proteins (2,632 in animal-, 1,292 in plant- and 786 in human-parasitic nematodes) (see Table 2), representing 4.6% of the total number of putative sequences identified. This result is comparable with earlier single organism studies of the bovine lungworm, D. viviparus [23], in which 85 secreted proteins were identified (representing 5.0% of 1685 peptides) and T. vitirinus [24], in which 40 secreted proteins were identified (representing 6.2% of 640 proteins). We manually examined the ES protein sequence data and found that 14 of 4710 entries were low quality sequences containing predominantly long stretches of unknown amino acids (X's), as a result of repeat masking, followed by conceptual translation. These sequences were from organisms like Meloidogyne chitwoodi and Pratylenchus vulnus which lack repeat libraries. These 14 sequences were functionally analysed and annotated in the EST2Secretome pipeline but could not be assigned any function. This step represents one of the challenges involved in the computational analysis of single pass reads from any organism which is not well characterized based on genomic data.
We employed EST2Secretome for the analysis of the entire proteome (23,624 sequences) of the model free living nematode, C. elegans, in the Wormpep database (18th February 2008). 2,649 (11.2%) sequences were predicted to be ES proteins, which is in the range of 8–20% suggested by Grimmond et al. [16]. These results independently validated the ability of the EST2Secretome pipeline to correctly identify ES proteins, using the Phase II filtering steps. The lower percentage of 4.6% ES proteins from EST data compared to 11.2% in C. elegans could be due the low coverage of the entire protein-coding gene set, compared to entire proteome comprising full length protein sequences in C. elegans, or to the low quality of some ESTs in public databases.
We carried out a comprehensive analysis of the 4,710 ES proteins predicted, using all relevant components of Phase III in EST2Secretome as well as some additional bioinformatic tools specific to nematodes (Figure 2). Functional annotation comprised the assignment of GO terms and pathway associations using KEGG pathways; mapping protein domains/motifs, with a particular focus on nematode-specificity and identifying protein interaction partners. Subsequently, we used comparative genomics approaches to identify orthologues in the free-living nematode C. elegans, with their associated loss-of-function RNAi phenotypes. From database comparisons with human, other animal and plant host sequences, we predicted several ES proteins that were either absent from their host or distantly related to host homologues, which might represent potential novel drug or vaccine targets for parasite intervention. Results of these analyses are described in the following sections.
C. elegans represents the best characterized nematode in many respects, particularly in terms of its genome, genetics, biology, physiology and biochemistry [31],[54],[55]. In addition, C. elegans (non-wild-type or loss-of-function) RNAi phenotypes may provide indications of the relevance and function(s) of homologous genes in other nematodes (of animals) for which the complexity of an obligate parasitic life cycle and the lack of an effective in vitro culture system and/or an RNAi assay make high-throughput screening impractical [56]. Moreover, the set of genes with RNAi loss-of-function phenotypes constitutes a pool of significant and potentially essential C. elegans genes. The RNAi phenotype data, comprising, ∼62,000 entries (on 10 January 2008), is available to download through WormBase [31]. In this study, we compared the 4,710 predicted ES proteins to the C. elegans proteome using BLASTP program and predicted 2,490 (52.8%) homologues in C. elegans (threshold e-value of 1e-05). From these 2,490 C. elegans homologues, we retrieved exclusively protein entries that had been reported with any one of the following observed strong RNAi phenotypes: Emb (embryonic lethal, including pleiotropic defects severe early emb), Lvl (larval lethal), Lva (larval arrest), Stp (sterile progeny), Ste (maternal sterile) and Gro (slow growth). In the present dataset (available from Table S5), 267 C. elegans homologues were identified that had one or more observed “strong” loss-of-function phenotype in RNAi; selected examples are listed in Table 6. The latter RNAi phenotypes were selected as they have been inferred to be essential for nematode survival or growth [56],[57], also representing potential drug and/or vaccine targets.
Sequence-based searches were performed to classify the ES proteins, to identify the presence or absence of putative homologues in C. elegans, and to infer nematode-specific and parasite-specific genes. For parasitic nematodes, Parkinson et al. [40],[58] suggested previously that it is beneficial to make simultaneous three-way comparisons (using SimiTri) of a specific organism or a group of organisms with homologues in C. elegans (the ‘model nematode’), other nematode species as well as the host organism. Such an analysis provides a means for the rapid identification of genes/proteins conserved between any two datasets compared (e.g., between parasitic nematodes and free-living ones, or between parasitic nematode and its host). In the present study, we systematically compared inferred ES protein data with those available in three relevant databases. For the three ES protein datasets from nematodes parasitic in humans (786 proteins), animals (2,632 proteins) or plants (1,292 proteins), we selected C. elegans and parasitic nematode databases as well as databases specific to the host organisms for comparative analysis. For instance, data for parasitic nematodes of humans were matched with those of the human host, C. elegans and parasitic nematodes from other hosts. Similarly, ES proteins predicted for nematodes parasitic in animals or plants were compared against host datasets. Protein sequences available in the following three datasets (i) C. elegans (from Wormpep [31]), (ii) parasitic nematodes (constructed locally) and (iii) respective hosts (human, other animal and plants sequences from NCBI non-redundant protein database) were processed. Three-way comparison of the parasitic nematode database with homologues in C. elegans, their principal definitive host organism (human, other animal or plant) and the database of all available parasitic nematodes, have been presented using SimiTri [40] in Figure 4. In all three datasets for parasitic nematodes, inferred ES proteins congregated with parasitic nematodes rather than with C. elegans or with the host species (lower right hand corner of each triangle, coloured in red in Figure 4). Overall, 320 (40.7%), 789 (29.7%) and 581 (44.9%) ES proteins inferred from human-, other animal- and plant-parasitic nematodes were associated exclusively with parasitic nematodes and are interpreted to be parasite-specific, based on the data currently available. Of the homologues predicted to be nematode-specific (along the side of the triangle connecting C. elegans and parasitic nematodes), 585 (74.4%), 1,511 (57.4%) and 1,034 (80.0%) of the inferred ES proteins were confined to nematodes (based on currently available datasets). Based on these comparisons, we illustrate that a significant percentage of these proteins in parasitic nematodes are either parasite- or nematode-specific and are either absent from or very divergent in sequence from molecules in their host(s). These molecules might represent candidate targets for novel anthelmintics for parasite intervention. Importantly, their apparent specificity to parasitic nematodes or different groups within the phylum Nematoda renders them as important groups of molecules for future study, particularly in relation to the roles of these molecules in the host-parasite interplay, their involvement in inducing immune responses and disease in the host.
Based on evidence from the literature, we selected candidate molecules from parasitic nematodes which have already proven to be therapeutic or vaccine targets for scrutiny. Such targets are either in early phases of clinical trials or have been identified as candidates following detailed experimental study. Firstly, prominent anti-parasite vaccine candidates have been identified through the Human Hookworm Vaccine Initiative and include a family of pathogenesis-related (PR) proteins, such as the Ancylostoma-secreted proteins (ASPs) [59]. This initiative has characterized Na-ASP-2, a PR-1 protein, from Necator americanus [59] which is in Phase II clinical trials [60] and Ac-ASP-1 from Ancylostoma caninum which exhibits 97% identity to Na-ASP-2 [61]. Secondly, cathepsin L and Z-like cysteine proteases (known to have been implicated in moulting and tissue remodelling in free-living and parasitic nematodes) represent potential targets for onchocerciasis and have been studied in significant detail in Onchocerca volvulus [62],[63],[64]. Also, astacin-like metalloproteases (MTP) was selected, as L3s of parasitic nematodes secrete MTPs that are considered critical to invasion and establishment of the parasite in the host [65],[66]. Astacin-like MTPs, such as MTP-1, have been characterized mainly in Ancylostoma caninum and are secreted by infective hookworm larvae [66],[67]. The sequences for four such proteins were retrieved from NCBI and matched to the present ES dataset using BLASTP. We discovered likely homologues for all of these proteins in parasitic nematodes of humans, other animals and plants (Table 7); organisms for which there is published information on these proteins are indicated (in bold font). Based on the present analysis, we identified 12 homologues of Ancylostoma-secreted proteins (ASPs) (above the threshold e-value of 1e-05) in the datasets in following nematodes (Strongylida): Necator americanus, Ancylostoma duodenale, Ancylostoma caninum, Haemonchus contortus and Teladorsagia circumcincta. Of these, published reports are available for only Necator americanus, Ancylostoma caninum, Haemonchus contortus and Ostertagia ostertagi [7],[61],[65],[66], while the analysis, based exclusively on available data, showed that this group of proteins (inferred from ESTs) occurs in the parasitic nematodes Teladorsagia circumcincta and Meloidogyne chitwoodi. Moreover, we identified eleven cathepsin L-like cysteine proteases, nine cathepsin Z-like cysteine proteinases and eight astacin-like metalloproteases in ES protein datasets, providing novel, yet unpublished evidence for the presence of these proteins in a number of key parasitic nematodes of socio-economic importance.
In this study, based on a comprehensive, targeted analysis of almost 0.5 million publicly available ESTs, we have inferred and functionally annotated 4,710 putative ES proteins from 39 parasitic nematodes infecting humans, other animals or plants, using the EST2Secretome, a new workflow developed for the large-scale processing of EST and complete proteome data. Furthermore, EST2Secretome has been developed as a multi-purpose, high-throughput analysis pipeline for diverse applications. For instance, it is possible to conduct analyses of all predicted proteins containing only signal sequences by selecting only SignalP and deselecting the TMHMM option, or select only the TMHMM program to investigate transmembrane proteins. The option to enter protein sequence data alone into the pipeline is also useful following the direct sequencing of proteins in proteomic studies.
Detailed annotations of inferred ES proteins revealed several parasite-specific (being absent from C. elegans and the host) and nematode-specific molecules as potential drug or vaccine candidates. Included in this set of molecules are pathogen-related protein (PRP) domains and several novel, nematode-specific protein domains. Gene Ontology (GO) annotations, at the level of molecular function, revealed an overwhelming representation of binding (63.4%) and catalytic activity (54.1%), supporting the further biochemical, proteomic and/or functional characterization of the ES proteins inferred herein. Predicted protein interaction data for each ES protein enables the classification of molecules as essential for parasite existence or survival, with relative potential to serve as target for parasite intervention, based on the number of primary and secondary interaction partners, as well as those interactions that are specific to parasites, rendering such “hub proteins” as potential targets for functional studies. In order to predict which ES proteins are essential, we also categorised molecules according to “strong” loss-of-function RNAi phenotypes for corresponding homologues in C. elegans. ES proteins homologous to these “loss-of-function” phenotypes are considered the best candidates for functional characterization, and possibly linked to the survival of the parasites. Finally, we selected some proteins for further characterization based on their similarity to proteins currently under evaluation as vaccines or drug targets. The present, systematic approach of inferring ES protein data from EST data sets represents a starting point for understanding the role ES proteins in parasitic nematodes and serves as a useful tool for the future study of essentially any eukaryotic organism.
|
10.1371/journal.ppat.1000260 | Control of Stochastic Gene Expression by Host Factors at the HIV Promoter | The HIV promoter within the viral long terminal repeat (LTR) orchestrates many aspects of the viral life cycle, from the dynamics of viral gene expression and replication to the establishment of a latent state. In particular, after viral integration into the host genome, stochastic fluctuations in viral gene expression amplified by the Tat positive feedback loop can contribute to the formation of either a productive, transactivated state or an inactive state. In a significant fraction of cells harboring an integrated copy of the HIV-1 model provirus (LTR-GFP-IRES-Tat), this bimodal gene expression profile is dynamic, as cells spontaneously and continuously flip between active (Bright) and inactive (Off) expression modes. Furthermore, these switching dynamics may contribute to the establishment and maintenance of proviral latency, because after viral integration long delays in gene expression can occur before viral transactivation. The HIV-1 promoter contains cis-acting Sp1 and NF-κB elements that regulate gene expression via the recruitment of both activating and repressing complexes. We hypothesized that interplay in the recruitment of such positive and negative factors could modulate the stability of the Bright and Off modes and thereby alter the sensitivity of viral gene expression to stochastic fluctuations in the Tat feedback loop. Using model lentivirus variants with mutations introduced in the Sp1 and NF-κB elements, we employed flow cytometry, mRNA quantification, pharmacological perturbations, and chromatin immunoprecipitation to reveal significant functional differences in contributions of each site to viral gene regulation. Specifically, the Sp1 sites apparently stabilize both the Bright and the Off states, such that their mutation promotes noisy gene expression and reduction in the regulation of histone acetylation and deacetylation. Furthermore, the NF-κB sites exhibit distinct properties, with κB site I serving a stronger activating role than κB site II. Moreover, Sp1 site III plays a particularly important role in the recruitment of both p300 and RelA to the promoter. Finally, analysis of 362 clonal cell populations infected with the viral variants revealed that mutations in any of the Sp1 sites yield a 6-fold higher frequency of clonal bifurcation compared to that of the wild-type promoter. Thus, each Sp1 and NF-κB site differentially contributes to the regulation of viral gene expression, and Sp1 sites functionally “dampen” transcriptional noise and thereby modulate the frequency and maintenance of this model of viral latency. These results may have biomedical implications for the treatment of HIV latency.
| After HIV genome integration into the host chromosome, the viral promoter coordinates a complex set of inputs to control the establishment of viral latency, the onset of viral gene expression, and the ensuing gene expression levels. Among these inputs are chromatin structure at the site of integration, host transcription factors, and the virally encoded transcriptional regulator Tat. Importantly, transcriptional noise from host and viral transcriptional regulators may play a critical role in the decision between replication versus latency, because stochastic fluctuations in gene expression are amplified by a Tat-mediated positive transcriptional feedback loop. To evaluate the individual contributions of key transcription factor binding elements in gene expression dynamics, we employ model HIV viruses with mutations introduced into numerous promoter elements. Extensive analysis of gene expression dynamics and transcription factor recruitment to the viral promoter reveals that each site differentially contributes to viral gene expression and to the establishment of a low expression state that may contribute to viral latency. This systems-level approach elucidates the synergistic contributions of host and viral factors to the dynamics, magnitudes, and stochastic effects in viral gene expression, as well as provides insights into mechanisms that contribute to proviral latency.
| HIV-1 can establish rare, latent infections in cells, and the resulting viral reservoirs represent the most significant barrier to elimination of virus from a patient since they persist for decades and can reactivate at any time [1]. After HIV enters a cell, it integrates its genetic material into the host genome and utilizes host cell transcriptional machinery to regulate its gene expression. Briefly, initial expression from the HIV long terminal repeat (LTR) promoter is hindered by stalling of RNA polymerase II (RNAPII) [2], which results in a high frequency of abortive transcripts [3]. However, a low leaky or basal transcription generates a small fraction of fully elongated transcripts that yield viral mRNA encoding a positive regulator, the transcriptional activator (Tat) [4]. Tat binds to cyclin T1, a unit of the endogenous positive transcriptional elongation factor B (P-TEFb) [5],[6], and the Tat:P-TEFb complex binds to an RNA motif in stalled HIV transcripts known as the transactivation response element (TAR) [7]. In this complex, P-TEFb phosphorylates the C-terminal domain of RNAPII, thereby enhancing its processivity and enabling the efficient generation of fully elongated transcripts [2]. The net result is a strong positive feedback loop of Tat-mediated transactivation that amplifies viral transcriptional elongation nearly 100-fold [8].
We previously explored whether stochastic delays in the onset of HIV-1 Tat expression contribute to the formation of latent viral infections [9]. Genetic noise is an inherent and significant process in gene expression in bacteria [10],[11], yeast [12]–[15], and mammals [9],[16],[17]. In particular, stochastic effects most commonly become important in slow chemical reactions or with low concentrations of chemical species, both of which apply early in HIV gene expression when basal expression and Tat concentrations are low. Using a lentiviral model of the Tat-mediated positive feedback loop (LTR-GFP-IRES-Tat, or LGIT), we have demonstrated that random fluctuations in Tat levels could result in clonal cell populations that exhibited two distinct viral gene expression levels, Off and Bright—behavior we refer to as phenotypic bifurcation (PheB) [9]. Such bifurcating clonal populations, expanded from single cells each harboring a single viral integration position, exhibit dynamic gene expression behavior, with cells continuously switching between the two modes of gene expression. Moreover, integrated provirus can remain Off for extended periods of time before switching to a Bright expression level, suggesting that long delays in transactivation could contribute to postintegration viral latency [18],[19]. Here, we expand upon this work to study how host transcription factor binding sites at the HIV-1 LTR contribute to both the level of viral gene expression and noise in that gene expression, with a focus on potential implications for the establishment and persistence of viral latency.
Following preferential HIV-1 integration into regions of active chromatin [9],[18],[20],[21], transcription factor binding sites in the LTR recruit activating and repressing host cell transcription factors and thereby likely influence the basal viral gene expression, the maximal inducible rate of viral expression, and the dynamics of switching between these two states. In particular, binding sites for NF-κB, Sp1, YY1/LBP-1, AP-1, and other factors recruit chromatin modifying complexes to the HIV promoter (Figure 1A) [22],[23]. Activating complexes may recruit histone acetyltransferases (HATs) and thus contribute to stabilizing the transcriptionally active state of HIV in either a Tat- dependent or independent manner [24],[25]. Alternatively, numerous repressing complexes may recruit histone deacetylases (HDACs) that stabilize the transcriptionally inactive mode by chromatin deacetylation or via competition with activating complexes [3],[22].
In particular, the prototypical HIV clade B promoter contains two κB-binding sites and three tandem Sp1-binding sites (Figure 1A and 1B), all of which have the potential to recruit either repressing or activating complexes (Figure 1C). For example, the NF-κB p50-p50 homodimer complex binds to the κB binding sites and can recruit the repressive HDAC1 and HDAC3 factors [3],[26]. Alternatively, binding of the activating NF-κB p50-RelA heterodimer [27] enables interaction with p300 [28],[29], a HAT that is required for full Tat activity [24],[30]. The p50-RelA heterodimer can also interact with P-TEFb [31] and thereby aid RNAPII processivity [32],[33]. Similarly, Sp1 can interact with both HDACs and HATs [34],[35], and thus may mediate both repressing and activating transcriptional mechanisms.
Modulation of HIV gene expression with cytokines and other pharmacological agents that function via NF-κB or Sp1 dependent mechanisms further demonstrates the importance of these sites to promoter regulation. For example, tumor necrosis factor alpha (TNF-α) activates HIV transcription by increasing the nuclear concentration of RelA, thereby increasing the availability of p50-RelA to bind κB sites [36]. In addition, trichostatin A (TSA) activates transcription by inhibiting class I and II HDACs, which otherwise repress HIV gene expression by maintaining chromatin deacetylation [37],[38]. Since both Sp1 and κB sites facilitate recruitment of class I HDACs [3],[34],[39], both NF-κB- and Sp1-mediated repression are targets for TSA activation.
A number of important studies demonstrate that the deletion or mutation of any of the Sp1 or κB elements compromises the rates of gene expression and/or viral replication [40]–[45], though the effects of mutations or deletions on the establishment of latency were not explored. Moreover, although κB sites have been demonstrated to play important roles in both HIV activation and proviral latency [3],[32],[33],[46],[47], the interplay between multiple transcription factor binding sites, gene expression noise [9], and the choice between transcriptional activation and viral replication vs. genetic silencing and latency have not been examined. As we hypothesize that PheB integrants are likely poised at the edge between repressive and activating mechanisms, these proviruses may be highly sensitive to other sources of noise, including the dynamic competition between the recruitment of repressing and activating complexes at the Sp1 and κB sites (Figure 1C).
Here, we examine the roles of the κB and Sp1 elements in the context of a model of postintegration HIV latency to dissect the contributions of each site to gene expression dynamics, transcriptional activation and repression, noise in gene expression, and potentially proviral latency. Using gene expression analysis, pharmacological perturbations, chromatin immunoprecipitation, and analysis of transcriptional initiation and elongation, we demonstrate that each Sp1 site plays a significant role in the persistence of both active and inactive expression states. Furthermore, the two κB sites differentially recruit transcriptional regulators, where κB site I contributes more to transcriptional activation through the recruitment of p50-RelA heterodimer, while κB site II has a bias for the repressing p50-p50 complex. Interestingly, these sites play unique, and at times synergistic, roles in the transcriptional regulation events that underlie gene expression noise and potentially clinical HIV latency.
Inactivating point mutations [48]–[50] were introduced into each of the Sp1 and κB sites within the LTR of the LGIT virus plasmid (Figure 1B). These mutant versions of LGIT include an inactivating mutation of Sp1 site I (mutI Sp1), Sp1 site II (mutII Sp1), Sp1 site III (mutIII Sp1), all Sp1 sites (mutALL Sp1), κB site I (mutI NF-κB), κB site II (mutII NF-κB), a combination of κB sites I and II (mutI&II NF-κB), a full deletion of both κB sites (del NF-κB), and a combination of Sp1 site III and κB site I (mutIII Sp1/mutI NF-κB). After viral production, Jurkat cells were infected at a low multiplicity of infection (MOI ∼0.05–0.10), a level demonstrated to yield a polyclonal population of infected cells with a broad range of single viral integration sites per cell [9]. Viral LTR expression was monitored by flow cytometry for 21 days following the initial infection. Gene expression was detectable 48 hours post-infection, the first time point analyzed, and progressively increased over the course of a week. This timing is consistent with in vivo reports that reveal that viral production initiates approximately two days after infection following the viral “eclipse phase” [51]. Histograms for LGIT and mutant versions revealed a Bright, transactivated population and an Off population that included infected, inactive cells in addition to a larger population of uninfected cells (Figure S1B). However, for two LGIT variants, mutALL Sp1 and mutIII Sp1/mutI NF-κB, a Bright population of cells was not detected (Figure S1B), and these two mutant combinations were not further studied.
For the 21-day time course experiments, heat maps depicting the GFP intensity distribution of the infected cell populations indicate that mutations in the Sp1 sites substantially impact GFP expression (Figure S1A). The WT and mutant LGIT variants exhibited a similar temporal onset of gene expression and reached a maximum in the mean position of their bright peaks (Bright Mean)—a metric of gene expression in the Tat feedback loop—10 days after infection (Figure 2A). Importantly, mutation of any of the Sp1 sites (mutI Sp1, mutII Sp1, and mutIII Sp1) resulted in dramatic 40–50% decreases in the Bright Mean levels (Figure 2A). These results indicate that each Sp1 site has a considerable role in the transcriptional activation of the proviral LTR, with Sp1 site III appearing to have a slightly larger contribution than Sp1 site I or II (p<0.05 for each day after day 6).
While the Bright Mean characterizes the strength of Tat transactivation within the positive feedback loop, a smaller, less stable population of LGIT cells exhibits intermediate levels of gene expression. We have previously demonstrated that stochastic effects in gene expression are most evident at these intermediate levels of Tat and contribute to switching between Bright and Off modes [9]. Therefore, the fraction of cells that expresses GFP at intermediate or Mid fluorescence levels (i.e., the Mid:On ratio, where On is the sum of Mid and Bright regions, Figure 1C) is a measure of stochastic fluctuations in Tat expression. Mutations that further stabilize the Off or Bright mode would be predicted to result in a lower Mid:On ratio and reduced “flipping” between the two stabilized states. In contrast, mutations that destabilize the Off and Bright modes would yield an increase in the Mid:On ratio, via increasing the rate of flipping between the two “less stable” transcriptional states and thereby creating a noisier promoter. At early times after infection, the Mid:On ratio is high, as the gene expression of infected cells ramps up, but it later settles into an informative steady state value (Figure 2B). Over the three week time course, the Mid:On ratios for each of the Sp1 mutants remain 3- to 4-fold higher than WT. These data indicate that each of the Sp1 sites in the WT promoter may stabilize the Bright and potentially the Off mode, and that a reduction of this stabilization (consistent with the observed decrease in the Bright Mean position, Figure 2A) may increase the rates of switching between Off and Bright expression modes. Thus, based on the Mid:On ratio as a metric for stochastic behavior in the Tat-feedback circuit, the Sp1 sites appear to control promoter noise, with potential implications for viral latency.
In parallel experiments to the LGIT Sp1 mutants, mutation of each of the two κB sites in the HIV promoter reveals the roles of each site in stabilizing the Bright modes (Bright Mean) as well as dynamic flipping between modes (Mid:On ratio). Compared to WT LGIT, the κB site I mutant (mutI NF-κB) exhibited a decrease of the Bright Mean, whereas mutation of κB site II (mutII NF-κB) yielded a slight, but statistically significant (p<0.05 at two weeks after infection) increase of the Bright Mean (Figure 2A). Interestingly, these results indicate that the roles of the two κB sites in the HIV promoter are not redundant, with an intact site I serving an activating role and site II a slightly repressing role. Consistent with these observations, the double κB mutant (mutI&II NF-κB) exhibited gene expression levels closer to those of the WT promoter than mutI NF-κB, indicating that the loss of the repressing site II slightly counteracts the loss of the activating site I. In contrast to mutI&II NF-κB, del NF-κB exhibited a severe loss in gene expression, indicating that the complete deletion of the 24 nucleotides encompassing the κB sites had effects beyond the loss of NF-κB binding, perhaps through altered nucleosome spacing [36] or loss of the NFAT1 and GABP transcription factor binding sites at the 3′ ends of the κB sites [52],[53], which were not affected by the individual mutations in mutI&II NF-κB. To focus analysis specifically on the roles of κB recruitment, we did not pursue analysis of the variant with full deletion of both κB sites.
In contrast to the Sp1 mutants, mutation of the κB sites had modest effects on the Mid:On ratio compared to the WT LTR (Figure 2B). However, significant differences between κB sites I and II are evident, as mutII NF-κB had no change in the Mid:On ratio, but mutI NF-κB exhibited a 1.5-fold increase compared to the WT promoter. Thus, the observed decrease in the Bright Mean position of mutI NF-κB (Figure 2A) is consistent with destabilization of the Bright mode, resulting in noisier gene expression or an increased Mid:On ratio (Figure 2B).
Infecting cells at an MOI of 0.05–0.10 results in approximately 90–95% of cells being uninfected (Figure 3A, panel 1) as predicted by a Poisson distribution. However, a fraction of the infected cells may conceivably persist in the Off mode and thus be indistinguishable from the uninfected cells by flow cytometry. This fraction of “Infected but Off” cells provides additional insights into the relative stability of the Off and Bright modes for the different mutants. Specifically, increases in the fraction of Infected but Off cells suggest an increase in the stability of the Off mode or a decrease in the stability of the Bright mode, impeding cells from undergoing Tat transactivation. To measure the fraction of Infected but Off cells, we stimulated gene expression through simultaneous addition of exogenous Tat [18] and the hybrid polar compound hexamethylene bisacetamide (HMBA), which activates HIV transcription independent of the NF-κB pathway [54]. Six days after infection, cells were treated with 5 mM HMBA and Tat protein (8 µg per 3×105 cells) and incubated for 18 hours (Figure 3A, panel 2). This combined stimulation increased the fraction of the WT LGIT infected cells that expressed GFP by 17.0%±0.8% of infected cells, which would otherwise persist in the Off mode (Figure 3B).
Interestingly, all three Sp1 mutants exhibited considerably higher fractions of Infected but Off cells, peaking with mutIII Sp1 at 57.6%±3.7% (Figure 3B). In addition, mutation of κB site I (in both mutI NF-κB and mutI&II NF-κB), but not κB site II, resulted in a more modest but significant increase in the fraction of infected cells being Off (Figure 3B). Specifically, mutII NF-κB is indistinguishable from WT LGIT (p = 0.64), but mutI NF-κB exhibited statistically higher fractions of Infected but Off cells (p<0.01). These are consistent with our observations that the two κB sites are functionally different, with κB site I having a stronger activating role (Figure 2A). Collectively, these data indicate that loss of any of the Sp1 sites, and to a lesser degree κB site I, destabilizes the Bright, transactivated expression state.
To examine the relative stabilities and switching dynamics of the Bright and Off modes of expression, we sorted pure populations of infected cells that had persisted in the Bright mode (Bright sort, Figure 3A, panel 5) or relaxed into the Off mode (Off sort, Figure 3A, panel 6). The polyclonal Bright and Off sorts are phenotypically homogeneous populations of singly-integrated cells that represent a wide range (>105) of integration positions. The distribution of viral gene expression in Bright and Off modes was dynamic, and the stability of the Bright mode of the bimodal distribution of LGIT was determined by measuring the spontaneous inactivation or relaxation of Bright-sorted cells (Figure 3A, panel 5). Fourteen days after sorting the Bright population, the frequencies of spontaneous inactivation (%Loss of Bright) for each of the three individual Sp1 mutations (mutI Sp1, mutII Sp1, and mutIII Sp1) increased significantly compared to WT LGIT (Figure 3C). Consistent with previous findings (Figures 2A and 3B), this trend again indicates that each Sp1 site may contribute to the stability of the Bright mode. In contrast to the Sp1 mutants, the frequencies of spontaneous inactivation for the κB site mutants (mutI NF-κB and mutII NF-κB, and mutI&II NF-κB) were unchanged compared to WT LGIT (Figure 3C, p = 0.20 and 0.15, respectively), suggesting that the κB sites play a comparatively smaller role in the stability of the Bright mode.
To examine the stability of the Off mode we measured the spontaneous initiation of GFP expression from the Off-sorted cells (Figure 3A, panel 6), which we refer to as spontaneous activation. After 28 days of culturing the Off-sorted cells for WT LGIT, fewer than 3% of these cells activated out of the Off region (%Loss of Off), which demonstrates the stability of its Off mode. However, mutation of any of the three Sp1 sites resulted in dramatic increases (2- to 3-fold) in the rates of spontaneous activation compared to WT LGIT (Figure 3D), indicating that each of these three mutants has a destabilized Off mode. This result implies that in the Off state, Sp1 sites may be involved in a repressive mechanism, such as recruitment of HDAC complexes by individual Sp1 proteins [34],[39],[55]. This observation is surprising in light of earlier results suggesting that each Sp1 site is required for strong activation (Figure 3B and 3C). Each of the Sp1 sites may thus serve a repressing role in the Off mode and an activating role in the Bright mode, and the dynamic interplay between these roles may contribute to transcriptional noise and stochastic switching.
Analysis of the Off mode also revealed a reduction in spontaneous activation for mutI NF-κB and mutI&II NF-κB (by approximately 30% and 50%, respectively) relative to WT (Figure 3D), consistent with earlier observations that κB site I is important for the recruitment of an activating complex (Figures 2A and 3B). By contrast, κB site II did not affect spontaneous activation, as LGIT and mutII NF-κB are statistically indistinguishable (Figure 3D, p = 0.31), suggesting that in the Off state the second κB site does not recruit the same activating complex as the proximal site. Again, the roles of the two κB sites significantly differ (as in Figures 2 and 3B), with κB site I exhibiting a greater activating role than κB site II.
In addition to regulating the overall dynamics and steady states of viral gene expression, the individual Sp1 and κB elements may influence stochastic aspects of viral gene expression and thereby affect viral latency. We hypothesized that the dynamic balance in the recruitment of repressing and activating factors to individual promoter sites (Figure 1C) modulates the stabilities of the Off and Bright modes of gene expression, and mutation of these sites would therefore impact the frequency of phenotypic bifurcation (PheB), singly infected clonal cell populations that split into Off and Bright gene expression modes [9].
To analyze the role of individual transcription factor binding sites in the bifurcation phenotype, we isolated 362 individual clones from WT and mutant LGIT populations. Six days after infection, LGIT (and mutant) infected cells from Figure 2 were transiently stimulated with HMBA/Tat, and the infected (On) populations were isolated using fluorescence activated cell sorting (FACS) (Figure 3A, panel 3). These polyclonal populations were allowed to relax for one week, and single cells were then sorted from the Mid region, expanded, and analyzed by flow cytometry for heterogeneous expression and Phenotypic Bifurcation (PheB) in gene expression levels (Figure 3A, panel 7).
The frequency of bifurcation for WT LGIT was 1.77%±0.35% (Figure 4A), a level consistent with our prior findings [9]. In addition, all κB mutations yielded a PheB-clone frequency statistically indistinguishable from WT LGIT. Strikingly, however, all three Sp1 mutants exhibited a greater than 6-fold increase in the frequency of PheB. These results are consistent with the increased Mid:On ratio for Sp1 mutants (Figure 2B) and the increased dynamic switching between Off and Bright sorts (Figure 3C and 3D), and further indicate that the loss of any of the three Sp1 sites increases stochastic flipping between Off and Bright modes.
Mutation of any Sp1 site thus renders the viral promoter both weakly silenced (Off) and weakly transactivated (Bright), resulting in increased rates of spontaneous switching and phenotypic bifurcation. In agreement with this interpretation, there are compelling correlations between the frequency of PheB and the fraction of spontaneous inactivation (Figure 4B) and spontaneous activation (Figure 4C), indicating increased transcriptional noise and stochastic switching between two “less stable” states (Figure 1C). Together, these data reveal that each Sp1 site−and particularly Sp1 site III−plays an important role in the control of stochastic gene expression by regulating the active and inactive gene expression states via the recruitment of activating and repressing factors. This is the first demonstration that specific cis-regulatory elements within the HIV promoter contribute to transcriptional stochasticity and implicates the Sp1 sites as significant factors in the establishment and maintenance of proviral latency.
To further support our hypothesis that the Sp1 mutants have increased switching dynamics, we have examined the switching dynamics of Off and Bright sorts of PheB clones for the LGIT variants (Figure S2). We hypothesize that the clonal Off and Bright sorts may exhibit switching dynamics similar to the polyclonal populations (Figure 3C and 3D) and may partially converge back to the original bimodal distribution. Due to the rarity of clonal populations exhibiting PheB for all LGIT variants (∼2%–15% depending on mutant, Figure 4A)—and since gene expression profiles widely vary between different PheB clones—isolating and identifying different PheB clones that have identical gene expression profiles was not possible. However, we selected one PheB clone for each LGIT mutant that exhibited similar bimodality and isolated the Bright and Off modes using FACS (Figure S2). We have normalized the measured switching effects by the distribution from the unsorted clone, and the resulting “normalized switching” value provides a metric for the convergence to the original bimodal distribution. Values ranging from zero (no switching) to one (complete convergence) enable the evaluation of clonal switching dynamics for each Sp1 or κB mutant.
At four and seven days after FACS sorting, we have measured the GFP distributions for the unsorted, Off-sorted, and Bright-sorted fractions (Figure S2). The Bright fractions for each Sp1 mutant clone (S1.C1, S2. A3, and S3.B6) exhibit increased switching into the Off region (Figure 4D and 4E), which mimic increased spontaneous inactivation in polyclonal Bright sorts (Figure 3C). Similarly, Off-sorted fractions from the clones for mutII Sp1 (S2.A3) and mutIII Sp1 (S3.B6) have dramatically enhanced switching into the Bright region seven days after sorting (Figure 4E), which are consistent with spontaneous activation of the polyclonal Off sorts (Figure 3D). In contrast, the Off-sorted fraction from the clone for mutI NF-κB (N1.D5) exhibits decreased switching into the Bright region (Figure 4D and 4E), consistent with the observed polyclonal dynamics that this mutant has a stabilized Off mode (Figure 3D). Collectively, clonal and polyclonal switching dynamics reveal destabilization of the Off and Bright modes for the Sp1 mutants (Figure 3C and 3D).
To identify mechanistic differences in the roles of individual Sp1 and κB sites, we performed exogenous perturbations on each LGIT variant. Two weeks after infection with LGIT or mutants (the same unsorted populations analyzed in Figures 2 and 3A, panel 1), cells were stimulated with TNF-α (20 ng/ml) or TSA (400 nM) for 18 hours. The change in Bright Mean after perturbation revealed differential contributions for each site in the Bright mode (histograms in Figure S1B and non-normalized data in Table S4).
Although each of the three Sp1 mutants exhibited a lower Bright Mean than WT (Figure 2A), stimulation with TNF-α strongly increased the Bright Mean position of mutI Sp1, mutII Sp1, and mutIII Sp1 (Figure 5A, gray bars), confirming that these promoters are susceptible to activation via NF-κB dependent pathways. Furthermore, stimulation with the HDAC inhibitor TSA increased the Bright Mean almost 2-fold in mutI Sp1 and mutII Sp1 (Figure 5A, black bars). Since Sp1 has been shown to recruit class I HDACs to the HIV promoter [39], TSA inhibition of these HDACs may shift the chromatin modification balance towards acetylation by HATs. However, mutIII Sp1 was strikingly insensitive to TSA (Figure 5A), suggesting that this mutant may have minor regulation by HDACs or that it may not have sufficient HAT occupancy to take advantage of HDAC inhibition. Of these two possibilities, the former is consistent with a destabilized Off mode (Figures 3D and 4C), while the latter is consistent with a destabilized Bright mode (Figures 3C and 4B).
Both mutI NF-κB and mutII NF-κB were activated by TNF-α, though to a lesser extent than the WT promoter. However, mutI NF-κB was slightly but significantly less activated than mutII NF-κB (p<0.05), consistent with our prior findings that mutI NF-κB has lower levels of gene expression than mutII NF-κB (Figure 2A). TNF-α induced no relative change in the Bright Mean position for mutI&II NF-κB, confirming that this double mutation eliminated NF-κB-mediated activation of the HIV promoter (Figure 5A, gray bars). Stimulation with TSA strongly increased the Bright Mean position for all κB mutants, including mutI&II NF-κB, as its effects are not dependent upon NF-κB activation (Figure 5A, black bars). However, in contrast to TNF-α, TSA activated mutI NF-κB more strongly than mutII NF-κB or WT (p<0.05), suggesting that mutI NF-κB may be more heavily repressed by class I HDACs.
Using the Off-sorted polyclonal populations (Figure 3A, panel 6) as a model for HIV latency, we examined the stability of the Off mode by measuring the susceptibility of Off-sorted cells to activation by TNF-α and TSA. TNF-α activated approximately 33% of the Off-sorted LGIT cells, demonstrating that a large fraction of these “latent” cells is capable of reactivation via a NF-κB-dependent mechanism (Figure 5B, gray bars). Each of the Off-sorted Sp1 mutants responded more strongly to TNF-α induction than WT LGIT (Figure 5B). These results are consistent with the previously observed increase in Bright Mean position for these mutants (Figure 5A), suggesting that these mutants are deficient in recruiting RelA under unstimulated conditions.
TSA activates approximately 35% of the Off-sorted, “latent” cells of the mutI Sp1 and mutII Sp1 populations (Figure 5B, black bars) but only 13% of mutIII Sp1 cells, analogous to results in unsorted cells (Figure 5A, black bars). These results suggest that all these mutants are repressed by HDACs in the Off state, but that Sp1 site III is specifically required for an effective response to TSA, possibly because it plays a key role in recruitment of HAT complexes.
In contrast to WT LGIT and the corresponding Sp1 mutants, in which at least one-third of the Off cells were activated by TNF-α, both mutI NF-κB and mutI&II NF-κB were virtually insensitive to TNF-α stimulation, indicating that κB site I is essential for NF-κB-dependent activation of Off cells (Figure 5B, gray bars). However, 13% of mutII NF-κB cells responded to TNF-α stimulation (Figure 5B), further demonstrating that when intact, this κB site plays a significant, but weaker, role in NF-κB activation than site I. Finally, all three κB mutants exhibited reduced responses to TSA stimulation compared to WT LGIT (Figure 5B, black bars), suggesting that both κB sites have significant but unequal roles in the recruitment of p50-p50 homodimer and HDAC complexes in the latent state.
The gene expression and perturbation results thus far suggest that individual Sp1 binding sites differentially recruit activating and repressing transcription factors, thereby differentially stabilizing the Off and Bright expression modes and contributing to gene expression noise (Figure 4B and 4C). We used chromatin immunoprecipitation (ChIP) to measure p50, RelA, p300, Sp1, and HDAC1 protein occupancy at the LTR in populations sorted from Off (Figure 3A, panel 6) and Bright (Figure 3A, panel 5) regions. Additionally, we have analyzed Off and Bright sorts for acetylation of lysines 9 and 14 of the tail of histone 3 (AcH3, markers for active chromatin [54]) and trimethylation of lysine 9 (TriMetH3K9, a signature of repressed chromatin [56]). Performing ChIP on Off- and Bright-sorted populations is distinct from recent ChIP analyses on chromatin targets of transfected and/or integrated LTR, which used pharmacological factors including TNF-α, TSA, and phorbol esters to observe occupancy and histone acetylation patterns in the stimulated or unstimulated wild type LTR [3],[32],[46],[47],[57],[58]. Our work focuses instead on analyzing differences in the occupancies of chromatin regulators and transcription factors within Off and Bright modes of integrated viral mutants in unstimulated conditions. Such quantitative differences in LTR occupancy between two coexisting cell populations may influence and reflect the fate of the provirus towards transcriptional activation or repression and latency.
ChIP analysis revealed that RelA recruitment to the HIV promoter in Off-sorted cells was reduced approximately 10-fold for mutIII Sp1 as compared to WT (Figure 6B). In contrast, mutI Sp1 and mutII Sp1 promoters recruit RelA to similar extents as WT in both the Off and Bright populations (Figure 6B, p>0.20 compared to respective WT sorts). This finding is consistent with gene expression results suggesting that Sp1 site III is important for recruiting activating complexes, as its mutation led to a higher fraction of Infected but Off cells (Figure 3B), as well as insensitivity to TSA (Figure 5). Therefore, we conclude that Sp1 site III enhances the ability of the κB sites to recruit RelA, and destabilization of the Bright mode in mutIII Sp1 may in part be due to insufficient recruitment of RelA.
ChIP on mutI NF-κB revealed that recruitment of RelA decreased approximately 3-fold for mutI NF-κB as compared to WT but was unchanged for mutII NF-κB (Figure 6B), confirming distinct roles for the two sites. Also, WT LGIT, mutI NF-κB, and mutII NF-κB recruit RelA to similar extents in the Bright sort (Figure 6B), indicating that κB site I in particular is necessary for the recruitment of p50-RelA heterodimer in the Off mode, but that both κB sites can sufficiently recruit the heterodimer in the Bright mode. ChIP for p50 indicated that WT LGIT, mutI NF-κB, and mutII NF-κB variants recruit p50 to similar extents (Figure S3A), consistent with the fact that p50 is present as part of both the p50-p50 homodimer and the p50-RelA heterodimer (Figure S3B). Thus, ChIP data strongly support the prior hypothesis that κB site I recruits RelA to a greater extent than site II in the Off mode (Figure S3C and S3D). Collectively, our results demonstrate that the two κB sites have distinct roles in transcriptional regulation, and implicate unequal roles in the establishment and maintenance of latency.
Histone acetyltransferase p300 is a central factor in HIV transactivation [30] that is actively recruited to the HIV promoter [24] by Sp1 [59] and NF-κB [28],[60] complexes. Analysis of p300 by ChIP revealed similar levels of recruitment for all Bright populations (Figure 6C). However, we observed a ten-fold reduction in p300 recruitment for mutIII Sp1 relative to mutI Sp1, mutII Sp1, and WT in the Off-sorted populations, indicating that Sp1 site III is particularly important for recruiting p300 to the HIV promoter in the Off or latent state. Since mutIII Sp1 suffers a loss of p300 recruitment in the Off mode, the striking insensitivity to TSA stimulation for this mutant (Figure 5) is likely due to the inability to recruit this HAT after inhibition of HDAC activity.
We next analyzed the overall recruitment of Sp1 protein to each LTR in Off and Bright populations. In the Off fraction, the WT promoter and each of the individual Sp1 mutant promoters recruit Sp1 to similar extents (Figure 6D). However, in the Bright fractions, mutation of any of the Sp1 sites results in greater than 10-fold reduction in Sp1 recruitment. This loss of Sp1 in the transactivated (Bright) mode does not correlate with loss of p300 (Figure 6C), suggesting that other factors, including RelA [28],[60] and Tat protein [30],[61], may be involved in the localization of p300.
Transcriptional repression is commonly regulated by histone deacetylation, and HDAC1 is associated with p50-p50 homodimer [3] and Sp1 [39] at the HIV-1 LTR. Therefore, we performed ChIP against HDAC1 to determine its recruitment to each Sp1 and κB site and its role in transcriptional repression (Off sorts) vs. activation (Bright sorts). ChIP on the Off sorts revealed statistical decreases in HDAC1 occupancy for all mutants, except mutI NF-κB, when compared to WT LGIT (Figure 6E). In contrast, HDAC1 occupancy in the Bright sorts was statistically indistinguishable from WT for all mutants (p>0.1, Figure 6E). Additionally, WT and all mutants except mutIII Sp1 had elevated levels of HDAC1 occupancy in the Off sorts compared to the respective Bright sorts (p<0.1, Figure 6E). Collectively, these ChIP findings reveal that each Sp1 site and κB site II are important in the recruitment of HDAC1, and that mutation of Sp1 site III abolishes differential regulation of HDAC1 between Off and Bright modes.
To estimate differences in overall levels of transcriptional activation between Off and Bright sorts and between WT and mutant populations, we measured acetylation of the histone 3 tail at lysines 9 and 14 (AcH3) by ChIP and normalized to total histone 3 (Figure 6F). In the Off sort, each mutant was statistically indistinguishable to WT LGIT (p>0.05, Figure 6F). In contrast, the Bright sorts revealed significant decreases of AcH3 for all mutants compared to WT. These results show that each Sp1 and κB site is essential for maximum acetylation of H3. However, there are no significant differences in AcH3 between Off and Bright sorts of mutI Sp1, mutII Sp1, and mutIII Sp1 cells, indicating that each of the three Sp1 sites is required for the regulation of the deacetylated (Off) and acetylated (Bright) states.
Lastly, to examine an indicator of repressed chromatin beyond histone deacetylation, we performed ChIP for Off and Bright sorts to examine trimethylation of histone 3 at lysine 9 (TriMetH3K9). However, we observed undetectable levels of TriMetH3 for the Off and Bright sorts of WT LGIT and all Sp1 and κB mutants (Figure S4E), suggesting that trimethylation of H3K9 is not a significant factor in the phenotypes we observe in LGIT, including dynamic switching and stabilization of the Off mode.
Sp1 site III regulates recruitment of repressor HDAC1 (Figure 6E) and activators p300 (Figure 6C) and p50-RelA (Figure 6B). In contrast, RelA occupancy is not hindered by mutation of κB site II; however, mutation of κB site I decreases the occupancy of p50-RelA heterodimer but apparently not p50-p50 homodimer (Figure 6B and Figure S3C and S3D). The differences in p50-p50, p50-RelA, p300, and HDAC1 occupancies at these mutant promoters likely influence the frequencies of both transcription initiation and elongation at the LTR. In particular, an inactive LTR occupied with a repressive p50-p50 homodimer and HDAC1 would be unable to recruit RNAPII and would thus not initiate transcription [3]. In contrast, when p50-RelA heterodimer and p300 localize to the LTR, RNAPII is readily recruited and phosphorylated by P-TEFb, to generate fully elongated, productive transcripts [3],[19]. Competition between the recruitment of repressing p50-RelA heterodimer and p50-p50 homodimer may result in RNAPII initiating transcription and then stalling, which results in a large number of abortive transcripts and little GFP and Tat expression [3],[19]. Thus, we adapted a previously established RT-PCR method [3] to quantify the functional differences transcriptional initiation and elongation for the Off and Bright sorts of WT LGIT, mutIII Sp1, mutI NF-κB and mutII NF-κB.
The Off sorts for mutI NF-κB and mutIII Sp1 have respective decreases of 36% and 48% in the number of initiated transcripts compared to WT (Figure 6G), in agreement with ChIP observations that these two mutants have decreased p50-RelA occupancy and potentially increased recruitment of the repressing p50-p50 homodimer (Figures 6B and S3C). In contrast, there is no statistical change in transcriptional initiation for the mutII NF-κB Off sort (p = 0.88 vs. WT, Figure 6G), also in agreement with ChIP observations of no change in p50-RelA occupancy vs. WT (Figure 6B). Quantification of transcriptional initiation on Bright-sorted populations revealed a similar trend, in which mutII NF-κB exhibited a 35% increase in the number of initiated transcripts compared to WT, while mutI NF-κB and mutIII Sp1 have significant decreases (18% and 63%, respectively, Figure 6F).
Although transcription initiation was not hindered for mutII NF-κB (and was actually enhanced in the Bright mode), this mutant has a 48% decrease in the number of elongated transcripts vs. WT in the Off sort and a 15% decrease in the Bright sort (Figure 6E and 6F). Mutation of κB site I further decreased transcriptional elongation, with 57% and 38% decreases for Off and Bright sorts vs. WT, respectively (Figure 6E and 6F). These findings demonstrate that while both κB sites are required for full transcriptional elongation, κB site I has a greater contribution in both transcriptional initiation and elongation. Importantly, mutation of Sp1 site III has striking reduction of transcriptional elongation in the Off and Bright modes (53% and 70% decreases, respectively, Figure 6E and 6F), consistent with ChIP observations that Sp1 site III is necessary for recruitment of RelA in both Off and Bright modes and of p300 in the Off mode (Figure 6B). Moreover, these data indicate that κB site I and Sp1 site III combine to activate transcriptional initiation (likely in part via recruitment of p50-RelA heterodimer), and mutation of either site abrogates this role.
Collectively, the molecular and transcriptional phenotypes reveal that mutation of any Sp1 site destabilizes the Off and Bright gene expression modes and increases dynamic switching and phenotypic bifurcation. Alternatively, mutation of κB site I appears to slightly increase the stability of the Off mode, while mutation of κB site II may slightly stabilize the Bright mode. The molecular regulation of Off and Bright modes of each LGIT variant is summarized in Figure 7, which integrates the results from ChIP, perturbation, and transcriptional experiments into a model of transcription factor occupancies.
In Figure 7, the degree of shading (or transparency) of the individual molecules corresponds to hypothetical degrees of occupancy. For this study, these configurations are inferred by functional tests (transcriptional activities and responses to perturbation) and directly measured by ChIP, which are all summarized in Figure S6. We hypothesize that repressing markers, including p50-p50 homodimer, HDAC1, and deacetylation of lysines 9 and 14 of histone 3 (H3K9/14), indicate a stabilized Off mode and inactive transcription. In contrast, we hypothesize that a stabilized Bright mode and transcriptional activation are characterized by the presence of p50-RelA, p300, and histone acetylation of H3K9/14. However, the association of Sp1 with HDAC1 or p300 may govern its structural conformation and DNA-binding affinity [35],[55],[62], and we have illustrated these two conformations by the orientation of the Sp1 molecule (trapezoid). The strongest destabilizing phenotypes occur with mutIII Sp1, which displays the highest population of Infected but Off cells, highest Mid:On ratio, and lowest Bright Mean position; exhibits the strongest response to TNF-α but the weakest to TSA; yields the highest frequency of phenotypic bifurcation and greatest dynamic switching; and recruits decreased levels of RelA (Off mode), p300 (Off mode), HDAC1 (Off mode) and Sp1 (Bright mode) (Figures 7 and S6).
We have previously demonstrated that stochastic fluctuations in Tat levels may contribute to HIV-1 proviral latency by delaying the onset of viral gene expression and Tat feedback [9]. In this study, we explored how each Sp1 and κB element in the HIV LTR modulates these stochastic fluctuations via differential recruitment of activating and repressing factors. We found that each Sp1 and κB site contributes distinctly to the dynamics of switching between low and high gene expression modes and the frequency of phenotypic bifurcation—both of which have implications for viral latency.
This experimental system, based on single integrations of the LGIT provirus in CD4+ Jurkat cells [9], is similar in design to the J-Lat model—a clonal Jurkat cell line with a single integration of full-length HIV in which the viral gene Nef has been replaced by GFP—which has been used as an in vitro model for HIV latency [3],[32],[46],[63]. In this study, individual Sp1 and κB elements within the LTR of LGIT were disabled to dissect their contributions to transcriptional activation and repression. Furthermore, this system examines the dynamic behavior of single integrated proviruses rather than transient transfection reporter systems [40]–[42]—which provide valuable gene expression information but do not account for the integration site, chromatin environment, and low copy number of the provirus—or viral replication assays [43]–[45]—which measure replicative fitness but not gene expression or proviral latency. Additionally, by isolating and characterizing transactivated (Bright) and latent (Off) populations, we examine the contribution of each Sp1 and κB site to transactivated and “latent” states, as well as their responses to stimulation with pharmacological agents TNF-α and TSA.
Gene expression results (Figures 2–4) provide corroborating evidence that mutation of any one of the three Sp1 binding sites dramatically increases promoter sensitivity to transcriptional noise and stochastic effects in the Tat feedback loop. Moreover, we conclude that each Sp1 site—particularly Sp1 site III—plays an important role in the control of stochastic gene expression by stabilizing the inactive and active transcription states via the recruitment of activating and repressing factors to the LTR (Figure 6). In support of a destabilized Bright state, single Sp1 mutations and in particular site III mutation result in a significantly weaker promoter (Figure 2A), a 2- to 3-fold increase in the Off (latent) population of infected cells (Figure 3B), rapid switching from Bright to Off (Figure 3C), and a considerable loss of overall Sp1 binding to the promoter in the Bright state as measured by ChIP (Figure 6D). As support for a destabilized Off state, Sp1 mutations result in rapid switching from Off to Bright (Figure 3D), a considerable increase in the fraction of integrated provirus that responds to TNF-α (Figure 5), and a decrease in the LTR occupancy of HDAC1 as measured by ChIP (Figure 6E). In addition, as evidence of both destabilized Off and Bright modes for the Sp1 mutants, there were insignificant differences between the deacetylated vs. acetylated states of the respective Off and Bright sorts (Figure 6E). Collectively, these destabilizing effects result in a promoter more susceptible to transcriptional noise and phenotypic bifurcation (Figure 4A), with potentially important implications for viral latency.
We also demonstrate that the two κB sites, despite having identical sequences, provide distinct roles in transcriptional activation. κB site I plays a preferential role in promoter activation, likely because cooperativity between κB site I and the adjacent Sp1 site III promotes recruitment of p50-RelA and p300 (Figure 6B and 6C). In contrast, the distal κB site II provides a bias for the recruitment of p50-p50 homodimer (Figure S3C) and HDAC1 (Figure 6E). These observations bear similarities to TLR-induced genes in dendritic cells, in which subunit specificities of κB sites are governed by cooperative interactions with other factors, including CBP [64].
Binding cooperativity between Sp1 and RelA has been reported in biochemical studies [65],[66], and the distinct roles we have observed for the different Sp1 and κB binding sites raises the possibility that the individual enhancer sites may differentially contribute to this cooperativity. ChIP results show that p300 recruitment to the mutI NF-κB promoter is maintained, whereas RelA is lost; however, both p300 and RelA are lost in mutIII Sp1 (Figure 6B and 6C). This result implies that Sp1 site III may first recruit p300, whose presence enhances p50-RelA localization to κB site I. The proximity of κB site I to Sp1 site III may even underlie its biased recruitment of the p50-RelA heterodimer, which may be related to reported direct binding between RelA and Sp1 [67] or between RelA and p300 [28],[29].
Although Sp1 site III is required for p300 recruitment in the Off mode, other factors at the LTR may contribute to subsequent p300 maintenance at the promoter [35]. These may include p300 interacting with RelA [68], LEF-1 [38], YY1 [69], and SWI/SNF [25],[60], as well as p300 directly binding to DNA at motifs (i.e., GGGANT) found within the LTR, including in both κB elements [70]. RelA and p300 are bound to the LTR in all Bright sorts, and also in most Off sorts, supporting earlier models that RelA and p300 are necessary, but not sufficient, for Tat-transactivation, and recruitment of other factors (RNAP II, P-TEFb, PCAF, etc.) is required [31],[71]. Other potential factors contributing to RelA localization at the promoter in the Off mode include its binding to an inactive promoter [72], its potential interaction with various HDACs [71],[73],[74], or the competing repressive roles of other regulatory factors (YY1, LSF/LBP-1, etc.) [75]–[77].
Sp1 recruitment to WT LGIT was unchanged in Off vs. Bright cells, as assessed by ChIP (Figure 6D). This finding corroborates a recent study in transiently transfected cells that detected no change in Sp1 levels at the LTR with or without addition of exogenous Tat [58]. In contrast to the WT LTR, Sp1 recruitment to the Sp1 mutant promoters was compromised in the Bright sorts relative to the Off sorts, potentially due to different Sp1 binding affinities in the two expression modes. In the Bright sorts, the association of individual Sp1 molecules with p300 may weaken its binding affinity for DNA [35],[62]. In contrast, Sp1 interactions with HDAC1 appear to have no reduction in DNA binding [55]. The DNA binding affinities of Sp1 are increased by homomultimerization and synergy between Sp1 molecules [67],[78], and Sp1 EMSA analyses with the HIV-1 LTR have revealed that mutation of one of the three Sp1 sites reduced Sp1 recruitment, while mutation of two sites eliminated detection of Sp1 [79]. Thus, we hypothesize that the diminished Sp1 levels in Bright sorts of Sp1 mutants may result from decreased affinities of individual Sp1 molecules and failure to recruit multimerized Sp1 complexes.
Collectively, these findings demonstrate that the balance between the repressing and activating roles of each Sp1 site impacts transcriptional noise and the propensity for latent infections. Therefore, it appears that by recruiting activating and repressing host factors, intact Sp1 sites dampen noise in HIV gene expression. Although it remains to be determined whether such noise arises from events external to (extrinsic) or directly from (intrinsic) the mechanisms under study here [80], the Sp1 sites' regulation of promoter activity and chromatin dynamics agrees with a paradigm that eukaryotic promoters generate noise by localization of chromatin factors [12],[13]. Such local chromatin dynamics may yield transcriptional pulses [16],[17] that would be amplified by Tat feedback.
Although the Sp1 and κB sites enhance replicative fitness of the virus [43]–[45],[50], variations and mutations within these sites are often observed in isolates from subtype B cohorts [79],[81],[82]. Moreover, there is considerable sequence variability of Sp1 and κB elements across different HIV-1 subtypes [82],[83]. In addition to altering the replication fitness of the virus, our findings suggest that such evolutionary divergence within subtype B variants and across other subtypes likely impact viral transcriptional dynamics and propensities for latency. Thus, we postulate that variations in promoter architectures will have important unexplored epidemiological and therapeutic implications.
This work demonstrates the power of quantitative, dynamic phenotyping of viral mutants for dissecting regulatory inputs into the viral promoter in a proviral model of HIV. This approach revealed that each Sp1 site influences the control of stochastic gene expression by stabilizing both the active state—therefore likely playing a role in the regulation of bursts in viral gene expression—and the inactive state—thus playing a role in the establishment and maintenance of proviral latency. It remains to be determined which of these features are central to survival and propagation of the virus in a natural environment or under therapeutic challenge. Finally, this work may aid the future development of paradigms to predict the gene expression and latency phenotypes of HIV-1 isolates and subtypes, as well as draw important correlations between viral genotype and clinical outcomes and responses to antiviral therapies.
Construction of LGIT plasmids has been previously described [9]. Double and triple point mutations at the Sp1 and κB sites in the HIV LTR were performed using the Quikchange PCR method (Stratagene). The specific inactivating mutations used for κB [50] and Sp1 [43],[48],[49] were previously described, and primer sequences are listed in Table S1.
Jurkat cells were cultured in RPMI 1640 (Mediatech) medium supplemented with 10% fetal bovine serum, 100 U/ml penicillin-streptomycin, and 2 mM L-glutamine. Cells were grown at concentrations between 2×105 and 106 cells/ml in 5% CO2 at 37oC. HEK 293T cells, used for lentiviral packaging, were cultured in the same conditions as Jurkats but with Isocove's DMEM (Mediatech). For perturbation and viral titering experiments, the following factors were used in the specified concentrations: 20 ng/ml tumor necrosis factor-α (TNF-α, Sigma-Aldrich), 400 nM trichostatin A (TSA, Sigma-Aldrich), and 5 mM hexamethylene bisacetamide (HMBA, Sigma-Aldrich).
Lentiviral vectors were packaged and harvested in HEK 293T cells using 10 µg of pCLGIT (or mutant κB/Sp1 variants), 5 µg pMDLg/pRRE, 3.5 µg pVSV-G, and 1.5 µg pRSV-Rev, as previously described [9],[84], then concentrated by ultracentrifugation to yield between 107 and 108 infectious units/ml. For titering, 3×105 Jurkat cells in 12-well plates were infected with approximately 103–106 infectious units per well. Six days later, infected Jurkats were incubated with a combination of 5 mM HMBA, 20 ng/ml TNF-α, and 400 nM TSA for 18 hours and then analyzed by flow cytometry to determine infectious titer by GFP expression. This combination of agents was chosen to stimulate the promoter via P-TEFb [85], NF-κB [36], and Sp1 dependent mechanisms [37]. Titering curves were constructed to achieve infection of 5–10% of cells after maximum stimulation, corresponding to MOI ∼0.05–0.10.
Infected cultures were analyzed via flow cytometry on a Beckman-Coulter EPICS XL-MCL cytometer (http://biology.berkeley.edu/crl/cell_sorters_analysers.html). All flow measurements were performed in parallel with an uninfected Jurkat control, and perturbation experiments with TNF-α and TSA were performed in parallel with stimulated but uninfected Jurkat controls. To isolate infected, expressing populations, GFP+ cells were sorted on a DAKO-Cytomation MoFlo Sorter. As described in the text, bulk population (polyclonal) and single cell (clonal) sorts were performed for a range of different GFP positive regions, as follows: LGIT bulk and clonal sorts: “Off” region (∼0.1–2.0 Relative Fluorescence Units), “Mid” region (∼2.0–30 RFU), “Bright” region (∼30–1024 RFU), and “On” region (∼2.0–1024 RFU). Flow cytometry data analysis was performed with FlowJo (Tree Star, Inc.).
Gene expression levels were tracked over a 21-day time course by measuring the fluorescence intensity of GFP in LGIT and mutant cells. Cells were infected at an MOI ∼0.05–0.10, and a GFP+ population was detectable by flow cytometry 48 hours after infection. The time to peak activity occurred approximately one week after infection; however, a bimodal distribution of infected cells (“Off” and “Bright”) persisted throughout the three week experiment. The strength of Tat-transactivation was measured by examining the Bright population, and in particular, the mean of this population (Bright Mean) was used as a marker of the base efficiency of transactivated gene expression. The Bright Mean of infected LGIT and mutants was determined by calculating the average relative fluorescence of cells within the “Bright” region (∼30–1024 RFU).
For LGIT and all mutants, a small fraction of cells occurs in a critical region defined as the Mid region (∼2.0–30 RFU), which lies between Off (∼0.1–2.0 RFU) and Bright (∼30–1024 RFU) populations. Cells isolated from this region tend to turn Off or Bright in a random fashion, demonstrating an instability of that region [9]. The fraction of infected cells persisting in the Mid region at a specific time is represented by the Mid:On ratio, in which the On region (∼2.0–1024 RFU) is the sum of Mid and Bright subpopulations. We employ this ratio as a metric for transcriptional instability, such that a high Mid:On ratio suggests an unstable promoter and a high degree of stochastic switching.
FACS sorting was performed to isolate Off and Bright fractions of the LGIT and LGIT mutant cell lines, and 1×106 cells were acquired for each sort. Off, Bright, and original unsorted populations were expanded to achieve 5×107 cells, incubated in 1% formaldehyde for 10 minutes at room temperature for fixation, and subsequently incubated with 125 mM glycine for 5 minutes at room temperature to quench the formaldehyde. Upstate EZ ChIP (17–371) reagents and protocol were utilized for crosslinking, lysis, sonication, immunoprecipitation, elution, reverse crosslinking, and DNA purification procedures. Sonication was performed with the Branson Sonifier 450 for 15 cycles with power output of 2.5, 10% duty cycle, for 10–15 second pulses and 1 minute intervals on ice. DNA was sheared to achieve an average of 0.2–0.7 kb, as confirmed by DNA gel electrophoresis. Immunoprecipitations were performed with Upstate polyclonal antibodies anti-p50 (06–886), anti-p65 (06–418), anti-p300 (05–257), anti-Sp1 (07–645), and anti-AcH3H9/14 (06–599) and Abcam polyclonal antibodies anti-HDAC1 (ab7028), anti-H3 (ab1791), and anti-TriMetH3K9 (ab8898).
Immunoprecipitated DNA was quantified using quantitative polymerase chain reaction (QPCR) with primers within the HIV LTR which flank the κB and Sp1 elements [46]. To accurately assess the input of each QPCR reaction, and to normalize for the efficiency of immunoprecipitation of each antibody, we used endogenous promoters containing functional κB and/or Sp1 domains as normalization controls for RelA, p300, and Sp1. The endogenous TAP1/LMP2 regulatory domain (PubMed accession# NM_000593.5), which contains a single κB site four nucleotides downstream of an Sp1 site, was used to normalize QPCR data from RelA and p50 immunoprecipitations, as this promoter has been shown to constitutively recruit both p50-RelA and Sp1 [86]. Similarly, the endogenous BCL2L1 regulatory domain (PubMed accession# NW_001838664.2), which contains Sp1 elements and has been shown to strongly recruit p300 and Sp1 [87], was used for normalizing QPCR data from p300 and Sp1 immunoprecipitations. Non-normalized ChIP results are presented in Figure S4, and all ChIP primer sequences are provided in Table S2. HDAC1, AcH3, TriMeH3, and H3 immunoprecipitations were normalized by inputs, and acetylated histone 3 is reported as a ratio of AcH3:H3 immunoprecipitations.
Amplified DNA products from each primer set were cloned into the Invitrogen pCR2.1 plasmid (pCR2.1-TOPO-LTRκB, pCR2.1-TOPO-TAP1/LMP2, and pCR2.1-TOPO-BCL2L1) to create plasmids that were subsequently used to generate standard curves for all QPCR analyses. Linear regression of standard curves was achieved by serial dilutions ranging from ∼10 ng to ∼10−6 ng plasmid DNA, which corresponds to ∼2×109 to ∼2×102 copies per 20 µL reaction. Quantitative PCR was performed using the iCycler iQ Real-Time PCR Detection System (Bio-Rad, Hercules, CA), and SYBR Green I (Invitrogen) was used as the fluorescent nucleic acid stain.
As performed in ChIP experiments, FACS sorting was used to isolate Off and Bright fractions of the LGIT and LGIT mutant cell lines, and 1×106 cells were acquired for each sort. Off and Bright populations were expanded to achieve 1×107 cells, total mRNA was isolated using Trizol (Invitrogen), and transcripts were quantified using the QuantiTect SYBR Green RT-PCR kit (Qiagen) on the Bio-Rad iCycler. The total number of transcripts (initiated and elongated) were detected with TAR primers [3], and Tat primers were used to detect only elongated transcripts (sequences in Table S3). For each sample, initiated and elongated transcript levels were normalized by the corresponding levels of β-Actin mRNA (sequences in Table S3) [9]. Measurements and calculations of initiated, elongated, and truncated transcripts are provided in Figure S5. Triplicate RT-PCR measurements were performed for all samples for each primer set, and melt curves were performed on the Bio-Rad iCycler for all samples to confirm the specificity of QPCR reaction.
|
10.1371/journal.pntd.0000671 | Transmission of West Nile Virus by Culex quinquefasciatus Say Infected with Culex Flavivirus Izabal | The natural history and potential impact of mosquito-specific flaviviruses on the transmission efficiency of West Nile virus (WNV) is unknown. The objective of this study was to determine whether or not prior infection with Culex flavivirus (CxFV) Izabal altered the vector competence of Cx. quinquefasciatus Say for transmission of a co-circulating strain of West Nile virus (WNV) from Guatemala.
CxFV-negative Culex quinquefasciatus and those infected with CxFV Izabal by intrathoracic inoculation were administered WNV-infectious blood meals. Infection, dissemination, and transmission of WNV were measured by plaque titration on Vero cells of individual mosquito bodies, legs, or saliva, respectively, two weeks following WNV exposure. Additional groups of Cx. quinquefasciatus were intrathoracically inoculated with WNV alone or WNV+CxFV Izabal simultaneously, and saliva collected nine days post inoculation. Growth of WNV in Aedes albopictus C6/36 cells or Cx. quinquefasciatus was not inhibited by prior infection with CxFV Izabal. There was no significant difference in the vector competence of Cx. quinquefasciatus for WNV between mosquitoes uninfected or infected with CxFV Izabal across multiple WNV blood meal titers and two colonies of Cx. quinquefasciatus (p>0.05). However, significantly more Cx. quinquefasciatus from Honduras that were co-inoculated simultaneously with both viruses transmitted WNV than those inoculated with WNV alone (p = 0.0014). Co-inoculated mosquitoes that transmitted WNV also contained CxFV in their saliva, whereas mosquitoes inoculated with CxFV alone did not contain virus in their saliva.
In the sequential infection experiments, prior infection with CxFV Izabal had no significant impact on WNV replication, infection, dissemination, or transmission by Cx. quinquefasciatus, however WNV transmission was enhanced in the Honduras colony when mosquitoes were inoculated simultaneously with both viruses.
| Unlike most known flaviviruses (Family, Flaviviridae: Genus, Flavivirus), insect-only flaviviruses are a unique group of flaviviruses that only infect invertebrates. The study of insect-only flaviviruses has increased in recent years due to the discovery and characterization of numerous novel flaviviruses from a diversity of mosquito species around the world. The widespread discovery of these viruses has prompted questions regarding flavivirus evolution and the potential impact of these viruses on the transmission of flaviviruses of public health importance such as WNV. Therefore, we tested the effect of Culex flavivirus Izabal (CxFV Izabal), an insect-only flavivirus isolated from Culex quinquefasciatus mosquitoes in Guatemala, on the growth and transmission of a strain of WNV isolated concurrently from the same mosquito species and location. Prior infection of C6/36 (Aedes albopictus mosquito) cells or Cx. quinquefasciatus with CxFV Izabal did not alter the replication kinetics of WNV, nor did it significantly affect WNV infection, dissemination, or transmission rates in two different colonies of mosquitoes that were fed blood meals containing varying concentrations of WNV. These data demonstrate that CxFV probably does not have a significant effect on WNV transmission efficiency in nature.
| The majority of the >70 recognized flaviviruses (family Flaviviridae, genus Flavivirus) are arthropod-borne, and include some of the world's most historically- and medically-important viruses including Yellow fever (YFV) and the Dengue (DENV) viruses. Gaunt et al. [1] described four distinct evolutionary clades within the genus Flavivirus that correlated with geography, vector, and associated disease: tick-borne, Culex-borne, Aedes-borne, and no known vector. Basal to all of these groups was Cell fusing agent virus (CFAV), an insect-only flavivirus discovered in an Aedes aegypti cell line more than 30 years ago [2].
Recently, a number of novel flaviviruses which cluster phylogenetically with CFAV have been isolated and identified from a diversity of field-collected mosquitoes and ticks around the world, including known arbovirus vectors. These arthropod-specific viruses collectively represent a unique clade of flaviviruses and include Ngoye virus from Rhipicephalus ticks in Senegal [3], Kamiti River virus (KRV), isolated from Aedes mcintoshi Huang in Kenya [4], [5], CFAV isolated from Ae. aegypti in Thailand and Puerto Rico [6], [7], Quang Binh virus from Culex tritaeniorhynchus Giles in Vietnam [8], and Nounané virus from Uranotaenia mashonaensis Theobald in Côte d'Ivoire [9]. Additionally, many strains of Culex flavivirus (CxFV) have been isolated from Culex pipiens L. in Japan [10], and North America [11], Culex tarsalis Coquillett throughout the western United States and Canada [11]–[12] (Bolling et al., unpublished data), Cx. restuans Theobald from Texas [13], and Cx. quinquefasciatus Say from Guatemala [14], the Yucatan Peninsula [15], Texas and Trinidad [13]. While there has been extensive genetic characterization of these viruses, the natural history and potential impact of mosquito-specific flaviviruses on the transmission efficiency of arboviruses of public health importance such as West Nile virus (WNV) remains unclear.
Arbovirus superinfection in mosquitoes and mosquito cell culture has been previously studied [16]–[25]. Infection with one flavivirus has been shown to suppress infection and prevent transmission of a second, antigenically-similar flavivirus. This phenomenon was demonstrated for Japanese encephalitis virus and Murray Valley encephalitis (MVE) virus superinfections in Culex tritaeniorhynchus Giles [16], two different strains of WNV in Culex pipiens form molestus Forskal [17], and WNV and St. Louis encephalitis virus in Cx. quinquefasciatus [18]. Sabin [19] demonstrated that high doses of YFV administered to Ae. aegypti previously infected with DENV still resulted in transmission of YFV, although mosquitoes were less susceptible to secondary infection with YFV. Similar findings have been reported in cell culture, where homologous superinfections were inhibited but secondary infection with a heterologous virus was permitted [22]–[25]. Therefore, based on previous observations, a primary infection of mosquitoes with a mosquito-specific flavivirus has the potential to interfere with infection or transmission of WNV acquired secondarily.
West Nile virus activity has been documented in Guatemala since 2003, beginning with the detection of WNV seroconversions in horses [26]. Serological evidence of WNV transmission has since been found in wild birds and chickens (Morales-Betoulle et al., unpublished data) and WNV has been isolated from several species of Culex (Culex) mosquitoes including Cx. quinquefasciatus (Morales-Betoulle et al., unpublished data). Culex quinquefasciatus is abundant in the urban WNV transmission focus comprising the city of Puerto Barrios, Guatemala, however, there has been little evidence of WNV-associated human disease in Guatemala or elsewhere in Latin America [27]. CxFV Izabal strain has also been found co-circulating with WNV in Cx. quinquefasciatus in Guatemala [14]. Minimum infection rates of CxFV in Cx. quinquefasciatus in Latin America were 20.8 per 1000 in Mexico [15] and 4.7 per 1000 in Guatemala [14]. Prevalence of CxFV Izabal in Cx. quinquefasciatus and the potential for this mosquito-specific flavivirus to disrupt WNV transmission is one of several hypotheses that have been proposed to explain the lack of human disease attributable to WNV in Latin America [15].
The objective of this study was to determine if prior infection with CxFV Izabal altered the vector competence of Cx. quinquefasciatus for transmission of WNV. The specific aims of this work were to: 1) compare replication kinetics of a Guatemalan isolate of WNV in CxFV Izabal – infected (CxFV Izabal (+)) and CxFV Izabal-uninfected (CxFV Izabal (−)) C6/36 cells and female Culex quinquefasciatus, 2) compare infection, dissemination and transmission rates of WNV in Cx. quinquefasciatus either infected or uninfected with CxFV Izabal, 3) determine whether WNV transmission by CxFV Izabal (+) Cx. quinquefasciatus is influenced by WNV blood meal titer, mosquito colony, simultaneous inoculation with CxFV, or inoculation with heat-inactivated CxFV. These data test the null hypothesis that there is no difference between the replication or transmission of WNV in CxFV Izabal (+) and CxFV Izabal (−) cells or mosquitoes
All CxFV experiments utilized CxFV Izabal isolate GU-06-2692, passage 1, isolated from a pool of Cx. quinquefasciatus in Puerto Barrios, Guatemala, 2006 [14]. West Nile virus isolate GU-06-2256, passage 3, also isolated from Cx. quinquefasciatus in Puerto Barrios, Guatemala, was used for flavivirus co-infection and vector competence studies. The growth of CxFV Izabal in cell culture was compared to that of KRV, strain SR-75 [4], [5]. Aedes albopictus C6/36-ATCC cells (American Type Culture Collection, Manassas, VA) maintained at 28°C in were used for growth and plaque titration of CxFV Izabal, and Vero (African green monkey kidney) cells maintained at 37°C were used for WNV plaque titrations. Two strains of Cx. quinquefasciatus were used in this study. The Sebring colony was originally established from Florida in 1988 and has been in colony at the CDC in Fort Collins since 2004. In an attempt to utilize viruses and vectors from the same geographic region, Cx. quinquefasciatus from Tegucigalpa, Honduras were colonized in September 2008. Generations F5/6 and F12 of the Honduras colony were used in this study. Prior to use, both colonies were confirmed CxFV-negative by RT-PCR.
CxFV Izabal was quantified from cell culture supernatant and homogenized mosquitoes by plaque titration on C6/36 cells [28]. Plaque assays were performed on C6/36 cell monolayers in 6-well plates using a double overlay method in nutrient media (5× Earle's BSS, 6.6% yeast extract-lactalbumin hydrolysate, 6% sodium bicarbonate, 4% FBS, and 0.4% gentamycin) mixed 1∶1 with 2.6% low-melt Sea Plaque agarose. Second overlay containing neutral red was added at seven DPI. WNV was quantified by Vero cell plaque assay using the double overlay method [28]. Second overlay containing neutral red was added 2 DPI.
CxFV Izabal viral RNA was also quantified by qRT-PCR. RNA extractions were performed using the QIAamp Viral RNA Mini Kit according to manufacturer's instructions (Qiagen Inc., Valencia, CA) with an elution volume of 100 µl. Quantification of viral RNA from 10-fold dilutions of RNA extracted from 100 µl stock virus of known concentration was used for the qRT-PCR standard curve. Four qRT-PCR primer and probe sets were designed from NS5 and E gene regions of CxFV RNA using Primer Express software (Applied Biosystems Inc, Foster City, CA) (Table 1). The complete genome sequence of CxFV (GenBank Accession number NC_008604) [10] and available RNA sequence from CxFV Izabal (EU805805, EU805806) [14] were used to select primers. CxFV Izabal primer and probe sensitivity and specificity were evaluated by sequence comparison to CFAV and CxFV (Table 1), and by testing each primer and probe set on a dilution series of available isolates of CxFV Izabal, WNV and KRV [4], [5].
qRT-PCR assays were correlated to plaque titration on C6/36 cells. Ten-fold serial dilutions of CxFV Izabal were split such that RNA was extracted from 100 µl of each virus dilution for quantification by qRT-PCR, and the remaining sample was subjected to plaque titration on C6/36 cells as described above. RNA copies/mL determined for each virus concentration were plotted against the corresponding pfu/mL determined by C6/36 cell plaque titration.
All animals were handled in strict accordance with the standards and policies of the Department of Health and Human Services' Office of Laboratory Animal Welfare (OLAW) and the US Department of Agriculture's Animal Welfare Act. All animal work was approved by the Centers for Disease control Institutional Animal Care and Use Committee, Protocol # 06-011. Hyperimmune polyclonal antisera against CxFV Izabal was generated in Swiss-Webster mice. Twenty-five adult female mice were housed in groups of five animals per cage. Each of four groups of five mice was immunized intraperitoneally with 0.1 mL CxFV Izabal virus seed (infected C6/36 cell, passage 1, tissue culture supernatant) diluted either 1∶10 or 1∶100 in Dulbecco's phosphate buffered saline (PBS). Mice were administered boosters of the same virus stock and concentration 3 wks and 6 wks following the initial vaccination. The fifth group was sham-inoculated with 0.1 mL PBS. Hyperimmunized mice were bled out by cardiac puncture three weeks following the third immunization. Blood was collected directly into microtainer tubes and centrifuged for serum separation. Pooled and individual aliquots of hyperimmune sera were stored at −80°C.
Because some antibodies in the sera were found to bind to C6/36 cells during immunofluorescence assay (IFA), sera were 4× cross-adsorbed against sonicated C6/36 cells to remove C6/36-specific antibodies. Uninfected C6/36 cells in DMEM maintenance medium containing 2% FBS were harvested from a T25 flask and washed once with PBS. Washed cells were pelleted by centrifugation at 5000 rpm for 5 min at 4°C, and resuspended in 1 mL PBS. Aliquots of 50 µl cell suspension were transferred to 0.2 mL PCR tubes and sonicated for 5 min. Five hundred microliters of pooled sera was cross-adsorbed with 200 µl sonicated C6/36 cell suspension at room temperature for 1.5 hrs with continuous mixing. Cell debris was removed from adsorbed sera by centrifugation at 5000 rpm for 5 min. The supernatant was adsorbed three additional times as above using fresh sonicated cells. Clarified antiserum was stored at −20°C.
Immunofluorescence assay using polyclonal mouse anti-CxFV serum was optimized using slides spotted with CxFV Izabal (+) and CxFV Izabal (−) C6/36 cells. To generate infected cells for spot slides, a T25 flask was inoculated with CxFV Izabal at a multiplicity of infection (MOI) of 0.1. Virus was allowed to adsorb for 1 hour at 28°C in 1 mL DMEM with 2% FBS, rocking every 15 min. After one hour, the volume of medium was increased to 5 mL. Cells were harvested at 4 DPI, washed twice with cold PBS, and acetone-fixed to 12-well multispot slides for 10 min (Thermo Electron Corp., Pittsburgh, PA). Uninfected C6/36 cells were harvested and fixed to slides as negative control. Spot slides were incubated with serial 2-fold dilutions of polyclonal mouse anti-CxFV for 30 minutes at 37°C in a humid box. Slides were washed twice for 10 min in PBS and air dried. Slides were then incubated for 30 min at 37°C in a humid box with secondary antibody conjugate AlexaFluor 488 goat anti-mouse IgG H+L (Invitrogen, Molecular Probes, Eugene, OR), diluted 1∶1000 in PBS with 0.08% trypan blue. Again, slides were washed twice with PBS, rinsed briefly with distilled water, and air dried. Cover slips were mounted using SlowFade Gold mounting medium (Invitrogen, Molecular Probes, Eugene, OR) and visualized with a Zeiss AxioImager Z1 (Carl Zeiss Microimaging, Inc., Thornwood, NY).
For IFA on mosquito tissues, mosquitoes were dipped briefly in 70% EtOH to destroy hydrophobicity. Midguts and ovaries were dissected in PBS using fine forceps on a microscope slide. Dissected tissues were placed on poly-L-lysine-coated slides (Polysciences, Inc.Warrington, PA) and allowed to dry. Wells were drawn around each tissue using a TexPen plastic pen (Mark-Tex Corp., Englewood, NJ), and tissues were fixed in ice cold acetone for 10 min. Headsquashes were performed by squashing dissected heads directly onto clean spot slides with a coverslip and manually removing pieces of cuticle, followed by acetone fixation. Immunostaining of mosquito tissues was performed as described above. For staining of CxFV+WNV co-infected tissues, human anti-WNV IgG obtained from the CDC reference collection (CDC, DVBID, Fort Collins, CO) was used as a primary antibody in addition to the anti-CxFV serum. Both CxFV and WNV antisera were used at 1∶320 dilution. Secondary staining of co-infected tissues utilized AlexaFluor 594 goat anti-mouse IgG H+L and AlexaFluor 488 goat anti-human IgG H+L each diluted 1∶1000 in PBS/trypan blue.
CxFV Izabal growth was measured in C6/36 cells and in Cx. quinquefasciatus. To measure virus growth in cell culture, C6/36 cells were inoculated with CxFV Izabal at an MOI of 0.03 or 0.1. Virus was allowed to adsorb for one hour at 28°C in 1 mL DMEM containing 2% FBS. After one hour, cells were washed three times with PBS, and 5 mL cell culture maintenance medium was added. One flask of no-virus control was maintained simultaneously. Supernatant aliquots were harvested from each flask at 0, 1, 2, 4, 6, 8, 10, 12, and 14 DPI and stored at −80°C. Samples were clarified by centrifugation and titrated by C6/36 cell plaque assay as described above.
For growth in vivo, groups of Cx. quinquefasciatus Sebring strain mosquitoes were infected with CxFV Izabal either by intrathoracic inoculation [29] or per os. For inoculations, approximately one week-old female Cx. quinquefasciatus were inoculated with 1.9 log10 pfu±1.6 log10 pfu CxFV Izabal. Mosquitoes were housed in screened paperboard pint containers held at 28°C and 95% relative humidity. Three to five mosquitoes were harvested on Days 0, 2, 4, 8, 12, and 14 post inoculation. For virus infection per os, C6/36 cells were inoculated with CxFV Izabal at an MOI of 0.1, as above. Virus-infected cell culture supernatant was harvested 4 DPI and clarified by centrifugation at 8000 rpm for 10 min at 4°C. The artificial blood meal contained two parts freshly-harvested, clarified CxFV Izabal in cell culture supernatant, two parts defibrinated chicken blood (washed 3× in PBS), and one part FBS+10% sucrose, warmed to 37°C. Culex quinquefasciatus Sebring mosquitoes were allowed to feed for 30 min from a Hemotek feeder (Discovery Workshops, Accrington, Lancashire, UK). All unfed and partially fed mosquitoes were removed and discarded. An aliquot of the infected blood meal was reserved and held at 37°C during the length of the feed, then stored at −80°C for titration. Three to five mosquitoes were harvested at 0, 2, 4, 8, 12, and 14 DPI and processed as described above.
The effect of CxFV Izabal infection on WNV growth in cell culture and in mosquitoes was also determined. In cell culture, C6/36 cells were inoculated as above with CxFV Izabal at an MOI of 0.1. At 2 DPI, the supernatant was removed, cells were washed 3× with PBS and infected with WNV at an MOI of 0.1. WNV was adsorbed for 1 hour at 28°C. Cells were washed with PBS and 5 mL DMEM maintenance medium replaced. Concurrently, a control flask uninfected with CxFV Izabal was inoculated with WNV in the same manner. An aliquot of supernatant was harvested from each flask on Days 0, 2, 4, 6, 8, 10, and 14 following WNV infection, and WNV titer determined by Vero cell plaque assay. For WNV growth in mosquitoes, Cx. quinquefasciatus Sebring mosquitoes were divided into three groups. The first group was inoculated intrathoracically with 3.3 log10 pfu CxFV Izabal, the second group was mock-inoculated with an empty glass capillary needle, and the third group was not inoculated. Seven days post inoculation, each group was administered a WNV-infectious blood meal of 6.3 log10 pfu/mL. Three to five mosquitoes were harvested on Days 0, 2, 4, 8, and 10 days post infection with WNV and processed as described above. For each time point, the average WNV titers in CxFV Izabal (+), CxFV Izabal (−) and mock-inoculated groups were compared by 2-tailed pairwise Student's t-tests at the 5% significance level, assuming unequal variances.
For each growth curve, mosquitoes were homogenized individually in 2 mL conical microcentrifuge tubes containing a single copper bb and 1 mL DMEM with 10% FBS. Mosquitoes were ground for 4 min at 20 cy/s on a mixer mill MM300 (Retsch, Haan, Germany). Homogenates were clarified by centrifugation at 8,000 rpm for 10 minutes at 4°C. Supernatants were stored at −80°C until virus quantification.
The ability of WNV to infect, disseminate, and be transmitted by Cx. quinquefasciatus infected with CxFV Izabal was evaluated across multiple WNV blood meal titers, routes of exposure to WNV, strains of Cx. quinquefasciatus mosquito, and prior infection with viable or inactivated CxFV Izabal. Artificial, infectious WNV blood meals were prepared as described above. In each experiment, one week-old Cx. quinquefasciatus were exposed to CxFV Izabal by intrathoracic inoculation with 2.8–3.3 log10 pfu seven days prior to receiving an artificial, WNV-infectious blood meal. Each CxFV Izabal (−) group was held, uninoculated, for one week and given the same WNV-infectious blood meal as the CxFV (+) group. In the first experiment, groups of Sebring strain Cx. quinquefasciatus, infected and uninfected with CxFV, received a WNV-infectious blood meal of 7.8 log10 pfu WNV per mL. In the second experiment, CxFV (+) and CxFV (−) Sebring and Honduras strain Cx. quinquefasciatus received WNV infectious blood meals of 8.9 log pfu per mL. In the third experiment, CxFV-positive and –negative Sebring and Honduras strain Cx. quinquefasciatus received a high titer (7.4–7.5 log pfu/mL) or low titer (5.4–5.6 log pfu/mL) WNV-infectious blood meal. Additional groups of mosquitoes were also inoculated with heat-inactivated (56°C for 45 min) CxFV Izabal to determine whether or not actively-replicating virus was necessary for any observed interference with WNV transmission.
For each WNV-infectious blood meal, an aliquot was reserved for plaque titration on Vero cells. Fully engorged mosquitoes were double-caged and held at 28°C at 95% relative humidity, and provided either 5% sucrose solution or raisins. At 14 days following the WNV-infectious blood meal, bodies, legs, and saliva were harvested from each live remaining specimen in each group and assayed for WNV by Vero cell plaque assay. Bodies and legs were each homogenized separately as described above. For saliva collections, specimens were knocked down by freezing at −20°C for 1 min, then, inside a glove box, wings were clipped off and the proboscis of each specimen was inserted into a capillary tube containing 5 µl Spectrosol immersion oil. After 20 min of salivation, specimens were removed from the capillary tube, and legs and bodies were separated into individual tubes. The tip of each capillary tube containing salivary expectorate was clipped off into a 1.7 mL microcentrifuge tube containing 450 µl DMEM with 10% FBS. Salivas were centrifuged for 5 minutes at 5000 rpm at 4°C to draw the oil out of the capillary tube, and stored at −80°C. The percentage of CxFV (+) and CxFV (−) mosquitoes that became infected, disseminated, and transmitted WNV were compared by Fisher exact test. The mean WNV titers in mosquito bodies and saliva 14 DPI between CxFV (+) and CxFV (−) experimental groups were compared by 2-tailed Student's t-tests assuming unequal variances.
To further evaluate WNV transmission in Cx. quinquefasciatus with a known WNV-disseminated infection, groups of Cx. quinquefasciatus Sebring and Honduras strain were inoculated with either WNV only or inoculated simultaneously with CxFV Izabal and WNV. Sebring specimens were inoculated with either 4.0 log10 pfu WNV (n = 66) or with 4.0 log10 pfu WNV+3.6 log10 pfu CxFV (n = 27) per mosquito. Honduras specimens were inoculated with either 3.9 log10 pfu WNV (n = 36) or 3.9 log10 pfu WNV+3.3 log10 pfu CxFV (n = 53) per mosquito. Nine days post inoculation, saliva was collected from each specimen as described above, and bodies were stored whole at −80°C. Salivary expectorates were analyzed as above.
Four novel quantitative RT-PCR (qRT-PCR) primer and probe sets were designed to amplify CxFV Izabal (Table 1). No amplification was obtained from WNV or KRV with any of the primer/probe sets. Primer and probe sequences were aligned to available genome sequences for CxFV [10] and CFAV to further examine their specificity for CxFV Izabal (Table 1). Correlation between CxFV Izabal qRT-PCR and C6/36 plaque assays was >99% (r = .9992). The equation for the trendline fit to the data was y = 2.47x, with the y-intercept fixed at zero (Fig. 1).
Replication kinetics of CxFV Izabal were determined in C6/36 cells and in Cx. quinquefasciatus Sebring strain mosquitoes (Figs. 2,3). Replication of WNV was also monitored in CxFV Izabal (+) and (−) C6/36 cells, and in CxFV Izabal (+) and (−) Cx. quinquefasciatus Sebring (Figs. 4,5). In C6/36 cells, CxFV Izabal (passage 1) reached a maximum titer of approximately 7.0 log10 plaque forming units (pfu)/mL six days following infection at either multiplicity of infection (MOI) of 0.03 or MOI = 0.1, approximately one log less than that observed for KRV (Fig. 2). CxFV Izabal caused evident cytopathic effects (CPE) in C6/36 cells, completely destroying the cell monolayer by 8 days post infection (DPI) following inoculation at an MOI of 0.1. In Cx. quinquefasciatus Sebring mosquitoes exposed to CxFV Izabal by intrathoracic inoculation, CxFV Izabal reached a peak titer of 4.3 log10 pfu/mosquito approximately 8 DPI. Mosquitoes were not susceptible to CxFV Izabal infection following oral exposure (Fig. 3).
Growth of WNV was not inhibited by CxFV Izabal in either C6/36 cells or Cx. quinquefasciatus Sebring mosquitoes (Figs 4,5). WNV titers were not significantly different between CxFV Izabal (+), CxFV Izabal (−) or mock-inoculated Cx. quinquefasciatus at 1, 2, 8, or 10 days following per os infection with WNV (p>0.05) (Fig. 5). At 0 DPI the average WNV titer in CxFV Izabal (+) mosquitoes was significantly less than in the mock-infected group (p = 0.016). It is unclear why this might be since mosquitoes were harvested immediately post-feeding and all three treatment groups imbibed the same WNV-infectious blood meal. At 4 DPI the average WNV titer in the CxFV Izabal (+) group was significantly higher than in the CxFV (−) mosquitoes (p = 0.0048) (Fig. 5). The biological significance of this observation is unclear, as this difference disappeared at 8 and 10 DPI (Fig 5). It is probable that this difference between groups is an artifact of small sample sizes (3–5 mosquitoes per time point).
The percentage of CxFV (+) and CxFV (−) Cx. quinquefasciatus that became infected, developed a disseminated infection, and transmitted WNV were not significantly different for either mosquito strain or any WNV blood meal titer examined (Fisher Exact test, p>0.05, Table 2). Furthermore, WNV infection, dissemination, and transmission rates in mosquitoes inoculated with heat-inactivated CxFV Izabal did not differ significantly from those inoculated with live CxFV Izabal or uninfected with CxFV Izabal (Table 2). There was extensive variation in WNV body and saliva titers in each of these groups (Table 3). West Nile virus titers in mosquito bodies and salivary expectorates were not significantly different between CxFV Izabal (+) and CxFV Izabal (−) Cx. quinquefasciatus when mosquitoes were exposed orally to WNV (Table 3, Student's t-test, p>0.05). One group of Sebring Cx. quinquefasciatus and one group of Honduras F12 Cx. quinquefasciatus failed to become infected (Table 2). We speculate that these results were not due to experimental treatment, but rather to small sample sizes and the relatively low WNV titer in those particular blood meals, potentially approaching a threshold of infection of approximately 5 log10 pfu WNV/mL [30], [31].
A significantly higher percentage of Honduras Cx. quinquefasciatus transmitted WNV when co-inoculated simultaneously with CxFV Izabal (98%, n = 53) than when inoculated with WNV alone (69%, n = 36) (p = 0.0014, Fisher Exact test) (Fig. 6). The percentage of Sebring Cx. quinquefasciatus that transmitted WNV when co-inoculated simultaneously with CxFV Izabal (93%, n = 27) was not significantly different from those inoculated with WNV alone (88%, n = 66) (p>0.05, Fisher exact test) (Fig. 6). The percentage of intrathoracically-inoculated specimens that transmitted WNV alone was also significantly less in the Honduras colony as compared with the Sebring colony, suggesting a more effective salivary gland barrier to WNV in the Honduras colony (p = 0.033, Fisher exact test); 87% of Sebring specimens (n = 66) transmitted WNV compared with only 69% (n = 36) of the Honduras specimens. For the Sebring colony, the average WNV titer in salivary expectorates for specimens inoculated with WNV only was 4.4 log10 pfu (n = 58), and not significantly different from an average titer of 4.7 log10 pfu in the expectorates of WNV+CxFV Izabal group (n = 25) (Student's two-tailed t-test, p = 0.11). For the Honduras colony, the average WNV titer in salivary expectorates for specimens inoculated with WNV only was 4.6 log10 pfu (n = 25) compared with 4.8 log10 pfu in the WNV+CxFV Izabal group (n = 52) (Student's two-tailed t-test, p = 0.38). For these groups, co-inoculated mosquitoes that transmitted WNV also contained CxFV Izabal in their saliva and mosquitoes that did not transmit WNV also did not transmit CxFV (n = 12). Mosquitoes infected with CxFV Izabal only (n = 5) did not have CxFV Izabal in their saliva. Midgut (Fig. 7) and head tissues (Fig. 8) of mosquitoes inoculated simultaneously with CxFV Izabal and WNV were observed to be infected with both viruses by IFA.
In this study we demonstrated that sequential infection of C6/36 cells or Cx. quinquefasiatus mosquitoes with CxFV Izabal and WNV did not interfere with either growth or transmission of WNV. This finding is not surprising given that Culex flaviviruses are being discovered in mosquito populations around the world in locations where WNV and other flaviviruses circulate sympatrically. Therefore, the prevalence of CxFV Izabal in Cx. quinquefasciatus in Guatemala does not explain the lack of human disease attributable to WNV in this region.
Growth of WNV in C6/36 cells was not inhibited by prior infection of CxFV Izabal. The WNV titer in CxFV Izabal (+) C6/36 cells did not reach the maximum titer observed in CxFV Izabal (−) cells due to death of cells caused by CxFV Izabal (Fig. 4). As suggested by Hoshino et al. [10], CPE observed in C6/36 cells may be the result of an unnatural association between this Culex-derived virus and Aedes-derived cell line since CxFV apparently replicates avirulently in its mosquito host. Therefore, future studies should include utilization of Culex-derived cell lines. The natural host range of CxFV Izabal across mosquito species and genera is not known.
Data regarding the establishment of superinfection by homologous viruses in cell culture have been variable. C6/36 cells persistently infected with Aedes aegypti densonucleosis virus remained permissive to infection with Haemagogus equinus densovirus (HeDNV), arguing against the induction of an anti-viral or immune state in the cells that would otherwise inhibit superinfection by this a similar virus [32]. However, interference between superinfecting alphaviruses in mosquito cell culture has been documented multiple times [22]–[24]. The cellular and molecular mechanisms that support replication of WNV in CxFV (+) cells are not known and require further study.
Overall, neither growth nor transmission of WNV in Cx. quinquefasciatus was significantly affected by CxFV Izabal when viruses were administed sequentially. These findings are in contrast to what has been found previously for flavivirus - flavivirus superinfections involving WNV in Culex mosquitoes, however insect-only flaviviruses are fairly divergent from other vector-borne flavivirues such as WNV [5], [10], [33]. Previous studies have demonstrated that transmission of a superinfecting flavivirus was blocked if the secondary flavivirus was antigenically-similar to the primary infecting flavivirus [17], [18]. Interference to arbovirus superinfection in mosquitoes or mosquito cells by homologous viruses could be the result of RNA interference (RNAi). RNAi is a mechanism by which invertebrates respond to viral infection through the specific recognition and degradation of viral mRNA sequences by virus-derived small interfering RNAs (viRNA) [34]–[38]. However, our data demonstrate that prior infection with CxFV Izabal does not interfere with WNV replication when the viruses are inoculated simultaneously, or when mosquitoes are exposed to WNV one week following inoculation with CxFV Izabal. If an RNAi pathway was induced in Cx. quinquefasciatus by CxFV Izabal it would most likely still be effective seven days post-inoculation when mosquitoes were exposed to WNV as it has been previously reported that viRNAs targeting Sindbis virus in C6/36 cells were first detected 48h following infection and were still abundant 7 DPI [38], and viRNAs targeting WNV in Cx. quinquefasciatus midguts were detected 7 and 14 days post exposure to WNV [37].
There are numerous potential reasons why mosquitoes would be permissive to co-infecting flaviviruses. First, if CxFV Izabal induced an RNAi response in Cx. quinquefasciatus, the viRNAs generated might not be sufficiently homologous to WNV to interfere with the establishment of WNV infection. RNAi is a highly sequence-specific mechanism with little tolerance to mismatches between the viRNA trigger and mRNA target sequences [39], and the nucleotide sequence identity between CxFV and other vertebrate-infecting flaviviruses is relatively low. Kim et al. [13] reported between 25 and 52% nucleotide sequence homology between CxFV (TX24518) and WNV among structural and non-structural genes. Similarly, Hoshino et al. [10] reported 17–25% sequence identity for structural proteins and 17–40% identity among non-structural proteins between their Japanese isolate of CxFV and other flaviviruses. Therefore, any interference of CxFV Izabal viRNAs with WNV was probably minimal. Secondly, it is possible that over a history of co-evolution between insect-only flaviviruses and their mosquito hosts, these viruses have evolved a way to either evade or suppress an immune mechanism that would otherwise interfere with their own replication, or replication of a subsequently-infecting virus. Flock house virus encodes an RNAi suppressor protein, B2, that is necessary for establishment of viral infection in Drosophila S2 cells [40]. Virus-encoded suppressors of RNAi have also been found in plant viruses such as tobacco etch potyvirus [41]. The molecular mechanisms that permit co-existence of both WNV and CxFV Izabal, potentially even within the same tissues and cells, requires further study. Thirdly, CxFV replication in mosquito cells is presumably similar to that of other flaviviruses due to similar genome organization [10], [13]. Zebovitz and Brown [22] determined that interference of superinfecting alphaviruses in cell culture was due to competition for replication sites or metabolites, and that viral RNA synthesis was necessary for inhibition of alphavirus superinfection. The non-structural proteins encoded by flaviviruses play important roles in virus replication and maturation [42], and the 5′ and 3′ untranslated regions contain conserved nucleotide sequences and RNA secondary structures involved in virus replication and translation [43]. Camissa-Parks et al. [44] discovered that the 3′ stem loop structure of CFAV differed from that of other vertebrate-infecting flavivirus RNAs, and the 3′ pentanucleotide sequence, which is completely conserved among mosquito- and tick-borne flaviviruses contained a point mutation in cell fusing agent virus. The function of this pentanucleotide sequence element is thought to be involved in the binding of cellular or viral proteins to the 3′ stem loop structure during RNA replication [43]. The 3′ UTR of CxFV also was found to contain four tandem repeats, hypothesized to be specially adapted for replication in the mosquito host [10] since deletion of conserved tandem repeat sequences alters virus growth properties [43]. Finally, CxFV may target and replicate in different mosquito tissues than WNV or other flaviviruses. In Culex quinquefasciatus inoculated simultaneously with CxFV Izabal and WNV, midgut and head tissues became infected with both viruses, demonstrating a potential for physical interaction between CxFV Izabal and WNV (Figs. 7,8). However it is unclear if these tissue tropisms would be the same for mosquitoes naturally-infected with CxFV or if infection of these tissues is an artifact of inoculation, or inoculation simultaneously with WNV. More work is needed on characterizing the tissue tropisms of CxFV in naturally-infected mosquitoes and the mechanism by which this virus propagates and is transmitted.
One limitation of this study is that the natural mechanism by which mosquitoes become infected with CxFV has not yet been elucidated, so mosquitoes in this study were infected with CxFV Izabal by intrathoracic inoculation. It is unclear how or if the results of this study would be different using naturally-infected mosquitoes. Route of infection has been shown to affect the outcome of arbovirus superinfection studies. Most notably, Aedes triseriatus mosquitoes that were infected transovarially with LaCrosse virus remained permissive to secondary infection with a homologous or heterologous bunyavirus [45], whereas mosquitoes exposed to the primary infection by intrathoracic inoculation became refractory to superinfection after seven days [20]. Ideally, these studies should be repeated using mosquitoes naturally-infected with CxFV to fully understand the dynamics of interaction, or lack thereof, between these two flaviviruses within the mosquito vector. However the advantage of inoculations is that experiments can be standardized by infecting mosquitoes of the same age with approximately the same amount of virus, and 100% infection rates are assured.
Interestingly, both CxFV Izabal and WNV were found in saliva of co-infected specimens when mosquitoes were exposed to both viruses simultaneously by intrathoracic inoculation, but no CxFV Izabal was found in the saliva of singly-infected specimens. This observation suggests that CxFV Izabal may be infecting the salivary glands by “piggybacking”on WNV. This phenomenon has been suggested for expansion of cellular tropism by human immunodeficiency virus (HIV), whereby Epstein-Barr virus, cytomegalovirus, human t-lymphotrophic virus, and sperm proteins share large regions of similarity with the CD4 protein of T-helper lymphocytes, a cellular receptor used by HIV [46]. Because HIV binds to CD4, binding of HIV to CD4 homologues on other co-infecting viruses or sperm may allow HIV to “piggyback” into additional cell types which it normally would not infect [46]. The molecular basis for this interaction between CxFV Izabal and WNV and how these results compare to natural infection is unknown. Our intrathoracic inoculation data suggest that CxFV Izabal may have the potential to enhance WNV transmission in some mosquito populations; however WNV transmission was not enhanced in Honduras colony when mosquitoes were exposed per os (Table 2).
In summary, this is the first study to address the potential effect of an insect-specific flavivirus on transmission of WNV. We have demonstrated that CxFV Izabal does not interfere with growth of WNV in C6/36 cells or in Cx. quinquefasciatus, nor does it inhibit infection, dissemination, or transmission of WNV. These findings are in contrast to what would be expected based on previous studies following flavivirus – flavivirus superinfections. We hypothesize that both CxFV Izabal and WNV have evolved mechanisms for persistence and transmission by a common mosquito vector, Cx. quinquefasciatus, despite the presence of mosquito immune defenses and the prevalence of co-circulating flaviviruses. Future studies should address the effect of CxFV and WNV co-infection in mosquitoes naturally infected with CxFV, as well as the tissue tropisms and molecular mechanisms of CxFV replication and transmission in mosquitoes.
|
10.1371/journal.pbio.0060326 | Nutrient-Regulated Antisense and Intragenic RNAs Modulate a Signal Transduction Pathway in Yeast | The budding yeast Saccharomyces cerevisiae alters its gene expression profile in response to a change in nutrient availability. The PHO system is a well-studied case in the transcriptional regulation responding to nutritional changes in which a set of genes (PHO genes) is expressed to activate inorganic phosphate (Pi) metabolism for adaptation to Pi starvation. Pi starvation triggers an inhibition of Pho85 kinase, leading to migration of unphosphorylated Pho4 transcriptional activator into the nucleus and enabling expression of PHO genes. When Pi is sufficient, the Pho85 kinase phosphorylates Pho4, thereby excluding it from the nucleus and resulting in repression (i.e., lack of transcription) of PHO genes. The Pho85 kinase has a role in various cellular functions other than regulation of the PHO system in that Pho85 monitors whether environmental conditions are adequate for cell growth and represses inadequate (untimely) responses in these cellular processes. In contrast, Pho4 appears to activate some genes involved in stress response and is required for G1 arrest caused by DNA damage. These facts suggest the antagonistic function of these two players on a more general scale when yeast cells must cope with stress conditions. To explore general involvement of Pho4 in stress response, we tried to identify Pho4-dependent genes by a genome-wide mapping of Pho4 and Rpo21 binding (Rpo21 being the largest subunit of RNA polymerase II) using a yeast tiling array. In the course of this study, we found Pi- and Pho4-regulated intragenic and antisense RNAs that could modulate the Pi signal transduction pathway. Low-Pi signal is transmitted via certain inositol polyphosphate (IP) species (IP7) that are synthesized by Vip1 IP6 kinase. We have shown that Pho4 activates the transcription of antisense and intragenic RNAs in the KCS1 locus to down-regulate the Kcs1 activity, another IP6 kinase, by producing truncated Kcs1 protein via hybrid formation with the KCS1 mRNA and translation of the intragenic RNA, thereby enabling Vip1 to utilize more IP6 to synthesize IP7 functioning in low-Pi signaling. Because Kcs1 also can phosphorylate these IP7 species to synthesize IP8, reduction in Kcs1 activity can ensure accumulation of the IP7 species, leading to further stimulation of low-Pi signaling (i.e., forming a positive feedback loop). We also report that genes apparently not involved in the PHO system are regulated by Pho4 either dependent upon or independent of the Pi conditions, and many of the latter genes are involved in stress response. In S. cerevisiae, a large-scale cDNA analysis and mapping of RNA polymerase II binding using a high-resolution tiling array have identified a large number of antisense RNA species whose functions are yet to be clarified. Here we have shown that nutrient-regulated antisense and intragenic RNAs as well as direct regulation of structural gene transcription function in the response to nutrient availability. Our findings also imply that Pho4 is present in the nucleus even under high-Pi conditions to activate or repress transcription, which challenges our current understanding of Pho4 regulation.
| How does a microorganism adapt to changes in its environment? Phosphate metabolism in the budding yeast Saccharomyces cerevisiae serves as a model for investigating mechanisms involved in physiological adaptation. The nutrient inorganic phosphate (Pi) is essential for building nucleic acids and phospholipids; when yeast cells are deprived of Pi, genes required for scavenging the nutrient are activated. This activation is mediated by the Pho4 transcription factor through its migration into or out of nucleus. The Pi-starvation (low-Pi) signal is transmitted by a class of inositol polyphosphate (IP) species, IP7, which is synthesized by one of two IP6 kinases, Vip1 or Kcs1. However, the IP7 made primarily by Vip1 is key in the signaling pathway. Here we report that under Pi starvation Pho4 binds within the coding sequence of KCS1 to activate transcription of both intragenic and antisense RNAs, resulting in the production of a truncated Kcs1 protein and the likely down-regulation of Kcs1 activity. Consequently Vip1 can produce more IP7 to enhance the low-Pi signaling and thus form a positive feedback loop. We have also demonstrated that Pho4 regulates, both positively and negatively, transcription of genes apparently uninvolved in cellular response to Pi starvation and that it sometimes does so independently of Pi conditions. These findings reveal mechanisms that go beyond the currently held model of Pho4 regulation.
| When environmental conditions change, the budding yeast Saccharomyces cerevisiae, like other microorganisms, makes a decision about growth, cell division, and which responses to elicit in a coordinated fashion. Starvation for nutrients, alterations in temperature or salt concentration, and the presence of toxic agents are critical stresses for yeast cells and elicit signals that evoke cellular responses favoring survival under the new conditions. Nutrient status is probably the most important condition that must be accurately and rapidly sensed and responded to in order to ensure cell survival. In this process, nutrient-sensing kinases including cyclic-adenosine-monophosphate-dependent kinase, Snf1, Tor, and Pho85 kinases play important roles in regulation at levels ranging from transcription to the activity of individual enzymes [1,2]. Transcriptional regulation is the most fundamental process in the nutritional response, and DNA-binding transcription factors and genes under their control are extensively characterized by conventional genetic and biochemical analyses and, more recently, expression profiling via DNA microarray and chromatin immunoprecipitation (ChIP)-on-chip analysis. The PHO system is a well-studied case in which a set of genes (PHO genes) is expressed to activate inorganic phosphate (Pi) metabolism for adaptation to Pi starvation [3]. The Pho4 transcription factor that activates PHO genes is regulated by phosphorylation to alter its cellular localization: under high-Pi conditions, the Pho85 kinase phosphorylates Pho4, thereby excluding it from the nucleus and resulting in repression (i.e., lack of transcription) of PHO genes. Pi starvation triggers an inhibition of Pho85 kinase, leading to the migration of unphosphorylated Pho4 transcriptional activator into the nucleus and enabling expression of PHO genes [4–6]. Transcriptional regulation responding to nutrient change is also extensively studied in glucose repression and in amino acid starvation, cases in which a complex interplay between activators and repressors acting on the structural genes involved in the respective process is well documented [7,8].
Recent studies on transcriptional regulation have revealed the participation of novel regulators in addition to protein factors, specifically, an involvement of RNA in the regulation of protein expression responding to external signals including nutrient changes [9,10]. Prokaryotic mRNAs that change their conformation upon binding of specific metabolites can alter transcription elongation or translation initiation and are called riboswitches [11]. Noncoding (nc) RNAs including small inhibitory (si), micro (mi), and small nucleolar (sno) RNAs modify RNA species to regulate gene expression: siRNA and miRNA target mRNA to cause mRNA cleavage and inhibition of translation, respectively, whereas snoRNA targets rRNA. Numerous ncRNAs, however, have been found that do not show these known functions, including antisense (AS) RNAs and transcribed pseudogenes [10]. In S. cerevisiae, several ncRNAs involved in transcriptional regulation are reported: SRG1 intergenic RNA functions in repression of SER3 [12,13], and an AS RNA in the PHO5 locus appears to facilitate PHO5 transcription upon activation [14], whereas those in PHO84 and IME4 function in gene silencing in aging cells [15] and inhibition of transcription [16], respectively (Accession numbers for genes described in this article are listed in Table S1). Recent large-scale cDNA analysis [17,18] and mapping of RNA polymerase II binding using a high-resolution tiling array [19] have identified a large number of intergenic, intragenic, and AS RNA species whose functions are yet to be clarified.
The Pho85 kinase has a role in various cellular functions other than regulation of the PHO system via Pho4; these functions include nutrient sensing, cell cycle progression, stress response, and control of cell morphology [20–23]. Pho85 monitors whether environmental conditions are adequate for cell growth and represses inadequate (untimely) responses in these cellular processes [1]. When nutrient is sufficient, the kinase phosphorylates Gsy2, a glycogen synthase [24], and represses the expression of UGP1, which encodes an enzyme that catalyzes the production of UDP-glucose for glycogen synthesis [25]. Both of these events lead to the down-regulation of glycogen synthesis. Pho85 also facilitates the degradation of Gcn4, a transcription factor that activates genes involved in amino acid metabolism when amino acids are depleted [26].
In contrast, the known cellular function of Pho4 seems rather limited to the PHO system [3]. Recent microarray analysis, however, has demonstrated that some genes involved in stress response and various other metabolic functions are activated under Pi-limiting conditions [27], implying that Pho4 may activate these genes. Indeed, Pho4 is required for G1 arrest caused by DNA damage [28]. These observations suggest that Pho4 facilitates stress response by activating genes involved in the process. The fact that overproduction of Pho4 causes growth arrest of yeast cells in the absence of Pho85 [29] supports the antagonistic function of these two players on a more general scale when yeast cells must cope with stress conditions. To explore the general involvement of Pho4 in stress response, we tried to identify Pho4-dependent genes by a genome-wide mapping of Pho4 and Rpo21 binding (Rpo21 being the largest subunit of RNA polymerase II) using a yeast tiling array. In the course of this study, we found that Pho4- and Pi-dependent AS and intragenic RNAs modulate Pi signaling, leading to stimulation of expression of PHO genes, which demonstrates that nutrient-regulated RNA species other than mRNA are functioning in nutrient-responsive pathways in yeast cells. We also found that Pho4 was involved in transcriptional regulation of stress-responsive genes, either positively or negatively, and in some cases independently of environmental Pi conditions, which challenges the current model of Pho4 regulation [4,5].
To analyze the Pho4 binding sites in the yeast genome, we used two kinds of oligo-DNA arrays, an Affymetrix high-density oligo-DNA array harboring 25-mer oligonucleotides with 4-nucleotide spatial resolution (high-resolution [HR] chip) and an Agilent yeast whole genome 44K array that had 60-mer oligos with ca. 270-nucleotide spatial resolution (low resolution [LR] chip). The complete datasets of the HR chip analysis are found in the NCBI GEO database (http://www.ncbi.nlm.nih.gov/geo/) under accession number GSE13350. Analysis with the HR and LR arrays revealed that Pho4 bound to 51 and 57 genes, respectively, under low- but not high-Pi conditions (Table S2). Thirty-five genes were common in the two analyses, and all but two of these (KCS1 and SHE9) had a prospective Pho4 binding sequence (CACGTG/T or CTGCAC) in their upstream regions. For the two exceptions, a binding sequence was present within the ORF (Table S2). Among the 35 genes, 16 had already been reported as PHO genes, that is, genes regulated by Pho4 in a Pi-dependent manner, by genetic, biochemical, and microarray analyses [30,31]. Our results demonstrated that Pho4 actually binds to the upstream regions of these genes depending on Pi conditions. In addition to the known PHO genes, our analyses identified 19 genes as possible novel PHO-type genes. Among them, eight genes, PST1, MNN1, HOM3, HOR7, PTK2, CBF1, SUR1, and GLN1, showed the expected pattern of Rpo21-binding (Table S2). Representative results for PHO8 and PST1 are shown in Figure 1A: Pho4 binding to the upstream regions of these two genes depending upon Pi condition is demonstrated by the two different ChIP-on-chip analyses (Figure 1A), Pi- and Pho4-dependent transcription by Rpo21 binding (Figure 1A) and by northern analysis (Figure 1B), and a Pi-dependent in vivo binding of Pho4 to their promoter regions by gene-specific PCR of a chromatin-immunoprecipitated (ChIPed) fragment (Figure 1C). Pi- and Pho4-dependent transcription of MNN1, PTK2, CBF1, and PST1 was demonstrated by northern analysis (Figure 1B). In vivo binding of Pho4 to the upstream regions of these genes depending on Pi conditions was demonstrated by PCR using ChIPed DNA as a template and primers specific to the respective gene, which was in good accordance with the results of ChIP-on-chip analysis (Figure 1C and Table S2). AIR2 serves as a negative control for Pho4 binding, because it has no prospective Pho4 binding sites in its ORF and its expression is not affected by Pi conditions or Pho4. ARO9 shares a divergent promoter with SPL2, a known PHO-gene. Expression of ARO9 appeared dependent on both Pi and Pho4 (Figure 1B), and Pho4 bound to its promoter in a Pi-dependent fashion (Figure 1C). Binding of Pho4 to the CYC3-CDC19 region appeared to be Pi-dependent (Table S2), but CDC19 expression appeared independent of Pi conditions or Pho4 (Figure 1B and 1C), indicating that Pho4 binding to the upstream region of CDC19 did not play a major role in transcriptional regulation of CDC19.
ChIP-on-chip analyses with the HR and LR arrays revealed that 140 and 30 genes, respectively, showed Pi-independent binding of Pho4, and among them, nine genes (URA3, MUC1, CIS3, ILV3, PDC1, YPS3, YLR137W, HPF1, and ASN1) were commonly detected. Each of these nine genes has the prospective Pho4 binding site in its promoter or ORF (Table S3). We focused on ILV3, ASN1, CIS3, and YPS3 for further analysis. Analyses of the binding profiles of Pho4 to these genes by ChIP-on-chip analysis using the two different platforms are shown in Figure 2A. Although the binding profiles of Rpo21 in the wild-type (wt) or Δpho4 mutant exhibited some inconsistency with those expected from Pho4 binding profiles, northern analysis demonstrated Pi-independent but Pho4-dependent expression, either a decrease (ILV3 and ASN1) or an increase (CIS3 and YPS3) in the absence of Pho4 (Figure 2B). Gene-specific PCR demonstrated that Pho4 bound to the upstream region of these genes in vivo irrespective of Pi conditions (Figure 2C). These results raise possibilities that Pho4 can bind to a promoter under high-Pi conditions and that Pho4 binding can lead to transcriptional activation (ASN1 and ILV3) or repression (CIS3 and YPS3). The former hypothesis challenges the current model of Pho4 regulation in which, under high-Pi conditions, Pho4 is excluded from the nucleus through phosphorylation by Pho85-Pho80 [5]. Therefore, we further analyzed whether the activity of the ASN1 promoter was dependent on Pho4 under high-Pi conditions by measuring reporter activity in cells grown in a high-Pi medium. The wt promoter was active under high-Pi conditions (Figure 2D), and its activity level decreased to almost 50% when a prospective Pho4 binding site (at −451 with A of ATG as +1) in the promoter was mutated (ASN1mut). In the absence of Pho85 where Pho4 became active, the activity of the wt promoter was stimulated by 1.3-fold compared to that in the wt strain, but the ASN1mut promoter showed a level of the activity similar to that observed in the wt strain. In the absence of Pho4 (Δpho4), the activity of the wt ASN1 promoter decreased to 50% whereas that of the mutant promoter showed a similar level of the activity in the wt cells. Gcn4 activates ASN1 under amino-acid-starvation conditions [32], and in the absence of Gcn4 (Δgcn4), the promoter activity was decreased drastically (Figure 2D). These results further demonstrated that Pho4 can activate ASN1 regardless of Pi conditions, which requires the Pho4 binding sequences in its promoter.
Our ChIP-on-chip analyses using two different platforms demonstrated the binding of Pho4 within the KCS1 and SHE9 ORFs depending on Pi conditions (Figure 3A and Table S2). Because, to our knowledge, gene transcription mediated by the binding of a yeast transcription factor within an ORF is very rare with only three precedents [33–35], we further analyzed the regulation of KCS1 by Pho4. Whereas gene-specific PCR using a primer set specific to the KCS1 promoter (−908 to +70) failed to detect an enrichment of the Pho4-bound fragment (Figure 3B, top panel), the ORF-specific primer (+36 to +1102) could detect an enrichment of the Pho4-bound fragment prepared only from cells grown in low-Pi medium (Figure 3B, second panel from the top). In the absence of Pho85 where Pho4 became active, Pho4 binding to the ORF was detected under both high- and low-Pi conditions (Figure 3B, bottom panel) whereas that to the upstream region was not (Figure 3B, second panel from the bottom). These results demonstrated that Pho4 binding within the KCS1 ORF depends on Pi conditions. We also analyzed Rpo21 binding by ChIP-on-chip and found that Rpo21 was localized to the KCS1 locus in Pi- and Pho4-dependent manners (Figure 3A, bottom two panels). This result indicated the presence of Pi- and Pho4-dependent transcription in the KCS1 locus. To examine whether binding of Pho4 within the ORF could direct KCS1 mRNA transcription, initiation of transcription within the ORF (intragenic RNA), or synthesis of AS RNA, we carried out northern analysis using RNA probes specific to the sense or antisense strands of KCS1 (Figure 3C). With an RNA probe hybridizing with the KCS1 mRNA (Figure 3E), a transcript of ca. 3,200 nucleotides (nt), approximately the size of the KCS1 ORF (3,150 nt; Figure 3E), was detected, which was not dependent on either Pho4 or Pi conditions (Figure 3C, top panel, lanes 1 to 4, designated by upper arrow). Judging from its size, this RNA species is highly likely to be the KCS1 mRNA, because this band was not detected in Δkcs1 cells (unpublished data). In the absence of Pho85, a transcript with a smaller size (ca. 2,600 nt) was detected (Figure 3C, top panel, lanes 5 and 6, designated by lower arrow), together with the KCS1 mRNA and short transcripts ranging from 2,300 to 1,800 nt in length (lanes 2, 5, and 6, designated by a vertical bar). These short sense transcripts were also detected in the wt cells under low-Pi conditions, though weakly (lane 2), but not observed in a Δpho4 (lanes 3 and 4) or Δpho4 Δpho85 double mutant (lane 7). These results indicated that Pho4 binding within the KCS1 ORF can activate the transcription of RNAs shorter than the KCS1 mRNA in both Pi- and Pho4-dependent manners. It should be noted that the KCS1 mRNA level appeared reduced when these short RNA species were abundant in Δpho85 cells (lanes 5 and 6). On the other hand, an AS RNA probe covering from −715 to +295 of the KCS1 gene (Figure 3E) could detect a short transcript of ca. 500 nt, which was dependent on both Pho4 and Pi conditions (Figure 3C, middle panel, lanes 1 to 7), whereas that covering from −715 to −228 failed to do so (unpublished data). Therefore, the AS RNA was highly likely to be encoded between +295 and −228. The presence of AS RNA in PHO genes is reported in PHO5 [14] and PHO84 [15], and in both cases, RRP6 affects the stability of the AS RNA. We tested the effect of a Δrrp6 mutation and found that the mutation also stabilized the AS RNA in the KCS1 locus while not significantly affecting the amount of sense RNAs (Figure 3D).
To examine whether the 2,600 nt transcript is a processing product and to confirm the presence of the AS RNA, we determined transcriptional start points of the sense and AS RNAs in the Δpho85 strain by the 5′ rapid amplification of cDNA ends (RACE) method and found that the sense RNAs started mainly at −14 and +537 and that the AS transcription started at +235 (Figure 3E). The sense RNA starting at −14 is highly likely to be the KCS1 mRNA, because the transcription start points of the sense RNA in the wt were also mapped mainly to this point (unpublished data). The one starting at +537 can be ca. 2,600 nt in length when transcribed through the ORF, the size of which coincides well with the estimated size of the short transcript detected by northern analysis (Figure 3C, top panel, designated by lower arrow). This result supported the conclusion that the 2,600 nt RNA is not a processing product but is transcribed from within the KCS1 coding sequence.
Three prospective Pho4 binding sites (at +406, +1127, and +1193) are present in the 5′ half of the KCS1 ORF (Figure 3E), and the one closest to the 5′ end is sandwiched by the transcription start points of the intragenic (+537) and AS RNAs (+235), which suggests a possibility that binding of Pho4 to the +406 site activates both of the 2,600 nt sense and AS RNAs. We constructed the KCS1 mutant (KCS1mut) in which three prospective Pho4 binding sites were mutated while keeping the amino acid sequence intact and analyzed the wt or KCS1mut in a low copy (YCp) plasmid in Δkcs1 cells. We found that the KCS1 mRNA was normally produced (Figure 4A, top panel, lanes 1 to 4), whereas the AS and intragenic RNAs were produced from wt KCS1 but not from KCS1mut under low-Pi conditions (second panel, lanes 2 and 4). Both the AS and truncated sense RNAs were synthesized from wt KCS1 in Δkcs1 Δpho85 cells (lanes 5 and 6) whereas KCS1mut produced RNA of the wt size but not the intragenic or AS RNAs (lanes 7 and 8). These results indicated that the generation of the 2,600 nt intragenic as well as AS RNAs are activated by Pho4, which requires the presence of at least one of the prospective Pho4 binding sites. Short sense transcripts around 2,000 nt (Figure 4A, lanes 2, 5, and 6, designated by a vertical bar) decreased to below detectable levels with KCS1mut (lanes 7 and 8), which suggested that these transcripts are also dependent on Pho4.
Antisense RNA functions in cis to inhibit sense RNA transcription by transcriptional collision as reported in the IME4 case [16] and in trans to form a hybrid with sense RNA to inhibit its function [9,10]. To reveal the role of the KCS1 AS RNA that is regulated by Pho4 and Pi conditions, we first asked whether the AS RNA could affect the synthesis of the Kcs1 protein by immunoblotting. Kcs1 protein whose C-terminus was tagged with the c-myc epitope was produced from the wt KCS1 or KCS1mut gene in a YCp plasmid (Figure 4A, second panel from the bottom). Kcs1 protein of the wt size (1,143 amino acids including the myc tag) was detected in all cases (lanes 1 to 8), and Δkcs1 Δpho85 cells harboring the wt KCS1 plasmid produced a truncated protein together with the normal one (Figure 4A, second panel from the bottom, lanes 5 and 6). The band observed between 83 and 62 kDa markers was nonspecific staining with anti-myc antibody because it was also detected in the absence of Kcs1-myc protein in the extract (unpublished data). Because both the AS and truncated sense RNAs were produced in Δkcs1 Δpho85 cells, the AS RNA hybridizing to the 5′ region of the KCS1 mRNA might inhibit normal translation initiation or the truncated sense RNA might be translated, either of which could use the in-frame initiation codon at +676 (Figure 3E) to produce truncated Kcs1 protein composed of 825 amino acids plus 93 amino acids from the c-myc tag. To examine whether the AS RNA could act in trans, for example, by hybridizing to the 5′ region of the KCS1 mRNA, we constructed a plasmid in which the AS RNA was produced from the GAL1 promoter by placing a KCS1 fragment covering +295 to −950 downstream of the promoter and introduced it into a Δkcs1 strain harboring the KCS1mut gene, so that the AS RNA was provided only in trans and the short sense transcripts including the 2,600 nt intragenic RNA were not produced (Figure 4A, top panel, lane 7). The truncated Kcs1 protein was observed only when transcription of the AS RNA was induced from the plasmid in galactose medium (Figure 4B, second panel from the bottom, lanes 9 and 10). As expected, the short sense transcripts were not detected when the AS RNA was overproduced in trans (Figure 4B, top panel). These results indicated that the AS RNA can act in trans (i.e., inhibition of the normal translation initiation, possibly by hybrid formation with the KCS1 mRNA). To test hybrid formation, we tried to detect the presence of the double-stranded (ds) RNA in the total RNA sample by RNase protection analysis using single-stranded RNA-specific RNase, followed by reverse transcription (RT) and PCR amplification of the protected fragments. Reverse transcription was carried out using either sense- or antisense-strand-specific primers, hybridizing to from +243 to +223 and from −16 to +5 (sense and AS in Figure 4C, respectively). As shown in Figure 4C, dsRNA of ca. 250 bp in length protected from RNase digestion was detected in Δpho85 cells in which both sense and AS RNAs were present (Figure 4C, second panel from the top, lanes 13, 14, 17, and 18). When the RNA sample was not digested with RNase, the sense strand was successfully amplified, whereas the AS strand was not (lanes 16 and 20), indicating that the KCS1 mRNA was present but the AS RNA was not in the RNA sample tested. In Δpho85 cells in which both sense and AS RNAs were present (Figure 4C, lane 14), dsRNA of ca. 260 bp in length protected from RNase digestion was detected (Figure 4C, second panel from the top, lanes 13, 14, 17, and 18). The Δkcs1 mutant cells do not produce the KCS1 sense or AS RNAs, and accordingly the sense or AS primer failed to synthesize cDNA (Figure 4C, bottom panel). When KCS1mut was expressed in Δkcs1 cells under low-Pi conditions, the sense RNA was produced but not the AS (Figure 4A, lane 4; Figure 4C, the middle panel, lanes 16 and 20), and therefore dsRNA was not detected (lanes 13, 14, 17, and 18). Reciprocally, when the AS RNA was expressed in Δkcs1 cells, only the AS was detected (Figure 4C, second panel from the bottom, lanes 16 and 20), and the protected dsRNA fragments were, if any, below the detectable level (lanes 13, 14, 17, and 18). To confirm that the protected dsRNA was specifically amplified, we carried out the RT reaction using primers hybridizing upstream of the transcription start points of the sense or AS RNAs (at −14 and +235, respectively, in Figure 3E, and those designated by asterisks in Figure 4C). The RNA samples prepared from the strains producing the KCS1 mRNA (wt, Δpho85, and Δkcs1 + KCS1mut) could generate cDNA of ca. 430 bp in length when amplified with the sense*/AS primer pair only in the absence of RNase digestion (Figure 4C, lanes 21 and 23). Similarly, the AS*/sense primer pair could synthesize cDNA when the RNA samples containing the AS RNA were not digested by RNase (lanes 24 and 26). These results indicated that the protected dsRNA fragment was specifically transcribed and amplified by the RT-PCR reaction and therefore strongly suggested that the AS RNA can form a hybrid with the KCS1 mRNA in vivo. Such a hybrid may inhibit normal translation initiation, leading to the generation of the truncated Kcs1 protein.
Although we demonstrated that the truncated Kcs1 protein could be generated independently of the short sense RNAs, it is still possible that the truncated protein is translated from the 2,600 nt intragenic RNA. To test this possibility, we constructed plasmids harboring N-terminally truncated KCS1 or KCS1mut ORF fragments (+105 to +3150) and introduced them into Δkcs1 or Δkcs1 Δpho4 mutants. Because these KCS1 fragments lack the KCS1 promoter, the plasmids are unable to produce the full-length KCS1 mRNA. With a strand-specific RNA probe (Figure 3E), we could detect Pi- and Pho4-dependent transcripts of ca. 2,600 nt in length (Figure 4D, lanes 27, 28, 31, and 32). This transcript was not observed in the KCS1mut (lanes 29, 30, 33, and 34), indicating that the presence of the Pho4 binding sites is required for the production of this RNA species. Thus Pho4 could activate transcription from downstream of its binding site in the KCS1 ORF. Although this downstream transcription was plasmid-borne, the observed similarity in the size of transcript and its regulation strongly indicate that the 2,600 nt transcript is actually transcribed in the chromosomal KCS1 locus in a Pho4-dependent fashion and not a processing (or limited degradation) product. This plasmid-derived transcript could produce protein that had a size similar to that of the truncated Kcs1 protein (Figure 4E, lanes 36 and 39), indicating that the 2,600 nt intragenic RNA can encode the truncated Kcs1 protein. Taken together, these results indicated that Pho4 binding within the KCS1 ORF provokes transcription of both of the AS and the 2,600 nt intragenic RNAs, which may lead to production of the truncated Kcs1 protein by alteration of translation initiation through the formation of a hybrid with the KCS1 mRNA and by translation of the 2,600 nt intragenic transcript.
What is the biological relevance of Pi- and Pho4-dependent production of the AS and intragenic RNAs and consequently of the truncated Kcs1 protein? KCS1 codes for inositol hexakisphosphate (IP6) kinase synthesizing 5-diphospho myoinositol pentaphosphate (5-PP-IP5) [36]. The same substrate is used by another yeast IP6 kinase, Vip1, to synthesize IP7 isomers, 4- or 6-PP-IP5 that function in Pi signaling in the PHO system [37,38]. Therefore, it is conceivable that a decrease in the Kcs1 activity can supply more substrate for Vip1, thereby enhancing Pi signaling. The fact that overproduction of Kcs1 reduces the extent of PHO5 derepression whereas a deletion of KCS1 derepresses PHO5 under high-Pi conditions [39] supports this model. Although the levels of normal Kcs1 protein did not appear to be altered significantly in the presence of the AS RNA and 2,600 nt intragenic transcript (Figure 4A), the formation of a hybrid RNA could affect the normal level of the KCS1 mRNA, which may cause a slight difference in the level of the Kcs1 protein not detectable by western analysis. In addition, the presence of the truncated Kcs1 may perturb normal function of Kcs1. To test these hypotheses, we analyzed the effects of the AS RNA and the intragenic sense transcript on PHO84 and PHO5 expression by northern analysis (Figure 5). PHO84 responds more quickly to a change in Pi conditions than PHO5 [40]. In the wt cells, PHO5 and PHO84 were expressed only under low-Pi conditions (Figure 5, lanes 1 and 2) but not in the absence of Pho4 (lanes 5 and 6). The two genes were expressed under high-Pi conditions in Δkcs1 cells (Figure 5, lanes 7 and 8) as reported [39], which was suppressed by the wt KCS1 in a YCp plasmid (lanes 9 and 10). This result indicated that the plasmid-borne Kcs1 is functional. KCS1mut that produced neither the AS RNA nor the intragenic transcript (Figure 4A) showed a decreased expression level of the two genes under low-Pi conditions (Figure 5, lane 12), suggesting that the low-Pi signal was not transmitted sufficiently to activate Pho4. On the other hand, overexpression of the AS RNA in the wt cells resulted in a significant derepression of the two genes under high-Pi conditions (lanes 1 and 13), suggesting that the low-Pi signal was transmitted to activate Pho4 under high-Pi conditions in this case. This stimulatory function of the AS RNA was dependent on the presence of Vip1 IP6 kinase (lane 19) that functions in low-Pi signal transmission [38]. This result further supported the conclusion that the AS RNA functions in the low-Pi signal transduction pathway. We also overproduced a truncated Kcs1 protein (dKcs1, +670 to +3150) in the wt cells, which caused derepression of PHO84, albeit weakly, and barely detectable expression of PHO5 under high-Pi conditions (lane 15). The different expression levels of PHO84 and PHO5 can be attributed to different responsiveness of the two genes against the change in the environmental Pi level [40]. We also assayed the activity of acid phosphatase encoded by PHO5 in the strains with a combination of various plasmids as tested in northern analysis and found that the levels of the enzyme activities correlated with the mRNA level (unpublished data). These results indicated that the presence or absence of the AS RNA and the 2,600 nt intragenic RNA cause altered regulation of PHO5 and PHO84 responding to Pi conditions, and therefore it is likely that the Kcs1 activity was modulated by the RNAs, the truncated Kcs1 protein, or both. The apparently weak effect on Pi signaling of dKcs1 compared to that of the AS RNA suggests that the truncated protein is not solely responsible for the stimulation of the low-Pi signaling. The AS RNA could play a certain role in this stimulation process, possibly through modulation of the KCS1 mRNA and protein levels. Thus, Pho4 appears to enhance low-Pi signaling by expressing the AS and the intragenic RNAs from within the KCS1 ORF, thereby constituting the positive feedback loop in the Pi signaling pathway.
In this paper, we reported three novel findings derived from the ChIP-on-chip analyses of Pho4 and Rpo21 binding throughout the entire yeast genome: (i) the finding of novel PHO-type genes, (ii) the ability of Pho4 either to activate or to repress transcription independently of environmental Pi conditions, and (iii) the presence of Pi-regulated AS and intragenic RNAs that modulate Pi signal transmission. We demonstrated that 18 genes that had not been classified previously as involved in the PHO system showed Pho4 binding in a Pi-dependent fashion (Table S2 and Figure 1), and at least four of them, viz., MNN1, CBF1, PST1, and PTK2, clearly showed Pho4 binding to their promoters in vivo dependent on Pi conditions and consequently transcription that was dependent on both Pi conditions and Pho4 (Figure 1B and 1C). Harbison et al. reported Pho4 binding profiles under low-Pi conditions [41], and their results share MNN1 and PTK2 out of our 18 novel PHO-type genes. Gonze et al. predicted ARO9 and PST1 as PHO-type genes by computational analysis [42]. KCS1 expression is reported to increase under Pi-limiting conditions [27] and in the absence of Pho85 by microarray analysis [43], probably because of the use of an oligo-DNA array bearing 3′-nested probes that detects Pi- and Pho4-dependent intragenic RNA. Cross-regulation of phosphate and sulfate metabolism has been suggested [44], and in this context, it is noteworthy that we found CBF1, which encodes a transcription factor that regulates MET genes under the control of Pi conditions and Pho4.
Judging from the Gene Ontology terms of these 18 newly recognized PHO-type genes (Saccharomyces Genome Database [SGD], http://www.yeastgenome.org/), they apparently do not have any functional relationship to either Pi metabolism or Pi signaling and are not categorized in a specific functional group (Table S4). Their expression profiles by global analysis, however, showed some similarities in that 12 of them are induced by either nitrogen depletion or amino acid starvation [32,45] and 9 of them are induced in the stationary phase [45]. This raises the possibility that Pho4 is involved in the regulation of a certain set of genes that responds to these nutrient-limiting or stress conditions. Pho4 is also reported to activate the transcription of genes involved in G1 arrest caused by DNA damage [28]. Thus, Pho4 appears to activate the transcription of genes responding to various stress conditions. This notion implies possible cross talk between Pi starvation and other stress conditions, the requirement of function of some, if not all, of these 18 genes in the adaptation of yeast cells to Pi starvation, or both.
The results in this paper imply that Pho4 is present in the nucleus even under high-Pi conditions to activate or repress transcription (Figure 2), an implication that challenges our current understanding of Pho4 regulation. If the current model were correct, then Pho4 should somehow avoid phosphorylation by Pho85, or if phosphorylated, then the modified Pho4 should have much less affinity to the Msn5 exportin to remain in the nucleus. Recently, Zappacosta et al reported Pi-dependent phosphorylation of Pho4 at Ser242 and Ser243 by a kinase other than Pho85 [46]. Phosphorylation of these two sites, however, appears less dependent on Pi than that at those sites modified by Pho85 (i.e., Ser at 100, 114, 128, 152, and 223) [46]. We could imagine that, under high-Pi conditions, prior phosphorylation of the Ser242, Ser243, or both by this unknown kinase could prevent phosphorylation of the other Ser residues by Pho85, thereby decreasing the affinity of Pho4 for Msn5 while increasing the affinity to the target promoter, including ASN1. Alternatively, Pho4 modified at Ser242, Ser243, or both might have more affinity to a yet unknown factor than to the exportin, and the resulting complex might be recruited to the target promoter regardless of phosphorylation by Pho85.
The Pho4 transcription factor appeared to repress CIS3 and YPS3, both cell wall constituents. Expression of CIS3 is repressed by nitrogen starvation and in the stationary phase, implying that Pho4 can function as both an activator and a repressor under these stress conditions. The functioning of a yeast transcription factor as both an activator and a repressor has precedents (e.g., Rap1 and Abf1) [47,48]. Transcriptional repression by the two factors is often accompanied by silent chromatin structure. In a separate paper, we have reported that Pho4 negatively regulates the expression of SNZ1, a stationary phase-specific gene, and that this regulation is accompanied by alterations in chromatin structure evoked by Pho4 binding [49]. The mechanism underlying transcriptional repression of CIS3 and YPS3 by Pho4 is yet to be clarified, but we suppose that a similar mechanism with the SNZ1 may apply in these cases.
We demonstrated the presence of Pi-regulated AS RNA in the KCS1 locus. A large-scale cDNA sequencing by Miura et al. revealed the presence of many AS RNA species [17], including an AS RNA in the KCS1 locus transcribed from +293 to −43. Although its start point is different a little from our result by 5′-RACE analysis (Figure 3E), we think it highly likely that this AS RNA coincides with the Pi-regulated AS RNA that we have reported here. The Pi-regulated AS RNA in KCS1, however, did not appear to be coregulated with the KCS1 mRNA in the wt cells (Figure 3C), and this contrasts with the observation in higher eukaryotes that sense and AS pairs are frequently coregulated [50]. Although Miura et al. did not describe the regulation of the KCS1 AS RNA, they claim coregulation between the sense and AS RNA in the GAL10 locus. However, the fact that the fold induction of the AS RNA is much less compared to that of the GAL10 mRNA when cells are grown in galactose medium and that Gal4 dependency of the GAL10 AS RNA was not analyzed points out that more work is necessary to establish coregulation of the sense and AS RNA at the GAL10 locus.
With respect to biological function of ncRNA in yeast, a noncoding intergenic transcript (SRG1), originating from upstream of SER3 on the same strand and activated by Cha4 transcription factor in the presence of serine [13], inhibits the binding of activators to the SER3 upstream activating sequence and of TATA-binding protein to its TATA box, leading to repression of SER3 [12]. Yeast AS RNA has been reported in the IME4 locus, which is expressed only in the haploid state to inhibit the IME4 mRNA transcription by transcriptional collision and thereby determines cell fate (i.e., the entry into meiosis) [16]. The AS RNA at the PHO5 locus is constitutively expressed at a low level from ca. 1,400 bp downstream of the PHO5 TATA box through its promoter and is proposed to increase chromatin plasticity to enhance histone eviction upon a shift to low-Pi conditions [14]. Those in the PHO84 locus are suggested to recruit and/or stimulate Hda1 histone deacetylase for silencing of PHO84 in aging yeast cells [15]. Although these AS RNA species are found in PHO genes, they have not been reported to be regulated by environmental Pi conditions to facilitate the activation of PHO genes. The KCS1 case presents a different situation from them in that the AS and intragenic RNAs are activated by Pho4 in response to Pi starvation and may modulate the level of Kcs1 IP6 kinase to enhance Pi signaling, thereby stimulating the activation of PHO genes. We observed a decrease in the KCS1 mRNA level when Pho4 binds to the KCS1 ORF under low-Pi conditions or in a Δpho85 mutant (Figure 3C, lanes 2, 5, and 6). This observation suggests a possibility that transcriptional elongation of KCS1 mRNA is inhibited directly by Pho4 binding within the ORF. However, the scenario may not be so simple, because the KCS1 mRNA transcription itself can interfere with Pho4 binding, as reported in the SER3/SRG1 case [12]. Alternatively, the AS RNA could cause transcriptional collision with the mRNA, and hybrid formation of the AS RNA with the mRNA could lead to degradation of the mRNA. Both of these events could lead to a reduction in the KCS1 mRNA level.
The stimulation of low-Pi signaling by Pho4-dependent intragenic and AS RNA represents an autoregulation (induction) or positive feedback loop responding to Pi limitation that can be envisioned as follows. Upon Pi limitation, the low-Pi signal is transmitted to Pho81, leading to inhibition of Pho85-Pho80 and thereby stimulating Pho4 migration into the nucleus [6]. Pho4 then activates transcription of the AS and intragenic RNAs in the KCS1 locus: the AS RNA could reduce the KCS1 mRNA level by hybrid formation and possible transcriptional collision, which can lead to stabilization of Pho4 binding, resulting in the production of more AS and intragenic RNAs and consequently of more truncated Kcs1 protein using the downstream ATG codon at +676. These events could lead to down-regulation of Kcs1 activity, enabling Vip1 IP6 kinase to utilize more IP6 to synthesize 4- or 6-PP-IP5 functioning in low-Pi signaling. Because Kcs1 also can phosphorylate these IP7 species to synthesize 4,5- or 5,6-PP2-IP5 (IP8) [37], reduction of the Kcs1 level can ensure accumulation of the IP7 species to further stimulate low-Pi signaling, leading to complete inhibition of Pho85-Pho80. When Pi becomes sufficient, this loop runs in an opposite way for efficient inactivation of Pho4 and consequent repression of PHO genes. Though putative Pho4 binding sequences are present at −464 and −154 in its promoter, VIP1 expression was dependent on neither Pi condition nor Pho4 (unpublished data), as in the case of the KCS1 mRNA. Inositol polyphosphate (IP) plays an important role in intracellular signal transduction as second messengers. The absence of Kcs1 and Vip1 causes abnormal vacuolar function and cell morphology, respectively, suggesting that they bear important cellular function [37,51]. Therefore, it is reasonable that the genes involved in IP synthesis are not regulated directly by individual nutrients (in this case, Pi) but indirectly by AS and intragenic RNAs responding to the nutrient, so that signal transduction and normal cellular function are not easily perturbed by fluctuation in the status of an individual nutrient.
Positive feedback in the PHO system is also suggested to function in switching of Pi transporters [52], in which Spl2, activated by Pho4, down-regulates low-affinity Pi transporters, Pho87 and Pho90, whereas high-affinity Pi transporter Pho84 is activated by Pho4. When the intracellular Pi level increases via the high-affinity transporter, Pho4 is inactivated to switch the transporters.
Our finding expands the role of the Pho4 transcription factor beyond the regulation of the PHO system. The current consensus view is that it is the master regulator of the genes involved in the system, in that Pho4 activates transcription of the structural genes composing the PHO system to coordinate cellular response to Pi starvation [3]. Our findings indicate that Pho4 can modulate the activity levels of the products of apparently non-PHO genes by activating antisense and intragenic RNA expression to stimulate low-Pi signal transduction. A Δpho85 deletion causes pleiotropic mutant phenotypes [1], some of which could be based on otherwise dormant transcriptional initiation, either intergenic or intragenic, or on the AS strand, caused by hyperactive Pho4. In fact, we have also found Pi- and Pho4-regulated AS RNA in the GTO3 locus and intragenic sense transcript in the SHE9 locus (unpublished data). GTO3 encodes an omega class glutathione-S-transferase having glutaredoxin activity, which is suggested to maintain an adequate redox state of specific target proteins, not in the general defense against oxidative stress [53]. SHE9, also known as MDM33, encodes a mitochondrial inner membrane protein functioning to maintain mitochondrial morphology [54]. Although, at present, we are unable to elucidate whether the GTO3 AS RNA or the SHE9 intragenic transcript can affect the annotated function of the respective gene product, this line of work will lead us to uncover yet unknown protein functions in the cellular response to Pi starvation.
Regulation of gene expression and function by these nonconventional RNA species that are regulated by nutrient signals need not be restricted to the PHO system. Other inducible systems including GAL, glucose repression, and various stresses may well have these RNA species regulated by corresponding signals. High-throughput cDNA sequencing by Miura et al. and other works using microarrays [17–19] have revealed the presence of many intergenic, intragenic, and antisense RNAs in the yeast transcriptome. The finding of nutrient-regulated RNAs that are not coding annotated proteins adds more complexity to the intimately wired transcriptional regulation system by which yeast cells adapt to alterations in environmental conditions. High-resolution mapping of transcription factor and RNA polymerase II binding and very recent development of DNA–RNA hybridization techniques [55] will help to identify these regulatory RNAs. The yeast system, which can be manipulated by an array of genetic tools and for which there exists a substantial body of genetic information, will be the best resource to explore the complexity of the genetic network, including ncRNA species that function in responding to external signals.
Standard yeast genetics and media were used as described [56]. For phosphate-limited medium, Yeast Nitrogen Base (YNB) without phosphate (Q-Biogene) was used instead of normal YNB (SD medium) and was supplemented with 0.2 μM or 2 mM sodium phosphate to make low- and high-Pi media, respectively. The yeast strains used in this work are listed in Table S5.
Standard Escherichia coli and yeast protocols were employed [56,57]. Plasmids and primers used in this work are listed in Tables S5 and S6, respectively. A Δpho85::URA3 fragment [25] was used to disrupt the PHO85 locus of the BY4741 (MFY371) strain, and successful disruption was confirmed by PCR and constitutive expression of acid phosphatase (unpublished data). To disrupt the PHO4 locus, Pho4Δ-F and -R primers were used to amplify the LEU2 marker having PHO4 sequences (from +1 to +100 and from +830 to 929 with A of ATG as +1) at its termini, and the resulting fragment was introduced into MFY371. Successful disruption was confirmed by PCR and failure to express PHO5. For disruption of the VIP1 and RRP6 loci, the adaptamer-mediated PCR method was employed to prepare the DNA fragments for disruption [58]. The detailed methods are described in Text S1. Disruption of GCN4 is described elsewhere [49]. To construct PHO4-tagged strains, MFY376 and MFY377, a fragment containing PHO4 tagged with His × 6 and Flag × 3 was amplified using primers Pho4-Flag-F and -R and pUG6H3Flag plasmid as a template [59], followed by transformation of MFY371 and MFY373, respectively. Rpo21 fragments tagged with His × 6 and Flag × 3 (Rpo21-Flag-F and -R) were used to construct MFY378 and MFY379.
To mutagenize the prospective Pho4 binding site in the ASN1 promoter and in the KCS1 ORF, a QuickChange II site-directed mutagenesis kit (Stratagene) and appropriate primers were used. The detailed methods are described in Text S1. Successful mutagenesis and the whole sequence of the mutant ASN1 promoter and of the mutant KCS1 ORF were confirmed by DNA sequencing. The promoter (−920 to −1) and ORF (−1 to +3143) of KCS1 were amplified by PCR using MN1132/1133 and MN1134/1135 pairs, respectively, so that EcoRI-NcoI and NcoI-XhoI fragments containing the respective sequences were generated. The two fragments were then ligated through the NcoI site and introduced into pRS313 to generate the pMF1530 plasmid. The wt NcoI-BamHI (+1875) fragment had been replaced by the mutant fragment that lacked the three prospective Pho4 binding sites prior to incorporation of the EcoRI-XhoI KCS1 fragment into pRS313 to generate the pMF1531 plasmid. To construct plasmids pMF1527 and pMF1529 producing the wt and mutant Kcs1 protein tagged with six copies of the c-myc epitope at their C-termini, respectively, the EcoRI-XhoI fragment containing the wt or mutant KCS1 sequence was introduced into pRS316 containing a 6 × myc sequence. Plasmids pMF1540 and pMF1560 overexpressing the KCS1 AS RNA were constructed by placing the KpnI-EcoRI (+291 to −920) fragment downstream of the TDH3 and GAL1 promoters in the pRS323 plasmid, respectively. Plasmid pMF1563 overproducing N-terminally truncated Kcs1 protein (dKcs1) was constructed by placing the NcoI-XhoI fragment (+676 to +3150) that had been cloned by PCR downstream of the TDH3 promoter in pRS326.
Yeast cells producing tagged protein were cultivated in high- or low-Pi medium as described above to a cell density of A600 = 1.0–1.2, and chromatin immunoprecipitation was carried out as described [60]. ChIPed fragments were amplified by the T7 RNA polymerase-mediated method (T7RPM) followed by cDNA synthesis and the ligation-mediated PCR (LM-PCR) method for HR and LR analysis, respectively, essentially as described [41,61]. For T7RPM, ChIPed DNA (about 100 ng) was dephosphorylated in the reaction mixture (30 μl) containing 2 units of CIAP and 0.2 units of BAP at 37°C, followed by incubation at 50°C for 15 min each. DNA was purified using a MinElute Reaction CleanUp kit (Qiagen) and eluted from the column with 10 μl of elution buffer (10 mM Tris.HCl, pH8.0). This cleanup method was used throughout the following procedure except for the cleanup of reactions containing RNA. Dephosphorylated DNA was then subjected to poly(dT) tailing reaction in a reaction mixture (20 μl) containing terminal transferase buffer (Roche), 1.25 mM CoCl2, 2.5 μM dNTP, and 20 units of terminal transferase by incubating at 37°C for 15 min. T7A18B primer (GCATTAGCGGCCGCGAAATTAATACGACTCACTATAGGGAG[A]18B, where B refers to C, G, or T) was then annealed to the dT-tailed DNA by incubating at 94°C for 2 min, at 35°C for 2 min, and at 25°C, followed by extension reaction in a 50 μl reaction mixture containing 1 ng/μl tailed DNA, 0.5 mM dNTP, 1 unit of Klenow enzyme, and 5 units of Sequenase at 37°C for 60 min. DNA was then subjected to in vitro transcription using a T7 Megascript kit (Ambion) in 20 μl of reaction mixture at 37°C for 4 h, followed by cleanup with an RNEasy Mini kit (Qiagen). One microliter of 100 μM T7 degenerate primer (GGATCCTAATACGACTCACTATAGGAACAGACCACCNNNNNNNNN) was added to the RNA product, which was incubated at 70°C for 10 min, then on ice for 2 min, followed by cDNA synthesis. About 500 ng of cDNA was subjected to a second round of in vitro transcription and subsequent cDNA synthesis, followed by labeling using an in vitro transcription labeling kit (Affymetrix). Labeled cRNA was hybridized to Affymetrix high-density oligonucleotide arrays of S. cerevisiae whole genome (Watson strand) or of chromosomes 3, 4, 5, and 6, which were processed and analyzed as described [62]. For LM-PCR, phosphorylation of 5′ termini of ChIPed DNA fragments by T4 polynucleotide kinase and ATP was performed prior to the blunt-end reaction, followed by ligation of annealed linkers (MN974 and MN975) at 15°C for 16 h. The resulting fragments were amplified by PCR using MN974 primer, followed by PCR labeling with Cy3-dUTP and Cy5-dUTP for ChIPed and whole cell extract DNA, respectively. The data were analyzed with ChIP Analytics 3.0 software (Agilent).
The procedures for RNA isolation, northern and immunoblot analyses, and assay for β-galactosidase were as described [25,43]. DNA probes were prepared by PCR using digoxigenin (DIG)-PCR labeling mix (Roche). RNA probes were prepared by transcribing DNA fragment cloned in pSP72 or pSP73 (Promega) with T7 RNA polymerase and a DIG-RNA labeling mix (Roche). Gene-specific PCR was performed using primers listed in Table S6 and ChIPed DNA fragments or DNA in the whole cell extract (WCE) fraction as template under cycling condition as described [49]. A 5′-Full-RACE kit (TaKaRa) was used to determine the transcription start points in the KCS1 locus with total RNA from Δpho85 cells grown in high-Pi medium. Phosphorylated primers MN915 and MN1190 were used as a reverse transcription primer for the start points upstream of the initiating codon and within the ORF, respectively, and MN1134 for the start point of AS RNA. Amplified fragments were cloned using a TOPO TA cloning kit (Invitrogen) according to the manufacturer's protocol, and the transcription start points were determined by DNA sequencing. RNase protection assay and RT-PCR were carried out essentially as described [57]. Total RNA was digested with RNase ONE (Promega) at 30°C for 1 h and was recovered by precipitation in the presence of ethanol. First-strand cDNA was then synthesized using a primer specific to the sense or antisense strand, followed by PCR amplification after inactivation of reverse transcriptase and addition of appropriate reverse primer. The PCR products were separated by electrophoresis on a 5% polyacrylamide gel.
|
10.1371/journal.pntd.0001490 | High-Resolution Genotyping of the Endemic Salmonella Typhi Population during a Vi (Typhoid) Vaccination Trial in Kolkata | Typhoid fever, caused by Salmonella enterica serovar Typhi (S. Typhi), is a major health problem especially in developing countries. Vaccines against typhoid are commonly used by travelers but less so by residents of endemic areas.
We used single nucleotide polymorphism (SNP) typing to investigate the population structure of 372 S. Typhi isolated during a typhoid disease burden study and Vi vaccine trial in Kolkata, India. Approximately sixty thousand people were enrolled for fever surveillance for 19 months prior to, and 24 months following, Vi vaccination of one third of the study population (May 2003–December 2006, vaccinations given December 2004).
A diverse S. Typhi population was detected, including 21 haplotypes. The most common were of the H58 haplogroup (69%), which included all multidrug resistant isolates (defined as resistance to chloramphenicol, ampicillin and co-trimoxazole). Quinolone resistance was particularly high among H58-G isolates (97% Nalidixic acid resistant, 30% with reduced susceptibility to ciprofloxacin). Multiple typhoid fever episodes were detected in 22 households, however household clustering was not associated with specific S. Typhi haplotypes.
Typhoid fever in Kolkata is caused by a diverse population of S. Typhi, however H58 haplotypes dominate and are associated with multidrug and quinolone resistance. Vi vaccination did not obviously impact on the haplotype population structure of the S. Typhi circulating during the study period.
| Typhoid fever is caused by the bacterium Salmonella enterica serovar Typhi (S. Typhi) and is a major health problem especially in developing countries. Vaccines against typhoid are commonly used by travelers but less so by residents of endemic areas. We used single nucleotide polymorphism (SNP) typing to investigate the population structure of 372 S. Typhi bacteria isolated from typhoid patients during a typhoid disease burden study and Vi anti-typhoid vaccine trial in Kolkata, India. Approximately sixty thousand people were enrolled for fever surveillance for 19 months prior to, and 24 months following, vaccination of one third of the study population against typhoid (May 2003–December 2006, vaccinations given December 2004). We detected a diverse population of S. Typhi, including 21 different genetic forms (haplotypes) of the bacteria. The most common (69%) were of a haplogroup known as H58, which included all multidrug resistant isolates (bacteria resistant to the antibiotics chloramphenicol, ampicillin and co-trimoxazole). Resistance to quinolones, a class of antibiotics commonly used to treat typhoid fever, was particularly high among a subgroup of H58 (H58-G). Vi vaccination did not obviously impact on the haplotype distribution of the S. Typhi circulating during the study period.
| Salmonella enterica serovar Typhi (S. Typhi) is the bacterium responsible for typhoid fever, which affects more than 20 million people each year, resulting in over 200,000 deaths [1], [2]. As S. Typhi is transmitted by the fecal-oral route, the typhoid fever burden falls almost exclusively in developing areas where sanitation is poor [1], [3]. The current mainstay of typhoid fever treatment is antimicrobial therapy [4], however resistance to antimicrobials is common among S. Typhi [5], leading to prolonged bacterial clearance times and treatment failure [6], [7]. Children and young adults are the most vulnerable population for developing typhoid fever [1], [8], [9] and can be protected by vaccination against S. Typhi [10], [11]. However while vaccines against S. Typhi are frequently used by travelers to typhoid endemic areas [12], they are yet to be effectively harnessed for the protection of local, typhoid endemic populations [13].
S. Typhi is a highly clonal bacterium estimated to have entered the human population on a single occasion approximately 50,000 years ago [14]. We have recently identified hundreds of single nucleotide polymorphisms (SNPs) within the S. Typhi chromosome that are suitable for rapidly and informatively subtyping S. Typhi populations [15], [16]. As recombination is rare in S. Typhi, SNP typing allows individual S. Typhi isolates to be assigned unequivocally to unique haplotypes. Importantly, as haplotypes are defined by phylogenetically informative sequence variation, SNP typing also reveals information about genome sequence and the evolutionary relationship between isolates [15], [16]. As our SNP panel is designed to allow inference of phylogenetic relationships, it does not target SNPs that are likely to be under selection, such as drug resistance loci. SNP haplotyping studies in localized areas where typhoid is endemic, including Jakarta [17], Kathmandu [18], [19], the Mekong Delta [20] and Nairobi [21], have revealed that the typhoid burden in endemic areas is usually attributable to a diverse population of differentiable S. Typhi haplotypes, co-circulating within the local human population. These studies also revealed the clonal expansion of a S. Typhi haplogroup, H58, in South East Asia [16], [18], [20], as well as in Nairobi [21].
During 2003–2004, a typhoid burden study was conducted in a typhoid endemic area of Kolkata, India [8], [22], [23]. This was followed by a large-scale, cluster-randomized phase IV trial to determine the efficacy of the injectable Vi polysaccharide vaccine (Typherix, GlaxoSmithKline) among the local population (>60,000 persons). The study site was divided into 80 geographic clusters (40 clusters each randomly assigned to Vi vaccine or inactivated hepatitis A vaccine as a control) and in December 2004, eligible residents were vaccinated (mean 60% of the population vaccinated in each cluster) [11]. The primary results of the trial, namely 61% efficacy among vaccinees and indirect protection within and around Vi vaccinated geographic clusters, have been published elsewhere [11], [24]. Surveillance for fever was conducted uninterrupted throughout May 2003–December 2006, and typhoid fever was confirmed by positive blood culture of S. Typhi [8], [11], [22], [23]. A total of 372 typhoid cases were confirmed by blood culture during the study period, including 197 during the post-vaccination period. All S. Typhi isolates produced Vi during in vitro culture [11], however Vi expression is tightly regulated in S. Typhi growing on laboratory media and in vivo [25], [26], [27] and we consequently hypothesised that selection against Vi expression in Vi immunized individuals might result in differential efficacy of Vi vaccine against different S. Typhi phylogenetic lineages. Here we present an analysis of the 372 S. Typhi isolates collected during the study period, including SNP haplotyping, antimicrobial susceptibility profiling, analysis of intra-household transmission and determination of Vi vaccine efficacy for the most common circulating haplotypes.
A total of 372 S. Typhi were isolated during the typhoid disease burden study from May 2003 to December 2006, intervened by a Vi effectiveness trial (December 2004), conducted in Kolkata, India [8], [11], [22], [23]. S. Typhi were isolated from blood cultures of fever patients following standard techniques [28]. The institutional review boards at the International Vaccine Institute, the National Institute of Cholera and Enteric Diseases, and the Indian Council of Medical Research approved the protocol and monitored the progress of the studies. All subjects provided written informed consent for vaccination and oral informed consent for blood culture (for children, informed consent was provided by their guardian). The assayed isolates represent all confirmed typhoid cases during the study period May 2003 to December 2006, among subjects who were present in the field area at baseline, including 10 cases in non-vaccinees that were not included in the original vaccine report due to incomplete demographic data [11]. Confirmation of S. Typhi was done by agglutination with poly and monovalent antisera (BD diagnostics, US), Vi phenotype was checked by agglutination with monovalent Vi antisera.
Testing was performed using Kirby Baure's disc diffusion method using 11 antimicrobial discs from BD diagnostics (ampicillin, tetracycline, chloramphenicol, cotrimoxazole, nalidixic acid, ciprofloxacin, ofloxacin, ceftriaxone, amikacin, aztreonam, amoxicillin-clavulanic acid). MICs of antimicrobials were determined by E-test (AB Biodisk, Solna, Sweden) and interpreted following CLSI guidelines [29]. Multidrug resistance (MDR) was defined as simultaneous resistance to chloramphenicol (MIC>256 µg/mL), ampicillin (MIC>256 µg/mL) and co-trimoxazole (MIC>32 µg/mL).
DNA extraction was carried out from overnight LB culture of S. Typhi isolates using Promega DNA extraction kit following manufacturer's instructions. DNA samples were quantified using the Quant-IT kit (Qiagen, USA) and concentrations adjusted to 10 ng/µl using nuclease-free water (Ambicon, USA). SNP typing was performed using either GoldenGate or Sequenom assays (loci in Table S1). The former was performed using a GoldenGate custom array according to the manufacturer's standard protocols (Illumina, USA), covering 1,500 loci (Table S1) as described previously [18], [20], [21]. Briefly, DNA samples were arrayed in a 96-well plate along with a negative control (water) and positive control (sequenced Typhi), assayed using two custom oligo pools (200 SNPs included on both arrays for quality control) using the Illumina GoldenGate platform and analyzed using Illuminus-P [21]. Sequenom assays of 100 loci (Table S1) were performed using the iPLEX Gold assay (Sequenom Inc, USA), designed using the MassARRAY Assay Design software version 3.1 (Sequenom Inc, USA) as previously described [19]. Samples were amplified in multiplexed PCR reactions before allele specific extension. Allelic discrimination was obtained by analysis with a MassARRAY Analyzer Compact mass spectrometer. Genotypes were automatically assigned and manually confirmed using MassArray TyperAnalyzer software version 4.0 (Sequenom Inc, USA).
Phylogenetic analysis (Figure 1) was based on 81 SNPs common to both GoldenGate and Sequenom assays (Table S1), which include those dividing isolates into 48 major haplotypes (original defined in [16]) and further subdivision of the H58 haplogroup into subtypes (originally defined in [15]). Each isolate was assigned to a node in the previously defined S. Typhi phylogenetic tree based on alleles at these 81 SNP loci. Statistical analysis was performed in R [30]. Haplotype-specific typhoid isolation rates in Vi vaccinees vs hepatitis A vaccinees (Table 1) were compared using Fisher's exact test (two-tailed test).
All 372 S. Typhi isolates collected between May 2003 and December 2006 were subjected to SNP haplotyping using high-throughput Sequenom or Illumina GoldenGate platforms (Table S1). These two genotyping methods have been applied previously to study S. Typhi populations [17], [18], [19], [20]. Forty-five of the assayed loci were discovered by mutation analysis of 200 gene fragments within a global collection of S. Typhi [16] and provide medium-level resolution of the S. Typhi population, subdividing it into 48 distinct haplotypes (displayed as a phylogenetic tree in Figure 1). Eleven of these haplotypes, which are broadly distributed across the tree, were identified among the Kolkata S. Typhi (Figure 1, excluding shaded area). The globally dominant haplotype H58 was by far the most common (N = 260, 70%), followed by H42 (N = 65, 17%) and H14 (N = 25, 7%) (Figure 1). We assayed 50 additional SNP loci, discovered by whole genome sequence analysis of seven globally distributed S. Typhi H58 isolates [15], that provide greater resolution within the H58 haplogroup and subdivide it into 20 distinct subtypes (Figure 1, shaded area). Eleven H58 subtypes were identified among the Kolkata S. Typhi (Figure 1), however 97% of H58 isolates belonged to just four H58 subtypes: B (N = 148, 40% of all S. Typhi tested), G (N = 66, 18%), A (N = 22, 6%) and H64 (N = 17, 5%).
Resistance to the quinolone Nalidixic acid (Nal) was common (54% of all isolates), with Nal resistance observed among phylogenetically unlinked haplotypes (Table 2), indicating that Nal resistance arises frequently within distinct S. Typhi chromosomal backgrounds. Each common haplotype included isolates that were Nal resistant but susceptible to ciprofloxacin, as well as isolates that were Nal resistant and exhibiting reduced susceptibility to ciprofloxacin (MIC≥0.125 µg/mL) (Table 2). Interestingly the most common haplotype, H58-B, exhibited low rates of Nal resistance, with only 24% of H58-B isolates displaying resistance to Nal (significantly lower than other H58 (94% resistant), p<10−8 using Fisher's exact test). The highest rate of Nal resistance was observed among the second-most common haplotype, H58-G, with 97% of isolates resistant to Nal and 31% also exhibiting reduced susceptibility to ciprofloxacin (Table 2). Two isolates were ciprofloxacin resistant (MIC≥16 µg/mL) and have been described in detail elsewhere [31]. These isolates were of identical haplotype, H58-I1 (see Figure 1) and isolated from siblings (aged 3 and 5 years) on the same day in July 2004 [31]. No other isolates of this haplotype were detected during the study (2003–2006). Multiple drug resistance (MDR, defined as resistance to chloramphenicol, ampicillin and co-trimoxazole) was observed in 43 S. Typhi isolates (11.5%), of which most (N = 38) were also Nal resistant. The MDR S. Typhi isolates belonged to five H58 subtypes: H58-A (5 isolates), H58-B (12 isolates), H58-G (10 isolates), H64 (a sub-type of H58) (15 isolates) and H58-I4 (1 isolate). These subtypes occupy the internal nodes of the H58 phylogeny, including members of both major lineages (see Figure 1), indicating that MDR is widely distributed among the H58 haplogroup.
The incidence of typhoid fever remained high throughout the four-year study period, with a median of seven cases per month and no clear seasonal pattern (Figure 2). A total of 168 S. Typhi were isolated during May 2003–November 2004 (19 month pre-vaccination period), 7 during December 2004 (vaccination period) and 197 during January 2005–December 2006 (24 month post-vaccination period). The same haplotypes dominated throughout the study (Figure 2), indicating that the burden of typhoid fever in Kolkata was the result of a diverse range of co-circulating haplotypes. One exception to this pattern was a peak in typhoid cases in November 2005 involving 27 infections, of which 21 were S. Typhi H58-B consistent with a small outbreak (Figure 2), although no spatial clustering was evident. Only 2 of the 21 H58-B cases in this month occurred in clusters assigned to Vi vaccine, suggesting the vaccine was effective in providing protection during the outbreak (Figure 2).
As previously reported, the incidence of typhoid fever during the two years following vaccination was >60% lower among individuals who received the Vi typhoid vaccine than those who received hepatitis A vaccine (Table 1) [11]. Our haplotype data indicates this overall reduction was due to a statistically significant reduction in isolation rate across all S. Typhi haplotypes (H58, H42 and others, see Table 1). All S. Typhi isolated during the study, including those from individuals who had been vaccinated with Vi (‘breakthrough cases’), reacted strongly with commercially available Vi antisera (BD diagnostics, USA) in an agglutination test, indicating that all strains could express the vaccine target Vi. The 34 S. Typhi isolates from breakthrough cases belonged to several distinct SNP haplotypes and were also diverse in terms of antimicrobial resistance (Table 3).
There were 22 households from which multiple S. Typhi were isolated by blood culture (21 households with 2 positive cultures; 1 household with 3 positive cultures, total 45 positive cultures; Table 4). For three of these households, the paired isolates resulted from two blood cultures from the same individual, taken 3–5 weeks apart and thus representing possible cases of relapse or re-infection. Each of these isolate pairs displayed identical S. Typhi haplotypes and resistance phenotypes, consistent with relapse as opposed to re-infection with a distinct haplotype (Table 4). However different S. Typhi haplotypes were involved in each pair of these relapse cases, and displayed different antimicrobial resistance profiles (H14, NalR; H58-G, MDR; H64, NalR+MDR).
In the remaining 19 households with multiple cases, S. Typhi was isolated from different individuals, thus representing distinct typhoid cases within the same household. Among these, twelve households had more than one typhoid case occurring within two months (Table 4). In nearly all of these households, the same S. Typhi haplotype (displaying same resistance phenotype) was isolated from both cases, consistent with direct transmission between household members or a shared environmental source such as food or water (10/12 households, Table 4, p = 0.039 using Binomial test with equal probability of same or different haplotypes). Among households in which a second typhoid case occurred more than two months after the first, the later infection was most often caused by a distinct S. Typhi haplotype (5/7 households, Table 4). To examine whether Vi vaccination reduced intra-household transmission, we compared the proportion of cases for which apparent transmission was observed in the same household (defined as the same S. Typhi haplotype isolated from another member of the household one week to two months after the initial case), among Hepatitis A and Vi clusters in the post-vaccination period. While it is possible that two typhoid cases caused by the same haplotype in the same household could result from shared exposure to a common source of S. Typhi, it is more likely that infections separated by more than a week constitute transmission events. Using this definition of household transmission, six percent (8/135) of cases in the Hepatitis A cluster were linked to putative transmission within a household, while none of the 59 cases in the Vi cluster were obviously linked to transmission. While the numbers are low, this provides weak evidence for protection against person-to-person or direct transmission by the Vi vaccine (p = 0.045 using Fisher's exact test), which may be via direct protection of vaccinated individuals and/or indirect protection via herd immunity in clusters assigned to the Vi vaccine. The distribution of haplotypes among these likely transmission events was no different to that of haplotypes among all Hepatitis A clusters during the post-vaccination period (5 cases, 8% for H58-B; 1 case, 6% for H580-G; 1 case, 4% for H42; p = 1 using Fisher's exact test). Thus there is no evidence that any particular haplotype is more likely to be transmitted person-to-person.
SNP typing of S. Typhi isolated during 2003–2006 revealed a diverse range of haplotypes co-circulating in the study site, an urban slum area in eastern Kolkata. A similar level of diversity has been observed in previous studies in typhoid endemic areas [17], . The dominant S. Typhi haplotypes were subtypes of H58, collectively accounting for 70% of all S. Typhi isolated during the four-year study (Figure 1). The dominance of H58 has been reported in recent studies of S. Typhi infections in other typhoid endemic areas including Kathmandu, Nepal (69% H58, 2003–2004) [18], the Mekong Delta, Vietnam (98% H58, 2004–2005) [20] and Nairobi, Kenya (87% H58, 2001–2008) [21]. However there does appear to be greater diversity within the H58 group in Kolkata. We identified 11 distinct H58 haplotypes, including four with high frequency among S. Typhi from Kolkata (18–150 isolates each) including the ancestral node (A) and nodes from both major lineages of H58 (Figure 1). In neighbouring Nepal, two hospital-based studies of S. Typhi found 61–69% of isolates belonged to a single subtype of H58 lineage II, H58-G, and few other H58 isolates were detected [18], [19]. In the Mekong Delta, Vietnam, a large hospital-based study found 95% of S. Typhi isolated from adults and children with typhoid fever belonged to one of three closely related H58 lineage I subtypes, H58-C, -E1 and -E2 (see Figure 1) [20]. In that study, differentiation of the three subtypes was possible because the genome of an isolate from the study had been sequenced for the purpose of SNP discovery [15].
The H58 subtypes that were common in the present study in Kolkata are internal nodes of the H58 phylogenetic tree described by the assayed SNPs (A, B, G, H64, see Figure 1). This is not particularly surprising, since SNP discovery for our assays did not include analysis of any Kolkata strains, with the exception of two isolates of the H64 haplotype (actually part of the H58 group, see Figure 1) which were included in mutation detection within 200 gene fragments [16] but not at a genome-wide scale [15]. Since SNPs accumulate locally over time as bacteria replicate, we would expect that there is more diversity in the Kolkata S. Typhi population than we are able to detect in our SNP assays (i.e. mutations have occurred locally at genomic positions that we did not assay). If more Kolkata isolates had been included in SNP discovery, we would be able to differentiate among Kolkata isolates at higher resolution. This is known as SNP ascertainment bias [32], and implies that diversity which has accumulated in the local S. Typhi H58 population of Kolkata in the last decade or so is being collapsed into just a few haplotypes using our SNP typing method. Despite this, the fact that both major H58 lineages and the ancestral node were detected at high frequency in Kolkata indicates that H58 S. Typhi has been present in this location for some decades. This is similar to the pattern observed in Kenya where both H58 lineages have been observed at high frequency [21], but quite unlike Vietnam or Nepal where lineage I or II dominated, respectively.
In this study, as in others, antimicrobial resistance was frequent among H58 S. Typhi. All MDR isolates were from the H58 group, similar to recent observations in Vietnam [20], Kenya [21] and global collections [33]. Nal resistance was frequent among all H58 subtypes except H58-B (25% of H58-B; 95% of all other H58, see Table 2), although it was also frequent among common non-H58 haplotypes H42 (46%) and H14 (76%). Interestingly, S. Typhi H42 was also common in the Nepal study (19% of S. Typhi), yet Nal resistance in that location was observed only among H58-G isolates. Taken together, these observations suggest that while MDR is now largely restricted to H58 S. Typhi, Nal resistance arises frequently in S. Typhi of a diverse range of haplotypes.
Several households experienced more than one typhoid infection during the study. Among multiple cases occurring in the same household within a 2-month period, nearly all (10/12) were caused by identical infecting S. Typhi haplotypes, consistent with intra-household transmission or a common source (Table 4). In the post-vaccination period, eight such putative transmission events were detected in control clusters (assigned to Hepatitis A vaccine) and none were observed in Vi vaccine clusters, possibly reflecting protection via vaccination and/or herd immunity in these clusters. Three cases of relapse were identified (two infections with the same haplotype in a single individual). Although each pair of relapse isolates had an identical haplotype the haplotype was different in each individual, although all exhibited some form of antimicrobial resistance (NalR and/or MDR), suggesting that relapse may be associated with antimicrobial failure.
In addition to providing a snapshot of the S. Typhi population circulating in a localized region of Kolkata, this study offers the first insight into the impact of the introduction of Vi typhoid vaccine upon a local S. Typhi population. Our data indicate the incidence of all haplotypes of S. Typhi was similarly reduced among Vi vaccinated individuals (Table 1, Figure 2). S. Typhi isolated from Vi vaccinated individuals included several distinct haplotypes, which could be further differentiated by antimicrobial resistance phenotypes (Table 3). All S. Typhi isolates expressed Vi during laboratory culture. Thus, it is likely that ‘breakthrough’ cases of typhoid fever among vaccinees is due to subtle variations in the regulation of Vi expression in vivo and/or to host factors, and not to lineage-associated differences in Vi expression.
The S. Typhi population responsible for typhoid fever in Kolkata is genetically and phenotypically diverse, displaying a wide range of haplotypes and antimicrobial susceptibility phenotypes. However the H58 haplotype dominates, and is responsible for the majority of MDR and quinolone resistant S. Typhi infections. The Vi polysaccharide vaccine was effective against infections with all S. Typhi haplotypes.
|
10.1371/journal.ppat.1001028 | A Limited Number of Antibody Specificities Mediate Broad and Potent Serum Neutralization in Selected HIV-1 Infected Individuals | A protective vaccine against HIV-1 will likely require the elicitation of a broadly neutralizing antibody (bNAb) response. Although the development of an immunogen that elicits such antibodies remains elusive, a proportion of HIV-1 infected individuals evolve broadly neutralizing serum responses over time, demonstrating that the human immune system can recognize and generate NAbs to conserved epitopes on the virus. Understanding the specificities that mediate broad neutralization will provide insight into which epitopes should be targeted for immunogen design and aid in the isolation of broadly neutralizing monoclonal antibodies from these donors. Here, we have used a number of new and established technologies to map the bNAb specificities in the sera of 19 donors who exhibit among the most potent cross-clade serum neutralizing activities observed to date. The results suggest that broad and potent serum neutralization arises in most donors through a limited number of specificities (1–2 per donor). The major targets recognized are an epitope defined by the bNAbs PG9 and PG16 that is associated with conserved regions of the V1, V2 and V3 loops, an epitope overlapping the CD4 binding site and possibly the coreceptor binding site, an epitope sensitive to a loss of the glycan at N332 and distinct from that recognized by the bNAb 2G12 and an epitope sensitive to an I165A substitution. In approximately half of the donors, key N-linked glycans were critical for expression of the epitopes recognized by the bNAb specificities in the sera.
| The development of an immunogen that elicits antibodies that neutralize a wide range of global circulating HIV-1 isolates is a major goal of HIV-1 vaccine research. Unfortunately, even the most promising antibody-based vaccine candidates have only induced NAb responses that neutralize a limited number of these strains. However, recent studies have demonstrated that broad and potent NAb responses develop in the sera of a subset of HIV-1 infected individuals, and studying the nature of these responses may provide clues for the design of new vaccine immunogens. Here, we show that the broad neutralization in the sera of most of the individual donors that we studied can be associated with single or a small number of specificities. Across the donor panel, broad neutralization appears associated with 4–5 principal specificities.
| The hallmark of most successful anti-viral vaccines is the ability to induce neutralizing antibodies [1], [2], [3], [4]. For HIV-1, NAbs have been shown to provide protection against viral challenge in non-human primate models [5], [6], [7], [8], [9], [10], [11], [12], [13], suggesting that a vaccine capable of inducing similar types of antibodies would provide benefit upon exposure to the virus. However, due to the extraordinary genetic diversity of the HIV-1, a successful vaccine will require the induction of antibodies that neutralize a wide spectrum of global circulating viral isolates, i.e. broadly neutralizing antibodies (bNAbs) [14]. Unfortunately, the development of an immunogen capable of eliciting bNAbs has not been met with success to date. Importantly, although NAbs generated during natural HIV-1 infection usually target immunodominant variable regions of the virus, recent studies have shown that 10–30% of infected individuals develop moderate to broadly neutralizing sera [15], [16], [17], [18]. These individuals are of considerable interest from a vaccine standpoint; understanding the antibody specificities that mediate potent cross-clade serum neutralizing activity may illuminate potential targets for HIV-1 immunogen design. In addition, knowledge of the epitopes targeted by the bNAbs can assist in the design of reagents, “baits”, to facilitate the isolation of broadly neutralizing monoclonal antibodies (bnMAbs) from these donors. BnMAbs can be used in molecular studies to help direct vaccine design [19], [20], [21].
Several studies have previously been performed to systematically analyze the NAb specificities in HIV-1 positive sera displaying varying degrees of neutralization breadth and potency [15], [16], [17], [18], [22], [23], [24]. In all of these studies, a series of complementary methods, such as selective removal of certain antibody specificities using antigen-coated beads, inhibition of neutralizing activity using linear peptides, and the use of chimeric viruses displaying specific epitopes, were used to define the epitopes targeted by NAbs in broadly neutralizing sera. Although the breadth of serum neutralization could rarely be mapped exclusively to a single epitope, several sera appeared to contain CD4bs and co-receptor binding site (CRbs)-specific antibodies that contributed to the overall breadth of serum neutralizing activity [16], [17], [22], [25]. In a minority of cases, sera were found to contain NAbs to the membrane-proximal external region (MPER) [16], [17], [23]. Arguably, one of the most significant results from these studies was that a substantial fraction of the serum NAbs appeared to target unidentified viral epitopes. Considering that most of the reagents used for characterization were based on monomeric gp120 and linear peptides, one possibility here is that the serum neutralization breadth is mediated by NAbs that recognize quaternary epitopes preferentially expressed on trimeric Env. Two recently described broad and potent NAbs, PG9 and PG16, fall into this category [26].
An important question that has arisen from serum studies concerns the number of NAb specificities that mediate broad serum neutralization. A few scenarios are possible; broad serum neutralization could be mediated by a very large number of neutralizing antibodies with limited breadth [27], a few relatively broad and potent neutralizing antibody specificities, or a single, extraordinarily broad and potent, neutralizing antibody specificity. Although these scenarios are not mutually exclusive, the latter two are more attractive in terms of vaccine design, as it appears far more practical to focus the immune response on a small number of conserved epitopes with a vaccine rather than on a large number of more variable epitopes.
In a previous study, we screened sera from approximately 1,800 HIV-1 infected donors from Thailand, Australia, the United Kingdom, the United States, and several sub-Saharan African countries for neutralizing activity and identified donors who exhibit among the most broad and potent neutralizing serum activity observed to date [15]. The top 1% of samples screened, designated “elite neutralizers”, displayed particularly potent serum neutralizing activity against a cross-clade pseudovirus panel. These donors are valuable for understanding the development of broad responses and for the isolation of broad and potent neutralizing monoclonal antibodies. Notably, PG9 and PG16 were isolated from an individual who ranked in the top 5% of donors screened [26]. In this study, we have used a number of established and new techniques to map the broadly neutralizing antibody specificities in the serum of individuals who were ranked in the top 5% of neutralizers identified in our previous study, including elite neutralizers. Importantly, since many of our approaches rely on the use of functional assays, we defined the epitopes recognized by the broadly neutralizing serum antibodies in the context of the native trimer. Our results demonstrate that the broad neutralization in the sera of most of the individual donors can be associated with single or a small number of specificities. Across the donor panel, broad neutralization appears associated with 4–5 principal specificities.
A total of 19 volunteers from diverse HIV-1 epidemics were characterized. Of the 19 volunteers, 63% were female, 26% male and 11% unknown. The median age of all volunteers was 38 with a median CD4 count of 414 and median log viral load of 4.07. All volunteers were infected for at least 3 years. Of the 19 donors analyzed, 14 ranked in the top 1% of neutralizers identified in our previous report (elite neutralizers), and the remaining 5 ranked in the top 5% of neutralizers. Figure S1 shows the serum neutralization profiles of the selected donors. Serum neutralization was assessed using an indicator cross-clade pseudovirus panel that has previously been shown to be predictive of neutralization breadth and potency over a larger number of isolates [15].
As a first approach, we used a previously described serum adsorption method to determine whether the NAb specificities in these sera would react with recombinant monomeric gp120 [17]. Eighteen sera, which all neutralized HIV-1 YU2, were adsorbed with recombinant YU2 gp120 coupled beads or blank control beads (donor #37 was excluded from the analysis because the plasma did not neutralize YU2). After confirming depletion efficiencies by ELISAs, which showed that all detectable gp120-binding antibodies had been removed (Figure S2), the adsorbed fractions were tested for neutralizing activity against a cross-clade pseudovirus panel (Figure 1A). For ten donors (#74, #36, #20, #51, #26, #33, #57, #17, #14, and #23), most of the broad serum neutralizing activity was removed after gp120 adsorption, indicating that the serum neutralization breadth could be attributed to gp120-reactive NAbs. In contrast, for the remaining eight donors (#21, #30, #39, #15, #31, #24, #56, #29), a large proportion of the broad neutralizing serum activity was retained after removal of the YU2 gp120-specific Abs, suggesting the presence of NAbs that recognize epitopes that are not expressed on recombinant YU2 gp120. Based on these results, we next sought to determine whether the NAb specificities in the serum would react with a recombinant trimerized Env protein. The YU2 gp140-foldon trimer was chosen for these studies because it has been well characterized structurally and antigenically [28], [29], [30], [31], [32], [33], and it has been previously used for the isolation of NAbs from HIV-1 infected patients [27]. Using the same method as for gp120, we adsorbed the sera with YU2 gp140-foldon coupled beads or blank control beads (Figure 1B). As expected, we found that, if the broad neutralizing activity of a particular serum could be adsorbed with YU2 gp120, it could also be adsorbed with YU2 gp140. However, for a subset of donors (#39, #21, and #15), the broad neutralizing activity of the sera was absorbed more efficiently with YU2 gp140 than YU2 gp120. This result may suggest that certain gp120 epitopes are better presented on the YU2 gp140-foldon trimer. Alternatively, a significant fraction of the serum neutralization breadth in these donors could be mediated by NAbs directed against gp41.
We next investigated the contribution of gp41-directed NAbs to broad serum neutralizing activity. Since the MPER region of gp41 contains the epitopes recognized by three broadly neutralizing monoclonal antibodies, 2F5, 4E10, and Z13e1, and is the only known neutralizing determinant on gp41, we focused on determining whether NAbs directed against this region were mediating broad serum neutralizing activity. As a first step, we tested the sera for neutralizing activity against a chimeric HIV-2 virus containing the complete MPER region of gp41 [24]. Based on this assay, six sera appeared to contain MPER-reactive NAbs (Figure 2A). One of these donors (#36) also neutralized the parental HIV-2 virus, indicating that this donor may be co-infected with HIV-2 or contain anti-HIV-1 NAbs that cross-react with HIV-2. Notably, since HIV-2/MPER chimeras are 1 to 2-logs more sensitive to NAbs 4E10 and Z13e1 than HIV-1 primary isolates [22], this assay may overestimate the contribution of anti-MPER antibodies to serum neutralization breadth and potency. Indeed, for all six of these donors, a substantial fraction of the broad serum neutralization could be adsorbed with monomeric gp120, demonstrating that anti-MPER NAbs probably do not dominate the overall serum neutralization breadth and potency. Nonetheless, to further investigate the contribution of anti-MPER NAbs to the serum neutralization breadth in these six donors, we adsorbed the sera with MPER-coupled beads or blank control beads, as described previously [23]. As above, the adsorbed sera were then tested for neutralizing activity against a cross-clade pseudovirus panel (Figure 2B). In addition, the MPER-specific antibodies were eluted off the beads and tested for neutralizing activity against the clade B isolate JR-CSF (Figure 2C). As expected, for all six donors, the neutralizing activity of the serum after adsorption with MPER peptide-coated beads or blank beads were comparable, indicating that anti-MPER NAbs do not mediate a major fraction of the serum neutralization breadth and potency. However, for three donors, weak to moderate neutralizing activity against JR-CSF was observed in the fraction eluted from the MPER peptide-coupled beads, suggesting the presence of MPER-directed NAbs at low concentrations and/or low neutralizing potency in these sera.
Previous serum mapping studies have evaluated CD4bs and CRbs-directed neutralizing activity in sera by performing serum adsorptions with gp120 point mutants that fail to react with existing mAbs directed against these epitopes. However, a caveat to this approach is that some CD4bs or CRbs NAbs may be insensitive or only partially sensitive to these particular mutations. Notably, a recently reported broadly neutralizing CD4bs-directed NAb binds to the D368R gp120 variant, often considered a prototypic non-CD4bs Ab binding gp120, with higher affinity than the wild-type (WT) gp120 molecule [34]. Therefore, as an alternative to use of the D368R gp120 variant, we developed a serum adsorption method based on antibody competition. Using this method, serum adsorptions to Env are performed in the presence of saturating concentrations of a non-neutralizing competitor mAb (the competitor Ab must be non-neutralizing so its presence will not affect the results of the neutralization assay). In principle, Abs directed against epitopes overlapping that of the competitor Ab will fail to bind to the Env-coated beads. Since the non-neutralizing CD4bs-directed mAb b6 has been shown to compete with both CD4bs and CRbs-directed Abs for binding to gp120 [35], b6 was used as a competitor in these experiments.
To first validate this method, we adsorbed b12 (a CD4bs-directed bNAb) with YU2 gp140-coated beads in the presence or absence of saturating concentrations of b6 or blank control beads. Indeed, all of the b12 neutralizing activity could be adsorbed with YU2 gp140-coated beads, but none of neutralizing activity could be adsorbed when the assay was performed in the presence of saturating concentrations of b6 (Figure S3).
We next performed the assay using donor sera (Figure 3A). For one donor (#23), adsorption of the broad serum neutralizing activity with YU2 gp140 was completely inhibited by b6, suggesting that CD4bs or CRbs NAbs dominate the serum neutralization breadth and potency in this individual. For 4 additional donors, b6 inhibited 50–70% of broad serum neutralization indicating that CD4bs or CRbs Abs contribute significantly to the overall serum neutralization breadth and potency but not exclusively. For the remaining 12 donors, none or only a small fraction of the broad serum neutralization was blocked by b6, suggesting that CD4bs and CRbs-directed NAbs are of minor importance to the broad neutralizing responses in these individuals. A caveat to note is that anti-CD4bs or anti-CRbs NAbs that mediate the serum neutralization against some isolates (but not others) could be present in these 12 sera, but their epitopes may not be properly expressed in the context of YU2 gp140 used in the adsorption experiments.
Since the b6-blocking approach does not discriminate between CD4bs and CRbs-directed NAbs, we also performed serum adsorptions with a D368R gp120 variant that fails to bind CD4 and most, although not all, CD4bs-directed mAbs [25]. In two cases, a positive correlation was observed between the b6-inhibition adsorptions and the D368R adsorptions, indicating that the NAbs contributing to serum neutralization breadth in these donors are directed against the CD4bs (Figure 3B). However, for donors #20 and #57, the NAb specificities mediating serum neutralization breadth and potency competed with mAb b6 for gp140 binding yet did not exhibit sensitivity to the D368R mutation. It is possible that these NAbs are directed against the CRbs or novel epitopes that overlap the b6 epitope. Alternatively, these NAbs may be directed against the CD4bs but are insensitive to the D368R substitution.
The inability to adsorb a significant fraction of the broad serum neutralization with recombinant Env proteins in approximately one third of the donors prompted us to develop mapping strategies based on functional assays. As a first approach, we tested all of the sera for neutralizing activity against approximately 100 JR-CSF pseudoviruses incorporating single amino acid substitutions (Figure 4 and Table S1). In principle, if the serum neutralization against JR-CSF were mediated by a small number of NAb specificities, this activity would be diminished against pseudoviruses incorporating mutations that disrupt the epitopes targeted by these bNAbs. Indeed, for approximately 75% of the donors, potent serum neutralization against JR-CSF was abrogated by single amino acid substitutions. Notably, for the donor from whom PG9 and PG16 were isolated (#24), the N160K substitution in the context of JR-CSF Env resulted in complete viral escape from serum neutralization. Considering that this glycan is essential for PG9 and PG16 neutralizing activity [26], but does not affect the binding or neutralization profiles of any of the other Abs we tested (Table S2), this result suggests that this donor's potent serum neutralization against JR-CSF is entirely mediated by PG9, PG16, and similar antibodies. The N160K mutation also diminished serum neutralization against JR-CSF in four additional donors (#56, #29, #31, and #21), suggesting the presence of NAbs that target epitopes overlapping that of PG9 and PG16 in these sera. Alanine mutations in other regions of the V1/V2 and V3 loops also abrogated serum neutralizing activity against JR-CSF in these donors, further suggesting involvement of these regions in forming the epitopes recognized by these NAbs. Additionally, the broad and potent serum neutralizing activity in the five sera above could be not be efficiently adsorbed with monomeric gp120 or recombinantly trimerized gp140, indicating that these NAbs bind poorly to recombinant Env proteins. However, it is worth noting that a small number of gp120s have been identified that react weakly with PG9 [26], suggesting that it may be possible to adsorb these donor sera on certain gp120s.
In five different donors (#17, #51, #26, #14, #33), the N-linked glycan at position 332 at the base of the V3 loop of gp120 was critical for potent serum neutralization against JR-CSF. Since this glycan is also critical for 2G12 recognition [36], but does not significantly affect the neutralization profiles of other neutralizing mAbs we tested (Table S2), this result raised the possibility that the donor sera target epitopes overlapping that of 2G12. Interestingly, 2G12 also requires the glycan at position 295 for neutralizing activity, and one of the donors (#33) also exhibited sensitivity to this mutation. The glycan-dependent nature of the epitopes targeted by the NAbs in these sera is discussed below. Interestingly, a significant fraction of the broad serum neutralizing activity in all five of these sera could be adsorbed with monomeric gp120 and trimeric gp140, indicating that the epitopes targeted by these NAbs are expressed on recombinant forms of Env.
Approximately 21% of the donors exhibited significant sensitivity to the I165A mutation located in the V2 loop of gp120. To gain insight into the potential epitopes targeted by these NAbs, we tested a panel of neutralizing monoclonal antibodies for sensitivity to this mutation. Interestingly, only the trimer-specific NAbs 2909, 2.2G, and 2.3E required this residue for potent neutralizing activity (Table S2). Although these NAbs are strain-specific, they recognize quaternary epitopes involving the V2 and V3 loops of gp120 [37], [38]. For most of these donors, the broad neutralizing activity could not be adsorbed with monomeric gp120, suggesting that these NAbs may also target epitopes that are preferentially expressed on trimeric HIV-1 Env. Alternatively, the serum NAbs may target novel viral epitopes that are disrupted by the I165A mutation.
Of note, the contribution of CD4bs-directed NAbs could not be assessed using the mutant virus approach because pseudoviruses incorporating mutations in this region are non-infectious. Indeed, the donor sera identified above with CD4bs-directed neutralizing activity did not exhibit significant sensitivity to any of the pseudovirus mutants in our panel (Figure 4 and Table S1) consistent with the critical residues that form the epitopes recognized by these bNAb specificities being located in the CD4bs and the corresponding variants being absent from the pseudovirus panel. A second caveat of this assay is that certain mutations impart global sensitivity to serum neutralization, and therefore the effect of these residues in forming the epitopes recognized by the broadly neutralizing specificities in the sera could not be assessed. Interestingly, several of the substitutions that conferred a global neutralization sensitive phenotype were located in the V2 and V3 loops and the CRbs (Table S1), suggesting that these residues play a role in restricting antibody access to potentially neutralizing epitopes.
We next investigated whether the NAb specificities that were mediating the potent serum neutralization against JR-CSF were also mediating the breadth of serum neutralization. Since most of the sera exhibited sensitivity to the N160K, N332A, or I165A mutations, we introduced these mutations into a cross-clade pseudovirus panel and then tested the corresponding sera for neutralization against these variants (Figure 5A). In the context of 92RW020 (clade A), the N160K mutation resulted in a loss of viral infectivity and was therefore excluded from the N160K panel. For all of the donors, the single amino acid substitution that diminished serum neutralization against JR-CSF also abrogated cross-clade serum neutralization, suggesting that the broad and potent serum neutralization is mediated by a limited number of antibody specificities. In a minority of cases, the serum neutralizing activity against a particular isolate was not affected by the single amino acid substitution. However, these isolates were usually not potently neutralized by the serum (Figure 5B). For example, for donor #24, the N160K substitution only reduced the serum neutralizing activity against YU2 by 40%, but the serum neutralizing titer against this isolate was at least 10-fold lower than most of the other viruses on the panel. Thus, although there may be several bNAb specificities in these sera, the most potent neutralizing activity is likely only mediated by a small subset of bNAbs.
We next sought to determine whether 2G12-like antibodies were mediating the serum neutralization breadth and potency in donors who exhibited sensitivity to the N332A mutation. Since 2G12 binding to gp120 is completely inhibited by 1M mannose [36], we first evaluated whether high concentrations of mannose could inhibit binding of the serum NAbs to YU2 gp140. For this, we performed serum adsorptions in the presence of 1M mannose or 1M glucose (negative control). As expected, 2G12 neutralizing activity was retained after adsorption with gp120-coupled beads in the presence of 1M mannose and depleted when the adsorption was performed in the presence of 1M glucose (Table S3). In contrast, the neutralization depletion efficiency of all five N332A-sensitive sera was similar in the presence of both mannose and glucose, indicating that the monosaccharide mannose does not inhibit the interaction of the NAbs in the sera with YU2 gp140 (Table S3). Notably, the crystal structure of 2G12 complexed with oligomannoses shows that 2G12 primarily recognizes the terminal mannose on the D1 arm of Man9GlcNAc2 [39]. If the sera contain glycan-specific NAbs that require several mannose residues for high affinity interaction, the monosaccharide mannose may not efficiently inhibit binding of these antibodies to gp120. Therefore, in a second approach we investigated whether the N332A-sensitive NAbs in the sera would react with the high-mannose glycans presented on a heavily N-glycosylated yeast protein, referred to as TM-Pst1, that has been shown to bind 2G12 with high affinity and inhibit 2G12 neutralization of HIV-1 pseudoviruses [40]. For these experiments, we performed serum adsorptions with TM-Pst1-coupled beads or blank control beads. Since all of the sera contained TM-Pst1 binding antibodies, ELISA assays were used to confirm depletion efficiency (data not shown). As above, we next tested the adsorbed sera for neutralizing activity against a cross-clade pseudovirus panel (Figure 6 and Figure S4). Interestingly, for the donor that exhibited sensitivity to mutations at positions 295 and 332 (#33), 70–95% of the serum neutralization was adsorbed with the TM-Pst1 protein. Furthermore, the antibodies eluted off the beads bound to gp120, displayed cross-clade neutralizing activity, and exhibited sensitivity to the N332A mutation (Figure S5). Thus, it appears likely that this donor's broad and potent serum neutralizing activity is mediated by bNAbs that target the oligomannose cluster on the HIV-1 glycan shield. In contrast, for the remaining four donors, the broad neutralizing serum activity could not be adsorbed with TM-Pst1-coupled beads. One possible explanation is that these NAbs bind to glycan epitopes distinct from the 2G12 epitope. Alternatively, these NAbs may bind to protein epitopes that are conformationally dependent on the glycan at position 332. To further examine whether any of these sera contain anti-glycan NAbs that bind to epitopes overlapping that of 2G12, we performed competition ELISA experiments using biotinylated mAb 2G12 (Figure S6). None of the sera, including the serum that could be adsorbed on the yeast TM-PstI protein, decreased the binding of biotinylated mAb 2G12 to JR-CSF gp120, suggesting that any glycan-specific NAbs in these sera bind to epitopes distinct from that of 2G12.
Next, we further assessed the contribution of PG9 and PG16-like antibodies to serum neutralization breadth and potency in donors who exhibited sensitivity to the N160K mutation. We have previously observed that pseudoviruses produced in cells that have been treated with kifunensine, a mannose analogue that inhibits type-1 endoplasmic reticulum (ER) and Golgi α-mannosidases, are resistant to PG9 and PG16 neutralization (Doores et al., submitted). Surprisingly, we found that other NAbs, including those that bind to quaternary epitopes on trimeric Env, neutralized kifunensine-treated pseudoviruses with similar potency as wild-type (WT) pseudoviruses (Table S4). Therefore, to further investigate whether PG9 and PG16-antibodies mediate the potent serum neutralizing activity observed in donors who exhibit sensitivity to the N160K mutation, we tested these sera for neutralizing activity against JR-CSF pseudoviruses produced in the presence of kifunensine. For comparison, we also tested sera from donors whose serum neutralizing activity was unaffected by this mutation. Indeed, we found that only the five sera that exhibited sensitivity to the JR-CSF N160K mutation showed markedly diminished neutralizing activity against kifunensine-treated JR-CSF pseudoviruses (Figure 7). To determine whether the kifunensine-sensitive NAbs were also mediating broad serum neutralization, we next tested these five sera for neutralization against a cross-clade panel of pseudoviruses produced in the presence of kifunensine. For four out of the five donors, the broad serum neutralizing activity was almost completely abolished against kifunensine-treated pseudoviruses (Figure 8). Notably, for donor #29, both kifunensine treatment and the N160K mutation had only moderate effects on serum neutralization against the DU172 isolate, further suggesting a correlation between sensitivity to the N160K mutation and kifunensine treatment. Based on these results, it appears highly likely that PG9 and PG16-like antibodies mediate most of the broad and potent serum neutralizing activity in the four donors.
We have used several established and new approaches to map the serum NAb specificities in 19 donors, mostly infected with non-clade B viruses, who exhibit remarkably broad and potent serum neutralizing activity. This study extends our previous serum mapping studies, which only focused on a small number of patients with relatively limited serum neutralization breadth and potency [24]. Here, we found that the broad serum neutralizing activity in about one third of the donors in our cohort could not be efficiently adsorbed with recombinant monomeric gp120 or recombinantly trimerized gp140, suggesting the possible presence of bNAbs that target quaternary epitopes. Interestingly, in a subset of donors, the broad serum neutralizing activity was more efficiently adsorbed with recombinantly trimerized gp140 than monomeric gp120, indicating the presence of NAbs directed against gp120 epitopes that are somewhat better exposed on the YU2 gp140-foldon trimer or that the NAbs are directed against gp41. A caveat associated with these conclusions is that a single Env protein was used for the serum adsorption studies; it is possible that some of the broadly neutralizing specificities in these sera would react differently with Env proteins derived from other isolates.
MPER-directed neutralizing activity was detected in 30% of the donors, but these NAbs did not substantially contribute to the overall breadth and potency observed in the sera. Indeed, the broad and potent serum neutralizing activity for most donors in our cohort could either be adsorbed with monomeric gp120 and/or was abrogated by specific mutations located in gp120, further suggesting that gp41-directed NAb specificities do not mediate the overall serum neutralization breadth and potency in these donors. This result is in agreement with previous serum mapping studies that suggest NAbs directed against the MPER rarely mediate broad and potent serum neutralization [16], [17], [18], [22].
Previous serum mapping studies have shown that, in some individuals, NAbs directed against the CD4bs or CRbs mediate broad serum neutralization [16], [17], [25], [41], [42]. Furthermore, four broadly neutralizing CD4bs-directed mAbs have been isolated from HIV-1 infected donors [34], [43,Wu et al., submitted]. To evaluate whether CD4bs or CRbs-directed NAbs were mediating the serum neutralization breadth and potency in these donors, we performed serum adsorptions in the presence or absence of saturating concentrations of the non-neutralizing CD4bs-directed mAb b6. We found that, in 5 out of the 17 donors, a significant proportion of the overall broad and/or potent serum neutralizing activity was mediated by receptor binding site-directed NAbs, and in one donor, nearly all of the broad serum neutralizing activity could be attributed to the b6-competed fraction. In two donors, similar results were obtained when a CD4bs-altered mutant (D368R) was used for adsorptions, indicating that the NAbs contributing to the serum neutralization breadth in these donors are directed against the CD4bs.
Although serum adsorption studies provide valuable insight into the epitopes targeted by NAbs that react with recombinant Env or linear peptides, defining the epitopes recognized by trimer-specific antibodies requires the use of functional assays. By testing the sera for neutralizing activity against a large panel of JR-CSF pseudovirus mutants incorporating single amino acid substitutions, we were able to define the epitopes recognized by the bNAb specificities in the context of the native trimer. Interestingly, we found that single amino acid mutations in Env frequently generated viruses that were far less sensitive to serum neutralization than the wild-type JR-CSF virus. Incorporation of these single substitutions into a cross-clade pseudovirus panel similarly generated viruses that were far more resistant to neutralization by certain sera than the corresponding wild-type viruses. These findings suggest that the broad and potent serum neutralizing activity in these donors is mediated by a limited number of antibody specificities. Notably, for the donor from whom PG9 and PG16 were isolated, the neutralization properties of the serum mirrored that of PG9 and PG16. These results complement the previous observation that PG9 and PG16 could recapitulate this donor's broad and potent serum neutralization against most isolates [26]. Thus, it appears highly likely that PG9, PG16, and related antibodies mediate this donor's serum neutralization breadth and potency. In the context of the single amino acid variant viruses, it should be noted that although it is most likely that the substitutions directly define residues involved in bNAb binding, it is also possible that the effects are due to transmitted effects from substitutions distant from the neutralizing epitope.
Interestingly, for 9 out of the 19 donors, the potent cross-clade serum neutralizing activity was mediated by NAbs dependent on specific glycans for epitope recognition. In four of these donors, the glycan at position 160 at the base of the V2 loop was critical for serum neutralization breadth and potency, suggesting the presence of NAbs that target epitopes overlapping that of PG9 and PG16. To evaluate the contribution of PG9 and PG16-like antibodies to broad and potent serum neutralizing activity, we used the following criteria: 1) the broad and potent serum neutralizing activity could not be efficiently adsorbed with recombinant Env proteins, 2) the N-linked glycan at position 160 was essential for serum neutralization breadth and potency, and 3) the broad and potent serum neutralizing activity was diminished against pseudoviruses produced in the presence of the glycan processing inhibitor kifunensine. Based on these criteria, PG9 and PG16-like NAbs were identified in approximately 21% of the donors we studied, demonstrating that this specificity is relatively common in donors that develop broad and potent serum neutralization.
The I165A substitution located in the V2 loop of gp120 abrogated broad serum neutralizing activity in 4 out of the 19 donors studied. Given that the trimer-specific NAbs 2909, 2.2G, and 2.3E require this amino acid for potent neutralizing activity, and that the broad serum neutralizing activity in these 4 donors could not be adsorbed with monomeric gp120, it appears highly likely that these bNAb specificities also target epitopes preferentially expressed on trimeric HIV-1 Env. Indeed, this epitope may be linked to the epitope recognized by PG9 and PG16.
We found that the N-linked glycan at position 332 was critical for broad and potent serum neutralizing activity in approximately 25% of donors. Considering that this N-linked glycan is also important for formation of the 2G12 epitope, we investigated whether 2G12-like antibodies were mediating the serum neutralization breadth and potency in any of these donors. Notably, the presence of 2G12-like antibodies in HIV-1 positive sera has also been discussed recently by others [22], [44]. In one donor, a large fraction of the broad serum neutralizing activity cross-reacted with a yeast glycoprotein that expresses homogenous Man8GlcNac2 carbohydrates and binds 2G12 with high affinity, suggesting the presence of bNAbs that bind directly to the glycan shield. This donor's serum neutralizing activity was also dependent on the presence of the N-linked glycan at position 295, a second N-linked glycan required for 2G12 recognition [36]. Interestingly, the sera from this individual did not inhibit 2G12 binding to gp120, suggesting that this bNAb specificity may bind to a glycan epitope distinct from that of 2G12. Considering that protein-carbohydrate interactions are typically weak, perhaps this antibody specificity gains the necessary avidity by cross-linking two protomers within a trimer [45], [46]. Hence, low affinity binding to monomeric gp120 may also explain the lack of competition with 2G12. Another observation worth noting is that only a single individual in our study was found to have bNAb specificities targeting the glycan shield, indicating that these types of antibodies rarely mediate serum neutralization breadth and potency. However, it has recently been shown that 2G12 is unusually efficient in protection relative to its neutralizing ability [9], suggesting that the glycan shield may have advantages as a vaccine target. Thus, it will be of interest to isolate the glycan-specific NAbs from this donor and test their protective efficacy to determine whether this is a general property of NAbs directed against the glycan shield. For the remaining four donors that exhibited N332A sensitivity, adsorption with TM-Pst1 coupled beads had no effect on the neutralizing activity of the sera; it therefore appears likely that the NAbs mediating serum neutralization breadth and potency in these donors bind to protein epitopes that are conformationally dependent on the glycan at position 332.
The observation that only a limited number of antibody specificities mediate serum neutralization breadth and potency in these donors contrasts with the results of a previous study by Scheid et. al in which no single broadly neutralizing monoclonal antibodies were isolated using single B cell sorting [27]. The authors of that study concluded that the serum neutralization breadth in their donors was due to the combined activity of a large number of antibody specificities that individually display limited breadth and potency. Indeed, this may be the case for certain donors, particularly those with relatively weak serum neutralization breadth and potency. However, in donors with extraordinarily potent serum neutralization, it appears that this activity is usually only associated with one or a few different specificities. Of note, serum neutralization breadth found to be mediated by a single specificity may in fact require multiple different NAbs circulating in the plasma that recognize overlapping targets within the same region. For example, a number of different CD4bs-directed NAbs may mediate broad serum neutralization in donors where all of the serum neutralizing activity can be mapped to this region.
In summary, the data presented here show that the unusually potent cross-clade serum neutralizing activity observed in a selection of donors is mediated by a small number of antibody specificities that target conserved regions of Env to a significant extent (Figure S7 and Figure 9). Antibodies dependent on specific glycans for Env recognition were found to be responsible for this activity in approximately half of the donors we studied; 21% of these donors targeted epitopes overlapping those of PG9 and PG16 and 25% targeted epitopes that were dependent on the presence of an N-linked glycan at position 332. CD4bs and/or CRbs-directed NAbs contributed to serum neutralization breadth in 25% of donors, NAbs sensitive to the I165A substitution were identified in 21% of donors, and MPER-directed NAbs were present in a subset of donors but did not make a substantial contribution to the overall serum neutralization breadth and potency. Future studies, aimed at isolating and characterizing broadly neutralizing monoclonal antibodies from these donors, will be important for the molecular definition of the broadly neutralizing epitopes recognized by the donors.
After obtaining written informed consent, sera and plasma were collected from HIV-1 infected volunteers in Rwanda, Zambia, Ivory Coast, Thailand, Kenya, Uganda, United Kingdom, and the United States. Eligible participants were age 18 or older, were HIV-1 infected for at least 3 years prior to the day of screening, were clinically asymptomatic, without evidence of progression to AIDS based on WHO Stage III or IV criteria or CD4 count <200 cells/mm3 and were not on antiretroviral therapy (ART) for at least 1 year. The study was reviewed and approved by the Republic of Rwanda National Ethics Committee; Emory University Institutional Review Board; University of Zambia Research Ethics Committee; Charing Cross Research Ethics Committee; UVRI Science and Ethics Committee; University of New South Wales Research Ethics Committee; and St. Vincent's Hospital and Eastern Sydney Area Health Service; Kenyatta National Hospital Ethics and Research Committee; University of Cape Town Research Ethics Committee; International Institutional Review Board, Mahidol University Ethics Committee; Walter Reed Army Institute of Research (WRAIR) Institutional Review Board; Ivory Coast Comité National d'Ethique des Sciences de la Vie et de la Santé (CNESVS).
Sera and plasma were collected from a cohort of HIV-1 infected individuals (IAVI Protocol G), as previously described [15]. Samples were heat-inactivated at 55 °C for 1 h prior to use in neutralization assays.
The panel of MAbs directed to HIV-1 Env included MAb b12, b13, and 15e, directed to epitopes overlapping the CD4bs of gp120 [43], [47]; 2G12, directed to a cluster of oligomannose residues on gp120 [36], [48]; A32, directed against an epitope comprised of the C1/C2/C4/and CD4i domains [49]; C11, directed against the C1 domain; F425/b4e8, directed to the V3 loop on gp120 [50]; ×5, directed to the CRbs binding site of gp120 [51]; PG9, PG16, 2.2G, 2.3E, and 2909, directed to quaternary epitopes comprised of the V2 and V3 loops of gp120 [26], [37], [38]; and 2F5, 4E10, and Z13e1, directed to the gp41 MPER [48], [52], [53].
Pseudoviruses incorporating single alanine substitutions were generated by transfection of 293T cells with an Env-expressing plasmid and an Env-deficient genomic backbone plasmid (pSG3ΔEnv), as described previously [54]. Pseudoviruses were harvested 72 hours post transfection for use in neutralization assays. Neutralizing activity was assessed using a single round of replication pseudovirus assay and TZM-bl target cells, as described previously [54]. A chimeric HIV-2 clone containing the MPER of HIV-1 was derived from the parental HIV-2 7312A clone in which the HIV-2 Env MPER sequence QKLNSWDVFGNWFDLASWVKYIQ was replaced by the HIV-1 MPER sequence LALDKWASLWNWFDITKWLWYIK, as described [22]. To determine IC50 values, serial dilutions of plasma or mAb were incubated with virus and the dose-response curves were fitted using nonlinear regression. Serum antibody concentrations and antigen-binding activity were determined by ELISA, as described below.
The envelope glycoproteins were expressed by transfecting the 293T cell line in serum-free medium (Invitrogen, Carlsbad, CA). In brief, the 293T cells were seeded in T225 flasks at a density of 1×105 cells/cm2 and transfected with the expression plasmid YU2gp120/pcDNA3.1(−) for monomeric gp120 or YU2gp140-fibritin/pcDNA3.1(−) DNA for trimeric gp140 proteins. In addition, YU2 gp140-fibritin mutants were codon optimized for mammalian expression and synthesized (GeneArt AG, Germany). The D368R YU2 gp120 mutant was generated by Quik change mutagenesis (Stratagene). Four days post-transfection, cell culture supernatants were collected, clarified, filtered and two protease inhibitor tablets (Roche) per liter of supernatant were added to limit proteolysis. The gp120 or gp140-containing supernatants were stored at 4°C prior to purification. The supernatants containing the gp120 or gp140 proteins were applied to columns containing 10 ml of Galanthus nivalis lectin-bound agarose (Vector Laboratories). The column was then washed sequentially with 10 column volumes of phosphate-buffered saline (PBS) (pH 7.4) containing 0.5 M NaCl, followed by 10 column volumes of PBS (pH 7.4). The lectin-bound glycoproteins were eluted with a total of 12 column volumes of elution buffer (PBS buffer [pH 7.4] with 0.5 M methyl-D-mannopyranoside). The mannoside-eluted glycoproteins were pooled and the protein eluates were dialyzed against phosphate-buffered saline (PBS), pH 7.4, and concentrated with Amicon Ultra 30,000 MWCO centrifugal filter devices (Millipore, Bedford, MA). The purified proteins were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis and ELISA analysis, and protein purity was verified to approach 95% homogeneity.
Serum adsorptions with antigen-coupled beads were performed using tosyl-activated magnetic beads, as described previously [25]. 0.5 mg of gp120 and gp140 and 2 mg of TM-Pst1 were used for bead coupling. Two to three rounds of adsorption were performed to ensure complete removal of antigen-specific antibodies. Functional Abs were eluted from beads by exposing the beads to series of increasingly acidic conditions, as described [17]. For serum adsorptions performed in the presence of b6, gp140-coupled beads were pre-incubated with 200 µg/ml IgG b6 for 1 h at room temperature before adding serum. ). For adsorptions performed in the presence of monosaccharides, sera and mAb 2G12 were pre-incubated with 1M mannose or 1M glucose for 1 h at room temperature before adding to gp140 or gp120-coupled beads. Serum antibody concentrations and antigen-binding activity were determined by ELISA, as described below.
Ninety-six-well ELISA plates were coated overnight at 4 °C with 50 uL PBS containing 50 ng of goat anti-human IgG Fc (Pierce) or 100 ng gp120, gp140, or TM-Pst1 per well. The wells were washed four times with PBS containing 0.05% Tween 20 and blocked with 3% BSA at room temperature for 1 h. Serial dilutions of sera or mAb were then added to the wells, and the plates were incubated at room temperature for 1 hour. After washing four times, goat anti-human IgG F(ab')2 conjugated to alkaline phosphatase (Pierce), diluted 1∶1000 in PBS containing 1% BSA and 0.025% Tween 20, was added to the wells. The plate was incubated at room temperature for 1 h, washed four times, and the plate was developed by adding 50 uL of alkaline phosphatase substrate (Sigma) to 5 mL alkaline phosphatase staining buffer (pH 9.8), according to the manufacturer's instructions. The optical density at 405 nm was read on a microplate reader (Molecular Devices). Antibody concentration was calculated by linear regression using a standard concentration curve of purified IgG protein. Competition ELISAs were performed by pre-incubating 5-fold serum dilutions (starting at a 1∶25) on JR-CSF gp120-coated ELISA wells for 30 min at room temperature and then adding a concentration of biotinylated 2G12 previously determined to give a half-maximal binding signal. Biotinylated 2G12 was detected using alkaline phosphatase- conjugated streptavidin (1∶200 in dilution buffer; Pierce), and the plates were developed as described above. Endpoint titers of the plasma antibodies were defined as the last reciprocal serum dilution at which the OD signal was greater than twofold over the background signal.
|
10.1371/journal.pcbi.1005066 | Different Evolutionary Paths to Complexity for Small and Large Populations of Digital Organisms | A major aim of evolutionary biology is to explain the respective roles of adaptive versus non-adaptive changes in the evolution of complexity. While selection is certainly responsible for the spread and maintenance of complex phenotypes, this does not automatically imply that strong selection enhances the chance for the emergence of novel traits, that is, the origination of complexity. Population size is one parameter that alters the relative importance of adaptive and non-adaptive processes: as population size decreases, selection weakens and genetic drift grows in importance. Because of this relationship, many theories invoke a role for population size in the evolution of complexity. Such theories are difficult to test empirically because of the time required for the evolution of complexity in biological populations. Here, we used digital experimental evolution to test whether large or small asexual populations tend to evolve greater complexity. We find that both small and large—but not intermediate-sized—populations are favored to evolve larger genomes, which provides the opportunity for subsequent increases in phenotypic complexity. However, small and large populations followed different evolutionary paths towards these novel traits. Small populations evolved larger genomes by fixing slightly deleterious insertions, while large populations fixed rare beneficial insertions that increased genome size. These results demonstrate that genetic drift can lead to the evolution of complexity in small populations and that purifying selection is not powerful enough to prevent the evolution of complexity in large populations.
| Since the early days of theoretical population genetics. scientists have debated the role of population size in shaping evolutionary dynamics. Do large populations possess an evolutionary advantage towards complexity due to the strength of natural selection in these populations? Or do small populations have the advantage, as genetic drift allows small populations to cross fitness valleys that large populations are unlikely to traverse? There are many theories that predict whether large or small populations–those with strong selection or those with strong drift–should evolve the greatest complexity. Here, we use digital experimental evolution to examine the interplay between population size and the evolution of complexity. We found that genetic drift could lead to increased genome size and phenotypic complexity in very small populations. However, large populations also evolved large genomes and phenotypic complexity. Small populations evolved larger genomes through the fixation of slightly deleterious insertions, while large populations used rare beneficial insertions. Our results suggest that both strong drift and strong selection can allow populations to evolve similar complexity, but through different evolutionary trajectories.
| The relative importance of adaptive (i.e., selection) versus non-adaptive (i.e., drift) mechanisms in shaping the evolution of complexity is still a matter of contention among evolutionary biologists [1–6]. In molecular evolution, the role of non-adaptive evolutionary processes such as genetic drift and genetic draft are well-established [7–9]. Theoretical population-genetic principles argue that neutral evolution, not natural selection, drove the evolution of large, primarily non-functional, genomes [10–12]. Meanwhile, there exists abundant experimental evidence that natural selection is the main cause of evolutionary change [13–15], including the spread of novel adaptive phenotypes [16, 17] in experimental populations. However, it is still possible that non-adaptive processes play a significant role in the evolution of complexity. For instance, genetic drift (or relaxed selection) may allow for the accumulation of mutations that can later lead to the evolution of novel complexity [4, 18]. Much of the work demonstrating the role of selection in driving the evolution of novel complex traits is based on experiments with large populations and strong selection [19]. In much smaller populations (i.e., those with fewer than 104 individuals), selection is weaker, and genetic drift begins to alter evolutionary dynamics [15, 20]. Therefore, to explain the role of adaptive vs. non-adaptive process in the evolution of complexity, one must explore the role of population size in the evolution of complexity.
Both theoretical modeling and experiments suggest many possibilities for the relationship between population size and the evolution of complexity. There are two classes of evolutionary trajectories that would favor large populations in the evolution of complexity. First, populations could perform an adaptive walk (the fixation of a sequence of beneficial mutations) towards the evolution of a novel complex trait [21]. If this was the case, then larger populations would follow this trajectory faster than small populations due to their larger mutation supply. Experiments with microorganisms support the possible existence of adaptive trajectories towards complexity, as there is strong evidence that the mutations leading up to a phenotypic innovation in both Escherichia coli [22] and phage λ [23] were under positive selection. However, it is unclear whether adaptive mutations generally precede the evolution of complex traits or whether these large microbial populations can only take adaptive walks due to the intensity of selection in large populations. The second type of trajectory that favors large populations is the neutral walk (the fixation of a sequence of neutral mutations). While any individual neutral mutation has a low probability of fixation, a large population would be able to accumulate many neutral mutations at any given time allowing for the exploration of its fitness landscape. Work by Wagner and colleagues suggests that many phenotypic traits are connected to each other by sequences of phenotypically neutral mutations [18, 24].
If the evolution of complexity requires the fixation of deleterious mutations (for example, via valley-crossing), then the elimination of deleterious mutations by purifying selection may limit the evolutionary advantage large populations may have. Wright was the first to propose an evolutionary advantage of small populations due to valley-crossing [25]. More recently, scientists have explored under which conditions small populations have an evolutionary advantage over large populations [26, 27]. A prominent theory that predicts that small (but not large) populations should evolve the greatest genomic complexity (and subsequently organismal complexity) is the Mutational Burden (or Mutational Hazard) hypothesis, proposed by Lynch and colleagues [4, 28, 29]. This hypothesis argues that genome size should be inversely correlated with the product of the effective population size and the mutation rate [3, 28]. Strong purifying selection against excessive genome size streamlines the genomes in large populations [30–32]. Meanwhile, weakened purifying selection and increased genetic drift in small populations results in the accumulation of slightly deleterious excess genome content [3, 29]. At a later time, this slightly deleterious genome content may be mutated into novel beneficial traits [4, 33]. However, recent work on valley-crossing in asexual populations (and sexual populations with a low recombination rate) showed that both small and large populations cross valleys more than intermediate-sized populations [34–36]. Therefore, it is not clear whether large or small populations are expected to evolve the greatest complexity when deleterious mutations are required.
The long timescales required to observe the emergence of novelty and evolution of complexity make biological experiments to distinguish between these theories difficult to perform. To overcome this difficulty, we used digital experimental evolution [37] to test the role of population size on the evolution of genome size and phenotypic complexity in asexual organisms. Digital evolution has a long history of addressing macroevolutionary questions (such as the evolution of novel traits) experimentally [38, 39]. Digital populations can be manipulated in ways that biochemical organisms can not, making it possible to study aspects of the evolutionary process that are ordinarily too difficult to test [40]. In this regard, digital experimental evolution has the same goals as microbial experimental evolution: to use a well-controlled model system that is as simple as possible, to study “evolution in action” [41]. And while digital evolution studies cannot test hypotheses that depend on particular biochemical processes involved in cellular life, digital populations do undergo selection, drift, and mutation, allowing for their use in testing hypotheses derived from theoretical population genetics. Thus, digital experimental evolution represents a well-suited model system to test the population genetics-based theories concerning the role of population size in the evolution of complexity.
Here, we evolved populations ranging in size from 10 to 104 individuals, starting with a minimal-genome ancestor. We found that small populations do evolve greater genome sizes and phenotypic complexity (number of phenotypic traits) than intermediate-sized populations. These small populations evolve larger genomes primarily through increased fixation of slightly deleterious insertions. However, the small population sizes that enhance the evolution of phenotypic complexity also enhance the likelihood of population extinction. We also found that the largest populations evolved similar complexity to the smallest populations. Large populations evolved longer genomes and greater phenotypic complexity through the fixation of rare beneficial insertions instead. Large populations were able to discover these rare beneficial mutations due to an increased mutation supply. Finally, we found that a strong deletion bias can prevent the evolution of greater complexity in small, but not in large, populations.
To explore the effect of population size on the evolution of genome size and phenotypic complexity, we used the Avida digital evolution system [42]. Avida is a platform that allows researchers to perform evolution experiments inside of a computer, as the genetic code that evolves are actual computer programs of variable length. It has been used extensively in research in evolutionary biology [37, 43, 44], and is described in detail in Methods.
We evolved one hundred replicate populations across a range of population sizes (10 − 104 individuals) for 2.5 × 105 generations. Many of the smallest populations (those with ten individuals) did not survive the entire experiment. Therefore, we evolved one hundred additional small populations ranging from twenty individuals to ninety individuals in order to examine how the probability of extinction was related to the evolution of complexity. All populations with at least thirty individuals survived for the entire experiment. Forty-seven of the populations with ten individuals went extinct, while only one of one hundred populations underwent extinction in the populations with twenty individuals. Extinction was a consequence of populations evolving large genomes that accumulated deleterious mutations and led to the production of only non-viable offspring. These extinct populations were not included in the statistics described below.
Of the surviving populations, we first examined how genome size changes from the ancestral value of fifteen instructions. The size of the genome from every population size increased, on average (see Fig 1 and panel A in S1 Fig). However, both the smallest and the largest populations evolved the largest genomes. Populations with ten individuals evolved a median genome size of 35 instructions, while populations with ten thousand individuals evolved a median genome size of 36 instructions. The median final genome size decreased as population size increased for populations with between ten and fifty individuals. However, from populations with fifty individuals to populations with ten thousand individuals, the median final genome size increased as population size increased.
Next, we examined the dynamics of fixation of insertion mutations (insertions, for short) to explain why both the smallest and the largest populations evolved the largest genomes. For each experimental population, we counted every insertion that occurred on the fittest genotype’s ancestral lineage that went back to the ancestral genotype (the “line of descent”, see Methods). The median number of insertions fixed follows the same trend as the evolution of genome size (S2 Fig). A large fraction of these fixed insertions are slightly deleterious in populations with fewer than one hundred individuals (see Fig 2 and panel B in S1 Fig). However, no insertions are slightly deleterious, on average, in large populations with more than one hundred individuals. The opposite trend holds for beneficial insertions. The fraction of insertions that are under positive selection increases with increasing population size, with the largest populations usually fixing only beneficial insertions (Fig 3 and panel C in S1 Fig). These data demonstrate that small populations evolve larger genomes through the fixation of slightly deleterious insertions. However, large populations can evolve similarly large genomes through the fixation of rare beneficial insertions.
Next, we focus on the role of population size in the evolution of phenotypic complexity (defined as the number of phenotypic traits). In Avida, a phenotypic trait is a program’s ability to perform a certain mathematical operation on binary numbers (see Methods). The evolution of phenotypic complexity follows the same trend as the evolution of genome size (see Fig 4 and panel D in S1 Fig). Populations with ten individuals evolved a median of four traits, while populations with one thousand and ten thousand individuals evolved a median of one trait. The rest of the population sizes evolved a median of zero traits. As an avidian’s fitness is primarily determined by its phenotypic traits in the Avida environment used here, the evolution of fitness showed a similar trend to the evolution of phenotypic complexity (S3 Fig).
That the trend in genome size evolution and in phenotypic complexity evolution are mirrored suggests that the evolution of larger genomes enables the evolution of increased phenotypic complexity. To establish a link between the two, we performed two tests. First, we examined the correlation between genome size and phenotypic complexity across all populations. Phenotypic complexity is positively correlated with genome size (Fig 5, Spearman’s ρ ≈0.72; p < 2.3 x 10−57), suggesting that it was the increased genome size that allowed for the evolution of increased phenotypic complexity. However, there are two potential mechanisms that could cause an increased genome size to result in increased phenotypic complexity. On the one hand an increased genome size could simply allow more “room” for novel functional content. On the other hand, it could be that the increased genome size leads to a faster rate of evolution due to the increased genomic mutation rate. To examine the role of an increased mutation rate in driving the evolution of phenotypic complexity, we evolved a further one hundred populations of ten individuals with a fixed genomic mutation rate of 1.5 × 10−1 (i.e., the ancestral genomic mutation rate). Under this condition, no population went extinct (as opposed to forty-seven in the variable mutation rate treatment). The fixed genomic mutation rate populations evolved a median of 2 phenotypic traits compared to the variable genomic mutation rate populations that had evolved a median of 4 phenotypic traits (S4 Fig). These data demonstrate that the increased genomic mutation rate that follows from larger genomes does increase the evolution of phenotypic complexity. However, even with a fixed genomic mutation rate, the smallest populations still evolved a greater median number of traits (on average 2 traits) than every other population size. Thus, while an increased genomic mutation rate (due to increased sequence length) indeed enhances the evolution of phenotypic complexity, small populations still possess an evolutionary advantage due to drift-driven increases in genome size only.
In the previous experiments, large populations evolved larger genomes and greater phenotypic complexity because they fixed rare beneficial insertions. Next, we more closely examine the finding that beneficial insertions are necessary for the evolution of complexity in large populations. We repeated the experiments with the same population sizes and mutation rates, except we changed how insertions worked. Instead of inserting one of the twenty-six instructions that compose the Avida instruction set, we inserted “blank” instructions into the genome (see Methods for details). These blank instructions cannot be beneficial (on their own or in combination with existing instructions) and would have to be further mutated to lead to the evolution of phenotypic complexity. In this treatment, greater phenotypic complexity in large populations would require a two-step mutational process, as opposed to the single step in a beneficial insertion.
We saw no qualitative difference in the trend between these experiments and the original experiments (S5 Fig). Very small and large populations still both evolved the largest genomes and the greatest phenotypic complexity. Populations of all sizes evolved longer genomes and more phenotypic traits in this treatment (S5 Fig) than in the original treatment (Figs 1 and 4). The fraction of fixed insertions that were under positive selection decreased for every population size compared to the original experiments, as expected from the insertion of non-functional instructions (S6 Fig). We observed an increased rate of extinction in the very small populations, with only 2 populations with ten individuals and 25 populations with twenty individuals surviving the experiment. Population extinction was likely enhanced by the increased growth in genome size in these experiments as compared to the original experiments.
Finally, we performed experiments to test whether the effect of a deletion bias (a higher fraction of deletions among all indels) alters the relationship between population size and the evolution of complexity. A biased ratio of deletion to insertion mutations is found in biological organisms across the tree of life, especially in bacteria [45, 46]. In these experiments we set the ratio of deletions to insertions as 9:1, but kept the total indel mutation rate as in the original experiments. In this treatment, only one population with ten individuals went extinct, as opposed to 47 populations in the original treatment. However, the advantage towards evolving complexity previously enjoyed by small populations vanished (S7 Fig). The median genome size increased as the population size increased for all populations sizes. Only the largest populations evolved a median number of novel phenotypic traits greater than zero. These results suggest that it is not only the role of genetic drift, but the equal frequency of insertions and deletions that results in the increased genome size and phenotypic complexity in small populations.
The idea that small populations could have an evolutionary advantage over large populations dates back to Wright and his Shifting Balance theory [25]. More recently, a potential small-population advantage has been demonstrated both theoretically [27] and experimentally [26], but only in regard to short-term increases in fitness. The Mutational Burden hypothesis provides an evolutionary mechanism that gives small populations an advantage towards increased phenotypic complexity [4, 33]. However, an experimental demonstration of this advantage is lacking. Our study provides further insight into the conditions that give small populations such an evolutionary advantage. We confirmed that small populations do evolve larger genomes due to the increased fixation of slightly deleterious mutations, as predicted [28]. We also showed how small populations have an increased potential to later evolve increased phenotypic complexity in small populations through the larger genomes generated by increased genetic drift [3, 4]. As phenotypic traits are strongly beneficial in the Avida environment used here, these small populations used slightly deleterious genome expansions to cross fitness valleys and eventually reach novel fitness peaks.
Our work also shows that this evolutionary advantage of small populations is limited by an increased rate of population extinction. Such a trend between the evolution of large genomes and an increased rate of extinction is seen in some multicellular eukaryote clades [47, 48]. These small populations are still likely to have a larger risk of extinction beyond that caused by population-genetic risks such as Muller’s ratchet [49] and mutational meltdowns [50, 51]. Ecological stressors increase extinction risk [52] and small populations are less able to adapt to detrimental environmental changes [53]. Our results concerning extinction, combined with the risk of other factors not examined here, suggest that the likelihood of a small population using genetic drift to evolve greater complexity without an increased risk of extinction may be limited. However, it is possible that multiple small populations could reduce the risk of extinction without reducing the evolution of complexity; future work should consider the interplay between population size and the evolution of complexity within a metapopulation of small populations.
Large populations also evolved greater genome sizes and phenotypic complexity. In our original experiments, genome evolution in large populations was driven by the fixation of rare beneficial insertions (Fig 4). While it is likely that many gene duplications are not under positive selection and lost due to genetic drift and mutation accumulation [54], some, especially those resulting in the amplification of gene expression, can be immediately beneficial and later lead to increased phenotypic complexity [55–58]. Due to the increased mutation supply, these events would occur at a greater frequency in large populations [59] and possibly lead to an increased probability of the evolution of complexity there. However, we also found that large populations did not require this large supply of beneficial insertions. Even when insertion mutations added non-functional instructions and further point mutations were required to evolve functional traits, large populations still evolved complexity similar to that evolved in small populations. These results suggest that purifying selection may not limit the evolution of complexity in large populations. Finally, we found that when deletions occur at a much greater frequency than insertions, only large populations have an evolutionary advantage towards complexity. As many bacteria do have a bias towards deletions [60, 61], this result suggests that large microbial populations can have an evolutionary advantage over small microbial populations for evolving novel traits after all.
Such a trend where both large and small, but not intermediate-sized populations have an evolutionary advantage has already been theoretically proposed elsewhere. Weissman et al. showed that both small and large populations cross fitness valleys more easily than intermediate-sized populations [34]. Small populations valley-crossed due to genetic drift and large populations did so due to an increased supply of double mutants. Ochs and Desai also showed that intermediate-sized populations evolved to a lower fitness peak compared to small or large populations when valley-crossing was required for reaching a higher peak [36]. We found similar results, but from different evolutionary mechanisms. Here, populations needed to increase in genome size in order to evolve phenotypic complexity. Additionally, our populations evolved in a complex fitness landscape with many different possible paths to phenotypic complexity. While small populations did fix deleterious insertions to increase genome size, large populations evolved on a different path, either through beneficial insertions (Fig 3) or neutral insertions (S4 Fig). It is possible that even larger populations than those evolved here would fix more deleterious insertions, as the likelihood of a further, beneficial mutation arising on the background of a segregating deleterious mutation increases as population size increases. However, our results emphasize that large populations may not be dependent on valley-crossing in some fitness landscapes if alternative evolutionary trajectories exist, even if these trajectories are rare. While the first maps of fitness landscapes suggested mutational paths are small in number [62], more recent work suggests that many indirect evolutionary trajectories exist in larger fitness landscapes [63].
The population sizes that led to the evolution of greater phenotypic complexity via drift are very small (10 individuals). As biological populations of that size are unrealistic, we may wonder whether such populations can actually evolve greater complexity due to increased genetic drift. However, there are reasons to believe that these results would generally hold for biological systems. The limited range of small population sizes that led to complexity is an Avida-specific result due to the severe fitness effect of insertion mutations in avidians with small sequence length. We found that for those sequences, most insertions are lethal (about 80%), and the rest are significantly detrimental, of the order 10% to 90%. To overcome a detrimental effect of 20% via drift, populations must be as small as N = 10. Insertion mutations in biological genomes are not nearly as detrimental, and therefore the critical population size to see evolution of complexity via drift is much larger. In E. coli, for example, the deleterious effect of insertions is between 1% and 3% [64]. We can therefore expect to see the effect of increased complexity due to drift in biological populations that are small, but not unreasonably small.
Another possible avenue for future work suggested by this study is to use a simpler population genetics model to explore the same questions we attempted to answer here. Many previous theoretical studies have examined the relevance of valley-crossing to the evolution of complex traits in simple fitness landscapes [34–36]. One benefit of a simpler model is that it allows for a broader exploration of the relevant parameters involved in the interplay between population size, genome size, and the evolution of phenotypic complexity. While we were not able to perform large parameter searches using the Avida system, our work here establishes a possible relationship between the factors that influence the evolution of complexity in a fitness landscape with many possible mutational trajectories to novel traits [65]. These results should drive future theoretical studies on the evolution of genome size and phenotypic complexity using population genetics models with simpler fitness landscapes.
Here we studied the evolution of complexity in haploid asexual digital organisms with an ancestral minimal genome on a frequency-independent fitness landscape. While beyond the scope of this work, it is worth considering how adjusting these genotype characteristics would alter our results. It is likely that the ancestral minimal genomes are a requirement for small populations to evolve the same number of novel traits as large populations. If the ancestor organism had a significant amount of non-functional genome content, the mutation supply advantage that large populations have should result in an accelerated rate of phenotypic evolution in large populations [66]. The organisms used here, as in all Avida experiments, are haploid. It is possible that polyploidy would alter the results found here. However, the implementation of a ploidy cycle in Avida is non-trivial due to the mechanistic style of replication, and so presently other experimental systems would have to be used to explore the role of ploidy in the evolution of phenotypic complexity.
It is unclear how sexual, instead of asexual, reproduction would change the results. While sexual reproduction can enhance adaptation by combining beneficial mutations that arise in different background, it can also break up beneficial combinations of mutations [67]. One result that may be altered by sexual reproduction is the rate of extinction in small populations, as sex has been found to reduce the rate of mutational meltdowns [68]. Weissman et al. also demonstrate that the large population advantage towards valley-crossing does not exist under high recombination rates [35]. Sexual reproduction has previously been studied using Avida, but it is more akin to homologous recombination in bacteria [69] (as there is no ploidy cycle). Future work should address the role of sexual recombination on the results shown here. Finally, the experiments performed here had no frequency-dependent fitness effects. Previous Avida studies showed that frequency-dependent interactions enhanced the evolution of complexity for a given population size [70, 71]. It is worth exploring how the presence of frequency-dependent selection alters the evolution of complexity, especially in small populations. The benefits of the diversity seen in frequency-dependent fitness landscapes may be reduced in small populations. The extensions to the experiments performed here would provide a more complete understanding of the role of adaptive and non-adaptive evolutionary processes in the origins of complexity.
In order to experimentally test the role of population size and genetic drift in the evolution of complexity, we used the digital evolution system Avida version 2.14 [42]. In Avida, self-replicating computer programs (avidians) compete in a population for a limited supply of CPU (Central Processing Unit) time needed to successfully reproduce. Each avidian consists of a circular haploid genome of computer instructions. During its lifespan, an avidian executes the instructions that compose its genome. After executing certain instructions, it begins to copy its genome. This new copy will eventually be divided off from its mother (reproduction in most Avida experiments is asexual). Because an avidian passes on its genome to its descendants, there is heredity in Avida. As an avidian copies its genome, mutations may occur, resulting in imperfect transmission of hereditary information. This error-prone replication introduces variation into Avida populations. Finally, avidians that differ in instructions (their genetic code) also likely differ in their ability to self-replicate; this results in differential fitness. Therefore, because there is differential fitness, variation, and heredity, an Avida population undergoes evolution by natural selection [72]. This allows researchers to perform experimental evolution in Avida as in microbial systems [19, 73]. Avida has been successfully used as a model system to explore many topics concerning the evolution of complexity [2, 65, 71, 74, 75].
Twenty-six different instructions compose the Avida instruction set (see [42] for a more complete overview). These include instructions for genome replication, such as an instruction to allocate memory for a new daughter genome, an instruction to copy instructions from the mother genome into the daughter genome, and an instruction to divide off the new avidian. There are instructions that allow for the input, output, and manipulation of random numbers that are used in the performance of certain Boolean logic calculations (see below). There are also instructions for altering instruction execution, including conditional instructions and instructions for changing the next instruction location in the genome to be executed. It is important to note that the Avida instruction set was not designed to mimic any biological organism. Instead, it was created in order to have an organism with mechanistic reproduction in a non-specified fitness landscape that allows for studies of evolutionary dynamics.
The Avida world consists of a toroidal grid of N cells, where N is the (maximum) population size. When an avidian successfully divides, its offspring is placed into a cell in the population. While the default setting places the offspring into one of nine neighboring cells of the parent, here the offspring is placed into any cell in the entire population. This simulates a well-mixed environment without spatial structure. When there are empty cells in the population, new offspring are preferentially placed in an empty cell. However, if the population is at its carrying capacity, the individual who is currently occupying the selected cell is replaced by the new offspring (a new individual can also eliminate its parent if that cell is selected). This adds an element of genetic drift into the population as the individual to be removed is selected without regard to fitness. A population can also decrease in size by the death of individuals. An avidian will die without producing offspring if it executes 20L instructions without successfully undergoing division, where L is the avidian’s genome size. This can lead to population extinction in very small populations.
Time in Avida is divided into updates, not generations. This method of keeping time was implemented in order to allow individuals to execute their genomes in parallel. During one update, a fixed number of instructions is executed across the entire population. The resource that is necessary to execute instructions (the CPU “energy”) is measured in SIPs (single instruction processing) units. By default, there are 30N SIPs available to the entire population per update, where N is the population size. SIPs are distributed among the individual genotypes within a population in proportion to the trait or traits displayed by an individual. The total amount of SIPs garnered by an individual from traits is called the “merit”. In a homogeneous population of one genotype (clones) where each individual has the same merit, each individual will obtain approximately 30 SIPs per update. However, in a heterogeneous population where merit differs between individuals, SIPs will be distributed in an uneven manner. That way, individuals with a greater merit will execute and/or replicate a larger proportion of their genome per update and replicate faster, thus having a greater fitness. This places a strong selection pressure on evolving a greater merit. One generation has passed when the population has produced N offspring. Typically (depending on the complexity of an avidian) between 5 and 10 updates pass in one generation.
A genotype’s merit is increased through the evolution of certain phenotypic traits that form a “digital metabolism” [37]. These phenotypic traits are the ability (or lack there-of) to perform certain Boolean logic calculations on random binary numbers that the environment provides. To do this, an avidian must have the right “genes”–in this case, the right sequence of instructions. First, during an avidian’s lifespan, instructions that allow for the input and output of these random binary numbers must be executed. Further instructions should manipulate those numbers so as to perform the rewarded computations. When a number is then written to the output, the Avida program checks to see whether a logic operation was successfully performed. If so, the the individual that performed the computation consumes a resource tied to the performance of that trait (there are many different codes, that is, combinations of instructions, that will trigger the reward). Resource consumption causes the offspring of that individual to have their merit modified by a factor set by the experimenter. Here, we use the “Logic-9” environment to reward the performance of nine one- and two-input logic functions [65]; see S1 Table for the names and specific rewards of each function. Each individual only gains a benefit from performing each function once per generation. There is an infinite supply of resources for the performance of each logic function in the present experiments, making fitness frequency-independent. Because the performance of these logic functions increases merit, they also increase fitness and are under strong positive selection.
While increases in an individual’s merit increase replication speed and thus the individual’s fitness, fitness in Avida is implicit and not directly calculated. Unlike simulations of evolutionary dynamics, a genotype’s fitness is thus not set a priori by the experimenter. The only way to measure the fitness of an avidian is to run it through its lifecycle and examine its phenotype. This is similar in principle to how bacterial fitness cannot be calculated by examining an individual bacterium’s genome, but must be measured through a number of different experiments, such as competition assays [76]. A genotype’s fitness is determined by how many offspring it can produce per unit time. Genotypes that can reproduce faster will out-compete other genotypes, all else being equal. Therefore, evolution will increase a population’s fitness through two means. The first is that the population will evolve individuals with a greater number of phenotypic traits and thus with a greater merit, as explained above. The second way to increase replication speed is by optimizing (shortening) the replication time. This occurs either by shrinking the genome, which results in fewer instructions that need to be copied and replicated, or by optimizing genome architecture for faster replication. Fitness w in Avida can be estimated by the following equation:
w ≈ merit replication time (1)
For an avidian to be able to successfully reproduce, it must first allocate memory for the new individual, copy its genome into the allocated memory space, and then divide off the daughter organism. As instructions are copied, the avidian may inaccurately copy some instructions into the newly allocated memory at a rate set by the experimenter. Additionally, upon division, insertions and deletions of a single instructions occur at (possibly different) rates set by the experimenter. Finally, larger insertions or deletions (indels) can occur when an avidian divides into two daughter genomes if the division occurs unevenly. In most cases, this results in the creation of one larger and one smaller genome and both of these are non-viable. However, in rare cases, one of these new genotypes is able to reproduce, resulting in a large change in genome size in that individual’s descendants. Because this mutation through inaccurate division is a characteristic of a genome and thus emergent, the rate at which it occurs is not set by the experimenter.
We used four experimental designs (treatments) to explore how population size determines the evolution of complexity: the original experiments, the non-functional insertion experiments, the fixed genomic mutation rate experiments, and the deletion bias experiments. For all experiments, we evolved populations of size N = {10,100,1000,10000} for 2.5 × 105 generations under 100-fold replication. For the original treatment, we also performed experiments with population sizes of N = {20,30,40,50,60,70,80,90}. All populations were initiated at full size N with an altered version of the standard length-100 Avida start organism [42]. The alteration was the removal of all non-essential genome content (85 nop-c instructions). This reduced the genome size of the ancestor organisms from 100 instructions to only 15 instructions.
For the original experiments, point mutations occurred at a rate of 0.01 mutations per instruction copied, and insertions and deletions at 0.005 events per division. Insertions and deletions occur at most once per division. The ancestor thus started with a genomic mutation rate of 0.15 mutations per generation (0.01 mutations/instruction copied × fifteen instructions copied per generation), but this changes over the course of the experiment as genome size evolves. These experiments are similar to most standard Avida experiments, with the exception of a smaller genome size (fifteen instructions) for the ancestral organism.
For the remainder of the experimental settings, one of the above settings was changed to examine a specific effect. For the experiments where the genomic mutation rate was fixed, point mutations occurred at a rate of 0.15 mutations per division, independently of genome size, which fixes the mutation rate at 0.15 mutations/genome/generation. For the non-functional insertion experiments, the mutation rates were the same as in the original experiments. However, instead of inserting one of the twenty-six instructions from the Avida instruction set (see [42] for the Avida instruction set), “blank” instructions called nop-x were inserted. These instructions have no function on their own or in combination with any other instruction. Finally, for the deletion bias experiments, point mutations occurred at the same rate as in the standard experiments. However, insertions and deletions did not occur at the same rate. Insertions occurred at a rate of 0.001 per division and deletions occurred at a rate of 0.009 per division. This kept the total mutation rate equal to the other experimental treatments, while altering the ratio of insertions to deletions.
In order to analyze the evolution of complexity in each population, we extracted the individual with the greatest fitness at the end of each experiment (the “dominant” type). We then calculated relevant statistics for each of these genotypes by running them through Avida’s analyze mode. This mode allows us to run each genotype through its lifecycle in isolation, and calculate its fitness, its genome size, whether it performs any logic functions, and whether it produces viable offspring, among other characteristics. To measure the evolution of phenotypic complexity, we determined how many unique logic calculations each genotype could perform. Such a measure of complexity is similar to a measure of phenotypic complexity used previously [5] in population genetics. The relative fitness was calculated by dividing the analyzed fitness value by the ancestor’s fitness (0.244898).
To examine why certain population sizes evolved larger genomes, we examined the “line of descent” (LOD) of the fittest type [65]. An LOD contains every intermediate genotype between the final individual with the greatest fitness and the ancestral genotype that initialized each population. This line provides a perfect “fossil record” to examine all of the mutations, insertions, and deletions that led to the final fittest genotype for each population. We also calculated the selection coefficient s for each mutation, defined as the ratio of the offspring’s fitness to the parent’s fitness minus one. We defined beneficial mutations as those with s > 0 and deleterious mutations as those with s < 0 (this ignores classifying slightly beneficial and slightly deleterious mutations as neutral.) We determined the number of beneficial insertion mutations by counting those insertions on the LOD with s > 1 N, where N is the population size. These are beneficial mutations that are not nearly-neutral and hence should be under positive selection. We note that using s > 1 N is only an approximation, as the equation for a nearly neutral mutation is | s | ≪ 1 N e, where Ne is the effective population size [77]. We also examined those mutations that had a slightly deleterious effect on fitness, i.e., those whose selection coefficient was - 1 N < s < 0.
|
10.1371/journal.pntd.0005120 | Identifying Early Target Cells of Nipah Virus Infection in Syrian Hamsters | Nipah virus causes respiratory and neurologic disease with case fatality rates up to 100% in individual outbreaks. End stage lesions have been described in the respiratory and nervous systems, vasculature and often lymphoid organs in fatal human cases; however, the initial target organs of Nipah virus infection have not been identified. Here, we detected the initial target tissues and cells of Nipah virus and tracked virus dissemination during the early phase of infection in Syrian hamsters inoculated with a Nipah virus isolate from Malaysia (NiV-M) or Bangladesh (NiV-B).
Syrian hamsters were euthanized between 4 and 48 hours post intranasal inoculation and tissues were collected and analyzed for the presence of viral RNA, viral antigen and infectious virus.
Virus replication was first detected at 8 hours post inoculation (hpi). Nipah virus initially targeted type I pneumocytes, bronchiolar respiratory epithelium and alveolar macrophages in the lung and respiratory and olfactory epithelium lining the nasal turbinates. By 16 hpi, virus disseminated to epithelial cells lining the larynx and trachea. Although the pattern of viral dissemination was similar for both virus isolates, the rate of spread was slower for NiV-B. Infectious virus was not detected in the nervous system or blood and widespread vascular infection and lesions within lymphoid organs were not observed, even at 48 hpi.
Nipah virus initially targets the respiratory system. Virus replication in the brain and infection of blood vessels in non-respiratory tissues does not occur during the early phase of infection. However, virus replicates early in olfactory epithelium and may serve as the first step towards nervous system dissemination, suggesting that development of vaccines that block virus dissemination or treatments that can access the brain and spinal cord and directly inhibit virus replication may be necessary for preventing central nervous system pathology.
| Nipah virus is a highly fatal paramyxovirus that causes respiratory and neurologic disease with widespread vascular damage. Although end stage disease has been characterized in humans and in animal models, the early phase of infection has yet to be described. Thus, it is known where the virus replicates during the final stages of infection, but not how the virus is transported to these end stage tissues or which tissues are targeted first. Using Syrian hamsters that were intranasally inoculated with virus as a model of Nipah virus infection, we show here that Nipah virus initially targets the respiratory tract, specifically the nasal cavity and lung, with virus replication being detected as early as 8 hours after inoculation. During the first 48 hours of infection, the virus slowly penetrates into the underlying tissues of the respiratory tract, but viremia and virus replication in the brain are not detected. In the lung, Nipah virus initially infects epithelial cells lining airways, followed by spread to arterial smooth muscle cells, suggesting that localized spread of the virus within an organ can lead to viral dissemination into the vasculature. Likewise, replication in the olfactory epithelium likely precedes spread of Nipah virus along the olfactory nerve into the brain. Understanding how Nipah virus spreads through the host to cause encephalitis is important for developing effective countermeasures that can target Nipah virus replication and prevent viral dissemination.
| Nipah virus, a highly virulent paramyxovirus, was first identified in an outbreak in Malaysia and Singapore in 1998 and has since caused almost yearly outbreaks in humans in Bangladesh [1–6]. Nipah virus causes encephalitis and respiratory disease; however, the prevalence of respiratory symptoms and case fatality rates have varied among the reported outbreaks. Humans infected during the initial outbreak in Malaysia mainly exhibited neurologic symptoms while some also developed respiratory disease [7–10]. Outbreaks in Bangladesh have also resulted in neurologic disease, yet the incidence of respiratory disease has been higher, as have the case fatality rates [3, 11–13]. A definitive cause for the reported differences between these outbreaks has not yet been determined.
The pathology of Nipah virus in humans has only been described in fatal cases from the Malaysia outbreak; lesions were noted in the central nervous system, lung, vasculature and to a lesser extent the spleen and lymph nodes [1, 7, 9]. Nipah virus has been shown to affect the nervous, respiratory, vascular and immune systems, to varying degrees, by the end stage of disease in both humans and experimentally inoculated animals, including Syrian hamsters, ferrets, cats, guinea pigs and African green monkeys. Animal models have shown that by the late stages of disease, Nipah virus generally infects epithelial cells in the upper and lower respiratory tracts, endothelium and smooth muscle cells in arteries, neurons and mononuclear leukocytes regardless of inoculation route [14–20]. Although the end stage of Nipah virus disease is well characterized, the exact sequence and mechanism of virus dissemination through these different organ systems and cell types remains poorly defined for Nipah virus isolates from both Malaysia (NiV-M) and Bangladesh (NiV-B). Identifying the initial target organs and target cell types is vital to understanding the pathogenesis of Nipah virus and in potentially preventing virus dissemination.
Here, we identified the initial target tissues and cell types for Nipah virus during the first 48 hours post intranasal inoculation of Syrian hamsters. Syrian hamsters were used to model Nipah virus infection since they develop late stage lesions similar to those in humans [19–25]. Although both intranasal and intraperitoneal routes of Nipah virus inoculation have been described in Syrian hamsters, the intranasal route was chosen for this study since it more closely represents a potential natural route of Nipah virus infection in humans. In the Syrian hamsters, we showed that the lung and nasal turbinates were the initial target organs and that Nipah virus replication could be identified by 8 hours post inoculation (hpi) in type I pneumocytes, bronchiolar respiratory epithelial cells and alveolar macrophages and by 16 hpi in the nasal cavity respiratory and olfactory epithelial cells. Virus then appeared to disseminate to the trachea and larynx. We did not detect infectious virus or virus replication in the central nervous system or peripheral blood and there was no evidence of widespread viral infection of blood vessels, even at 48 hpi. These results suggest that Nipah virus first targets the respiratory system and that widespread virus dissemination and vascular and nervous system infection occur later.
All animal experiments were approved by the Institutional Animal Care and Use Committee of the Rocky Mountain Laboratories (RML; ASP#2014-088-E)) and carried out by certified staff in an Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) International accredited facility, according to the institution’s guidelines for animal use, and followed the guidelines and basic principles in the United States Public Health Service Policy on Humane Care and Use of Laboratory Animals (available from http://grants.nih.gov/grants/olaw/references/PHSPolicyLabAnimals.pdf), and the Guide for the Care and Use of Laboratory Animals (available from https://grants.nih.gov/grants/olaw/Guide-for-the-Care-and-use-of-laboratory-animals.pdf). All infectious work with Nipah virus was performed in a biosafety level 4 (BSL4) laboratory in the Integrated Research Facility at RML. Sample inactivation and removal of samples from the BSL4 laboratory was performed according to established standard operating procedures [26] approved by the Institutional Biosafety Committee (IBC).
Nipah virus isolates from Bangladesh and Malaysia were kindly provided by the Special Pathogens Branch of the Centers for Disease Control and Prevention, Atlanta, GA. NiV-B was isolated from a throat swab collected from a fatal human case in 2004 and had been passaged three times in VeroE6 cells. NiV-M was isolated from the cerebrum of a fatal human case in 1999 and had been passaged four times in VeroE6 cells.
Two groups of twenty-four 6- to 8-week-old female Syrian hamsters (HsdHan:AURA; Harlan Laboratories, Haslett, MI) were intranasally inoculated with 5 x 106 TCID50 (50% tissue culture infectious dose) of either NiV-B or NiV-M in a total volume of 80 μl (40 μl per nostril). All hamsters were evaluated daily for clinical signs of disease. Four hamsters from each group were euthanized at 4, 8, 16, 24, 32 and 48 hpi. A terminal heart blood sample was collected from each hamster before necropsy. The nasal cavity, larynx, trachea, lung, cervical lymph nodes, spleen, brain and spinal cord were collected for histologic and virologic analysis.
Necropsies and tissue sampling were performed according to a standard protocol approved by the IBC. Tissues were fixed for a minimum of 7 days in 10% neutral-buffered formalin and embedded in paraffin. The sections through the nasal turbinates, which were contained within the nasal cavity, and the spinal cord, which was contained within the vertebrae, were decalcified using a 20% EDTA solution in sucrose prior to paraffin embedding.
Leukocytes were isolated from terminal blood samples using centrifugation over a histopaque gradient (Sigma-Aldrich, St. Louis, MO) in conjunction with erythrolysis using ACK lysing buffer (Thermo Fisher Scientific, Waltham, MA) according to manufacturer instructions. The resulting leukocyte pellets were fixed for a minimum of 24 hours in 10% neutral-buffered formalin. Leukocytes from hamsters inoculated with the same Nipah virus isolate and which were euthanized at the same time point were pooled, then processed in HistoGel (Thermo Fisher Scientific) according to manufacturer instructions and embedded in paraffin to form a cell block.
Routine hematoxylin and eosin (H&E) staining, immunohistochemistry (IHC) and in situ hybridization (ISH) were performed on tissue sections and cell blocks. Nipah virus antigen was detected by IHC; tissue sections were labeled with a rabbit polyclonal antiserum against Nipah virus nucleoprotein (1:5000; kindly provided by L. Wang, Duke-NUS Medical School, Singapore) [27]. Nipah virus replication was detected in tissue sections by ISH using probes specific for positive sense Nipah virus nucleoprotein RNA using a previously described method [28]. All slides were evaluated by a board certified veterinary anatomic pathologist.
Viral RNA was isolated from hamster tissues using the RNeasy Mini kit (Qiagen, Valencia, CA) or from hamster blood using the QIAamp Viral RNA Mini kit (Qiagen), according to manufacturer instructions. 5 μl of RNA was used in a one-step real-time RT-PCR targeting the Nipah virus nucleoprotein, as described previously [25], using the QuantiFast kit (Qiagen) according to manufacturer instructions. In each run, standard dilutions of RNA extracted from a titered virus stock were run in parallel, to calculate TCID50 equivalents in the samples.
Virus titrations were performed by end-point titration in Vero C1008 cells. Vero C1008 cells were inoculated with tenfold serial dilutions of tissue homogenates. One hour after inoculation of cells with tissue homogenates, the inoculum was removed and replaced with 200 μl DMEM (Sigma-Aldrich) supplemented with 2% fetal bovine serum (HyClone, Logan, UT), 1 mM L-glutamine (Lonza, Walkersville, MD), 50 U/ml penicillin and 50 μg/ml streptomycin (Thermo Fisher Scientific). Five days after inoculation with tissue homogenates from NiV-B inoculated hamsters and three days after inoculation with tissue homogenates from NiV-M inoculated hamsters, when full cytopathic effect (CPE) was reached in Vero cells, CPE was scored and the TCID50 was calculated from 4 replicates by the Spearman-Karber method [29].
To determine which tissues Nipah virus initially targets, we intranasally inoculated Syrian hamsters with 5 x 106 TCID50 of NiV-M and analyzed tissues from the respiratory tract, central nervous system and immune system for the presence of Nipah virus RNA and infectious virus at multiple time points up to 48 hpi. Although viral RNA was detected by qRT-PCR in the lung, nasal turbinates, trachea, larynx, cervical lymph nodes and brain, but not the spleen and spinal cord at 4 hpi, infectious virus could be isolated only from the lung and nasal turbinates by 8 hpi, and from the trachea and larynx by 16 hpi (Fig 1). ISH using probes that targeted the positive sense RNA of the Nipah virus nucleoprotein, which is only observed when virus replication occurs, first detected virus replication at 8 hpi in the lung and 16 hpi in the nasal turbinates (Fig 2, Table 1). The lack of positive sense viral RNA at 4 hpi in all tissues examined, in addition to the decrease in detected mean viral loads in most tissues between 4 and 8 hpi, suggests that viral RNA detected at 4 hpi by qRT-PCR represented residual administered viral inoculum. Virus replication was not identified in the central nervous system or lymphoid organs at any of the time points tested. The absence of positive sense viral RNA and infectious virus from all tissues except the lung and nasal turbinates at 8 hpi suggests that the lung and nasal cavity were the initial target tissues for Nipah virus.
To assess which cell types in the lung and nasal turbinates Nipah virus initially binds to and replicates in, we examined these tissues for the presence of positive sense viral RNA using ISH and Nipah virus nucleoprotein antigen using IHC. Positive sense viral RNA, indicating Nipah virus had bound to, infected and replicated in a cell, was first detected at 8 hpi in the lung, where it was observed in type I pneumocytes, bronchiolar respiratory epithelium and alveolar macrophages in 2 out of 4 hamsters (Table 1). As early as 8 hpi, the location of viral antigen on IHC labeled slides was often shown to correspond to foci of acute minimal bronchointerstitial pneumonia that were identified on H&E stained slides. IHC was used to track virus dissemination throughout the lung between 4 and 48 hpi; increasing amounts of viral antigen were detected at subsequent time points (Fig 3, Table 2). By 16 hpi, Nipah virus had disseminated from alveoli and bronchioles to larger airways and viral antigen and virus replication were observed in the bronchial respiratory epithelium in 2 out of 4 hamsters. Within bronchioles, viral antigen spread from the bronchiolar epithelium to the underlying bronchiolar smooth muscle in 1 out of 4 hamsters at 16 hpi. Dissemination of viral antigen to arterial smooth muscle cells in the lung was only noted at 32 and 48 hpi in 1 out of 4 hamsters and affected one to a few small-caliber arteries in areas exhibiting bronchointerstitial pneumonia; viral antigen was not observed in the endothelial cells of these vessels.
In the nasal cavity, virus replication was first observed in respiratory and olfactory epithelium lining the nasal turbinates at 16 hpi (Table 1). Tracking of viral antigen in the nasal cavity over time showed that respiratory and olfactory epithelium were infected first, followed by the submucosal gland epithelium underlying both respiratory and olfactory epithelial cells at 24 hpi (Fig 3, Table 2). A mild neutrophilic rhinitis was identified on H&E slides as early as 16 hpi. Foci of rhinitis were often observed in locations where virus antigen was detected by IHC. Viral antigen was not observed in the vascular smooth muscle cells or endothelium in the nasal turbinates at any time point.
Although infectious virus was detected at 8 hpi in the nasal cavity and lung of hamsters intranasally inoculated with NiV-M, it was not until later time points that Nipah virus was detected in the trachea and larynx. At 16 hpi, positive sense viral RNA, as analyzed by ISH, and Nipah virus antigen, detected by IHC, were identified in the trachea in 2 out of 4 hamsters (Fig 4). Infectious virus was isolated from both the trachea and larynx by 16 hpi (Fig 1). However, viral antigen was first detected in the larynx at 32 hpi and positive sense viral RNA was not identified in this tissue at any time point examined (Fig 4) despite detection of infectious virus and viral RNA, by qRT-PCR (Fig 1). As ISH and IHC are evaluated histologically, it is possible that positive sense viral RNA and viral antigen may have also been present in the larynx at 16 hpi, but were not found in the exact tissue sections that were labeled and examined. The lack of virus replication, infectious virus and viral RNA at 8 hpi in both of these tissues suggests that Nipah virus does not initially target the trachea and larynx. When viral antigen and virus replication were present in the trachea and larynx, they were identified in epithelial cells lining these airways and were not observed in blood vessels, suggesting that Nipah virus secondarily spreads from one or both of its initial target tissues, the nasal cavity and lung, to the trachea and larynx through the airways. Replication of virus in the nasal cavity and lung leads to increasing virus titers in these tissues and potentially increased numbers of virus particles disseminating to the larynx and trachea.
Since lesions have occasionally been identified in the spleen and lymph nodes of fatal human cases of Nipah virus infection [7], we examined these lymphoid organs for the presence of infectious virus, viral RNA and viral antigen in hamsters. Although viral RNA could be detected by qRT-PCR in cervical lymph nodes of hamsters from 4 hpi onwards, infectious virus was only detected at 48 hpi in the cervical lymph nodes of 2 out of 4 hamsters (Fig 1). Viral RNA was detected sporadically in the spleen at early time points and infectious virus was detected in the spleen of 1 out of 4 hamsters at 48 hpi (Fig 1). Positive sense viral RNA and viral antigen were not observed in the lymphoid organs at any time point.
To determine if Nipah virus had disseminated to the central nervous system, we analyzed both the brain and spinal cord of NiV-M inoculated hamsters for the presence of viral RNA, infectious virus and viral antigen. Low viral loads were detected by qRT-PCR in the brain and spinal cord of a few hamsters, with an increase in mean viral loads between 32 and 48 hpi (Fig 1). However, neither infectious virus (Fig 1), viral antigen, nor positive sense viral RNA were identified in the central nervous system tissues at any time point.
It has been shown that Nipah virus can bind to human lymphocytes and hamster mononuclear leukocytes which may then transfer the virus to permissive cells, such as endothelial cells, thereby potentially resulting in systemic viral dissemination [30]. To determine if Nipah virus binds to or replicates in circulating leukocytes during the early stage of infection, we performed IHC and ISH on cell blocks made up of leukocytes collected during terminal heart bleeds. Viral antigen and positive sense viral RNA were not observed in leukocytes in the peripheral blood at any time point. Additionally, viral RNA was not detected in the peripheral blood by qRT-PCR. These results suggest that Nipah virus either did not bind to, or replicate in, leukocytes circulating in blood or the level of virus in the blood was below the limit of detection. Furthermore, even at 48 hpi, viral antigen and positive sense viral RNA were not identified in the vasculature of any organ examined, other than the lung. Additionally, lesions associated with Nipah virus infection of blood vessels, including vasculitis, fibrinoid change and fibrin thrombi, were not detected histologically at any time point in any organ, except the lung. In the lung, viral antigen was only detected in the vascular wall of blood vessels located in foci of bronchointerstitial pneumonia. This suggests that the virus first has to infect and replicate in superficial cells lining the respiratory tract before the virus infiltrates deeper structures such as blood vessels (Fig 5), where it appears to invade from the outer vascular wall towards the endothelium and then finally enter the bloodstream, which may explain why viremia was still below the detection limit at 48 hpi.
Since the human case fatality rates and prevalence of respiratory disease were different between Nipah virus outbreaks in Malaysia and Bangladesh [8, 10, 12, 13], we sought to determine whether there were differences in the early pathogenesis of NiV-M and NiV-B infections.
Similar to hamsters inoculated with NiV-M, viral RNA was detected by qRT-PCR in the respiratory and central nervous systems and cervical lymph nodes at 4 hpi in hamsters intranasally inoculated with NiV-B (S1 Fig). Despite the detection of viral RNA in tissues from multiple organ systems at 4 hpi, positive sense viral RNA, indicative of virus replication, was only detected in the nasal turbinates and lung from 8 hpi onwards (S1 Fig and S2 Fig). Virus replication was observed at 8 hpi in type I pneumocytes and alveolar macrophages in 4 out of 4 hamsters and in bronchiolar respiratory epithelium in 2 out of 4 hamsters (S1 Fig, S2 Fig and Fig 5). Similar to NiV-M inoculated hamsters, the respiratory and olfactory epithelium in the nasal turbinates were also early targets for virus replication (S1 Table, S2 Fig and Fig 5). The presence of Nipah virus antigen in the nasal turbinates and lung, as detected by IHC, mirrored what was detected by ISH (S2 Table, S3 Fig). Unlike NiV-M inoculated hamsters, dissemination of viral antigen to pulmonary bronchiolar and arterial smooth muscle cells was not observed, even at 48 hpi. Spread of viral antigen in the nasal cavity from the respiratory and olfactory epithelial cells to the underlying submucosal gland epithelial cells took longer in NiV-B inoculated hamsters and was not detected until 48 hpi, compared to 24 hpi in NiV-M inoculated hamsters. Similar to NiV-M, NiV-B appeared to spread from the nasal cavity or lung to the trachea and larynx (S4 Fig and Fig 5). Widespread vascular dissemination was not detected, nor was infectious virus, viral antigen or virus replication identified in the brain, spinal cord or lymphoid organs at any time point in NiV-B inoculated hamsters. Additionally, viral antigen, virus replication and viral RNA were not detected in peripheral blood leukocytes at any time point.
In this study, we identified the early target tissues and cells of Nipah virus and evaluated viral dissemination during the early stages of infection for both NiV-M and NiV-B in intranasally inoculated Syrian hamsters. Previous Nipah virus studies in Syrian hamsters have utilized either an intranasal or intraperitoneal route of inoculation and have shown that end stage lesions in Syrian hamsters inoculated by either route are similar [21, 22], suggesting that organ tropism is not affected by inoculation route. Moreover, the intranasal inoculation route likely represents a more natural route of inoculation for humans, as compared to intraperitoneal inoculation. In our study, the nasal cavity and lung were the initial target tissues for both virus isolates. Within these tissues, Nipah virus initially exhibited epitheliotropism and, in the lung, a predilection for alveolar macrophages. We showed that virus replication, as indicated by the presence of positive sense viral RNA, could be identified by 8 hpi in the lung in type I pneumocytes, bronchiolar respiratory epithelial cells and alveolar macrophages. The presence of positive sense viral RNA in alveolar macrophages suggests virus replication occurred early in this cell type; however, phagocytosis is a major function of alveolar macrophages and if virus replication in epithelial cells began between the 4 and 8 hpi time points, it cannot be ruled out that the positive sense viral RNA detected in alveolar macrophages simply represented phagocytosis of virus infected epithelial cells [31].
In the nasal cavity and lung, virus infection and replication were first identified in epithelial cells that lined air spaces. Once in the nasal cavity and lung, Nipah virus disseminated from superficial epithelial cells to adjacent underlying cells, including bronchiolar smooth muscle cells, arterial smooth muscle cells in the lung and submucosal gland epithelial cells in the nasal cavity. Interestingly, our results showed that in the early phase of infection the spread of Nipah virus from epithelial cells that lined airways in the lung and nasal cavity to underlying cells appeared to be faster in hamsters inoculated with NiV-M than with NiV-B, suggesting that development of severe disease may occur faster for NiV-M. However, in studies evaluating late stage Nipah virus disease in Syrian hamsters, results were contradictory as to whether disease progression in the respiratory system was faster for NiV-M or for NiV-B [20, 21]. Additionally, end stage lesions in the respiratory tract appeared to plateau at the same severity level by 4 days post inoculation in Syrian hamsters inoculated with either Nipah virus isolate [20]. Similar studies analyzing late stage disease have been performed in ferrets and African green monkeys. Inoculation of ferrets with either NiV-M or NiV-B resulted in comparable terminal lesions and clinical signs, despite differences in oral viral shedding and viremia [16], while in African green monkeys NiV-B appeared to replicate more quickly and end stage respiratory lesions were more severe in animals inoculated with NiV-B than NiV-M [32]. The results from the ferret and hamster models of Nipah virus suggest that differences in the rate of virus replication or dissemination noted prior to end stage disease may not in fact significantly impact the overall clinical disease severity or outcome, while the African green monkey model suggests otherwise and further evaluation of the early pathogenesis of Nipah virus in African green monkeys is needed to determine if differences exist between these models.
Although widespread viral infection of blood vessels with development of histologic lesions has been reported in fatal human cases [1, 7], this was not identified at any time point in the hamsters in this experiment, implying that widespread vascular involvement does not occur during the early phase of infection. In some of the later time points in NiV-M inoculated hamsters in our study, Nipah virus antigen was present in pneumocytes adjacent to multiple arteries in the lung and Nipah virus antigen was identified in the arterial tunica media, but not endothelial cells in these blood vessels. These results suggest that Nipah virus spread from the adjacent pulmonary parenchyma to the outer vascular wall of arteries, with successive dissemination towards the vascular lumen. If time points subsequent to 48 hpi were analyzed it is likely that Nipah virus would have spread from the arterial smooth muscle cells to the overlying endothelial cells in the lung, then into the bloodstream with subsequent infection of blood vessels throughout the body, as evidenced by detection of viral antigen in endothelial cells in multiple tissues in Syrian hamster studies focusing on the late stages of disease [19–22]. Although the exact route of systemic virus dissemination to tissues and vasculature throughout the body is still unclear, it is suspected that circulating lymphocytes may play a role. Even though we were not able to detect virus binding to, or infecting, peripheral blood leukocytes during the early stage of infection in hamsters, it has been shown that Nipah virus can bind to human lymphocytes and hamster mononuclear leukocytes and that these leukocytes can carry and transfer the virus to permissive cells [30]. As such, lymphocytes may transfer virus directly to endothelial cells in blood vessels throughout the body, with secondary viral spread from blood vessels to adjacent parenchymal cells, including epithelial cells and neurons, or transmigration of lymphocytes through vascular walls and direct transfer of the virus to parenchymal cells during the later stages of infection [33].
Subsequent to the initial infection of cells in the nasal cavity and lung, epithelial cells lining the larynx and trachea were targeted by Nipah virus. Based on the lack of vascular involvement at this time point, the dissemination of Nipah virus throughout the respiratory tract was likely caused by the upward or downward spread of virus particles through the airways during respiration, or from the bronchi to the trachea and larynx by way of the mucociliary apparatus rather than via a hematogenous route. Other viruses that target the respiratory tract, including influenza virus, have been shown to damage the mucociliary apparatus, thereby decreasing the speed and effectiveness with which pathogens are cleared from the respiratory tract [34]. Alternatively, the epithelium lining the larynx and trachea may have been exposed to Nipah virus at the same time point as the nasal cavity and lung, but the rate of viral entry and replication in the larynx and trachea may have been slower resulting in delayed detection of viral antigen and replication.
Since histologic lesions have been reported in the central nervous system, and to a lesser extent, lymphoid organs in fatal human cases [1, 7], we evaluated whether there was virus dissemination to non-respiratory tract tissues during the early stages of Nipah virus infection in Syrian hamsters. Low viral loads were observed in the brain and spinal cord in a few NiV-B and NiV-M inoculated hamsters at 24 or 32 hpi, yet infectious virus, viral antigen, and virus replication were not detected. The low viral loads in these tissues may signify early virus dissemination to the nervous system at a level below the detection limit of IHC and virus titrations. It has previously been shown that IHC can detect viral antigen in the brain of Syrian hamsters by 4 days post inoculation and that virus disseminated from the olfactory epithelium to the brain through the olfactory nerve [23]. Our results likely represent the initial stages of virus dissemination to the nervous system; however, since viral loads in the brain were low, we were unable to detect, or track movement of, viral antigen into the brain by IHC. Spread of virus from the brain to the spinal cord may have occurred through the cerebrospinal fluid. Nipah virus has been detected in the cerebrospinal fluid of infected humans [35]. In the lymphoid organs, infectious virus was detected at 48 hpi in the cervical lymph nodes of NiV-M inoculated hamsters, but was not present in NiV-B inoculated hamsters. Lymphatics from the head and neck drain into the cervical lymph nodes, as such, virus could be transported from the nasal cavity to the cervical lymph nodes through lymphatic drainage. Infectious virus was only detected at 48 hpi in the spleen of a single NiV-M inoculated hamster and was not identified in any NiV-B inoculated hamsters. Spread of Nipah virus to the spleen may have occurred by a hematogenous route. Although viremia was not detected at 48 hpi, it is possible that the overall viral load in the blood was below the level of detection; however, since the spleen functions to remove pathogens from the blood as blood is filtered through the spleen [36–38], Nipah virus may have accumulated to high enough levels in the spleen that virus could be detected there.
In summary, our results provide evidence that Nipah virus initially disseminates throughout the upper and lower respiratory tracts via airways during the early phase of infection after intranasal inoculation. This may be followed by dissemination of virus from the nasal cavity to the nervous system by neural route and dissemination of virus to lymph nodes via lymphatic drainage. Dissemination of virus to blood vessels outside of the respiratory tract does not appear to occur during the first 48 hpi and is considered a late stage event. Since Nipah virus can cause encephalitis that results in severe neurologic disease in humans and brainstem damage that often causes death, either by directly infecting neurons or infecting nervous system blood vessels with secondary development of infarcts, it is important to prevent virus dissemination and replication in the brain and vasculature [1, 5, 7, 9]. Our data indicate that virus replication and dissemination are initiated rapidly after viral infection, thus making it difficult to prevent the virus from spreading to non-respiratory tract tissues such as the central nervous system. This suggests that development of vaccines that block dissemination or treatments that can access the brain and spinal cord and directly inhibit viral entrance into cells and/or virus replication may be necessary for prevention of central nervous system pathology.
|
10.1371/journal.pmed.1002489 | Estimated mortality on HIV treatment among active patients and patients lost to follow-up in 4 provinces of Zambia: Findings from a multistage sampling-based survey | Survival represents the single most important indicator of successful HIV treatment. Routine monitoring fails to capture most deaths. As a result, both regional assessments of the impact of HIV services and identification of hotspots for improvement efforts are limited. We sought to assess true mortality on treatment, characterize the extent under-reporting of mortality in routine health information systems in Zambia, and identify drivers of mortality across sites and over time using a multistage, regionally representative sampling approach.
We enumerated all HIV infected adults on antiretroviral therapy (ART) who visited any one of 64 facilities across 4 provinces in Zambia during the 24-month period from 1 August 2013 to 31 July 2015. We identified a probability sample of patients who were lost to follow-up through selecting facilities probability proportional to size and then a simple random sample of lost patients. Outcomes among patients lost to follow-up were incorporated into survival analysis and multivariate regression through probability weights. Of 165,464 individuals (64% female, median age 39 years (IQR 33–46), median CD4 201 cells/mm3 (IQR 111–312), the 2-year cumulative incidence of mortality increased from 1.9% (95% CI 1.7%–2.0%) to a corrected rate of 7.0% (95% CI 5.7%–8.4%) (all ART users) and from 2.1% (95% CI 1.8%–2.4%) to 8.3% (95% CI 6.1%–10.7%) (new ART users). Revised provincial mortality rates ranged from 3–9 times higher than naïve rates for new ART users and were lowest in Lusaka Province (4.6 per 100 person-years) and highest in Western Province (8.7 per 100 person-years) after correction. Corrected mortality rates varied markedly by clinic, with an IQR of 3.5 to 7.5 deaths per 100 person-years and a high of 13.4 deaths per 100 person-years among new ART users, even after adjustment for clinical (e.g., pretherapy CD4) and contextual (e.g., province and clinic size) factors. Mortality rates (all ART users) were highest year 1 after treatment at 4.6/100 person-years (95% CI 3.9–5.5), 2.9/100 person-years (95% CI 2.1–3.9) in year 2, and approximately 1.6% per year through 8 years on treatment. In multivariate analysis, patient-level factors including male sex and pretherapy CD4 levels and WHO stage were associated with higher mortality among new ART users, while male sex and HIV disclosure were associated with mortality among all ART users. In both cases, being late (>14 days late for appointment) or lost (>90 days late for an appointment) was associated with deaths. We were unable to ascertain the vital status of about one-quarter of those lost and selected for tracing and did not adjudicate causes of death.
HIV treatment in Zambia is not optimally effective. The high and sustained mortality rates and marked under-reporting of mortality at the provincial-level and unexplained heterogeneity between regions and sites suggest opportunities for the use of corrected mortality rates for quality improvement. A regionally representative sampling-based approach can bring gaps and opportunities for programs into clear epidemiological focus for local and global decision makers.
| Previous studies from cohorts in South Africa and parts of East Africa have suggested that site-level reporting of mortality is incomplete.
We wanted to understand the degree to which this phenomenon was impacting HIV outcomes at a broader scale, in this case at the provincial level in Zambia, a country with one of the highest burdens of HIV.
We also wanted to gain an in-depth understanding of differences between the outcomes of clinical care sites in order to assess the role of mortality as a potential quality improvement target.
From a source population of patients in 4 provinces (Lusaka, Southern, Eastern, and Western) who visited government-operated HIV treatment sites in these provinces, we conducted a multistage sampling approach of a stratified selection of sites and a random sample of patients lost to follow-up.
Lost patients were traced and their vital status was ascertained, which was used to enable a corrected regionally representative estimate of survival after starting antiretroviral therapy (ART) as well as corrected site-specific mortality estimates.
Of 165,464 individuals, the 2-year cumulative incidence of mortality increased from 1.9% to 7.0% for all ART users and from 2.1% to 8.3% for new ART users, and provincial-level mortality rates rose 3- to 8-fold once corrected for true outcomes.
Being late (>14 days late for appointment) or lost (>90 days late for an appointment) was associated with death.
Deaths are under-reported within the Zambian HIV program, and mortality rates are highly variable across sites and provinces.
Our findings enable national- and global-level policy makers to correct existing underestimates of mortality, link these data to quality improvement efforts, and reprioritize interventions to target regional and site-level reductions in mortality as a goal of HIV programs.
We have also established that this methodology is feasible for use as a representative surveillance tool for accurate monitoring of provincial and potentially national levels of mortality, even as vital status registries and data systems are further developed and strengthened.
| As the global response to HIV moves towards universal treatment, with the intent of ending AIDS as a public health crisis by 2030, a rigorous understanding of contemporary program effectiveness demonstrates progress as well as identifies important opportunities for improvement [1]. Although other metrics also contain important information, mortality captures program effectiveness better than any other single process or clinical outcome. For example, although plasma HIV RNA suppression is a critical measure of HIV control and uncontrolled viremia indicates immunological deterioration, suppressed HIV RNA does not reflect benefits from other aspects of care such as isoniazid preventative therapy, management of side effects, or treatment of opportunistic infections. Of note, even 90-90-90 [2]—a centrally important set of metrics of national program performance—does not include monitoring of a program’s success in saving lives.
Despite the importance of mortality, relatively little data exist about long-term survival after initiation of HIV treatment in lower- and middle-income countries [3]. In South Africa, a new National Population Register for recording vital status has enabled an advanced understanding of mortality, but most high-burden settings lack such registries [4,5]. Routine monitoring and evaluation of treatment programs also fail to capture most deaths because they often happen outside of the health system in patients considered “lost to follow up” [6,7]. A nomogram has been proposed that incorporates the results of some of these tracing studies and uses the fraction of individuals lost to assess mortality, but it is not always highly accurate at site-level estimates [8]. Efforts have been undertaken to understand the “true” mortality rates through tracing random samples of patients lost to follow-up in convenience samples of clinics in multiple studies [9–11], but this approach has only been examined in the context of a limited number of conveniently sampled facilities. Mathematical models for the Joint United Nations Programme on HIV/AIDS (UNAIDS) and others report on mortality but depend on a number of assumptions and limited underlying epidemiological data [12].
In order to rigorously assess mortality, we apply for the first time, to our knowledge, a multistage sampling design to obtain representative estimates of mortality in 4 provinces in Zambia as well as site-level estimates with enough precision to identify site-to-site variation. This approach offers a rigorous estimate of (1) the magnitude of deaths after starting antiretroviral therapy (ART) among those already on treatment at the start of the study and among new treatment initiates, (2) when deaths occur, (3) which groups are at highest risk of death, and (4) whether these factors differ by region, facility, or other important factors such as prior healthcare utilization patterns. Both the approach and findings have implications for assessment of program effectiveness and program improvement efforts.
The protocol and study were approved by the University of Zambia Biomedical Research Ethics Committee (UNZABREC) (004-06-14) and the IRB of the University of Alabama, Birmingham, School of Medicine (F160122006), and the submitted protocol is available in the supplemental material. In this paper, we used a number of analytic innovations not specified in the analysis protocol, including the use of a modified Lorenz curve to depict the distribution of deaths across facilities and survival analyses of mortality in which deaths are classified by retention histories. In addition, analyses of pretreatment experience as specified in the protocol are also planned but not included in this paper.
The target population for this analysis is the contemporary population of HIV-infected adults on ART in Zambia. We considered “contemporary” to be patients who had a clinical encounter while on ART (including the ART initiation visit) in the 24 months before the evaluation, and they are referred to as “all ART users” and provide the most complete picture of treatment outcomes. We also designate a subpopulation of this group as “new ART users,” i.e., those who initiated ART during the 24-month period prior to the evaluation—this group is more readily comparable to outcomes from other analyses of those starting ART. Our source population consisted of patients in 4 provinces (Lusaka, Southern, Eastern, and Western) who visited government-operated HIV treatment sites in these provinces, each of which receives technical assistance support from Centre for Infectious Disease Research in Zambia (CIDRZ), a local Zambian organization.
The analysis population was enumerated using a multistage sampling approach to yield estimates representative of the entire population of patients on treatment as well as to enable comparison, through stratification, across 4 provinces and 3 facility types (hospital, urban health center, and rural health center). We selected using probabilities proportional to size a minimum of 2 to 10 facilities from each of 12 strata defined by facility type and province for a total of 32 facilities. In each selected facility, we enumerated all adults on ART who made a visit in the previous 24 months and identified loss to follow-up from this cohort (defined as >90 days late at the time of sampling). From patients lost to follow-up, we selected a simple random sample for tracing inverse proportional to size. This approach seeks to enable both a regionally representative estimate of survival after starting ART and site-specific estimates with approximately 95% confidence estimates from 5% to 15% in a setting where mortality is estimated to be 10% [9]. We additionally used simulation to optimize the sampling strategy to achieve these aims. A diagram detailing the sampling process and tracing outcomes is presented in the supplementary materials (S1 Fig).
As described in previous work, we used data from an electronic medical records system to enumerate the cohort and to obtain sociodemographic and clinical data. Peer health workers traced lost patients intensively in the community to ascertain vital status [13]. Tracing included in-depth review of paper and electronic medical records, phone calls, and in-person tracing in the community. Tracers used bicycles, public transportation, study vehicles, or motorcycles as needed. A minimum of 3 tracing attempts was required. If patients were found to be dead, death dates, cause of death, and location of death were solicited from family and other close contacts. The list of patients to be traced was ranked in random order and released in blocks of 25 to each facility to preserve the benefits of a random sample.
We estimated the cumulative incidence of mortality using the Kaplan Meier (KM) approach as well as simple mortality rates. In KM estimates, we treated time zero as the date of ART initiation. Patients who started ART before the observation period were left-truncated, which yields an estimate of survival after ART initiation during the 2-year period of observation of this study that is contingent on surviving into this period. This approach is analogous to life expectancy estimates. Mortality estimates were carried out for all ART users as well as stratified by new ART initiators during the 2-year observation period and those already on ART at the start of the current period. These analyses are intended to convey the totality of the contemporary experience of the patient population, including both that of those on treatment for extended periods of time and that of important subpopulations. We carried out mortality estimates both using only deaths known to the program before tracing and after incorporating findings from tracing lost patients to “correct” estimates through use of probability weights. These individual weights were combined with facility-level weights for estimates in the entire patient population.
In additional survival analyses, we used the Aalen Johansen method to estimate the cumulative incidence of death events classified by their current and previous engagement history using the reweighted data. We defined patients as having “died in care” if they had died within 30 days of their last visit; the remaining patients were classified as having “died out of care.” In addition, patients were categorized as “previously lost” if they had had at least 1 episode of being out of care for more than 90 days after an appointment in the past and returned to care prior to our study sampling and “previously late” if they had at least 1 episode of being late for an appointment by more than 30 days but less than 90 days prior to our study sampling. These categories yielded 6 engagement states for the analysis [14,15]. We examined the sociodemographic (e.g., sex and age), clinical (e.g., CD4 level at ART initiation), and health systems factors (e.g., facility size) associated with mortality through simple stratification and multivariate Poisson regression. Analytic weights inverse to the probability of missingness were used to address missing predictor side variables under the missing-at-random assumption [16]. We quantified site-to-site variability in mortality by examining the distribution of site-specific mortality estimates. In order to summarize the unequal contribution to total mortality across facilities, we used a modification of the Lorenz curve in which the y-axis displays the cumulative proportion of excess deaths and the x-axis the cumulative population proportion after adjustment for patient (e.g., sex and CD4 level at ART initiation) as well as facility factors (e.g., facility volume) [17]. Excess deaths are defined as deaths above a threshold of 1% at 2 years by facility—the approximate death rate in cohorts from Europe of stable patients used to represent “ideal” outcomes [18]. The Lorenz analysis was restricted to new ART initiators in order to minimize the effect of local, clinic-specific distribution of patient time of treatment.
We used 3 approaches to characterize deaths over time. First, we graphically displayed the retention history with all patients who died represented by a single horizontal bar. The length of the bar represents time since ART initiation, and colors represent time spent in care, late for care (>30 days and <90 days late for a visit), or lost (>90 days late for last appointment). Second, we examined mortality rates in each year after ART initiation among all patients as well as only among those patients who had made a visit in the last 2 years (and therefore survived into the current era) and tested for significance in the change over time. Third, we added time-varying covariates—(1) engaged, never late or lost; (2) engaged, but with a history of being late or lost; (3) late without a history of previous episodes of being late or lost; (4) late with a history of being lost; and (5) being lost (defined as being >90 days since last appointment)—in multivariable regression to capture the association between lapses in retention states and mortality. Based on the revised estimates, we quantified the fraction of deaths that fell into each category of visit histories and in each of the segments of time after ART initiation. We calculated the “population attributable fraction” of all deaths that occurred in each of these states—an estimate of how much mortality could be avoided if all measured characteristics were set at the lowest level of risk.
In a network of 64 HIV delivery facilities operated by the Zambian Ministry of Health and supported by CIDRZ across 4 provinces, we identified a total of 165,464 individuals on ART and who had made a visit in the 2 years before the project period (i.e., 1 August 2013 to 31 July 2015), which included 49,129 individuals newly initiating ART during that period and 116,335 individuals already on treatment at the opening of the observation period (Fig 1).
In the entire cohort (Table 1), 64% were female, the median age was 39 years (IQR 33–46), 56% attended urban facilities, about half were from Lusaka Provence (52%), median CD4 at ART initiation was 201 cells/mm3 (IQR 111–312), and 58% were WHO stage 1 or 2 at ART initiation. New ART users were similar demographically but had a higher median CD4 level at ART initiation of 262 cells/mm3 (IQR 138–288), and 65% were WHO stage 1 or 2 at enrollment. The median time on ART was 1,142 days (IQR 390–2,139) for all ART users in the cohort and 255 days (IQR 63–407) in new ART starters. The median date of ART initiation for all users was 14 February 2012 (IQR 03 June 2009–23 January 2014) and for new ART users was 24 July 2014 (IQR 3 February 2014–23 December 2014).
Routine program data before the tracing exercise suggested that 1% of all patients had died. A total of 17% of all patients were lost to follow-up, and this fraction lost was highly variable across clinic sites (IQR 10%–31%). Among the 28,111 lost patients from the entire cohort, we made an attempt to trace 2,892 (10%). Tracing efforts led to updated vital status in 2,163/2,892 (75%) of patients, but ascertainment of vital status varied from 54% to 93% across all sites (S1 Table). Among patients lost to follow-up from the entire cohort, 17% (95% CI 15%–19%) had died. Among patients lost from the group newly starting ART during the 2-year observation period, 21% had died (95% CI 17%–25%). After incorporating outcomes from tracing into underlying data known to the program, we found the 2-year cumulative incidence of mortality among all ART users increased from a naive estimate of 1.9% (95% CI 1.7%–2.0%) to a sample-revised estimate of 7.0% (95% CI 5.7%–8.4%) (Fig 2). Among new ART users, sampling changed the cumulative incidence of mortality at 2 years from a naïve estimate of 2.1% (95% CI 1.8%–2.4%) to a revised estimate of 8.3% (95% CI 6.1%–10.7%) (Fig 2). Facility-level revised mortality estimates ranged widely, with an IQR from 1.6 to 2.6 deaths per 100 person-years among all patients and an IQR of 3.5 to 7.5 deaths per 100 person-years and a high of 13.4 deaths per 100 person-years among new ART initiators (Fig 2). Most facilities showed marked differences between naïve and revised estimates, with 1 site showing a 23-fold difference in mortality among all patients and another site showing a 14-fold difference for new ART users.
The deaths during the first year on treatment occur mostly in persons who have not missed appointments. There is an increase over time in the fraction of deaths that occur in patients who are either late for a visit or who have a history of being late or lost from care (Fig 3). The cumulative incidence estimates and confidence intervals for this analysis are presented in the supplementary materials (S2 Table). Furthermore, general categories of reported causes of death were available for about half of the patients who died (S3 Table).
Mortality varied markedly across geographical areas and facilities. At the provincial level, revised estimates of mortality were higher than naïve estimates, ranging in all ART users from 3-fold higher in Eastern and Southern Province to 8-times higher in Lusaka Province and in new ART users from 3-fold higher in Eastern Province to 9-fold higher in Lusaka Province (Fig 4, S4 Table). The highest absolute mortality rates, prior to revision, were in Eastern Province (all ART users) and Southern Province (new ART users), whereas the highest revised mortality rates were in Western Province, followed by Southern, Eastern, and Lusaka Provinces, for both all ART users and new ART users. Facility-level mortality estimates were highly variable even within province and across facilities with similar median CD4 levels at ART initiation (Fig 5).
Multivariable regression models using patient and facility characteristics at ART initiation found that male sex, CD4 level (≥200 cells/μL) at ART initiation, and higher WHO stage were associated with higher mortality in new ART users, whereas facility size and type (health center) were protective. For those already on ART at the start of observation, male sex; province; facility size, type, and setting; and disclosure of HIV status at treatment initiation were associated with mortality (Table 2).
A modified Lorenz curve to capture how “concentrated” mortality was across facilities showed that approximately half (57%) of excess deaths occurred in 15% of the population (Fig 6). The Gini coefficient was 0.714, suggesting that 71.4% of deaths were not evenly distributed across the population.
Mortality in all ART users was highest during the first year of treatment at 4.6 per 100 person-years (95% CI 4.0–5.5), fell to 2.9 per 100 person-years (95% CI 2.1–3.9) in the second year, and then remained steady at approximately 1.6% per year for up to 8 years on treatment. The overall mortality rate in this group was 1.7 per 100 person-years (95% CI 1.5–2.0), and a test of difference in mortality rates by year on treatment for up to 10 years was not statistically significant (p-value = 0.46). When examined alone, patients who started ART during the observation time for this study (new ART users) displayed the highest mortality in the first 90 days of treatment, at 7.9 per 100 person-years (95% CI 6.3–10.0); this fell to 4.6 per 100 person-years from 91–180 days and then fell to 3.3 per 100 person-years (95% CI 2.0–5.8). Regression models with time-varying factors capturing changing retention status found that being late and being lost were both associated with mortality, but a history of lateness or loss among those in care did not predict mortality after adjustment for other sociodemographic and clinical factors. These associations were more pronounced among all ART users than among new ART users (Table 2).
We estimated that 38.1% (95% CI 33.1%–43.4%) of mortality occurred among individuals in their first year on treatment, while the remaining 61.8% (95% CI 56.6%–66. 9%) was distributed among patients who had been on treatment for longer periods of time (Table 3). For patients newly starting treatment, 50.3% (95% CI 42.4%–58.2%) of all deaths occurred among those who had no history of missed visits, while 14.0% (95% CI 9.2%–20.7%) occurred in individuals while they were late, and 31.7% (95% CI 24.5%–39.9%) occurred among individuals while they were lost (S2 Fig). For those already on ART in the current period, 8.8% (95% CI 6.4%–12.1%) of deaths occurred among those who were always in care at the time of death, 13.7% (95% CI 9.7%–18.9%) occurred in those with a history of lateness (but who were in care at the time of death), 29.1% (95% CI 23.1%–35.9%) occurred in those who were in care but had a history of being lost, 11.7% (95% CI 8.1%–16.7%) occurred in individuals while they were late, and another 36.7% (95% CI 30.1%–43.8%) occurred among individuals while they were lost (S2 Fig).
Despite the success of implementing HIV treatment in high-burden settings in Africa, we have identified high rates of mortality among those on treatment in Zambia, substantial under-reporting of deaths, and marked heterogeneity among provinces and sites. We applied for the first time, to our knowledge, a multistage sampling approach that included a sample of both sites and patients lost to follow-up in order to develop both provincially representative and site-level revised mortality estimates for an HIV program. In a sample representing over 160,000 patients across 4 provinces in Zambia, we found that 7% of all patients and 8% of new ART starters died within 2 years—from 10- to 20-fold higher than treated patients in Europe [18]. Provincial mortality rates developed from routine program data under-estimated true mortality rates in the provinces for new ART users by 3- to 9-fold, and revised mortality rates revealed the highest mortality rates for the HIV program in Western Province, followed by Southern, Eastern, and Lusaka. We identified important patient characteristics associated with short-term (e.g., CD4 level at initiation and male sex) and long-term mortality (e.g., male sex and disclosure), as others have described [9,19]. At the site level, care in a hospital and care in a nonurban setting were associated with a protective effect on mortality risk for all ART users, and the size of the facility had a statistically significant but modest protective effect. Importantly, we also found that the facility itself (regardless of province) is a critical driver of survival: even after adjustment for patients and facility characteristics, mortality across facilities ranged from 0 per 100 person years to 13.4 per 100 person years, even though the facilities in theory receive similar support (e.g., human resources and infrastructure). We also found that although mortality rates were highest in the first months of treatment, death rates among those on treatment for longer periods of time remain steady and high. None of these epidemiological observations would be possible using routinely collected death data, which differed markedly from sample-revised estimates.
Although mortality from childbirth, road accidents, and other factors are indeed higher in Africa, over 95% of the deaths ascertained in this cohort were due to “illness” rather than death from these competing causes. We believe the potential drivers of deaths include unmet needs for more advanced medical care, especially given the protective effect in our multivariable model of being treated in a hospital versus a health center, which may indicate a better capacity for advanced diagnostics and care in hospital-based clinics. Indeed, approximately half of deaths occur relatively shortly after a clinic visit. In a previous analysis, we found that the majority of patients who died after a recent visit had some opportunity for medical intervention that was missed—a more detailed assessment of the clinical encounters before death in these patients should be urgently undertaken [13]. At a time when public health approaches are focused on deintensifying care through differentiated service delivery [20,21], it should be recalled that in many cases, improving outcomes means identifying appropriate patients in whom to escalate care.
The provincially representative nature of the study enabled a new and more nuanced understanding of regional mortality among those newly starting ART and among the cross-section of all ART users in a province. First, the ratio of the revised mortality rate versus the naïve rate was highest in Lusaka Province—8-fold higher for all ART users and 9-fold higher for new ART users—compared with the other 3 provinces (ratios between 3–5). One interpretation of this is that populations in dense urban population centers such as Lusaka may be more likely to have their mortality under-reported. Lusaka Province is dominated by the capital city and transport hub of Lusaka, which has an urban population of 2.4 million people. It is conceivable that high mobility and/or fewer community linkages to health facilities result in less passive reporting of mortality than in the less urban provinces. The highest absolute revised mortality rates were in Western Province, which is a largely rural floodplain of the Zambezi River with the lowest population density of the 4 provinces, an HIV prevalence equal to that in Lusaka Province (16%), and limited economic activity other than cattle farming. Although our overall multivariable analysis suggested an association of nonurban facility setting with reduced mortality, health services in much of Western Province region are notoriously difficult to deliver and up-referrals for advanced care can be particularly challenging in the rainy season because of flooding and long distances to higher-level facilities. We believe these and potentially other factors may account for the high revised mortality rates in Western Province, although further investigation is needed at the provincial and site level. Indeed, the differences between the revised provincial mortality rates and between the ratios of the naïve and corrected morality rates in the provinces serve as starting points for both quality improvement work by provincial health offices and for further study. For example, future investigation could examine the system- and human-level causes of low levels of passive mortality reporting that we found across all 4 provinces, but particularly in Lusaka Province, and could investigate differential utilization of up-referral for sick patients by province and facility type and the association of provincial mortality rates with the quality and availability of advanced care services in the region.
Our results, drawn from a sample of patients visiting HIV treatment clinics between 2013 and 2015, indicate that even in an era of expanded eligibility for therapy and rising CD4 cell count thresholds, CD4 cell counts remain low at initiation and are a risk factor for mortality [22,23]. With clinical guidelines in Zambia now recommending universal treatment, we expect CD4 cell counts at treatment initiation will increase and mortality rates will decline overall. However, we anticipate that inconsistent healthcare utilization will continue to be a threat that could undermine potential gains. Although mortality amongst samples of lost patients has declined in recent years [11], perhaps reflecting an overall increase in CD4 cell counts at treatment initiation, our study also demonstrated that death rates remain unacceptably high even after 1 or more years of therapy, a time period when many might assume patients to be stable. An elevated risk of death, even years after initiating ART, was also seen in South African cohorts with high ascertainment of mortality through their national vital status registry. Standardized mortality ratios (SMRs) (observed deaths divided by the number of deaths expected if all patients were HIV negative) demonstrated that men in particular had elevated SMRs even after 48 months of therapy and having started therapy with a CD4 cell count >200 cells/mm3; women’s rates approached those of HIV-uninfected individuals by 48 months [4], a gender differential that has been widely noted [19]. These observations problematize a number of assumptions. First, it has been assumed that most patients on treatment for longer periods of time will be stable and can be placed in deintensified, demedicalized groups that emphasize social support and community-based treatment. Yet, high mortality rates in this group suggest that time on treatment itself cannot be a reliable marker of stability. Furthermore, it is assumed that intensive adherence support is critical only at initiation for most patients. Persistent deaths after 2 years on treatment imply that a subset of patients is poorly adherent or retained and that progression of HIV disease on treatment appears to be common. Therefore, effective support for adherence that is attuned to mitigating threats to adherence over time remains a crucial issue [24]. Previous work has suggested that adherence, as measured by medical possession ratio, is not optimal [25], and the high death rates on treatment may reflect disease progression. Finally, treatment does not fully protect individuals against tuberculosis; thus, rolling out isoniazid preventative therapy is a priority [26].
The marked heterogeneity of mortality across facilities is a crucial and novel observation made possible by our sampling scheme. The fact that mortality differs across facilities, even when higher-level systems support (e.g., supply chain, information systems, and central laboratory services) is relatively uniform, implies that the next generation of improvement efforts should be focused at the facility level and with access to improved data for better decision-making. In South Africa, where facilities are now linked to the national vital status registry, this is already possible, but in most settings in the region, it is not. Another implication of the site-to-site variability is that attention should be targeted to the high-volume, high-mortality-rate facilities because the absolute reductions in mortality are likely to be the biggest through improvements at these sites. Third, the presence of high-performing facilities suggests that strategies such as exemplars by positive deviants—in which high-performing facilities are brought in to assess and guide lower-performing facilities—should be explored [27]. Unexplained variability in small geographical areas supported by largely uniform systems spurred a generation of improvement science in the United States through the Dartmouth Atlas and other investigators and should be a crucial driver of improvement in the AIDS response as well [28,29].
Overall, our findings also imply that in Zambia and other similar settings, even if the regular use of site and subnational region-level mortality data for improvement were attempted, inaccuracy in the underlying data would undermine the utility of the approach. As we have shown, under-recording of mortality is the norm, and variability is masked under typical approaches to monitoring and evaluation, as evidenced by the differences in the ordering of mortality burden by province with use of naïve versus revised mortality rates. Although Demographic and Health Surveys (DHSs) and other door-to-door surveys provide representative data that can be used to evaluate trends in mortality, they are often not linked to sites and are not conducted with the frequency required to monitor for program performance and to measure the proximal effects of health systems interventions [30]. We propose 2 potential approaches to this issue. The first is to renew existing calls for nationwide death and birth registries, an approach successfully taken to scale in South Africa [31] and now being piloted in Zambia. Where this is not immediately possible, a scalable version of the representative multistage sampling performed in this study would be a feasible approach to incorporate into regional surveillance. Indeed, this approach is currently being adapted for broader use in Zambia, and we will shortly be publishing a cost analysis and a toolkit to encourage its broader use in the region. If efforts to expand this approach to improving our understanding of mortality are successful and supported by governments and funders, we would ideally be able to equip every site, district, and country with up-to-date access to accurate mortality data. In addition, incorporation of corrected provincial mortality rates into projection models could further improve their ability to inform resource needs and epidemic trends.
The methodology is predicated on tracing a random sample of individuals lost to care. We were unable to ascertain the vital status of about one-quarter of those lost and traced, which could have biased our results. However, we did not find differences in the sociodemographic or clinical characteristics among those who were traced and found compared with those we were unable to find. We were also not able to adjudicate causes of death and therefore are unable to distinguish clearly all-cause mortality versus HIV-related mortality. Further studies are needed to better understand clinical causes of death among HIV-infected individuals on ART in Zambia. Lastly, although the study included a comprehensive sample of Ministry of Health (MOH) ART sites in the 4 provinces, faith-based and other sites were not included in the estimates.
To our knowledge, this study is the first to apply a multistage sampling-based approach to generating provincially representative mortality rates for HIV programs in low- and middle-income countries (LMICs). In addition to demonstrating that provincial mortality is substantially under-reported and variable among HIV-infected individuals on ART, we found unexpectedly high site-to-site variability in mortality after starting HIV treatment under routine health service delivery conditions. This heterogeneity among sites suggests that service delivery has much room to improve and that greater attention to systematically generating and using corrected mortality rates is a logical next step for ongoing quality improvement for national programs and major donors. In addition, given the associations of late/missed visits with mortality, greater attention to patient-centered and differentiated care approaches that are also sensitive to the need for care intensification may have a role in not only improving retention but addressing the markedly high death rates, despite the widespread use and availability of ART.
|
10.1371/journal.pcbi.1001044 | Neural Development Features: Spatio-Temporal Development of the Caenorhabditis elegans Neuronal Network | The nematode Caenorhabditis elegans, with information on neural connectivity, three-dimensional position and cell linage, provides a unique system for understanding the development of neural networks. Although C. elegans has been widely studied in the past, we present the first statistical study from a developmental perspective, with findings that raise interesting suggestions on the establishment of long-distance connections and network hubs. Here, we analyze the neuro-development for temporal and spatial features, using birth times of neurons and their three-dimensional positions. Comparisons of growth in C. elegans with random spatial network growth highlight two findings relevant to neural network development. First, most neurons which are linked by long-distance connections are born around the same time and early on, suggesting the possibility of early contact or interaction between connected neurons during development. Second, early-born neurons are more highly connected (tendency to form hubs) than later-born neurons. This indicates that the longer time frame available to them might underlie high connectivity. Both outcomes are not observed for random connection formation. The study finds that around one-third of electrically coupled long-range connections are late forming, raising the question of what mechanisms are involved in ensuring their accuracy, particularly in light of the extremely invariant connectivity observed in C. elegans. In conclusion, the sequence of neural network development highlights the possibility of early contact or interaction in securing long-distance and high-degree connectivity.
| Long-distance connections are crucial for information processing in neural systems, and changes in long-distance connectivity have been shown for many brain diseases ranging from Alzheimer's to schizophrenia. How do long-distance connections develop? Traditionally, connections can be formed over long distances using guidance cues for steering axonal growth. Subsequently, other connections can follow those pioneer axons to a target location. Alternatively, two neurons can establish a connection early on, which turns into a long-distance connection as the neural system grows. However, the relative contribution of both mechanisms previously remained unclear. Here, we study long-distance connection development in the neuronal network of the roundworm C. elegans. We find that most neurons that are connected by a long-distance connection could interact and establish contact early on. This suggests that early formation could be an influential factor for establishing long-distance connectivity, with a hypothetical role in neuronal wiring accuracy. Reducing the need for axonal guidance is also likely to reduce metabolic costs during development. We also find that highly-connected neurons (hubs) are born early on, potentially giving them more time to host and establish connections. Therefore, neuron birth times can be an important developmental factor for the spatial and topological properties of neural circuits.
| The complexity of the nervous system continues to protract efforts to understand its development. The relatively simple invertebrate neural systems have been the subject of intense study in the last few decades [1], helping to shed light on mechanisms involved in development like axon guidance and molecular cues. We looked at the development of the neuronal network of C. elegans using information from 279 of the 302 neurons (see methods and [2]). Information on embryonic and post-embryonic lineages of the neurons [3], [4], [5], [6] formed the basis of our developmental data.
Our work marks the first attempt to computationally and statistically represent the neural development of C. elegans based on available biological data, enabling a spatio-temporal analysis of the developing neuronal network. Graph theory [7] is increasingly being applied to elucidate the function based on the structures of complex networks like the brain [8], [9] and here we carry out a structural analysis of neuronal networks during different stages of C. elegans development. We observe neuronal growth (see Video S1), development times of different classes of neurons, and the time windows for establishing short- and long-distance connectivity.
It is now known that C. elegans displays a higher than expected wiring cost with the total wiring length of all connections being twice as high as for an optimized network with spatially rearranged neurons [2], [10]. Instead, processing speed indicated by the number of intermediate neurons in a pathway and made possible by long-distance connections, seems to be a critical constraint [2]. Establishing accurate connectivity, particularly long-distance, is a critical challenge in development. While the role of guidance molecules is well established in correctly wiring neurons, our results suggest that temporal and possible spatial closeness permitting early neuron-neuron interaction could also have an important role to play in the process. It may be conjectured that this would serve to reduce metabolic costs during development and also increase the probability of accurately establishing long-distance connections. Long-distance connections are vital and known to be affected in several neurological disorders in humans [11].
Our analysis was primarily four-fold. First, we traced the time course of neuron establishment. Second, we looked at temporal patterns emerging from birth times of neurons and their known connectivity details. Third, we looked at spatial network features in particular focusing on the onset of short, medium, and long distance connection-pairs. The analysis was further refined by segregating connections based on the nature of transmission (i.e. gap junctions or chemical synapses) as well as their membership in functional circuits. Finally, we analyzed possible networks at the various stages identified.
The growth in the neuronal network was visualized with respect to the spatial positions of neurons in the nematode body. In the development phase, neurons are born in two bursts – a relatively brief embryonic burst lasting around four hours and a longer post-embryonic phase stretching across seventeen hours. These intervals have been sub-divided into smaller time intervals to enable a more detailed visualization of the sequential appearance of neurons during these bursts. In Figure 1 neurons appearing at a particular stage are indicated in black, while previously existing neurons are coloured in gray. In the first stage, at the end of 350 minutes of embryonic growth, neurons that eventually reside in the head, along the ventral cord (classes DAn, DBn and DDn) and tail of the nematode have appeared, with neurons in the head constituting the majority. By the end of the embryonic burst nearly all of the neurons in the head have appeared. Ventral cord neurons belonging to the other five classes namely, ASn, VAn, VBn, VCn and VDn as well as the rest of the tail neurons are born in the latter phase, towards the end of the first larval stage. A relevant question is whether the spatial clustering of neurons in the head, body and tail of the worm also relates to their topology. We find that neurons in the head are mostly connected to neurons in the head (74% of all connections) while 46% of connections of neurons in the body are to other neurons also in the body. Within spatial cluster connectivity is least for neurons in the tail at 29%. Figure S1 shows the successive appearance of neurons with respect to their final 3-dimensional positions in the adult body on a colour scale.
We plotted a histogram of the differences in birth-times of connected neurons to assess the time interval available for formation of connections (Figure 2A). For comparison, we looked at the outcomes from twenty random networks (see Methods for details). The values from the trials were separated into ten time bins and the mean and standard deviation values were calculated for each of the bins. It can be seen that approximately two-thirds of connected neurons appear less than 200 minutes apart and this proportion is much higher than that seen in random networks. A zoom-in on time differences of up to 500 minutes shows that most connected neurons within this interval are born less than 50 minutes apart (Figure 2B).
Neurons appearing in the embryonic burst go on to make most of their connections, approximately 80%, with other neurons born in the same phase, while this figure is around 46% for neurons born in the post-embryonic burst. These values in randomly shuffled networks are on an average 62.1% and 38.2%, for the embryonic and post-embryonic phases, respectively. This observation could well be attributed to high degree neurons being born in the embryonic phase (discussed below), however a statistical comparison of the ratio of connections made within and outside the temporal cluster for the embryonic and post-embryonic phases to random development, highlights significance with p value less than 0.001 in a one sample t-test.
Early stages are also crucial for the formation of highly connected neurons (hubs). Figure 3 shows node degree (the number of connections of a neuron) with respect to birth times of neurons and their positions along the body of the worm, namely head, body or tail. Neurons that appear early on in development are more likely to have a higher degree than those that are born later. As would be expected, the highest degree is possessed by neurons in the head. There are 112 neurons with 20 or more connections. This number drops sharply to 46 and then to 9, for degrees more than or equal to 30 and 60, respectively. Of all neurons with degree 20 or more, two thirds appear before hatching (840 minutes), and nearly all neurons (∼98%) with more than 30 connections are born before hatching. It needs to be noted however, that correlation of early birth and higher degree is less obvious for neurons with less than 30 connections. To gauge the significance of these results we repeated the test for twenty random networks (details in Methods). For this random connection formation, approximately 74% (±5.5) of neurons with degrees above 30 could appear before hatching, which using two-sided t-test corresponded to a p value of less than 0.001 indicating significant difference between actual and random networks. We also analyzed how the connected neighbors of a neuron appeared over time, particularly to examine if longer time windows served to receive connections from late appearing neurons. This was indeed the case: all neurons with a degree of more than 60 were connected to more than one late forming neighbour. However in random networks, this was on average true for 60% of the cases (based on 20 random observations) and had a statistically lower connectivity with late appearing neurons, with p = 0.001. Thus availability of time between neuron birth and nematode maturation appears to be important to hub neurons.
The time differences of bilaterally paired neurons were also compared for examining symmetry during development. All bilateral pairs of neurons were born within approximately ten minutes of each other. Figure S2 shows the development of motor, sensory and interneurons. More than 80% of sensory and interneurons appear before hatching whereas this figure was found to be lower at around 50% for motor neurons. However, several of these motor neurons were polymodal, functioning also as inter-neurons. Only around 30% of exclusive motor neurons appeared before hatching. During such early development, there is less need for motor control: for example, earliest movements inside the shell do not involve neural coordination and first signs of neural activity are only observed around 30 minutes before hatching [12].
The distance between any two connected neurons was calculated in three-dimensional space to determine the approximate length of the connection between them. Numerically, this was calculated as the three-dimensional Euclidean distance between two connected neurons and provided a useful measure to assess the length distribution of all edges. Although there is little information on the timing of synaptogenesis, we wanted to visualize how many of the neuron pairs forming short-, medium-, or long-distance connections in the adult were present at each stage. Hence, when we use the term ‘connection pair’ - it does not imply synaptogenesis and is merely indicative of the birth of both neurons that will eventually connect. For the adult nematode, out of the 2,990 actual connections pairs 391 are long-distance (∼14%), 298 are medium-distance (∼10%) and 2,301 (∼76%) are short-distance (see Methods for details).
Figure 4A shows the percentage of short, medium and long-range connection-pairs appearing at each stage of development. By the time of hatching (840 minutes), approximately 73% of short, 36% of medium and 68% of long-range connections-pairs had appeared, accounting for 69% of total connections pairs. For twenty random networks (Figure 4B) on the other hand, 52% (±3.5), 53% (±6.7) and 51% (±5.7) of short-, medium-, and long-distance connection-pairs respectively, occurred before hatching. In none of the cases did the actual value fall within the data domain of random tests. Further analysis revealed that the observed difference between the percentage of connection-pairs appearing before hatching in real and random systems was statistically significant, with p<0.001 in each case.
To determine whether there was any relation between the degree of neuron and the proportion of short or long distance connections that they possessed, we computed the Pearson's coefficient for degree versus short and long connections. No significant correlation was found, with the correlation coefficient of degree with short and long connections being −0.03 and 0.11 respectively.
We then segregated the three types of junctions, namely, gap junction, chemical synapse or a combination of both. (If both gap junctions and chemical synapses existed between any two neurons, then the connection was termed as a combination.). The proportion of short-, medium- and long-length connections in each category is listed in Supplementary Table S1. Figure 5 compares the time of appearance of connection-pairs linked by gap junctions and chemical synapses in the adult, for short and long-distance connection lengths. Around 35% of short- and 27% of long-distance connections in C. elegans are electrically coupled (Figure 5A), together constituting more than one-third of all connections in the nematode. Interestingly, electrically and chemically connected neuron pairs appear at the same rate with approximately 70% of each category appearing before hatching.
A bar chart was also plotted to visualize the distribution of long, medium, and short-distance connection-pairs during each of the developmental stages (Figure S3). Short distance connection-pairs are most abundant during all stages of development. The frequency of short-distance connections is followed by long-distance and then by medium-distance connection-pairs.
The neurons of C. elegans have membership in various functional circuits [6] and we analyzed the appearance of connection-pairs in relation to this functional segregation. The percentage of each circuit that had appeared was measured in relation to the birth of connection-pairs. If a connection-pair involved neurons from two different circuits, then it was considered to belong to both circuits. Figure 6 shows how the various circuits appear over time. In the analysis of temporal features it was shown that approximately 80% of sensory and 50% of motor neurons appeared before hatching. Here, more specifically, analysis showed that the connection-pairs associated with circuitry of amphids, motoneurons in the nerve ring and other sensory receptors in the head, which are understandably more relevant to the early life of the worm are born sooner than other circuits like that of egg-laying, not required in the embryonic stages. Hence although we observe connection pairs belonging to all circuits present to various degrees in the earliest embryonic stages, the order of functional precedence may also influence the likelihood of early contact or interaction.
The networks at each stage were predicted based on the earliest possible time of synaptogenesis - when both neurons had appeared. Here, the network represents the neurons present at each stage with adult connectivity (see Methods). Our motivation behind this was to present the potential of network analysis in inferring development features like significant periods, from even discrete data. The global network over time was analyzed for topological changes such as number of nodes and edges (Figure S4A,B), clustering coefficient (Figure S4C,E), and characteristic path length (Figure S4D,F). The ratio of actual clustering coefficient to random of more than 2 and actual average path length to random of less than 1.5 are considered to signify a small-world network. The adult C. elegans network has already been shown to have small-world characteristics [13], here we find that all the networks display small-world characteristics with the ratio of actual clustering coefficient to that of random networks being above 4 at all times (Figure S4C), and the ratio of characteristic path length in the actual to the random network being consistently below 1.15 (Figure S4D). Although this analysis does not consider the effect of pruning, as the ratios are well beyond the characteristic ratios, the results are likely to be robust for small changes. The random networks created for these calculations were Erdős-Rényi networks which only maintained the total number of neurons and connections without preserving the degree distribution.
Detailed analysis on the developing neural network in C. elegans revealed that timings of neurons as well as the connections between them were reflected in characteristic patterns of development. Based on the three-dimensional positions of the neurons and their known connectivity data, we showed that long-distance connection-pairs often appear in the early stages of development. At these times before hatching, neurons are nearby so that there is less need for long-distance cues and axon guidance. Finally, early-born neurons were highly connected (hubs) in contrast to neurons born later, which could be attributed to the longer time frame available to them to establish connections.
A significant proportion of connected neuron births were temporally close, with over half of the connected neurons appearing within an hour of each other. At the time of hatching (840 minutes), when the worm would be less than 20% of its final size [5], approximately 68% of long-length and 73% of short-length connection-pairs had already appeared. In contrast, random growth of a network of similar degree distribution as the actual network resulted in approximately half of all connection lengths appearing by the time of hatching. It may be noted that the connection length throughout this study refers to the spatial separation of the soma of the neurons that are connected and is not with respect to the location of synapse. The axonal length may be different from the overall connection length depending on the proximity or separation of the pre-synaptic neuron from the synapse or junction.
Temporal closeness in early development would imply spatial proximity, therefore reducing hurdles in wiring neurons that are far apart in the adult. It is therefore likely that early contact with subsequent axon extension will be preferred. This presumably would aid in minimizing metabolic costs sustained in wiring establishment during development, as axon outgrowth is an inherently more resource intensive pathway than axon extension [14], [15]. Note that this cost minimization is only with respect to establishing the wiring. It should not be compared with sustained metabolic cost optimization through alternative wiring or cell migration costs that would be incurred irrespective of the time of synaptogenesis. The C. elegans neuronal network that has been the subject of several studies, is understood to be wired for efficiency of processing [2], [10]. It has also been shown that while this spatial organization is not the one that minimizes total wiring length, the neurons are placed close to their optimal positions within the worm body for reducing wiring [16]. Here our results suggest that C. elegans has also evolved a pattern of development that enables it to establish this known connection pattern (particularly the substantial number of long-range connections) efficiently during growth.
The need for early contact has also been observed in other systems including the development of the excretory canal in C. elegans [17]. The canal tip is guided to and makes contact with the basement membrane early on in development. Subsequent growth involves a passive extension of the canal along the hypodermal membrane. It has been found that absence of this early contact, results in a slowing of growth in the larval stage, with the canal being unable to reach its target region [18]. The influence of early guidance has also been observed in the migration of the sex myoblast [18], [19]. There are two mechanisms involved in guiding the sex myoblasts to the gonads, one acting early and the other much later, with absence of early guidance resulting in delays in the migration.
It is generally accepted that the most significant aspect of forming a synapse is contact between the neurons or neurites. While the role of pioneer axons has been studied in C. elegans [20], the fact that nearly 70% of connected neurons appear much before hatching, makes it more likely that cell-cell contact or interaction is established early on, particularly within circuits having functional significance in early life. There is evidence showing that several guidance molecules like Netrin and Nerfin-1 are expressed in the early stages of development. Netrin is a protein involved in axonal guidance in vertebrates as well as in invertebrates [21], [22], [23] and is specifically known to influence early path-finding events [23], [24], [25]. Nerfin-1 belonging to a highly conserved family of Zn-finger proteins, is found transiently expressed in neuron precursors and plays a role in early path finding. Studies involving pioneering neurons in the central nervous system of Drosophila melanogaster have shown that Nerfin-1, whose expression is spatially and temporally regulated [25], is essential in early axonal guidance. Disruption of genes involved in early guidance and migration could result in improper connectivity [26], [27], as could delays in cell division. Our results suggest that neuron-neuron contact in early development could be important for achieving accurate connectivity. This would be significant not just in invertebrates but also in vertebrates where similar temporal expression profiles are seen for genes involved in early guidance. Inaccurately timed appearances of long-range connections or pioneer neurons would also influence subsequent connectivity. In addition, in C. elegans, as over two-thirds of long distance connection-pairs are born early on, it is likely that the interactions among the growing axons will play a role in fasciculation as observed in Drosophila [26].
The analysis of the formation of connection-pairs with respect to the nature of the junction has highlighted another interesting feature while also raising an important question. Approximately a third of all junctions are electrically coupled with around 70% each of long- and short-range connection-pairs appearing early in development. While little is known about the formation of gap junctions, they are most often seen in adjacent cells and hence it is likely that the 70% of connection-pairs appearing early will connect early. However, the interesting question raised here is of the remaining 30% of long-range connections coupled by gap junctions that appear in the later stages of development. Specifically, how are the neurites guided to their correct targets for the formation of gap junctions? The role of connexin proteins [28] and cell adhesion molecules like cadherins [29] in gap junction formation has been studied, however they would be insufficient to explain the formation of long-range axonal gap-junctions between neurons that are far apart. Theses gap junctions are mostly between late-forming ventral cord neurons and the existing pharyngeal ring neurons in the head. As the ventral cord neurons appear at later stages in development when the worm has already grown to approximately half of its final size, neuronal contact with subsequent migration is highly unlikely. While it is interesting to consider a dependence on chemically guided axons through fasciculation, in the absence of experimental data, these remain speculations.
Another important finding was that availability of time would be an important factor in the generation of network hubs. In the C. elegans network, nearly all neurons (∼98%) with degree more than 30 were born before hatching, whereas only 74% (±5.5) would have occurred for random growth. Developmental time windows have been identified in the past [30] as playing a role in generating network hubs. A high probability in random networks, with an even higher value in the actual network, is indicative of the importance of time availability in establishing hubs. Indeed functional significance cannot be undermined and it needs to be emphasized that while being essential for achieving high number of connections, time is unlikely to be the deterministic factor attributing high degree connectivity to a particular neuron. Additionally, bilaterally symmetric neurons were born close to each other in time. Hence although time of birth is known to play an important role [31], [32], [33], in C. elegans, the asymmetry in function is less likely to be a result of environment-mediated change.
The neuronal network of C. elegans, being the only fully characterized connectome to date, has provided the opportunity to observe changes occurring during the course of its neural development. Based on available data, we have extracted time of creation of neurons to capture changes during growth and have identified features of neural development that would be significant in establishing long-range connections and network hubs. Continuous monitoring of synaptic connectivity during development is a steep challenge that goes beyond the current approaches for determining the adult connectome of different species [34], [35]. We present an alternative, more feasible approach of employing global analysis on networks existing at discrete times of growth. With the aid of recent advances, obtaining connectivity information for these time stages could be possible in the near future. Availability of such data on connectivity will permit more detailed analyses, to give an insight into the structural changes unfolding during development. An interesting question that has been raised here is what mechanism ensures accuracy of wiring in late-forming, electrically coupled long-distance connections, and a clear answer is as yet unavailable. We hope that this work will stimulate further experimental and theoretical work on the network development of neural systems and C. elegans in particular.
We have produced a spatial representation of the neuronal network of C. elegans in three-dimensional space, so that the network resembled the anatomical network as much as possible. Three-dimensional coordinates were based on the two-dimensional spatial information of C. elegans neurons [36], which were updated with spatial information of three neurons that had been excluded. The data for neurons that did not have associated spatial information were obtained based on the spatial data of corresponding bilateral counterparts. The connectivity details from earlier studies [12], [16], [37] published in the Worm Atlas was used for the analysis. The ventral cord neuron VC6 that only makes connections through neuro-muscular junctions was not included here. Three loop connections (connections of a neuron to itself) were also excluded; as such connections did not influence our spatial and topological measures. Thus a total of 279 neurons and corresponding 2,990 connections were used. This included 1,584 uni-directional and 1,406 bi-directional connections. Biologically, they represent 672 gap junctions, 1962 chemical sysnapses and 376 connections where both gap junctions and chemical synapses exist between the neuron pairs. The latter were represented as bi-directional edges in the connection matrix. Neuro-muscular junctions were not included in the analysis. The coordinate information represents the positions of the soma of the neurons in three-dimensional space. The third dimension of each of the neurons was obtained by treating the body of the worm as a cylinder, guided by the actual shape of the worm. The third spatial coordinate for any neuron was then calculated as a function of the radius of the body of the worm and its known position along the y-axis, as follows:where, r – radius of the worm, was assumed to be constant along the length of the nematode (50 µm). This returned a positive value of z, and as many neurons in C. elegans have a left or right orientation, the physiological information available [5] was also used in determining the third coordinate. The three-dimensional coordinate was computed so that the spatial properties were closer to reality. Although as C. elegans has a very high length to diameter ratio, the results are unlikely to be affected even if the data had been two-dimensional. The left and right neurons were differentiated into positive and negative values, while those lying along the dorsal and ventral line had their third coordinate as zero.
To trace the growth of the network over time, we used the time estimates provided by Sulston et al. (1977 & 1983), in the cell lineage charts. The image files representing the lineage charts were read in and then analyzed to obtain the time of creation of each of the neurons. Based on this information, the neuronal network of C. elegans was obtained at different stages of growth. The margin of error in the embryonic lineage, as published, is 10% and 2% in the post-embryonic lineages. We produced spatial representations of the network at times of 350, 400, 500, 800, 2000 and 2700 minutes after fertilization. The embryo hatches at around 840 minutes [3], and the network at that time stage was found to be identical to that at 600 minutes in terms of neurons present. The choice of the time interval between the successive stages was based on the number of neurons appearing at different stages. It can be seen from the postembryonic lineage chart that there are very few neurons being created within first 20 hours after hatching (1200 minutes after hatching or 2000 minutes from fertilization). We therefore chose 2000 minutes post-fertilization as the next stage of development. As all but two neurons in C. elegans derive from the AB cell lineage, cross lineage comparisons were not performed. Although the network contained chemical synapses connecting one neuron to another as well as gap junctions coupling both neurons, the networks were treated as unweighted and directed as more than half of the connections (53%) were unidirectional. Gap junctions were represented as bi-directional. The node degree included both the incoming and outgoing connections.
Random networks created for comparative analysis had the same degree distribution as the actual C. elegans network. Neuron identities were randomly shuffled, so that all quantities estimated such as connection-length, birth-time difference, etc, would be modified. A neuron's identity referred to its spatial position and birth-time.
At any given time, a connection-pair existed if both neurons forming that connection (as in the adult) were present, without however, implying the formation of a synapse between them. The pair-wise, time-difference in the birth of neurons forming each of the 2990 connection-pairs was determined. The values were then separated into ten bins and mean of the time-differences in each bin was computed for enabling comparisons with random networks. The connection length between two connected neurons was the Euclidean distance between them in three-dimensional space. The lengths at each stage were separated into ten bins of size 0.12 mm each, the first three were considered to be short distance (i.e. shorter than 0.36 mm), the middle four as medium-range and the final three as long-range connections (i.e. longer than 0.84 mm). This classification was used as the first three bins and the final three bins displayed a very high frequency of connections, whereas the intermediate bins were sparsely populated.
|
10.1371/journal.pntd.0001518 | The Diverse and Dynamic Nature of Leishmania Parasitophorous Vacuoles Studied by Multidimensional Imaging | An important area in the cell biology of intracellular parasitism is the customization of parasitophorous vacuoles (PVs) by prokaryotic or eukaryotic intracellular microorganisms. We were curious to compare PV biogenesis in primary mouse bone marrow-derived macrophages exposed to carefully prepared amastigotes of either Leishmania major or L. amazonensis. While tight-fitting PVs are housing one or two L. major amastigotes, giant PVs are housing many L. amazonensis amastigotes. In this study, using multidimensional imaging of live cells, we compare and characterize the PV biogenesis/remodeling of macrophages i) hosting amastigotes of either L. major or L. amazonensis and ii) loaded with Lysotracker, a lysosomotropic fluorescent probe. Three dynamic features of Leishmania amastigote-hosting PVs are documented: they range from i) entry of Lysotracker transients within tight-fitting, fission-prone L. major amastigote-housing PVs; ii) the decrease in the number of macrophage acidic vesicles during the L. major PV fission or L. amazonensis PV enlargement; to iii) the L. amazonensis PV remodeling after homotypic fusion. The high content information of multidimensional images allowed the updating of our understanding of the Leishmania species-specific differences in PV biogenesis/remodeling and could be useful for the study of other intracellular microorganisms.
| Leishmania parasites lodge in host cells within phagolysosome-like structures called parasitophorous vacuoles (PVs). Depending on the species, amastigote forms can be individually hosted within small, tight-fitting PVs or grouped within loose, spacious PVs. Using multidimensional live cell imaging, we examined the biogenesis of the two PV phenotypes in macrophages exposed to L. major (a representative of the tight PV phenotype) or L. amazonensis (an example of the loose PV phenotype) amastigotes. L. major PVs undergo fission as parasites divide; we demonstrate that in the course of fission there are transients of the lysosomotropic fluorescent probe Lysotracker. In contrast, during the course of amastigote population size expansion, L. amazonensis PVs do accumulate Lysotracker while increasing in diameter and volume. The large PVs fuse together, and the products of fusion undergo size and shape remodeling. The biogenesis/remodeling of the two types of Leishmania PVs is accompanied by a reduction in the number of macrophage acidic vesicles. The present imaging study adds new morphometric information to the cell biology of Leishmania amastigote intracellular parasitism.
| Leishmania spp. are dimorphic trypanosomatid parasites that alternate between extracellular promastigote forms found in insect vectors and intracellular amastigote forms found in mammalian hosts. In infected cells, Leishmania amastigotes are sheltered within phagolysosome-like structures called parasitophorous vacuoles (PVs). The PV membranes and contents change as PVs fuse with the endoplasmic reticulum (ER), late endosomes, lysosomes, or other host cell vesicular elements, conferring to them distinctive properties and a hybrid nature [1]–[3].
In the majority of Leishmania species, including L. major, one or two amastigotes are enclosed within PVs, which display a modest vacuolar space. In contrast, the large PVs that shelter parasites of the L. mexicana complex, such as L. amazonensis, can contain numerous amastigotes, often bound by their posterior poles to the internal face of the PVs [4]. The biogenesis of these two types of PVs involves the acquisition of host cell late endosomes membrane markers, as shown in infected cells immunostained for lysosome-associated membrane proteins (LAMPs), Rab GTPases, cathepsin, proton ATPases, and MHC class II molecules [2], [5]–[8]. The acquisition of these markers is a coordinated event that results in a “mature” PV, which is presumably required for the survival and multiplication of the parasites.
Because the relatively large dimensions of their PVs that allow them to be easily recognized at low magnification, L. amazonensis and L. mexicana PVs have often been used in studies of the fusogenic properties of Leishmania PVs. These PVs could be demonstrated to selectively fuse with each other or with phagosomes containing macromolecules, colloids, inert particles, and other live parasitic microorganisms [9]–[17]. The fusogenicity and easy access of certain particles and molecules to these large structures increase with the duration of infection [1]. The spacious PVs also incorporate acidic pH markers such as Lysotracker and neutral red [17], [18] and may be probed using pH-sensitive dyes [19], [20]. By taking advantage of differences in fluorescein emission under different pH conditions, it was reported that the pH of L. amazonensis PVs falls from approximately 5.2 at 24 h to 4.8 after 48 h of intracellular infection, whereas the pH of secondary lysosomes of around 5.4 remained constant in non-infected control cells [19].
These studies led to the characterization of the biochemical and functional features of Leishmania PVs, which may not apply to the majority of Leishmania species studied that are lodged in tight PVs and assumed to undergo fission as parasites divide [21], [22]. The accessibility of particles, macromolecules, and probes to these tight-fitting PVs and the identification of their contents are hindered by the limited vacuolar space available between the parasites and their PV membranes. Chang and Dwyer [21] and Berman and colleagues [23] observed by electron microscopy that thorium dioxide (“Thorotrast”) particles, which were pre-loaded in lysosomes, were transferred to the small vacuolar space of L. donovani and L. major PVs, respectively. The granules were absent within PVs when parasites and PV membranes were only in close contact. These studies suggest that tight-fitting Leishmania PVs can fuse with lysosomes, although the retention of lysosomal markers differs accordingly to PV dimensions. Additionally, the pH in tight-fitting PVs may be different from that within loose vacuoles: the pH within L. donovani tight PVs was reported to reach 5.5 after 2 h of infection [24].
Most of the available information on Leishmania PV biogenesis has been obtained by experiments on fixed cells, a drawback we sought to overcome in the present study. We examined the biogenesis of large or tight-fitting, membrane-bound Leishmania PVs recorded by the multidimensional imaging of live infected macrophages. The fission of Leishmania tight-fitting PVs was studied for the first time in live infected cells and characterized as a two-step process that involves the replication of amastigotes in a single PV prior to separation into two distinct PVs that accumulate transient amounts of lysosomotropic probe. The process is accompanied by the depletion of macrophage small acidic compartments, as previously described for Leishmania large vacuoles [25]. The biogenesis of these large structures was also studied and revealed to involve PV enlargement in volume and diameter to the detriment of other PVs in the same infected cells. The homotypic PV fusion between L. amazonensis PVs was recorded and involves PV volume restoration.
All experiments involving animal work were conducted under Brazilian National Commitee on Ethics in Research (CONEP) and French National Committee on Ethics and Animal Experimentation (CNREEA) ethic guidelines, which are in accordance with international standards (CIOMS/OMS, 1985). The present study was approved by CEP/UNIFESP (Comitê de Ética em Pesquisa da Universidade Federal de São Paulo/Hospital São Paulo) under the protocol number 0856/07.
BALB/c and BALB/c nude mice (8 weeks of age) were used as sources of bone marrow macrophage precursor cells and lesion-derived Leishmania amastigotes. Macrophages were obtained from bone marrow precursor cell suspensions cultivated in vitro for 7 days in RPMI 1640 medium with 10% fetal calf serum, 5% L929 cell conditioned medium, 100 U/ml penicillin, and 100 µg/ml streptomycin [7]. RAW 264.7 macrophage-like cells were cotransfected with LAMP1-GFP and Rab7-GFP plasmids (using FuGene HD transfection reagent, ROCHE) kindly donated by Dr. Norma Andrews (Maryland University) and employed in infection experiments in order to observe PV membranes in live recordings.
Macrophages were transferred to glass coverslips or round dishes (ibidi, GmbH or Mattek Corporation) suitable for maintaining living cells in incubators coupled to microscopes. Before their use in experiments, cultures were incubated overnight at 37°C in a humidified air atmosphere containing 5% CO2.
BALB/c nude mice footpads were inoculated with wild-type L. (L.) amazonensis LV79 (MPRO/BR/72/M1841) or DsRed2-transfected L. (L.) major NIH173 (MHOM/IR/-/173). Isolation of amastigotes from footpad lesions was performed as previously described [26] after 2 months of inoculation.
Leishmania amastigotes were added to macrophage cultures at a multiplicity of infection of 5 and incubated at 34°C and 5% CO2 in complete medium for different periods according to the experiment. Cultures were washed with Hanks' Buffered Salt Solution to remove free parasites and cultivated in complete medium at 34°C in a 5% CO2 atmosphere. Observation and image acquisition of live or fixed macrophage cultures under the employed microscopes started after periods ranging from 2 to 48 h depending on the experiment. Cultures were maintained at 34°C and 5% CO2 within the incubators coupled to the microscopes.
Macrophages on coverslips were washed and fixed for 1 h with 3.5% formaldehyde in phosphate-buffered saline (PBS). Leishmania PVs and other compartments were identified by immunolabeling of the membrane proteins LAMP1 and LAMP2 (monoclonal antibodies obtained from DSHB, Iowa University, USA). L. amazonensis amastigotes were immunostained with 2A3-26 antibody conjugated to FITC (kindly provided by Dr. Eric Prina, Institut Pasteur, France) or loaded with 5 µM 5,6-carboxyfluorescein diacetate succinimidyl ester (CFSE, Invitrogen, Life Technologies). Samples were stained for 15 min with 100 µg/ml 4′,6-diamidino-2-phenylindole (DAPI, Invitrogen, Life Technologies) and mounted with 50% glycerol in PBS containing 0.01% p-phenylenediamine. Confocal images were obtained using a Bio-Rad 1024 UV system coupled to a Zeiss Axiovert 100 microscope or a Leica TCS SP5 II system. Images acquired with a 100× (1.44 NA) oil immersion objective were rendered with Imaris Software (Bitplane AG) by using Blend filters.
Live imaging of cultures was performed using a Nikon Biostation IM-Q Live cell recorder system (Nikon Corporation), a Perkin-Elmer UltraView RS Nipkow-disk system (PerkinElmer Inc.) attached to a Zeiss Axiovert 200 M microscope with a Hamamatsu ORCA II ER CCD camera, or in a Leica TCS SP5 II system (Leica Microsystems). To identify Leishmania PVs, 50 nM Lysotracker green DND-26 (Invitrogen, Life Technologies), a lysosomotropic probe for acidic compartments, was added to complete medium 1 h before microscopic recordings and maintained throughout image acquisition. In contrast with L. amazonensis PVs that were rich in Lysotracker, vacuoles containing a single L. major-DsRed2 parasite displayed a feeble Lysotracker signal, possibly due to the lower acidity of their PVs. The effect of a tight vacuolar space between PV membranes and amastigotes on Lysotracker intensity is not discarded, although Lysotracker is a relatively small molecule (MW = 398.69). Alternatively, 1 mg/ml FITC-dextran (average mol wt 42,000, Sigma-Aldrich) was used as lysosomotropic probe, with a 1 hour pulse, removal of the probe by 6 washings prior to image acquisition.
The Nikon Biostation IM-Q was used to acquire, in 10 different microscopic fields, serial images of infected macrophage cultures in dishes. The Biostation acquired images in phase contrast and in two fluorescent channels (for Lysotracker- and DsRed2-labeled parasites) with a 40× (0.8 NA) objective in 5-min intervals. Points in time of time-lapse image acquisitions are displayed as day, hours, and minutes (d:hh:mm). Images of the DsRed2 fluorescence signal displayed by L. major were processed by Acapella software (PerkinElmer Inc.) for algorithm-based quantification of these parasites during infection in macrophage cultures [17].
The Perkin-Elmer UltraView RS and the Leica TCS SP5 II system were used to acquire stacks of 20 to 30 optical sections from live infected cells in 5 to 12 microscopic fields. Stacks along the z- axis (z-stacks) were obtained with an optical section separation (z-interval) of 0.2 to 1 µm.
We acquired images of infected cell cultures after different post-infection times. Some image acquisitions began 2 h after parasite addition to macrophages; other started after 24 h or 48 h. Thus, we chose to present temporal data as “time of image acquisition” instead of “time of infection” due to different time ranges of intracellular infection. The time of multidimensional acquisition is displayed as d:hh:mm.
All statistics were performed using SPSS software (SPSS Inc.). The statistical tests employed are indicated in the figure legends, and they were chosen on the basis of normal and non-normal distributions and equal and non-equal variances. Figures represent the same result reproduced in at least 2/3 out of all analyzed multidimensional images acquired in 3 different experiments, using 2 different equipments.
PVs are the host intracellular compartments in which Leishmania parasites differentiate and multiply. They are lined by a dynamic membrane originated at the host cell plasma membrane and formed by successive and coordinated fusion/fission events with vesicles from early/late endocytic pathways, secondary lysosomes, ER, and possibly autophagic vesicles, resulting in an acidic compartment similar to phagolysosomes [3], [5], [7], [19], [27]. Most of this information has been obtained in experiments with Leishmania cell-cycling amastigotes of the L. mexicana group in fixed preparations.
This study presents high-resolution morphological characterization of the main features displayed by Leishmania PVs in live infected macrophages, such as PV volumetric expansion/retraction and homotypic PV fusion/fission, using the fluorescent lysosomotropic probe Lysotracker. Additionally, a software-based methodology permitted the quantification of host cell acidic vesicles, which were distinguished from Leishmania PVs in the same fluorescence channel. As secondary lysosomes are considered the main source of membrane for Leishmania PV biogenesis, we investigated how macrophage acidic vesicles reservoirs (which include, but are not restricted to secondary lysosomes) were related to Leishmania PV biogenesis.
The development of spacious PVs is an exception rather than a rule for most Leishmania parasites studied. Soon after the internalization of L. amazonensis, PV membranes acquire Rab5 and EEA1, two early endosomes markers; by membrane remodeling, PV membranes lose these early markers and rapidly acquire the late endosome and lysosome markers Rab7p and LAMP1 [7]. In the first hours of intracellular infection, fusion between PVs and lysosomes could be a determinant of PV enlargement, as the depletion of secondary lysosomes [25] and acidic vesicles, as shown in this report, is observed during parasite establishment. Additionally, the up-regulation of host cell lipid biosynthesis triggered by the parasite [28] could increase the repertoire of membrane donors to Leishmania PVs development. This up-regulation was not observed in the analysis of gene expression of macrophages infected with L. major promastigotes [29].
After lysosomal marker acquisition (by means of fusion with acidic vesicles, lysosomes and/or late endosomes), the membrane input into Leishmania PVs and their stability as large compartments could be related to continuous fusion with ER-derived or Golgi-derived vesicles [1], [3] instead of fusion with secondary lysosomes, late endosomes, or vicinal L. amazonensis PVs.
The homotypic fusion between L. amazonensis PVs was described by Real and colleagues and occurs mainly after 24–48 h of intracellular infection [16]. Fusion between PVs before 12 h of infection was seldom observed. Indeed, Lippuner and colleagues found that a minority of Rab5-positive L. mexicana PVs fuse with each other in the first hours of intracellular infection [30]. The fusion between L. amazonensis PVs did not contribute to sustain PV enlargement because fused PVs regained their dimensions, a process probably related to compensatory membrane recycling. Although the results suggest a selective nature for PV membrane composition, the host cell/parasite components retained by large PVs to maintain their homeostasis remain to be elucidated. The output of membrane from large PVs could be incorporated by the plasma membrane (displaying parasite proteins on host cell surface) and other cytoplasmic organelles such as the ER.
The spacious PV represents a strategy to subvert host cell defenses, providing an environment with lower activities and/or concentrations of hydrolytic enzymes [9], [31]–[33]. Considering that Leishmania from the L. mexicana complex are resistant to IFN-γ–mediated macrophage activation and high NO concentrations [34], the unique morphology of large vacuoles could be vital for the establishment of these parasites. Recent studies documented that L. major procyclic promastigotes survive and multiply within L. amazonensis PVs after interspecific PV fusion in doubly infected macrophages [17]. As expected, the procyclic promastigotes were destroyed within their own phagolysosomes but were spared from destruction once they entered the spacious L. amazonensis PVs.
L. major, the first organism in the genus to have its genome sequenced [35], is one of several Leishmania species in which amastigotes develop tight PVs that undergo fission during parasite multiplication in host cells. These amastigotes are generally individualized in PVs that maintain their dimensions and possibly require different strategies to avoid host cell defenses. PV fission is displayed by several important pathogens that live in vacuoles throughout their intracellular life cycle, such as Leishmania and Mycobacterium [36]. The process is likely to require the mobilization of membrane sources because replicating amastigotes require increases in PV dimensions prior to division, which implies an increase in PV membrane surface area and volume.
After parasite division, there is an intermediary state of double-occupancy PVs in which a vacuolar interface between recently-divided parasites and the PV membrane is visible using lysosomotropic probes. This detectable vacuolar space is polarized in agreement with the polarized event of amastigote division, which suggests a period of hours in which two dividing L. major amastigotes share the same PV with a small, acidic lumen. Indeed, in the late 1970s, a Thorotrast-rich vacuolar space located between two dividing L. major amastigotes was documented by electron microscopy [23]. Amastigote division within tight PVs could cause increase in PV fusogenicity with small vesicles by modifying the membrane curvature, which could physically assist SNARE machinery for membrane fusion [37]. An alternative interpretation of the Lysotracker clusters on dividing L. major amastigotes is that acidic vesicles could be mobilized to PV membrane sites in hotspots, where membrane input preferentially occurs.
The numeric decrease of host acidic vesicles during L. major PV fission or L. amazonensis PV enlargement is likely related to a higher demand of host membrane sources for the biogenesis of Leishmania PVs, at least in the first 4 days of intracellular infection. The host cell reservoirs of acidic vesicles would partially account for membrane incorporation of fission-prone L. major PV and partition in two new PVs or massive incorporation of vesicles into enlarging L. amazonensis PVs. Considering the membrane surface area of Leishmania PVs, a large L. amazonensis PV with 20 µm of diameter has the approximate surface area of 16 L. major tight-fitting PVs, indicating that Leishmania PVs would require approximate amounts of host cell membrane for their biogenesis regardless of the different PV architectures (tight-fitting vs. loose vacuoles). The contribution of each acidic vesicle to PV biogenesis may be also hypothesized: approximately 50 detected acidic vesicles (1 µm in diameter) in macrophages infected with L. major or L. amazonensis are consumed in the course of intracellular infection. This represents a hypothetical volume contribution of 30 µm3 to Leishmania PVs, which only partially accounts for L. major or L. amazonensis PV dimensional doubling.
Indeed, other mechanisms of Leishmania PV volume control were addressed in the literature and include fusion with ER vesicles, acquisition of water and ion transport channels, and parasite secretion of exosomes and macromolecules inserted in PV membranes, or displayed in parasite membrane surfaces [3], [33], [38]–[41]. Additionally, Leishmania can internalize PV membrane components: in a process resembling host membrane “clearance,” L. amazonensis can actively internalize and digest MHC class II, possibly by its posterior pole that interacts with PV membranes [42]. This process could also be conserved in Leishmania with membrane-bound PVs participating in the control of membrane input into the PVs. The polarization of lysotracker clusters in dividing L. major parasites suggests that parasite poles could also participate in the PV biogenesis of tight-fitting phenotype.
Although L. major PVs incorporate phagolysosomal markers on their membranes, they were unable to retain amounts of lysosomotropic content probes comparable to other tight-fitting phagosomes (i.e. latex beads or aldehyde-fixed amastigotes). A less acidic environment could account for Lysotracker-negative L. major PVs, which could present lower abundance of vesicular proton ATPases than L. amazonensis PVs. Osorio y Fortea and colleagues [28] showed that eight isoforms of vesicular proton ATPases subunits are up-regulated in macrophages infected with L. amazonensis amastigotes; in contrast, Gregory and colleagues [29] showed that only one isoform of these ATPases, V1 subunit H, is up-regulated in macrophages infected with L. major promastigotes. In these same experiments, a V0 subunit A2 is down-regulated and other lysosomal components (such as two isoforms of a lysosomal-associated transmembrane protein 5 and an isoform of acid phosphatase 2) are down-regulated. Non-acidic phagosomes displaying phagolysosomal markers on their membranes were associated with exclusion of these proton ATPases in Mycobacterium phagosomes [43]. It is possible that L. major PVs maintain the exclusion of vesicular proton ATPases for amastigote replication, in a process similar to what occurs in L. donovani promastigote PVs [44].
By presenting these fundamental pre-mechanistic data obtained by live imaging, we highlighted that Leishmania PVs are dynamic structures that remodel their shapes, allowing these structures to develop a privileged intracellular niche where Leishmania parasites survive and multiply. Several pathogens can utilize this process; describing and understanding the nature of these vacuoles in live infected cells are crucial steps towards the understanding of how parasites evade host immune responses [45].
|
10.1371/journal.pgen.1006037 | Molecular Characterization of Three Canine Models of Human Rare Bone Diseases: Caffey, van den Ende-Gupta, and Raine Syndromes | One to two percent of all children are born with a developmental disorder requiring pediatric hospital admissions. For many such syndromes, the molecular pathogenesis remains poorly characterized. Parallel developmental disorders in other species could provide complementary models for human rare diseases by uncovering new candidate genes, improving the understanding of the molecular mechanisms and opening possibilities for therapeutic trials. We performed various experiments, e.g. combined genome-wide association and next generation sequencing, to investigate the clinico-pathological features and genetic causes of three developmental syndromes in dogs, including craniomandibular osteopathy (CMO), a previously undescribed skeletal syndrome, and dental hypomineralization, for which we identified pathogenic variants in the canine SLC37A2 (truncating splicing enhancer variant), SCARF2 (truncating 2-bp deletion) and FAM20C (missense variant) genes, respectively. CMO is a clinical equivalent to an infantile cortical hyperostosis (Caffey disease), for which SLC37A2 is a new candidate gene. SLC37A2 is a poorly characterized member of a glucose-phosphate transporter family without previous disease associations. It is expressed in many tissues, including cells of the macrophage lineage, e.g. osteoclasts, and suggests a disease mechanism, in which an impaired glucose homeostasis in osteoclasts compromises their function in the developing bone, leading to hyperostosis. Mutations in SCARF2 and FAM20C have been associated with the human van den Ende-Gupta and Raine syndromes that include numerous features similar to the affected dogs. Given the growing interest in the molecular characterization and treatment of human rare diseases, our study presents three novel physiologically relevant models for further research and therapy approaches, while providing the molecular identity for the canine conditions.
| Rare developmental disorders make a major contribution to pediatric hospital admissions and mortality. There is a growing interest in the development of therapeutics for these conditions, but that requires understanding of the genetic cause and pathology. Research can be facilitated by physiologically relevant models, such as dogs with corresponding disorders. We have characterized the clinical features and genetic causes of three developmental syndromes in dogs, including craniomandibular osteopathy (CMO), a previously undescribed skeletal syndrome, and dental hypomineralization, for which we identified mutations in the canine SLC37A2, SCARF2 and FAM20C genes, respectively. CMO is a clinical equivalent to an infantile cortical hyperostosis (Caffey disease) for which SLC37A2 is a new candidate gene. SLC37A2 is a glucose-phosphate transporter in osteoclasts, and its defect suggests an impaired glucose homeostasis in developing bone, leading to hyperostosis. Mutations in the SCARF2 and FAM20C genes have been associated with the human van den Ende-Gupta and Raine syndromes. Our study provides molecular identity for the canine conditions and presents three novel physiologically relevant models of human rare diseases.
| One to two percent of all children are born with a developmental disorder, such as a heart defect, skeletal abnormality, or mental retardation as a result of errors in embryogenesis and early neurodevelopment. These disorders make a major contribution to pediatric hospital admissions and mortality [1]. Rare pediatric disorders are typically with homozygous, compound heterozygous, or de novo pathogenic variants. The advent of cost-efficient next generation sequencing (NGS) technologies drives gene discovery in many disorders without a cognate gene [2] and thousands of rare variants have been described (www.orpha.net). There is also a growing interest in the development of therapeutics for rare diseases, which requires the identification of the genetic defects, comprehensive understanding of the molecular pathology and access to physiologically relevant animal models.
Developmental disorders are frequent also in other species, including dogs, which as large animals bear very close physiologic and genetic resemblance with us. Dogs give birth to litters with multiple puppies. However, often some of the littermates are affected by developmental or other abnormalities, and perinatal mortality (stillbirth, fetal and neonatal death) is common with prevalence ranging from 5 to 35% [3]. The causes include respiratory distress syndrome/hypoxia, infectious diseases, severe malformations and suspected hereditary diseases. Culling is a common practice among dog breeders and the deceased or abnormal puppies are not always presented to the veterinarian. Therefore, numerous syndromes with likely genetic origin remain unknown. It would be important to investigate the extent of this common phenomenon at the clinical and molecular level to better understand the diverse causes of the morbidity and to better manage it through advised breeding programs. At the same time, the identification of the causative gene could inform gene functions, disease etiology, molecular pathology and phenotypic overlap across species. Importantly, physiological similarity of dogs with human would establish relevant therapeutic models to human rare disorders. Online Mendelian Inheritance in Animals (OMIA), a catalogue of inherited disorders and associated genes in animals, reports more than 350 inherited diseases in dogs as potential models for human disease (http://omia.angis.org.au/).
NGS approaches are rapidly changing the diagnostic landscape in veterinary medicine in companion animals and enable now a feasible approach to tackle the molecular background of developmental conditions in small pedigrees with translational potential to human rare disease. For example, we have previously discovered a new gene (ATG4D) responsible for a neurodegenerative vacuolar storage disease in Lagotto Romagnolos [4] and a missense variant in canine FAM83G causing palmoplantar hyperkeratosis and demonstrating its role in maintaining the integrity of the palmoplantar epidermis [5]. A CNGB1 frameshift variant has been identified to cause a progressive retinal atrophy in dogs [6] and the same gene has been associated with retinal degeneration in human as well [7]. Similarly, we have found that a congenital skeletal disease in Brazilian Terriers is caused by a pathogenic variant in the GUSB orthologue responsible for human pediatric disorder, mucopolysaccharidosis type VII [8].
This study addressed the clinical and genetic background of three developmental disorders in dogs; craniomandibular osteopathy (CMO; OMIA: 000236–9615) in West Highland White Terriers (WHWT), Cairn Terriers and Scottish Terriers, a previously undescribed developmental syndrome in Wire Fox Terriers, and dental hypomineralization in Border Collies. CMO is a self-limiting proliferative bone disease seen in young dogs [9]. It manifests between 4 to 8 months of age with typical signs including swelling of the jaw, periodical fever, lack of appetite, pain, difficulty opening the mouth and dysphagia. The excessive proliferation causes bony lesions primarily on the skull bones, especially on the mandible and tympanic bulla, but occasionally also on the metaphyses of long bones. Signs of the disease usually resolve with time, when the growth period is finished. CMO exists in several breeds with the highest frequency in WHWT and has been suggested to be an autosomal recessive trait [10–12]. Canine CMO corresponds to human Caffey disease [MIM: 114000] and its genetic characterization might reveal insights into similar painful human swelling disorders [10]. We identified a novel CMO gene that represents a candidate gene for human Caffey disease. The two other syndromes have not previously been reported in dogs. We describe the detailed clinical features for them in our study. Moreover, we demonstrate their shared genetic etiology with the corresponding human syndromes, van den Ende-Gupta (VDEGS [MIM: 600920]) and Raine syndromes [MIM: 259775].
We performed a genome-wide association study (GWAS) to map the CMO locus with Illumina’s 22K canine SNP chips in a cohort of 51 WHWTs, including 10 cases (diagnosed by radiography; Fig 1A) and 41 controls. A case-control association test revealed significant association on CFA5 with the best SNP (BICF2S23544899) at 8,953,507 Mb (praw = 1.2 x 10−7, pgenome = 0.02) (genomic inflation factor λ = 1.17) (Fig 1B). Manual assessment of genotypes at the CFA5 locus revealed a shared 1.9-Mb homozygosity block in all affected dogs spanning from 7,764,955 bp to 9,707,794 bp. The same homozygosity block was seen in five controls as well. The associated region was replicated and fine-mapped using 105 additional SNPs in 88 samples from three related breeds, WHWT, Cairn Terriers and Scottish Terriers. Fine mapping confirmed the association in all three breeds with the best SNP BICF2S23134295 at 8,183,669 (p = 2.09x10-15).
To identify candidate variants, the associated and fine mapped region (1.8 Mb) was captured from two affected and two healthy WHWT with opposite haplotypes followed by a paired-end NGS by HiSeq2000. The shared variants identified in the two affected WHWT dogs were filtered against the two WHWT controls and 32 additional healthy dogs from four different breeds. We found altogether three homozygous variants shared in cases (S1 Table); two intergenic indels and a synonymous variant in exon 15 of solute carrier family 37 member 2 gene (SLC37A2 c.1332 C>T) (Fig 1C). As an additional independent verification to avoid potential targeted capture biases, we performed whole genome sequencing in one CMO-affected WHWT and 188 control dogs from other breeds (S2 Table) to compare the variants in the associating region. This analysis yielded a single case-specific variant (chr5 g.9,387,327G>A) that was the same as the one identified in the capture experiment.
Although the variant, c.1332C>T in exon 15 of SLC37A2 is synonymous, it was predicted to affect a splicing enhancer element based on ESEfinder analysis (Fig 1D). The mutant T allele eliminates a potential binding site for the splicing factor ASF/SF-2. To confirm the predicted effect on splicing, we amplified the region between exons 7 and 18 in lymphocyte mRNAs from three cases, six carriers and seven controls. RT-PCR experiments showed that two alternatively spliced SLC37A2 transcripts were expressed in all dogs that carried one or two copies of the mutant T allele at the SLC37A2 SNP, regardless of disease status (Fig 1E). Sequencing of the RT-PCR products indicated that the smaller band corresponded to a mutant SLC37A2 transcript lacking 79 bp from exon 15. The splice variant resulted in a frameshift and premature stop codon at the beginning of exon 16. This altered splicing was predicted to lead to a C-terminally truncated protein lacking 75 amino acids compared to the wild-type SLC37A2 protein (Fig 1F). The RT-PCR indicated that both wild-type and mutant transcripts were expressed in even the CMO affected dogs, although the expression of the wild-type transcript was significantly reduced in the affected homozygous dogs (Fig 1E). The reduction of the wild type transcript level was more moderate in healthy heterozygous carrier dogs. None of the examined wild-type dogs expressed the mutant transcript. Overall, these results demonstrate the leaky nature of the splice site mutation.
To investigate the segregation and frequency of the variant across the CMO affected breeds, we performed a large variant screening by genotyping the c.1332C>T variant altogether in 1052 dogs, including 695 WHWT, 249 Scottish Terriers and 108 Cairn Terriers (S3 Table). We found 123 homozygous dogs in the WHWT breed, of which 48% had been reported with CMO. About 40% of WHWT (275 dogs) carried the pathogenic variant, of which 10 dogs (3,6%) were reported with CMO. In Scottish Terriers, 10 dogs (4%) were homozygous and all were reported with CMO, and 43 dogs (17%) were carriers, of which 3 dogs (7%) were reported with CMO. In Cairn Terriers, 9 dogs (8%) were homozygous and reported with CMO, and 15 dogs (14%) carried the pathogenic variant, from which 3 dogs (20%) were reported with CMO. We found one wild-type dog in both Scottish and Cairn Terriers with CMO, and screened the coding regions of the entire SLC37A2 gene in these two dogs for possible other pathogenic variants, but did not find any. This suggests phenocopies, misdiagnoses or genetic heterogeneity. The analysis of the pathogenic variant in the three main breeds with 96 cases resulted in a highly significant association (p = 6.62x10-303) with CMO.
In addition to the above three breeds, we screened the c.1332C>T variant in 458 dogs in 124 breeds, but found only a single heterozygous carrier dog in Jack Russell Terrier breed (S3 Table). The phenotype information for this dog was not available. The variant was also screened in the known CMO cases from seven breeds (two Bull Terriers, one Curly Coated Retriever, two Border Collies, one Australian Terrier, one Basset, one German Wirehaired Pointer, one Old English Sheepdog), but they did not have it.
Collectively, our results suggest that CMO is inherited as dominant disease with incomplete penetrance. Canine CMO is equivalent to Caffey disease and our data reveals a novel candidate gene, SLC37A2, for the syndrome.
Wire Fox Terrier breeders contacted us for help in the characterization of an unknown congenital syndrome with severe mandibular prognathia and other skeletal features, mainly severe patellar luxation, in the breed. Two affected 7-week-old puppies from different litters, two unaffected littermates and two affected adult Wire Fox Terriers were examined by radiography. Additionally, a computed tomography (CT) study of the skull was made on two affected dogs (one adult and one puppy), and three dogs were studied for general clinical characteristics and neurological examination.
A prominent underbite with short maxilla (brachygnathia superior) was evident in all affected Wire Fox Terriers except one adult dog (Fig 2A and 2B). The caudodorsal border of the maxilla was slightly convex in all affected animals. In CT images, the nasal septum deviated prominently to the left at the level of the dorsocaudal frontal bone in both examined dogs (Fig 2C). The number and position of the vertebrae were normal, but the mid-thoracic spinous processes were thinner, longer and more horizontally aligned in the affected than in the normal dogs. One adult dog had an abnormally wide second rib. The adult dog had marked spondylosis of the spine. One affected puppy had unilateral congenital elbow luxation (Fig 2D), and in the other the secondary ossification centers of the olecranon were non-mineralized. The secondary ossification centers of the tibial tuberosities were small in both affected puppies, when compared to a healthy puppy (Fig 2E and 2F). The proximal epiphyses of the fibulae were not mineralized in the affected puppies and the patellae were medially luxated. The femurs of the affected dogs had medial bowing of mid-shafts of the bone. In an eye examination of a puppy and two affected adult dogs, the eyes appeared small and the sclera thinner than normal.
Clinical examination of three affected dogs indicated swollen knee joints and patellar luxation. Neurologically, all the examined dogs were alert and exhibited no remarkable neurological deficits. Postmortem examination of two puppies (one newborn and one 7 weeks old) did not reveal any additional gross abnormalities.
To identify the genetic cause of the syndrome, we performed GWAS with Illumina’s CanineHD array in a cohort of 12 Wire Fox Terriers including 4 cases and 8 controls. A case-control association test revealed association on chromosome 26 with seven nearby SNPs at 29,607,333 to 31,863,083 Mb (praw = 7.74x10-6, pgenome = 0.05) (genomic inflation factor λ = 1.10) (Fig 3A). Manual assessment of genotypes at the CFA5 locus revealed a shared homozygosity segment of ~3 Mb in the affected dogs spanning from 29,176,909 to 32,226,403 bp (Fig 3B). The associated region was captured and resequenced from five samples including two affected, two healthy and one obligate carrier Wire Fox Terrier (S4 Table). We identified a 2-bp homozygous deletion in exon 6 of the SCARF2 gene in the affected dogs after filtering the data according to an autosomal recessive model and against additional 169 unaffected control dogs from different breeds (S2 Table). The identified SCARF2 c.865_866delTC variant results in a frameshift and a premature stop codon, (p.S289Gfs*15), leading to a truncated protein in the first half of the coding region (Fig 3C).
Screening of the variant within the Wire Fox Terrier breed (57 dogs) confirmed full segregation with the disease (S5 Table). The four cases from the GWA study and an additional case were homozygous for the variant, while obligate carriers were heterozygous. The larger screening for the mutation revealed one homozygous dog that was a littermate of one of the genotyped affected dogs. This dog was then clinically examined, including neurological examination, radiography and CT scanning. The clinical findings confirmed the disease, including mandibular prognathia and patellar luxation. The carrier frequency among the population controls (n = 45) was 22%.
The effect of the 2-bp deletion on the stability of the SCARF2 mRNA was investigated by RT-PCR in postmortem skin samples from one affected and one unaffected dog (Fig 3D). The SCARF2 mRNA was detected both in affected and unaffected samples suggesting that the mutated transcript is not affected by nonsense-mediated RNA decay in the studied tissue.
SCARF2 defects have been reported in the rare human bone disease van den Ende-Gupta syndrome. Our results thus established an orthologous canine model with clinical similarity.
We were approached by a Border Collie breeder with a family of several affected dogs that suffered from severe tooth wear resulting in pulpitis and requiring extraction of those teeth. Further inspection of the tooth problem in the breed identified additional related cases, suggesting an autosomal recessive mode of inheritance (S1 Fig). Two affected dogs were subjected to a clinical study, including dental examination and radiography, as well as to histology of the extracted teeth and were regularly followed up in the next years. In addition, dental radiographs were available from two other cases.
Dental examination of a neutered 9-year-old female Border Collie revealed that all remaining teeth had significant wear. The previous dental treatment was performed 2 years earlier and multiple teeth were extracted. The length of the crowns was reduced. Lower incisor teeth were worn close to gingival margin. The enamel appeared dull and had light brown discoloration. The worn occlusal surfaces were discolored dark brown and there was reparative dentin formation. There were five teeth that had pulp exposure and pulpitis as a result of the wear (Fig 4A). The dog’s occlusion was normal and, therefore, the dental wear was not caused by abnormal tooth-to-tooth contact (attrition). Calcitriol (1,25(OH)2D3), phosphate and alkaline phosphatase levels in blood were normal. The other dental examination was performed for a neutered 10-year-old male Border Collie, revealing similar findings (Fig 4B and 4C). Other external causes such as abrasive hard chews were excluded as a cause of dental wear in both affected dogs.
Extracted teeth from two affected dogs, a 9-year-old female Border Collie described above and its female littermate, were submitted to histopathological examinations. The analysis of ground sections did not reveal structural aberrations but the enamel of the incisor was smooth and slightly hypoplastic as compared to the unaffected control dog (Fig 4D and 4E). The enamel of the premolar had largely worn and cracked, but the cervical enamel, which was preserved, showed no structural defects. Coronal dentin of both teeth comprised three distinct, circumferential zones. The tubular pattern and the structure of the matrix in the thick peripheral zone subjacent to the enamel were regular. The middlemost zone was pronouncedly globular (Fig 4E and 4F). The neighboring globules largely adapted to each other and the tubules ran uninterrupted. Wider interglobular spaces were filled with air. The globules diminished and disappeared in the central direction. The central dentin zone next to the reduced pulp showed no matrix defects, but the tubular pattern was slightly irregular (Fig 4F). The proportion of globular dentin gradually reduced in the apical direction.
The analysis of paraffin sections demonstrated significant wear of the dentin. The structure of the peripheral zone was regular, whereas the middlemost zone was globular. Between the globules there were wide, contiguous defects with an angular contour, void of matrix and tubules and filled with amorphous, barely detectable material (Fig 4G). A pulp chamber did not become visible at the level of the sections. Tubules in the central dentin zone were slightly irregular. Pulp tissue in the root canal was necrotic. A patchy, chronic inflammatory cell infiltrate with plasma cells predominating was present apically. The structure of the acellular cementum was regular. Lacunae in the cellular cementum and at the periphery of the alveolar bone trabeculae were in places obliterated. Demarcated areas in the periodontal ligament facing the alveolar bone showed an amorphous, fibrotic texture. Overall, clinical and histopathological analyses indicate severe hypomineralization of teeth in the affected dogs.
To identify the cause of the mineralization defect we sequenced the whole genomes of three affected dogs. The variants of the affected dogs were filtered against two unaffected obligate carriers and fourteen other unaffected Border Collie genomes assuming recessive transmission, resulting in the identification of a case-specific non-synonymous homozygous variant, c.899C>T, in the FAM20C gene. This leads to a missense change, p.A300V, in a highly conserved position in the kinase domain of the FAM20C protein (Fig 5). Bioinformatic predictions by SIFT (with a score of 0.00) and Polyphen2 (with a HumVar score of 0.992) suggested pathogenicity. Genotyping the pedigree and additional Border Collies (191 dogs) demonstrated complete segregation of the variant with the disease phenotype (S1 Fig and S6 Table) and showed 11% carrier frequency in the breed. We also genotyped 186 dogs from 20 additional breeds (S6 Table), but did not find any carriers in the other breeds suggesting that this pathogenic variant is specific for Border Collies.
Defects in FAM20C have been associated with a rare human disease, Raine syndrome/FGF23-related hypophosphatemia characterized by dental and bone hypomineralization. Our results indicated a causative variant in the kinase domain of FAM20C and established a canine model for human Raine syndrome.
This study demonstrates the power of NGS approaches to establish molecular diagnosis and categorization for unknown congenital canine disorders with high relevance to rare human diseases. We identified a novel candidate gene, SLC37A2, for the corresponding human disease, infantile cortical hyperostosis, also known as Caffey disease, and implicated SCARF2 and FAM20C variants in the canine forms of van den Ende-Gupta and Raine syndromes, respectively. All three spontaneous canine models closely resemble the human syndromes and provide physiologically relevant models to better understand poorly characterized gene functions in each condition and the entire molecular pathologies. Given the growing interest in the development of new therapies for rare human diseases, our study highlights the fact that dogs carry clinically and genetically close counterparts of rare human diseases that could be better utilized to advance molecular research and to develop efficient preclinical trials in large animals with spontaneous diseases [13, 14]. Many human rare disease models exist in dogs, and encouraging private pet owners to participate in clinical trials could facilitate the validation of human treatment approaches, while also benefitting canine health.
Our study unraveled a physiological function of SLC37A2 and provided new insights into infantile swelling diseases, which may be related to disturbances in the intracellular glucose homeostasis during bone development. We identified an unusual leaky splicing defect in the SLC37A2 gene in CMO-affected dogs. CMO is a self-limiting hyperostosis in multiple bones in young dogs. The most prominent sign in the affected puppies included painful swelling of the jaw, leading to dysphagia and difficulty in opening the mouth. CMO is clinically equivalent to human infantile cortical hyperostosis [10, 15]. A common missense variant in COL1A1 has been found in several patients with an autosomal dominant condition with incomplete penetrance [16, 17]. Interestingly, the mechanism how the defected collagen leads to self-limiting hyperostotic bone lesions is still unknown. Early molecular diagnosis of Caffey patients would avoid invasive procedures, however, the molecular etiology remains unknown in many cases.
SLC37A2 represents a new functional candidate gene. It belongs to the SLC37 family of four ER-associated glucose-phosphate transporters [18]. SLC37A2 is ubiquitously expressed, but transcript and protein levels are particularly high in bone-related tissues such as bone marrow and hematopoietic cell linages such as osteoclasts and macrophages [19, 20]. Murine Slc37a2 was shown to be one of the genes strongly involved in the osteoclast differentiation, suggesting that it plays a role in osteoclast function and differentiation [20]. Therefore, SLC37A2 may play a central role in glucose homeostasis in the key cell types that participate in osteogenesis. For example, an impaired function of SLC37A2 due to a truncating splice variant might disturb proper glucose supply in the osteoclasts, decreasing their overall activity, which in turn would result in an imbalance between osteoblastic and osteoclastic functions in the developing bones eventually leading to hyperostosis. Defects in SLC37A4, glucose-6-phosphate transporter (G6PT), have been associated with glycogen storage disease 1b and 1c, characterized by recurrent infections and neutropenia due to disturbed blood glucose metabolism [21–23]. Our results link SLC37A2 to bone physiology and disease, and we propose SLC37A2 as an excellent candidate for genetic screening in Caffey patients. Meanwhile, the affected dogs provide unique resources for future experiments to address SLC37A2-related mechanisms in osteogenesis biology. A recent study in hematopoietic cells identified SLC37A2 as a primary vitamin D target with a conserved vitamin D receptor-binding site [24]. This may open investigations to study the opportunity to use vitamin D as a therapeutic booster to regulate diminished expression of wild type expression of SLC37A2 in the affected dogs to alleviate clinical signs.
Caffey disease is an autosomal dominant disease with incomplete penetrance, although rare cases of recessively inherited Caffey disease have also been reported [25]. Corresponding canine diseases exist in several terrier breeds with the highest frequency in West Highland White Terriers. The determination of the exact mode of inheritance in dogs is not straightforward due to the nature of the leaky splice variant and mild self-limiting phenotype that may remain unobserved and prevent retrospective diagnosis. We found some dogs that were homozygous for the variant but had no reported clinical signs. However, we observed a considerable level of the wild-type SLC37A2 transcript in homozygous dogs in the peripheral blood due to the splicing leakage, suggesting that the leaky expression is sufficient to avoid a clinical phenotype in some cases. We also found several heterozygous dogs that had developed CMO. We found that heterozygous dogs had lower levels of wild-type SLC37A2 transcript compared to the unaffected dogs with individual variation of expression between dogs. This result suggested a dominant disease with incomplete penetrance that could help to explain the reported differences in the severity and duration of CMO among the affected dogs, although alternative models of inheritance cannot be completely ruled out yet. The dominant phenotype could be due to a dominant-negative effect, but this hypothesis requires further experimental validation to better understand the details of the gene, its regulation and protein function, including potential pairing with other proteins as described for SLC37A4/G6Pase complexes [18]. The in vivo function of SLC37A4 has been shown to depend upon its ability to couple functionally with either G6Pase-a or G6Pase-b [18, 26].
We identified a 2-bp deletion in SCARF2 in dogs with severe mandibular prognathia and other skeletal abnormalities and established a canine model for van den Ende-Gupta syndrome (VDEGS). VDEGS is a very rare disease with less than 30 reported cases [27]. It is characterized by a heterogeneous variety of craniofacial and skeletal abnormalities including blepharophimosis, a flat and wide nasal bridge, narrow and beaked nose, hypoplastic maxilla with or without cleft palate and everted lower lip, prominent deformed ears, down-slanting eyes, arachnodactyly, and camptodactyly. Patients may present congenital joint contractures that improve without intervention, and have normal growth and development. Enlarged cerebellum is an infrequent finding yet intelligence is normal. Some patients experience respiratory problems due to laryngeal abnormalities. Human and canine VDEGS patients share many similarities including hypoplastic maxilla, dislocated radial head, patellar dislocation, and deviated nasal septum. Both have small eyes.
It remains unknown how the loss of function of SCARF2 leads to VDEGS. SCARF2 is a poorly characterized member of the scavenger receptor type F family [28]. Besides epidermis, Scarf2 is expressed in branchial arches, mandible, maxilla and urogenital ridge tissue of developing mouse embryos [29, 30]. It is a single-pass transmembrane protein with homology to calmodulin (CaM)-like Ca2+-binding protein genes. The extracellular domain contains several putative epidermal growth factor-like (EGF) domains, and it has a number of positively charged residues within the intracellular domain, suggesting a role in intracellular signaling. The 2-bp deletion of the canine SCARF2 gene in one of the extracellular EGF domains leads to a severely truncated protein that completely lacks the transmembrane and intracellular domains. The lack of a transgenic mouse model and scarcity of human patients highlight the role of affected dogs as a novel resource to understand SCARF2 functions and molecular pathology. As some of the affected dogs survive past 10 years, they could potentially serve also as preclinical models.
Whole genome sequencing of a family of several affected dogs that suffered from a severe dental wear and loss of teeth revealed a recessive missense variant in the kinase domain of the FAM20C gene. FAM20C defects cause autosomal recessive osteosclerotic bone dysplasia (Raine syndrome) in humans. This rare syndrome with less than 40 reported cases was originally described to be neonatal lethal, but recently there have been several reports of cases surviving into childhood with variable severity and clinical heterogeneity [31–38]. Typical characteristics in Raine syndrome include craniofacial anomalies, such as exophthalmos, abnormal and hypomineralized teeth, midface hypoplasia, microcephaly and cleft palate, as well as gingival hyperplasia, generalized osteosclerosis and intracerebral calcifications. Variable extent of hypophosphatemia has been observed, sometimes as the primary diagnosis [32, 35, 38]. Canine findings in our cohorts were more limited to severe hypomineralization of teeth, leading to extensive wear and inflammation as prominent features. We did not observe some of the typical gross changes described in Raine patients such as hypophosphatemia and craniofacial anomalies. However, there is a significant clinical heterogeneity in the symptoms between human patients and more detailed radiographic analyses should be performed in dogs to observe potential mild changes outside the dental phenotype.
FAM20C is a Golgi casein kinase that phosphorylates secretory proteins such as FGF23 and SIBLING (Small Integrin-Binding Ligand, N-linked Glycoprotein) family [39, 40]. Fam20c is significantly expressed in mouse teeth and bone and transgenic mice studies have indicated a role in differentiation and mineralization of odontoblasts, ameloblasts, osteoblasts and osteocytes during tooth and bone development. FAM20C-deficient mice also have a prominent dental phenotype [41–43]. Ablation of the Fam20c gene in conditional knockout mice affects tooth and bone development by downregulation of SIBLING family of proteins such as DMP1 and DSPP and by increasing FGF23 in serum and promoting phosphate excretion and hypophosphatemia [44]. Our FAM20C-deficient dogs have a dental phenotype similar to mice and humans and provide a new research and preclinical model for this rare human bone disease. Unlike rodents, dogs have dental physiology similar to human, having both deciduous and permanent dentition.
In summary, we describe here the clinical and genetic characteristics of three new canine models for rare human bone disorders. While highlighting clinical and genetic similarities between canine and human conditions, our study have several implications; it indicates new physiological functions for the identified genes and provides new candidate genes to rare human diseases, establishes potential preclinical models, and finally enables the development of genetic tests for veterinary diagnostics and breeding purposes.
Sample collection in Finland was ethically approved by the Animal Ethics Committee of State Provincial Office of Southern Finland (Finland, ESAVI/6054/04.10.03/ 2012). The collection of blood samples in Switzerland was approved by the “Cantonal Committee For Animal Experiments” (Canton of Bern; permit 23/10).
EDTA-blood and tissue samples were collected from privately owned dogs in Finland, US and Switzerland. The samples were stored at -20°C until genomic DNA was extracted using the semi-automated Chemagen extraction robot (PerkinElmer Chemagen Technologie GmbH). DNA concentration was determined either with the NanoDrop ND-1000 UV/Vis Spectrophotometer or Qubit 3.0 Fluorometer (Thermo Fisher Scientific Inc.). Pedigrees were drawn by the GenoPro genealogy software (http://www.genopro.com/), and utilizing the public dog registry by the Finnish Kennel Club (http://jalostus.kennelliitto.fi).
Clinical examinations for each condition are described in detail in the results section. CMO cases were diagnosed by radiography by local veterinarians. The developmental disorder in Wire Fox Terriers was investigated by radiography and CT. Neurological examination was performed for three and ophthalmoscopy for two of the affected dogs. A specialized dental veterinarian examined two of the affected Border Collies with dental hypomineralization, while the others were examined by local veterinarians. Clinical phenotype information was not available for dogs used as population controls.
Genome-wide association studies were performed in the CMO and VDEGS projects. For CMO, altogether 51 dogs including 10 affected and 41 control dogs, were genotyped using Illumina’s CanineSNP20 BeadChip of 22,362 validated SNPs. For VDEGS, a total of 15 dogs including 4 affected and 11 unaffected Wire Fox Terriers, were genotyped using Illumina’s HD array. The genotype data in both projects was filtered with a SNP call rate of >95%, array call rate of >95% and minor allele frequency of >5%. No individual dogs were removed for low genotyping and no SNPs were removed because of significant deviations from the Hardy-Weinberg equilibrium (p ≤ 0.0001). After frequency and genotyping pruning, 14,835 and 69,694 SNPs remained for analyses for CMO and VDEGS data, respectively. Basic case-control association test was performed by PLINK [45]. Genome-wide significance was ascertained with phenotype permutation testing (n = 10,000 for CMO and n = 100,000 for VDEGS). Fine mapping of the identified CMO locus was performed with 105 selected SNPs from a 1.9-Mb region (7,764,955–9,707,794 bp) on CFA5 (based on CanFam3.1). The SNPs were selected using Broad Institute SNP collection CanFam2.0. Genotyping was performed using the Sequenom (San Diego, CA, USA) iPLEX methodology at our local core facility in the FIMM Technology Centre, University of Helsinki, Finland. A total of 88 samples were genotyped including 8 cases and 8 controls in Cairn Terriers, 9 cases and 8 controls in Scottish Terriers and 29 cases and 26 controls in WHWTs. Association analysis was performed with PLINK using a single-marker association analysis.
We performed a targeted sequence capture and next generation sequencing to identify the pathogenic variants. We used NimbleGen’s in-solution capture technology to enrich the target regions for sequencing (Roche NimbleGen, Madison, WI, USA). We captured 1.8-Mb region of CMO associated locus at position CFA5: 10,750,000–12,550,000 using two WHWT cases and controls with opposite haplotypes. The haplotypes were assessed manually using SNP genotype data. The same targeting experiment also contained samples from our other targeting projects including 8 Border Terriers, 12 Duck Tolling Retrievers, 8 Schipperkes and 4 Brazilian Terriers and these samples were used as additional controls. For target enrichment and sequencing of associated locus in Wire Fox Terriers, we captured a 3.3-Mb region at CFA26: 29,030,700–32,328,700 using two affected and two controls with opposite haplotypes and one obligate carrier. The filtered case-specific variants were further checked from 169 additional dogs in our variant database. Probes in the target regions were designed by Roche NimbleGen (Roche NimbleGen). Target enrichment, alignment and variant calling pipeline were performed as previously described [8]. Further data analysis was performed using open source R language and environment (http://www.r-project.org). Canine genome build CanFam3.1 was used as a reference sequence.
The genetic causes of CMO in WHWTs and tooth attrition in Border Collies were studied by whole genome sequencing. In the CMO study, we performed whole genome sequencing of one affected WHWT dog and used 188 other available whole genomes as controls (S2 Table). A fragment library was prepared with a 290 bp insert size and collected a single lane of Illumina HiSeq2000 paired-end reads (2 x 100 bp). The reads were mapped to the dog reference genome using the Burrows-Wheeler Aligner (BWA) version 0.5.9-r16 with default settings. The Picard tools (http://sourceforge.net/projects/picard/) were used to sort the mapped reads by the sequence coordinates and to label the PCR duplicates. The Genome Analysis Tool Kit (GATK version v2.3–6) was used to perform local realignment and to produce a cleaned BAM file. Variant calls were then made using the unified genotyper module of GATK. Variant data was obtained in variant call format (version 4.0) as raw calls for all samples and sites flagged using the variant filtration module of GATK. Variant calls that failed to pass the following filters were labeled accordingly in the call set: (i) Hard to Validate MQ0 ≥ 4 & ((MQ0 / (1.0 * DP)) > 0.1); (ii) strand bias (low Quality scores) QUAL < 30.0 || (Quality by depth) QD < 5.0 || (homopolymer runs) HRun > 5 || (strand bias) SB > 0.00; (iii) SNP cluster window size 10. The SnpEff software together with the CanFam 3.1 annotation was used to predict the functional effects of detected variants. In addition to the SNP and short indel variant calling, large deletions contained in the candidate region were searched by visual inspection of the BAM file using the Integrative Genomics Viewer (IGV). In the Border Collie study, we whole genome sequenced altogether nineteen dogs, including three affected dogs, two carriers and fourteen unaffected Border Collies, in the Science for Life Laboratory in Stockholm, Sweden. The reads were processed using speedseq align module available in SpeedSeq suite to produce a duplicate-marked, sorted and indexed BAM file. The Genome Analysis Tool kit (version = 3.3.0-g37228af) was used to perform realignment around potential indel sites and base quality score recalibration using the known SNP variation available at the Broad Institute (https://www.broadinstitute.org/ftp/pub/vgb/dog/trackHub/canFam3/variation). Dual algorithms, Samtools mpileup (version samtools-1.2) and GATK haplotype caller module were used to detect variants and the variants from both algorithms were merged into variant call format (VCFv4.1). Annovar and SnpEff tools were used to annotate the variants to Ensembl, NCBI and Broad annotation databases to predict the functional effects of the variants. We identified on average ~6 million variants per sample and the sequencing coverage varied between 22-49x. The three affected dogs shared ~1.5 million homozygous variants. Filtering under recessive model against two carriers and fourteen controls left 2690 homozygous variants in total, of which five variants were in the predicted coding regions (1 indel, 2 non-synonymous, 1 synonymous).
Large numbers of dogs were genotyped for the identified variants using various protocols. Genotyping of individual dogs for the CMO variant was performed either by TaqMan assay (Applied Biosystems) or by sequencing a 786-bp PCR product using a forward primer (5-GGCTCCAGTCTAAGCCAGGT-3) and a reverse primer (5-AAGGAGTGCGCTCAAGACAG-3) flanking the SLC37A2 SNP. The PCR products were amplified with AmpliTaqGold360Mastermix (Life Technologies), and the products were directly sequenced using the PCR primers on an ABI 3730 capillary sequencer (Life Technologies) after treatment with exonuclease I (New England Biolabs) and rapid alkaline phosphatase (Roche). The sequence data were analyzed using Sequencher 5.1 (GeneCodes). Potential exonic splice enhancer (ESE) motifs were detected with ESEfinder 3.0 [46, 47].
Pathogenicity of the FAM20C c.899C>T variant was evaluated using web-based bioinformatic prediction tools SIFT [48] and PolyPhen-2 (genetics.bwh.harvard.edu/pph2) [49] was applied to evaluate the pathogenic effect of the mutation. SIFT score ranges from 0 to 1. The amino acid substitution is predicted to be damaging if the score is smaller than 0.05. PolyPhen-2 score ranges from 0 to 1. The amino acid substitution is predicted to be damaging if the score is bigger than 0.85.
Genotyping of FAM20C and SCARF2 variants was performed by standard PCR with the following primers: FAM20C: 5-GCTTCTATGGCGAGTGTTCC-3 and 5-CCGGGATGTCTGAGTAAGGA-3; SCARF2:5-CAATCCCCGAGTGCTCTCC-3 and 5-AGGAAACTGCCCCCAAAGAG-3.
Expression analyses were performed for the CMO and VDEGS projects. Blood samples were collected into PAXgene tubes (PreAnalytix) and RNA was isolated using PAXgene Blood RNA Kit (Qiagen). The cDNA synthesis was performed using SuperScriptIII enzyme (Invitrogen) with an oligo d(T)24V primer according to manufacturer’s instructions. To investigate possible aberrant splicing events, we amplified cDNAs from the junction of exons 7 and 8 to the junction of exons 17 and 18 of the SLC37A2 gene (810 bp) using SequalPrep Long Polymerase (Invitrogen) with a forward primer GAATACCCAGAAGACGTGGAC and a reverse primer CCTCTGTCTCTGTTCAGGAATG in 16 WHWTs. B2M was used as a loading control. The identity of the amplicons was confirmed by Sanger sequencing. In the VDEGS project, the possible effect of the 2-bp deletion on the stability of the SCARF2 transcript was investigated by RT-PCR. Total RNAs were isolated from skin samples from one affected and unaffected dog obtained in the postmortem autopsies. The cDNA synthesis was performed as described above. Forward 5-CAACCACGTCACTGGCAAGT-3 and reverse 5-TTACAGTGGGG CCCGTGG-3 primers were designed to amplify a 188-bp region between exon 6 and exon 8 of SLC37A2. Semi-quantitative analysis of the expression was determined and visualized by electrophoresis.
Permanent and deciduous incisor and premolar teeth were removed for therapeutic reasons from two Border Collie bitches with clinically affected teeth. The dogs were from the same litter. Three teeth, two incisors and one premolar, were obtained from one dog. One incisor and the premolar were processed to ground sections and the other incisor to paraffin sections. From the other dog, four teeth, two incisors and two premolars, were obtained. One incisor and one premolar were processed to ground sections and the other incisor and the other premolar to paraffin sections. For comparison, deciduous teeth obtained from a healthy Border Collie were studied. One incisor was processed to ground sections and one tooth to paraffin sections. The teeth to be processed to ground sections (a procedure preserving the enamel with a proportionally high mineral content and sparse organic matrix) were fixed with 10% neutral buffered formalin, dehydrated and embedded in liquid methylmethacrylate monomer. After complete polymerization, started up by benzoylperoxide (2 g/l), the teeth were serially cut to 100–150 μm thick longitudinal sections with a rotating diamond-coated saw microtome, let dry and mounted unstained with DePex (Gurr, BDH, Poole, UK). For preparation to paraffin sections, the teeth were fixed with formalin, demineralized with EDTA (0.33 mol/l, pH 7.2), which leads to the loss of the enamel, and embedded in paraffin. A representative series of sections were longitudinally cut at 7 μm and stained with haematoxylin and eosin (HE).
|
10.1371/journal.pgen.1000849 | Genome-Wide Identification of Susceptibility Alleles for Viral Infections through a Population Genetics Approach | Viruses have exerted a constant and potent selective pressure on human genes throughout evolution. We utilized the marks left by selection on allele frequency to identify viral infection-associated allelic variants. Virus diversity (the number of different viruses in a geographic region) was used to measure virus-driven selective pressure. Results showed an excess of variants correlated with virus diversity in genes involved in immune response and in the biosynthesis of glycan structures functioning as viral receptors; a significantly higher than expected number of variants was also seen in genes encoding proteins that directly interact with viral components. Genome-wide analyses identified 441 variants significantly associated with virus-diversity; these are more frequently located within gene regions than expected, and they map to 139 human genes. Analysis of functional relationships among genes subjected to virus-driven selective pressure identified a complex network enriched in viral products-interacting proteins. The novel approach to the study of infectious disease epidemiology presented herein may represent an alternative to classic genome-wide association studies and provides a large set of candidate susceptibility variants for viral infections.
| Viruses have represented a constant threat to human communities throughout their history, therefore, human genes involved in anti-viral response can be thought of as targets of virus-driven selective pressure. Here we utilized the marks left by selection to identify viral infection-associated allelic variants. We analyzed more than 660,000 single nucleotide polymorphisms (SNPs) genotyped in 52 human populations, and we used virus diversity (the number of different viruses in a geographic region) to measure virus-driven selective pressure. Results showed that genes involved in immune response and in the biosynthesis of glycan structures functioning as viral receptors display more variants associated with virus diversity than expected by chance. The same holds true for genes encoding proteins that directly interact with viral components. Genome-wide analysis identified 441 variants, mapping to 139 human genes, significantly associated with virus-diversity. We analyzed the functional relationships among genes subjected to virus-driven selective pressure and identified a complex interaction network enriched in viral products-interacting proteins. Therefore, we describe a novel approach for the identification of gene variants that may be involved in the susceptibility to viral infections.
| Infectious diseases represent one of the major threats to human populations, are still the first cause of death in developing countries [1], and are therefore a powerful selective force. In particular, viruses have affected humans before they emerged as a species, as testified by the fact that roughly 8% of the human genome is represented by recognizable endogenous retroviruses [2] which represent the fossil remnants of past infections. Also, viruses have probably acted as a formidable challenge to our immune system due to their fast evolutionary rates [3]. Indeed, higher eukaryotes have evolved mechanisms to sense and oppose viral infections; the recent identification of the antiviral activity of particular proteins such as APOBEC, tetherin, and TRIM5 has shed light on some of these mechanisms. Genes involved in anti-viral response have therefore been presumably subjected to an enormous, continuous selective pressure.
Despite the relevance of viral infection for human health, only few genome-wide association studies (GWAS) have been performed in the attempt to identify variants associated with increased susceptibility to infection or faster disease progression [4]–[5]. These studies have shown the presence of a small number of variants, mostly located in the HLA region. This possibly reflects the low power of GWAS to identify variants with a small effect. An alternative approach to discover variants that modulate susceptibility to viral infection is based on the identification of SNPs subjected to virus-driven selective pressure. Indeed, even a small fitness advantage can, on an evolutionary timescale, leave a signature on the allele frequency spectrum and allow identification of candidate polymorphisms. To this aim we exploited the availability of more than 660,000 SNPs genotyped in 52 human populations distributed world-wide (HGDP-CEPH panel) [6] and of epidemiological data stored in the Gideon database.
Previous studies [7]–[9] have suggested that the number of the different pathogen species transmitted in a given geographic location is a good estimate of pathogen-driven selection for populations living in that area. Indeed, pathogen diversity is largely dependent on climatic factors [10] and might more closely reflect historical pressures than other estimates such as the prevalence of specific infections. We therefore reasoned that virus diversity can be used as a measure of the selective pressure exerted by virus-borne diseases on human populations and, as a consequence, that SNPs showing an unusually strong correlation with virus diversity can be considered genetic modulators of infection susceptibility or progression. To explore this possibility we used a large set of SNPs that have been genotyped in the HGDP-CEPH panel, a collection of DNAs from almost 950 individuals sampled throughout the world (Table 1). Virus diversity estimates were derived from the Global Infectious Disease and Epidemiology Network database: for each country where HGDP-CEPH populations are located we counted the number of different virus species (or genera/family as described in materials and methods) that are naturally transmitted (Table 1).
One simple prediction of our hypothesis whereby virus diversity is a reliable estimator of virus-driven selective pressure is that genes known to be involved in immune response are enriched in SNPs significantly associated with virus richness. In order to verify whether this is the case we analysed the InnateDB gene list which contains 2,915 genes involved in immune response and showing the presence of at least one SNP in the HGDP-CEPH panel. Correlations with virus richness were calculated using Kendall's partial rank correlation; since allele frequency spectra in human populations are known to be affected by demographic factors in addition to selective forces [11]–[12], each SNP was assigned a percentile rank in the distribution of τ values calculated for all SNPs having a minor allele frequency (MAF) similar (in the 1% range) to that of the SNP being analysed. A SNP was considered to be significantly associated with virus diversity if it displayed a significant correlation (after Bonferroni correction with α = 0.01) and a rank higher than 0.99. As shown in Table 2, 104 SNPs in InnateDB genes showed a significant association with virus diversity. All SNPs in InnateDB genes that correlated with virus diversity are listed in Table S1. By performing 10,000 re-samplings of 2,915 randomly selected human genes (see materials and methods for details) we verified that the empirical probability of obtaining 104 significantly associated SNPs amounts to 0.010, indicating that genes in the InnateDB list display more virus-associated SNPs than expected.
It is worth mentioning that amongst these genes, UNG (MIM 191525), encoding uracil DNA glycosylase, functions downstream of APOBEC3G (MIM 607113) to mediate the degradation of nascent HIV-1 DNA [13]. SERPING1 (MIM 606860), a regulator of the complement cascade, is also involved in HIV-1 infection (MIM 609423) as its expression is dysregulated in immature dendritic cells by Tat [14]; moreover, the protein product of SERPING1 is cleaved by HCV and HIV-1 proteases [15]–[16].
Genes involved in the biosynthesis of glycan structures have also been considered as possible modulators of infection susceptibility. Indeed, since Haldane's prediction in 1949 [17] that antigens constituted of protein-carbohydrates molecules modulate the resistance/susceptibility to pathogen infection, protein glycolsylation has been shown to play a pivotal role in viral recognition of host targets [18], as well as in antigen uptake and processing and in immune modulation [19]–[20]. We therefore computed a list of genes involved in glycan biosynthesis from KEGG pathways and Gene Ontology annotations. Again these genes displayed significantly more virus-associated SNPs than expected if randomness alone were responsible (empirical p = 0.0138) (Table 2 and Table S2). Several virus-associated SNPs were located in genes coding for sialyltransferases (ST6GAL1 (MIM 109675), ST3GAL3 (MIM 606494), ST6GALNAC3 (MIM 610133), ST8SIA1 (MIM 601123), ST3GAL1 (MIM 607187) and ST8SIA6 (MIM 610139)). Notably, sialic acids represent the most prevalent terminal monosaccharides on the surface of human cells and determine the host range of different viruses including influenza A [21]–[22], polyomaviruses (i.e JCV and BKV in humans) [23], and rotaviruses (the leading cause of childhood diarrhea) [24].
Sialyltransferases also play central roles in B and T cell communication and function. In particular, the generation of influenza-specific humoral responses is impaired in mice lacking ST6GAL1 [25], while ST3GAL1 regulates apoptosis of CD8+ T cells [20]. Interestingly, ST8SIA6 is expressed in NK cells, possibly playing a role in the regulation of Siglec-7 lectin inhibitory function in these cells [26]. Four other genes (XYLT1 (MIM 608124), HS3ST3A1 (MIM 604057), UST (MIM 610752) and CHSY3 (MIM 609963)) carrying SNPs associated with virus diversity are involved in the biosynthesis of either heparan sulphate or chondroitin sulphate. The former is an ubiquitously expressed glycosaminoglycan serving as the cell entry route for herpesviruses [27], HTLV-1 [28] and papillomaviruses [29]. Chondroitin sulphate is similarly expressed on a wide array of cell types and functions as an auxiliary receptor for binding of herpes simplex virus [30] as well as a facilitator of HIV-1 entry into brain microvascular endothelial cells [31]. Finally, we identified LARGE (MIM 603590) among the genes subjected to virus-driven selective pressure (Table 2). Recent studies have demonstrated that the post-translational modification of α-dystroglycan by LARGE is critical for the binding of arenaviruses of different phylogenetic origin including Lassa fever virus and lymphocytic-choriomeningitis virus [32]–[33]. Therefore our data support the previously proposed hypothesis whereby viruses represent the selective pressure underlying the strong signal of positive selection at the LARGE locus [34].
Since genes involved in immune response and in the biosynthesis of glycan structures are likely to be subjected to selective pressures exerted by pathogens other than viruses, we verified whether a set of genes directly involved in interaction with viral proteins also displays more SNPs significantly correlated with virus diversity. To this aim we retrieved a list of 1,916 genes known to interact with at least one viral product and displaying at least one genotyped SNP in the HGDP-CEPH panel (see materials and methods). In order to perform a non-redundant analysis, genes included in the InnateDB list and involved in glycan biosynthesis were removed; the remaining 987 genes displayed 80 SNPs correlated with virus diversity, corresponding to an empirical p value of 0.017 (Table 2 and Table S3). Notably, when this same analysis was performed using the diversity of pathogens other than viruses (bacteria, protozoa and helminths), no significant excess of correlated SNPs was found (all empirical p values>0.05).
Given these results, we wished to identify SNPs significantly associated with virus richness on a genome-wide base. We therefore calculated Kendall's rank correlations between allele frequency and virus diversity for all the SNPs (n = 660,832) typed in the HGDP-CEPH panel. We next searched for instances which withstood Bonferroni correction (with α = 0.05) and displayed a τ percentile rank higher than the 99th among MAF-matched SNPs. A total of 441 SNPs mapping to 139 distinct genes satisfied both requirements. Table 3 shows the 30 top SNPs (or SNP clusters) located within genic regions and associated with virus diversity, while the full list of SNPs subjected to virus-driven selective pressure is available on Table S4. It is worth noting that the SNP dataset we used contains less than 200 variants mapping to HLA genes (both class I and II), therefore covering a minor fraction of genetic variability at these loci; as a consequence HLA genes cannot be expected to be identified as targets of virus-driven selective pressure using the approach we describe herein.
We next verified whether the correlations detected between the SNPs we identified and virus diversity could be secondary to climatic variables. Hence, for all countries where HGDP-CEPH populations are located we obtained (see materials and methods) the following parameters: average annual minimum and maximum temperature, and short wave (UV) radiation flux. Results showed that none of the SNPs associated with virus diversity significantly correlated with any of these variables (Table S5).
Previous works have reported an enrichment of selection signatures within or in close proximity to human genes [12],[35]. In line with these data we verified that virus-associated SNPs are more frequently located within gene regions compared to a control set of MAF-matched variants (χ2 test, p = 0.026).
We investigated the role and functional relationship among genes subjected to virus-driven selective pressure using the Ingenuity Pathway Analysis (IPA, Ingenuity Systems) and the PANTHER classification system [36]–[37]. Unsupervised IPA analysis retrieved two networks with significant scores (p = 10−17 and p = 10−12) which were merged into a single interaction network (Figure 1). The network contains 23 genes showing a significant correlation with virus diversity and, among these, 10 encode proteins interacting with viral products (Figure 1). Based on the number of observed human-virus interactions, this finding is unlikely to occur by chance (χ2 test, p = 0.0013) as 2.88 human-virus interactions would be expected for 23 genes. Analysis of the whole network indicated that a 31 of 66 genes encode proteins interacting with viral products (Figure 1): again this number is higher than expected (expected interactions = 8.27; χ2 test, p = 2.8×10−10). Thus, the interaction network we have identified is enriched in genes subjected to virus-driven selective pressure and in genes coding for proteins interacting with viral products. It is worth mentioning that, in agreement with previous findings [38], many viral-interacting proteins represent hubs in the network. Conversely, most of the genes we found to be subjected to virus-driven selective pressure, irrespective of their ability to interact with viral proteins, tend to display very low connectivity (low-degree nodes). This observation might be consistent with previous indications [39]–[41] that in eukaryotes hub genes are more selectively constrained compared to low-degree nodes, these latter being more likely to evolve in response to environmental pressures.
In addition to proteins directly interacting with viral products, several network genes showing correlation with virus diversity might play central roles during viral infection. DNMT1 (MIM 126375) and MGMT (MIM 156569) are involved in DNA methylation and repair, respectively, two processes that are often dysregulated during viral infection. In particular, altered expression of DNMT1 is induced by diverse viruses including HIV-1 [42], EBV [43], BKV and adenovirsuses [44]; also, DNMT1 plays a pivotal role in the expansion of effector CD8+ T cell following viral infection [45]. A relevant role in HIV-1 infection is also played by HSPG2 (MIM 142461), the gene coding for perlecan, a cell surface heparan sulfate proteoglycan which mediates the internalization of Tat protein [46].
We next investigated the over-representation of PANTHER classification categories among genes subjected to virus-driven selective pressure. Table 4 shows the significantly over-represented PANTHER molecular functions and biological processes with the contributing genes. In line with the results we reported above, genes involved in immune response, as well as genes coding for proteins involved in cell adhesion and extracellular matrix components, resulted to be over-represented; these latter genes might mediate viral-cellular interaction and facilitate viral entry.
The identification of non-neutrally evolving loci with a role in immunity can be regarded as a strategy complementary to classic clinical and epidemiological studies in providing insight into the mechanisms of host defense [47]. Here we propose that susceptibility genes for viral infections can be identified by searching for SNPs that display a strong correlation with the diversity of virus species/genera transmitted in different geographic areas. Similar approaches have previously been applied to study the adaptation to climate for genes involved in metabolism and sodium handling [48]–[50]. These analyses, including the one we describe herein, rely on similar assumptions and imply some caveats. First, we implicitly considered virus diversity, as we measure it nowadays, a good proxy for long-term selective pressure. This clearly represents an oversimplification, as new viral pathogens have recently emerged and the virulence of different viral species or genera might have changed over time. Still, previous studies have indicated that the geographic distribution of virus diversity is strongly influenced by climatic variables such as temperature and precipitation rates [10], suggesting that, despite significant changes in prevalence and virulence, virus diversity might have remained relatively constant across different geographic areas, possibly representing the best possible estimate of long-standing pressure. In line with these considerations, we calculated virus diversity as the number of all viral species (or genera/families) that can cause a disease in humans, irrespective of virulence or pathogenicity (Table S6).
The second issue relevant to the data we present herein is that environmental variables tend to co-vary across geographic regions: the distribution of different pathogens (e.g. parasitic worms and viruses/bacteria/protozoa) is correlated across HGDP-CEPH populations [9] and, as reported above, virus diversity is influenced by climatic factors. Therefore, our genome-wide search was preceded by analyses aimed at verifying whether virus diversity is a reliable and specific estimator of virus-driven selective pressure. In particular, we verified that genes involved in immune response and in the biosynthesis of glycans display significantly more variants associated with virus diversity than randomly selected human genes; this finding supports the idea that pathogens rather than climate or demography has driven the genetic variability at these loci. Notably, we also analysed genes that encode proteins interacting with viral components: since loci involved in immune response and in glycan biosynthesis were removed from this list, the remaining genes are expected to be specific targets of viral-driven selective pressure; consistently, we verified that a significant excess of SNPs correlating with virus diversity map to these loci. Conversely, a SNP excess was not noticed when the diversity of other human pathogens was used for the analysis, suggesting that, despite the correlation among different pathogen species across geographic locations [9], the selective pressure imposed by viruses can be distinguished from that exerted by other organisms.
As a further control for the possible confounding effects of other environmental factors, we verified that the variants we identified at the genome-wide level do not correlate with climate (temperature) and UV radiation. This analysis was motivated by the known association of virus diversity and biodiversity in general, with temperature [10],[51] and by the fact that both climate and UV exposure have long been considered among the strongest selective pressures in humans [52]. Since none of the SNPs we identified correlated with either short wave radiation flux or temperature, we consider that their geographic distribution is likely to have been shaped by virus-driven selective pressure. In this respect it is worth mentioning that UV irradiation has been shown to be immunosuppressive in mice (reviewed in [53]–[54]), but the effect of sun exposure on immune functions in humans is still poorly understood. Yet, herpes viruses (both simplex and zoster) and some papillomavirus types have been shown to be reactivated by UV exposure, suggesting that the link between short wave radiation flux and virus-driven selective pressure might be more complex than simply predicted on the basis of geographic variation.
Our genome wide search for genes subjected to virus-driven selection allowed the identification of a gene interaction network that is enriched in both genes associated with virus diversity and in genes encoding proteins that interact with viral products. Many of the genes included in the identified network are of great interest as they are known to be involved in the activation of mechanisms that have direct or indirect protective effects against viruses. Thus, beside the well known activities of IL1A (MIM 147760) and B (MIM 147720), IL4 (MIM 147780), TGFB1 (MIM 190180), IL16 (MIM 603035), IFNG (MIM 147570) and TNF (MIM 191160), OAS2 (MIM 603350) encodes a protein that activates latent RNases, resulting in the degradation of viral RNA and in the inhibition of viral replication [55]. CCL17 (MIM 601520) induces T lymphocytes chemotaxis, thus potentiating the immune responses, and PPP3CA (MIM 114105), also known as calcineurin, activates NFATc [56], a key factor in the up-regulation of IL2 (MIM 147680) [57], the main cytokine responsible for T lymphocytes growth and differentiation. Finally, ULBP2 (MIM 605698) encodes an MHC1-related protein that binds to NKG2D (MIM 602893) [58], an activating receptor expressed on CD8 T cells as well as on NK cells, NKT cells and γδ T cells. In the light of the viral pathogenesis of a growing number of neoplasia, it is very interesting that other members of the network play a well described role in the inhibition of tumoral growth. In particular, E2F1 (MIM 189971) is known to have a pivotal role in the control of cell cycle and in the activation of tumour suppressor proteins and, together with TP53I3, TADA3L, and TP53BP2 mediates p53-dependent and independent apoptosis [59]–[60]. CCND3 (MIM 123834) is involved in cell cycle progression through the G2 phase, whereas RAD23A (MIM 600061) up-regulates the nucleotide excision activity of 3-methyladenine-DNA glycosylase [61], therefore playing a role in DNA damage recognition in base excision repair. Finally, NR4A2 (MIM 601828) encodes a nuclear orphan receptor expressed in T cells and involved in apoptosis [62]. NR4A2 is also known to play a central role in eliciting the production of inflammatory cytokines in multiple sclerosis (MS (MIM 126200)) [63]. Notably, variants in PPP3CA (Figure 1) have recently been reported to correlate with MS severity as well [64]. We therefore investigated whether other genes carrying SNPs which correlate with virus diversity have been identified in GWAS for MS susceptibility or severity. Three additional genes, JMJD2C (MIM 605469), C20orf133 (also known as MACROD2, (MIM 611567)) and CSMD1 (MIM 608397) have been associated with MS [64] and display SNPs significantly correlated with virus diversity (Table S1). While the function of C20orf133 is unknown, JMJD2C encodes a histone demethylase expressed at very high levels in B cells and cytotoxic lymphocytes (see materials and methods), a pattern consistent with its being subjected to virus-driven selective pressure. Finally, CSMD1, in analogy to the aforementioned SERPING1, acts as a regulator of the complement system [65]; notably, complement activation plays a central role in both response to viruses and inflammatory reactions, particularly in the central nervous system [66].
Analysis of the 30 stronger associations (Table 3) indicated that several genes are part of the network described above or have been involved in immune response (see InnateDB gene list, Table 2). Conversely, others encode relatively unknown products (e.g. KIAA1529 (MIM 611258), LHFPL3 (MIM 609719), LOC51149, RNF217, TMEM132B, LEPREL1 (MIM 610341), ANKFN1, MYO5C (MIM 610022), ANXA4 (MIM 106491) and SCRN3). Among these genes, MYO5C, ANXA4 and SCRN3 are involved in membrane trafficking events along exocytotic and endocytotic pathways, suggesting that they might play a role in either viral cell entry [67] or lytic granule exocytosis; this might be the case for ANXA4 which is expressed at high levels in NK cells (see materials and methods). Most interestingly, EYA4 (MIM 603550) (Table 3) has recently been described as a phosphatase involved in triggering innate immune responses against viruses [68]. Finally, both PDE2A (MIM 602658) and SCNN1A (MIM 600228) might play a role in maintaining lung epithelial barrier homoeostasis during viral infection. Indeed, both genes can be induced by TNF-alpha in lung epithelial cells [69]–[70] and can influence lung fluid reabsorption and, therefore, edema formation. In line with these observations, expression of the amiloride-sensitive epithelial Na+ channel (SCNN1A codes for the α subunit) is affected by infection with influenza virus, severe acute respiratory syndrome coronavirus and respiratory syncitial virus.
In humans, resistance to infectious diseases is thought to be under complex, multigenic control with single loci playing a small protective role [47]. This concept also holds for viral infection as demonstrated by the role of genetic variants in modulating the susceptibility to HIV infection or disease progression (reviewed in [71]). Classic GWAS offer a powerful resource to identify susceptibility loci for infectious diseases; yet GWAS typically have limited power to detect variants with a low frequency or a small effect. Indeed, recent GWAS for SNPs determining the host control of HIV-1 [4]–[5] failed to identify most known loci with a role in AIDS progression. The alternative approach we have proposed here is based on the identification of variants subjected to virus-driven selective pressure. Similarly to the GWAS results mentioned above we did not identify well known antiviral-response genes. Still, we noticed that variants in TRIM5 (MIM 608487) (rs2291845, τ = 0.44, p = 1.86×10−5, rank = 0.97) and IFIH1 (MIM 606951) (also known as MDA5, rs10439256, τ = 0.51, p = 5.4×10−7, rank = 0.99) showed significant associations with virus-diversity, although they did not withstood genome-wide analysis. Also, it is worth mentioning that variants with a well established role in resistance to viral infections may be neutrally evolving; this is the case for the Δ32 allele of CCR5 (MIM 601373) for example, which confers protection against HIV-1 infection and possibly against other pathogens, but displays no selection signature [72]. This is possibly due to how long and how strong the selective pressure has been exerted. Conversely, variants subjected to selective pressure must have (or have had along human history) some selective advantage, indicating that the SNPs we have identified can be regarded as candidate modulators of infection susceptibility or disease progression.
Virus absence/presence matrices for the 21 countries where HGDP-CEPH populations are located were derived from the Global Infectious Disease and Epidemiology Network database (Gideon, http://www.gideononline.com), a global infectious disease knowledge tool. Information in Gideon is weekly updated and derives from World Health Organization reports, National Health Ministries, PubMed searches and epidemiology meetings. The Gideon Epidemiology module follows the status of known infectious diseases globally, as well as in individual countries, with specific notes indicating the disease's history, incidence and distribution per country. We manually curated virus absence/presence matrices by extracting information from single Gideon entries. These may refer to either species, genera or families (in case data are not available for different species of a same genus/family). Following previous suggestions [7]–[9], we recorded only viruses that are transmitted in the 21 countries, meaning that cases of transmission due to tourism and immigration were not taken into account; also, species that have recently been eradicated as a result, for example, of vaccination campaigns, were recorded as present in the matrix. A total of 81 virus species/genera/families were retrieved (Table S6). The same approach was applied to calculate the diversity of other pathogens, namely bacteria, protozoa and helminths [9]. The annual minimum and maximum temperature were retrieved from the NCEP/NCAR database (http://www.ngdc.noaa.gov/ecosys/cdroms/ged_iia/datasets/a04/, Legates and Willmott Average, re-gridded dataset) using the geographic coordinates reported by HGDP-CEPH website for each population (http://www.cephb.fr/en/hgdp/table.php). Similarly, net short wave radiation flux data were obtained from NCEP/NCAR (http://www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis.surfaceflux.html, Reanalysis 1: Surface Flux); these data were read using Grid Analysis and Display System (GrADS, http://www.iges.org/grads/). Daily values for four years (1948–1951) were averaged to obtain an annual mean.
Since virus diversity, due to data organization in Gideon, can only be calculated per country (rather than per population), the same procedure was applied to climatic variables. Therefore the values of annual temperature and radiation flux were averaged for populations located in the same country. This assures that a similar number of ties is maintained in all correlation analyses.
Data concerning the HGDP-CEPH panel derive from a previous work [6]. Atypical or duplicated samples and pairs of close relatives were removed [73].
A SNP was ascribed to a specific gene if it was located within the transcribed region or no farther than 500 bp upstream the transcription start site. MAF for any single SNP was calculated as the average over all populations. The list of immune response genes was derived from the InnateDB website (http://www.innatedb.com/) and it contains a non-redundant list of 5,070 immune genes derived from ImmPort, IRIS, Septic Shock Group, MAPK/NFKB Network and Immunome Database; it only includes genes derived from curated immune gene lists.
Genes involved in glycan biosynthesis were obtained by merging genes from two KEGG pathways (“Glycan structures - biosynthesis 1” and “Glycan structures - biosynthesis 2”). Additional genes were identified by searching Gene Ontology categories for genes that act as glycosyltransferases (GO:0016757) and are located in either the Golgi or the endoplasmic reticulum (GO:0005783, GO:0005793 and GO:0005794). The list of human genes coding for proteins interacting with viral products was derived from three sources: a previously published study [38], the VirHostNet website [74] (http://pbildb1.univ-lyon1.fr/virhostnet/) and the HIV-1 Human Protein Interaction Database [75] (http://www.ncbi.nlm.nih.gov/RefSeq/HIVInteractions/).
Expression data were obtained from SymAtlas (http://symatlas.gnf.org/). The location of genomic elements that are highly conserved among vertebrates was derived from UCSC annotation tables (http://genome.ucsc.edu/; “PhastCons Conserved Elements, 44-way Vertebrate Multiz Alignment” track).
All correlations were calculated by Kendall's rank correlation coefficient (τ), a non-parametric statistic used to measure the degree of correspondence between two rankings. The reason for using this test is that even in the presence of ties, the sampling distribution of τ satisfactorily converges to a normal distribution for values of n larger than 10 [76].
In order to estimate the probability of obtaining n SNPs located within m genes and significantly associated with virus diversity, we applied a re-sampling approach: samples of m genes were randomly extracted from a list of all genes covered by at least one SNP in the HGDP-CEPH panel (number of genes = 15,280) and for each sample the number of SNPs significantly associated with virus diversity was counted. The empirical probability of obtaining n SNPs was then calculated from the distribution of counts deriving from 10,000 random samples. A SNP was ascribed to a gene if it was located within the transcribed region or in the 500 upstream nucleotides.
Analysis of PANTHER over-represented functional categories and pathways was performed using the “Compare Classifications of Lists” tool available at the PANTHER classification system website [77] (http://www.pantherdb.org/). Briefly, gene lists are compared to the reference list using the binomial test for each molecular function, biological process, or pathway term in PANTHER.
All calculation were performed in the R environment [78] (http://www.r-project.org/).
Biological network analysis was performed with Ingenuity Pathways Analysis (IPA) software using an unsupervised analysis (www.ingenuity.com). IPA builds networks by querying the Ingenuity Pathways Knowledge Base for interactions between the identified genes and all other gene objects stored in the knowledge base; it then generates networks with a maximum network size of 35 genes/proteins. We used all genes showing at least one significantly associated SNP as the input set; in this case a SNP was ascribed to a gene if it was located within the transcribed region or in the 25 kb upstream. All network edges are supported by at least one published reference or from canonical information stored in the Ingenuity Pathways Knowledge Base. To determine the probability of the analysed genes to be found together in a network from Ingenuity Pathways Knowledge Base due to random chance alone, IPA applies a Fisher's exact test. The network score represents the -log (p value).
|
10.1371/journal.pgen.1004078 | Single Nucleus Genome Sequencing Reveals High Similarity among Nuclei of an Endomycorrhizal Fungus | Nuclei of arbuscular endomycorrhizal fungi have been described as highly diverse due to their asexual nature and absence of a single cell stage with only one nucleus. This has raised fundamental questions concerning speciation, selection and transmission of the genetic make-up to next generations. Although this concept has become textbook knowledge, it is only based on studying a few loci, including 45S rDNA. To provide a more comprehensive insight into the genetic makeup of arbuscular endomycorrhizal fungi, we applied de novo genome sequencing of individual nuclei of Rhizophagus irregularis. This revealed a surprisingly low level of polymorphism between nuclei. In contrast, within a nucleus, the 45S rDNA repeat unit turned out to be highly diverged. This finding demystifies a long-lasting hypothesis on the complex genetic makeup of arbuscular endomycorrhizal fungi. Subsequent genome assembly resulted in the first draft reference genome sequence of an arbuscular endomycorrhizal fungus. Its length is 141 Mbps, representing over 27,000 protein-coding gene models. We used the genomic sequence to reinvestigate the phylogenetic relationships of Rhizophagus irregularis with other fungal phyla. This unambiguously demonstrated that Glomeromycota are more closely related to Mucoromycotina than to its postulated sister Dikarya.
| Endomycorrhizal fungi are known for their symbiosis with the vast majority of land plants. The fungus penetrates the root and facilitates uptake of nutrients for the plant. For a long time it was hypothesized that endomycorrhizal fungi have a complex genetic makeup, as they are asexual organisms. Their hyphae do not consist of individual cells, but rather form a continuous compartment in which numerous nuclei migrate. Several studies indicated that these nuclei are genetically highly diverse, suggesting that endomycorrhizal fungi evolved a unique genome structure. By sequencing individual nuclei of a single individual of the reference fungus Rhizophagus, we demystify this hypothesis and show that the nuclei are highly similar. Furthermore, we created the first genome sequence of these ancient fungi that will serve as a valuable resource to further understand and exploit this agriculturally and ecologically vital symbiosis.
| The interaction of arbuscular endomycorrhizal (AM) fungi and land plants is a very successful symbiosis as it is ancient (∼450 million years), and maintained by the vast majority of plant species [1]. AM fungi are obligate biotrophs that infect roots and form highly branched structures (arbuscules) inside root cortical cells [1]. These arbuscules are connected to an extensive network of extraradical mycelium that facilitates uptake of nutrients from the soil, e.g. immobile phosphates.
AM hyphal networks form a continuous coenocytic compartment with numerous nuclei. AM fungi are considered to be ancient asexual organisms [2]–[4] and propagation occurs via spores that become filled with multiple nuclei that subsequently divide [5]. AM fungal individuals can be heterokaryotic, i.e. consist of genetically divergent nuclei, because single nucleus cellular stages never occur during the lifecycle, and because hyphae of different fungal individuals can fuse and exchange nuclei by anastomosis [6], [7]. Our knowledge of the genome structure of AM fungi is rudimentary. For instance, the degree to which a minimal gene set is present in a single nucleus, or is distributed over genetically distinct nuclei is unknown [2], [8]–[11]. Although there is evidence for genetic variability within single spores, the genomic organization of this variation remains elusive. Two competing hypotheses have been advocated. The genetic variation may be present in a single, possibly polyploid, nucleus [9], or it could be distributed over multiple nuclei in a single individual [8], [10]. However, in reality these hypotheses may represent extremes along a continuum of genetic variation among and within nuclei [2].
Extensive efforts to sequence the genome of the reference AM fungal species Rhizophagus irregularis DAOM197198 (previously known as Glomus intraradices [12], [13] have not been successful, possibly because of its heterokaryotic nature [14]. To address this issue and determine the extent to which nuclei are indeed markedly different, we conducted de novo genome sequencing of individual nuclei of an R. irregularis line isolated from the reference strain DAOM197198 (designated DAOM197198w). The resulting R. irregularis genome sequence revealed a surprisingly low level of polymorphism between nuclei.
Spores of a mycorrhized root culture of chicory (Cichorium intybus) were stained by 10 µM Sytox Green (Fig. 1A). Single nuclei were collected from a supernatant of crushed spores using a micromanipulator (Fig. 1B). Individual nuclei were immediately processed for whole genome amplification. To verify the quality of the amplified nuclear DNA ten randomly selected loci were PCR amplified, and also the extent of bacterial contamination was monitored. Four amplified single nucleus genomes were processed for sequencing, resulting in assembled genomes of 115, 90, 71 and 95 Mbps, respectively (Tables S1 and S2). The different sizes of the assemblies are likely reflecting variation in the whole genome amplification efficiencies among the four samples. First comparative analyses detected surprisingly few SNPs and indels across the four nuclei. This suggested that nuclei are markedly more similar than was expected. Therefore we decided to sequence also two DNA samples extracted from mycelium. The generated sequences of these DNA samples (designated DNA1 and DNA2) were assembled individually, resulting in genome assemblies of 116 and 117 Mbps, respectively (Table 1). Additionally, the six genome sequences were assembled together resulting in a reference genome for R. irregularis of 141 Mbps. A self-alignment of this reference genome revealed little redundancy ruling out the occurrence of (significant) artificial duplifications within the assembly (Fig. S1). By comparative genomic analysis, only 28,872 SNPs and 12,315 indels were detected across the six assemblies when compared to the reference genome (Fig. 1C, Table S2). Furthermore, a reference-independent comparison of the four single nuclei and the two mycelial samples also revealed a comparable low level of polymorphisms (Table S3). This indicates that more than 99.97% of the (aligned) genome sequence is identical between different nuclei. Furthermore, as the size of the assembled genome is in line with previous estimates of the DNA content of nuclei [15], we conclude that R. irregularis nuclei are haploid.
Several loci have previously been used to determine genetic polymorphisms within AM individuals. These include Binding Protein (BIP), SSR marker Bg112, the internal transcribed spacers (ITS1 and ITS2) of the 45S rDNA locus in R. irregularis and POL1-Like Sequence (PLS) in Glomus etunicatum [8], [9], [16]. We compared these loci in the different genome assemblies. Only a single PLS homolog was identified in R. irregularis (RiPLS, RirG174000), whereas G. etunicatum has multiple copies that belong to two main types, of which the highly polymorphic PLS1 likely represents a pseudogene [9], [17]. No polymorphisms were found for RiPLS in the different assemblies (Fig. S2). For BIP three loci were identified and designated RiBIP1 (RirG196040), RiBIP2 (RirG160690) and RiBIP3 (RirG043980). Sequence and structure of these genes is highly conserved and homologous to a Rhizopus delemar 70 kD Heat shock protein (GenBank: EIE83965). RiBIP1, RiBIP2 and RiBIP3 are present also in nucleus 6 without allelic variation when compared to the DNA1 and DNA2 genome assemblies. This holds true also for the other three sequenced nuclei, though not all three BIP loci were covered in the genome assemblies, which can be attributed to incomplete amplification (Fig. S3). Next, we studied Bg112 for which three loci were identified. Again, no allelic variation was detected among the four nuclei (Fig. S4). The polymorphism of the ITS region of the multi-copy 45S rDNA locus was studied within each of the 4 nuclei. By mapping sequence reads to a reference R. irregularis ITS sequence (Genbank JF439109), many variants reported previously for strain DAOM197198 were identified within individual nuclei (Fig. 2) [8], [12]. This demonstrates that, in addition to reported intraspecific ITS variability within single R. irregularis spores [12], [18], the ITS region in the multi-repeat 45S rDNA locus is extremely variable even within individual nuclei, and that different nuclei can show quantitative variation in polymorphic ITS variants. In general, multi-repeat loci such as rDNA sequences are thought to be homogenized through concerted evolution [19], which presumably is most effective during meiosis [20], [21]. Therefore, the high level of heterogeneity among the copies within a single repeat seems to be consistent with ancient asexuality. However, also in several sexual fungal species varying levels of intra-individual polymorphism have been found [22], and R. irregularis may be an extreme case, although exact percentages cannot be deduced from the Illumina read data. Given the high level of ITS variability within single nuclei, we conclude that the 45rDNA ITS sequence is less suited for comparative studies of Glomeromycota. Based on the whole genome comparison of individual nuclei we conclude that the organization of the R. irregularis genome of the used reference culture DAOM197198w is basically homokaryotic. The high divergence observed among copies of the 45S rDNA repeat occurs within a single nucleus, indicating that this region is unsuited to claim that nuclei within a strain are highly divergent [8]. However, the presence of a low level of polymorphisms suggests that genetically, slightly divergent nuclei can arise and coexist in a single mycelium.
The reference genome assembly of DAOM197198w covers about 97% of the current R. irregularis EST collection [23] indicating that it represents nearly the complete genic region of the genome. This is further supported by a survey of core eukaryotic genes (CEG), which shows that among the 248 CEG proteins 229 (92.3%) are included in the predicted protein-coding genes (Table S4). Genome annotation using EVidenceModeler resulted in 27,392 protein-coding gene models representing 30,003 putative transcripts. Of these models 11,145 are supported by at least one R. irregularis EST, whereas an additional 5,586 protein-coding gene models find support by homology to available protein sequences. Using an AHRD functional annotation pipeline we could assign putative functions to 14,073 protein-coding gene models (Table S5). To obtain insight into the R. irregularis gene repertoire a comparative approach using OrthoMCL was conducted on 10 species representing all five fungal phyla (Fig. 3). This resulted in 19,300 putative orthology groups (Table S6), of which 1,370 contained exclusively R. irregularis gene models that may represent genes unique for AM fungi (14,742 gene models in total). Of these 6,014 were functionally annotated (Table S7). A summary of the top ten Interpro domains is shown in Table S8. Interestingly, about 28% of these putative genes are predicted to encode proteins with a kinase domain, underling a striking overrepresentation of these signaling proteins in the R. irregularis genome. The second largest group (∼25%) that seems to be enriched especially in R. irregularis are BTB domain containing proteins (BTB-POZ (PF00651) and BTB-Kelch (BACK; PF07707)). Both findings are supported a recent transcriptome study [23].
We observed a high level of putative/predicted (retro-)transposable (TE) elements in the R. irregularis genome. In addition to well-known TE classes, representing 1.1% of the genome based on the Repbase [24] TE library (Table S9), potential novel TE repeats were identified, revealing that TE repeats represented ∼40% of the genome (Table S10). The presence of potential deleterious TE elements is difficult to reconcile with the ancient asexuality of Glomeromycota, as an uncontrolled accumulation of such elements would cause a deleterious load that leads to extinction [25], [26]. Therefore, the presence of such TE elements [27], together with the identification of meiotic recombination proteins [3] and signatures of recombination within populations [28]–[30], argues for the potential rare occurrence of so far unidentified sexual reproduction in R. irregularis [25], [26]. As an alternative, parasexual cycles where nuclei fuse and undergo recombination, together with observed exchange of nuclei through anastomoses, may explain both the spread of TE elements as well as restrain their intragenomic proliferation [2], [11].
We noted that the gene repertoire of R. irregularis overlaps the most with the repertoire of sequenced Mucoromycotina species. Mucoromycotina have traditionally been classified as Zygomycota, which also have coenocytic hyphae, similarly as those in AM fungi. In general they are saprotrophic fungi, but some isolates can also act as opportunistic pathogens. A reconstruction of the early evolution of fungi largely based on the 45S rDNA locus suggested that the Zygomycota phylum is paraphyletic and that Glomeromycota are sister to the Dikarya phyla Ascomycota and Basidiomycota [31]. However, this has only limited statistical support, and analyses based on protein coding genes gave conflicting results [3], [32], [33]. As our data, together with that from others [12], [18] revealed that the 45S rDNA locus of R. irregularis is highly polymorphic we reinvestigated the phylogenetic relationships of R. irregularis within the fungi. To do so, we analysed a supermatrix of 35 highly conserved, putative single copy nuclear genes proposed by Capella-Gutiérez et al. [34], totaling a concatenated length of 26,604 aligned amino acids from 23 fungal species and 4 outgroups (Table S11). Phylogenetic analysis of this supermatrix using maximum-likelihood (ML) revealed that R. irregularis is related to Mucoromycotina rather than to the Dikarya phyla Basidiomycota and Ascomycota (Fig. 3). This phylogenetic placement of R. irregularis received maximal bootstrap support (100%; Fig. 3) and alternative placements resulted in significantly lower likelihoods (p< = 0.004; see Table S12). This finding is in concordance with gene repertoire reconstructions presented here, as well as phylogenetic studies based on genes encoding (meiotic) DNA repair proteins [3], [35], [36]. We note, however, that our taxonomic sampling includes Mucorales only. Additional lineages within Mucoromycotina (i.e. Mortierellales, Endogonales) and especially other currently unplaced subphyla traditionally classified as Zygomycota (e.g. Kickxellomycotina, Zoopagomycotina, Entomophthoromycotina) may better resolve the precise relationships of R. irregularis, as genome sequences for these members will become available in the future.
In comparison to pathogenic fungi, AM fungi have an extremely broad host range. Pathogenic fungi suppress defence responses of their host by secreting effectors that interfere with this defence. This raises the question whether a particular repertoire of secreted putative effector proteins underlies the broad host range of AM fungi. From the deduced proteome of 30,003 putative proteins, we predicted the secretome to contain 299 proteins (1% of proteome) using stringent bioinformatics criteria, and 566 proteins (1.9% of proteome) using more relaxed criteria (Table S13). In relative sense, this is rather low compared with averages of other fungal secretomes such as plant pathogens (7.4%), animal pathogens (4.7%), and non-pathogens (5.3%) (Fig. 4). It is remarkable that AM fungi are able to colonize a broad range of plants despite the fact that it has a small secretome suggesting more research is needed on the effectors. The relative small secretome may have resulted from adaptation to a symbiotic lifestyle in which the secretome has been streamlined through the loss of unnecessary secreted protein genes. The proteins in the R. irregularis secretome identified with relaxed criteria were grouped into 254 tribes based on sequence similarity, annotated, and ranked based on potential effector features (Table S13). The top 100 tribes that are likely to contain effectors highlighted five protein tribes containing thirteen sequences with similarity to the known R. irregularis effector protein SP7 (Fig. 5) [37]. Alignment of these protein sequences identified conserved features also present in SP7 (Fig. S5), indicating that these proteins are good candidates to display effector functionality. To further analyze potential R. irregularis specific features, we compared the number of predicted secreted proteins of R. irregularis in each tribe with those of selected pathogenic and symbiotic fungi (Fig. S6). A survey of top 100 tribes, containing 16–134 members, revealed that R. irregularis was represented in only 26 tribes compared to for example 76 tribes and 64 tribes for the fungi Magnaporthe oryzae and Laccaria bicolor, respectively. This suggests that not only the secretome of R. irregularis is reduced, but also that it is missing some secreted proteins that are present in other fungi compared in this analysis. However, there is a 22-member tribe composed of R. irregularis proteins only (Tribe 62 based on the numbering of Fig. S6, equivalent to the largest R. irregularis Tribe 1 of Table S13). It is tempting to speculate that such effectors play important roles in the AM symbiosis.
Among the putative effectors, a protein with a so-called Crinkler (CRN) domain was present (RirT087480; tribe 245). Secreted CRN domain effectors are abundantly present in oomycete plant pathogens of the Phytophthora genus [38], [39]. We searched the R. irregularis deduced proteome for proteins containing CRN domains using amino acid sequences of canonical CRN proteins from the potato blight pathogen Phytophthora infestans as query. This resulted in 42 sequences with positive scores for the so-called N-terminal LFLAK domain that is common to all CRN proteins (Table S14). Within this set, we also identified additional CRN domains (Fig. S7, Table S14). Among these 42 CRN-like proteins, only five have a putative signal peptide, similar as the canonical CRN proteins from P. infestans. Similar CRN domain effector-like proteins were identified in the Chytrid fungus Batrachochytrium dendrobatidis, but not yet in other sequenced fungal genomes. This led to speculations of horizontal acquisitions of these genes by this pathogenic fungus [40]. However, the occurrence of CRN genes in the R. irregularis genome makes a vertical descent equally well possible, and indicates that these proteins are encoded by an ancient eukaryotic gene family.
Genome sequencing of individual cells has previously been used for example to determine the genome of individual cancer cells [41]. However in these cases a reference genome was already available. Our study shows that it is possible to obtain a de novo genome sequence starting from a single haploid nucleus. This approach can be attractive for genomes of species with high heterozygocity that are notoriously difficult to assemble. We applied a single nucleus genome sequence approach on the AM fungus R. irregularis and provide solid evidence for the occurrence of homokaryosis in this strain. This demystifies the long lasting hypothesis that nuclei of a single Rhizophagus isolate are markedly different. The sequences of four nuclei, in combination with the reference genome sequence will provide the basis for future studies on AM fungi to address issues such as genetic selection, long-term persistence of asexuality, obligate endosymbiosis, adaptation to host plants and suppression of plant defense.
A monoxenic culture of Agrobacterium rhizogenes (RiT-DNA) transformed chicory (Cichorium intybus) roots mycorrhized with the fungus R. irregularis DAOM197198 was obtained from Dr. Paola Bonfante and Dr. Andrea Genre (University of Torino) (originally obtained from GINCO (MUCL 43194)). This root culture was designated DAOM197198w and grown in a split-plate setup, where the fungus is allowed to grow into a compartment containing liquid M medium to allow easy collection of spores and extraradical mycelium [42].
Genomic R. irregularis DNA, used for meta-genome sequencing, was isolated from extraradical mycelium containing spores using the DNeasy Plant kit (Qiagen). Mycelium containing spores was washed 10× in sterile water. Spores were carefully teased out using forceps, washed by transferring through a series of (at least 5) sterile water droplets, and finally transferred to a small drop of 10 µM Sytox Green (Invitrogen) in Citifluor (Citifluor Ltd). To release the nuclei, spores were crushed using a teflon coated dounce and transferred to an eppendorf tube. The volume was adjusted to 25 µl with 10 µM Sytox Green. To remove cell debris, the crushed spore suspension was centrifuged for 1 min. at 4000 rpm.
Spore suspensions were loaded onto cover slips, from which individual nuclei were collected using a Narishige micromanipulator mounted to an inverted PASCAL Zeiss Confocal Laser Scanning microscope (excitation 480 nm; emission 505–530 nm). Individual isolated nuclei were transferred to a PCR tube containing 5 µl 1× ALB (200 mM KOH, 0.5 mM DTT) buffer, by breaking the tip of the glass microinjection needle containing the captured nucleus.
Whole genome amplification (WGA) was performed using the REPLI-g UltraFast midi-kit (Qiagen) according to the manufacturers instructions. Amplified DNA was diluted 100×. To verify the efficiency of the WGA a set of 10 selected amplicons was amplified using Premix Taq (Ex TaqVersion 2.0) polymerase (Takara Bio Inc). Amplicons could not be amplified from WGA-amplified control suspension lacking single nuclei. The extent of contamination of the WGA amplified DNA with bacterial DNA was checked by amplification of 16S rDNA amplicons. Primers for selected amplicons are listed in Table S15. From in total 40 WGA samples, 4 samples that allowed amplification of the selected R. irregularis amplicons and showed minimal bacterial contamination were selected for Illumina sequencing.
OrthoMCL [65] was used to identify orthologous groups among the set of protein sequences extracted from the following eleven completely sequenced genomes: R. irregularis, Neurospora crassa, Tuber melanosporum, Saccharomyces cerevisiae, Laccaria bicolor, Ustilago maydis, Rhizopus oryzae, Phycomyces blakesleeanus, Batrachochytrium dendrobatidis, Magnaporthe grisea and Monosiga brevicollis [66]–[75]. Only the longest sequence of each protein-coding gene was chosen in the further analysis. The set contains 171,398 sequences. Three steps took as follows: (1) all-against-all comparison strategy was applied to the set of protein sequences by BLASTP with an E-value cutoff of 1e-5; (2) The distance matrix among all proteins was constructed by the OrthoMCL algorithm; (3) The orthologous groups were generated by MCL [76] (I = 1.5) algorithm based on the distance matrix. The software versions used in this process were: OrthoMCL version 2.02, MCL version mcl 10–201, and NCBI BLAST version 2.2.15.
We reinvestigated the phylogenetic placement of R. irregularis within the fungi based on a set of 52 low-copy genes proposed by [34] with addition of orthologs from R. irregularis, Magnaporthe orzyzae, Tuber melanosporum, Ustilago maydis, and the Cryptomycete Rozella allomyces [77]. Amino acid sequences were aligned using MAFFT [78] and positions covering less than three species were trimmed. Seventeen gene alignments supported paralogy shared among different fungal lineages and were excluded from the analysis, leaving in a total number of 35 gene alignments that were concatenated into a supermatrix of 26,604 amino acids. Table S14 lists all included protein sequences. We then estimated a ML phylogenetic tree based on the supermatrix using RAxML 7.2.8 [79] applying the amino acid substitution model with the best fit on a maximum parsimony tree (rtREV; [80] with empirical frequencies and gamma-distributed rate heterogeneity (-m PROTGAMMARTREVF). Clade support was assessed using the rapid bootstrapping algorithm [81] with 100 alignment replicates.
To test alternative hypotheses of monophyly we imposed three alternative topological constraints on parallel RAxML analyses, with R. irregularis forming a clade with either Dikarya, Chytridiomycota, or Microsporidia and Cryptomycota. Branch lengths were optimized and all competing hypotheses were compared with an unconstrained analysis using the eight bootstrap probability tests implemented in CONSEL [82]; Table S12).
The sequence data have been deposited into Genbank with accession number PRJNA230015. The R. irregularis reference genome and assemblies are also available at http://cmb.bnu.edu.cn/Rhizophagus_irregularis_v10/.
|
10.1371/journal.pbio.1001605 | Quantitative Assessment of the Importance of Phenotypic Plasticity in Adaptation to Climate Change in Wild Bird Populations | Predictions about the fate of species or populations under climate change scenarios typically neglect adaptive evolution and phenotypic plasticity, the two major mechanisms by which organisms can adapt to changing local conditions. As a consequence, we have little understanding of the scope for organisms to track changing environments by in situ adaptation. Here, we use a detailed individual-specific long-term population study of great tits (Parus major) breeding in Wytham Woods, Oxford, UK to parameterise a mechanistic model and thus directly estimate the rate of environmental change to which in situ adaptation is possible. Using the effect of changes in early spring temperature on temporal synchrony between birds and a critical food resource, we focus in particular on the contribution of phenotypic plasticity to population persistence. Despite using conservative estimates for evolutionary and reproductive potential, our results suggest little risk of population extinction under projected local temperature change; however, this conclusion relies heavily on the extent to which phenotypic plasticity tracks the changing environment. Extrapolating the model to a broad range of life histories in birds suggests that the importance of phenotypic plasticity for adjustment to projected rates of temperature change increases with slower life histories, owing to lower evolutionary potential. Understanding the determinants and constraints on phenotypic plasticity in natural populations is thus crucial for characterising the risks that rapidly changing environments pose for the persistence of such populations.
| Predictions about the effect of climate change on organisms often ignore the possibility that organisms can evolve, or that they have an inbuilt capacity to cope with changing conditions. In order to understand the potential for existing populations to adapt to climate change, and the relative risks of extinction, such processes need to be modelled together with projected changes in climate. In this paper, we use data from a long-term study (51 years) of a small bird, the great tit, to model how birds can match the time they breed each year with the time their food is most abundant, and how this match can change with a changing climate. We found that evolution offers the chance for short-lived birds to adapt at the rate of climate change that is expected over the next century, but that the most important way that birds can cope with climate change is their evolved ability to adjust their behaviour depending on the environment they experience (“plasticity”). We extrapolated the model to other bird species, to estimate their relative vulnerability to changing climate. The model shows that longer-lived species (which also tend to have fewer offspring and take longer to reach sexual maturity) are more vulnerable to extinction because their evolutionary potential is lower. For such species, the importance of close adjustment to their environment becomes even greater. Hence, knowledge of the causes and limits of individual adjustment to the environment are crucial to predict the fate of populations under climate change.
| Evidence that climate change influences many properties of wild populations of animals and plants is now ubiquitous [1]–[4]. As a consequence, there is widespread concern about the demographic and evolutionary effects of changing climate for the long-term viability of populations. A popular approach to study the impact of climate change on population viability is the use of “climate envelope models” or “niche models.” These models take environmental correlates of species presence, combined with climate change projections, to predict range shifts and extinction rates (e.g., [5]–[7]). However, such projections do not take a population's ability to adapt to changing environmental conditions into account [8]–[10]. Further, since habitat fragmentation potentially constrains range shifts to track the optimal environment, populations of many species will have to adapt in situ to a changing environment to avoid extinction. Such models may therefore not be ideally suited to predict sustainable rates of climate change for existing populations.
In contrast, mechanistic population models focus specifically on those population attributes that underlie population persistence. By assessing how phenotypic traits that influence population growth rate are affected by environmental variables, predictions of the fate of populations under varying rates of environmental change can be made [11],[12]. Recently, Chevin et al. [13] proposed a mechanistic population model that predicts the critical rate of environmental change that allows long-term population persistence by local adaptation. The main novelty of the model lies in the fact that it allows local adaptation by both genetic change (i.e., micro-evolution) and phenotypic plasticity (the potential for a given genotype to be expressed differently in different environments [14]). Since phenotypic plasticity is currently recognized as being responsible for the majority of adaptive phenotypic changes in response to climate change [15]–[18], this model is an important step forward in predicting effects of climate change on population persistence. The model combines demographic population properties (e.g., generation time, maximum intrinsic growth rate) with quantitative genetic measures (e.g., additive genetic variance, strength of stabilising selection on traits sensitive to climate change), and allows for phenotypic plasticity by incorporating the effect of the environment on the trait. Since the purpose of the model is to make predictions about the fate of wild populations, the required parameters should ideally also be estimated using data from those same populations. To do so may be challenging, as it requires long-term data describing responses to the environment, as well as extensive pedigree and fitness data, a combination of information typically only found in long-term studies of marked individuals [19].
A long-term population study on great tits (Parus major) breeding in Wytham Woods near Oxford (UK) offers a rare opportunity to parameterise the model of Chevin et al. [13] for a single population, and hence to investigate the projected effects of climate change on population viability allowing for plasticity and evolution. For many wild bird species—both marine and terrestrial species—reproduction is restricted to a short annual period, in which there is sufficient food available to meet the needs of offspring production. This period varies annually and is set by the responses of lower trophic levels to abiotic factors, which are ultimately shaped to maximise productivity [20],[21]. Although timing of this period is sensitive to ambient temperature, there is no a priori expectation that different trophic levels respond similarly to change in temperature. Hence, climate change has the potential to upset synchrony between food availability and timing of reproduction in birds, which may have important consequences for population viability [20],[21].
Successful reproduction in great tits depends to a large extent on synchronization of offspring food demand with a brief annual peak in caterpillar abundance. This can be achieved by individual adjustment of laying date to early spring temperature, which predicts the timing of the peak in food availability [22]. Repeated observation of females breeding in multiple years yields observations of individual laying dates under different spring temperatures, providing a measure of phenotypic plasticity, or the “reaction norm” to temperature [23],[24]. In addition, long-term monitoring of the annual timing of peak abundance of caterpillars feeding on newly emerged pedunculate oak (Quercus robur) leaves provides an estimate of how the optimal great tit laying date changes with temperature. An estimate of the optimum derived from an independent aspect of the environment is preferable to one derived from direct observations of birds, as it is unaffected by a potential constitutive cost of plasticity or differences in intrinsic individual quality of birds with different laying dates.
Here we parameterise Chevin et al.'s [13] model with estimates from the long-term study of Wytham Woods' great tits, and so calculate the maximum rate of sustained change in early spring temperature that allows long-term persistence of this population. We also use the model to explore the dependence of population persistence on currently observed phenotypic plasticity, and further to explore the interactions between life-history variation and plasticity as a key element in persistence of populations facing environmental change. Our aim was thus to use the model as an heuristic tool to understand the importance of phenotypic plasticity in adaptation to climate change.
Inter-annual changes in the spring temperature experienced by individuals had, as expected, a pronounced effect on great tit laying date (χ2 = 101.25; Δdf = 1; p<0.001) with individual females laying an estimated 4.98 (±0.49 standard error [SE]) days earlier for each 1°C rise in spring temperature (Figure 1). The within-individual response to spring temperature was similar to the difference in laying date between individuals that experienced different spring temperature, as averaged over all their reproductive attempts (estimate ± SE = −4.31±0.50; χ2 = 75.39; Δdf = 1; p<0.001), indicating that the relationship between annual population average laying date and spring temperature is predominantly caused by phenotypic plasticity (Figure 1), as found previously [22]; note that any evolutionary response to selection would be captured in the between-individual term. Phenotypic plasticity in response to increasing mean spring temperature has resulted in an advance of average laying date by about 2 wk in the last half century [22].
Caterpillar half-fall date (an index for timing of peak food availability; see “Materials and Methods”) also reacted strongly to spring temperature (χ2 = 90.10; Δdf = 1; p<0.001), with half-fall date advancing an estimated 5.30 (±0.56 SE) days per 1°C rise in spring temperature (Figure 1), a rate only slightly more rapid than the response of great tits over the same period. The effect of spring temperature on half-fall date did not change over time (spring temperature×year; estimate ± SE = −0.05±0.04; χ2 = 1.57; Δdf = 1; p = 0.21), and we thus assume that the reliability of spring temperature as a cue for the optimal laying date has been constant. Overall we conclude that the response in laying date of individual great tits to spring temperature (corresponding to b in Chevin et al.'s model; see Table 1) closely matches the optimal response (the term represented by B in their model).
Combining parameter estimates for Chevin et al.'s model (Table 1), the Wytham great tit population is predicted to be able to adapt to a maximum long-term rate of increase in spring temperature of 0.47°C y−1, i.e. >15 times the rate of temperature change of 0.030°C y−1 predicted under a high emissions scenario for this location and time in the annual cycle [25]. However, this estimate does not take uncertainty in parameter estimates into account. To calculate the probability that the modelled critical rate of change (ηc) will fall below 0.030°C y−1 while accounting for parameter uncertainty, we ran 100,000 simulations, with each simulation randomly sampling from a normal error distribution of parameters σ2h2, γ, T, B, and b. This resulted in an estimated probability of 0.001 that ηc falls below 0.030°C y−1 (Figure 2a), and hence again very little likelihood of extinction due to predicted temperature change. If we assume that there is no phenotypic plasticity in great tit laying date (hence: |B−b| = 5.30) the point estimate of ηc is 0.028°C y−1, with a 60% probability of population extinction (ηc<0.030) when the error around the parameter estimates of σ2h2, γ, and T is taken into account (Figure 2b). Hence, the likelihood of population persistence in a changing environment depends heavily on the presence of phenotypic plasticity, as extinction risk is >500-fold higher in the absence of phenotypic plasticity.
We explored the sensitivity of the probability of population extinction for other species with different life histories, assuming similar rates of change in the environment (see Discussion), by varying the demographic and life-history parameters T (generation time) and rmax (maximum rate of annual population growth) while holding other parameters in the model constant; these effects are illustrated with contour plots in Figure 3a and 3b. This exercise revealed that with a difference in observed and optimal reaction norm equivalent to that seen in Wytham great tits (|B−b| = 0.32), which we take as indicative of a population showing close matching to the environment (note that, when |B−b| = 0 [perfect tracking of the environment], ηc is undefined), the model suggests little concern about a population being unable adapt to a rate of environmental change equivalent to an increase in spring temperature of 0.030°C y−1, over most of the range of T and rmax (Figure 3a). However, since the fundamental life-history trade-off between survival and reproduction leads, in general, to a negative correlation between T and rmax [26],[27], organisms with the slowest life histories (i.e., high T, low rmax) are, even with quite close phenotypic matching (Figure 1), not far from the region at which risk begins to be appreciable.
It is not plausible that great tit life history parameters such as generation time would evolve rapidly enough to the extent that the risk of population extinction would become substantial with the observed phenotypic plasticity. However, by setting phenotypic plasticity to zero (|B−b| = 5.30), we can explore the importance of phenotypic plasticity, and the extinction risk given these rates of environmental change, across the life-history continuum for other birds. Plotting T and rmax values for 13 species of birds [28] in Figure 3b shows a general pattern (rmax = 0.92T−0.92) under which populations of other species with longer generation times are much less likely to adapt to increasing temperatures in the absence of phenotypic plasticity, assuming that the quantitative genetic parameters determining evolvability (σ2h2 and γ) are similar to that of the studied population of great tits (see also Figure 4).
We then explored the sensitivity of our conclusions to varying evolvability of populations while holding other quantities constant. Figure 3c shows that, with the observed life history and phenotypic plasticity in laying date, our conclusions about the ability of this great tit population to adjust to the high emissions scenario projected temperature change of 0.030°C y−1 are quite robust to variation in the estimated genetic variance (σ2h2) in laying date and strength of stabilising selection (γ) on laying date. In the absence of phenotypic plasticity, the population is at the threshold at which the additive genetic variance (σ2h2) in laying date is insufficient for the population to remain viable (Figure 3d). Equally, if the strength of stabilising selection on the match with the environment were weaker, extinction risk would also be elevated. However, in general it appears that a relatively fast life history provides sufficient potential to considerably reduce the risk of population extinction due to climate change.
In this study we explored the viability of a well-studied wild bird population to changes in climate predicted to the end of this century, by parameterising a mechanistic model by Chevin et al. [13]. We further explored the dependence of population viability on phenotypic plasticity as a form of adaptation to the environment, and the extent to which these conclusions depend on life history, and on evolvability. Our general conclusions are that the importance of phenotypic plasticity in adaptation to climate change is strongly dependent on life history. Short-lived species, with high reproductive rates, are more resilient to expected rates of climate change even with relatively little phenotypic plasticity, and while phenotypic plasticity is likely to be an adaptive response to environmental uncertainty in such species, it is not the only potential form of adaptation to climate change unless generation time encompasses multiple years and the rate of reproduction is slow. While the parameters we fitted to the model were determined by the specific details of our study system, we discuss below the extent to which our conclusions can be generalised.
Like all models, the model by Chevin et al. [13] makes assumptions to simplify reality. For example, the model assumes no stochastic variation in optimal timing of reproduction. Stochastic variation occurring over time scales shorter than a species' generation time can only be countered with phenotypic plasticity, and as such the model may underestimate the importance of phenotypic plasticity. Our conclusions should therefore be interpreted with respect to long-term directional climate change only, assuming that population demography is buffered against environmental stochasticity. Such buffering, in the present system, may be accomplished by the fact that generations overlap and adult survival is largely independent of the match with the environment [29],[30]. This possibility is not accounted for by the model as it assumes non-overlapping generations. Further, if adult survival is independent of the match to the environment, any evolutionary response to directional change is likely to be retarded. Moreover, in applying the model we have assumed that both the response to environmental cues and the dependence of the environment on climate can be extrapolated outside the ranges currently observed. In the case of the three trophic levels studied here (oaks, caterpillars, and great tits) the possibility remains that they exhibit differential phenotypic responses or physiological tolerances to increased temperature. If so, it is questionable whether the degree of matching can be assumed to be fixed over time. In this respect it is noteworthy that the model also allows for overcompensation, which would be just as detrimental as under-compensation, and causes a modification of predictions when parameter error is incorporated, as this results in a skewed error distribution of |B−b| (Figure 2a).
Although we incorporated error in parameter estimates for our predictions of extinction probability, this does not exclude the possibility that certain parameters and associated errors are systematically over- or underestimated. Estimates of the additive genetic variance for laying date in birds have been derived in several ways, from different study species with a range of life histories (reviewed in [31]; see also Text S1 for further discussion). While there is considerable variability in the estimates, it is likely that many estimates are inflated by a failure to control effectively for common environmental effects between relatives, which can be expected to be considerable for a trait with a strong link to environmentally determined phenology (see also [32],[33]). In this study we used an estimate of σ2h2 derived from a very low heritability estimate (0.03) from an animal model controlling for several types of environmental variance [31]. We suspect that estimates of the additive genetic variance for time of breeding will be closer to this value than many previous estimates once appropriate environmental control is built into models. Sex-limited expression of traits will reduce the response to selection. While laying date is a phenotype only expressed by pairs of birds, in many, but not all, species it is primarily determined by the female [34]. Hence, the predicted evolutionary response to selection can be over-estimated if sex-limitation is not considered. The strength of stabilising selection on timing of breeding used here (γ) is more likely to be an underestimate as this is based on observational data at the level of the population. Two likely additional sources of stabilising selection that are not considered by such data result from, first, the extent to which individuals optimise timing of breeding to the phenology of their local environment. If there are different optima for different locations, then birds in the tails of the population phenotypic distribution may be closer to their local optimum than assumed: hence phenotypes should be measured at the appropriate relative scale. A second effect that will underestimate stabilising selection is the extent to which apparent directional selection on laying date results from phenotypic covariance between other aspects of individual quality and breeding date [35]. Figure 3d suggests that, if the match between organisms and the environment is poor, the outcomes of the model may be sensitive to variation in the strength of stabilising selection, or the additive genetic variance. However, the model assumes a fixed strength of stabilising selection, whereas it might be expected that as the match between a population and a changing environment became poorer, the strength of stabilising selection would increase. Lastly, the estimate of rmax (0.49) employed here may be an underestimate, as this does not include immigrants, which compensate for recruits that have dispersed from the population [28],[36]. In summary, with the other parameters being relatively straightforward to estimate, any systematic bias in parameter estimates is most likely in the direction such that the potential for micro-evolutionary adjustment to climate change is underestimated.
Extrapolating the model using parameters derived from a single great tit population to other bird species suggests that species at the faster end of the life-history continuum would have sufficient evolutionary potential to adapt phenology to a temperature change of the order of 0.030°C y−1 (Figures 3a, 3b, and 4). The predicted rates of change for the study area from United Kingdom Climate Projections 2009 (UKCP09) [25] are broadly comparable to predicted rates of global temperature change, as IPCC [37] scenarios predict similar or less temperature change for this century. However, how representative are the parameter estimates derived from this single population for other species and populations? Current knowledge suggests that evolutionary potential of most bird species in terms of phenological adaptation should be broadly similar, since heritability for laying date is not likely to be much greater than the value used here [31], and predictions are not very sensitive to values of γ (Figures 3c). While heritability may decrease under adverse environmental conditions [38],[39] the opposite may also apply [40],[41] and at present there is no evidence of climate-related dependence of the heritability of laying date in our study population [40],[42].
Estimates of the optimal phenotypic response to changing environmental conditions (in the present study, the optimal response in laying date to temperature [B], as determined from the response of the timing of caterpillar peak abundance to temperature) are not widely available. An estimate of B for another very well-studied Dutch population of great tits is lower than the one for our population (−4.01 versus −5.30; [43]), and this is a population for which the phenotypic response of the birds is also lower (see [40] for a comparison), suggesting that |B−b| would be larger than in the Wytham population. To the best of our knowledge, there are few comparable estimates from other systems, though see [44]. In general, one can expect that optimal responses are determined by the response of lower trophic levels in the food chain [21],[45]. In that respect, observations for 1,558 largely Northern hemisphere wild plant species suggesting an average advance in spring leafing and flowering of 5–6 d per °C [46], suggest that our estimate of B (which is also in units of days per °C) is quite representative of terrestrial systems in the Northern hemisphere. Rates of change in higher trophic levels (i.e., b) may be more variable. A large-scale analysis of data from three decades across environments in the UK by Thackeray et al. [3] suggested that while primary producers and consumers have shown broadly comparable rates of advance with climate change, secondary consumers have on average advanced at only about half the rate. Hence, the general expectation might be that B and b will not be very closely matched, and that a scenario intermediate to the two we modelled (close match between B and b; no plasticity at all) is most common. It should be noted that our conclusions are drawn from analysis of plasticity in phenology, and given considerable annual variability, phenological traits may have a very high degree of plasticity. Other traits, for instance thermal tolerance, or migration timing, might show less plasticity, but we are not aware of studies of other classes of trait that would support analysis in the framework used here.
Recently a similar approach to calculate the risk of extinction for a Dutch population of great tits yielded a more pessimistic outcome [47]. This is predominantly caused by the combination of lower plasticity, weaker selection, and more extreme climate change scenarios (up to 0.067°C y−1) [47]. However, in contrast with our study population, where average offspring recruitment is lower in years with stronger selection on relative laying date [22] and about 13% of annual population growth can be explained by the population's match with the food peak (unpublished data), population growth is hardly affected by the match with the food peak in the Dutch population [30],[47],[48]. This illustrates that even when the match with the food peak is the single most important factor explaining relative fitness, other ecological processes that determine population growth or absolute fitness (e.g., density dependence)—the effects of which on population viability in response to climate change are less straightforward to estimate—can potentially mitigate adverse population effects [30],[48].
In contrast to cases where there is a close tracking of the environment, inability to adjust phenotypically to a gradual shift in optimal timing caused by climate change suggests very high risks of population extinction in species with long generation times (Figures 3b and 4). Such risks could potentially be buffered with higher evolvability, but we are unaware of any evidence for a link between life history and genetic variance. The greater vulnerability of species with slower life histories contrasts with predictions of Morris et al. [49] who suggested longevity should act as a buffer against climatic variability. This raises the question of whether longer-lived species will have already evolved a sufficiently plastic response in timing of reproduction, to variation in temperature, to cope with the relatively fast directional change that is predicted for the future. This is especially relevant as our results show that their long generation time limits their potential to respond with genetic adaptation to climate change.
In conclusion, parameterisation of Chevin et al.'s [13] model with conservative estimates from an extensively studied wild bird population suggests little risk of extinction of that population due to future change in temperature as predicted by climate models. By varying terms in the model we estimated that the absence of phenotypic plasticity would increase the likelihood of population extinction approximately 500-fold. For birds with longer generation times, vulnerability to extinction is considerably higher even for only moderate mismatches of phenotypic plasticity with the rate of environmental change, as they may exhibit insufficient evolutionary potential to adjust to relatively fast change. For those species, phenotypic plasticity in timing of reproduction is likely to be by far the most effective mechanism to cope with constantly increasing temperatures. However, relatively less is known about the determinants and limits on plasticity in such organisms, and increased focus on this area, as well as work on the link between phenotypic plasticity and life history would be very valuable.
Great tits are small (14–22 g) passerine birds, common in large parts of Europe, Asia, and Northern Africa [50]. They are socially monogamous and breed in cavities, but readily accept nestboxes, if provided. Wytham Woods (Oxfordshire, UK, 51°46′ N 1°20′W) consists of ca 385 ha mixed deciduous woodland with an excess of nestboxes (n = 1,020) available since 1960. On average 217 nestboxes are occupied annually by great tits [51], although population size has increased in recent decades. Second broods are rare (<3%) and typically excluded from analyses (e.g., [22]). Data collection in the breeding season (April–June) consists of weekly nestbox checks in the laying phase to record first egg date (here referred to as “laying date”) and clutch size. Occupied nestboxes are checked every 2 d around the anticipated hatching date to infer hatching date and allow ringing of nestlings (for future identification) at a standard age of 15 d. At least 5 d later, nestboxes are checked for successful fledging of nestlings. Parents are caught in the nestbox while feeding nestlings, and identified by their ring, or newly ringed if immigrant. Recruits to the natal population are defined as locally hatched birds that were caught as a parent in subsequent years. For analyses in this paper, we use data collected between 1960–2010, as field protocols were standardised over this period.
Chevin et al.'s model [13] extends an earlier model by Lynch and Lande [11], by incorporating plasticity in a phenotypic trait (z, here first egg-laying date) that mediates adaptation to a changing continuous environmental parameter (ε, here temperature). It predicts the maximum rate with which ε can change (at a constant rate in time) to allow long-term population persistence, referred to as the critical rate of environmental change (ηc). In the original model [11] ηc depended only on the phenotypic variance (σ2) in z, the heritability (h2) of z (together comprising the additive genetic variance for z), the strength of stabilising selection (γ, [52]) on z, and the maximum intrinsic rate of population growth (rmax). Note that γ refers to selection on unstandardised phenotypic variation, assumes the absence of strong directional selection, and a positive value represents stabilising selection, rather than disruptive selection. The extended model also includes the species' generation time (T), with T being expressed on the same units of time scale as ηc and rmax (here in years; rmax is measured in years−1). Furthermore, it includes the environmental sensitivity of selection (B), which reflects how the optimal value of z (laying date) depends on ε (temperature), and the degree of phenotypic plasticity or reaction norm (b), which quantifies the effect of ε on z, within individuals. Altogether the critical rate of change is modelled as:We refer to Chevin et al. [13] for a more detailed description of the model and its rationale.
We used a range of previously published estimates and new analyses to parameterise the model, all of them specific to the Wytham Woods study population. All parameter estimates are listed in Table 1 and, for cases where we used previous estimates from this population, we refer to Figure S1, Text S1, and the specific publication for exact methodological details. Some parameters have been estimated multiple times, and can vary because of different data inclusion criteria, different time spans, different assumptions, or different statistical estimation procedures. In such cases we used the most recent estimate of the respective parameter, as these generally used most data, and employed the most appropriate estimation procedures (see Figure S1 and Text S1 for more discussion).
We estimated the strength of stabilising selection on laying date (γ) with the following equation: −(ω2+σ2)−1 = γ−β2 [53]. The width of the fitness function (ω) for laying date was estimated by calculating year-centred laying dates (i.e., subtracting annual average laying date, n = 8,646 laying dates in 51 y), categorising them in 10 equally spaced intervals, and calculating the average number of recruits per breeding attempt for each category. A Gaussian function (Figure S1) was fitted to these average numbers and ω was estimated as the “standard deviation” of the function (ω = 11.62). Phenotypic variance (σ2) in laying date was estimated as the average of all annual values (σ = 5.39). Since the model by Chevin et al. [13] assumes that the population is initially well adapted, we set the strength of directional selection (β) at zero, and calculated γ as −0.0061. The assumption of an initially well adapted population, and thus zero directional selection, is required by the model, yet depending on the match with the food peak there can be strong directional phenotypic selection on laying date observed [22]. Since we have no indication that the population is currently poorly adapted, the observed phenotypic selection on laying date may be biased by phenotypic covariance between other aspects of individual quality and laying date (see also Discussion). Using a bootstrapping procedure we estimated the standard error of γ as 0.0010. Note that we use the absolute value of γ in the model.
A recent study by Husby et al. [40] showed that the average temperature between 15 February and 25 April (here referred to as “spring temperature”) is the best predictor of average annual laying date; we thus used the individual response in laying date to this environmental variable as an estimate of phenotypic plasticity, and the response in the date of standardised caterpillar abundance as an estimate of environmental sensitivity (see details below). A similar exercise to that of Husby et al. [40] had been performed earlier, but based on a longer time series and a slightly different environmental variable, i.e., “warmth sum” (the sum of the daily maximum temperatures between 1 March and 25 April, [22]). We chose to conduct analyses with the average temperature, as used by Husby et al. [40], to permit more straightforward comparison between the modelled critical rate of environmental change and predictions about future climate change; see [22] for detailed information on how great tit laying date and peak caterpillar abundance date have changed over time.
We used the daily average of minimum and maximum temperatures (in °C) that were collected by the Radcliffe Observatory in Oxford, 5 km east of Wytham Woods, for our measure of spring temperature. The date by which 50% of the seasonal total of winter moth caterpillars (Operopthera brumata larvae, the main source of food for great tit nestlings; [54]–[56]) had descended from trees to pupate on the ground (here referred to as “caterpillar half-fall date”) was recorded in Wytham Woods in the majority of years from 1961 onwards (n = 43), and gives a good indication of the timing of the peak in caterpillar biomass (see [22] for more details). Given a fixed period between great tit laying date and peak offspring food demand, this serves as a proxy for the optimal response in laying date to spring temperature [22]. Hence, environmental sensitivity of selection (B) was accordingly calculated as the slope of the linear function of caterpillar half-fall date in response to spring temperature.
Phenotypic plasticity, or the average within-individual response in laying date to changes in spring temperature, was calculated from a dataset restricted to females that bred at least twice (n = 4,742 reproductive attempts of 1,874 females, in 51 y). The within-individual slope was calculated by using the difference between the spring temperature a female experienced before a specific reproductive attempt with the average of the spring temperatures a female experienced before all her reproductive attempts, as explanatory variable in a model on laying date (following [57]). In the model we also included the average of the spring temperatures a female experienced before all her reproductive attempts as explanatory variable, to account for potential micro-evolution or selective (dis)appearance of individuals with higher, or lower, average spring temperature experience. Female identity, year, and sector of the wood (Wytham Woods consists of nine different sectors with different vegetation types and management regimes, see [58]) were included as random effects, to correct for an uneven distribution of repeated measures of individuals, inter-annual variation (not due to spring temperature) and environmental heterogeneity, respectively. Models were fitted with a normal error distribution and a Markov Chain Monte Carlo estimation algorithm with 100,000 iterations, using MLwiN version 2.02 [59],[60]. Significance of explanatory terms was determined using the Wald statistic, which approximates the χ2 distribution.
We used projections from the United Kingdom Climate Projections 2009 (UKCP09, [25]) to compare our results against the predicted rate of average temperature change for the Wytham Woods area. To this end, we used the average temperature change predictions for the 25-km grid box that contained Wytham Woods (number 1,547) for the 2070–2099 time period, under the low, medium, and high emissions scenario, for the months February, March, and April. We weighted the predictions per month according to their number of days contained in our measure of spring temperature (see above). To calculate an annual rate of change we used the midpoint of 2070–2099 relative to the midpoint of the baseline period (1961–1990). This resulted in a predicted rate of increase of spring temperature of 0.021, 0.025, and 0.030°C y−1 for the low, medium, and high emissions scenario, respectively.
|
10.1371/journal.pcbi.1006351 | The role of the encapsulated cargo in microcompartment assembly | Bacterial microcompartments are large, roughly icosahedral shells that assemble around enzymes and reactants involved in certain metabolic pathways in bacteria. Motivated by microcompartment assembly, we use coarse-grained computational and theoretical modeling to study the factors that control the size and morphology of a protein shell assembling around hundreds to thousands of molecules. We perform dynamical simulations of shell assembly in the presence and absence of cargo over a range of interaction strengths, subunit and cargo stoichiometries, and the shell spontaneous curvature. Depending on these parameters, we find that the presence of a cargo can either increase or decrease the size of a shell relative to its intrinsic spontaneous curvature, as seen in recent experiments. These features are controlled by a balance of kinetic and thermodynamic effects, and the shell size is assembly pathway dependent. We discuss implications of these results for synthetic biology efforts to target new enzymes to microcompartment interiors.
| Bacterial microcompartments are protein shells that encase enzymes and reactants to enable bacteria to perform vital reactions, such as breaking down chemicals for energy or converting the products of photosynthesis into sugars. Microcompartments are essential for many bacteria, including human pathogens. Thus, there is great interest in understanding how microcompartment shells assemble around their cargo (the interior enzymes and reactants), and what determines the structure and size of a microcompartment. These questions are difficult to answer with experiments alone, because most intermediates in the assembly process are too short-lived to characterize in experiments. Therefore, this article describes theoretical and computational models for microcompartments, which predict assembly pathways and how the sizes of assembled shells depend on factors such as protein interactions and concentrations. The simulations show that the properties of the cargo are an important factor for determining shell size, and suggest an explanation for recent experimental results showing that cargo can either increase or decrease shell size. In addition to helping to understand the natural behavior of microcompartments, the simulations provide guidance to researchers working to reengineer microcompartments to produce drugs or biofuels.
| While it has long been recognized that membrane-bound organelles organize the cytoplasm of eukaryotes, it is now evident that protein-based compartments play a similar role in many organisms. For example, bacterial microcompartments (BMCs) are icosahedral proteinaceous organelles that assemble around enzymes and reactants to compartmentalize certain metabolic pathways [1–10]. BMCs are found in at least 20% of bacterial species [2, 11, 12], where they enable functions such as growth, pathogenesis, and carbon fixation [1, 10, 13–16]. Other protein shells act as compartments in bacteria and archea, such as encapsulins [17] and gas vesicles [17, 18], and even in eukaryotes (e.g. vault particles [19]). Understanding the factors that control the assembly of BMCs and other protein-based organelles is a fundamental aspect of cell biology. From a synthetic biology perspective, understanding factors that control packaging of the interior cargo will allow reengineering BMCs as nanocompartments that encapsulate a programmable set of enzymes, to introduce new or improved metabolic pathways into bacteria or other organisms (e.g. [10, 20–29])]. More broadly, understanding how the properties of a cargo affect the assembly of its encapsulating container is important for drug delivery and nanomaterials applications.
Despite atomic resolution structures of BMC shell proteins [1, 10, 30, 31], the factors that control the size and morphology of assembled shells remain incompletely understood. BMCs are large and polydisperse (40-600 nm diameter), with a roughly icosahedral protein shell surrounding up to thousands of copies of enzymes [1, 7–9, 30, 32, 33]. For example, the best studied BMC is the carboxysome, which encapsulates RuBisCO and carbonic anhydrase to facilitate carbon fixation in cyanobacteria [1, 30, 32, 34]. BMC shells assemble from multiple paralogous protein species, which respectively form homo-pentameric, homo-hexameric, and pseudo-hexameric (homo-trimeric) oligomers [1, 30, 31]. Sutter et al. [31] recently obtained an atomic-resolution structure of a complete BMC shell in a recombinant system that assembles small (40 nm) empty shells (containing no cargo). The structure follows the geometric principles of icosahedral virus capsids, exhibiting T = 9 icosahedral symmetry in the Caspar-Klug nomenclature [35, 36] (meaning there are 9 proteins in the asymmetric unit). The pentamers, hexamers, and pseudo-hexamers occupy different local symmetry environments.
Although the Sutter et al. [31] structure marks a major advance in understanding microcompartment architectures, it is uncertain how this construction principle extends to natural microcompartments, which are large (100-600 nm), polydisperse, and lack perfect icosahedral symmetry. Moreover, the effect of cargo on BMC shell size is hard to interpret from experiments. In some BMC systems, empty shells are smaller and more monodisperse than full shells [23, 28, 31, 37], whereas in other systems empty shells are larger than full ones [38]. Thus, the cargo may increase or decrease shell size.
The encapsulated cargo can also affect BMC assembly pathways. Microscopy experiments showed that β-carboxysomes (which encapsulate form 1B RuBisCO) undergo two-step assembly: first the enzymes coalesce into a ‘procarboxysome’, then shells assemble on and bud from the procarboxysome [39, 40]. In contrast, electron micrographs suggest that α-carboxysomes (another type of carboxysome that encapsulates form 1A RuBisCO) assemble in one step, with simultaneous shell assembly and cargo coalescence [33, 41]. Our recent computational study [42] suggested that the assembly pathway depends on the affinity between cargo molecules. However, that study was restricted to a single shell size, and thus could not investigate correlations between assembly pathway and shell size.
Numerous modeling studies have identified factors controlling the thermodynamic stability [43–45] or dynamical formation [46–54] of empty icosahedral shells with different sizes. For example, Wagner and Zandi showed that icosahedral shells can form when subunits sequentially and irreversibly add to a growing shell at positions which globally minimize the elastic energy, with the preferred shell size determined by the interplay of elastic moduli and protein spontaneous curvature. Several studies have also investigated the effect of templating by an encapsulated nanoparticle or RNA molecule on preferred shell size [50, 55–57]. However, the many-molecule cargo of a microcompartment is topologically different from a nucleic acid or nanoparticle, and does not template for a specific curvature or shell size.
Rotskoff and Geissler recently proposed that microcompartment size is determined by kinetic effects arising from templating by the cargo [58]. Using an elegant Monte Carlo (MC) algorithm they showed that proteins without spontaneous curvature, which form polydisperse aggregates in the absence of cargo, can form kinetically trapped closed shells around a cargo globule. However, there are reasons to question the universality of this mechanism for microcompartment size control. Firstly, several recombinant BMC systems form small, monodisperse empty shells [23, 28, 31, 37], suggesting that the shell proteins have a non-zero spontaneous curvature even without cargo templating. Secondly, when Cameron et al. [39] overexpressed RuBisCO to form ‘supersized’ procarboxysomes, carboxysome shells encapsulated only part of the complex, suggesting that there is a maximum radius of curvature that can be accommodated by the shell proteins. Thirdly, the kinetic mechanism is restricted to systems in which rates of shell association vastly exceed cargo coalescence rates, a condition which may not apply in biological microcompartment systems. Thus, despite this and other recent simulation studies of microcompartments [42, 58, 59], the factors which control BMC size and amount of encapsulated cargo remain unclear.
In this article we use equilibrium calculations and Brownian dynamics (BD) simulations on a minimal model to identify the factors that control the size of a microcompartment shell. Although computationally more expensive than the MC algorithm of Ref. [58], BD better describes cooperative cargo-shell motions and thus allows for any type of assembly pathway. Using this capability, we explore the effect of cargo on shell size and morphology over a range of parameters leading to one-step or two-step assembly pathways. To understand the interplay between shell curvature and cargo templating, we consider two limits of shell protein interaction geometries: zero spontaneous curvature and high spontaneous curvature, which respectively form flat sheets or small icosahedral shells in the absence of cargo.
Our calculations find that the presence of cargo can increase or decrease shell size, depending on the stoichiometry of cargo and shell proteins, and the protein spontaneous curvature. For shell proteins with high spontaneous curvature, we observe a strong correlation between assembly pathway and shell size, with two-step assembly leading to larger shells than single-step pathways or empty shell assembly. This result is consistent with the fact that β-carboxysomes tend to be larger than α-carboxysomes. For shell proteins with zero spontaneous curvature, we find that introducing cargo can result in a well-defined shell size through several mechanisms, including the kinetic mechanism of Ref. [58] and the ‘finite-pool’ effect due to a limited number of cargo particles available within the cell. However, spontaneous curvature of the shell proteins allows for robust shell formation over a wider range of parameter space.
We simulated assembly dynamics using the Langevin dynamics algorithm in HOOMD (which uses GPUs to efficiently simulate dynamics [81]), and periodic boundary conditions to represent a bulk system. The subunits are modeled as rigid bodies [82]. Each simulation was performed in the NVT ensemble, using a set of fundamental units [83] with σ0 defined as the circumradius of the pentagonal subunit (the cargo diameter is also set to σ0), and energies given in units of the thermal energy, kBT. The simulation time step was 0.005 in dimensionless time units, and we performed 3 × 106 timesteps in each simulation unless mentioned otherwise.
Initial conditions. We considered two types of initial conditions. Except where stated otherwise, simulations started from the ‘homogeneous’ initial condition, in which subunits and (if present) cargo were initialized with random positions and orientations, excluding high-energy overlaps. In the ‘pre-equilibrated globule’ initial condition, we first initialized cargo particles with random positions (excluding high-energy overlaps), and performed 105 simulation timesteps to equilibrate the cargo particles. Shell subunits were then added to the simulation box with random positions and orientations, excluding high-energy overlaps.
Systems. We simulated several systems as follows. For shell subunits with spontaneous curvature we set pentamer-hexamer and hexamer-hexamer angles consistent with the T = 3 geometry (see Estimating the shell bending modulus in section S2 Text), and we set εangle = 0.5. We first performed a set of empty-shell assembly simulations, with 360 hexamers, and varying number of pentamers, in a cubic box with side length 60σ0, with εHH = 2.6kBT (the smallest interaction strength for which nucleation occurred). These simulations were performed for 107 timesteps to obtain sufficient statistics at low pentamer concentrations despite nucleation being rare.
For cargo encapsulation by subunits with spontaneous curvature, we simulated 2060 cargo particles, 180 pentamers, and 360 hexamers in a cubic box with side length 60σ0. Other parameters were the same as for the empty-shell simulations, except that we varied εPH, εSC, and εSC as described in the main text. All simulations with spontaneous curvature used εPH ≥ 1.3εHH to ensure that the shells with the T = 3 geometry (or asymmetric shells with similar sizes) were favored in the absence of cargo. We note that our results generalize to other ranges of shell interaction parameters, but this choice distinguishes effects due to cargo from those due to changes in the inherent preferred shell geometry. Simulations with strong cargo-cargo and cargo-shell interactions (εCC ≥ 1.55 and εSC < 8.75) required a long timescale for pentamers to fill pentameric vacancies in the hexamer shell (discussed in Results). To observe pentamer adsorption, these simulations were run for up to 9 × 106 simulation timesteps.
For simulations of ‘flat’ subunits (with no spontaneous curvature), we considered a range of system sizes at fixed steady state cargo chemical potential, with the number of cargo particles varying from 409 to 3275, and the box side length varying from 35σ0 to 70σ0. Since these were NVT simulations, we ensured that the final hexamer chemical potential was the same at each system size by setting the number of hexamers so that the concentration of free hexamers remaining after assembly of a complete shell was constant (10−3 subunits/σ 0 3). The resulting number of hexamers varied from 109 to 581 in boxes with side lengths 35σ0 to 70σ0. The assembly outcomes were unchanged if instead we kept the total hexamer subunit concentration the same across all simulations. For each of these system sizes we performed simulations over a range of εangle to identify the maximum value of κs at which assembly of a complete shell could occur. Simulations were stopped upon completion of a shell or after the maximum simulation time tmax with tmax = 3 × 106 timesteps for boxes with side length ≤ 55σ0 and tmax = 8 × 106 for boxes with side length ≥ 55σ0. The maximum simulation time was increased for large system sizes because the minimum time required for assembly of a complete shell increases linearly with the shell size [84].
To estimate the relationship between the shell bending modulus κs and the parameter εangle we performed additional simulations, in which we measured the total interaction energy of completely assembled shells as a function of εangle (see ‘Estimating the shell bending modulus’ in section S2 Text).
Sample sizes. For simulations of shells with spontaneous curvature, we performed a minimum of 10 independent trials at each parameter set. To enable satisfactory statistics on shell size and morphology for parameter sets that result in at most one complete shell in the simulation box 3, we performed additional trials such that at least 10 complete shells were simulated. For flat subunits (Fig 1F and 1G), we identified the maximum εangle for which a complete shell forms at each system size as follows. We first performed independent simulations over a range of εangle values, separated by increments in εangle of 0.02 for systems with box side length ≤ 55σ0, and increments of 0.05 for systems with side length ≥ 55σ0. We performed 10 independent trials at each value of εangle. For the largest value of εangle at which at least one of these trials resulted in a complete shell, we then performed 10 additional trials to obtain a more accurate estimate of the shell bending modulus κs at the maximum εangle.
To simulate the dynamics of microcompartment assembly, we build on the model developed by Perlmutter et al. [42], which allowed only a single energy minimum shell geometry, corresponding to a T = 3 icosahedral shell containing 12 pentamers and 20 hexamers. We have now extended the model to allow for closed shells of any size. Based on AFM experiments showing that BMC shell facets assemble from pre-formed hexamers [60], and the fact that carboxysome major shell proteins crystallize as pentamers and hexamers [30], our model considers pentamers and hexamers as the basic assembly units. These are modeled as rigid bodies with short-range attractions along their edges, which drive hexamer-hexamer and hexamer-pentamer association. Repulsive subunit-subunit interactions control the preferred angle of subunit-subunit interactions, which sets the shell protein spontaneous curvature (Fig 1A and 1B). To minimize the number of model parameters, we do not explicitly consider pseudo-hexamers; thus, the model hexamers play the role of both hexamers and pseudo-hexamers.
We particularly focus on carboxysomes, for which the most experimental evidence is available, although our model is sufficiently general that results are relevant to other microcompartment systems. In carboxysomes, interactions between the RuBisCO cargo and shell proteins are mediated by non-shell proteins containing ‘encapsulation peptides’ [39, 41, 71, 85–88]. For simplicity we model these interactions as direct-pair attractions between model cargo particles and shell subunits. Because there is no evidence that encapsulation peptides interact with pentamers, in our model the cargo only interacts with hexamers. Further details of the model and a thermodynamic analysis are given in section Materials and methods and section S2 Text.
There are numerous parameters which can affect shell size, including the interaction strengths among the various species of cargo and shell subunits, shell protein spontaneous curvature and bending modulus, and the concentration of each species. To facilitate interpretation of results from this vast parameter space, we focus our simulations on two extreme limits. In the first limit, we consider shell subunits with a spontaneous curvature that favors assembly of the smallest icosahedral shell, the T = 3 structure with 12 pentamers and 20 hexamers (Fig 1D). In the second limit we consider a system containing only hexamer subunits with no preferred curvature, which form flat sheets without cargo (Fig 1F).
We begin by considering shells with T = 3 spontaneous curvature (Fig 1D). To isolate the effects of cargo on shell size, we consider shell-shell interaction parameters which favor pentamer insertion (setting the ratio of pentamer-hexamer and hexamer-hexamer affinities εPH/εHH ≥ 1.3) so that assembly without cargo results in primarily T = 3 empty shells for our ratio of pentamer to hexamer concentrations, ρp/ρh = 0.5, and results in shells close in size to the T = 3 geometry at all of the stoichiometries we consider here. A typical assembly trajectory without cargo is shown in Fig 2A. When simulating assembly around cargo, we set the hexamer-hexamer affinity εHH ≤ 2.2 (while maintaining εPH/εHH ≥ 1.3) so that assembly occurs only in the presence of cargo, and we vary cargo-cargo εCC and cargo-shell εSC interaction strengths. Throughout this article, all energy values are given in units of the thermal energy, kBT. Except where mentioned otherwise, values of our simulation shell bending modulus κs fall within the range estimated for β−carboxysomes from AFM nanoindention experiments κs ∈ [1, 25]kBT (see Ref. [89] and section ‘Determination of parameter values’ in S2 Text; simulations with shell spontaneous curvature use κs = 10 − 16kBT.
We now consider the opposite limit: a system of ‘flat’ hexamer subunits, which have zero spontaneous curvature and thus favor formation of flat sheets (Fig 5A). Fig 5B shows a typical assembly trajectory for flat subunits with εCC = 1.7, in which the cargo rapidly coalesces followed by adsorption and assembly of the hexamers. Interestingly, the shapes of assembly intermediates reflect the lack hexamer spontaneous curvature—hexamers initially assemble into flat sheet wrapped around the globule, deforming the spherical globule into a cigar shape. Eventually the two sides of the sheet meet, creating a seam with an unfavorable line tension due to unsatisfied subunit contacts. As the seam gradually fills in, the elastic energy associated with such an acute deformation forces the complex toward a more spherical shape. As in systems with spontaneous curvature, the hexamer shells exhibit the 12 five-fold vacancy defects required by topology. If pentamers are present they eventually fill these holes (as in Fig 2 above), but for simplicity we consider systems containing only hexamers here. The large shells are roughly but not perfectly icosahedral, presumably reflecting slow defect reorganization on assembly timescales.
The size of the assembled shell is limited by the finite system size of our simulations. Importantly, the same limitation occurs within cells when the cargo undergoes phase separation into a single complex whose size is limited by the enzyme copy number (e.g. the procarboxysome precursor to carboxysome assembly [39, 40]). We therefore investigated the dependence of assembly morphologies on system size, as a function of the shell bending modulus, κs (controlled by the parameter εangle). Specifically, at each value of κs we performed a series of simulations in which the maximum size of the cargo globule was controlled by changing the system size with fixed total cargo concentration and hexamer chemical potential (section Materials and methods). An example assembly trajectory for a small system is shown in Fig 5C.
As shown in Fig 6, we observe a minimum globule size required for complete shell assembly, which linearly increases with κs. We observe complete wrapping for all system sizes above this threshold. Below the threshold size, assembly stalls with one or more open seams remaining; examples of this configuration are shown for a low and high bending modulus in Fig 6. Interestingly, while the pentameric defects are roughly equally spaced within large shells, small shells assembled with extremely low values of κs tend to exhibit adjacent vacancy pairs (Fig 5C, final frame). This defect morphology focuses curvature in a region with no elastic energy (the vacancy) while reducing the number of unsatisfied hexamer edges.
To understand these results, in section S2 Text we present a calculation of the equilibrium shell size distribution for subunits with no spontaneous curvature and stoichiometrically limiting cargo. We restrict the ensemble to spherical shells as observed in the simulations. While the aggregates are large and polydisperse without cargo, the calculation shows that cargo leads to a minimum free energy spherical shell size (S8 and S9 Figs).
The linear relationship between minimum shell size and bending modulus can be understood from our equilibrium model by comparing the excess free energy difference ΔΩwrap between the complete shell and an unwrapped globule (see section S2 Text). For the simulated conditions, the size and shape of the cargo globule is essentially the same in each of these states, and thus the free energy difference for a globule wrapped by nh hexamers in Eq. (S2.9) simplifies to
Δ Ω wrap = 8 π κ s + Δ G p + Δ μ h n h (1)
with Δμh = ghh + ghc − μh, ΔGp as the free energy due to the 12 pentameric vacancies, ghh(εHH) as the hexamer-hexamer interaction free energy, ghc(εSC) as the hexamer-cargo free energy, and μh = kBT log(ρh) the chemical potential of unassembled hexamers at concentration ρh. The term 8πκs describes the bending energy of the complete shell. The minimum globule size n* corresponds to the locus of parameter values at which ΔΩwrap = 0, giving
n * = 8 π - Δ μ h κ s + Δ G p - Δ μ h (2)
A linear fit to the simulation results for n* results in Δμh = −2.4 and ΔGp = 80.5kBT, or 6.7kBT per pentameric defect. Plugging in ρh = 10−3 subunits/σ 0 3 and ghc = −8.1kBT for εSC = 7.0 (using the estimate from Perlmutter et al. [42]) then results in ghh ≈ −0.45kBT. This value and the fit value of ΔGp are reasonably close to interactions estimated from the relationship between the shell-shell dimerization free energy ghh and potential well-depth εHH for a similar model in Perlmutter et al. [42]. Thus, the simulation results are consistent with the minimum stable shell size predicted by the theory.
We have used computational and theoretical modeling to investigate factors that control the assembly of a protein shell around a fluid cargo. We have focused on two limiting regimes of protein interaction geometries—high spontaneous curvature that drives the formation of small shells, and zero spontaneous curvature that favors assembly of flat sheets or polydisperse shells. In both regimes the presence of cargo can significantly alter the size distribution of assembled shells. For high spontaneous curvature, encapsulated cargo tends to increase shell size, whereas for shell proteins with low (or zero) spontaneous curvature cargo templating provides a mechanism to drive shell curvature and thus tends to reduce shell size. These results could provide a qualitative explanation for experimental observations on different systems in which full microcompartment shells were either larger or smaller than empty shells [23, 28, 31, 37, 38].
Our simulations identify a combination of kinetic and thermodynamic mechanisms governing microcompartment size control. At equilibrium, the shell size is determined by the stoichiometry between cargo and shell subunits, with an excess of cargo or shell protein respectively favoring larger or smaller shells. Similarly, a high surface energy (high cargo surface tension and weak shell-cargo interactions) favors larger shells whereas a strong shell bending modulus favors shells closer to the preferred size. Although dynamical simulations exhibit similar qualitative trends to these equilibrium results, we observe significant kinetic effects as well. Fast cargo coalescence relative to rates of shell assembly favors larger shells, since closure of an assembling shell prevents further cargo aggregation. Thus, the shell size is strongly correlated to the assembly pathway, with two-step assembly leading to larger shells than single-step pathways. Although many factors likely control shell size in biological systems, this result is consistent with the observations of small empty shell assemblies [23, 28, 31, 37] and the fact that β-carboxysomes (which assemble by two step pathways [39, 40]) tend to be larger and more polydisperse than α-carboxysomes (which experiments suggest assemble by one-step pathways [33, 41]).
Our results for shell proteins without spontaneous curvature build upon Rotskoff and Geissler [58], which identified a kinetic mechanism in which cargo templating drives shell curvature, and shell closure eventually arrests assembly. Their mechanism proceeds by two-step assembly, with initial nucleation of a cargo globule followed by assembly of shell subunits, but requires that rates of subunit arrival are at least 10 times faster than cargo arrival rates [58]. However, it is unclear how many physical microcompartment systems may fit this criteria, and our results suggest other mechanisms may play important roles in microcompartment assembly. Firstly, if cargo is stoichiometrically limiting then the finite-pool mechanism can result in finite shell sizes, with the coalesced cargo still providing a template for shell curvature. Secondly, subunits with spontaneous curvature can form complete shells even under conditions of excess cargo or fast coalescence rates that lead to large cargo aggregates (Fig 3D), as observed for carboxysome assembly in cells [39]. Thus, biological microcompartments with some degree of preferred shell curvature could robustly assemble over a much wider parameter space than systems without spontaneous curvature. Intriguingly, the recent atomic-resolution microcompartment structure from Sutter et al. [31] suggests that different hexamer or pseudo-hexamer species have different preferred subunit-subunit angles, and thus the spontaneous curvature may depend on the shell composition. We will investigate this in a future work.
The importance of spontaneous curvature to a particular BMC system could be investigated by comparing our computational predictions to experimental shell size distributions measured for varying cargo/shell protein stoichiometries and interaction strengths. While such tests would be most straightforward to perform in vitro, they could be performed in vivo by varying expression levels of various shell proteins or the enzymatic cargoes. Of particular interest would be a comparison between the shell size distribution in the presence and absence of cargo. However, note that we have focused on extreme limits (high spontaneous curvature or zero spontaneous curvature); systems with moderate shell spontaneous curvature may exhibit less dramatic cargo effects. Also note that the effective shell spontaneous curvature depends on the stoichiometries of different shell protein species; e.g., overexpressing pentamers would shift the size distribution toward smaller shells (Fig 2D).
These results have implications for targeting new core enzymes to BMC interiors. Recent experiments have shown that alternative cargoes can be targeted to BMC interiors by incorporating encapsulation peptides that mediate cargo-shell interactions, but that relatively small amounts of cargo were packaged [21–23, 96]. Our previous simulations showed that assembly of full shells requires both cargo-shell and cargo-cargo (direct or mediated) interactions. Here, we see that the strength of cargo-cargo interactions can not only affect the efficiency of cargo loading, but also the size of the containing shell.
|
10.1371/journal.pntd.0005780 | Characterizing the malaria rural-to-urban transmission interface: The importance of reactive case detection | Reported urban malaria cases are increasing in Latin America, however, evidence of such trend remains insufficient. Here, we propose an integrated approach that allows characterizing malaria transmission at the rural-to-urban interface by combining epidemiological, entomological, and parasite genotyping methods.
A descriptive study that combines active (ACD), passive (PCD), and reactive (RCD) case detection was performed in urban and peri-urban neighborhoods of Quibdó, Colombia. Heads of households were interviewed and epidemiological surveys were conducted to assess malaria prevalence and identify potential risk factors. Sixteen primary cases, eight by ACD and eight by PCD were recruited for RCD. Using the RCD strategy, prevalence of 1% by microscopy (6/604) and 9% by quantitative polymerase chain reaction (qPCR) (52/604) were found. A total of 73 houses and 289 volunteers were screened leading to 41 secondary cases, all of them in peri-urban settings (14% prevalence). Most secondary cases were genetically distinct from primary cases indicating that there were independent occurrences. Plasmodium vivax was the predominant species (76.3%, 71/93), most of them being asymptomatic (46/71). Urban and peri-urban neighborhoods had significant sociodemographic differences. Twenty-four potential breeding sites were identified, all in peri-urban areas. The predominant vectors for 1,305 adults were Anopheles nuneztovari (56,2%) and An. Darlingi (42,5%). One An. nuneztovari specimen was confirmed naturally infected with P. falciparum by ELISA.
This study found no evidence supporting the existence of urban malaria transmission in Quibdó. RCD strategy was more efficient for identifying malaria cases than ACD alone in areas where malaria transmission is variable and unstable. Incorporating parasite genotyping allows discovering hidden patterns of malaria transmission that cannot be detected otherwise. We propose to use the term “focal case” for those primary cases that lead to discovery of secondary but genetically unrelated malaria cases indicating undetected malaria transmission.
| Malaria is a disease of rural areas in developing countries. Although a rise in urban malaria cases has been noted during the last decade, this trend could be an artifact due to lack of solid data. In order to better understand “urban” and “peri-urban” malaria, we developed a rigorous and systematic methodology that allows characterizing malaria risk in such settings. Our approach is based on cross-sectional studies using active and reactive case detection strategies, genotyping of parasite isolates in order to better understand transmission patterns, and the local assessment of the entomological factors that allow active transmission in urban and peri-urban neighborhoods. This approach was tested in Quibdó, Colombia. No evidence of malaria transmission in urban areas was found. However, we found solid evidence indicating transmission in peri-urban areas due to Plasmodium vivax (86%). This was supported by the identification of Anopheles mosquitoes and their breeding places. Our results show that reactive case detection is not only an effective strategy to identify cases in areas where transmission is variable and unstable, but also allows the detection of hidden transmission when combined with genotyping methods. Such patterns are undetected by traditional surveillance methods.
| Malaria remains a major public health problem that affects 106 countries worldwide mostly in tropical and subtropical regions where ~3.4 billion people are at risk of infection and death [1]. Although malaria is mainly transmitted in rural areas where there are suitable environments for Anopheles mosquitoes breeding sites, malaria transmission in urban areas of endemic countries has been increasingly reported over the last three decades [2,3]. Unfortunately, the factors driving urban and peri-urban malaria transmission remain poorly characterized. As urban malaria cases are likely to be found at a broader range of primary care/diagnostic facilities, including hospitals and private laboratories failing to report them to the central surveillance system [4–7], urban malaria control by National Malaria Control Programs (NMCP) requires important administrative changes. Furthermore, there are no clear definitions of “urban”, “peri-urban”, and “rural” settings that properly describe the socioeconomic and ecological contexts where malaria transmission occurs. Thus, there is a need for a rigorous and systematic approach to characterize malaria transmission in the rural-to-urban interface that could provide solid information to assess disease risk in such contexts.
Like other epidemiological settings, urban malaria transmission is influenced by population movements from rural to urban and peri-urban areas. This rural population influx into urban and peri-urban areas facilitates the introduction of malaria from places where the disease is of high prevalence such as those where illegal mining and logging are common[8]. Furthermore, these underserved populations practice subsistence farming and inhabit poor housing with limited access to health services; such social dynamics favor mosquito breeding in areas considered administratively urban [9]. Here, we propose an integrated approach that aims to characterize epidemiologic and entomologic drivers of “urban” and “peri-urban” malarias in settings that are commonly found in Latin America. Our approach was tested in an endemic area of Colombia.
Despite the fact that malaria prevalence is decreasing in Colombia with a 75% reduction in the number of cases since 2000 [1], the National Surveillance System (SIVIGILA) reported an accelerated increase in urban malaria cases from 5.9% in 2011 [9] to 30% in 2015 [10]. Although this increase may be explained by population displacement due to political unrest and illegal crops and mining, there is still the possibility of autochthonous urban transmission. Because only a few studies have focused on urban malaria transmission in Colombia, the growing number of reports on urban malaria cases generates concerns and demands to unequivocally confirm the extent of urban and peri-urban transmission and to establish the corresponding control strategies. Thus, our integrated approach was used to study patterns of malaria transmission in five neighborhoods of Quibdó, the capital of the department of Choco (Colombia)[9,11]. These areas report the greatest number of so called “urban” cases providing an ideal setting for this investigation.
The study protocol was reviewed and approved by the institutional review board of Caucaseco Scientific Research Center (CECIV, Cali-Colombia) before initiation. Written informed consent (IC) was obtained from each volunteer at enrolment. Parents or legal guardians were asked to consent for children (<18-year-old) to participate in the study, and children older than seven years were asked to sign an informed assent if they wanted to participate. Information obtained from the participants was managed on principles of confidentiality. Immediately after blood sample processing, malaria-positive volunteers were informed and assessed during administration of appropriate anti-malarial treatment at the corresponding point of care. Asymptomatic volunteers did not receive treatment in concordance with the Colombian Ministry of Health (MOH) malaria treatment guidelines.
This study was conducted in Quibdó, which is currently the municipality of Colombia with the highest reported number of malaria cases [10,12]. It is in the Department of Chocó, in the northern area of the Pacific coast in the border with Panama. It has an area of 3,337 km2 between the jungle of Darien and the Atrato and San Juan river basins[11]. It consists mostly of a dense tropical rain forest with warm weather (average temperature of 28°C), relative humidity of 90% and an annual rainfall of 8,000–6,000 mm. It has an estimated total population of ~500,000 habitants (2015), with a geographical dispersion 5.4 times higher than the rest of the country [13].
Five sentinel sites (SS) were selected based on location, urbanization and history of malaria cases: La Yesquita, Silencio and Roma which are neighborhoods with urban characteristics located in the center of the city have paved streets, public services and no vegetation close to the houses. On the other hand, Casa Blanca and Cabí are classified as peri-urban, located in the North and South ends of the municipality, respectively, with unpaved streets, variable housing infrastructure, lack of a sewage system and abundant vegetation.
We have considered an urban area as some groups of buildings and contiguous structures grouped in blocks, which are delimited by streets or avenues, with a number of essential services such as aqueducts, sewage, electrical energy, and hospitals and schools. Capital cities and the remaining municipal administrative headings are urban areas. In contrast, a rural area is characterized by the dispersed disposition of houses and agricultural holdings, and a lack of road structure and public services. A peri-urban area is one that combines characteristics both urban and rural, usually located in areas outside the city (DANE (2005) [14]
A total of 1mL of blood was collected by venipuncture from every subject, of which ~50 μL were used for thick blood smear (TBS) and the remainder was stored in tubes containing EDTA, refrigerated at 4°C and transported by airplane to the laboratory in Cali for later qPCR analysis and microsatellite (STRs) genotyping. Samples were handled as potential biohazards and all laboratory staff strictly followed bio-safety standardized procedures.
We characterized the urban-to-rural malaria interface by integrating epidemiological and entomological approaches. To determine the prevalence of malaria, we first identified eight malaria positive volunteers among individuals seeking diagnosis at the Ismael Roldán hospital in Quibdó. These volunteers reported to live in urban or peri-urban sites of Quibdó and were further selected as sentinel sites (SS) for active case detection (ACD) and reactive case detection (RCD). Then, a cross-sectional survey was performed in five SS from Quibdó, three of them considered urban neighborhoods and two peri-urban. A visit to their neighbourhoods of origin was performed to classify them as urban (2/8) or peri-urban (6/8). For RCD, four houses, closest to the primary case were selected in each neighborhood, using the Vector Born Diseases Program (ETV) census. Eight primary cases identified by ACD were studied with the RCD strategy as well. Men, women and children above one year of age who lived in the household and were present at the moment of the visit were enrolled. A Household was defined as the place where people enrolled in the study lived, including family members, servants, tenants, and others.
Those who agreed to participate answered a symptoms survey and donated a blood sample. Epidemiological questionnaires were answered by the head of the household. A malarial household was defined as one with at least one infected person.
The sample size (n) was calculated for each SS using an estimated prevalence (P) of 2.6% with a confidence level of 95%, 2.5% error (d) according to the following equation n0 = (zα2), where α = 1.96; P (1—P)/d2. Then, it was adjusted according to the population of each neighborhood (N) considering the equation n = n0/(1+(n0-1)/N)[15].
Collection of adult specimens: Mosquitoes collection was performed using Human Landing Catches (HLC) [22] from 18:00 hours to 6:00 hours in the households of infected volunteers diagnosed by RCD. Collections were carried out simultaneously indoors and outdoors for each house for two consecutive nights. Data on relative humidity and temperature were recorded. Mosquitoes were kept in cups labeled with the date, neighborhood, house code, capture time, mosquitoes quantity and collector's name. Specimens were sacrificed with tri-ethylamine and subsequently individually packaged in 1,5mL vials with a perforated lid, and conserved in airtight bags with silica gel. Technicians in charge of mosquito catches signed an informed consent prior HLC. Adult and immature mosquito specimens were determined using dichotonous keys for Anopheles of Colombia [23].
Detection of natural infection with Plasmodium spp: After taxonomic identification, mosquitoes’ head and thorax belonging to the same species and capture hour were pooled. Samples were macerated following the MR4 protocol and insert specifications. Circumsporozoite protein (CS) from P. falciparum, P. vivax VK-210 and VK-247 variants were detected by Enzyme-Linked ImmunoSorbent Assay (ELISA) using the kit distributed by the Center for Disease Control and Prevention (CDC, Atlanta, USA) [24,25]
Immature collection: We searched SS for open water bodies. The larval habitats search was performed around the households, within a 500-meters radius. Each potential breeding site was georeferenced and characteristics such as size, vegetation type, water type, water body type and water use were recorded. Sampling was carried out using a standard dipper (350 mL) taking ten dips per square meter. Larvae were stored in vials with ethanol for preservation. Each larval container was labeled with date, code, larval number, neighborhood and collector's name.
Data were recorded in REDCap (v.6.9.4) web application and analyzed using the software R for statistical analysis (v3.3.0) for variables like age, gender, sociodemographic features, living conditions and malaria positive cases, by site and as a total. Non-parametric tests and Chi-square and Fisher’s exact test were performed to check for differences between categories and calculate associations. A level of statistical significance of 5% was used and 95% confidence intervals were calculated for proportions.
The limitations of the study were: A memory bias occurred while conducting the surveys, as for some volunteers it was very difficult to remember exactly the activities carried out weeks before the survey. Volunteers’ displacement was questioned only for the last month. Follow up of some infected cases after diagnosis was difficult because some of them had to leave the city due to their informal jobs”.
A total of 717 volunteers (60.0% female) were surveyed using ACD detection in 135 households visited, 58.5% of them located in the peri-urban areas and the rest of them in urban areas. Significant differences were found in sociodemographic variables between urban and peri-urban neighborhoods (Table 1).
Median ages were significantly higher in urban neighborhoods. Overall the Afro-American ethnic group was predominant (77.7%), indigenous population was significantly higher in both areas peri-urban areas (21.4%); especially in Cabí, where 91.6% of the population share this ethnic origin. Education level was significantly lower in peri-urban areas, with 25.1% of the population being illiterate. Most frequent occupations were merchant in urban areas and housewife in peri-urban areas.
Housing conditions were also significantly different (Table 1). In the urban neighborhoods, most houses were made of brick, 66.1% had aqueducts and the majority of them had access to electricity (100%), garbage collection system (89.3%) and sewage system (67.9%). In the peri-urban neighborhoods, the predominant housing material was wood (59.7%), and none of the houses had an aqueduct service or sewage system. Water supply was obtained from rain in 74 of 79 houses while the remaining obtained water directly from the river or had a well. Only three houses in Casa Blanca and one in Cabí had a garbage collection system.
Eight primary cases were recruited in a hospital in Quibdó where they attended to seek diagnosis. Four of them were caused by P. vivax, three by P. falciparum and one was a mixed infection. Two cases came from urban sites and the other six from peri-urban neighborhoods. Most of the cases were in Afro-descendants, with complete to incomplete secondary education and with informal jobs. Five of the eight cases were men.
The overall prevalence of malaria by ACD was 1% (6/604) using microscopy and 9% (52/604) by qPCR. A total of 44 cases (85%) of the 52 detected by qPCR were due to P. vivax being the predominant species. Ninety-six percent of the detected infections were in peri-urban areas, presenting a significant difference with those originated in urban neighborhoods (p<0.0001). Cabí was the neighborhood with the highest prevalence (29% ±8 SE), followed by Casa Blanca (9% ±5 SE qPCR). In the urban neighborhoods, only two submicroscopic infections were diagnosed (Fig 1).
Eight primary asymptomatic infections from ACD were selected for RCD, six for P. vivax and two for P. falciparum. During the RCD, 33 houses were studied, 18 in Cabí, 10 in Casa Blanca and five in La Yesquita, and a total of 113 volunteers were surveyed. Twenty-seven secondary cases (three by TBS and 27 by qPCR) were identified, all of them from peri-urban areas. The overall percentage of positives among the screened people was 2.7% by microscopy and 24% by qPCR. A total of 13 malarial houses were found.
In addition, a second round of RCD was performed around the eight symptomatic cases recruited by PCD. A total of 175 volunteers among family members and neighbors were included. Fourteen secondary infections were detected (eight by TBS and 14 by qPCR), representing an infection rate of 4.6% by microscopy and 8% by PCR. Twelve cases were caused by P. falciparum and the other two by P. vivax.
Twelve of the 14 secondary cases were female, most of them Afro-descendant, four housewives, three students and one with another type of job. Seven were children under eleven years, 7/14 were older than 17 years, one of them 47 years old. One 10-year-old child was an asymptomatic positive case by PCR from an urban site, he did not have recent history of recent displacement outside Quibdó but had moved to other peri-urban neighborhood close to his home. A total of 10 malarial houses were detected in this survey.
In total, using the RCD strategy we found 41 malaria cases, 27 by P. vivax (66%) and 14 P. falciparum (34%) and 22 malarial houses. No mixed infections were detected.
A high number of asymptomatic volunteers were identified, i.e. 58 of the 93 cases diagnosed by qPCR did not report any symptoms at the time of blood examination and none reported to have had malaria symptoms for the last 15 days. Asymptomatic infection was more frequent with P. vivax (46/71). Of the nine volunteers with infection diagnosed by microscopy, seven presented with symptoms. Thirty-five cases showed symptoms; the most common symptoms were fever (25/35), headache (25/35), chills (11/35), muscles pain (9/35), malaise (6/35), and profuse sweating (3/35).
Out of 8 primary cases (6 P. vivax and 2 P. falciparum), 27 secondary cases were detected in Cabí: 21 multiple infections (>1 allele in at least 1 microsatellite locus) and only 6 single infections for the set of microsatellite loci used. These multiple infections included 11 (52.4%) with 2 alleles in at least 1 locus, 6 (28.6%) with at least 2 alleles in 2 loci, and 4 (19%) with > 2 alleles in 3 or more loci (Fig 2)[26].
None of the genotypes found in the primary cases matched those found in the secondary cases. Furthermore, two primary cases were P. falciparum but all the detected secondary cases were P. vivax. This pattern indicates that the secondary cases were not related to the primary cases that allowed their detection. Nevertheless, some genotypes were shared among the secondary cases. Specifically, related genotypes were found in two patients that inhabited the same house (4), as well as patients that inhabited houses that were near (4 and 7 separated by ~43 mts, 4 and 5 separated by ~13 mts, and 204 and 207 separated by ~18 mts) (Fig 2).
Entomological studies were performed in 15 houses where at least one malaria case was detected. These houses were distributed in three SS, one urban (La Yesquita) and two peri-urban (Casa Blanca and Cabí). Neither mosquitoes nor breeding sites were found in the urban setting. On the contrary, both peri-urban areas registered the presence of immature and adult Anopheles mosquitoes.
Natural infection by P. falciparum and P. vivax (VK-210 and VK-247 variants) was analyzed in 1,305 adult Anopheles mosquitoes; An. nuneztovari (n = 734), An. darlingi (n = 555), An. triannulatus (n = 15) and An. apicimacula (n = 1). One An. nuneztovari obtained in Cabí was found infected with P. falciparum (1/590), corresponding to an infection rate of 0.17% in this locality. This infected mosquito was captured biting indoors between 22:00h and 23:00h in the Urada indigenous community.
A total of 24 open water bodies were examined and georeferenced. From these, four were positive for Anopheles mosquito larvae in Casa Blanca and six in Cabí. Of the ten positive larval habitats, seven were excavation site type, two were stream and one was a puddle. A total of 100 larvae were collected in the breeding sites from these 40 late (3rd and 4th) instar larvae were identified belonging to the two most abundant species. An. nuneztovari was found in excavation sites and An. darlingi in streams and puddles (Table 2).
Our integrated approach shows a complex pattern of malaria transmission in the rural-to-urban interface in this Colombian community. Although all the selected SS were administratively classified as urban neighborhoods, they showed significant differences in terms of sociodemographic characteristics, prevalence of Plasmodium species, and distribution of the different vector species. Based on the criteria developed here, we can unequivocally classify three SS (La Yesquita, Roma and Silencio) as urban. However, the others (Cabí, Casa Blanca) should be considered peri-urban/rural settings. This highlights the need to reach a consensus for the administrative classification of the territory. Social and ecological characteristics of each studied neighborhood would be most relevant for an accurate classification as well as in terms of designing malaria control programs.
Our results also show that malaria continues to be a disease that mainly affects vulnerable communities. Indeed, 97.5% of the cases were diagnosed in peri-urban areas inhabited by young and less educated people living in precarious conditions. The significant greater prevalence of malaria among children could be partially explained by the age structure of the population [1]. Lower education levels have also been associated with higher malaria risk [27]. In terms of occupations, 78.6% of the adult volunteers were housewives, unemployed population, students and merchants. This finding supports the hypothesis that infection is occurring mostly in homes rather than work places, reducing the possibilities of bias related to occupational risks.
As expected, the number of cases detected by PCR (93 cases) was significantly higher (5.6 times) than that detected by microscopy (16 cases), strengthening the need to re-evaluate the diagnosis methods used in this type of epidemiological settings. The confirmed presence of submicroscopic infections represents an important public health problem, as the unidentified positive cases will not receive treatment, and will keep contributing to transmission maintenance [28]. Therefore, it is of paramount importance to accurately determine parasite prevalence for monitoring malaria interventions.
P. vivax was the predominant species and most cases were diagnosed in Cabí, where the population was predominantly indigenous. This could be explained by their Duffy positive (Fy+) genotype, which makes this population susceptible to P. vivax infection [29] as compared to the Afro-descendant population which in this region display a Fy- prevalence of 38.9% [30].
This investigation also shows that RCD combined with parasite genotyping allows a better assessment of the transmission patterns which are far more complex than the ones inferred from epidemiologic data alone [31,32]. RCD usually involves detection of a primary case that is further referred to as “index case”. This term is commonly used in the context of infectious diseases to denote a first detected case that allows the identification of secondary cases that are usually part of the same transmission tree. Loosely used, however, a primary case can simply indicate an environment that can sustain transmission where secondary cases are spatially clustered. Thus, in such context, the epidemiological evidence does not make a distinction between a cluster of cases that is the result of a recent introduction or simply ongoing transmission that has not been detected [32,33]. Considering that RCD is laborious in areas of low transmission [32], the incorporation of parasite genotyping is essential to increase the information yield by RCD. It allows separating between recently introduced cases [34,35] and hidden ongoing transmission that simply remains elusive to the control program. Such distinction is of importance during the elimination phase or wherever control measures should be deployed to control malaria in urban or peri-urban settings. This investigation indicates that, in the context of these study sites, the primary case detected areas with ongoing transmission, but secondary cases were not part of the same transmission tree as expected under a scenario of recent introduction. Indeed, two primary cases were caused by P. falciparum and all the secondary were P. vivax (Fig 2). This indicates that the social and environmental conditions in the study areas can sustain asymptomatic-submicroscopic malaria patients that could allow for the re-emergence of malaria whenever control strategies are relaxed or environmental conditions change [36].
Based on the patterns detected in this investigation, we propose to reserve the use of the term index case for those primary cases that allow the detection of secondary cases that are part of the same genetically related cluster indicating a single (or few) reintroductions. Accordingly, we propose using the term ‘focal case’ to denote a primary case that allows for the identification of hidden (ongoing) malaria transmission as evidenced by a cluster of unrelated secondary cases. At least in the context of P. vivax, we could also use the proportion of multiple infections as a metric for ongoing transmission since those indicate superinfections [37]. In this context, patients with complex genotypes (multiple alleles in >2 loci, see [26]) may indicate two or more infectious bites providing additional evidence of stable undetectable transmission.
Previous studies in Quibdó have shown that there are malaria transmission hot spots, which are located in peri-urban areas, and that neighborhoods such as Cabi in the South and Casa Blanca in the North are areas with a high incidence and prevalence of malaria, as confirmed in the present study. It would be important to follow these transmission hotspots with active surveillance and RCD, including molecular diagnostic methods to properly assess the effect of interventions on malaria transmission. These interventions should be evaluated to confirm its efficacy in terms of reducing morbidity and mortality associated with malaria transmission. The importance of this hidden malaria transmission is evidence by the fact that Chocó significantly contributed to the malaria rebound that occurred in Colombia during 2016, with an expansion of 62% in the number of cases, of those, 16% were reported from Quibdó [10,38].
Finally, a worldwide effort to understand urban malaria has been undertaken by the International Centers of Excellence for Malaria Research-ICEMRs. In a recent publication authored by this group, one of the pillars to this understanding includes the definition of what constitutes “urban” and “peri-urban” malaria. There is no global consensus definition of urban malaria and most countries use an administrative definition [4]. In Colombia, it would be important to redefine the concept of "administrative urban municipality", which includes urban and peri-urban areas in terms of the charts used by the notification system (SIVIGILA), to classify reported cases correctly.
No evidence for urban malaria transmission was found in Quibdó. The cases found in urban areas were imported from other cities or peri-urban neighborhoods with high prevalence. Thus, malaria transmission is mainly peri-urban, and autochthonous transmission occurs mainly in indigenous communities. The implementation of RCD with molecular diagnostics and genotyping, allow the detection of hidden malaria transmission clusters. This approach is suitable to better understand the efficacy of the malaria control programs interventions.
|
10.1371/journal.ppat.1000686 | Innate Immune Recognition of Yersinia pseudotuberculosis Type III Secretion | Specialized protein translocation systems are used by many bacterial pathogens to deliver effector proteins into host cells that interfere with normal cellular functions. How the host immune system recognizes and responds to this intrusive event is not understood. To address these questions, we determined the mammalian cellular response to the virulence-associated type III secretion system (T3SS) of the human pathogen Yersinia pseudotuberculosis. We found that macrophages devoid of Toll-like receptor (TLR) signaling regulate expression of 266 genes following recognition of the Y. pseudotuberculosis T3SS. This analysis revealed two temporally distinct responses that could be separated into activation of NFκB- and type I IFN-regulated genes. Extracellular bacteria were capable of triggering these signaling events, as inhibition of bacterial uptake had no effect on the ensuing innate immune response. The cytosolic peptidoglycan sensors Nod1 and Nod2 and the inflammasome component caspase-1 were not involved in NFκB activation following recognition of the Y. pseudotuberculosis T3SS. However, caspase-1 was required for secretion of the inflammatory cytokine IL-1β in response to T3SS-positive Y. pseudotuberculosis. In order to characterize the bacterial requirements for induction of this novel TLR-, Nod1/2-, and caspase-1-independent response, we used Y. pseudotuberculosis strains lacking specific components of the T3SS. Formation of a functional T3SS pore was required, as bacteria expressing a secretion needle, but lacking the pore-forming proteins YopB or YopD, did not trigger these signaling events. However, nonspecific membrane disruption could not recapitulate the NFκB signaling triggered by Y. pseudotuberculosis expressing a functional T3SS pore. Although host cell recognition of the T3SS did not require known translocated substrates, the ensuing response could be modulated by effectors such as YopJ and YopT, as YopT amplified the response, while YopJ dampened it. Collectively, these data suggest that combined recognition of the T3SS pore and YopBD-mediated delivery of immune activating ligands into the host cytosol informs the host cell of pathogenic challenge. This leads to a unique, multifactorial response distinct from the canonical immune response to a bacterium lacking a T3SS.
| Most multicellular organisms have immune sensors that recognize molecules common among microorganisms. Recognition of such molecules informs the host that invading microbes are present, triggering an immune response. Many known innate immune sensors, however, do not appear to distinguish commensals from pathogens. This is in spite of the fact that the host must clear pathogens while simultaneously avoiding a response to benign or beneficial microbes. There are few molecular explanations for how this discrimination occurs in mammalian hosts. To address this problem, we analyzed the response of mammalian cells to the gut pathogen Yersinia pseudotuberculosis. We found that Yersinia expressing a virulence-associated secretion system caused a transcriptional response in host cells that was very different from the response to a strain with a nonfunctional version of the secretion system. This transcriptional response included several distinct signaling pathways leading to production of mediators of innate immunity, including cytokines such as type I interferon and TNF-α. A large number of pathogens express specialized secretion systems similar to that in Yersinia, so these findings provide evidence that there is a mammalian immune response to alterations in host cells that results from pathogen attack, supporting known systems for recognition of common microbial molecules.
| The ability to detect pathogens while maintaining beneficial commensal bacteria is important for the health of animal hosts [1]. In order to defend itself against potentially injurious intruders, the host must recognize and respond to previously unencountered microbes. This is achieved in part by recognition of molecules unique to microbes but common among subgroups of bacteria, viruses, or fungi [2],[3]. Detection is accomplished partially by a group of 13 mammalian surface-associated proteins called Toll-like receptors (TLRs) which induce an immune response upon recognition of microbial molecules such as lipopolysaccharide and flagellin. Beneficial commensal bacteria contain at least some TLR ligands [1], yet aberrant TLR stimulation by commensals may cause intestinal inflammation [4]. Therefore, mechanisms must be in place to allow the host to distinguish between pathogenic and commensal bacteria. Compartmentalization of TLRs facilitates this process. For example, TLR5, which recognizes flagellin, is only expressed on the basolateral side of intestinal epithelial cells, from which commensal bacteria are normally absent [5]. In contrast, pathogens possess virulence factors that allow them to penetrate into deeper tissues sites where receptors such as TLR5 can access them [1].
In addition to TLRs, there are innate immune sensors residing in the cytosol of mammalian cells that allow recognition of microbial molecules at this site or damage caused by pathogenic bacteria. One example is activation of the innate immune sensor Nod1 by the presence of bacterially-derived peptidoglycan [6]. For instance, Nod1 activation is triggered by the pathogen Shigella flexneri, which is thought to shed peptidoglycan during its residence in the host cytosol [7]. In addition, Helicobacter pylori may introduce peptidoglycan into host cells in a process dependent on a specialized secretion system, activating Nod1 [8]. Both of these microbial strategies, gaining entry into the host cytosol and utilizing a specialized secretion system, are thought to be more common among pathogens than commensals.
One such specialized secretion system found in a number of pathogenic bacteria is the type III secretion system (T3SS), which forms small pores in target host cells and delivers bacterial proteins into the host cytosol [9]. A common result of this “injection” is perturbation of normal host processes, to the benefit of the pathogen. One human pathogen that requires a T3SS for virulence is Yersinia pseudotuberculosis, which causes inflammation of the gastrointestinal tract exemplified by symptoms such as fever and swelling of gastric tissues and lymph nodes [10]. In healthy individuals, enteropathogenic yersiniosis is usually a self-limiting disease. However, in immunocompromised patients, the case fatality rate approaches 50% as a result of bacterial dissemination [11].
The Y. pseudotuberculosis T3SS is encoded on a virulence plasmid that is also found in the closely related human pathogens Y. enterocolitica and Y. pestis [12]. The Yersinia T3SS is composed of three protein subgroups: those that make up the injectisome, translocator Yops (Yersinia outer membrane proteins), and effector Yops. The injectisome is a needle-like structure that is evolutionarily related to the flagellar apparatus and has a central pore of about 20 Å [13],[14]. This needle apparatus is all that is required for secretion of the effector Yops, but is not sufficient for their translocation across the target cell plasma membrane. Targeting of effector Yops into the host cell cytosol requires the translocator proteins YopB, YopD, and LcrV, which are secreted through the T3SS apparatus and act to form channels in host cell membranes [15]. LcrV can be found associated with the tip of the needle apparatus [16] where it is thought to form a scaffold for the pore-forming proteins YopBD. T3SS effector Yops presumably travel through the type III needle and then through the pore made by YopBD in the host cell membrane. When the entire T3SS is functional, Yersinia translocate a group of five to six effector proteins into the host cytosol that interfere with target cell functions [17]. YopE, YopT, YopH, and YopO/YpkA target the host actin cytoskeleton, inhibiting phagocytosis and allowing the bacteria to remain largely extracellular. YopJ/YopP inhibits several inflammatory signaling pathways and influences the viability of a subset of host cells [18]–[20], while the function of YopM remains unknown [21].
The Yersinia T3SS pore, which forms during translocation of effector Yops, was recently suggested to trigger processing of the cytokines IL-1β and IL-18 in macrophages by the protease caspase-1 [22],[23]. Maturation of these cytokines has been linked to activation of a cytosolic innate immune complex called an inflammasome [24]. This type of complex is involved in detection of pore formation caused by a number of bacterial toxins [25]–[27]. Because other pathogens expressing specialized secretion systems, such as H. pylori, induce cytosolic innate immune sensors distinct from those found associated with inflammasomes [8],[28], we hypothesized that other host pathways may also be involved in recognizing the Y. pseudotuberculosis T3SS. Intriguingly, Y. pseudotuberculosis was previously reported to induce production of the chemokine IL-8 during infection of HeLa cells and this was dependent on expression of the Y. pseudotuberculosis T3SS [29]. Therefore, we tested the ability of several mammalian cell types to distinguish between Y. pseudotuberculosis expressing or lacking a functional T3SS. We show here that T3SS-competent Y. pseudotuberculosis triggers a TLR-independent transcriptional response that includes NFκB activation as well as induction of type I IFN-regulated genes, which are not known to be downstream of inflammasome activation. Furthermore, the NFκB pathway activated by the T3SS was independent of caspase-1-inflammasome activity. Recognition of T3SS-positive Y. pseudotuberculosis required bacterial expression of YopBD, but did not require any of the known effector Yops. Our results suggest that virulent Yersinia activate multiple cytosolic immune surveillance pathways as a consequence of utilizing a specialized secretion system.
To determine whether cytosolic innate immune sensors contribute to detection of Yersinia, an infection model of mammalian cells lacking TLR signaling was established. TLR activation generates strong innate immune responses, many of which overlap with those generated by cytosolic innate immune sensors [30]. In addition, Yersinia contain many TLR ligands. To circumvent this, we used bone marrow-derived macrophages (BMDMs) from mice that lack the two main TLR adaptor proteins, MyD88 and Trif, and hence lack TLR signaling (MyD88−/−/Trif−/−) [31]. The macrophage is a major in vivo target of Yersinia type III secretion [32], making it a good candidate for our experiments.
We initially examined the production of the inflammatory cytokine TNF-α because it is produced downstream of several cytosolic innate immune sensors [28], but not following inflammasome activation [33],[34]. MyD88−/−/Trif−/− BMDMs produced TNF-α mRNA in response to wildtype Y. pseudotuberculosis, but not to Y. pseudotuberculosis lacking a T3SS apparatus (ΔyscNU) (Table 1; Fig. 1A). This indicates that macrophages can distinguish Yersinia expressing a T3SS from one that lacks it.
To identify the component of the T3SS apparatus recognized by macrophages, we challenged MyD88−/−/Trif−/− BMDMs with Y. pseudotuberculosis Δyop6 lacking all six known T3SS effector proteins. TNF-α was induced even more robustly (Fig. 1A), indicating that these effector proteins are not required for the macrophage response to the Yersinia T3SS. In contrast, Y. pseudotuberculosis Δ6/ΔyopB lacking the T3SS translocator component YopB did not induce any TNF-α in MyD88−/−/Trif−/− BMDMs (Fig. 1A). In the absence of YopB, Yersinia cannot induce pore formation in target host cell membranes and cannot translocate any molecules into the host cell cytosol [35],[36]. However, the Y. pseudotuberculosis Δ6/ΔyopB strain does express a T3SS needle on its surface and can secrete the translocator component YopD under type III secretion-inducing conditions in vitro (Fig. S1A).
Since YopB was necessary for TLR-independent recognition of Yersinia, we reasoned that either the YopB protein itself was recognized by macrophages or that a YopB-mediated event, such as pore formation or translocation of an immune activating ligand into the host cytosol, was detected by macrophages. To distinguish between these possibilities, we assessed whether a strain of Y. pseudotuberculosis that was incapable of type III translocation, but still expressed YopB, could induce TLR-independent TNF-α production. A Y. pseudotuberculosis strain that carries a frameshift mutation in the translocator component YopD is unable to induce pore formation or effector protein translocation into target host cells [36], but can still secrete YopB into the culture supernatant (Fig. S1A). Similar to the strain lacking YopB, this YopD-deficient strain also did not induce TNF-α in the absence of TLR signaling (Fig. S1B). This indicates that rather than direct recognition of YopB, TLR-independent recognition of Yersinia expressing a functional T3SS involves detection of a YopBD-mediated event.
To test whether Y. pseudotuberculosis must be internalized in order for TLR-independent detection to occur, MyD88−/−/Trif−/− BMDMs were preincubated with the actin cytoskeletal depolymerizing agent cytochalasin D (cytD) to inhibit phagocytosis. This treatment allowed 99% of the Yersinia to remain extracellular (data not shown). However, cytD treatment did not inhibit the TNF-α mRNA response of MyD88−/−/Trif−/− BMDMs to Yersinia (Fig. 1B). This indicates that extracellular Yersinia expressing a functional T3SS can trigger TLR-independent innate immune signaling.
The T3SS effector protein YopJ has been shown to inhibit NFκB and MAP kinase signaling, leading to dampening of TLR-induced cytokine expression in cultured cells [18]. To determine if TLR-independent cytokine production is also inhibited by YopJ, we challenged MyD88−/−/Trif−/− BMDMs with Y. pseudotuberculosis ΔyopJ (Fig. 1A). This strain induced an elevated level of TNF-α mRNA relative to the wildtype control, indicating that YopJ interferes with TLR-independent TNF-α production. Interestingly, Y. pseudotuberculosis Δyop6 lacking all six known T3SS effector proteins (Δyop6) induced TNF-α mRNA to a level three times greater than the wildtype Y. pseudotuberculosis strain, but 1.5-fold less than the ΔyopJ Y. pseudotuberculosis strain (p = 0.009; Fig. 1A). Therefore, in the absence of YopJ, other translocated effector proteins enhance the TLR-independent response to the T3SS. In fact, the presence of a single effector Yop was sufficient to recapitulate this stimulation. Y. pseudotuberculosis Δ6/pYopT expressing only the effector protein YopT induced 1.7-fold more TNF-α mRNA (p = 0.04; Fig. 1C) and five-fold more TNF-α protein (p = 0.008; Fig. 1D) than the Y. pseudotuberculosis Δyop6 strain. Furthermore, Y. pseudotuberculosis Δ6/pYopTC139S expressing a catalytically inactive point mutant of YopT did not induce this enhanced TNF-α production (Fig. 1C–D). This indicates that while the effector protein YopJ partially inhibits TLR-independent TNF-α production triggered by Yersinia, the catalytic activity of YopT enhances TNF-α production.
The above results indicated that there may be a global response to the insertion of the Y. pseudotuberculosis T3SS into the host cell membrane. To identify the spectrum of host genes regulated in response to the Y. pseudotuberculosis T3SS, we determined the gene expression profile of MyD88−/−/Trif−/− BMDMs after challenge with Y. pseudotuberculosis Δyop6 expressing a functional T3SS translocator compared to Y. pseudotuberculosis Δ6/ΔyopB expressing a T3SS needle but defective in type III translocation (Materials and Methods). We used strains lacking the known T3SS effector proteins to exclude macrophage genes controlled by the catalytic activity of these effectors. Even in the absence of the T3SS effector proteins, we found a large number of macrophage genes significantly regulated following challenge with Y. pseudotuberculosis compared to the uninfected condition (Fig. S2).
The Y. pseudotuberculosis adhesin invasin has been reported to induce host cell signaling, consistent with the fact that it engages integrin receptors [37]. Because both the Y. pseudotuberculosis Δyop6 and Δ6/ΔyopB strains express invasin, integrin engagement may account for at least some of the alterations in host gene expression that are common to these two strains. Instead, we chose to focus on host genes that were differentially regulated in response to these two strains. Indeed, 266 genes were regulated by Y. pseudotuberculosis Δyop6 (compared to the uninfected condition) at least four-fold or more than by the Δ6/ΔyopB strain (Fig. 2; Datasets S1 and S2). These genes predominantly clustered into five groups based on their expression pattern (cluster I–V). Most of these genes were upregulated by the Y. pseudotuberculosis Δyop6 strain compared to the uninfected condition, although cluster IV contains downregulated genes. Other than TNF-α, a number of cytokines, chemokines, and costimulatory molecules were induced specifically in the presence of an intact T3SS translocator (Table 2). Furthermore, a number of genes involved in the NFκB signaling cascade were preferentially upregulated by Y. pseudotuberculosis Δyop6 compared to the Δ6/ΔyopB strain (Table 3). There was a similar pattern of expression for many of the genes controlled by this transcription factor, as transcription was induced by two hours post-inoculation with the Y. pseudotuberculosis Δyop6 strain and then subsequently tapered (Fig. 3A shows three examples). This was similar to the TNF-α expression pattern as determined by quantitative PCR analysis (qPCR; Fig. S3). Consistent with the panel of genes that were identified in this fashion, NFκB could indeed be preferentially activated by Y. pseudotuberculosis Δyop6 compared to Y. pseudotuberculosis Δ6/ΔyopB, based on monitoring infected 293T cells expressing a NFκB luciferase reporter (see below).
A subset of genes represented in the heatmap in Fig. 2 are known to be controlled by cytokines of the type I interferon (IFN) family and had a distinct expression pattern. These genes were not induced significantly until four hours post-inoculation with the Y. pseudotuberculosis Δyop6 strain (Fig. 3B shows three examples). This was consistent with the induction pattern of the cytokine IFN-inducible protein of 10 kDa (IP10) that we obtained using qPCR (Fig. S4A). In addition, most of the type I IFN-inducible genes identified by microarray analysis fell into cluster II, where expression was highest at four hours post-inoculation (Fig. 2; Table 4). Consistent with induction of type I IFN-regulated genes at later time points, mRNA levels of the gene encoding the type I IFN family member IFNβ were significantly upregulated (p<0.0001) over uninfected levels by Y. pseudotuberculosis Δyop6 (5.4-fold±0.9) compared to the Δ6/ΔyopB strain (1.2-fold±0.2) two hours post-inoculation (Fig. S4B). Similar to TNF-α, the observed induction of IFN-β by Y. pseudotuberculosis Δyop6 did not depend on the bacteria being internalized into macrophages because the level of IFNβ mRNA was not reduced in the presence of cytochalasin D (data not shown).
To validate some of our microarray results, we used qPCR to verify that a subset of the genes identified in our microarray analysis were indeed induced by Y. pseudotuberculosis Δyop6 independently of TLRs. We infected MyD88−/−/Trif−/− BMDMs with Y. pseudotuberculosis Δyop6 or Δ6/ΔyopB and analyzed mRNA levels of select genes two hours post-inoculation. We found that, similar to TNF-α, genes encoding the anti-inflammatory cytokine IL-10, the cysteine- and serine-rich nuclear protein Axud1, and the transcription factor early growth response protein 1 (Egr1) were transcribed by macrophages in response to Y. pseudotuberculosis Δyop6 but not to the Δ6/ΔyopB strain (Fig. 4; Fig. 5A). These results support those obtained from our microarray analysis.
To determine if the presence of TLR signaling alters the expression of genes induced in response to an intact T3SS translocator, wildtype C57Bl/6 macrophages were challenged with Y. pseudotuberculosis Δyop6 or Δ6/ΔyopB. In the presence of TLR signaling, macrophages transcribed the genes encoding TNF-α, IL-10, and Axud1 in response to both Y. pseudotuberculosis strains (Fig. 4). This indicates that TLR ligands, such as LPS, trigger macrophage transcription of these genes independently of the T3SS.
The egr1 and egr2 genes were induced preferentially by the T3SS translocator-positive Y. pseudotuberculosis Δyop6 strain in macrophages lacking TLR signaling (Fig. 5A and microarray data not shown). However, while egr1 was induced in wildtype macrophages in response to both Y. pseudotuberculosis Δyop6 and Δ6/ΔyopB, the mRNA levels were 11-fold lower than that induced by the Δyop6 strain in MyD88−/−/Trif−/− BMDMs. This indicates that TLR signaling may negatively regulate egr1 expression. In order to test this hypothesis, we tolerized wildtype macrophages to TLR ligands by incubating them overnight with a low level of heat-killed Y. pseudotuberculosis. This type of treatment has been used previously to dampen subsequent TLR responses [38]. We then challenged these tolerized macrophages with live Y. pseudotuberculosis Δyop6 and measured Egr1 mRNA levels. While untolerized wildtype macrophages produced only low levels of Egr1 mRNA, tolerized wildtype macrophages produced Erg1 mRNA levels similar to that produced by MyD88−/−/Trif−/− BMDMs (Fig. 5B). Similar results were obtained for Egr2 (data not shown). Tolerization treatment of MyD88−/−/Trif−/− BMDMs did not affect Egr1 expression (data not shown). These data suggest that cross-talk exists between the TLR-dependent and -independent signaling pathways, modulating the response of a subset of genes to the T3SS.
The cytoplasmic innate immune sensors Nod1 and Nod2, which are activated by peptidoglycan moieties, have been shown to activate NFκB [6]. Rip2 is a kinase that is downstream of both Nod1 and Nod2 and is required for induction of genes that respond to these sensors [39],[40]. We hypothesized that Nod1 and/or Nod2 could be involved in NFκB activation in response to Yersinia infection, as peptidoglycan fragments could have access to the host cell cytosol via the T3SS. To test this, we knocked down expression of Nod1 and Rip2 in 293T cells using shRNAs and monitored NFκB activation using a luciferase reporter. We found that NFκB-driven luciferase expression in 293T cells could be triggered by an acylated derivative of the Nod1 ligand iE-DAP (C12-iE-DAP), which can enter the cell cytosol when added exogenously to the cell culture (Fig. 6AB). Compared to a control shRNA, shRNA against Nod1 (Fig. 6A) or Rip2 (Fig. 6B) caused a clear decrease in C12-iE-DAP-induced NFκB activation. However, knocking down either Nod1 or Rip2 did not significantly decrease NFκB activation induced by the Y. pseudotuberculosis Δyop6 strain. In fact, in two out of four experiments the level of NFκB activation increased slightly upon Y. pseudotuberculosis Δyop6 infection during Rip2 knockdown compared to the LacZ shRNA control (data not shown).
To gain further evidence that the NFκB signaling observed in response to Y. pseudotuberculosis was independent of a peptidoglycan activation pathway, BMDMs from MyD88−/−, MyD88−/−/Rip2−/−, and MyD88−/−/Trif−/− mice were analyzed. TNF-α mRNA levels were determined after challenge of BMDMs with either Y. pseudotuberculosis Δyop6 or Δ6/ΔyopB in order to analyze activation of an NFκB-dependent gene in the absence of MyD88 signaling (Fig. 6C). Rip2 was dispensable for signaling downstream of both Y. pseudotuberculosis strains, regardless of the presence or absence of a functional T3SS translocator (Fig. 6C, MyD88−/− compared to MyD88−/−/Rip2−/− cells). In fact, MyD88−/−/Rip2−/− cells produced somewhat more TNF-α mRNA in response to either strain of Y. pseudotuberculosis compared to MyD88−/−/Rip2+/+ strains. In the case of the bacterial strain lacking the T3SS translocator (Δ6/ΔyopB), however, TNF-α signaling appeared to be solely TLR-dependent, as Trif was necessary for signaling in the absence of MyD88 (Fig. 6C, MyD88−/− compared to MyD88−/−/Trif−/− cells). In contrast, T3SS translocator-dependent signaling occurred in BMDMs from all the relevant mouse genotypes, regardless of intact TLR or Rip2 signaling. Therefore, we conclude that NFκB activation that occurs in response to T3SS translocator-positive Y. pseudotuberculosis does not involve Nod1 or Nod2 signaling.
T3SS-positive strains of Yersinia have been shown to trigger maturation and secretion of the inflammatory cytokine IL-1β in a manner that is dependent on the host protease caspase-1 [22],[41]. To determine whether caspase-1 is also involved in NFκB signaling triggered by T3SS translocator-positive Y. pseudotuberculosis, we infected MyD88−/−/Trif−/− macrophages with Y. pseudotuberculosis in the presence or absence of the caspase-1 inhibitor Z-YVAD(OMe)-FMK, and measured levels of secreted TNF-α (Table 5). TNF-α levels remained unchanged in the presence of caspase-1 inhibitor. This was in contrast to IL-1β secretion, which was inhibited by Z-YVAD(OMe)-FMK by three-fold (p<0.01). These data indicate that, as previously reported, the Yersinia T3SS induces host caspase-1 activation and IL-1β release [22],[41]. However, the Yersinia T3SS also induces host production and secretion of TNF-α and the signaling pathway responsible is independent of caspase-1.
T3SS translocator-dependent cytokine induction may stem from recognition of a YopBD pore or from introduction of innate immune stimulating molecules into the host cytosol. To distinguish between these two possibilities, we set up two systems in which products secreted by Y. pseudotuberculosis via the T3SS were introduced into the host cytosol independently of the YopBD pore, using either scrape-loading [42] or transfection. 293T cells were chosen for scrape-loading because they form loose adhesions to the extracellular matrix and have a high survival rate after scrape loading (unpublished observations). To determine the sensitivity of the 293T cell response to T3SS translocator-positive Yersinia, we infected 293T cells with varying numbers of live Y. pseudotuberculosis and measured induction of the NFκB-regulated cytokine IL-8. An inoculum of as little as one Y. pseudotuberculosis Δyop6 per five 293T cells (multiplicity of infection, MOI = 0.2) induced IL-8 mRNA production after two hours (Fig. 7A). In contrast, even a MOI = 20 of Y. pseudotuberculosis Δ6/ΔyopB did not induce IL-8. These results indicate that 293T cells produce detectable IL-8 levels in response to a small number of Yersinia expressing a functional T3SS translocator.
To test specifically whether molecules secreted via the Y. pseudotuberculosis T3SS could induce cytokine production in the absence of YopBD-mediated pore formation, we collected culture supernatant from Y. pseudotuberculosis grown under type III secretion-inducing conditions and introduced it into 293T cells via scrape-loading (see Materials and Methods). Under in vitro type III secretion-inducing conditions, Y. pseudotuberculosis strains that express a T3SS needle apparatus secrete proteins and possibly other molecules into the supernatant (see Fig. S1A). Scrape-loading of culture supernatant from Y. pseudotuberculosis Δyop6, equivalent to that produced by 50 bacteria per 293T cell, did not result in an increase in IL-8 mRNA (Fig. 7B). In contrast, lysate from heat-killed Y. pseudotuberculosis (HKYP), which contains molecules such as peptidoglycan known to induce cytosolic innate immune signaling, triggered IL-8 production when delivered inside 293T cells via scrape loading (Fig. 7B). This was achieved even when the equivalent of one HKYP was scrape-loaded into one 293T cell. However, even the equivalent of 100 HKYP per 293T cell was insufficient to induce IL-8 production in the absence of scrape-loading. Importantly, scrape-loading alone (i.e.-nonspecific membrane disruption) did not trigger IL-8 production.
To support these results, we used transfection of primary macrophages as a second method to deliver culture supernatants into the host cell cytosol independently of the YopBD pore. In order to demonstrate the sensitivity of the macrophages to T3SS translocator-positive Y. pseudotuberculosis, we challenged macrophages with varying numbers of Y. pseudotuberculosis Δyop6. A MOI = 0.1 (for TNF-α) or a MOI = 1.0 (for IFNβ) induced detectable cytokine mRNA levels (Fig. S5A,D). To test specifically whether molecules secreted by the T3SS could induce cytokine production in the absence of YopBD-mediated pore formation, we transfected MyD88−/−/Trif−/− macrophages with supernatants from bacterial cultures grown under type III secretion-inducing conditions. Neither TNF-α nor IFNβ was detected even after transfection with supernatant equivalent to about 70 Y. pseudotuberculosis per macrophage (Fig. S5B,E). In contrast, transfection of small amounts of HKYP lysate into macrophages elicited both a significant TNF-α and IFNβ response (Fig. S5C,F).
Collectively, these data indicate that neither introduction of Y. pseudotuberculosis type III secreted products into the cytosol of mammalian cells independently of the YopBD pore nor nonspecific membrane disruption can recapitulate induction of NFκB-regulated cytokines or type I IFN triggered by Y. pseudotuberculosis expressing a functional T3SS translocator.
We have characterized a previously unknown mammalian innate immune response to the enteropathogen Y. pseudotuberculosis. Unlike TLR recognition of Yersinia, the TLR-, Nod1/2-, and caspase-1-independent pathway described here is able to distinguish between Y. pseudotuberculosis expressing a functional T3SS and Y. pseudotuberculosis lacking this essential virulence determinant. Because the ensuing immune response includes at least two distinct signaling branches with different expression profiles, we propose the involvement of multiple cytosolic immune sensors. Bacterial uptake into host cells was not required for T3SS-stimulated NFκB activation or type I IFN production, indicating that extracellular Yersinia capable of type III translocation can activate diverse cytosolic innate immune signaling pathways. Although the host response that is detailed here does not require the presence of known T3SS effectors, cytokine production triggered by these recognition events is modulated in opposing directions by at least two such effector proteins, YopJ and YopT. Collectively, these results suggest multiple layers of host manipulation carried out by the Y. pseudotuberculosis T3SS.
Dozens of animal and plant pathogens use specialized secretion systems to deliver bacterial proteins into the cytosol of target host cells in order to manipulate host functions and promote virulence [9],[43]. We have shown here that the mammalian innate immune system possesses the ability to specifically recognize and respond to pathogenic Y. pseudotuberculosis based on bacterial utilization of one such virulence-associated secretion system. We hypothesize that different features of the Y. pseudotuberculosis secretion system are recognized by distinct cytosolic innate immune receptors, launching a unique immune response.
Inflammasome complexes, which contain such cytosolic sensors, are involved in caspase-1 activation, facilitating release of the cytokines IL-18 and IL-1β in response to Y. pseudotuberculosis, Y. pestis, Pseudomonas aeruginosa, Salmonella typhimurium, Legionella pneumophila, and Burkholderia pseudomallei that express functional T3/4SSs [22], [44]–[52]. Many of these inflammasome activation events also require bacterial flagellin, suggesting that flagellin is delivered into the cytosol of target host cells through these specialized secretion systems [53], triggering an immune response. While Y. pestis encodes an inactive flhD allele that results in lack of flagellin expression, T3SS-competent Y. pestis still activates caspase-1 [22], perhaps because the pore-forming ability of the Yersinia T3SS is sufficient to stimulate inflammasome activity. Such a mechanism of stimulation would be similar to the response of inflammasome complexes to a diverse set of pore-forming toxins [25]. Interestingly, when the inactive, Y. pestis allele of flhD is crossed into the Y. pseudotuberculosis Δyop6 background, TNF-α is still induced in a TLR-independent manner (data not shown). This suggests that NFκB activation triggered by the Yersinia T3SS is also independent of flagellin. Furthermore, we could find no role for the inflammasome-associated cytosolic sensor Nalp3 in Y. pseudotuberculosis-induced TNF-α production. In addition, the potassium ionophore nigericin, which induces IL-1β secretion via the Nalp3 inflammasome, did not promote TNF-α production (data not shown). Lastly, we found that NFκB induced by T3SS-positive Y. pseudotuberculosis is independent of caspase-1, ruling out a role for caspase-1-inflammasome complexes in this pathway (Table 5).
Y. pseudotuberculosis was previously shown to induce production of the NFκB-regulated chemokine IL-8 in HeLa cells and this was dependent on YopB expression [29]. Here we show that in addition to stimulation of IL-8 production in 293T cells (Fig. 7) and HeLa cells [29], T3SS-competent Y. pseudotuberculosis triggers a large-scale transcriptional response in mouse macrophages. This represents a much broader response to a T3SS than previously appreciated. Interestingly, T3SS-competent Citrobacter rodentium has been shown to induce IL-8 and TNF-α production in HT29 human intestinal epithelial cells dependent on the peptidoglycan sensors Nod1 and Nod2, respectively [28]. We also observed Y. pseudotuberculosis-induced IL-8 production in HT29 cells dependent on an intact T3SS (data not shown). However, we did not find a role for Nod1 or Nod2 in T3SS-dependent NFκB activation in our 293T cell reporter system or in primary macrophages (Fig. 6). Similarly, T3SS-competent B. pseudomallei triggers IL-8 production in 293T cells and this response is also independent of Nod1 [54]. The nature of the putative ligand(s) that triggers NFκB-dependent cytokine production by T3SS-competent Y. pseudotuberculosis and B. pseudomallei remains to be determined. Intriguingly, induction of IL-8 by B. pseudomallei does not occur in the absence of a functional T3SS even though this bacterium is able to access the host cytosol [54],[55]. This indicates that the presence of a bacterium in the host cell cytosol does not necessarily lead to immune detection. Rather, it is the active process of type III translocation into host cells that allows recognition to occur.
In contrast to innate immune ligands such as peptidoglycan, microbially-derived nucleic acids induce a distinct mammalian immune response that includes type I IFN family members [56]. Expression of these cytokines involves not only NFκB, but also the transcription factor IRF3 [57]. Following production and secretion of type I IFNs, the type I IFN receptor becomes activated in a paracrine and autocrine manner. This leads to the upregulation of a number of type I IFN-responsive genes, such as the chemokine IP10. Both cytosolic RNA and DNA can induce this type of amplified type I IFN response [56],[58]. Specific RNA moieties are recognized by two sensors, Mda5 and RIG-I, that reside in the mammalian cytosol [59]. However, the identity of the cytosolic DNA sensors remain unclear. The protein DAI/DLM-1/ZBP1 has shown to be involved in upregulation of type I IFN in response to DNA in some cell types but not in others [60]–[62], suggesting possible redundancy with other, as yet unidentified receptors. Indeed, several cell types from DAI−/− mice still responded to cytosolic DNA by producing type I IFN [63]. Another cytosolic DNA receptor, AIM2, has been described that induces NFκB and caspase-1 activation, but not type I IFN [64]–[67]. While at least some DNA detection pathways are independent of the RNA sensors RIG-I and Mda5, another pathway exists that depends on RNA polymerase III-mediated transcription of cytosolic DNA templates, leading to synthesis of RNA intermediates that trigger RIG-I-dependent type I IFN production [68],[69]. Interestingly, over 60% of the genes induced by introducing DNA into mouse embryonic fibroblasts [70] were also regulated by T3SS translocator-positive Yersinia in MyD88−/−/Trif−/− macrophages (data not shown). It is tempting to speculate that nucleic acids may enter the host cell cytosol during Y. pseudotuberculosis type III translocation, triggering a type I IFN response.
The T4SSs of two other pathogens, Brucella abortus and L. pneumophila, have been implicated in induction of type I IFNs [71],[72]. Because T4SSs are evolutionarily related to conjugation machinery, it is feasible that DNA is translocated into the target host cell, activating a cytosolic innate immune sensor. However, the nature of the Y. pseudotuberculosis type I IFN-inducing ligand is unclear, as no evidence exists that nucleic acids are secreted or translocated through a T3SS. Interestingly, we found that adding exogenous, synthetic RNA to MyD88−/−/Trif−/− macrophages infected with Y. pseudotuberculosis led to an enhanced IFNβ response (Fig. S6). Importantly, this was dependent on the bacteria expressing a functional T3SS translocator, indicating that extracellular nucleic acids can leak into the cytosol of mammalian cells during the process of type III translocation.
While an extracellular pathogen, group B streptococcus, was recently shown to trigger type I IFN production in a TLR-independent manner, bacterial phagocytosis and phagosomal membrane disruption were required [73]. To our knowledge, the Yersinia T3SS-dependent induction of a type I IFN response shown here is the first demonstration of an extracellular pathogen inducing TLR-independent type I IFN in the absence of bacterial internalization into host cells. This indicates that Y. pseudotuberculosis may employ a novel mechanism of innate immune activation. Whether type I IFN induction by the T3SS hinders or aids Y. pseudotuberculosis survival in vivo remains to be determined since the type I IFN receptor has been shown to facilitate or inhibit bacterial growth in mice depending on the type of pathogen used [74]–[78]. It will be important to determine the contribution of T3SS-induced type I IFN signaling to survival of mice during infection with Yersinia.
Previous work has shown that the Yersinia T3SS effector protein YopJ inhibits the NFκB, MAP kinase, and IRF3 signaling pathways in target host cells [79]. Consistent with these data, we observed reduced TNF-α levels triggered by Y. pseudotuberculosis expressing YopJ. Surprisingly, we also found that Y. pseudotuberculosis expressing either the T3SS effector protein YopT (Fig. 1) or YopE (data not shown) induced elevated TNF-α levels. This was dependent on the catalytic activity of these proteins because YopT and YopE point mutants that lack a catalytically active residue did not trigger these enhanced TNF-α levels (Fig. 1 and data not shown). YopT is a cysteine protease that cleaves the prenyl group of RhoGTPases, mislocalizing them from the plasma membrane [80]. YopE is a Rho GTPase activating protein (GAP) that accelerates the hydrolysis of GTP on RhoGTPases [17]. Both YopT and YopE lead to a loss of activated RhoGTPases from the plasma membrane and to actin cytoskeletal rearrangements. Interestingly, the RhoGTPase Rac1 was recently shown to negatively regulate Nod2 signaling, possibly by altering its subcellular localization [81],[82]. In addition, proteins that modulate the actin cytoskeleton have been linked to Nod1-dependent activation of NFκB [83]. While we could not find a role for Nod1 or Nod2 in Yersinia-induced NFκB activation (Fig. 6), it is possible that RhoGTPases may affect the activity of other cytosolic innate immune sensors.
Based on the data presented here, we propose that extracellular Y. pseudotuberculosis activates the cytosolic innate immune system during the process of type III translocation in a manner that is dependent on both YopBD-mediated pore formation as well as entrance of innate immune activating molecules into the target host cell cytosol (Fig. 8). The ability of extracellular Yersinia to trigger this response indicates that phagosomal degradation of the bacteria is not required to release the active ligands. In addition, the diverse transcriptional response to T3SS translocator-positive Yersinia in the absence of TLR signaling points to involvement of cytosolic innate immune sensors since no other known surface-associated proteins are capable of inducing such a robust, de novo response [84]–[86]. If the putative receptor(s) recognizing T3SS-competent Yersinia are not surface-associated, then the active components must penetrate the host cell to trigger the observed signaling. It is possible that the Yersinia T3SS induces the opening of a surface channel that allows an immune activating ligand to enter the cytosol. Alternatively, YopBD-mediated pore formation may allow molecules other than T3SS effector proteins inside target host cells. A possible model is that more than one such molecule enters the host cell, explaining the varied nature of the innate immune response. However, it is possible that one molecule is responsible for initiating the entire response and that its cognate receptor is capable of triggering several different signaling pathways with distinct kinetics. We could not recapitulate the signaling triggered by T3SS-competent Yersinia by introducing type III secreted molecules into host cells independently of the T3SS nor by inducing nonspecific membrane disruption. It is possible that the immune activating ligands are not secreted via the Yersinia T3SS during growth in bacterial culture media, but are translocated into host cells. Alternatively, the act of translocation (YopBD-mediated entry of activating ligands) may be specifically required. The nature of these putative ligands remains unclear, as we did not find a role for peptidoglycan or flagellin in innate immune recognition of the Yersinia T3SS.
The data presented here describing induction of inflammatory cytokines in response to T3SS translocator-positive Y. pseudotuberculosis is consistent with the well-established ability of enteropathogenic Yersinia to cause acute, localized gut inflammation [10]. Our results demonstrating cross-talk between TLR signaling and T3SS-dependent cytosolic signaling in response to Yersinia (Fig. 5) indicate that both pathways contribute to the overall immune response observed during Yersinia infection. Furthermore, in cells that are normally unresponsive to some TLR ligands (such as intestinal epithelial cells), recognition of a functional T3SS may play a primary role in specifically recognizing pathogenic bacteria. It will be important to determine the specific contribution of TLR-independent recognition of Yersinia in terms of both virulence of the pathogen as well as the intestinal inflammation caused by infection.
All animal use procedures were in strict accordance with the NIH Guide for the Care and Use of Laboratory Animals and were approved by the Tufts University Institutional Animal Care and Use Committee.
The Y. pseudotuberculosis strains used in this study (Table 1) were derived from the serogroup III strain IP2666pIB1 (Bliska et al. PNAS 1991). The Δyop6 and Δ6/ΔyopB strains were constructed using the suicide plasmids described by Logsdon and Mecsas [87]. The Δ6/ΔyopN strain was constructed using the suicide plasmid described by Davis and Mecsas [88]. The Δ6/pYopT and Δ6/pYopTC139S strains were constructed by electroporating pPHYopT and mating pPHYopTC139S, kind gifts from Dr. James Bliska, into the Δyop6 strain. These plasmids express wildtype YopT or YopT carrying a substitution mutation in a catalytically active residue under the control of the YopH promoter. The Δ6/pYopT strain was able to cause rounding of macrophages (indicative of Yop intoxication) and caused the entry of Rac1 into the nucleus of Cos1 cells (data not shown), while the Δ6/pYopTC139S strain did not. The Δ6/yopDΔstyI strain was constructed by mating pSB890 encoding the yopDΔstyI allele, a kind gift from Dr. Tessa Bergsbaken and Dr. Brad Cookson, into the Δyop6 strain. The yopDΔstyI allele was constructed according to the strategy described by Viboud et al. [29] and harbors a frameshift mutation in yopD. Y. pseudotuberculosis strains were grown in 2× YT overnight at 26°C with agitation for incubation with macrophages or 293T cells. The overnight cultures were diluted into 2× YT containing 20 mM sodium oxalate and 20 mM MgCl2 (low calcium medium) in order to obtain an OD600 of 0.2. These back-diluted cultures were grown at 26°C for 1.5 h with agitation, and transferred to 37°C for an additional 1.5 h with agitation to induce the expression of the type III secretion system [89]. In order to make culture fitrates for scrape-loading, the low calcium cultures were filtered through a 0.22 µm syringe filter (Millipore). The filtrate was then either left at its original concentration or was concentrated two-fold in a speed-vac centrifuge.
MyD88−/−/Trif−/− mice were a gift from Dr. Shizuo Akira. C57Bl/6 mice were purchased from Jackson Laboratories. MyD88−/− [90] and Rip2−/− [39] mice have been described.
Bone marrow derived macrophages (BMDMs) were obtained as previously described [74], except engineered NIH-3T3 cells were used as a source of CSF [38]. The harvested macrophages were frozen and stored in liquid nitrogen. Frozen aliquots were thawed and plated one day prior to use in Dulbecco's modified Eagle medium (Gibco) supplemented with 10% fetal bovine serum (HyClone) and L-glutamine (Gibco). 293T cells were passaged and plated in the identical medium.
BMDMs were plated onto six-well plates (Falcon) at a density of 106 cells per ml and incubated at 37° with 5% CO2 overnight. One day later, the cells were inoculated with approximately 1–3×107 Y. pseudotuberculosis grown under type III secretion-inducing conditions (see Bacterial Strains section above) for a multiplicity of infection (MOI) of approximately 10∶1 (unless otherwise specified). The bacteria were then allowed to settle onto the macrophage monolayer for 30 minutes at 37° with 5% CO2. This treatment allowed approximately two-thirds of the macrophages to be associated with one or two bacteria thirty minutes post-inoculation (data not shown). The media was then filtered through a 0.22 µm syringe filter (Millipore). This was done in order to remove any non-attached or non-internalized bacteria without removing any cytokines secreted by the macrophages during the initial 30 minutes of incubation. This treatment prevented the macrophage monolayer from becoming overwhelmed with bacteria by the end of the inoculation period and yet allowed any extracellular bacteria attached to macrophages and in the process of type III secretion to remain viable. Two hours post-inoculation, macrophage supernatants were collected and stored at −80°C. Macrophage monolayers were then washed once with phosphate-buffered saline and total RNA was harvested using the RNAqueous kit (Ambion) according to the manufacturer's instructions. For inhibition of phagocytosis, BMDMs were treated with 2 µM cytochalasin D (Sigma) 30 minutes prior to infection.
The DNA-free kit (Ambion) was used to remove any contaminating genomic DNA from the total RNA samples harvested as described above. Total RNA yield was calculated using a NanoDrop ND-1000 spectrophotometer (Thermo Scientific) and 2 µg RNA was used to make cDNA as previously described [74]. SYBR Green PCR master mix (Applied Biosystems) was used for qPCR reactions according to the manufacturer's instructions using a Mx3005P (Stratagene) or DNA Engine Opticon 2 (BioRad) qPCR machine and a 60°C annealing temperature. The results were analyzed using the Mx3005P or Opticon 2 software. QPCR primers used in this study are described in Table S1. All qPCR primers were validated in silico using NCBI mapviewer (www.ncbi.nlm.nih.gov/projects/mapview/).
Y. pseudotuberculosis was grown under type III secretion inducing conditions as described in the Bacterial Strains section above, except the cultures were incubated at 37°C for an additional 30 minutes (two hours total). The cultures were plated on LB plates to determine CFU/ml. 900 µl of culture was pelleted and 800 µl supernatant was removed. Trichloroacetic acid was added to the supernatant for a final concentration of 10% and the mixture was incubated on ice for 15 minutes. The mixture was pelleted for 15 minutes at 13,000 rpm and the pellet was washed with acetone. The washed pellet was resuspended in 50 µl Laemmli buffer and frozen at −80°C. The samples were thawed and boiled for five minutes. Approximately half of each sample (the samples were normalized for CFU/ml) was run on a 12.5% polyacrylamide gel and the protein bands visualized with coomassie blue staining.
MyD88−/−/Trif/−/− BMDMs were infected as described above with Y. pseudotuberculosis Δyop6 or Δ6/ΔyopB or were left uninfected. After 45 minutes, two hours, or four hours post-inoculation, total RNA was harvested as described above except that two wells of macrophages were pooled per condition to yield enough RNA for microarray analysis. The infection was then repeated on a separate day to yield two biological replicates per condition. 5 µg total RNA was used to make probes for GeneChip Mouse Genome 430 2.0 arrays using the One-Cycle cDNA synthesis protocol and array hybridization was performed according to the manufacturer's instructions (Affymetrix). A GeneChip Fluidics Station was used to wash the arrays and a GeneChip Scanner was used to read the arrays (Affymetrix). GeneChip Operating Software was used to analyze the quality of the hybridizations (Affymetrix). GeneSpring GX software (Agilent) was used to analyze the microarray data and the GC Robust Multiarray Averaging method was used for data normalization.
To make heat-killed Y. pseudotuberculosis, the wildtype IP2666 strain was grown overnight at 26° with agitation in Luria-Bertani broth. The overnight culture was heat-killed at 60°C for one hour, aliquoted, and frozen at −80°C in the absence of glycerol. Aliquots were thawed at room temperature before use. Very few intact bacteria could be visualized by microscopy (unpublished observations), indicating lysis occurred upon freeze/thawing. The culture was plated before and after heat-killing to calculate the live colony forming unit (CFU) equivalents in the heat-killed mixture and to ensure complete killing.
C57Bl/6 BMDMs were plated onto six-well plates (Falcon) at a density of 2×106 per well in 2 ml of media containing 2×105 heat-killed Y. pseudotuberculosis. The cells were incubated overnight at 37° with 5% CO2. The next day, the tolerized macrophages were inoculated with live Y. pseudotuberculosis or left uninfected as described in the Macrophage infections section above.
293T cells were plated at a density of 2.5×104 per 100 µl in 96 well plates (Corning). One day later, the cells were transfected with 400 ng per well of plasmid encoding shRNA specific for Nod1, Rip2 or control shRNA (EGFP for Nod1 and LacZ for Rip2) using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. The Nod1 and EGFP shRNA constructs were cloned into the pRNAT-U6.1/Neo vector backbone (GenScript). The Nod1 shRNA construct #1 and shRNA EGFP control were based on previously published siRNA sequences (Viala et al. 2004). The Nod1 shRNA construct #2 (5′- GATCCCGTCAAAGGCAGCACGGAAGTGCTTGATATCCGGCACTTCCGTGCTGCCTTTGATTTTTTCCAAA -3′) was designed using GenScript siRNA Design Center (V. Losick and R. Isberg). The Rip2 and LacZ shRNA plasmids were purchased from InvivoGen. 200 ng per well of an NFκB luciferase reporter plasmid (Stratagene) was transfected into 293T cells either simultaneously with the Nod1 or EGFP shRNA plasmids or 24 hours after the Rip2 or LacZ shRNA plasmids. After 48 hours of shRNA plasmid transfection, the 293T cells were infected with 3–4×105 Y. pseudotuberculosis Δyop6 or Δ6/ΔyopB per well. Since the 293T cells doubled at least twice during the 48 hour transfection period, the MOI was approximately 3∶1. Alternatively, transfected 293T cells were treated with 1 µg/ml C12-iE-DAP, an acylated derivative of the iE-DAP Nod1 ligand that is able to enter the cytosol of cultured cells (InvivoGen). After four hours of infection with Y. pseudotuberculosis or treatment with C12-iE-DAP, luminescence was measured using the SteadyLite Plus reporter gene assay system (PerkinElmer) and a SpectraMax M5 microplate reader (Molecular Devices).
BMDMs were plated onto 24 well plates (Falcon) at a concentration of 5.75×105 per ml and incubated overnight. The irreversible caspase-1 inhibitor Z-YVAD(OMe)-FMK (Santa Cruz Biotechnology) was added at a concentration of 20 µM two hours prior to infection. The cells were inoculated with approximately 1–4×107 Y. pseudotuberculosis grown under type III secretion-inducing conditions (see Bacterial Strains section above) for a multiplicity of infection (MOI) of approximately 20∶1. Four hours post-inoculation, supernatants were collected and frozen at −80°C.
For Fig. 1D, the amount of TNF-α protein in supernatants of infected macrophages was measured as described in Auerbuch et al. [74] using the Cytometric Bead Array Mouse Inflammation Kit (BD Biosciences). For Table 5, frozen supernatants were centrifuged at 13,000 rpm for 1 min and TNF-α and IL-1β protein levels were measured using a plate-based ELISA (eBioscience) and a Victor 3 microplate reader (PerkinElmer).
The scrape-loading protocol used was adapted from McNeil et al. [42]. 293T cells were plated at a density of 0.5–1×106 cells per ml in six well plates. One day later, the cells were washed once with warm HBSS. 300 µl warm HBSS plus ligand were added to each well. MOI equivalents of 0.1–100 HKYP or 150 µl of culture filtrate from live Y. pseudotuberculosis grown under type III secretion-inducing conditions (see Bacterial Strains section above) were used as ligands. Cell monolayers were then scraped with a cell scraper and incubated at 37°C for five to ten minutes. The cells were then washed off the plastic with 5 ml of warm HBSS and pelleted. The cells were resuspended in 2 ml warm media, plated on fresh six well plates, and incubated at 37°C for two hours. Alternatively, 293T cells were incubated with HKYP (MOI 100) in the absence of scraping for two hours as a negative control. Total RNA was then harvested as described for infected macrophage monolayers and IL-8 mRNA levels measured by qPCR as described above.
293T cells were plated at a density of 5×105 cells per ml in six well plates. One day later, the cells were inoculated with Y. pseudotuberculosis grown under type III secretion-inducing conditions at a MOI of 1∶10 to 10∶1 (Yersinia:293Ts). The monolayer supernatants were filtered after 30 minutes as described for macrophage infections. IL-8 mRNA was measured after two hours of total infection as described above.
Primary macrophages were plated onto six well plates at a concentration of 106 per ml. One day later, MOI equivalents of 0.1–10 HKYP or 250 µl per well of culture filtrate from live Y. pseudotuberculosis grown under type III secretion-inducing conditions were transfected into the macrophages using Lipofectamine 2000 (Invitrogen) according to the manufacturer's recommendations (5 µl Lipofectamine 2000 reagent per well). The cells were then incubated for two hours. Total RNA was harvested and tnfa or ifnb mRNA levels measured as described above. The culture fitrates did not inhibit the transfection process because transfection of both the synthetic RNA poly(I:C) and culture filtrate together into macrophages did not lead to reduced ifnb mRNA levels compared to poly(I:C) alone (data not shown).
|
10.1371/journal.ppat.1006715 | The full transcription map of mouse papillomavirus type 1 (MmuPV1) in mouse wart tissues | Mouse papillomavirus type 1 (MmuPV1) provides, for the first time, the opportunity to study infection and pathogenesis of papillomaviruses in the context of laboratory mice. In this report, we define the transcriptome of MmuPV1 genome present in papillomas arising in experimentally infected mice using a combination of RNA-seq, PacBio Iso-seq, 5’ RACE, 3’ RACE, primer-walking RT-PCR, RNase protection, Northern blot and in situ hybridization analyses. We demonstrate that the MmuPV1 genome is transcribed unidirectionally from five major promoters (P) or transcription start sites (TSS) and polyadenylates its transcripts at two major polyadenylation (pA) sites. We designate the P7503, P360 and P859 as “early” promoters because they give rise to transcripts mostly utilizing the polyadenylation signal at nt 3844 and therefore can only encode early genes, and P7107 and P533 as “late” promoters because they give rise to transcripts utilizing polyadenylation signals at either nt 3844 or nt 7047, the latter being able to encode late, capsid proteins. MmuPV1 genome contains five splice donor sites and three acceptor sites that produce thirty-six RNA isoforms deduced to express seven predicted early gene products (E6, E7, E1, E1^M1, E1^M2, E2 and E8^E2) and three predicted late gene products (E1^E4, L2 and L1). The majority of the viral early transcripts are spliced once from nt 757 to 3139, while viral late transcripts, which are predicted to encode L1, are spliced twice, first from nt 7243 to either nt 3139 (P7107) or nt 757 to 3139 (P533) and second from nt 3431 to nt 5372. Thirteen of these viral transcripts were detectable by Northern blot analysis, with the P533-derived late E1^E4 transcripts being the most abundant. The late transcripts could be detected in highly differentiated keratinocytes of MmuPV1-infected tissues as early as ten days after MmuPV1 inoculation and correlated with detection of L1 protein and viral DNA amplification. In mature warts, detection of L1 was also found in more poorly differentiated cells, as previously reported. Subclinical infections were also observed. The comprehensive transcription map of MmuPV1 generated in this study provides further evidence that MmuPV1 is similar to high-risk cutaneous beta human papillomaviruses. The knowledge revealed will facilitate the use of MmuPV1 as an animal virus model for understanding of human papillomavirus gene expression, pathogenesis and immunology.
| Papillomavirus (PV) infections lead to development of both benign warts and cancers. Because PVs are epitheliotropic and species specific, it has been extremely challenging to study PV infection in the context of a naturally occurring infection in a tractable laboratory animal. The recent discovery of the papillomavirus, MmuPV1, that infects laboratory mice, provides an important new animal model system for understanding the pathogenesis of papillomavirus-associated diseases. By using state of the art RNA-seq to provide deep sequencing analysis of what regions of the viral genome are transcribed and PacBio Iso-seq that produces longer reads to define the complete sequences of individual transcripts in combination with several conventional technologies to confirm transcription starts sites, splice sites, and polyadenylation sites, we provide the first detailed description of the MmuPV1 transcript map using RNA from MmuPV1-induced mouse warts. This study reveals the presence of mRNA transcripts capable of coding for ten protein products in the MmuPV1 genome and leads to correctly re-assigning the E1^E4, L2 and L1 coding regions. We were able to detect individual transcripts from the infected wart tissues by RT-PCR, Northern blot and RNA ISH, to define the temporal onset of productive viral infection and to ectopically express a predicted viral protein for functional studies. The constructed MmuPV1 transcript map provides a foundation to advance our understanding of papillomavirus biology and pathogenesis.
| Human papillomaviruses are a group of small, non-enveloped, epitheliotropic DNA tumor viruses whose infection can result in benign lesions (called warts or papillomas) and in some cases cause malignancies. Certain genotypes of HPVs, such as HPV-16, HPV-18, and HPV-31, that infect mucosal epithelia, have been recognized as causative agents of anogenital cancers that include cervical and anal cancers, as well as a growing subset of head and neck cancers, particularly those arising in the oropharynx [1,2]. Papillomaviruses are species-specific. Although papillomavirus infection models in large animal species such as rabbits, dogs and cows have been used to study the molecular biology and pathogenesis of papillomavirus infections, a laboratory mouse model would greatly facilitate the study of papillomavirus-associated warts and cancers. The recent identification of the murine papillomavirus (MmuPV1) that can infect laboratory strains of mice now provides us with such a tractable laboratory animal-based infection model system [3–6]. The MmuPV1 circular, double stranded DNA genome is 7510-bp in length and encodes at least seven translational open reading frames (ORFs), designated E1, E2, E4, E6, E7, L1 and L2 based upon their conserved position within the viral genome and length comparable to ORFs of other papillomaviruses [7,8]. To date, a transcription map of MmuPV1 has not been described. Such a map would greatly facilitate understanding the MmuPV1genome structure and viral gene expression capacities and thereby information on nature of viral factors that contribute to papillomavirus infection and associated pathogenesis.
In this report, we describe a comprehensive map of MmuPV1 transcripts based upon a multi-pronged analysis of viral mRNAs isolated from tumor tissues derived from MmuPV1 infected mice. The viral transcription start sites (TSS) were mapped by 5’ rapid amplification of cDNA ends (5’-RACE) [9,10] in combination with PacBio Iso-seq and confirmed by TA cloning and Sanger sequencing. The polyadenylation cleavage sites of viral early and late transcripts were mapped by 3’-RACE [9,10]. Viral genome expression and RNA splicing was profiled by RNA-seq and primer-walking RT-PCR [9–11]. Accordingly, we assigned the coding regions of multiple potential viral gene products, E1^E4, E1^M1, E1^M2, E8^E2, L2 and L1 based upon the deduced structure of mRNA transcripts, and performed in situ hybridization studies in which we could detect a subset of these viral early and late transcripts in MmuPV1 infection-derived wart tissues by RNA-ISH in correlation with the presence of viral proteins and viral genomic DNA.
Athymic FoxN1nu/nu mice were infected with MmuPV1 as follows: inner (right) ear, muzzle, and three spots on the tail following scarification (Fig 1A). Resulting warts were harvested six months following infection and total RNA was isolated. Papillomatosis was confirmed by histopathological analysis (S1A Fig). Papillomas exhibited fibrillary projections accompanied with hyperkeratosis and were exophytic in morphology. Koilocytes, considered as hallmarks of papillomavirus infection, were seen throughout the papillomas. Productive phase of viral life cycle was confirmed by L1 positivity by immunofluorescence and detection of amplified viral DNA by fluorescence in situ hybridization. L1 and amplified viral DNA were detected throughout the papilloma including the terminally differentiating epithelia (S1A Fig).
To elucidate the presence and relative abundance of RNA transcripts arising from the MmuPV1 genome from these wart tissues (S1A Fig), ribosomal RNA-depleted total RNA isolated from each wart tissue at three anatomical sites of three infected animals was analyzed by RNA-seq. In addition, the uninfected ear from the same animal was used as an uninfected control. Approximately 100 million paired-end reads with high quality were obtained from each tissue sample (Table 1). By mapping the RNA-seq raw reads from each lesion to the newly arranged linear MmuPV1 genome starting from nt 7088 and ending at nt 7087 using RNA sequence aligner TopHat [12], we obtained ~0.4–3.3 million viral reads for each wart sample (GEO Accession No. GSE104118), varying among lesions and animals, which accounts for ~0.4%-3.4% of total RNA reads obtained from each sample and with the muzzle tissues from animal #2 containing the most RNA reads (Table 1). Surprisingly, we also obtained many MmuPV1 reads from the uninfected ear tissues (control ears) in all three wart-bearing animals (Table 1), with the animal #1 uninfected ear (left ear) displaying the viral reads similar to that of the infected right ear while appearing normal by visual scoring. MmuPV1 L1 protein and DNA were detected in this tissue confirming the presence of subclinical infections in these control ear tissues (S1B Fig). By uploading these uniquely mapped viral RNA reads obtained from individual samples to the Integrative Genomics Viewer (IGV) program to visualize reads coverage profile along with the MmuPV1 genome, we found three major coverage peaks, one in the E7 region, one in the E4 region and one between E4 and L2, make the last two peaks as a V shape (Fig 1B) among all wart-tissues and sub-clinically infected ear tissues obtained from the MmuPV1 wart-bearing animals. The V shape was attributed to (1) fewer uniquely mapped viral RNA reads, but more host-viral chimeric reads being excluded from the mapping in this region, and (2) RNA splicing by using a 5’ splice site at nt 3431 (see more detailed description later in this report). We also saw a small drop of viral RNA reads within the E4 ORF in particular in the control animal ears and this might be a result from increased E1^E4 splicing in proportion in the control animal ears when compared with the tumor ear RNA. By analyzing the orientation of the uniquely mapped viral RNA reads from the muzzle tissues obtained from mouse #2 which appeared the highest viral reads coverage, we conclude that the vast majority of the viral reads were of the sense strand (98.6%). Antisense-specific reads were of low abundance (~1.4% of all viral reads), although a few peaks at various points along with the viral genome were evident (Fig 1C). We view these antisense reads being background noise from cDNA library construction, sequencing errors or mapping artifacts. A step-wise zoom-in view further showed that the viral sense transcripts derived from the LCR (long control region), E6, E1, E2, L2 and L1 regions were less abundant than the reads from the E7 and E4 regions, with the L1 reads a little more than the others (Fig 1D).
Considering majority of eukaryotic RNA has a 5’-end cap structure, the current RNA-seq protocol lacks a de-capping step and fragments RNA into 200–350 nt pieces for adaptor ligation and amplification. These methods create an unfavorable bias for adaptor ligation to the RNA 5’ end and deletion of the RNA 5’ end reads smaller than 200 nts, making the RNA-seq unattainable for mapping the RNA start site [13]. To map the transcription start sites (TSS) of MmuPV1 transcripts, 5′ RACE analyses were performed on total RNA isolated from MmuPV1-infected ear lesions using virus-specific antisense primer Pr7237, Pr352, Pr518, Pr687, Pr738, Pr1123, Pr3299, or Pr5452 (Fig 2A, S2 Fig and S1 Table). The 5′ RACE products were visualized either by a DNA Bioanalyzer (Fig 2B) or by agarose gel electrophoresis (Fig 2E and S2 Fig). The 5’ RACE products derived from Pr3299 and Pr5452 were also subjected to long read single-molecular, real-time sequencing using PacBio (Pacific Biosciences) Iso-seq technology (Fig 2C and 2D, S3A Fig), for detection of existing full-length transcripts and overcoming the RNA-seq unfavorable bias on the fragmented RNA 5’ end. In addition, all 5’ RACE products were gel-purified, cloned and sequenced (Fig 2E, S2 and S3B Figs). From PacBio Iso-eq sequencing, approximate five thousands of the full-length viral transcripts were detected from each primer-derived 5’ RACE products (Table 2). By uploading these Iso-seq-identified individual viral transcripts to the IGV program to visualize their coverage profile along with MmuPV1 genome, we found that the Pr3299-derived 5’ RACE products were mainly transcribed from two TSS, one at nt 533 (~50%) and the other at nt 7503 (~20%) in the viral genome, although other scattered TSS positions were also identified immediately downstream of the nt 533 position (Fig 2C and 2D, Table 2). In contrast, the Pr5452-derived 5’ RACE products were predominantly transcribed from the nt 7107 position (>55%) and secondarily from the nt 533 position (>30%) on the viral genome (Fig 2C and 2D, Table 2). Similar to the Pr3299-derived products, the Pr5452-derived 5’ RACE products displayed scattered TSS positions (mostly from the nt 576 to nt 607) immediately downstream of the nt 533 position (Fig 2C). Both 5’ RACE products gave the same TSS products initiating at nt 859 (Fig 2C and 2D), while the TSS at the nt 360 was most associated with Pr3299-derived products (Fig 2C and 2D, S3A Fig and Table 2).
The TSS of viral transcripts was also assessed by TA cloning and Sanger sequencing of individual primer-derived 5’ RACE products. As shown in Fig 2E, S2 Fig and S2 Table, one 5’ RACE product obtained from the Pr7237 (lane 1) was mapped primarily to the nt 7107 (6/9 colonies). Three RACE products were detected from the Pr352 (lane 2) and they were defined to have TSS at nt 260 (7/14 colonies), nt 7503 (6/10 colonies) and nt 7107 (direct sequencing of PCR products). Pr518 exhibited a similar 5’ RACE product profile, 166-bp larger than that of Pr352. Direct sequencing of four Pr518 5’ RACE products (lane 3) identified TSS at nt 360, 260, 7503 and 7107. The Pr687 and Pr738 also showed a similar 5’ RACE product profile, with the two Pr738-derived 5’ RACE products being 51-nts larger than that of Pr692 (compare lane 4 to lane 5). Direct sequencing showed the main product from Pr687 had a TSS at nt 533 (lane 4). Cloning and sequencing of the Pr738-derived products 1 or 2 indicated that product 1 had a TSS at nt 576/579/589 (7/20 colonies) and product 2 had a TSS at nt 533 (6/21 colonies) (lane 5). The Pr3299 generated two major spliced 5’ RACE products containing splice junctions at nts 757/3139 or nts 7243/3139 (lane 6), with TSS for product 1 mainly placed at nt 7107 (12/18 colonies) and for product 2 at nts 533 or 576 (6/17 colonies). Pr5452 (lane 7) had a similar 5’ RACE profile to that of Pr3299, but gave three faster migrating faint bands. TA cloning and sequencing of the Pr5452 5’ RACE products revealed that its product 1 had a TSS mainly at nt 7107 (3/8 colonies) and product 2 had a TSS between nt 533–589 (6/8 colonies). Both were double spliced products either from nt 7243 to 3139 or from nt 757 to 3139 and then from nt 3431 to 5372. We did not clone and sequence the faster migrating faint bands which are most likely the products of Pr7107 (216 bp), Pr360 (477 bp) and Pr533 (304 bp) being spliced from nt 7243 or nt 757 to 5372. Using the Pr1123 for 5’ RACE (lane 8), cloning and sequencing revealed two smaller products being, respectively, transcribed from the nt 859 (7/12 colonies) and nt 760 (3/5 colonies).
Together, these 5’ RACE experiments identified nt 7107, 7503, 360, 533 and 859 as preferred TSS for MmuPV1 gene expression. The mapped TSS all started at a purine A or G, which is in agreement with conserved TSS having a purine in eukaryotes [14]. The TSS at nt 7503 drives MmuPV1 E6 transcription, the TSS at nt 360 drives MmuPV1 E7 transcription and the TSS at nt 859 drives E2 expression. The TSS at nts 7107 and 533 drive MmuPV1 late transcription. Hereafter, each of the preferred TSS is named as a promoter: P7107, P7503, P360, P533 or P859. Analyses of the region 5′ to each mapped TSS show that the P7503 has a TATA box (a eukaryotic core promoter motif) 25-bp upstream of its TSS, but other two promoters for viral early gene expression do not (S3B Fig). Two viral late promoters either bear a TATA-like box for P7107 or a TATA box 110-bp upstream of the promoter P533 (S3B Fig). These features of viral promoters perhaps account for the observed heterogeneity of their transcription initiation as seen in the expression of HPV-18 [10] and other eukaryotic genes [15].
The P7107 late promoter identified above by 5’ RACE was confirmed by RNase protection assays (RPA) performed on total RNA isolated from MmuPV1-infected lesions using a 32P-labeled antisense RNA probe from nt 6846 to 7237 that covers the mapped TSS around nt 7107. The RPA products were analyzed by electrophoresis in a denaturing 8% PAGE gel, along with the DNA sequencing ladders generated from MmuPV1 genome by a 32P-labeled antisense primer Pr7237. As shown in S4 Fig, the protected RPA products from the 6846–7237 probe showed two major bands, one (arrow) of 131 nts in length corresponds to the TSS at nt 7107. The other product of 217 nts (arrowhead) corresponds to the L1 mRNA cleaved at nt 7063 for the late polyadenylation (see below).
Genome analyses of MmuPV1 suggest a putative early poly(A) signal (PAS) AAUAAA downstream of the viral E2 ORF at nt 3844 presumably for early polyadenylation and two putative late PAS, at nt 5609 and 7047, responsible presumably for polyadenylation of viral L2 and L1 transcripts (Fig 3A). To determine further the early polyadenylation cleavage sites (CS), total RNA isolated from MmuPV1-infected lesions was analyzed by 3′ RACE using a MmuPV1-specific sense primer Pr3277 located within the E4 ORF. Following gel purification, cloning, and sequencing of a 3′ RACE product of ~750 -bp (Fig 3B), we found that 11/23 sequenced bacterial colonies exhibited a product with a 3′-end at nt 3864, 15 nt downstream of the putative nt 3844 PAS (Fig 3B). Additional 3′ RACE with a MmuPV1-specific sense primer Pr116 and Pr522, located within the E6 and E7 ORF, respectively, also determined the early cleavage site around nt 3864, confirming that MmuPV1 early transcripts are cleaved at nt 3864 for RNA polyadenylation using the nt 3844 PAS AAUAAA although its 3′ downstream sequence has no U/GU motifs, the highly conserved recognition sites for CSF (cleavage stimulation factor) binding in context of the RNA polyadenylation [16,17].
To detect the late CS, we carried out a 3′ RACE with MmuPV1-specific sense primers Pr6846 or Pr5433 located in the L1 ORF. Following gel purification and cloning of the Pr6846-derived 3′ RACE products of ~310 bp, we sequenced 16 bacterial colonies and found the usage of multiple cleavage sites for polyadenylation of MmuPV1 late transcripts. Eight clones exhibited a product with a 3′end at nt 7063, 11 nt downstream of the putative nt 7047 PAS AAUAAA motif (Fig 3C). 3′ RACE with an additional MmuPV1-specific sense primer Pr5433 also determined the late CS primarily mapping to nt 7063 (S5 Fig). Analysis of the region downstream of this cleavage site shows three overlapping U/GU motifs from nts 7109 to 7152. RPA analysis of the total cell RNA isolated from MmuPV1-infected lesions using the 32P-labeled antisense RNA probe from nt 6846 to 7237 further confirmed the presence of the late CS around the nt 7063 (S4 Fig).
In addition, we identified infrequent usage of the nt 5609 PAS AAUAAA for RNA polyadenylation at nt 5627 from Pr5433-derived 3’ RACE products (S5 Fig). This PAS could be useful for the expression of L2, but would lead to produce a truncated L1 protein. Because of its low frequency of usage, we consider that it represents a cryptic polyadenylation site of unknown function.
Given the presence of multiple splice sites in both early and late transcripts of papillomaviruses, we used a snout (muzzle) wart sample from the animal #2 that gave 3.28 million viral reads, the highest among all twelve samples, to elucidate all possible usage of viral splice sites in the MmuPV1 genome. STAR aligner program [18] was used to explore the potential exon-exon splicing junctions with a threshold of minimal overhang >30 nts for non-canonical junctions and >10 nts for canonical junctions. As shown in Fig 4A, we identified from this sample 324,535 junction reads (10% of total viral reads) that defined nine splicing junctions, with frequency by numbers of the junction reads in the order of nts 757/3139 >3431/5372 >7243/3139>757/2493 >7243/5372 >757/5372 >7243/2493 >1194/3139 >1125/3139 (S3 Table). The most common splice junction read, at nts 757/3139, accounted for 90% of total junction reads. The least common reads were those for the 1125/3139 splice, accounting for only 0.02% of the total junction reads. The identified splice junctions of 7243/3139, 757/3139, and 3431/5372 (Fig 4A) confirmed the findings from the 5’ RACE analyses using Pr3299 and Pr5452 primers (Fig 2E). IGV Sashimi plot visualized that the obtained viral junction reads were derived from five different 5’ splice sites (donor sites) to three separate 3’ splice sites (acceptor sites) in the MmuPV1 genome (Fig 4A). Analysis of the intron sequences between the exon-exon junctions confirmed that all introns contain a consensus GU dinucleotide in the intron 5′ end and a consensus AG dinucleotide in the intron 3′ end. The similar frequency patterns of the detected splice junctions were observed in all remaining samples, except the splice junction 1125/3139 which was detected only in 4 out of 9 tumor samples (S3 Table). We noticed that the 757/3139 junction reads were proportionally higher in the sub-clinically infected control ears than that seen in the tumor ears (S3 Table).
Primer walking RT-PCR using various combinations of primer pairs (Fig 4B) on total mRNA isolated from MmuPV1-infected lesions was used to further validate the exon-exon junctions identified by RNA-seq and by 5’ RACE for MmuPV1 early and late transcripts. Gel-purification and sequencing of each RT-PCR product confirmed all of the splice junction identified by RNA-seq. Using a forward primer of Pr7140 downstream of the late promoter P7107 in combination with a backward primer of Pr5452 (L1), Pr4647 (L2), Pr3299 (E4) or Pr2978 (E2), as shown in Fig 4C, we detected two L1 products (lane 3) and one each for L2, E4 and E2 (lanes 5, 7 and 9). Sequencing of the gel-purified RT-PCR products indicated that two L1 products (185-bp and 478-bp, lane 3) were the singly spliced (at 7243/5372) and doubly spliced (at 7243/3139 and 3431/5372) transcripts, respectively (Fig 4D); both an L2 product (1612-bp, Fig 4C, lane 5) and an E4 product (265-bp, Fig 4C, lane 7) were singly spliced (7243/3139) transcripts (Fig 4D); and an E2 product (590-bp, Fig 4C, lane 9) was a singly spliced (7243/2493) transcript (Fig 4D). Using a forward primer of Pr665 within the E7 ORF in combination with the same sets of backward primers also showed two L1 products (Fig 4C, lane 12) and one each for L2, E4 and E2 (Fig 4C, lanes 14, 16 and 18). L1 product 1 (174-bp, Fig 4C, lane 12) was a singly spliced (757/5372) transcript, whereas the L1 product 2 (467-bp, Fig 4C, lane 12) was a doubly spliced (757/3139 and 3431/5372) transcript (Fig 4D); both L2 (1602-bp) and E4 (254-bp) products (Fig 4C, lanes 14 and 16) were singly spliced (757/3139) transcripts; the E2 product (579-bp, Fig 4C, lane 18) was a singly spliced 757/2493 transcript (Fig 4D).
Using a backward primer Pr3299 in combination of a forward primer Pr1031 or Pr1141, we detected three RT-PCR products in sizes of 257-bp (product 1), 326-bp (product 2) and 2269-bp (product 3) for Pr1031 and two products of 214-bp (product 1) and 2158-bp (product 2) for Pr1141 (Fig 4C, lanes 20 and 22). We found both 257-bp and 326-bp products were the spliced products of 1125/3139 and 1194/3139 and the 214-bp product was also a spliced product of 1194/3139 (Fig 4D). Both the 2269-bp and 2158-bp products were unspliced products from this region presumably for E1 expression.
Intron retention is one of the common alternative splicing events during RNA splicing and is essential for the expression of E1 and L2 in all papillomaviruses. In high-risk HPVs, intron retention is also necessary for viral E6 expression. From RNA-seq analysis, a small fraction of viral reads spanned over the entire MmuPV1 E1 and L2 ORF regions (Figs 1D and 4A). These were further confirmed by primer-walking RT-PCR with a series of combined primer pairs (Fig 5A) from total RNA extracted from the MmuPV1-infected muzzle wart tissues of animal #2. As shown in Fig 5B, the primer-walking RT-PCR using a forward primer Pr7140 in combination with a reverse primer Pr135, Pr352, Pr738, Pr827 or Pr1938 for detection of the P7107-derived late transcripts, which are predominantly spliced from the nt 7243 5’ss to nt 3139 3’ss, led to the amplification of transcripts with an intron retained from nt 7244 to 3138 (lanes 2, 4, 6, 8 and 10). Using a forward primer Pr1263 in combination with a backward primer Pr4647 for L2 or Pr2978 for E2 (lanes 13–14), we obtained a 1715-bp product (lane 14) from the Pr2978, but none from the Pr4647 (lane 13), indicating retention of the intron within the E1 and E2 ORFs, but not together with retention of another intron in the L2 region. Using a forward primer of Pr1141 in combination with a backward primer of Pr3299, we detected a 214-bp product spliced from nt 1194 to 3139 (lane 18). The other primer pairs used in the same detection did not give any obvious RT-PCR products (lanes 16, 17, 19 and 20).
Next, we compared the primer-walking RT-PCR using fewer amplification cycles (25 cycles) and using an exonic or intronic backward primer in combination with the same sets of forward primers (Pr7140, Pr116, Pr522 or Pr665) to detect intron-containing viral transcripts derived from individual promoters in the same muzzle wart tissues. As shown in Fig 5C, we detected two major spliced products from the exonic backward primer Pr5452 (L1) plus the forward primer Pr7140, Pr522 or Pr665 (lanes 2, 6 and 8), the larger amplicon arose from double splicing (either 7243/3139 or 757/3139 and then 3431/5372) and the smaller amplicon arose from single splicing (7243/5372 or 757/5372), depending on the promoter usage. In this case, only a very few products could be amplified from the primer pair of Pr5452 plus Pr116 (lane 4). However, the primer pair of Pr5452 plus Pr522 gave much less amount of the spliced L1 products (lane 6) over the primer pair of Pr5452 plus Pr665 (lane 8), but relatively more than the primer pair of Pr5452 plus Pr116 (lane 4), the data suggest that L1 messages are the spliced product primarily transcribed from two promoters P7107 and P533. But few L1 might be also derived from the promoter P360. The identity of a faint band above 650-bp size from the primer pair of Pr5452 and Pr7140 (lane 2) was unknown and might be nonspecific. When an intronic backward primer Pr4647 in the L2 region was used in combination with the same sets of the forward primers, we obtained only one major product predominantly transcribed from two promoters P7107 and P533 (lanes 10 and 16) and spliced from 7243/3139 (P7107) or 757/3139 (P533), but very little from the promoter P7503 or P360 (lanes 12 and 14), suggesting that these two promoter-derived transcripts are used for L2 production. To detect intron retention in the viral early region required for the expression of E1, the primer-walking RT-PCR results from an exonic backward primer Pr3299 (lanes 19–26) were compared with backward primers Pr2978 (lanes 27–34), Pr1938 (lanes 36–43) or Pr827 (lanes 44–51) in combination of the same sets of primers described above for the detection. We found that the Pr3299 amplified one single spliced RT-PCR product of 7243/3139 for P7107-derived transcripts (lane 19) or 757/3139 for P7503-, P360- or P533-derived transcripts (lanes 21, 23 and 25). The Pr2978 detected a single spliced, weak product of 7243/2493 for P7107-derived transcripts (lane 27) or 757/2493 for P360- (lane 31), but a little more of 757/2493 for P533-derived transcripts (lane 33) and none from the promoter P7503 (lane 29). The Pr1938 detected at low levels an unspliced E1 transcripts derived from the P360 promoter (lane 40), but more from the P533 promoter (lane 42) and none from promoter P7107 (lane 36) or P7503 (lane 38). In contrast, the primer Pr827 in combination of each forward primer led to detection of an E1 intron-containing product mostly from three promoters P7503, P360 and P533 (lanes 46, 48 and 50), but little from P7107 (lane 44). We also confirmed in an MmPV1-induced ear wart of animal #3 that the L1 messages were the spliced products primarily transcribed from the promoter P533, but less from the promoter Pr360 (compare lanes 2 and 4 in Fig 5D with lanes 6 and 8 in Fig 5C). Moreover, the intron-containing E1 transcripts were confirmed to be derived from the promoter Pr7107 and Pr7503 (lanes 6 and 8 in Fig 5D and lanes 44 and 46 in Fig 5C). Further studies using an E1-specific primer Pr1263 for 3’ RACE demonstrated that the unspliced E1 transcripts are polyadenylated at nt 3864 position, an early CS site (Fig 5E).
Based on the mapped TSS, polyadenylation cleavage sites, and RNA splice sites of MmuPV1 early and late transcripts, we constructed a full transcription map from the MmuPV1 genome. As shown in Fig 6. MmuPV1 expresses at least 36 RNA isoforms spanning the entire MmuPV1 genome, with the majority of them being polycistronic transcripts that can potentially translate multiple gene products. MmuPV1 early transcription mainly starts at nt 7503 for E6 polycistronic or nt 360 for E7 polycistronic RNAs. Both are polyadenylated at nt 3864 using a PAS at nt 3844. MmuPV1 late transcription mainly starts either at nt 7107 or nt 533 and polyadenylates either at nt 3864 using the nt 3844 PAS for E1^E4 expression or at nt 7063 using a PAS at nt 7047, the latter encoding L1 and L2 proteins. Most viral early transcripts contain two exons and one major intron spanning the entire E1 ORF and partial E2 ORF. Although the majority of the viral early transcripts have intron 1 spliced out, a small fraction of early transcripts retains this intron (Fig 5B, lane 14, Fig 5C, lanes 40 and 42) and could be further spliced (Fig 4C, lanes 20 and 22; Fig 5B, lane 18 and Fig 5C, lanes 31 and 33). Most viral late transcripts have three exons and two major introns and their 5’ portions overlap with the viral early transcripts. Depending on which late promoter drives transcription, the 5’ end of the first intron can start either at nt 7244 or 758. If starting from nt 758, the first intron of the late transcripts is the same as found in the viral early transcripts. The second intron of the late transcripts is invariably from nt 3432 to nt 5371 and covers the entire L2 ORF and the early PAS. Retention of the intron 2 (Fig 6, products C, D, G, H, Z) and therefore capacity to encode L2 were found only in a small fraction of the late transcripts (Fig 4C, lanes 5 and 14; Fig 5C, lanes 10 and 16). The more abundant viral late transcripts are the ones that encode for E1^E4 and/or L1 (Fig 6, products A, X, AB).
As shown in Fig 6, both MmuPV1 early and late transcripts are alternatively spliced, and the coding capacities of each RNA species may be inferred from the ORF(s) included in the mRNA. We reassigned E4 as E1^E4 with an AUG codon starting from nt 742 position in the viral transcript exon 1 instead of the nt 3101 from a previous publication [4] and thus expression of the E4 ORF requires RNA splicing. We also reassigned the L2 ORF AUG start codon to be at nt 3745 instead of at nt 3735 in the prior publication [4] and the L1 ORF AUG start codon to be at nt 5372 instead of at nt 5291 in the prior publication [4]. To remove the intron 2 during RNA splicing, the viral late transcript exon 2 is spliced right to nt 5372 of the exon 3 and thus the 5291 AUG codon does not exist in the L1 mRNA after RNA splicing. Moreover, we identified the E8^E2 ORF as a spliced ORF (nt 1094-1125/nt 3139–3685) and two small E1 ORF variants, E1^M1 (nt742-1194/nt3139-3231) and E1^M2 (nt742-1125/nt3139-3231) not previously described. A detailed analysis of the upstream sequences of these ORFs indicated that the first AUG codon of each contains a strong Kozak consensus sequence of either ANNaugN or GNNaugG [19,20].
As shown in Figs 1B and 2E, most RNA reads or transcripts appear to be derived from the promoter P533. The relative expression levels of viral transcripts in the infected wart tissues were also measured semi-quantitatively by primer-walking RT-PCR with fewer cycle (25 cycles) amplification (Fig 5C) and the data in this study also suggested that both L1 and L2 are transcribed from either promoter P7107 or P533 (Fig 5C, lanes 2, 8, 10 and 16), but most viral RNA transcripts are P533 transcripts (Fig 5C, compare lane 25 to lanes 23, 21, and 19). To quantify the expression levels of the existing viral transcripts better in the infected tissues, Northern blot analysis using 32P-labeled oligo probes, Pr7237, Pr352, Pr687, Pr3299, Pr3682 or Pr5452 (Fig 7A), was further used to evaluate the quantitative levels of the existing viral RNA transcripts from individual promoters. In this study, the total RNA from mouse ears without MmuPV1 subclinical infection served as a MmuPV1-negative RNA control and was pooled RNA from ears of two naïve, freshly arrived female mice in ~4 months of age, with no detectable MmuPV1 reads by RNA-seq analysis. As shown in Fig 7B, we found that the late-transcript-specific probes Pr7237 (lane 2) and Pr5452 (lane 12) in Northern blotting displayed a comparable expression profile of the viral late region as did from the viral early region detected by the early transcript-specific probes Pr687 (lane 6) and Pr3299 (lane 8), with the more hybridization signals seen from the Pr3299 and Pr5452 than the corresponding comparable probes. Size analyses of individual transcripts detected by each probe demonstrated that L1 transcripts of ~2.3-kb in size are predominantly transcribed from the promoter P7107 (product A in lanes 2, 8 and 12) or P533 (product X in lanes 6, 8 and 12) and are preferentially doubly spliced from nt 7243 to 3139 (P7107) or from nt 757 to 3139 (P533) and then from nt 3431 to 5372, accompanied by a lower abundance of singly spliced L1 transcripts from nt 7243 (P7017, product B in lanes 2 and 12) or nt 757 (P533, product Y in lane 12) to 5372. In principle, a minor form of the doubly spliced L1 transcript Q derived from promoter P360 also exist in these Northern blot assays (lanes 6, 8 and 12). As expected, these doubly spliced L1 transcripts (A/Q/X) were not detectable by the probe Pr3682 (lane 10) hybridizing to a downstream region of the 3431 5’ donor site. A small fraction of 4.2-kb products (product C in lanes 2, 8, 10 and 12; product Z in lanes 8, 10 and 12) are L2 mRNA arising from nt 7243 splicing to nt 3139 in the case of P7107-derived transcripts or from nt 757 splicing to nt 3139 in the case of P533-derived transcripts. The ~1-kb L1 products (lanes 2 and 12 in a question mark) appear to be a product of a cryptic PA usage in the L1 ORF (S5 Fig).
As predicted from RNA-seq analysis in Fig 4A, the majority of the P533-derived RNA transcripts are spliced from nt 757 to 3139, therefore encoding E1^E4, and utilize the CS at nt 3864 for RNA polyadenylation. As shown in Fig 7B, Northern blotting using 32P-labeled oligo probe Pr687 (lane 6) or Pr3299 (lane 8) was able to detect the abundant E1^E4 transcripts (~1.2-kb, products AB) and the minor forms of E6 (~1.4-kb, product M), E7 (~1.1-kb, product T), E2 (~1.8-kb, products L/S), E1 (~4.2-kb, products K or K/P), L1 (~2.3-kb, products Q/X or A/Q/X) and L2 (~4.3-kb, products C and Z). Oligo probe Pr3682 in this assay (Lane 10) displayed the similar detection profile to these two probes except for the doubly spliced L1 which lacks the target sequence for the probe Pr3682 hybridization. Northern blotting using a 32P-labeled oligo probe Pr352 exhibited a predominant E6 RNA (~1.4-kb, product M) transcribed from P7503 as detected by PacBio Iso-seq (Fig 2C and 2D, Table 2) and spliced from nt 757 to 3139 in addition to the E1 RNA of ~3.9-kb (products P) from this promoter (lane 4). The ~0.8-kb products (lanes 8 and 10 in a question mark) appear to be the products of alternative late promoters around nt 576 to 607 (Fig 2C), which are spliced from nt 757 to nt 3139, but polyadenylated at nt 3864 or are doubly spliced and polyadenylated at a cryptic poly(A) site in the L1 ORF (S5 Fig). Thus, our data from the Northern blotting are consistent with the conclusion from RNA-seq analysis (Fig 4A) and semi-quantitative RT-PCR (Fig 5C) that the majority of viral transcripts are spliced products of 757/3139 and are polyadenylated at nt 3864, using an early PAS at nt 3844 for expression of the early region, and the fewer are polyadenylated at nt 7063, using a late PAS at nt 7047 for expression of the late region.
Subsequently, we also examined viral late transcripts in MuPV1-induced papillomas by RNA-ISH (in situ hybridization). Using antisense probes to the E4 and L1 regions (S6A Fig) and the RNAscope methodology which is highly sensitive in detection of both viral RNA and DNA of MmuPV1, we detected E4 and L1 signals primarily in the highly differentiated granular layers of the infected, hyperproliferative ear skin (S6B Fig). These patterns of E1^E4 and L1 expression were also seen in the tail papillomas (S6C Fig). Although these probes in RNAscope technology detect both viral RNA and DNA, pretreatment of tissue sections with DNase and/or RNase allowed us to distinguish between the DNA-derived and RNA-derived signals using this methodology (S6D and S6E Fig). We found that the detected viral E1^E4 transcripts appeared more cytoplasmic distribution than the L1 transcripts did (S6C Fig), particularly after removal of viral genomic DNA by DNase I treatment of the tissue sections (S6D Fig).
To understand the expression dynamics of viral L1 and MmuPV1 DNA over time, nude mice were infected (three spots on the tail) with equivalent amounts of MmuPV1 per site (108 VGE) and tissue was harvested by sacrificing animals at different time points until papillomas were overtly seen (28 days post-infection). These tissues collected at each time point were analyzed for appearance of L1 viral protein and amplified viral DNA, hallmarks of the productive phase of the viral life cycle. As shown in Fig 8A and S7 Fig, no obvious L1 expression or viral DNA was detected in any of the infected sites (0/3 infected sites) at 4 days post-infection. Hypertrophic scarring of terminal epithelia with intact basement membrane was observed. At day 10, both L1 and MmuPV1 DNA were detected in one of the infected sites (that closest to base of tail—shown in Fig 8A). L1 expression was found in the suprabasal layers along with terminally differentiating epithelia whereas MmuPV1 DNA appeared to be present only in suprabasal layers. Microscopically, this site had evidence for hyperplasia, koilocytes and some fibrillary projections; however, the overt appearance was that of a raised scar. At day 21, again one site of infection (shown in Fig 8A), that closest to the tail base, showed evidence of productive infection with an abundance of L1-positive and viral DNA-positive cells and more prominent fibrillary projections. The remaining 2 infected sites appeared mostly normal and no L1 or FISH positive nuclei were detected. At 28 days post infection, all three infected sites showed overt appearance of warts with L1 positivity (S8 Fig). While sites were infected at the same time, papillomas grew asynchronously as is evident by variation in size of the papillomas. We consistently observe that papillomas near the base of the tail grow fastest and show the most robust features of a productive viral infection. The pattern of L1 expression and viral DNA amplification at day 28 was similar to that observed at day 21 (Fig 8A). A similar timing and pattern of detection of viral RNA/DNA species was observed using the RNAscope in situ hybridization methodology with a probe to the L1 region (Fig 8B).
Next, we attempted to express each of the predicted MmuPV1 proteins based upon the transcript map in HEK293 and HeLa cells transiently transfected with individual FLAG-tagged viral ORF cDNA expression vectors, including MmuPV1 ORF E6 (GenBank Accession #MF350298), E7 (GenBank Accession #MF350299), E1 (GenBank Accession #MF350300), E2 (GenBank Accession #MF350301), L2 (GenBank Accession #MF350302), L1 (GenBank Accession #MF350303), E1^E4 (GenBank Accession #MF350304), E8^E2 (GenBank Accession #MF350305), E1^M1 (GenBank Accession #MF350306) and E1^M2 (GenBank Accession #MF350307) (Fig 9A and 9B). As shown in Fig 9C, we were able to detect the expression of E2 (lane 2), E1^E4 (lane 3), E1^M1 (lane 4), E1^M2 (lane 5), E7 (lane 9) and E8^E2 (lane 10) in HEK293 cells, but were unable to detect E1 (lane 1), L1 (lane 6), L2 (lane 7) and E6 (lane 8) by FLAG-specific immunoblot analysis. However, both viral E6 and E7 could be better detected when HEK293 cells were treated with proteasome inhibitor MG132 (Fig 9D), indicating they are likely degraded via the proteasome. We also failed to express E1, L1 and L2 in mouse epithelial keratinocytes (S9A Fig) and HEK293TT or HEK293FT cells. By Northern blot analysis of the total RNA extracted from transfected HEK293 cells, we were unable to detect both L1 and L2 RNA, but identified two spliced E1 RNA in smaller sizes and all other expected sizes of viral ORF-derived RNA transcripts (S9B Fig). Using FLAG-specific immunofluorescence (Fig 9E), we detected the expression of E6, E7, E2, and E8^E2 mainly in the nucleus of HeLa cells, E1^E4 as a filamentous protein in the cytoplasm, E1^M1 either in the cytoplasm or nucleus or both, and E1^M2 primarily in the cytoplasm. Distribution of viral E6 and E7 in the cells could be only slightly altered in the presence of MG132 (Fig 9E).
Understanding of the structure and coding capacity of transcripts is critical for disclosing genome function and biology of any organism. In this report, we have utilized two cutting-edge technologies, RNA-seq and PacBio Iso-seq, in combination with various conventional technologies to analyze the structure and expression of MmuPV1 genome in MmuPV1-induced wart tissues with productive MmuPV1 infection. We have constructed the first full transcription map for MmuPV1 and demonstrated that MmuPV1 genome encodes ten ORFs and utilizes five major promoters, two polyadenylation sites and eight splice sites for its expression of thirty-six RNA isoforms during virus infection. Similar to other papillomaviruses [8,10,21–28], this nature of the genome structure with alternative usage of promoters, pA sites and splice sites empowers a highly compact viral genome to express multiple gene products in a temporally and spatially organized manner within its viral life cycle.
The delineated MmuPV1 genome structure and expression of MmuPV1 is more close to that of bovine papillomavirus type 1 (BPV-1), cottontail rabbit papillomavirus (CRPV), cutaneous HPVs and some low-risk mucosotropic HPVs. Like all other papillomaviruses [7,8], the MmuPV1 genome transcribes unidirectionally from one DNA strand, with ~99% of RNA-seq reads mapping to viral sense transcripts. Similar to cutaneous HPVs and low-risk mucosotropic HPVs [25,29,30], MmuPV1 employs two separate early promoters for expression of viral E6 and E7, the promoter P7503 for the E6 expression and the promoter P360 for the E7 expression. These two promoters could be also responsible for expression of other viral early proteins, including E1, E1^M1, E1^M2, E2 and E8^E2. This strategy for expression of MmuPV1 E6 and E7 is different from high-risk HPVs in that their expression of both viral E6 and E7 is exerted by a single early promoter upstream of the E6 ORF [10,31,32] and the expression of viral E7 requires RNA splicing of an E6 intron in this early transcript [19,33,34]. Also similar to low-risk HPVs [25,29,30] and other animal papillomaviruses [23,24], MmuPV1 has no intron in the E6 ORF region. We found that the promoter P7503 is the only early promoter bearing a classical TATA box (a eukaryotic core promoter motif for binding of RNA polymerase II) 25-nts upstream of the promoter TSS. Based on our 5’ RACE and PacBio Iso-seq data, we conclude that the P7503 is stronger than the other two early promoters, P360 and P859. The third early promoter P859 is a weak promoter most likely driving the expression of E2 and E8^E2. Because of anticipated RNA-seq bias on eukaryotic RNA 5’ ends [13], fewer RNA reads next to the P7503 and P360 promoters were noticed (Fig 4A).
MmuPV1 transcribes its late transcripts from two late promoters, P7107 in the LCR region downstream of L1 ORF and P533 in the E7 ORF region. Utilization of a late promoter in the LCR region for the expression of viral late gene L1 and L2 is a characteristic feature for BPV-1 [22], CRPV [24] and some skin-tropic HPVs such as HPV-1 [35,36] and HPV-5 [29], but not for Mastomys natalensis papillomavirus (MnPV) [23] and other HPVs [10,30–32]. In contrast, high-risk HPVs express their late genes mainly from a late promoter in the E7 ORF [10,31,32,37]. Similar to high-risk HPVs, the late promoter P533 in the MmuPV1 genome is most likely responsible for E1^E4 expression, but also for L1 and L2 expression. Although the transcripts derived from promoter P360 or P859 might have the potential to encode L1, they were scarcely detectable from the infected tissues and could be negligible. We found that the P7107 has a TATA-like box 57-nts upstream of its TSS and the P533 bears a TATA box at 110-nts upstream of its TSS.
Posttranscriptional RNA processing, including RNA capping, splicing, polyadenylation and export, provides multiple layers of regulation to guide efficient expression of eukaryotic genes [38,39]. By using 3′ RACE, we mapped the cleavage sites of both viral early and late transcripts for RNA polyadenylation and demonstrated that viral early transcripts are polyadenylated primarily at nt 3864 by using a PAS at nt 3844 and viral late transcripts are polyadenylated primarily at nt 7063 by using a PAS at nt 7047. In addition, we have identified a few late transcripts being polyadenylated from nt 5627 by using a PAS at nt 5609 in the L1 ORF. Analyses of the sequences 3’ downstream of each mapped CS site for a highly conserved recognition site U/GU for CSF (cleavage stimulation factor) binding in RNA polyadenylation [16,17] showed three U/GU motifs in this region of the nt 7063, but not so for the nt 3864. Thus, what motif guides the polyadenylation cleavage of the nt 3864 remains unknown. Nevertheless, the mapped polyadenylation site usage for the expression of MmuPV1 early and late transcripts resembles to that of all papillomaviruses [8,10,22–24,29,30,32].
Extensive alternative RNA splicing contributes to the expression of multiple genes by papillomaviruses [8,10,22–24,29,30,32]. Our study revealed this feature in MmuPV1 gene expression by analyzing exon-exon splice-junction reads from RNA-seq and by primer-walking RT-PCR analyses of RNA extracted from MmuPV1-induced warts. We demonstrated that MmuPV1 employs five 5’ splice sites (donor sites) and three 3’ splice sites (acceptor sites) for expression of both viral early and late genes from its five promoters and produces at least thirty-six different RNA isoforms by alternative RNA splicing, of which thirteen were detectable from wart tissues by less sensitive Northern blotting. Thus, the primary MmuPV1 transcripts could have 2 or 3 exons and 1 or 2 introns, with viral E1 and L2 residing in the most commonly excised introns, as seen in other papillomaviruses. As with other papillomaviruses, a small fraction of MmuPV1 RNA transcripts does retain the capacity to express viral E1 and L2; the mechanism by which these RNA transcripts retain the introns for expression of the E1 or L2 should be an attractive area for future investigation.
By construction of this full transcription map, we conclude that MmuPV1 has the potential to express 10 gene products: E6, E7, E1, E1^M1, E1^M2, E1^E4, E2, E8^E2, L2 and L1. Like cutaneous HPVs [40], MmuPV1 does not contain an E5 ORF and therefore no E5 gene product is predicted. The coding regions for the E1^E4, L2 and L1 gene products have been reassigned from what was originally predicted based solely on viral DNA sequence information [4]. E1^M1, E1^M2 and E8^E2 have not been described before for MmuPV1. The E1^M1 and E1^M2 in MmuPV1 might be similar to the E1Ma and E1M in HPV-11 [25]. The MmuPV1 E8^E2 gene product is likely similar to the E8^E2 gene products characterized in BPV-1 [41–44] and HPVs [45,46] and predicted to be encoded by most papillomaviruses [47]. Although the first AUG codon for each of the predicted coding regions has a strong Kozak sequence [20], we were unable to detect the expression of MmuPV1 E1, L2 and L1 proteins from a common eukaryotic expression vector in HEK293, MEK, 293FT or 293TT cells and unable to detect by Northern blotting a full-length RNA expressed from E1, L2 and L1 vector in HEK293 cells. Several possible reasons for this, based upon studies of other papillomaviruses, include rare codon usage [48–50], RNA splicing or instability [8,51,52] and protein stability [53–56]. For example, the E1 ORF in the expression vector contains three splice donor sites (nt 757, nt 1125 and nt 1194 splice donor sites) and one acceptor site (nt 2493 splice acceptor site). The RNA expressed from this vector in HEK293 cells that we could detect by Northern analysis was all spliced forms; no unspliced mRNA capable of expressing full-length E1 could be detected. Among seven viral proteins detected, MmuPV1 E6 and E7 were also found to be increased in their steady state by using a proteasome inhibitor, suggestive that they are normally subjected to proteasomal degradation as seen for HPV-16 [53].
Several lines of evidence indicate that MmuPV1 shares a number of genomic, molecular and pathological features with high-risk cutaneous, beta HPVs and thus could be a useful model to study these important human pathogens. These include our observation that MmuPV1 contains separate promoters for E6 and E7, and the additional facts that MmuPV1 causes squamous cell carcinomas at cutaneous sites [57], lacks an E5 ORF [4] and encodes an E6 protein that shares with HPV8 E6 the ability to bind MAML1 and SMAD2/SMAD3 but not E6AP and p53 [58].
In this study, we also performed a careful assessment of the timing of onset of viral gene expression and productive amplification of viral DNA in the context of emerging warts. These studies used a fixed amount of virus (108 VGE per infection site) applied to three scarified sites on the tail of each nude mouse. We found that both viral late gene expression (L1 RNA and capsid) and viral DNA amplification could be observed in differentiated cell compartment as early as 10 days post infection, well before overt warts can be observed, which starts at 4 weeks with this dose of virus and this strain of mouse. We observed variability in the onset of detectable infection, be it at a microscopic or overt level, with the some infection sites not showing any microscopic evidence of infection until the 4-week time point when overt warts first appeared. Interestingly, we observed a reproducible spatial pattern in which sites of infection at the base of the tail gave rise to faster growing warts than sites infected at the tip of the tail. The reason for this is unknown, but some possibilities may include blood flow, temperature, and grooming behavior.
Another feature of the time course study is that L1 positive cells were throughout the epithelium, including basal cells in mature warts harvested at 3 or 6 months post-infection, an uncommon feature of MmuPV1 observed by others [6]. However, the L1-positive cells were restricted to differentiated cells in the early time points, out to day 28 post-infection (Fig 8A), suggesting that the complete nature of the viral life cycle is realized at times later than 4 weeks post-infection. Interestingly, canine oral papillomavirus (COPV) [59] is also found to be amplified in basal epithelial cells, though the timing at which this first appears is slightly different. COPV, while a mucosal papillomavirus, is closely related to cutaneous HPVs, HPV-1 and HPV-63, that cause plantar warts [60,61]. Basal cells have been seen to support viral DNA amplification in lesions caused by HPV-1 and HPV-63 [62].
Lastly, we made the interesting observation that in experimentally infected nude mice that have developed warts at the sites of infection, other areas of the epidermis can show evidence for subclinical infections. These subclinical infections showed the same distribution pattern of viral RNA-seq reads as the experimentally infected sites with warts when the reads mapped to the reference MmuPV1 genome (Fig 1B), and showed evidence for productive infection, based upon the detection of viral DNA amplification and L1 expression (S1B Fig). This raises the intriguing possibility that subclinical infections may be common in immunodeficient or immunosuppressed contexts. In this regard, organ transplant patients are known to have an increased abundance of HPV DNA in randomly sampled hair follicles from clinically normal skin [63].
In conclusion, we observe a similar transcription pattern for MmuPV1 as observed with animal papillomaviruses and some HPVs. We believe this carefully mapped landscape of MmuPV1 transcription from MmuPV1-induced warts will provide a solid foundation for future understanding of MmuPV1 molecular biology, pathogenesis and immunology.
Immunodeficient athymic BALB/c FoxN1nu/nu used in this study were obtained from Harlan (currently Envigo, Indianapolis, IN). All infected mice (6–8 weeks old at the time of infection) were housed in aseptic conditions in micro-isolator cages. Animals were handled only by designated personnel and personal protection gear was changed between cages to prevent any virus cross-contamination. Experimental infection was performed using quasivirions containing MmuPV1 synthetic genome as described previously [57,64,65]. The synthetic MmuPV1 genome (Gift from Dr. Chris Buck, NCI) is identical to the original wild type genome and has been described previously [6]. Briefly, 293FT cells (Thermo Fisher Scientific, Waltham, MA), a fast growing 293T cell line, were co-transfected with a codon optimized MmuPV1 capsid protein expression plasmid (pMusSheLL, gift from Dr. Chris Buck, NCI) [6,66] and recircularized MmuPV1 synthetic genome for encapsidation. The cells were harvested 48 h after cotranscfection and virions were purified using Optiprep (Sigma-Aldrich, St. Louis, MO) gradient centrifugation. The generated quasivirions were quantified by estimating viral genome equivalents (VGE) by comparing the amount of encapsidated viral DNA in the viral stock by Southern blot analysis using MmuPV1-specific probes, followed by quantification using ImageJ software as described previously [57]. BALB/c FoxN1nu/nu mice (6–8 weeks old) were infected with 2×108 VGE MmuPV1 per site after scarifying skin of tail, ear or muzzle as described previously [57,58]. The wart tissues from three anatomical sites (ear, tail and muzzle) were collected from each animal 6 months post-infection and snap-frozen in liquid nitrogen for RNA isolation and a portion of the papillomas was excised, fixed in 10% neutral buffered formalin and embedded in paraffin. Serial sections (5 μm thick) were stained with hematoxylin and eosin (H&E) and evaluated for histopathological features and processed for subsequent analyses.
All animal experiments were performed in full compliance with standards outlined in the "Guide for the Care and Use of Laboratory Animals” by the Laboratory Animal Resources (LAR) as specified by the Animal Welfare Act (AWA) and Office of Laboratory Animal Welfare (OLAW) and approved by the Governing Board of the National Research Council (NRC). Mice were housed at McArdle Laboratory Animal Care Unit in strict accordance with guidelines approved by the Association for Assessment of Laboratory Animal Care (AALAC), at the University of Wisconsin Medical School. All protocols for animal work were approved by the University of Wisconsin Medical School Institutional Animal Care and Use Committee (IACUC, Protocol number: M02478).
A home-based tyramide-based signal amplification (TSA) method was developed as described previously to detect MmuPV1 L1 [67,68]. Formalin-fixed paraffin embedded tissue slides were deparaffinized after 3 changes of xylene followed by rehydration in ethanol series (100%, 95%, 70%, 50% and finally double distilled water). Endogenous peroxidase activity was blocked using 0.3% hydrogen peroxide in methanol. Antigen retrieval was performed using antigen retrieval buffer (pH = 9.0, Abcam, Cambridge, MA, #ab93684) for 20 minutes in a microwave. Slides were cooled to room temperature and blocked for 1 h at room temperature in blocking buffer (Perkin Elmer, Fermont, CA, #FP1012). Rabbit sera against MmuPV1 L1 (Gift from Dr. Chris Buck, NIH) [6,69] was diluted at 1:5000 in blocking buffer and applied to sections overnight at 4°C. Samples were incubated with goat anti-rabbit-HRP secondary antibody (at 1:500 dilution) in blocking buffer for 1 h at room temperature. Subsequently, the secondary antibody was biotinylated by incubating with biotin-tyramide (10 μg/ml) for ten minutes as described previously [68]. Slides were rinsed with PBS (phosphate-buffered saline) containing 0.1% Tween-20 and a cytokeratin cocktail containing equal amounts of anti-K14 (BioLegend, San Diego, CA, #PRB-155P) and anti-K10 (BioLegend, #PRB-159P) at 1:1000 dilution was applied at room temperature for 1 hour. Slides were rinsed with PBS containing 0.1% Tween-20 and incubated with secondary detection reagents as follows—anti-rabbit conjugated with Alexa Fluor488 (Thermo Fisher Scientific, #A11008) at 1:500 to detect K10 and K14, and Streptavidin-Alexa Fluor-594 (Thermo Fisher Scientific #S-32356) at 1:500 to detect biotinylated L1 for 1 hour at room temperature. Tissues were counter stained with Hoechst for cellular DNA and coverslips were mounted using ProLong gold antifade (Thermo Fisher Scientific, #P36930).
MmuPV1 DNA FISH was performed as described previously [57,70]. This protocol has been adapted from a DNA FISH protocol used to detect Epstein Barr Virus (EBV) DNA in monolayer cells and is described in detail at: https://mcardle.oncology.wisc.edu/sugden/protocols.html. Briefly, formalin-fixed paraffin embedded tissue slides were baked at 65°C overnight and deparaffinized using xylene followed by treatment with 100% ethanol. Slides were then boiled in 10 mM sodium citrate buffer (pH = 6.0) for 30 minutes in a microwave. Slides were rinsed with PBS and completely dried before pre-hybridizing with 2 x SCC containing RNase A and 0.5% IPEGAL (pH = 7.0) for 30 min at 37°C. Slides were dehydrated using a series of ice cold ethanol (70%, 80%, 95%) for 2 min each. Slides were dried by placing them in an empty container at 50°C for 5 min and then placed in denaturation solution [28 ml formamide, 4 ml 20 x SSC (pH = 5.3) and 8 ml water] at 72°C for 2 min. The ethanol series was repeated again, and after drying the sections, denatured probe was added to the slides. A biotin-16-dUTP (Sigma-Aldrich, #11093070910) labeled probe was hybridized to tissue overnight at 37°C in a humidified chamber. To make the probe, nick translation was used to label the entire MmuPV1 plasmid DNA (pMusPV [6]) with biotin. Slides were then washed twice for 30 min with 2 x SSC and 50% formamide at 50°C followed by two washes for 30 min with 2 x SSC at 50°C. Signals were detected with streptavidin conjugated to Cyanine-3 (Sigma-Aldrich, #S6402) at 1% by volume in STM solution (4 x SSC, 5% non-fat dried milk, 0.05% Tween-20, 0.002% sodium azide) for 30 min at 37°C. Nuclei were counterstained with Hoechst coverslips were mounted using ProLong gold antifade (Thermo Fisher Scientific, #P36930).
High resolution wide-field fluorescent images were acquired by means of a super-resolution Leica SP8 STED confocal microscope equipped with a motorized stage. This microscope is equipped with PMT and HyD lasers. All images were taken by means of a 20X objective lens (Specifications: HC PL APO 20x/0.75 CS2, Dry). The images were acquired by tile-scanning by marking positions around the region of interest on the LAS-X suite (version: 2.0.1). The merged wide-field image was obtained by automatic stitching of individual styles by means of in-built auto stitching algorithm part of the LAS-X suite. All other images for tissue analyses were captured using a Zeiss AxioImager M2 microscope and AxioVision software version 4.8.2 (Jena, Germany).
The tissues were homogenized in TriPure reagent (Roche, Indianapolis, IN, #11667165001). Total RNA was extracted according to TriPure extraction protocol and treated with TURBO DNA-free Kit (Thermo Fisher Scientific, Waltham, MA, AM1907) to eliminate all traces of viral DNA. The RNA concentration and integrity were assessed by Bioanalyzer 2100 (Agilent, Santa Clara, CA). After removal of ribosomal RNA, the total RNA sequence libraries were prepared using Illumina Stranded Total RNA (Illumina, RS-122-2201, San Diego, CA) protocol with TruSeq V4 chemistry and sequenced in the Sequencing Facility of NCI on Illumina HiSeq 2500 with 2×125 nts modality and depth of 100 million reads per sample. The obtained reads were trimmed of adapters and low-quality bases and aligned to MmuPV1 reference genome (NC_014326; GI:301173443) with start site at nt 7088 using STAR aligner package [18,71]. This arrangement makes the linear MmuPV1 to end at nt 7087, approximately 186 nts from L1 stop codon. Thus, all nucleotide positions described in this report refer to the reference genome sequence (GenBank Acc. # NC_014326) [4]. Integrative Genomics Viewer (IGV, Broad Institute) program was used to visualize MmuPV1 reads coverage. The data discussed in this publication have been deposited in NCBI’s Gene Expression Omnibus and are accessible through GEO Series accession number GSE104118 (www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE104118). Additional criteria were used to identify the real splice junctions extracted by STAR aligner: (1) a threshold of median spliced read alignment overhang >10 nt; (2) the number of uniquely mapping reads crossing the junction >50 to ensure filtering out sporadic false junctions; (3) entropy of overhang length distribution >1 to filter out junctions with unevenly distributed overhangs. A Sashimi plot for splice junction visualization was generated by IGV.
The 5′ and 3′ RACE assays were carried out using a Smart RACE cDNA amplification kit (Clontech, Mountain View, CA, #634858) according to the manufacturer’s instructions using 1 μg/reaction of total RNA as template [9]. The primers used in the assays are in S1 Table (see supplemental materials). The final PCR products were gel purified, cloned into pCR2.1-TOPO vector (Thermo Fisher Scientific) and sequenced by Sanger sequencing (Macrogen USA, Rockville, MD). To obtain the comprehensive coverage of viral TSS, the 5’ RACE products obtained by Pr3299 and Pr5452 were subjected to single molecule, real-time sequencing using PacBio Iso-seq technology (Pacific Biosciences, Menlo Park, CA). The mouse papillomavirus amplicons produced using a 5’ RACE method were used as the input for PacBio Iso-seq sequencing. Using the SMRTbell Template Prep Kit 1.0 (Pacific Biosciences), the cDNA underwent damage repair, A-tailing, and A/T hairpin adaptor ligation. Following adaptor ligation, the libraries were digested with Exonuclease III and VII to remove non-ligated and nicked SMRTbell molecules. AMPure PB beads (Pacific Biosciences) were used at a 0.6 x ratio to clean up all enzymatic reactions throughout library construction. The DNA/Polymerase Binding Kit P6 v2 (Pacific Biosciences) was used for annealing the sequencing primer and binding the polymerase to the final libraries. The MagBead loading Kit (Pacific Biosciences) was used for loading the polymerase-bound SMRTbell molecules onto the SMRT Cell. Each SMRT Cell was sequenced for six hours using the DNA Sequencing Reagent Kit 4.0 v2 (Pacific Biosciences) on a PacBio RS II sequencer. The obtained sequence reads were trimmed of adaptors and only the reads containing a specific adaptor at 5’ end introduced by 5’RACE were considered the full-length and used for mapping to MmuPV1 reference genome in further analysis and visualized in the IGV program. The position of 5’ends of the full-length reads were extracted and quantified in promotor usage analysis.
PacBio SMRT Analysis Package (smrtpipe.py v1.87, with default settings) was used to process raw data into circular consensus sequence (CCS) for further analyses. First, the raw data was processed into error corrected reads of insert (ROI’s) by RS_Read Of Interest (ROI).1 protocol provided in the SMRT Analysis Package. The ROI’s were then processed using the Classify module with default parameters to remove adapter sequences, poly (A) tails, artificial chimeras, and 3′ truncated transcript sequences which resulted into full-length non-chimeric (FLNC) reads by RS_IsoSeq.1 protocol. For further analyses, we mapped the CCS reads and FLNC reads into the MmuPV1 genome and identified the transcription start sites by BLAT (with-fine option otherwise default options) [72]. Additional computational analyses were performed with Python (version 3.5, https://www.python.org/).
To remove contaminated genomic DNA, the total RNA was treated with TURBO DNA-free Kit (Thermo Fisher Scientific). Reverse transcription (RT) was performed with the SuperScript II kit (Thermo Fisher Scientific, #11904–018). Amplification of reversed transcribed cDNA was performed by PCR using the Platinum SuperFi Taq Polymerase Kit (Thermo Fisher Scientific, #12351–050) according to the manufacturer’s protocols. The MmuPV1-specific primers (S1 Table, see supplemental material) were used to detect viral transcripts. GAPDH RNA served as a sample loading control by using a mouse GAPDH-specific primer pair (forward oMA1, 5’-ATGTTCCAGTATGACTCCAC-3’ and backward oMA2, 5’-TGACAATCTTGAGTGAGTTG-3’). All PCR amplifications were performed under same conditions: on a primary denaturation step at 94°C for 2 min, followed by 25 or 35 cycles of 30 sec at 94°C, 45 sec at 55°C and 60 sec at 72°C, and final extension for 10 min at 72°C.
Total RNA used for Northern blot analysis was isolated from mouse ears with or without MmuPV1 infection or extracted from regular HEK293 cells (ATCC, Manassas, VA) transfected with individual MmuPV1 ORF expression vectors. Total RNA from mouse ears without MmuPV1 subclinical infection served as a MmuPV1-negative RNA control and was pooled RNA isolated from ears of two naïve, freshly arrived female mice from Harlan lab in ~4 months of age, with no detectable MmuPV1 reads by RNA-seq analysis. In general, total 5 μg of RNA from each sample was mixed with NorthernMax Formaldehyde loading dye (Thermo Fisher Scientific, #AM8552)) and denatured at 75°C for 15 min. The RNA samples were then separated in 1% (wt/vol) formaldehyde-containing agarose gels in 1× morpholinepropanesulfonic acid (MOPS) running buffer, transferred onto a GeneScreen Plus hybridization transfer membrane (Perkin Elmer, Waltham, MA, #NEF987001PK) and UV light crosslinked and stained by ethidium bromide for 18S ribosome level as an internal loading control. The membrane was then prehybridized with PerfectHyb Plus hybridization buffer (Sigma-Aldrich, #H7003) for 2 h at 42°C followed by overnight hybridization with a MmuPV1-specific oligo probe as described [73]. MmuPV1-specific oligo probes and a U6-specific probe (oST 197, 5’-AAAATATGGAACGCTTCACGA-3’) were prepared by end-labeling of antisense oligos (S1 Table) with γ-32P using T4 PNK (Thermo Fisher Scientific, #18004–010). After hybridization, the membrane was washed once with a 2× SSPE (1× SSPE: 0.18 M NaCl, 10 mM NaH2PO4 and 1 mM EDTA [pH 7.7])-0.1% SDS solution for 5 min at room temperature and twice with 0.1× SSPE-0.1% SDS for 15 min at 42° and then exposed to a PhosphorImager screen and X-ray film.
The radioactive RNA probes were prepared by in vitro transcription in the presence of [α-32P]CTP with Riboprobe System-T7 (Promega, Madison, WI, #P1446), using PCR products with a built-in T7 promoter as DNA templates. The following primers were used for MmuPV-1 DNA template preparation: oXYX-12 and oXYX-23 (S1 Table). The RNase protection assay (RPA) was performed with an RPA III kit (Ambion, Austin, TX, #1414) according to the manufacturer’s instructions with minor modifications. Briefly, 4 ng of each probe (specific activity, 35,000 cpm/ng) was hybridized overnight at 50°C with 30 μg of total tissue RNA in hybridization buffer and then digested with an RNase A-T1 mixture for 30 min at 37°C. Five micrograms of yeast RNA was used as a negative control (MmuPV1 -). Protected RNA fragments were separated in a denaturing 8% polyacrylamide gel containing 8 M urea. MmuPV-1 DNA sequencing ladders generated with the 32P-labeled Primer Pr7237 (oXYX-28) (S1 Table) were used as size markers and run along with the RPA products as described [73]. Autoradiographic data were captured with a Typhoon Imaging System (GE Healthcare Life Sciences, Pittsburgh, PA) and analyzed with ImageQuant software (GE Healthcare Life Sciences).
For in situ detection of viral transcripts, the tissues were fixed in 10% neutral buffer formalin for 20 h at room temperature, dehydrated, and embedded in paraffin. The sections were cut into 5 μm slides and subjected to RNA-ISH using RNAscope technology (Advanced Cell Diagnostics, Newark, CA) as recommended by manufacturer. Two custom designed probes derived from MmuPV1 genome were used: E1^E4 (nt 3139–3419) and L1 (nt 5372–6901). The signal was detected by colorimetric staining using RNAscope 2.5 HD Assay—BROWN followed by hematoxylin Gill’s No. 1 solution (Sigma-Aldrich, #GHS116) counterstaining. The slides were dehydrated, mounted in Cytoseal XYL (Thermo Scientific, #8312–4), and scanned at 40× resolution using Aperio CS2 Digital Pathology Scanner (Leica Biosystem, Buffalo Grove, IL). To distinguish viral RNA signal from viral genomic DNA signal, the MmuPV1-infected tissue sections with or without pre-treatment with DNase I or both DNase I and RNase A/T1 were compared in parallel in the RNAscope assays. To carry out DNase I or RNase treatment, all tissue sections after rehydration were digested first with RNAscope Protease Plus for 30 min and then followed by 20 units of DNase I (Thermo Fisher Scientific, cat. No. #EN0521) diluted in 1 x reaction buffer with MgCl2 for 30 min at 40°C or by 20 units of DNase I and 500 ug of RNase A (Qiagen, #1006657) plus 2000 units of RNase T1 (Fermentas, Waltham, MA, #EN0542) diluted in 1 x reaction buffer with MgCl2 for 30 min at 40°C.
To express viral proteins, the cDNAs of individual ORF under the optimized Kozak context were amplified by RT-PCR from total RNA isolated from infected tissues and cloned into pFLAG-CMV-5.1 (Sigma-Aldrich) vector in frame with a C-terminal FLAG tag. The obtained plasmid DNA (2 μg) was utilized to transfect HEK293 cells (2.5 × 105) plated in a 12-well plate using LipoD293 transfection reagent (SignaGen Laboratories, Rockville, MD, #SL100668). In some cases the cells were treated 24 h after transfection with 10 μM proteasome inhibitor MG132 (Sigma-Aldrich, #474790) for 6 h. The primary mouse keratinocytes were cultivated as described [74] in the presence of Rho kinase inhibitor Y-27632 (Enzo Life Sciences, Farmingdale, NY, #ALX-270-333). The mouse keratinocytes (1 × 105) were transfected with 1μg of plasmid DNA using Amaxa P3 Primary Cells 4D Nucleofector X Kit S (Lonza, Walkersville, MD, #V4XP-3032) and program DS-138 as recommended by manufacturer. After transfection, the keratinocytes were plated into 24-well plate containing mitomycin-treated feeder 3T3 cells and incubated for 45 hours in the absence of Rho kinase inhibitor.
Total protein extracts and total RNA were prepared 24 h after transfection of HEK293 cells and the expressed individual viral proteins were determined by Western blotting with a rabbit polyclonal anti-FLAG antibody (Sigma-Aldrich, #F7425). Total RNA was resolved on a 1% formaldehyde-agarose gel, stained by ethidium bromide for 18S ribosomal RNA as a sample loading control, and examined by Northern blotting for individual viral gene transcripts expressed from the transfected plasmid by a γ-32P-labeled oligo probe (oVM79, 5’-GGGCACTGGAGTGGCAAC-3’) which hybridizes to a common 3’ UTR region downstream of the FLAG-tag, but upstream of the poly(A) site. U6 snRNA served as a loading control and was detected using a γ-32P-labeled, U6-specific oligo probe oST197.
HeLa cells (2.5 × 105, ATCC) growing on the coverslips were transfected with each vector (0.5–1 μg) encoding a FLAG-tagged viral protein by using LipoD293 transfection reagent (SignaGen Laboratories). The cells at 24 h after transfection were fixed, permeabilized, and stained with a monoclonal anti-FLAG M2 antibody (Sigma-Aldrich, # F1804) in combination with Alexa Flour488-labeled anti-mouse secondary antibody (Thermo Fisher Scientific, #A11029) as described before [75,76]. The cell nuclei were counterstained by Hoechst 33342 dye (Thermo Fisher Scientific, #H3570).
|
10.1371/journal.pgen.1003572 | A Six Months Exercise Intervention Influences the Genome-wide DNA Methylation Pattern in Human Adipose Tissue | Epigenetic mechanisms are implicated in gene regulation and the development of different diseases. The epigenome differs between cell types and has until now only been characterized for a few human tissues. Environmental factors potentially alter the epigenome. Here we describe the genome-wide pattern of DNA methylation in human adipose tissue from 23 healthy men, with a previous low level of physical activity, before and after a six months exercise intervention. We also investigate the differences in adipose tissue DNA methylation between 31 individuals with or without a family history of type 2 diabetes. DNA methylation was analyzed using Infinium HumanMethylation450 BeadChip, an array containing 485,577 probes covering 99% RefSeq genes. Global DNA methylation changed and 17,975 individual CpG sites in 7,663 unique genes showed altered levels of DNA methylation after the exercise intervention (q<0.05). Differential mRNA expression was present in 1/3 of gene regions with altered DNA methylation, including RALBP1, HDAC4 and NCOR2 (q<0.05). Using a luciferase assay, we could show that increased DNA methylation in vitro of the RALBP1 promoter suppressed the transcriptional activity (p = 0.03). Moreover, 18 obesity and 21 type 2 diabetes candidate genes had CpG sites with differences in adipose tissue DNA methylation in response to exercise (q<0.05), including TCF7L2 (6 CpG sites) and KCNQ1 (10 CpG sites). A simultaneous change in mRNA expression was seen for 6 of those genes. To understand if genes that exhibit differential DNA methylation and mRNA expression in human adipose tissue in vivo affect adipocyte metabolism, we silenced Hdac4 and Ncor2 respectively in 3T3-L1 adipocytes, which resulted in increased lipogenesis both in the basal and insulin stimulated state. In conclusion, exercise induces genome-wide changes in DNA methylation in human adipose tissue, potentially affecting adipocyte metabolism.
| Given the important role of epigenetics in gene regulation and disease development, we here present the genome-wide DNA methylation pattern of 476,753 CpG sites in adipose tissue obtained from healthy men. Since environmental factors potentially change metabolism through epigenetic modifications, we examined if a six months exercise intervention alters the DNA methylation pattern as well as gene expression in human adipose tissue. Our results show that global DNA methylation changes and 17,975 individual CpG sites alter the levels of DNA methylation in response to exercise. We also found differential DNA methylation of 39 candidate genes for obesity and type 2 diabetes in human adipose tissue after exercise. Additionally, we provide functional proof that genes, which exhibit both differential DNA methylation and gene expression in human adipose tissue in response to exercise, influence adipocyte metabolism. Together, this study provides the first detailed map of the genome-wide DNA methylation pattern in human adipose tissue and links exercise to altered adipose tissue DNA methylation, potentially affecting adipocyte metabolism.
| A sedentary lifestyle, a poor diet and new technologies that reduce physical activity cause health problems worldwide, as reduced energy expenditure together with increased energy intake lead to weight gain and increased cardiometabolic health risks [1]. Obesity is an important predictor for the development of both type 2 diabetes (T2D) and cardiovascular diseases, which suggests a central role for adipose tissue in the development of these conditions [2]. Adipose tissue is an endocrine organ affecting many metabolic pathways, contributing to total glucose homeostasis [2]. T2D is caused by a complex interplay of genetic and lifestyle factors [3], and a family history of T2D has been associated with reduced physical fitness and an increased risk of the disease [4]–[6]. Individuals with high risk of developing T2D strongly benefit from non-pharmacological interventions, involving diet and exercise [7], [8]. Exercise is important for physical health, including weight maintenance and its beneficial effects on triglycerides, cholesterol and blood pressure, suggestively by activating a complex program of transcriptional changes in target tissues.
Epigenetic mechanisms such as DNA methylation are considered to be important in phenotype transmission and the development of different diseases [9]. The epigenetic pattern is mainly established early in life and thereafter maintained in differentiated cells, but age-dependent alterations still have the potential to modulate gene expression and translate environmental factors into phenotypic traits [10]–[13]. In differentiated mammalian cells, DNA methylation usually occurs in the context of CG dinucleotides (CpGs) and is associated with gene repression [14]. Changes in epigenetic profiles are more common than genetic mutations and may occur in response to environmental, behavioural, psychological and pathological stimuli [15]. Furthermore, genetic variation not associated with a phenotype could nonetheless affect the extent of variability of that phenotype through epigenetic mechanisms, such as DNA methylation. It is not known whether epigenetic modifications contribute to the cause or transmission of T2D between generations. Recent studies in human skeletal muscle and pancreatic islets point towards the involvement of epigenetic modifications in the regulation of genes important for glucose metabolism and the pathogenesis of T2D [11], [12], [16]–[21]. However, there is limited information about the regulation of the epigenome in human adipose tissue [22].
The mechanisms behind the long-lasting effects of regular exercise are not fully understood, and most studies have focused on cellular and molecular changes in skeletal muscle. Recently, a global study of DNA methylation in human skeletal muscle showed changes in the epigenetic pattern in response to long-term exercise [23]. The aims of this study were to: 1) explore genome-wide levels of DNA methylation before and after a six months exercise intervention in adipose tissue from healthy, but previously sedentary men; 2) investigate the differences in adipose tissue DNA methylation between individuals with or without a family history of T2D; 3) relate changes in DNA methylation to adipose tissue mRNA expression and metabolic phenotypes in vitro.
A total of 31 men, 15 FH+ and 16 FH−, had subcutaneous adipose tissue biopsies taken at baseline. The FH+ and FH− individuals were group-wise matched for age, gender, BMI and VO2max at inclusion, and there were no significant differences between FH+ and FH− individuals, respectively (Table S1). DNA methylation in the adipose tissue was analyzed using the Infinium HumanMethylation450 BeadChip array. After quality control (QC), DNA methylation data was obtained for a total number of 476,753 sites. No individual CpG site showed a significant difference in DNA methylation between FH+ and FH− men after false discovery rate (FDR) correction (q>0.05) [24]. Additionally, there were no global differences between the FH+ and FH− individuals when calculating the average DNA methylation based on genomic regions (Figure 1a) or CpG content (Figure 1b; q>0.05).
Subcutaneous adipose tissue biopsies were taken from 23 men both before and after exercise, followed by successful DNA extraction and analysis of DNA methylation using the Infinium HumanMethylation450 BeadChip array. Since we found no significant differences in DNA methylation between FH+ and FH− men at baseline, the two groups were combined when examining the impact of exercise on DNA methylation in adipose tissue. In Table 1 the clinical and metabolic outcomes of the exercise intervention are presented for these 23 men, showing a significant decrease in waist circumference, waist/hip ratio, diastolic blood pressure, and resting heart rate, whereas a significant increase was seen for VO2max and HDL.
To evaluate the global human methylome in adipose tissue, we first calculated the average level of DNA methylation in groups based on either the functional genome distribution (Figure 1a), or the CpG content and neighbourhood context (Figure 1b). We also present the average level of DNA methylation separately for the Infinium I (n = 126,804) and Infinium II (n = 326,640) assays due to different β-value distributions for these assays [25]. When evaluating Infinium I assays in relation to nearest gene, the global level of DNA methylation after exercise increased in the 3′ untranslated region (UTR; q<0.05), whereas a decrease was seen in the region 1500–200 bp upstream of transcription start (TSS1500), TSS200, 5′UTR and within the first exon (1st Exon; q<0.05). The global DNA methylation level of Infinium II assays increased significantly (q<0.05) after exercise within all regions except TSS200 (Figure 1c and Table S2). In general, the average level of DNA methylation was low in the region from TSS1500 to the 1st Exon (5–36%), whereas the gene body, the 3′UTR and intergenic region displayed average DNA methylation levels ranging from 43–72% (Figure 1c and Table S2). When evaluating global DNA methylation based on CpG content and distance to CpG islands, average DNA methylation for Infinium I assays decreased significantly after exercise in CpG islands, whereas an increase was seen in northern and southern shelves (regions 2000–4000 bp distant from CpG islands) as well as in the open sea (regions further away from a CpG island) (q<0.05; Figure 1d and Table S2). For Infinium II assays, average DNA methylation was significantly increased in all regions after the exercise intervention (q<0.05; Figure 1d and Table S2). The global level of DNA methylation was low within CpG islands (9–21%), intermediate within the shores (2000 bp regions flanking the CpG islands; 31–44%), whereas the shelves and the open sea showed the highest level of DNA methylation (67–76%; Figure 1d and Table S2). Although technical variation between probe types has been reported for the Infinium HumanMethylation450 BeadChip array, seen as a divergence between the β-values distribution retrieved from the Infinium I and II assays [25], the global differences in DNA methylation we observe between probe types are more likely a result of skewed GC content due to the design criteria of the two different assays. Infinium I assays have significantly more CpGs within the probe body than the Infinium II assays, and 57% are annotated to CpG islands, whereas most Infinium II assays have less than three underlying CpGs in the probe and only 21% are designated as CpG islands [26].
We next investigated if there was a difference in DNA methylation in any of the 476,753 analyzed individual CpG sites in adipose tissue in response to exercise. A flowchart of the analysis process is found in Figure 2. SNPs within the probe were not a criterion for exclusion in this analysis, as the participants are their own controls, thereby excluding genetic variation within the tested pairs. Applying FDR correction (q<0.05) resulted in 17,975 CpG sites, corresponding to 7,663 unique genes, that exhibit differential DNA methylation in adipose tissue after exercise. Among these 17,975 individual sites, 16,470 increased and 1,505 decreased the level of DNA methylation in response to exercise, with absolute changes in DNA methylation ranging from 0.2–10.9% (Figure 3a–b). Aiming for biological relevance, we further filtered our results requiring the average change in DNA methylation (β-value) for each CpG site to be ≥5% before vs. after exercise. Adding the criteria with a ≥5% change in DNA methylation resulted in 1,009 significant individual CpG sites: 911 with increased and 98 with decreased levels of DNA methylation in response to the six months exercise intervention. Of those, 723 sites are annotated to one or more genes, and correspond to 641 unique gene IDs. A comparison of our 1,009 significant CpG sites with Infinium probes reported to cross-react to alternative genomic locations [27] showed only one probe with 50 bases and 14 probes with 49 bases matching to an alternative genomic location. Data of the most significant CpG sites (q<0.005) and the sites that exhibit the greatest change in adipose tissue DNA methylation (difference in DNA methylation >8%) in response to exercise are presented in Table 2–3 and included ITPR2 and TSTD1 for increased, and LTBP4 for decreased DNA methylation. We found 7 CpG sites in this list to be targeted by Infinium probes reported to cross-react to alternative genomic locations (47 or 48 bases) [27]. Additionally, to investigate the possibility that the changes we see in response to exercise is rather an effect of epigenetic drift over time, we compared our 1,009 differentially methylated CpG sites (q<0.05, difference in β-value>5%) with three studies reporting aging-differentially methylated regions (a-DMRs) in a total of 597 unique positions [28]–[30]. Secondly we tested for association between age and the level of DNA methylation in the 31 individuals included at baseline in this study, representing a more valid age range (30–45 years) and tissue for the current hypothesis. We found no overlap between previously published a-DMRs or the age-associated CpG sites within our study (18 CpG sites; p<1×10−5), and the CpG sites differentially methylated after the exercise intervention.
The genomic distribution of individual CpG sites with a significant change in DNA methylation ≥5% with exercise is shown in Figure 3c–d, in comparison to all probes located on the Infinium HumanMethylation450 BeadChip and passing QC. The distribution is based on location in relation to the functional genome distribution (Figure 3c) or CpG content and distance to CpG islands (Figure 3d). We found that the CpG sites with altered level of DNA methylation in response to exercise were enriched within the gene body and in intergenic regions, while the proximal promoter, in particular TSS200 and the 1st exon, had a low proportion of differentially methylated CpG sites (p = 7×10−20; Figure 3c). In relation to CpG content and distance to CpG islands, the region with the highest proportion of significant CpG sites compared to the distribution on the array was in the open sea, i.e., regions more distant from a CpG island than 4000 bp. In contrast, the number of significant CpG sites found within the CpG islands was only half of what would be expected (p = 2×10−31; Figure 3d).
An increased level of DNA methylation has previously been associated with transcription repression [14]. We therefore related changes in adipose tissue DNA methylation of individual CpG-sites (q<0.05 and difference in mean β-values ≥5%) with changes in mRNA expression of the same gene (q<0.05) in response to exercise (Figure 2). We identified 236 CpG sites in 197 individual gene regions that exhibit differential DNA methylation together with a significant change in adipose tissue mRNA expression of the corresponding gene after exercise. Of these, 143 CpG sites (61%) connected to 115 genes showed an inverse relation to mRNA expression. After exercise, 139 CpG sites showed an increase in DNA methylation and a corresponding decrease in mRNA expression, including a gene for one of the GABA receptors (GABBR1), several genes encoding histone modifying enzymes (EHMT1, EHMT2 and HDAC4) and a transcriptional co-repressor (NCOR2). Only four CpG sites were found to decrease in the level of DNA methylation with a concomitant increase in mRNA expression. Table S3 shows all significant results of DNA methylation sites with an inverse relation to mRNA expression in human adipose tissue before vs. after exercise.
RALBP1 belongs to the genes that exhibit increased DNA methylation in the promoter region in parallel with decreased mRNA expression in adipose tissue in response to exercise (Figure 4a–b and Table S3). It has previously been shown to play a central role in the pathogenesis of metabolic syndrome [31] and to be involved in insulin-stimulated Glut4 trafficking [32]. We proceeded to functionally test if increased DNA methylation of the promoter of RALBP1 may cause decreased gene expression using a reporter gene construct in which 1500 bp of DNA of the human RALBP1 promoter was inserted into a luciferase expression plasmid that completely lacks CpG dinucleotides. The reporter construct could thereby be used to study the effect of promoter DNA methylation on the transcriptional activity. The construct was methylated using two different methyltransferases; SssI and HhaI, which methylate all CpG sites or only the internal cytosine residue in a GCGC sequence, respectively.
Increased DNA methylation of the RALBP1 promoter, as measured by luciferase activity, suppressed the transcriptional activity of the promoter (p = 0.028, Figure 4c). When the RALBP1 reporter construct was methylated in vitro using SssI (CG, 94 CpG sites), the transcriptional activity was almost completely disrupted (1.4±0.5), whereas the HhaI enzyme (GCGC, methylating 14 CpG sites) suppressed the transcriptional activity to a lesser extent (23.4±11.6), compared with the transcriptional activity of the mock-methylated control construct (448.2±201.7; Figure 4c).
We proceeded to investigate if candidate genes for obesity or T2D, identified using genome-wide association studies [3], are found among the genes exhibiting changed levels of DNA methylation in adipose tissue in response to six months exercise. Among all 476,753 CpG sites analyzed on the Infinium HumanMethylation450 BeadChip and passing QC, 1,351 sites mapped to 53 genes suggested to contribute to obesity in the review by McCarthy, and 1,315 sites mapped to 39 genes suggested to contribute to T2D [3]. We found 24 CpG sites located within 18 of the candidate genes for obesity with a difference in DNA methylation in adipose tissue in response to the exercise intervention (q<0.05, Table 4). Additionally, two of those genes (CPEB4 and SDCCAG8) showed concurrent inverse change in mRNA expression after exercise (q<0.05). Among the T2D candidate genes, 45 CpG sites in 21 different genes were differentially methylated (q<0.05) in adipose tissue before vs. after exercise (Table 5). Of note, 10 of these CpG sites mapped to KCNQ1 and 6 sites mapped to TCF7L2. A simultaneous change in mRNA expression was seen for four of the T2D candidate genes (HHEX, IGF2BP2, JAZF1 and TCF7L2) where mRNA expression decreased while DNA methylation increased in response to exercise (q<0.05, Table 5).
To further understand if the genes that exhibit differential DNA methylation and mRNA expression in adipose tissue in vivo affect adipocyte metabolism, we silenced the expression of selected genes in 3T3-L1 adipocytes using siRNA and studied its effect on lipogenesis. Two of the genes where we found increased DNA methylation in parallel with decreased mRNA expression in human adipose tissue in response to exercise (Figure 5a–d and Table S3) were selected for functional studies in a 3T3-L1 adipocyte cell line. HDAC4 was further a strong candidate due to multiple affected CpG sites within the gene, and both HDAC4 and NCOR2 are biologically interesting candidates in adipose tissue and the pathogenesis of obesity and type 2 diabetes [33]–[35]. Silencing of Hdac4 and Ncor2 in the 3T3-L1 adipocytes resulted in 74% reduction in the Hdac4 protein level (1.00±0.50 vs. 0.26±0.20, p = 0.043; Figure 5e) while the Ncor2 mRNA level was reduced by 56% (1.00±0.19 vs. 0.44±0.08, p = 0.043; Figure 5f) of control after transfection with siRNA for 72 hours and 24 h, respectively. Lipogenesis was nominally increased in the basal state (1.00±0.26 vs. 1.44±0.42, p = 0.079) and significantly increased in response to 0.1 nM insulin (1.16±0.30 vs. 1.52±0.34, p = 0.043) in 3T3-L1 adipocytes with decreased Hdac4 levels (Figure 5g). Decreased Ncor2 levels also resulted in increased lipogenesis in the basal (1.00±0.19 vs. 1.19±0.19, p = 0.043) and insulin stimulated (1 nM; 1.38±0.17 vs. 1.73±0.32, p = 0.043) state (Figure 5h).
To technically validate the DNA methylation data from the Infinium HumanMethylation450 BeadChips, we compared the genome-wide DNA methylation data from one adipose tissue sample analyzed at four different occasions. Technical reproducibility was observed between all samples, with Pearson's correlation coefficients >0.99 (p<2.2×10−16, Figure S1a). Secondly, we re-analyzed DNA methylation of four CpG sites using Pyrosequencing (PyroMark Q96ID, Qiagen) in adipose tissue of all 23 men both before and after exercise (Table S4). We observed a significant correlation between the two methods for each CpG site (p<0.05; Figure S1b), and combining all data points gives a correlation factor of 0.77 between the two methods (p<0.0001; Figure S1c).
This study highlights the dynamic feature of DNA methylation, described using a genome-wide analysis in human adipose tissue before and after exercise. We show a general global increase in adipose tissue DNA methylation in response to 6 months exercise, but also changes on the level of individual CpG sites, with significant absolute differences ranging from 0.2–10.9%. This data, generated using human adipose tissue biopsies, demonstrate an important role for epigenetic changes in human metabolic processes. Additionally, this study provides a first reference for the DNA methylome in adipose tissue from healthy, middle aged men.
Changes in DNA methylation have been suggested to be a biological mechanism behind the beneficial effects of physical activity [18], [36]. In line with this theory, a nominal association between physical activity level and global LINE-1 methylation in leukocytes was recently reported [37]. More important from a metabolic point-of-view, a study investigating the impact of long term exercise intervention on genome-wide DNA methylation in human skeletal muscle was recently published, and showed epigenetic alterations of genes important for T2D pathogenesis and muscle physiology [23]. This relationship between exercise and altered DNA methylation is here expanded to include human adipose tissue, as our data show 17,975 individual CpG sites that exhibit differential DNA methylation in adipose tissue after an exercise intervention, corresponding to 7,663 unique genes throughout the genome. Genome-wide association studies have identified multiple SNPs strongly associated with disease, but still the effect sizes of the common variants influencing for example risk of T2D are modest and in total only explain a small proportion of the predisposition. Importantly, although each variant only contributes with a small risk, these findings have led to improved understanding of the biological basis of disease [3]. Similarly, the absolute changes in DNA methylation observed in response to the exercise intervention are modest, but the large number of affected sites may in combination potentially contribute to a physiological response. Moreover, if the exercise induced differences in DNA methylation is expressed as fold-change instead of absolute differences, we observe changes ranging from 6 to 38%.
In regard to the distribution of analyzed CpG sites, most of the differentially methylated sites were found within the gene bodies and in intergenic regions, and fewer than expected was found in the promoter regions and CpG islands. This is in agreement with previous studies showing that differential DNA methylation is often found in regions other than CpG islands. For example, it was shown that tissue-specific differentially methylated regions in the 5′UTR are strongly underrepresented within CpG islands [38] and that most tissue-specific DNA methylation occurs at CpG island shores rather than the within CpG islands, and also in regions more distant than 2 kb from CpG islands [39]. It has further been proposed that non-CpG island DNA methylation is more dynamic than methylation within CpG islands [40]. The importance of differential DNA methylation within gene bodies is supported by multiple studies showing a positive correlation between gene body methylation and active transcription [40], and that DNA methylation may regulate exon splicing [41], [42]. In this study, the exercise intervention associated with a decrease in waist circumference and waist-hip ratio, which suggests reduced abdominal obesity, a phenotype known to be associated with reduced risk of metabolic diseases [43]. Indeed, increased levels of DNA methylation were observed after exercise both in the promoter region and in the gene body of ITPR2, a locus previously associated with waist-hip ratio [44]. Furthermore, in addition to increased VO2max, the study participants responded to exercise with a decrease in diastolic blood pressure and heart rate, and an improvement in HDL levels, which are some of the different mechanisms through which exercise is known to reduce the risk for T2D and cardiovascular disease [43]. Adipose tissue comprises not only of adipocytes but a mixture of different cell types. To evaluate if the cellular composition of adipose tissue may change during exercise, we looked at the mRNA expression for a number of cell type specific markers before and after the exercise intervention. None of these showed any difference in adipose tissue mRNA expression before vs. after exercise (q>0.05; LEP, PNPLA2, FAS, LIPE and PPARG as markers of adipocytes; SEBPA/B/D and DLK1 as markers of preadipocytes, PRDM16 and UCP1 as markers of brown adipocytes; ITGAX, EMR1, ITGAM as markers of macrophages; TNF and IL6 representing cytokines and finally CCL2 and CASP7 as markers for inflammation). Although this result suggests that there is no a major change in the cellular composition of the adipose tissue studied before compared with after the exercise intervention, future studies should investigate the methylome in isolated adipocytes. Additionally, in previous studies of DNA methylation in human pancreatic islets, the differences observed in the mixed-cell tissue were also detected in clonal beta cells exposed to hyperglycemia [20], [21], suggesting that in at least some tissues, the effects are transferable from the relevant cell type to the tissue of interest for human biology.
The impact of this study is further strengthened by our results showing altered DNA methylation of genes or loci previously associated with obesity and T2D. Although there was no enrichment of differential DNA methylation in those genes compared to the whole dataset, this result may provide a link to the mechanisms for how the loci associated with common diseases exert their functions [18]. 18 obesity and 21 T2D candidate genes had one or more CpG sites which significantly changed in adipose tissue DNA methylation after exercise. 10 CpG sites were found to have altered DNA methylation in response to exercise within the gene body of KCNQ1, a gene encoding a potassium channel and known to be involved in the pathogenesis of T2D, and also subject to parental imprinting [45]. Moreover, exercise associated with changes in DNA methylation of six intragenic CpG sites in TCF7L2, the T2D candidate gene harbouring a common variant with the greatest described effect on the risk of T2D [3]. This is of particular interest considering that TCF7L2 is subject to alternative splicing [46], [47] and the fact that gene exons are more highly methylated than introns, with DNA methylation spikes at splice junctions, suggesting a possible role for differential DNA methylation in transcript splicing [42]. In addition to differential DNA methylation, we also observed an inverse change in adipose tissue mRNA expression for some of these candidate genes, including TCF7L2, HHEX, IGF2BP2, JAZF1, CPEB4 and SDCCAG8 in response to exercise.
The understanding of the human methylome is incomplete although recently developed methods for genome-wide analysis of DNA methylation already have made, and are likely to continue to make, tremendous advances [48]. High coverage data describing differences in the levels of DNA methylation between certain human tissues or cell types [38], as well as differences observed during development [42], have started to emerge. Regardless, deeper knowledge about the epigenetic architecture and regulation in human adipose tissue has been missing until now. We found that the genetic region with the highest average level of DNA methylation in adipose tissue was the 3′UTR, followed by the gene body and intergenic regions, and those regions also increased the level of DNA methylation in response to exercise. This supports the view that the human methylome can dynamically respond to changes in the environment [14], [15]. One explanation for the low average levels of DNA methylation observed in the promoter region (TSS1500/200), 5′UTR and the first exon, may be that these regions often overlap with CpG islands, which are generally known to be unmethylated. Indeed, our results show a very low level of DNA methylation within the CpG islands, and how the level then increases with increasing distances to a CpG island.
It has long been debated if increased DNA methylation precedes gene silencing, or if it is rather a consequence of altered gene activity [40]. The luciferase assay experiments from this study and others [21], [23] suggest that DNA methylation may have a causal role, as increased promoter DNA methylation leads to reduced transcriptional activity. Here we further related our findings of altered DNA methylation to mRNA expression, and we identified 197 genes where both DNA methylation and mRNA expression significantly changed in adipose tissue after exercise. Of these, 115 genes (58%) showed an inverse relation, 97% showing an increase in the level of DNA methylation and a decrease in mRNA expression. It should be noted that epigenetic processes are likely to influence more aspects of gene expression, including accessibility of the gene, posttranscriptional RNA processing and stability, splicing and also translation [49]. For example, DNA methylation within the gene body has previously been linked to active gene transcription, suggestively by improving transcription efficiency [42].
Two genes, HDAC4 and NCOR2, with biological relevance in adipose tissue metabolism were selected for functional validation. HDAC4 is a histone deacetylase regulated by phosphorylation, and known to repress GLUT4 transcription in adipocytes [35]. In skeletal muscle, HDAC4 has been found to be exported from the nucleus during exercise, suggesting that removal of the transcriptional repressive function could be a mechanism for exercise adaptation [50]. For HDAC4, we observed increased levels of DNA methylation and a simultaneous decrease in mRNA expression in adipose tissue in response to the exercise intervention. Additionally, the functional experiments in cultured adipocytes suggested increased lipogenesis when Hdac4 expression was reduced. This could be an indicator of reduced repressive activity on GLUT4, leading to an increase in adipocyte glucose uptake and subsequent incorporation of glucose into triglycerides in the process of lipogenesis. NCOR2 also exhibited increased levels of DNA methylation and a simultaneous decrease in mRNA expression in adipose tissue in response to the exercise intervention, and furthermore we observed increased lipogenesis when Ncor2 expression was down regulated in the 3T3-L1 cell line. NCOR2 is a nuclear co-repressor, involved in the regulation of genes important for adipogenesis and lipid metabolism, and with the ability to recruit different histone deacetylase enzymes, including HDAC4 [51]. These results may be of clinical importance, since HDAC inhibitors have been suggested in the treatment of obesity and T2D [18], [52].
In summary, this study provides a detailed map of the human methylome in adipose tissue, which can be used as a reference for further studies. We have also found evidence for an association between differential DNA methylation and mRNA expression in response to exercise, as well as a connection to genes known to be involved in the pathogenesis of obesity and T2D. Finally, functional validation in adipocytes links DNA methylation via gene expression to altered metabolism, supporting the role of histone deacetylase enzymes as a potential candidate in clinical interventions.
Written informed consent was obtained from all participants and the research protocol was approved by the local human research ethics committee.
This study included a total of 31 men from Malmö, Sweden, recruited for a six months exercise intervention study, as previously described [23], [53]. Fifteen of the individuals had a first-degree family history of T2D (FH+), whereas sixteen individuals had no family history of diabetes (FH−). They were all sedentary, but healthy, with a mean age of 37.4 years and a mean BMI of 27.8 kg/m2 at inclusion. All subjects underwent a physical examination, an oral glucose tolerance test and a submaximal exercise stress test. Bioimpedance was determined to estimate fat mass with a BIA 101 Body Impedance Analyzer (Akern Srl, Pontassieve, Italy). To directly assess the maximal oxygen uptake (VO2max), an ergometer bicycle (Ergomedic 828E, Monark, Sweden) was used together with heart rate monitoration (Polar T61, POLAR, Finland) [53]. FH+ and FH− men were group-wise matched for age, BMI and physical fitness (VO2max) at baseline. Subcutaneous biopsies of adipose tissue from the right thigh were obtained during the fasting state under local anaesthesia (1% Lidocaine) using a 6 mm Bergström needle (Stille AB, Sweden) from all participants before and from 23 participants after the six months exercise intervention (>48 hours after the last exercise session). The weekly group training program included one session of 1 hour spinning and two sessions of 1 hour aerobics and was led by a certified instructor. The participation level was on average 42.8±4.5 sessions, which equals to 1.8 sessions/week of this endurance exercise intervention. The study participants were requested to not change their diet and daily activity level during the intervention.
DNA methylation was analyzed in DNA extracted from adipose tissue, using the Infinium HumanMethylation450 BeadChip assay (Illumina, San Diego, CA, USA). This array contains 485,577 probes, which cover 21,231 (99%) RefSeq genes [25], [54]. Genomic DNA (500 ng) from adipose tissue was bisulfite treated using the EZ DNA methylation kit (Zymo Research, Orange, CA, USA). Analysis of DNA methylation with the Infinium assay was carried out on the total amount of bisulfite-converted DNA, with all other procedures following the standard Infinium HD Assay Methylation Protocol Guide (Part #15019519, Illumina). The BeadChips' images were captured using the Illumina iScan. The raw methylation score for each probe represented as methylation β-values was calculated using GenomeStudio Methylation module software (β = intensity of the Methylated allele (M)/intensity of the Unmethylated allele (U)+intensity of the Methylated allele (M)+100). All included samples showed a high quality bisulfite conversion efficiency (intensity signal >4000) [55], and also passed all GenomeStudio quality control steps based on built in control probes for staining, hybridization, extension and specificity. Individual probes were then filtered based on Illumina detection p-value and all CpG sites with a mean p<0.01 were considered detected and used for subsequent analysis. In total we obtained DNA methylation data for 476,753 CpG sites from adipose tissue of 31 men before and 23 men after the exercise intervention. Before further analysis, the DNA methylation data was exported from GenomeStudio and subsequently analyzed using Bioconductor [56] and the lumi package [57]. β-values were converted to M-values (M = log2(β/(1-β))), a more statistically valid method for conducting differential methylation analysis [58]. Next, data was background corrected by subtracting the median M-value of the 600 built in negative controls and was further normalized using quantile normalization. Correction for batch effects within the methylation array data was performed using COMBAT [59]. For the calculations of global DNA methylation, quantile normalization was omitted and probes reported to be cross-reactive (≥49 bases) or directly affected by a SNP (MAF>5%) were removed [27]. Due to different performance of Infinium I and Infinium II assays [25], the results based on average DNA methylation are calculated and presented separately for each probe type. To control for technical variability within the experiment, one adipose tissue sample was included and run on four different occasions (Figure S1a). As the β-value is easier to interpret biologically, M-values were reconverted to β-values when describing the results and creating the figures.
RNA extracted from the subcutaneous adipose tissue biopsies was used for a microarray analysis, performed using the GeneChip Human Gene 1.0 ST whole transcript based array (Affymetrix, Santa Clara, CA, USA), following the Affymetrix standard protocol. Basic Affymetrix chip and experimental quality analyses were performed using the Expression Console Software, and the robust multi-array average method (RMA) was used for background correction, data normalization and probe summarization [60].
The human promoter fragment containing 1500 bp of DNA upstream of the transcription start site for RALBP1 (Chr18:9474030–9475529, GRCh37/hg19) was inserted into a CpG-free luciferase reporter vector (pCpGL-basic) as previously described [21]. The construct was methylated using two different DNA methyltransferases; SssI which methylates all cytosine residues within the double-stranded dinucleotide recognition sequence CG, and HhaI which methylates only the internal cytosine residue in the GCGC sequence (New England Biolabs, Frankfurt, Germany). INS-1 cells were co-transfected with 100 ng of the pCpGL-vector without (control) or with any of the three RALBP1 inserts (no DNA methyltransferase, SssI, HhaI) together with 2 ng of pRL renilla luciferase control reporter vector as a control for transfection efficiency (Promega, Madison, WI, USA). Firefly luciferase activity, as a value of expression, was measured for each construct and normalized against renilla luciferase activity using the TD-20/20 luminometer (Turner Designs, Sunnyvale, CA, USA). The results represent the mean of three independent experiments, and the values in each experiment are the mean of five replicates. Cells transfected with an empty pCpGL-vector were used as background control in each experiment.
For detailed description of siRNA and lipogenesis experiments see Methods S1. Briefly, 3T3-L1 fibroblasts were cultured at sub-confluence in DMEM containing 10% (v/v) FCS, 100 U/ml penicillin and 100 µg/ml streptomycin at 37°C and 95% air/5% CO2. Two-day post-confluent cells were incubated for 72 h in DMEM supplemented with 0.5 mM IBMX, 10 µg/ml insulin and 1 µM dexamethasone, after which the cells were cultured in normal growth medium. Seven days post-differentiation, cells were transfected by electroporation with 2 nmol of each siRNA sequence/gene (Table S5). 0.2 nmol scrambled siRNA of each low GC-, medium GC- and high GC-complex were mixed as control. The cells were replated after transfection and incubated for 72 hours (siRNA against Hdac4) or 24 hours (siRNA against Ncor2).
Cells harvested for western blot analysis were solubilized and homogenized, and 20 µg protein was subjected to SDS-PAGE (4–12% gradient) and subsequent transferred to nitrocellulose membranes. The primary antibody (rabbit polyclonal anti-hdac4; ab12172, Abcam, Cambridge, UK) was diluted in 5 ml 5% BSA/TBST and incubated overnight in 4°C. The secondary antibody (goat anti-rabbit IgG conjugated to horseradish peroxidase; ALI4404, BioSource, Life Technologies Ltd, Paisley, UK) was diluted 1∶20,000 in 5% milk/TBST. Protein was detected using Super Signal and ChemiDoc (BioRad, Hercules, CA, USA).
Quantitative PCR (Q-PCR) analyses were performed in triplicate on an ABI7900 using Assays on demand with TaqMan technology (Mm00448796_m1, Applied Biosystems, Carlsbad, CA, USA). The mRNA expression was normalized to the expression of the endogenous control gene Hprt (Mm01545399_m1, Applied Biosystems).
To measure lipogenesis, 10 µl tritium labelled ([3H]) glucose (Perkin Elmer, Waltham, MA, USA) was added followed by insulin of different concentrations; 0, 0.1, and 1 nM for Hdac4 siRNA and 0 and 1 nM for Ncor2 siRNA experiments, respectively. All concentrations were tested in duplicates. After 1 hour, incorporation of [3H] glucose into cellular lipids was measured by scintillation counting. Lipogenesis is expressed as fold of basal lipogenesis.
PyroSequencing (PyroMark Q96ID, Qiagen, Hilden, Germany) was used to technically validate data from the genome-wide DNA methylation analysis. PCR and sequencing primers were either designed using PyroMark Assay Design 2.0 or ordered as pre-designed methylation assays (Qiagen, Table S4), and all procedures were performed according to recommended protocols. Briefly, 100 ng genomic DNA from adipose tissue of 23 individuals both before and after the exercise intervention was bisulfite converted using Qiagen's EpiTect kit. With one primer biotinylated at its 5′ end, bisulfite-converted DNA was amplified by PCR using the PyroMark PCR Master Mix kit (Qiagen). Biotinylated PCR products were immobilized onto streptavidin-coated beads (GE Healthcare, Uppsala, Sweden) and DNA strands were separated using denaturation buffer. After washing and neutralizing using PyroMark Q96 Vacuum Workstation, the sequencing primer was annealed to the immobilized strand. PyroSequencing was performed with the PyroMark Gold Q96 reagents and data were analyzed using the PyroMark Q96 (version 2.5.8) software (Qiagen).
Clinical data is presented as mean ± SD, and comparisons based on a t-test and two-tailed p-values. Genome-wide DNA methylation data from the Infinium HumanMethylation450 BeadChip before vs. after the six month exercise intervention was analyzed using a paired non-parametric test, whereas a paired t-test was used to compare the mRNA expression. DNA methylation and mRNA expression data are expressed as mean ± SD. To account for multiple testing and reduce the number of false positives, we applied q-values to measure the false discovery rate (FDR) on our genome-wide analyses of DNA methylation and mRNA expression [24]. Luciferase activity was analyzed using the Friedman test (paired, non-parametric test on dependent samples) and presented as mean ± SEM. Data from 3T3-L1 adipocyte experiments showing protein, mRNA and lipogenesis levels are presented as mean ± SEM, and the results are based on Wilcoxon signed-rank test.
|
10.1371/journal.ppat.1003174 | Comparative Analysis of Measures of Viral Reservoirs in HIV-1 Eradication Studies | HIV-1 reservoirs preclude virus eradication in patients receiving highly active antiretroviral therapy (HAART). The best characterized reservoir is a small, difficult-to-quantify pool of resting memory CD4+ T cells carrying latent but replication-competent viral genomes. Because strategies targeting this latent reservoir are now being tested in clinical trials, well-validated high-throughput assays that quantify this reservoir are urgently needed. Here we compare eleven different approaches for quantitating persistent HIV-1 in 30 patients on HAART, using the original viral outgrowth assay for resting CD4+ T cells carrying inducible, replication-competent viral genomes as a standard for comparison. PCR-based assays for cells containing HIV-1 DNA gave infected cell frequencies at least 2 logs higher than the viral outgrowth assay, even in subjects who started HAART during acute/early infection. This difference may reflect defective viral genomes. The ratio of infected cell frequencies determined by viral outgrowth and PCR-based assays varied dramatically between patients. Although strong correlations with the viral outgrowth assay could not be formally excluded for most assays, correlations achieved statistical significance only for integrated HIV-1 DNA in peripheral blood mononuclear cells and HIV-1 RNA/DNA ratio in rectal CD4+ T cells. Residual viremia was below the limit of detection in many subjects and did not correlate with the viral outgrowth assays. The dramatic differences in infected cell frequencies and the lack of a precise correlation between culture and PCR-based assays raise the possibility that the successful clearance of latently infected cells may be masked by a larger and variable pool of cells with defective proviruses. These defective proviruses are detected by PCR but may not be affected by reactivation strategies and may not require eradication to accomplish an effective cure. A molecular understanding of the discrepancy between infected cell frequencies measured by viral outgrowth versus PCR assays is an urgent priority in HIV-1 cure research.
| Efforts to cure HIV-1 infection have focused on a small pool of CD4+ T cells that carry viral genetic information in a latent form. These cells persist even in patients on optimal antiretroviral therapy. Novel therapeutic strategies targeting latently infected cells are being developed, and therefore practical assays for measuring latently infected cells are urgently needed. These cells were discovered using a virus culture assay in which the cells are induced to release virus particles that are then expanded in culture. This assay is difficult, time-consuming, and expensive. Here we evaluate alternative approaches for measuring persistent HIV-1, all of which rely on the detection of viral genetic information using the polymerase chain reaction (PCR). None of the PCR-based assays correlated precisely with the virus culture assay. The fundamental problem is that infected cell frequencies determined by PCR are at least 2 logs higher than frequencies determined by the culture assay. Much of this difference may be due to cells carrying defective forms of the virus. These cells may not be eliminated by strategies designed to target latently infected cells. In this situation, successful clearance of latently infected cells might be masked by a large unchanging pool of cells carrying defective HIV-1.
| Treatment of HIV-1 infection with highly active antiretroviral therapy (HAART) can reduce plasma HIV-1 RNA levels in treated patients to below the detection limit of clinical assays (50 copies of HIV-1 RNA/ml) [1]–[3]. The effective suppression of viremia initially encouraged hopes that the virus could be eradicated with two to three years of HAART [3]. However, a latent form of HIV-1 infection persists in vivo [4], [5]. A small fraction of resting memory CD4+ T cells carry integrated viral genomes. These cells do not produce virus particles while in the resting state, but can give rise to replication-competent virus following cellular activation [4], [5]. These latently infected cells are rare but stable, even in patients on prolonged HAART [6]–[11]. Interruption of HAART leads to a rebound in viremia [12], [13], typically from an archival variant [14]. The latent reservoir is widely recognized as the major barrier to HIV-1 eradication [15].
Strategies aimed at reactivating latent virus and thereby accelerating the clearance of the latent reservoir are now in advanced pre-clinical testing or early clinical trials [15]. Approaches for the reactivation of latent HIV-1 include T cell activating cytokines [16]–[19], T cell receptor and T cell receptor signaling pathway agonists [20]–[24], histone deacetylase inhibitors [25]–[27], DNA methylase inhibitors [28], [29], and compounds like 5-hydroxynaphthalene-1,4-dione [30] and disulfiram [31]. A single dose of the histone deacetylase inhibitor suberoylanilide hydroxamic acid (SAHA) has recently been shown to increase cell-associated HIV-1 RNA in CD4+ T cells from patients on HAART [32].
A major question for current and future trials of eradication strategies is how to evaluate the effectiveness of the interventions. The principal approach for quantifying HIV-1 persistence during HAART is a viral outgrowth assay performed on highly purified resting CD4+ T cells. These cells do not produce virus without stimulation [4]. In the assay, limiting dilutions of resting CD4+ T cells are stimulated with the mitogen phytohemagglutinin (PHA) or with anti-CD3 plus anti-CD28 antibodies in the presence of irradiated allogeneic peripheral blood mononuclear cells (PBMC) [5], [6], [9], [10], [33]. These stimuli induce global T cell activation, which reverses latency at least in a fraction of cells carrying integrated HIV-1 genomes. The viruses released from these cells are expanded in CD4+ lymphoblasts from HIV-1-negative donors and detected after 2–3 weeks by an ELISA assay for HIV-1 p24 antigen in the supernatant. This assay detects individual latently infected cells that release replication-competent virus following cellular activation. The frequency of latently infected cells, expressed in terms of infectious units per million (IUPM) resting CD4+ T cells, is determined using Poisson statistics and is on the order of 0.1–10 IUPM in most patients on long term HAART. The value of this assay is that it detects cells that can, when activated, release viruses capable of robust replication. It therefore provides a minimum estimate of the frequency of latently infected cells that that must be eliminated to ensure eradication and is used here as a standard for comparison. In principle, this assay can also detect resting CD4+ T cells harboring labile unintegrated forms of HIV-1 DNA [34], [35], although the frequency of cells containing unintegrated DNA during HAART is low [36]–[38]. Although the viral outgrowth assay has important advantages, it is expensive and labor intensive, and it requires large amounts of blood (120–180 ml).
Alternative approaches generally involve PCR assays for HIV-1 DNA. Some of these assays distinguish between integrated proviruses and unintegrated HIV-1 DNA [39], [40]. A problem with all PCR-based assays is that they fail to distinguish between replication-competent and defective forms of the viral genome. A significant but poorly characterized proportion of infected resting CD4+ T cells contain proviruses that are defective, hypermutated, or silenced [41], [42]. PCR assays are now also being used to quantify HIV-1 DNA in CD4+ T cells from the gut associated lymphoid tissue (GALT), where the frequency of HIV-1 infection is generally higher than in the blood [43], [44]. Highly sensitive PCR methods are also now used to quantify HIV-1 RNA in cells [32]. Free virus particles are also found in the plasma of patients on HAART [45]–[47]. This residual viremia is an important indication of ongoing virus production. Several studies have shown that residual viremia is not reduced by treatment intensification [48]–[50], and thus it is likely to reflect virus production from stable reservoirs. For example, residual viremia could in part reflect virus production by latently infected cells that have become activated.
It is currently unclear which assay(s) should be used to monitor HIV-1 reservoirs in clinical trials of eradication strategies. The development of a high-throughput scalable assay to measure the latent reservoir in patients has been identified as a key priority in HIV-1 eradication research [51]. Here we present a comparative analysis of eleven different approaches for measuring for HIV-1 reservoirs in two well characterized cohorts of patients on long term HAART. The goal of the study was to determine how these assays correlate with the viral outgrowth assay. The results provide insights into how reservoirs should be evaluated in future clinical trials aimed at curing HIV-1 infection.
The baseline characteristics of the cohort are shown in Table 1. Of 30 study participants, 10 started HAART during acute or early HIV-1 infection while the remaining 20 started HAART during chronic infection. The mean (±SD) age was 53.2±9.6 years. Patients starting therapy during acute/early infection were slightly younger (47.8±9.3 vs. 55.9±8.7 years). The majority (76.7%) of study subjects were white/non-Hispanic. The current CD4+ T cell counts were not significantly different between patients starting HAART during acute/early vs. chronic infection (727±287 vs. 672±144, P = 0.58). For patients starting HAART during chronic infection, the CD4 nadir was 202±138 cells/µl. The average duration of viral load suppression on HAART was 5.8±2.5 years for patients starting HAART during acute/early infection and 8.0±4.2 years for patients starting HAART during chronic infection. No patient in either cohort had a documented “blip” above 40 copies RNA/mL in the year preceding the blood draw.
Samples were processed, split, and sent to laboratories with expertise in the assays described above. Each assay was developed independently by the relevant laboratory, with different input material, assay methodology, normalization method, statistical characteristics, and caveats. These are indicated in Table 2. The cell types analyzed and viral species detected in each assay are indicated in Table 3. Except for the single copy assay for plasma HIV-1 RNA, assay results are presented in the form of infected cell frequencies to facilitate cross-assay comparisons. However, the cell populations included and viral species detected in each assay (Table 3) must be kept in mind in interpreting these frequencies. The statistical characteristics of each assay, such as the coefficient of variation, the limit of detection, and the dynamic range (summarized in Table 2) are important considerations in choosing assays to monitor viral persistence. For example, the coefficient of variation for the viral outgrowth assay is considerably higher than that for most PCR-based assays. Statistical characteristics must also be considered in evaluating the correlations between the results of different assays because problems related assay precision, accuracy, and sensitivity can obscure correlations. Table 2 also describes the primers used for each PCR assay. Negative results for any single PCR assay on a given patient sample can reflect sequence variation in the primer binding site [47].
Replication-competent HIV-1 was isolated from purified resting CD4+ T cells from peripheral blood in 29/30 study participants (Figure 1). Infected cell frequencies showed a log normal distribution with a geometric mean frequency of 0.64 IUPM, consistent with previous reports [6], [9], [10], [52]. Latently infected cells were readily detected by this method in patients starting HAART during acute/early HIV-1 infection (mean = 0.28 IUPM). The mean frequency of latently infected cells was significantly lower in patients starting HAART in acute/early HIV-1 infection compared to those starting during chronic infection (0.28 vs. 0.97 IUPM, P = 0.048), although there was substantial overlap between the two populations (Figure 1). The frequency of latently infected cells was not correlated with the time between infection and the initiation of a suppressive HAART regimen (r = 0.18, P = 0.34), suggesting that the size of the reservoir does not increase continuously during untreated HIV-1 infection. For a single study participant, replication-competent virus was not detected even after repeat culture. Based on input cell number, the frequency of latently infected cells in this patient was <0.06 IUPM.
The limit of detection of this assay is determined by the input number of resting CD4+ T cells. With a 180 ml blood sample, the average yield of resting CD4+ T cells in millions was 28.3±14.7. With therapeutic strategies that reduce the reservoir by more than 1.5 logs, latently infected cells would no longer be detectable in the majority of patients unless larger blood volumes or leukopheresis samples were used (Figure 1).
A simple approach for quantifying persistent HIV-1 is to measure HIV-1 DNA using PCR in unfractionated PBMC. This was done using a droplet digital PCR approach that has greater accuracy than standard real time PCR methods, particularly with low template numbers [53]. HIV-1 DNA was detected in 28/30 PBMC samples (Figure 1). Values varied over a ∼2 log range with a geometric mean value of 74 copies/106 PBMC for the entire cohort. Two subjects (2113 and 2453) had values that were below the limit of detection (2 copies/106 PBMC). Interestingly, for these two subjects, cells with replication-competent virus were readily measured in the virus outgrowth assay (5.4 and 0.1 IUPM, respectively). HIV-1 DNA values were generally lower in subjects starting HAART in acute/early HIV-1 infection compared to those starting during chronic infection (geometric mean values 47 vs. 93 copies/106; P = 0.30). The frequency of PBMC with HIV-1 DNA showed a modest but significant correlation with the time between infection and the initiation of a suppressive HAART regimen (r = 0.38, P = 0.037).
To determine whether HIV-1 DNA levels in PBMC could be used as a surrogate measure of the size of the latent reservoir, we examined the correlation between results of the viral outgrowth assay on purified resting CD4+ T cells and the droplet digital PCR assay for HIV-1 DNA in PBMC from the same blood sample. As shown in Figure 2A and Table 4, there was essentially no correlation between the two assays (r = 0.20, P = 0.29) for the combined study population. Based on 95% confidence intervals for r, a strong correlation (r>0.6) could be excluded (Table 4). For patients treated during acute/early infection, a modest correlation that did not reach statistical significance was observed (r = 0.46, P = 0.18), but there was no correlation between the two assays for patients initiating HAART during chronic infection (r = −0.038, P = 0.87).
We next examined whether the correlation between the results of the viral outgrowth and PCR assays might be improved if HIV-1 DNA levels were measured in purified resting CD4+ T cells instead of unfractionated PBMC. HIV-1 DNA was detected in 14/16 samples (Figure 1). The geometric mean level of HIV-1 DNA was 186 copies/106 resting CD4+ T cells. In two patients (#2013, 2418), levels were below the limit of detection (2 copies/106 resting CD4+ T cells). For these two subjects, cells with replication-competent virus were readily measured in the virus outgrowth assay (1.05 and 0.04 IUPM, respectively). As a measure of infected cell frequency, the PCR assay gave substantially higher values than did the viral outgrowth assay performed on the same blood samples (186 vs. 0.62 infected cells/106 resting CD4+ T cells, P<0.0001). This difference has been noted previously [5] and may reflect the high fraction of proviruses that are defective [41], [42].
Cells harboring defective viral genomes could accumulate over time during untreated disease. We did not, however, observed a significant correlation between the frequency of resting CD4+ T cells with HIV-1 DNA and the time between initial infection and suppression of viral replication on HAART (r = 0.33, P = 0.20), the time on a suppressive HAART regimen (r = 0.076, P = 0.78), or the total time since infection (r = 0.30, P = 0.26).
If the proportion of defective proviruses was the same in different patients, then the measurement of HIV-1 DNA in resting CD4+ T cells from patients on HAART might provide a simpler surrogate measure of the latent reservoir. However, as is shown in Figure 3, the ratio of infected cells detected in the PCR vs. viral outgrowth assays is highly variable from patient to patient (range <2 to 3540) even when both assays are performed on the same sample of purified resting CD4+ T cells. A subset of patients who initiated therapy during chronic infection showed very high ratios (>1000∶1). For this reason, there is only a modest correlation between the results of the two assays for the combined population (Figure 2B and Table 4, r = 0.45, P = 0.08). In patients treated during chronic infection, no correlation is observed (r = 0.10, P = 0.76) The levels of HIV-1 DNA in unfractionated PBMC and in purified resting CD4+ T cells showed a strong correlation (r = 0.78, P = 0.0004). This reflects the fact that in patients on HAART, the stable reservoir for HIV-1 is located primarily in resting CD4+ T cells [6]. For most patients, infected cell frequencies were higher in resting CD4+ T cells than in PBMC (Figure 1). This is the expected result if most of the infected cells in the blood are resting CD4+ T cells. However, if substantial numbers of activated T cells and monocytes are infected, then the resting CD4+ T cell value may not be higher. Chun and colleagues have shown that even in patients on HAART, infected cells frequencies as measured by DNA PCR can be higher among activated than resting CD4+ T cells [54].
Levels of integrated HIV-1 DNA were also measured in PBMC and purified resting CD4+ T cells from study participants using a previously described Alu PCR assay [39], [40], [55]. As shown in Figure 1, integrated HIV-1 DNA was detected in 19/19 PBMC samples, at a geometric mean frequency of 186 copies/106 PBMC. The frequencies were significantly lower in patients starting HAART in acute/early vs. chronic infection (84 vs 286 copies/106 PBMC, P = 0.04). As was observed with the droplet digital PCR assay for HIV-1 DNA, levels of integrated HIV-1 DNA were higher in purified resting CD4+ T cells than in unfractionated PBMC (geometric mean values 604 vs 186 copies/106 cells, Figure 1). Also consistent with the results of the droplet digital PCR assay was the finding that the frequency of resting CD4+ T cells with integrated HIV-1 DNA was much higher than the frequency of latently infected resting CD4+ T cells detected in the viral outgrowth assay performed on the same sample (by 1000 fold). In paired samples, the mean infected cell frequencies were 604 vs. 0.61/106 resting CD4+ T cells (P<0.0001).
Measurements of integrated HIV-1 DNA by Alu PCR and of total HIV-1 DNA by droplet digital PCR correlated well with each other both for samples of PBMC (Figure 2C, r = 0.63, P = 0.0042) and resting CD4+ T cells (r = 0.85, P = 0.0079). These results are consistent with the conclusion that in patients on long term HAART, most of the HIV-1 DNA is integrated, with unintegrated forms making only a minor contribution (see below) [38]. The fact that infected cell frequencies detected by Alu PCR were higher (186 vs 46 copies/106 PBMC, P = 0.0003) may reflect differences in assay standardization. This assay provides definitive detection of integrated HIV-1 DNA through the use of an initial PCR with a primer in an Alu element and a primer in HIV-1. For each infected cell, the distance to the nearest Alu element is different [56]. This influences amplification efficiency, and some proviruses are located too far from an Alu element to be detected. To circumvent this problem, a correction is applied based on an integration standard containing proviruses integrated at different distances from the nearest Alu element [39], [40], [55]. The accuracy of the method depends on how closely the distribution of provirus-Alu element distances in the standard matches the distribution of distances in a particular patient sample. It is possible that the higher levels of integrated HIV-1 DNA detected in some patients result from over-correction for proviruses missed by Alu-PCR. Alternative explanations include underestimation of infected cell frequency by droplet digital PCR or issues with the normalization methods used in one or both assays.
Interestingly, there was a highly significant positive correlation between the level of integrated HIV-1 DNA in PBMC and the frequency of latently infected resting CD4+ T cells determined in the viral outgrowth assay on the same blood sample (Figure 2D and Table 4, r = 0.70, P = 0.0008). When only patients starting HAART during chronic infection were considered, the correlation remained significant (r = 0.76, P = 0.0038). However, when the Alu PCR and viral outgrowth assays were performed on the same sample of purified resting CD4+ T cells, the correlation was weaker and not statistically significant (r = 0.41, P = 0.13), possibly related to the small number of samples and small number of genomes assayed.
Because the GALT is the site of very active viral replication in untreated patients with acute HIV-1 infection [57]–[61], we measured levels of HIV-1 DNA and RNA in the rectal biopsy samples from 19 study participants using qPCR. CD4+ T cells were enumerated in each sample by flow cytometry, and results were expressed as copies per CD4+ T cell. HIV-1 DNA was readily detectable in all samples (Figure 1). The geometric mean value for these 19 samples was 3015 copies/106 CD4+ T cells. In paired comparisons, these values were significantly higher than the levels of infection of resting CD4+ T cells from the peripheral blood detected by digital droplet PCR (4282 vs. 263 copies/106 cells, paired t-test: P<0.0001) or by Alu PCR (2977 vs. 600 copies/106 cells, paired t-test: P = 0.0008). The level of HIV-1 DNA in rectal CD4+ T cells was not significantly correlated with the frequency of latently infected resting CD4+ T cells in blood as measured by the viral outgrowth assay (Figure 2E, r = 0.26, P = 0.28). However, a strong correlation could not be formally excluded based on 95% confidence intervals (r = −0.22 to 0.64, Table 4). There was a significant correlation between the level of HIV-1 DNA in rectal biopsy samples and the frequency of infected cells in the peripheral blood as measured by digital droplet PCR (r = 0.58, P = 0.015) or by Alu PCR (r = 0.65. P = 0.016).
HIV-1 RNA was also detected by qRT-PCR in all samples. The geometric mean level was 1985 copies/106 rectal CD4+ T cells. These values cannot be used to establish infected cell frequencies because of the likely presence of multiple HIV-1 RNA molecules in some individual infected cells. Levels of HIV-1 RNA correlated well with measures of HIV-1 DNA in the same samples (r = 0.8811, P<0.0001). The geometric mean RNA/DNA ratio was 0.68. RNA/DNA ratios in rectal biopsies were significantly correlated with the frequency of resting CD4+ T cells in peripheral blood that scored in the viral outgrowth assay (r = 0.57. P = 0.013).
2-LTR circles represent abortive integration events. They have been used as a measure of recent infection in some studies, although controversy remains about their stability [62]–[66]. 2-LTR circles were measured in PBMC and purified resting CD4+ T cells from study subjects using droplet digital PCR (Figure 1). Circles were detected in only 9 of 30 PBMC samples. Among these the geometric mean level was 6.8 copies/106 PBMC. This was 27 fold lower than the total level of HIV-1 DNA in the same samples measured by droplet digital PCR (162 copies/106 PBMC, P<0.0001). As expected, levels of 2-LTR circles in purified resting CD4+ T cells from peripheral blood were higher than levels in unfractionated PBMC (geometric mean 13 copies/106 resting CD4+ T cells), but again this level was much lower (by 34 fold) than the total level of HIV-1 DNA in the same samples measured by droplet digital PCR (467 copies/106 resting CD4+ T cells, P<0.0001). These results demonstrate that 2-LTR circles make up only a small fraction of the total HIV-1 DNA measured by PCR in patients receiving on HAART. By Spearman's rank correlation analysis, levels of 2-LTR circles in PBMC or resting CD4+ T cells were not significantly correlated with infected cell frequencies measured in the viral outgrowth assay (rho = 0.19, P = 0.31 and rho = 0.38, P = 0.15 respectively).
Residual viremia was detectable in 20/30 study subjects, 13/20 in the chronic cohort and 7/10 in the acute cohort (Figure 1). Among the patients with detectable residual viremia, the geometric mean level was 0.78 copies/ml, consistent with previous studies [46], [47], [67]. There was no significant difference between the acute and chronic cohorts in the proportion of patients with detectable residual viremia (7/10 vs. 13/20) or in the level of detectable residual viremia (0.84 vs. 0.75 copies/ml, P = 0.75). Scatter plots show no obvious correlation between residual viremia and the viral outgrowth assay (Figure 2F). We considered the possibility that such a correlation could have been obscured by the failure to detect residual viremia in one third of the subjects. Negative results in the single copy assay could be due to primer mismatch or levels of residual viremia below the detection limit [47]. If patients for whom the single copy assay was negative are excluded, the correlation between residual viremia and the frequency of latently infected cells in the viral outgrowth assay is weak (r = 0.24, P = 0.33). If all patients are included in the analysis and a single copy assay value of 0.1 copies/ml is assumed for those patients with negative results in this assay, the rank correlation is also very weak (rho = 0.070, P = 0.71). Thus regardless of whether the negative values in the single copy assay represent primer mismatch or values below the limit of detection, it is difficult to construct a scenario where there is strong correlation between residual viremia and the viral outgrowth assay. A strong correlation could be formally excluded for the total population and for subpopulations starting HAART during acute/early or chronic infection. Residual viremia was not correlated with the level of infection in PBMC by droplet digital PCR (rho = −0.25, P = 0.18), the level of infection of CD4+ T cells in the rectal biopsies (rho = 0.12, P = 0.61), or the level of 2LTR circles (rho = −0.109, P = 0.57). Because the limit of detection of this assay with a sample volume of 8 ml is 0.2 copies/ml, a one log reduction in the size of viral reservoirs contributing to residual viremia would render values unmeasurable.
With the discovery of new agents that reactivate latent HIV-1, clinical trials of HIV-1 eradication strategies have begun [32]. No available clinical assay measures the size of the latent reservoir. Patients being enrolled in eradication studies have been on HAART for years and already have undetectable levels of viremia by standard clinical assays (detection limit 50 copies/ml). Therefore, there is an urgent need for laboratory assays to determine the efficacy of eradication strategies. Here we compare eleven different approaches for quantitating persistent HIV-1 in patients on HAART. The analysis involved seven analytical approaches and four different kinds of tissue samples. Assays were carried out using well characterized patients on long term stable HAART. Results were compared to the viral outgrowth assay that was originally used to define the latent reservoir [5], [6], [9], [10], [33].
We first evaluated PCR assays for HIV-1 DNA in unfractionated PBMC or resting CD4+ T cells. PCR quantitation of HIV-1 DNA in unfractionated PBMC perhaps offers the best chance for a scalable, high throughput assay for the latent reservoir. Using a novel droplet digital PCR assay, we detected HIV-1 DNA in PBMC from 28/30 subjects, with a mean infected cell frequency more than 100 fold higher than the mean frequency of resting CD4+ T cells that release replication-competent virus in the viral outgrowth assay. An even greater difference was observed when the viral outgrowth and droplet digital PCR assays were run on the same sample of purified resting CD4+ T cells. In part, this difference is likely to reflect the detection by PCR methods of cells harboring defective viral genomes such as cells with viral genomes that have been lethally hypermutated by APOBEC3G [42]. The difference between viral outgrowth and PCR-based assays was also observed in patients who start HAART in acute/early HIV-1 infection, suggesting that it is not simply the result of accumulation of cells with defective viruses over time. If a constant proportion of the infected resting CD4+ T cells contained defective genomes, then PCR measurements might provide a reliable surrogate measure of reservoir size in cross-sectional analysis. However, for the whole study population and the subset initiating HAART during chronic infection, a strong correlation (r>0.6) between results of the PCR assay for HIV-1 DNA and the viral outgrowth assay can be formally excluded. A larger sample size is needed to determine whether some correlation exists for patients initiating HAART early. It is possible that as the infection progresses, cells with defective viral sequences accumulate at different rates in different patients so that there is eventually very little correlation between the viral outgrowth assay and PCR-based assays. The ratio of infected cell frequencies determined by the two different assays varies by over 3 logs, with a subset of the patients who initiated therapy during chronic infection showing very high ratios. Taken together, these results are consistent with the hypothesis that differential accumulation of resting CD4+ T cells with defective viral sequences obscures the relationship between the frequency of cells detected in the virologic and molecular assays.
Most patients start therapy during chronic infection, and it is problematic that readily scalable PCR assays for total HIV-1 DNA on PBMC do not provide a precise reflection of the size of the latent reservoir, at least for cross-sectional analysis. It remains possible that PCR assays will be useful in following individual patients participating in eradication trials, but it is not yet clear that latency-reversing strategies will cause proportionate reductions in the latent reservoir and in the total pool of cells with HIV-1 DNA. For example, some cells with defective viral genomes may not express viral genes in response to latency-reversing agents. Hypermutated genomes typically have stop codons in every open reading frame [42], and thus cells carrying hypermutated genome may not be eliminated by latency reversing strategies that depend on viral cytopathic effects or CTL-mediated clearance. These cells might still be detected by PCR-based assays even when cells with replication-competent viral genomes were being eliminated.
Initial studies established that the latent reservoir in resting CD4+ T cells consists of cells with stably integrated viral genomes, [4], [5]. We therefore evaluated a well established Alu PCR assay specific for integrated viral genomes. Infected cell frequencies determined by this method were similar to and well correlated with frequencies determined by the droplet digital PCR assay, which does not distinguish integrated and unintegrated viral genomes (r = 0.85, P = 0.0079 for resting CD4+ T cells). These results are consistent with the notion that most of the HIV-1 DNA in resting CD4+ T cells of patients on HAART is integrated. Linear unintegrated forms, which are prevalent in untreated patients [35], are labile [37], [68] and are not seen in the absence of ongoing viral replication. Circular unintegrated forms (2-LTR circles), when detected, were present only at extremely low levels. Although levels of integrated HIV-1 DNA correlate well with measurements of HIV-1 DNA by the droplet digital PCR assay, both assays can detect defective as well as replication-competent proviruses. Among the approaches evaluated, analysis of integrated HIV-1 DNA in PBMC showed the best correlation with results of the viral outgrowth assay on purified resting CD4+ T cells (Table 4, r = 0.70, P = 0.0008). The correlation was weaker when integrated HIV-1 DNA was measured in purified resting CD4+ T cells.
Because the GALT provides a major site for HIV-1 replication [57]–[61], we also measured the HIV-1 DNA and RNA levels in rectal biopsy samples. When normalized for the number of CD4+ T cells present, the DNA PCR assay gave infected CD4+ T cell frequencies that were significantly higher than infected cell frequencies in the blood, consistent with previous studies [43], [44]. HIV-DNA levels in rectal biopsy samples showed a modest correlation with HIV-1 DNA levels in cells from the peripheral blood, but not with results of the viral outgrowth assay. Overall, these results highlight the potential importance of the GALT as an HIV-1 reservoir. However several critical questions remain. It is important to understand the fraction of CD4+ T cells in the GALT that can produce replication-competent virus and whether or not the virus is generally latent in these sites. The HIV-1 RNA/DNA ratios measured in these samples were generally <1, but these values are difficult to interpret because of uncertainty regarding the distribution of RNA molecules among infected cells. Interestingly, RNA/DNA ratios in rectal biopsy samples showed a statistically significant correlation with the viral outgrowth assay (r = 0.57, P = 0.013). This may reflect the fact that the RNA∶DNA ratio provides some indication of the number of infected cells that have the capacity to produce viral RNA.
As a measure of viral persistence, the single copy assay for residual viremia is of particular interest because it detects ongoing virus production in patients on suppressive HAART regimens. Residual viremia was only detectable in two thirds of study subjects and did not correlate with infected cell frequencies as measured by the viral outgrowth assay. A precise correlation might be expected if the activation of latently infected resting CD4+ T cells was the major source of residual viremia. However, recent studies [69]–[71] have show that in many patients the residual viremia is dominated by viral clones that are profoundly underrepresented in resting CD4+ T cells from the peripheral blood, suggesting an additional source of residual viremia.
Another important factor in evaluating assays for persistent HIV-1 is the dynamic range. Specifically, it is important to understand how much of a reduction in the reservoir would be measurable with each assay assuming reasonable sample volumes. Among the assays evaluated, the single copy assay for residual viremia has the lowest operating range. The viral outgrowth assay is already run on large blood volumes (180 ml), and its dynamic range cannot be extended much further. PCR-based assays perhaps offer the largest dynamic range, but suffer from the caveats discussed above.
Other studies have compared assays for persistent HIV-1, and the results differ to some extent from the findings presented here. In an early study, Anton et al. compared several measures of persistent HIV-1 infection in patients on HAART [72]. Although these investigators noted some modest correlations between the results of different culture and PCR-based assays, their findings cannot be readily compared to those of the present study because the culture assays were not run on purified resting CD4+ T cells and the PCR-based assays were not normalized for input CD4 cell number. Chun et al. reported a weak but statistically significant correlation (r = 0.29, P = 0.001) between linear values for residual viremia measured with an assay that has a detection limit of 20 copies/ml and the level of HIV-1 DNA in resting CD4+ T cells [73]. Murray et al. [74] observed an accumulation of cells with HIV-1 DNA during the first two years of untreated HIV-1 infection, consistent with the differences in infected cells frequencies between the acute and chronic cohorts in our study. Cells with defective proviruses may fail to express viral proteins and may therefore be protected from viral cytopathic effects and host CTL. The cells will thus have a chance to accumulate. This may happen at different rates in different patients, greatly complicating measurement of the reservoir by PCR-based approaches. Of note, no other study has shown a precise correlation between the results of the viral outgrowth assay and any simpler assay.
Overall, the results of this study suggest that no PCR-based assay provides a precise and internally consistent indication of the amount of replication-competent HIV-1 in resting CD4+ T cells. These findings raise the important issue of how to quantify decreases in the latent reservoir in future HIV-1 eradication trials. The fundamental problem is that infected cell frequencies determined by PCR-based assays are at least 2 logs higher than infected cell frequencies determined by the viral outgrowth assay. Much of this difference may be due to cells carrying defective proviruses, for example those that have been lethally hypermutated by APOBEC3G. These defective proviruses may not be eliminated by strategies designed to target latently infected cells. In this situation, successful clearance of latently infected cells might be masked by a larger and unchanging pool of cells with defective proviruses. While PCR-based assays may overestimate the size of the reservoir, the viral outgrowth assay provides only a minimal estimate of the frequency of cells harboring replication-competent virus. The assay conditions were carefully chosen to induce blast transformation in 100% of the input resting CD4+ T cells [6]. In this situation, the failure of an infected cell to produce replication-competent virus can be due to defects in the provirus such as APOBEC3G-induced hypermutation [42] or large internal deletions [41]. However, epigenetic silencing [75], transcriptional interference [76], [77], and other factors could also in principle prevent some proviruses without intrinsic defects from scoring in the viral out growth assay. Importantly, no other culture assay, including those that use alternative T cell activating stimuli [7], [8], has detected a higher frequency of cells with replication-competent virus. Nevertheless, the development of assays that precisely quantitate the number of the latently infected cells that have the potential to release replication-competent virus in vivo is an important goal in eradication research.
This study enrolled 30 patients from two well established cohorts at the University of California San Francisco (UCSF). All study subjects provided informed consent. Twenty patients were from the SCOPE cohort, an ongoing longitudinal study of ∼1500 HIV-1-infected and uninfected adults. Infected individuals in this cohort started HAART during chronic infection (>180 days from estimated date of infection). Subjects were seen and interviewed at four-month intervals to ascertain: (1) current medications, (2) medication adherence, (3) recent intercurrent illnesses, and (4) recent diagnoses or hospitalizations. Plasma HIV-1 RNA levels and routine T cell immunophenotyping were performed at each visit.
Ten patients were recruited from the OPTIONS cohort, an ongoing longitudinal study of adults enrolled within 8 months of HIV-1 infection. The evaluation of patients with possible acute infection included detailed HIV-1 testing and exposure history and the following laboratory studies: (1) a high-sensitivity assay for plasma HIV-1 RNA, (2) a standard HIV-1 antibody EIA with western blot confirmation if positive, and (3) a less-sensitive (detuned) antibody EIA (LS-EIA). Screened subjects met one or more of the following criteria to be defined as having primary or recent HIV-1 infection: (1) repeated plasma HIV-1 RNA >5,000 copies/ml combined with a negative or indeterminate HIV-1 antibody test; (2) seroconversion within 6 months of a documented negative HIV-1 antibody test or (3) a history compatible with primary HIV-1 infection (including no prior positive HIV-1 antibody tests) and laboratory testing consistent with recent infection on the “detuned” antibody EIA. Eligible subjects were followed approximately every 3 months.
Eligibility criteria for all patients enrolled in the present study included: (1) confirmed HIV-1 infection, (2) documented prior initiation of one of the Department of Health and Human Services recommended/alternative HAART regimens [78], (3) at least 36 months of continuous HAART at study entry with no regimen changes in the preceding 24 weeks, (4) maintenance of plasma HIV-1 RNA levels below the limit of detection of conventional assays for at least 36 months (intermittent isolated episodes of detectable low-level viremia were allowed), (5) most plasma HIV-1 RNA levels below the level of detection (<40 copies RNA/ml), and (6) documented CD4+ T-cell count above 350 cells/µl for preceding 24 weeks. We excluded subjects who (1) had recent hospitalization, (2) recent infection requiring systemic antibiotics, (3) recent vaccination or (4) exposure to any immunomodulatory drug (including maraviroc) in the preceding 16 weeks.
All subjects who met entry criteria for screening and agreed to participate in the study had an initial structured interview, phlebotomy, and HIV-1 testing using the established SCOPE and OPTIONS infrastructure. Once eligibility for the large volume blood draw (220 ml) and gut biopsy procedures were determined, blood was collected in tubes containing acid-citrate-dextrose (ACD), and 180 ml of the sample were shipped overnight at ambient temperature to the Johns Hopkins School of Medicine where peripheral blood mononuclear cells (PBMC) and resting CD4+ T lymphocytes were isolated for quantitative studies of proviral DNA and replication competent virus. Extensive previous studies have shown that cells can be recovered with high viability for virus culture assays after overnight shipment [79]. The remaining peripheral blood (40 ml) was sent to the UCSF AIDS Specimen Bank for studies of the host response and additional virologic studies.
At John Hopkins, PBMC were isolated using density gradient centrifugation. Supernatant plasma was frozen at −80°C and sent on dry ice to Dr. Sarah Palmer of the Karolinska Institute in Stockholm, Sweden for the single copy assay for plasma HIV-1 RNA [46]. Aliquots of PBMC were frozen as cell pellets for analysis of total and integrated HIV-1 DNA. Resting CD4+ T cells were purified from PBMC by negative depletion using biotinylated antibodies and anti-biotin magnetic beads. Briefly, CD4+ T lymphocytes were first isolated from PBMC by removing unwanted cell populations (CD4+ T cell Isolation Kit II; Miltenyi). Non-CD4+ T cells were labeled first with a cocktail of biotin-conjugated monoclonal antibodies followed by incubation with anti-biotin-conjugated magnetic microbeads. Unwanted cells were then removed using a LS MACS Column with the MACS Separator magnet (Miltenyi). Activated CD4+ T lymphocytes were then removed from the total CD4+ T cell population by labeling unwanted cells with biotin-conjugated antibodies to CD25, CD69, and HLA-DR and anti-biotin microbeads (Miltenyi). Labeled cells were magnetically removed using MACS MS columns with the MACS separator. The resting CD4+ T cell population was typically 98–99% pure as assessed by FACS analysis. Latent HIV-1 in resting CD4+ T cells was quantitated using the viral outgrowth assay (see below). Aliquots of purified resting CD4+ T cells were frozen as cell pellets and sent along with aliquots of unfractionated PBMC to UCSD and the University of Pennsylvania for assays of total and integrated HIV-1 DNA. Details of individual assays are given below.
A subset of study subjects underwent a rectosigmoid biopsy at San Francisco General Hospital. Up to 30 3-mm biopsies were obtained 10–30 cm above the anus with a disposable biopsy forceps (3.3-mm outside diameter). Biopsy specimens were suspended in RPMI 1640 containing 10% fetal calf serum, piperacillin–tazobactam (500 µg/ml), and amphotericin B (1.25 µg/ml), and transported within 2–3 h to the UCSF Core Immunology Laboratory where the tissue was digested with collagenase and needle shearing (see below). Cells were counted and frozen as cell pellets until analyzed for HIV-1 DNA and RNA.
Blood and rectal biopsies from patients were obtained through protocols approved by the UCSF Committee on Human Research. For the virus culture assay, blood was obtained from healthy donors through a protocol approved by the Johns Hopkins University School of Medicine Internal Review Board #4. All study subjects provided written informed consent prior to participation in the study.
Acute/early and chronic cohorts were compared using 2-tailed t tests for independent samples. Log transformed virologic data met the D'Agostino-Pearson test for Normal distribution, and correlations were performed on log transformed data. For DNA PCR measurements of cell associated HIV-1 genomes, we assumed that infected cells carried a single provirus [84] and expressed the results as the frequency of infected cells. For culture assay and HIV-1 DNA data, one or two samples were below the limit of detection of the relevant assays. For these, an imputed value representing the lower of the limit of detection or a value corresponding to 1 percentile of a log normal distribution fitted to the measured values was used in the calculation of the Pearson correlation coefficient. In the case of the single copy assay for HIV-1 RNA in plasma, for which one-third of the measurements were below the limit of detection (0.2 copies/ml), a low imputed value of 0.1 was used to calculate the Spearman rank correlation coefficient. Similar results were obtained when an imputed value of 0.01 was used. Data analysis was done using Microsoft Excel and MedCalc software.
|
10.1371/journal.pcbi.1005725 | Deciphering HLA-I motifs across HLA peptidomes improves neo-antigen predictions and identifies allostery regulating HLA specificity | The precise identification of Human Leukocyte Antigen class I (HLA-I) binding motifs plays a central role in our ability to understand and predict (neo-)antigen presentation in infectious diseases and cancer. Here, by exploiting co-occurrence of HLA-I alleles across ten newly generated as well as forty public HLA peptidomics datasets comprising more than 115,000 unique peptides, we show that we can rapidly and accurately identify many HLA-I binding motifs and map them to their corresponding alleles without any a priori knowledge of HLA-I binding specificity. Our approach recapitulates and refines known motifs for 43 of the most frequent alleles, uncovers new motifs for 9 alleles that up to now had less than five known ligands and provides a scalable framework to incorporate additional HLA peptidomics studies in the future. The refined motifs improve neo-antigen and cancer testis antigen predictions, indicating that unbiased HLA peptidomics data are ideal for in silico predictions of neo-antigens from tumor exome sequencing data. The new motifs further reveal distant modulation of the binding specificity at P2 for some HLA-I alleles by residues in the HLA-I binding site but outside of the B-pocket and we unravel the underlying mechanisms by protein structure analysis, mutagenesis and in vitro binding assays.
| Predicting the differences between cancer and normal cells that are visible to the immune system is of central importance for cancer immunotherapy. Here we introduce a novel computational framework to harness the wealth of data from in-depth HLA peptidomics studies, including ten novel high quality (<1% FDR) datasets generated for this work, to improve predictions of peptides displayed on HLA-I molecules. These high-throughput and unbiased data enable us to refine models of HLA-I binding specificity for many alleles (including some that had no ligand until this study) and improve predictions of neo-antigens from exome sequencing data in melanoma and lung cancer samples. Moreover, the refined description of HLA-I binding specificity reveals cases of allosteric modulation of HLA-I binding specificity at the second amino acid position (P2) of their ligands by residues that are part of the HLA-I binding site but outside of the B pocket.
| HLA-I molecules play a central role in defence mechanisms against pathogens and immune recognition of cancer cells. Their main functionality is to bind short peptides (mainly 9- to 12-mers) coming from degradation products of endogenous or viral proteins. The peptides are cleaved in the proteasome, transported by the transporter associated with antigen processing (TAP) complex, loaded onto the HLA-I molecules in the ER and presented at the cell surface [1]. Non-self peptides presented on HLA-I molecules, such as those derived from degradation of viral proteins, mutated proteins (referred to as neo-antigens), and other cancer specific and abnormally expressed proteins can then be recognized by CD8 T cells and elicit cytolytic activity. Neo-antigens have recently emerged as promising targets for development of personalized cancer immunotherapy [2].
Human cells express three HLA-I genes (HLA-A/B/C). These genes are the most polymorphic of the human genome and currently more than 12,000 different alleles have been observed [3]. Such a high polymorphism makes it challenging to model the different binding specificities of each allele and predict antigens presented at the cell surface. Information about binding motifs (mathematically defined here as Position Weight Matrices and graphically represented as sequence logos) of HLA-I molecules has been mainly obtained from biochemical assays where chemically synthesized peptides are tested in vitro for binding. This in vitro approach is experimentally laborious, time consuming and expensive. Currently, the most frequent HLA-I alleles have thousands of known ligands that provide a detailed description of their binding specificity. Many of these ligands are stored in very important resources such as IEDB [4,5] and have been used to train machine learning algorithms for HLA-I ligand predictions [6–11]. As a result, for frequent alleles in Caucasian populations, much is known about their binding specificity. However, the vast majority (>95%) of HLA-I alleles still lack documented ligands and despite very valuable algorithmic developments to generalize prediction methods to any allele [12], it remains more challenging to make accurate predictions for alleles without known ligands. Importantly, although these alleles are only found in a small fraction of the Caucasian population, they are more frequently encountered in other ethnic groups and gaining more understanding of their binding specificity would be desirable for expanding the scope of therapeutic strategies relying on HLA-I ligand predictions. Moreover, many alleles with known motifs are supported by only a few tens of peptides and some of these ligands have been selected based on a priori expectations of the binding specificity rather than unbiased screening of random peptide libraries. Such potentially biased datasets can be sub-optimal for training HLA-I ligand predictors.
Mass-spectrometry (MS) analysis of HLA-I binding peptides eluted from cell lines or tissue samples is a promising alternative to the use of HLA-I ligand interaction predictions [13] and MS is increasingly used to directly identify viral [14,15] or cancer-specific (neo-)antigens [13,16–20]. For neo-antigen discovery, tumor exome sequencing is first performed. Patient-specific non-synonymous somatic alterations are included in a customized database. MS-MS spectra of HLA ligands eluted from the patient’s tissue samples are searched against this expanded database permitting either the wild-type or the mutated peptide sequences to be identified. Neo-antigens directly identified this way, or alternatively predicted in-silico, may be further validated using targeted mass-spectrometry approaches in which isotopically heavy-labeled synthetic peptides are spiked into the HLA ligands eluted from the patient’s tumor tissue. Identification of co-eluting pairs of heavy (standard) and light (endogenous) peptides validate the presentation of the neo-antigen in the investigated tissue [19]. However, this technique is only applicable to a small fraction of samples due to the large amount of material that is required for MS analysis (typically 1cm3 of tissue material, or 1x108 cells in culture) and the complexity of these experiments.
In addition to potentially immunologically relevant (neo-)antigens, tens of thousands of endogenous peptides naturally presented on HLA-I molecules are identified in such HLA peptidomics studies, providing a unique opportunity to collect very large numbers of HLA-I ligands that can be used to better understand the binding properties of HLA-I molecules. The challenge in studying HLA-I motifs based on such pooled peptidomics data from unmodified cell lines or tissue samples is to determine the allele on which each peptide was displayed. The most widely used approach is to predict binding affinity of each peptide to each allele present in a sample [21]. Recent studies by ourselves and others have shown that HLA-I motifs can be identified in HLA peptidomics datasets in an unsupervised way by grouping peptides based on their sequence similarity [17,22–25]. However, this strategy still relies on previous information about HLA-I binding specificity when associating predicted motifs with HLA-I alleles and is therefore restricted to alleles whose motifs have been already characterized.
Here, we describe a computational framework for direct identification and annotation of dozens of HLA-I motifs without any a priori information about HLA-I binding specificity by taking advantage of co-occurrence of HLA-I alleles across both newly generated and publicly available HLA peptidomics datasets. Our approach recapitulates and refines motifs for many common alleles and uncovers new motifs for eight alleles for which, until this study, no ligand had been documented. Importantly, this approach is highly scalable and will enable continuous refinement of motifs for known alleles and determination of novel motifs for uncharacterized alleles as more HLA peptidomics data will be acquired in the future. Training HLA-I ligand predictors based on the refined motifs significantly improves neo-antigen predictions in tumor samples with experimentally determined neo-antigens. Our large collection of HLA-I ligands further allowed us to unravel some of the molecular determinants of HLA-I binding motifs and revealed allosteric modulation of HLA-I binding specificity. To elucidate the underlying molecular mechanisms, we show how a single point mutation (W97R) in HLA-B14:02 outside of the B pocket significantly changes the amino acid preferences at P2 in the ligands.
To study the binding properties of HLA-I alleles without relying on a priori assumption on their binding specificity and investigate whether this unbiased approach could improve neo-antigen predictions from exome sequencing data, we reasoned that HLA-I binding motifs might be identified across samples with in-depth and accurate HLA peptidomics data by taking advantage of co-occurrence of HLA-I alleles. To this end, we measured the HLA peptidome eluted from six B cell lines, two in vitro expanded tumor-infiltrating lymphocytes (TILs) samples and two leukapheresis samples (peripheral blood mononuclear cells) selected based on their high diversity of HLA-I alleles (see Methods and S1 Dataset). By applying a stringent false discovery rate for peptides identification of 1%, we accurately identified 47,023 unique peptides displayed on 32 HLA-I molecules. To expand the coverage of HLA-I alleles, we further collected 40 publicly available high-quality HLA peptidomics datasets [17,18,22,23,26–28] (see Methods and S2 Dataset). Our final data consists of a total of 50 HLA peptidomics datasets covering 66 different HLA-I alleles (18 HLA-A, 32 HLA-B and 16 HLA-C alleles, see Table A in S1 Supporting Information). The number of unique HLA-I ligand interactions across all samples reaches 252,165 for a total of 119,035 unique peptides (9- to 14-mers), which makes it, to our knowledge, the largest currently available collection of HLA peptidomics datasets both in terms of number of peptides and diversity of HLA-I molecules. Binding motifs in each HLA peptidomics dataset were identified for 9- and 10-mers (see Fig A in S1 Supporting Information) using a motif discovery algorithm initially developed for multiple specificity analysis [29,30] and recently applied to the analysis of a small dataset of seven HLA peptidomics studies [24]. Importantly, this method does not rely on HLA-I peptide interaction predictions (see Methods).
To assign each motif to its allele even in the absence of a priori information about the alleles’ binding specificity, we developed a novel computational strategy illustrated in Fig 1. In this example, one allele (HLA-A24:02) was shared between all three samples. Remarkably, exactly one identical motif was shared between the three samples. As such, one can predict that this motif corresponds to the shared allele. Similarly, two alleles (HLA-A01:01 and HLA-C06:02) were shared between exactly two samples and here again two motifs were shared among the corresponding samples, and could therefore be annotated to their corresponding alleles. Moreover, if one sample shares all but one allele with another sample, it can be inferred that the motif that is not shared corresponds to the unshared allele (see example in Fig B in S1 Supporting Information), even if some of the shared motifs have not been annotated yet. Finally, if all but one motif had been annotated in a sample to all but one allele, one can infer that the remaining motif corresponds to the remaining allele. These three ideas can then be recursively applied to identify HLA-I motifs across our large collection HLA peptidomics datasets (see Methods). Of note, motifs identified in distinct samples that have some alleles in common show very high similarity (Fig 1 and Fig B in S1 Supporting Information) and our new approach builds upon this remarkable inherent reproducibility of in-depth and accurate HLA peptidomics data.
We applied our algorithm to the 50 HLA peptidomics datasets considered in this study. In total, motifs could be found for 44 different alleles without relying on any a priori assumption of HLA-I binding specificity (Fig 2A). These include seven alleles (HLA-B13:02, HLA-B14:01, HLA-B15:11, HLA-B15:18, HLA-B18:03, HLA-B39:24 and HLA-C07:04) that did not have known ligands in IEDB, and 5 additional ones (HLA-B38:01, HLA-B39:06, HLA-B41:01, HLA-B56:01, HLA-C07:01) that had less than 50 known ligands. To validate our predictions, we compared the motifs predicted by our fully unsupervised method with known motifs derived from IEDB [4], when available. Despite some differences (e.g. HLA-A25:01 motif at P9) affecting especially alleles with low number of ligands in IEDB (Fig C in S1 Supporting Information), we observed an overall high similarity confirming the reliability of our predicted motifs (Fig 2A). However, it is important to realize that even small differences in the motifs can have important effects on the performance of predictors that are trained on such data when ranking very large lists of potential epitopes. When comparing with data recently obtained by HLA peptidomics analysis of mono-allelic cell lines [31], a very high similarity was also observed (Fig 2B, stars in Fig C in S1 Supporting Information), which further validates our computational approach for HLA-I motif identification and annotation from in-depth pooled HLA peptidomics data.
As expected from many previous studies, alleles with the same two first digits code showed high similarity in their binding specificity, apart from HLA-B15 alleles which are known to be more diverse [32]. This includes many of the new motifs (e.g., HLA-B14:01 vs HLA-B14:02; HLA-B18:03 vs HLA-B18:01; HLA-B39:24 vs HLA-B39:06), which provides further evidences of the accuracy of our predictions for these uncharacterized alleles.
For the most frequent HLA-I alleles, including several shown in Fig 2A, a good description of their binding motifs can be already obtained from existing databases [4]. To further expand our collection of HLA-I binding motifs, we used similarity to the binding motifs derived from IEDB ligands to annotate motifs that could not be assigned to their corresponding allele by the fully unsupervised algorithm, following the approach previously introduced by ourselves in ref [24] (see Methods). This enabled us to determine the binding motifs of 8 additional alleles (Fig 2C). Of note, the new motif of HLA-A02:20 was predicted by observing that it was the only motif not annotated in one sample (RA957) and could only be annotated to this allele based on the motifs identified in other samples for all the other alleles (see Fig A in S1 Supporting Information). The final list of motifs for the 52 alleles and detailed comparison with IEDB derived data, when available, is shown in Fig D in S1 Supporting Information. Importantly, for the majority of alleles considered in this study, the motifs are supported by significantly more ligands than what is available in existing databases (Fig E in S1 Supporting Information), and in total our approach enabled us to collect 88’051 unique 9-mer peptide HLA-I interactions for all alleles annotated in this work, compared to the 57’651 interactions available in IEDB for the same set of alleles. 14 out of a total of 195 motifs (corresponding to 5’703 9-mer peptide-HLA interactions) could not be annotated by our approach (see Fig A in S1 Supporting Information).
To investigate how our approach depends on the number of samples in which a motif is found, we show in Fig F in S1 Supporting Information the distance between our predicted motifs and those derived from mono-allelic cell lines (Fig F(a) in S1 Supporting Information) or those pooled from all samples (Fig F(b) in S1 Supporting Information) as a function of the number of samples (i.e. sub-sampling). As expected, higher similarity could be observed by integrating several samples, justifying our idea of collecting as much data as possible from different HLA peptidomics datasets to refine our motifs. But overall, all distances are very small (D2 < 0.03), highlighting the excellent reproducibility of the motifs deconvoluted from HLA peptidomics datasets.
To explore the statistical significance of the motifs associated to the same alleles, we followed the approach of ref. [33] and compared the similarity (both Euclidean distance and BLiC score) between each pair of motifs (mi, mj) annotated to the same allele h (i = 1, …Nh, j = 1…Nh, i≠j with Nh the number of motif annotated to allele h) to the distribution of similarity values between motif mi and all known HLA-I motifs [33] (see Methods). Fig G in S1 Supporting Information shows that more than 99% of the pairs of motifs associated to the same alleles have a statistically significant similarity (P<0.05), confirming the few examples shown in Fig 1. Exceptions consist mainly of motifs annotated to HLA-C alleles, which are more degenerate and therefore more difficult to deconvolute [24]. We therefore recognize that our motifs are likely less accurate for HLA-C alleles, but emphasize that these alleles are also poorly described in existing databases or literature.
We further explored the effect of the threshold T = 0.078 on Euclidean distance manually defined in this work (see Methods). As expected lower values of T still results in highly similar motifs annotated to the same alleles, but in fewer alleles to which motifs can be annotated (Fig H in S1 Supporting Information). Reversely, larger values tend to increase the number of alleles with annotated motifs, but at some point (T>0.09) more distinct motifs (P>0.05, based on Euclidean distance) become associated to the same alleles (Fig H in S1 Supporting Information).
The full pipeline was also applied on 10-mers identified by MS across the 50 HLA peptidomics studies and revealed six new motifs for poorly characterized alleles in IEDB (Fig I in S1 Supporting Information).
Different technical biases may affect MS data, which could undermine their use for training HLA-I ligand predictors. To investigate this potential issue, we computed amino acid frequencies at non-anchor positions (P4 to P7) in our HLA peptidomics data, excluding alleles displaying anchor residues at these positions (see Methods and Table B in S1 Supporting Information). The reason for focusing on middle positions is that they display low specificity (especially in 9-mers, see discussion in [24,34,35] for longer peptides) and therefore could provide a global view of potential MS biases on amino acid frequencies that is not affected by the constraints of binding to HLA-I molecules.
As expected, we observed a good correlation between amino acid frequencies at non-anchor positions in our HLA peptidomics data and in the human proteome (r = 0.85) (Fig 3 and Fig J in S1 Supporting Information). The most important difference that could strongly affect predictors trained on such data was found for cysteine, which is prone to post-translational modifications that are typically not included in database searches and was observed at very low frequency in the HLA peptidomics data (the same observation was recently made in mono-allelic cell lines [31]). Moreover, IEDB data clearly indicate that cysteine can be found at non-anchor positions, including for immunogenic epitopes, and therefore the low frequency observed in MS data very likely corresponds to a technical bias. Other amino acids were less under- or over-represented and the differences observed with the human proteome may also reflect some residual specificity at non-anchor positions. Moreover, no clear pattern emerged from these data with respect to amino acid biophysical properties (e.g., charge, hydrophobicity, size). Overall, our results suggest that HLA peptidomics data do not show strong technical biases, apart from under-representation of cysteine (see next section for a proposed method on how to compensate this bias), and therefore could provide ideal data for training HLA-I peptide interaction predictors, especially for ligands coming from human cells like neo-antigens.
To test whether our unique dataset of naturally presented peptides could help predicting HLA-I ligands, including neo-antigens in tumors based on exome sequencing data, we trained a predictor of HLA-I ligands (referred to as MixMHCpred). As MS only includes positive examples and HLA-I ligands in general do not show strong amino acid correlations between different positions (see discussion in [24] for some exceptions), we built Position Weight Matrices (PWMs) for each of the 52 alleles. These PWMs were built by pooling together all peptides assigned to each allele across all our HLA peptidomics datasets (see Methods and Fig 1). We further included MS data from mono-allelic cell lines for 6 rare alleles that were not present in our dataset, resulting in a total of 58 alleles available in our predictor. To correct for the low detection of cysteine observed in HLA peptidomics data we further renormalized our predictions by amino acid frequencies at non-anchor positions (see Methods and Table B in S1 Supporting Information).
As a first validation, we attempted to re-predict naturally presented peptides experimentally identified in ten mono-allelic cell lines whose alleles overlapped with our dataset [31]. For this analysis, we did not include data from these mono-allelic cell lines in our training set. To assess our ability to predict naturally presented peptides, we added 99-fold excess of decoy peptides randomly selected from the human proteome to each mono-allelic cell line dataset and measured the fraction of Positives among the top 1% Predictions (PP1%, i.e., True Positive Rate among the top 1%), which in the case of 99-fold excess of decoy is equivalent to both the Precision and the Recall, since the number of predictions (top 1%) is equal to the number of actual positives (1%). For all but one allele, our algorithm showed higher or equal predictive power compared to standard HLA-I ligand predictors [8,12,36] (Fig 4A). We further measured the average Area Under the Curve (AUC) for these alleles and obtained quite similar values (0.978 for MixMHCpred, 0.976 for NetMHC, 0.979 for NetMHCpan and 0.977 for NetMHCstabpan). However, we emphasize that most random peptides used as negatives are quite distinct from the positives, which can explain the very high AUC values and we suggest that precision values for top 1% of the predictions shown in Fig 4A are more representative of the actual performance of the algorithms.
We then collected currently available datasets that included direct identification of neo-antigens displayed on cancer cells as well as exome sequencing data (Mel5, Mel8, Mel15 from [17] and 12T from [20]) for a total of ten 9- and 10-mers mutated peptides experimentally found to be presented on cancer cells (see Table 1). This dataset has the unique advantage of not being restricted to peptides selected based on in silico predictions, and is therefore an ideal testing set for benchmarking our predictor. Moreover, as these studies are quite recent, the neo-antigens used here as testing set are not part of the training set of any existing algorithm. In particular, they are not part of the large training set used in this study since we only included wild-type human peptides in our pipeline. We then retrieved all possible 9- and 10-mer peptides that encompassed each missense mutation (S3 Dataset) and ranked separately for each patient these potential neo-antigens based on the score of our predictor (see Methods and Table 1). Remarkably, six of the ten neo-antigens fell among the top 25 predicted peptides, suggesting that by testing as few as 25 mutated peptides per sample, we could identify more than half of the neo-antigens identified by MS (Table 1). Considering that the total number of potential neo-antigens (i.e. 9- and 10-mers containing a missense mutation) can be as large as 25,000 for tumors with high mutational load, our predictor trained on naturally presented human HLA-I ligands clearly enabled us to significantly reduce the number of peptides that would need to be experimentally tested to identify neo-antigens from exome sequencing data. We further added two datasets of neo-antigens identified in lung cancer patients (L011 and L013) [37], although only peptides pre-selected based on binding affinities predicted with existing tools [12] were tested in this study. Here again, our predictor ranked one neo-antigen in the top 25 predicted peptides in both samples (Table C in S1 Supporting Information). When comparing with standard tools that are widely used to narrow-down the list of potential neo-antigens predicted from exome sequencing data [8,12,36], our method trained on HLA peptidomics data showed clear improvement (Table 1) with a mean AUC value of 0.979, compared to 0.932 for NetMHC [8], 0.942 for NetMHCpan [12] and 0.945 for NetMHCstabpan [36] (Fig 4B) and increased number of neo-antigens in the top 1% of the predictions (i.e., typically what is experimentally tested for immunogenicity) across all six samples (Fig 4C). This is especially clear for the 12T sample, where the single neo-antigen was very well predicted by our model and poorly predicted by existing tools (>6’000nM with HLA-B51:01, see also Table 1 and [20]). We still note that, due to the low number of neo-antigens publicly available together with exome sequencing data, performance metrics in Fig 4B and 4C can be sensitive to one neo-antigen being better or less well predicted and we stress that the values shown in Fig 4B and 4C are simply a graphical way of looking at data shown in Table 1 and Table C in S1 Supporting Information. Importantly, even if we did not include in the training of our predictor MS data (i.e. wild-type peptides) from the samples in which the neo-antigens were identified, neo-antigens were still more accurately predicted compared to other tools (see Fig K in S1 Supporting Information). This demonstrates that our approach for neo-antigen predictions from the list of somatic mutations identified by exome sequencing of tumors does not require HLA peptidomics data from the same sample where neo-antigens had been identified. Nevertheless, both our predictor and standard prediction tools failed to identify some neo-antigens (e.g., KLILWRGLK from NCAPG2 P333L mutation, see Table 1). This suggests that, when enough tumor material is available for HLA peptidomics analyses, direct identification of neo-antigens with MS should still be performed to optimally enrich in true positives the list of ligands to be experimentally tested for immunogenicity [16,19].
The number of studies reporting both neo-antigens and exome sequencing results is still limited. To benchmark our algorithm with larger datasets of immunologically relevant tumor antigens, we tested our ability to predict epitopes from cancer testis antigens. We retrieved all epitopes listed in the CT database [38] (see Methods and Table D in S1 Supporting Information). We then assessed how our predictor could prioritize these epitopes from all possible peptides encoded by these cancer testis antigens. Although we cannot exclude that some of these epitopes had been selected for experimental testing after prediction by older versions of HLA-I ligand predictors, we still observed improvement using our predictor trained only on naturally presented HLA-I ligands, both in terms of AUC and fraction of true positives that fall in the top 1% of the predictions (Fig 4D and 4E). This indicates that improvement in prediction accuracy is not restricted to elution data (see similar results in [24]).
MS data can contain false positives for many different reasons, such as co-eluting peptide contaminations or errors in the computational identification of peptides from the spectra. Therefore, despite the high quality of HLA peptidomics datasets generated in this study (<1% FDR), we do expect our data to contain some noise. To test the robustness towards contaminations of our motif discovery and annotation pipeline, and our HLA-I ligand predictor, we incorporated 5% of random peptides from the human proteome into all HLA peptidomics datasets considered in this work and rerun the whole motif annotation pipeline and training of the predictor. Remarkably, the accuracy of the predictions was only very modestly affected by this noise and predictions were still better than with other existing tools (Fig L in S1 Supporting Information). To explore the effect of wrong peptide identification, we reprocessed with MaxQuant [39] the three MS samples shown in Fig 1 and chose the second best hit for 1% of the peptides. Overall the motifs predicted by our approach remained almost unchanged (Fig M in S1 Supporting Information). This suggests that our pipeline is robust and indicates that the wealth of unbiased and accurate data provided by MS can compensate the inherent contaminations, when using these data for training HLA-I ligand predictors.
An important step in our predictor is the renormalization by amino acid frequencies observed at non-anchor positions, which was designed to correct for biases in MS data. As expected, doing this renormalization step with amino acid frequencies observed in the human proteome (or no renormalization at all) results in very low frequency of cysteine-containing peptides among the top predicted ligands. As such, it improves the predictions of MS data (see especially Fig N(a) in S1 Supporting Information), but decreases the performance in other datasets (e.g., Fig N(b-c) in S1 Supporting Information for L011 and L013). These results highlight the importance of carefully considering MS biases when including such data to train predictors in order to avoid over-fitting elution data. We anticipate that additional work may further improve this step, such as inclusion of cysteine modifications in spectral searches [31] or better estimations of the expected baseline amino acid frequencies in HLA peptidomics datasets.
One of our novel HLA-I motifs describes the binding specificity of HLA-A02:20 (Fig 2C). HLA-A02 binding motifs have been widely studied. However, HLA-A02:20 motif differs from standard HLA-A02 motifs at P1, with a clear preference for charged residues (Fig 5A). Interestingly, HLA-A02:20 is among the very few (<2%) HLA-A02 alleles that do not have a conserved lysine pointing towards P1 at position 66 (residue numbering follows 2BNQ X-ray structure hereafter). Instead an asparagine is found there (Fig 5A), and this residue is the only difference with the sequence of the very common A02:01 allele. To explore how the absence of lysine at position 66 could explain the observed difference in binding specificity, we collected all HLA-I alleles showing preference for charged amino acids at P1 (see Fig O in S1 Supporting Information). All of them had either asparagine or isoleucine at position 66. We then explored available crystal structures of HLA-I peptide complexes with charged residues at P1. HLA-B57:03 was crystalized with such a ligand (KAFSPEVI) [40]. Superposing the crystal structure of this complex with the structure of HLA-A02:01 provides a possible mechanism for understanding the change in binding specificity at P1. In HLA-A02:01, lysine at position 66 interacts with the hydroxyl group of serine at P1 (Fig 5A, green sidechains). Such a conformation would not be compatible with a longer residue. Reversely, when asparagine was found at position 66, it did not point towards P1 (Fig 5A, pink sidechains), thereby freeing space for larger sidechains like lysine or arginine at P1. Overall, our analysis indicates that the presence of asparagine at residue 66 may be responsible for the change in binding specificity between HLA-A02:01 and HLA-A02:20. More generally, our results suggest that lysine at residue 66 in HLA-I alleles strongly disfavours charged residues at P1.
The new motif identified for HLA-B15:18 (Fig 2A) displayed strong preference for histidine at P2, which is not often observed in HLA-I ligands. To gain insights into the mechanisms underlying this less common binding motif, we surveyed all alleles that show preference for histidine at P2 (Fig P(a) in S1 Supporting Information). Sequence and structure analysis showed that all of them have a conserved P2 binding site, commonly referred to as the B pocket (see Fig 5B). However, several HLA-B14 alleles have exactly the same B pocket but show specificity for arginine at P2 (Fig P(b) in S1 Supporting Information). Among them, HLA-B14:02 had the highest sequence similarity to HLA-B15:18, with only 8 different residues in the peptide binding domain, none of them making any contact with arginine at P2 in the crystal structure of HLA-B14:02 (orange residues in Fig 5C). This suggests that the difference in binding specificity at P2 between HLA-B14:02 and HLA-B15:18 is likely explained by allosteric mechanisms. Of particular interest is residue 97 (W in HLA-B14:02 and R in HLA-B15:18), which is in the HLA-I binding site and contacts the peptide (mainly P3 to P6, Fig 5C) but is more than 7Å away from the arginine sidechain at P2. This residue is part of a network of aligned aromatic residues (Y9, W97 and F116) in HLA-B14:02 (Fig 5C) compatible with pi-pi interactions. Interestingly, X-ray structures with Arg at position 97 (e.g., 4O2C) show a flip in the orientation of Y9 sidechain, which reduces the size of the B pocket. We therefore hypothesized that mutating residue 97 into arginine in HLA-B14:02 may indirectly modify the binding specificity at P2 and explain the preference for histidine observed in the HLA-B15:18 motif.
To test our hypothesis, we generated a construct for HLA-B14:02 wild-type (wt) and W97R mutant. We tested several ligands of HLA-B15:18 identified in our HLA peptidomics data with histidine at P2, which were predicted to show enhanced binding to HLA-B14:02 W97R. As expected, a strong decrease in binding stability was observed between HLA-B14:02 W97R and HLA-B14:02 wt (Fig 5D). Reversely, when testing the same peptides with arginine at P2, a significant increase in stability was observed between HLA-B14:02 W97R and HLA-B14:02 wt (Fig 5D). For instance, binding of the peptide AHTKPRPAL was fully abolished in HLA-B14:02 wt, but was rescued when changing histidine to arginine at P2. Although other residues may also play a role in the binding specificity differences between HLA-B14:02 and HLA-B15:18, all of them are further away from P2. Overall, our results show that HLA-I binding specificity at P2 can be modulated by amino acids outside of the B pocket, and that residue 97 can act as a switch of the binding specificity at P2. These binding experiments further confirm the motifs predicted for HLA-B14:01 and HLA-B15:18 alleles.
Despite decades of work to characterize the binding motifs of the most common HLA-I alleles, unbiased peptide screening approaches have not been commonly used in the past. This is mainly because both the N- and the C-terminus of the peptides are engaged in binding to HLA-I molecules, thereby preventing the use of high-throughput techniques for peptide screening like phage display. To address this issue, we developed a novel algorithm to rapidly identify and annotate HLA-I binding motifs in a fully unsupervised way using in-depth and accurate HLA peptidomics data from unmodified cell lines and tissue samples. This enabled us to refine models of binding specificity for many alleles with few ligands in existing databases and characterize the binding properties of eight HLA-I alleles that had no known ligands until this study. Our approach is conceptually similar to existing approaches to deconvolute peptide epitopes by identifying shared peptides between different pools showing T cell reactivity in ELISpot experiments, but had never been applied to motif annotation across HLA peptidomics datasets.
Remarkably, our predicted motifs displayed high similarity with known motifs for common alleles, including motifs derived from HLA peptidomics analyses of mono-allelic cell lines [31] (Fig 2), and the MS-induced technical bias (mainly low detection of cysteine) could be compensated by renormalization with expected amino acid frequencies. This suggests that HLA peptidomics data are optimal to train HLA-I ligand interaction predictors, as confirmed by our ability to accurately predict from exome sequencing data several neo-antigens identified in tumor samples. These observations are in line with recent results obtained with predictors trained on HLA peptidomics data from mono-allelic cell lines for 16 human class I alleles [31] and mouse class II alleles [43]. Moreover, similar results seem to be observed when including MS data in the training set of NetMHCpan tools [44]. Although mass spectrometry may miss a fraction of the actually presented and immunogenic neo-antigens, those detected by MS are likely presented at high level on cancer cells. Therefore, accurately predicting such dominant neo-antigens is promising to prioritize targets for cancer immunotherapy.
Importantly, the improvement in prediction accuracy we report here comes primarily from refinement of known HLA-I motifs, since the less frequent alleles for which we uncovered new motifs were in general not part of our testing sets. For instance, differences are observed at P9 between the motif of HLA-A25:01 obtained from HLA peptidomics data (preference for F/W/Y/L) and from IEDB ligands (preference for Y/L/M/F) (Fig 2A). This likely explains why the neo-antigen ETSKQVTRW was poorly predicted by standard tools [8,12] (predicted IC50 > 3,000nM). Similar observations apply for the neo-antigen DANSFLQSV found in [20] (predicted IC50 > 6,000nM with HLA-B51:01). Although these differences may look relatively small on the logos (Fig 2A), they play an important role when ranking tens of thousands of potential epitopes with HLA-I ligand predictors. Moreover, as our predictor is only trained on naturally presented ligands, it may also capture some features of antigen presentation of endogenous peptides beyond the binding to HLA-I molecules. Along this line, it is interesting to note that a smaller improvement in prediction accuracy had been observed when attempting to predict ligands from the SYFPEITHI database [45] (including a large fraction of viral peptides) with HLA peptidomics data [24]. Although the dataset used in this previous study was significantly smaller, this observation suggests that HLA peptidomics data may be especially well suited for training predictors of human endogenous or mutated HLA-I ligands. Overall, our work highlights the importance of carefully determining HLA-I motifs, including for alleles that already have some known ligands, based on unsupervised analysis of naturally presented human HLA-I ligands for neo-antigen discovery.
In this work, we did not attempt to optimize the prediction algorithm itself, but rather focused on optimizing the training data and carefully correcting for MS biases, which in our view is more important for improving predictions of HLA-I ligands, since HLA-I ligands do not display strong correlations among the different residue positions. Nevertheless, we cannot exclude that using neural networks or other machine learning tools may further improve the predictions, and we anticipate that our large collection and assignment of unbiased HLA-I ligands may help exploring new amino acid correlation patterns among HLA-I ligands (see example in [24]).
Currently our predictor is limited to 9- and 10-mers, which is the most common length of HLA-I ligands and accounts for more than 80% of the HLA peptidome [17]. Although motif identification may work in some cases for 11-mers [24], the automated motif deconvolution and annotation becomes less accurate, especially for samples with less than 10’000 peptides identified by MS. Therefore, rather than including sub-optimal motifs in our predictor, we focused in this work on 9- and 10-mers. We anticipate that manual curation of HLA-I motifs in pooled HLA peptidomics datasets or the use of mono-allelic cell lines [31] may be more appropriate for training predictors for longer peptides. Importantly, not including 11-mers has no influence on the predictions for 9- and 10-mers, since peptides of different length are treated separately in the current framework (see ref. [8] for possible algorithmic extensions to include peptides of different length in the training set of HLA-I ligand predictors).
Our work enabled us to identify motifs for uncharacterized alleles and is to date the predictor entirely trained on naturally presented peptides with the largest allelic coverage, including all frequent HLA-I alleles in the Caucasian population. However, the number of HLA-I alleles for which predictions are available (58 in total) is still smaller than what other tools can do (especially tools like NetMHCpan [12] that can make predictions for any allele). Despite this limitation, we emphasize that our work provides the first scalable framework to integrate HLA peptidomics datasets that will be or are being generated for neo-antigen discovery in cancer immunotherapy and therefore will enable increasing the allele coverage as new studies are published. Moreover, our results suggest that improving existing models describing the binding specificity of relatively common HLA-I alleles may be as important as expanding allele coverage to rare alleles for neo-antigen discovery.
All HLA peptidomics datasets used in this work were generated with only 1% FDR and are of high purity. For this reason, and also to prevent including potential biases or removing important data, we decided not to filter our data with existing HLA-I ligand predictors, but we expect some contaminants in our large sets of peptides. Moreover, in a few cases, the motifs for some alleles were not detectable (see HLA-C12:03 in Fig 1). This suggests that the (few) peptides binding to this allele may contaminate the other motifs. This is a known situation when analysing HLA peptidomics data with unsupervised approaches [24,25,46]. As previously observed, it affects especially HLA-C alleles which are often poorly expressed and whose binding specificities are more redundant [24,47]. However, our results show that some level of noise is tolerated for training our predictor and can still lead to improvement over existing tools (Fig 4 and Fig L in S1 Supporting Information). We also stress that no existing HLA-ligand interaction dataset used for training predictors is free of false-positives and for many alleles the number of ligands is significantly lower (Fig E in S1 Supporting Information).
In a few cases, the motifs of two alleles could not be split because of the very high binding specificity similarity of these alleles (e.g., HLA-C07:01 and HLA-C07:02 in Fig 1). We emphasize that this does not preclude the use of HLA peptidomics data for training HLA-I ligand predictors, since many other samples in our dataset contained only one of these two alleles together with other non-overlapping ones. As such our strategy takes advantage of our large collection of HLA peptidomics datasets to naturally overcome cases where the deconvolution could not be fully achieved in one given sample.
Direct identification of neo-antigens with MS shows higher specificity compared to our predictions based on exome sequencing data [16,17], as expected. However, it is important to realize that these experiments are challenging and can be carried out only in a small subset of patients with enough tumor material. Moreover, in many cases, no neo-antigen is found by MS. As such, our work provides a scalable approach to capitalize on large MS data obtained from some patients or cell lines in order to improve predictions of neo-antigens in other patients where MS analysis of the immunopeptidome could not be carried out and only exome sequencing data are available. This work may also find applications in other jawed vertebrate species where reference motifs for MHC-I alleles are poorly described, since the only requirement of our approach is the availability of MHC-I specific antibodies and MHC typing information.
Previous computational approaches for neo-antigen predictions have shown that incorporation of gene expression data could improve accuracy [31,48]. Considering that our predictor using only exome sequencing information (i.e., the list of somatic mutations) already enabled us to improve predictions in the cancer samples analysed in this work, we anticipate that integrating additional information such as gene expression, when available, may lead to even more accurate predictions of neo-antigens.
Results shown in Fig 5 further emphasize the power of in-depth sampling of the HLA-I ligand space to inform us about molecular mechanisms underlying HLA-I binding properties [49]. Considering the rapid expansion of HLA peptidomics experiments performed in cancer immunotherapy research [13,17,18,20,27,28], we anticipate that our approach for HLA-I motif identification and annotation will enable similar analyses in the future to uncover other molecular determinants of HLA-I binding specificity.
Overall, our work shows for the first time that HLA-I motifs can be reliably identified across in-depth and accurate HLA peptidomics datasets without relying on HLA-I interaction prediction tools or a priori knowledge of HLA-I binding specificity. This unsupervised and scalable approach refines known HLA-I binding motifs and expands our understanding of HLA-I binding specificities to a few additional alleles without documented ligands. As such, this work is a powerful alternative to synthesizing every peptide for in vitro binding assays, or to genetically modifying [31] or transfecting cell lines with soluble HLA-I alleles [34,35,50], and may save substantially amount of money and time for HLA-I motif determination. Our results further contribute to our global understanding of HLA-I binding properties and improve neo-antigen predictions from exome sequencing data. This work may therefore facilitate identification of clinically relevant targets for cancer immunotherapy, especially when direct identification of neo-antigens with MS cannot be experimentally done.
Informed consent of the participants was obtained following requirements of the institutional review board (Ethics Commission, University Hospital of Lausanne (CHUV)).
We carefully selected ten donors expressing a broad range of HLA-I alleles and generated novel HLA peptidomics data (Table A in S1 Supporting Information). EBV-transformed human B-cell lines CD165, GD149, PD42, CM467, RA957 and MD155 were maintained in RPMI 1640 + GlutaMAX medium (Gibco, Paisley, UK) supplemented with 10% FBS (Gibco) and 1% Penicillin/Streptomycin Solution (BioConcept, Allschwil, Switzerland). TIL were expanded from two melanoma tumors following established protocols [51,52]. Informed consent of the participants was obtained following requirements of the institutional review board (Ethics Commission, University Hospital of Lausanne (CHUV)). Briefly, fresh tumor samples were cut in small fragments and placed in 24-well plate containing RPMI CTS grade (Life Technologies Europe BV, Switzerland), 10% Human serum (Valley Biomedical, USA), 0.025 M HEPES (Life Technologies Europe BV, Switzerland), 55 μmol/L 2-Mercaptoethanol (Life Technologies Europe BV, Switzerland) and supplemented with a high concentration of IL-2 (Proleukin, 6,000 IU/mL, Novartis, Switzerland) for three to five weeks. Following this initial pre-REP, TIL were then expanded in using a REP approach. To do so, 25 x106 TIL were stimulated with irradiated feeder cells, anti-CD3 (OKT3, 30 ng/mL, Miltenyi biotech) and high dose IL-2 (3,000 IU/mL) for 14 days. The final cell product was washed and prepared using a cell harvester (LoVo, Fresenius Kabi). Leukapheresis samples (Apher1 and 6) were obtained from blood donors from the Service régional vaudois de transfusion sanguine, Lausanne. Upon receival of TIL and leukapheresis samples, the cells were washed with PBS on ice, aliquoted and stored as dry pellets at -80°C until use. High resolution 4-digit HLA-I typing was performed at the Laboratory of Diagnostics, Service of Immunology and Allergy, CHUV, Lausanne.
W6/32 monoclonal antibodies were purified from the supernatant of HB95 cells grown in CELLLine CL-1000 flasks (Sigma-Aldrich, Missouri, USA) using Protein-A Sepharose (Invitrogen, California, USA).
We extracted the HLA-I peptidome from 2–5 biological replicates per cell line or patient material. The cell counts ranged from 1 x 108 to 3 x 108 cells per replicate. Lysis was performed with 0.25% sodium deoxycholate (Sigma-Aldrich), 0.2 mM iodoacetamide (Sigma-Aldrich), 1 mM EDTA, 1:200 Protease Inhibitors Cocktail (Sigma, Missouri, USA), 1 mM Phenylmethylsulfonylfluoride (Roche, Mannheim, Germany), 1% octyl-beta-D glucopyranoside (Sigma) in PBS at 4°C for 1 hr. The lysates were cleared by centrifugation with a table-top centrifuge (Eppendorf Centrifuge 5430R, Schönenbuch, Switzerland) at 4°C at 14200 rpm for 20 min. Immuno-affinity purification was performed by passing the cleared lysates through Protein-A Sepharose covalently bound to W6-32 antibodies. Affinity columns were then washed with at least 6 column volumes of 150 mM NaCl and 20 mM Tris HCl (buffer A), 6 column volumes of 400 mM NaCl and 20 mM Tris HCl and lastly with another 6 column washes of buffer A. Finally, affinity columns were washed with at least 2 column volumes of 20 mM Tris HCl, pH 8. HLA-I complexes were eluted by addition of 1% trifluoroacetic acid (TFA, Merck, Darmstadt, Switzerland) for each sample.
HLA-I complexes with HLA-I peptides were loaded on Sep-Pak tC18 (Waters, Massachusetts, USA) cartridges which were pre-washed with 80% acetonitrile (ACN, Merck) in 0.1% TFA and 0.1% TFA only. After loading, cartridges were washed twice with 0.1% TFA before separation and elution of HLA-I peptides from the more hydrophobic HLA-I heavy chains with 30% ACN in 0.1% TFA. The HLA-I peptides were dried using vacuum centrifugation (Eppendorf Concentrator Plus, Schönenbuch, Switzerland) and re-suspended in a final volume of 12 uL 0.1% TFA. For MS analysis, we injected 5 uL of these peptides per run.
Measurements of HLA-I peptidomics samples were acquired using the nanoflow UHPLC Easy nLC 1200 (Thermo Fisher Scientific, Germering, Germany) coupled online to a Q Exactive HF Orbitrap mass spectrometer (Thermo Fischer Scientific, Bremen, Germany) or with Dionex Ultimate RSLC3000 nanoLC (Thermo Fischer Scientific, Sunnyvale, CA) coupled online to an Orbitrap Fusion Mass Spectrometer (Thermo Fischer Scientific, San Jose, CA), both with a nanoelectrospray ion source. We packed an uncoated PicoTip 8μm tip opening with diameter of 50 cm x 75 um with a ReproSil-Pur C18 1.9 μm particles and 120 Å pore size resin (Dr. Maisch GmbH, Ammerbuch-Entringen, Germany) re-suspended in Methanol. The analytical column was heated to 50°C using a column oven. Peptides were eluted with a linear gradient of 2–30% buffer B (80% ACN and 0.1% formic acid) at a flow rate of 250 nl/min over 90 min.
Data was acquired with data-dependent “top10” method, which isolates the ten most intense ions and fragments them by higher-energy collisional dissociation (HCD) with a normalized collision energy of 27% and 32% for the Q Exactive HF and Fusion instruments, respectively. For the Q Exactive HF instrument the MS scan range was set to 300 to 1,650 m/z with a resolution of 60,000 (200 m/z) and a target value of 3e6 ions. The ten most intense ions were sequentially isolated and accumulated to an AGC target value of 1e5 with a maximum injection time of 120 ms and MS/MS resolution was 15,000 (200 m/z). For the Fusion, a resolution of 120,000 (200 m/z) and a target value of 4e5 ions were set. The ten most intense ions accumulated to an AGC target value of 1e5 with a maximum injection time of 120 ms and MS/MS resolution was 15,000 (200 m/z). The peptide match option was disabled. Dynamic exclusion of fragmented m/z values from further selection was set for 20 or 30 seconds with the Q Exactive HF and Fusion instruments, respectively.
We employed the MaxQuant computational proteomics platform [39] version 1.5.3.2 to search the peak lists against the UniProt databases (Human 85,919 entries, May 2014) and a file containing 247 frequently observed contaminants. N-terminal acetylation (42.010565 Da) and methionine oxidation (15.994915 Da) were set as variable modifications. The second peptide identification option in Andromeda was enabled. A false discovery rate of 0.01 was required for peptides and no protein false discovery rate was set. The enzyme specificity was set as unspecific. Possible sequence matches were restricted to 8 to 15 amino acids, a maximum peptides mass of 1,500 Da and a maximum charge state of three. The initial allowed mass deviation of the precursor ion was set to 6 ppm and the maximum fragment mass deviation was set to 20 ppm. We enabled the ‘match between runs’ option, which allows matching of identifications across different replicates of the same biological sample in a time window of 0.5 min and an initial alignment time window of 20 min.
To expand the number of samples and survey an even broader range of HLA-I alleles, we included in this study forty publicly available HLA peptidomics data from seven recent studies [17,18,22,23,26–28]. Only samples with HLA-I typing were used. Peptides identified in the recent study [18] in different repeats and under different treatments were pooled together to generate one list of unique peptides per sample. Since the published peptidomics datasets from Pearson et al. [28] were filtered to include only peptides with predicted affinity scores of less or equal to 1250 nM, we re-processed the mass spectrometer raw data using MaxQuant with similar settings as mentioned above except that peptide length was set to 8–25 mers (S2 Dataset). HLA typing information was retrieved from the original publications. In one case (THP-1 cell lines), the typing is controversial [53]. In this work, we used the typing determined by the authors of the HLA peptidomics study where the data came from [23]. The high fraction (>50% of the peptides) displaying a clear A24:02 and B35:03 motif based on our unsupervised deconvolution further indicates that these alleles are truly expressed in the sample on which the HLA peptidomics analysis was performed.
Known HLA-I ligands were retrieved from IEDB (mhc_ligand_full.csv file, as of Oct 2016) [4]. All ligands annotated as positives with a given HLA-I allele (i.e., “Positive-High”, “Positive-Intermediate”, “Positive-Low” and “Positive”) were used to build the IEDB reference motifs (Fig 2A and 2C). Ligands coming from HLA peptidomics studies analysed in this work were not considered to prevent circularity in the motif comparisons and because the HLA-I alleles to which these peptides bind were not experimentally determined. Position Weight matrices (PWMs) representing binding motifs in IEDB and used to compare with motifs derived from our deconvolution of HLA peptidomics datasets were built by computing the frequency of each amino acid at each position and using a random count of 1 for each amino acid at each position.
All 16 HLA peptidomics datasets obtained from mono-allelic cell lines were downloaded from [31]. Motifs used in the comparison presented in Fig 2B were built in the same way as for IEDB data. When benchmarking our ability to re-predict such data with, we considered 9-mers from the ten alleles that overlapped our set of deconvoluted HLA peptidomics data and added 99-fold excess of random peptides from the human proteome. The fraction of positives (i.e., MS peptides identified in these mono-allelic cell lines) predicted in the top 1% was used to assess the prediction accuracy with different methods. Peptides with a clear tryptic peptide signature (i.e., R/K at the last position for all alleles except HLA-A03:01 and HLA-A31:01) were manually removed.
An algorithm based on mixture models and initially developed for multiple specificity analysis in peptide ligands [29,30] was used to identify binding motifs in each dataset analysed in this work. Briefly, all peptides pooled by mass spectrometry analysis of eluted peptide-HLA-I complexes in a given sample were first split into different groups according to their size (9–10 mers). All 9- and 10-mers ligands were then modelled using multiple PWMs [24]. The results of such analysis consist of a set of PWMs that describe distinct motifs for each HLA peptidomics datasets (see Fig A in S1 Supporting Information) and probabilities (i.e., responsibilities) for each peptide to be associated with each motif. Sequence logos representing HLA-I motifs were generated with the LoLa software (http://baderlab.org/Software/LOLA). The command-line script for the mixture model (“MixMHCp”) can be downloaded at: https://github.com/GfellerLab/MixMHCp.
The availability of high-quality HLA peptidomics data from several samples with diverse HLA-I alleles suggest that one could infer which motifs correspond to which HLA-I alleles without relying on comparison with known motifs. For instance, if two samples share exactly one HLA-I allele, it is expected that the shared motif will originate from the shared allele (Fig 1). To exploit this type of patterns of shared HLA-I alleles, we designed the following algorithm:
Comparison of motifs was performed using (squared) Euclidean distance between the corresponding PWMs: D2=1L∑i=120∑A=1L(MiA−MiA′)2, where M and M’ stands for two PWMs (i.e., 20 x L matrices) to be compared and L is the peptide length. A threshold of T = 0.078 was used to define similar motifs based on visual inspection of the results. Cases of inconsistencies (i.e., distances larger that T) between motifs mapped to the same allele were automatically eliminated. The final binding motif for each HLA-I allele (Fig 2 and Fig D in S1 Supporting Information) was built by combining peptides from each sample that had been associated with the corresponding allele. Other measures of similarity between PWMs [33] did not improve the results (see for instance comparison between Euclidean distance and Jensen-Shannon divergence in Fig Q in S1 Supporting Information for the example shown in Fig 1) and we therefore used the Euclidean distance throughout this study. To further study the statistical significance of the similarity between motifs annotated to the same alleles, we computed the P-value of the similarity for each pair of motifs annotated to the same allele. To this end, we used both the Euclidean distance and the BLiC score with standard Dirichlet priors and hyper-parameters αj (j = 1…20) equal to one, as introduced in ref. [33]. This BLiC score was re-implemented using 20 amino acids (instead of the 4 nucleotides) and the background distribution was taken as the expected amino acid frequencies in MS data (see Table B in S1 Supporting Information). For a pair of motifs (mi, mj), empirical P-values were estimated by comparing the similarity between motifs mi and mj to the similarity between motif mi and all known HLA-I motifs (i.e, 107 in total, taking only alleles with more than 20 unique ligands in IEDB and excluding the allele to which the motif mi had been annotated). The distributions of P-values for all pairs of motifs annotated to the same alleles (i.e., for each allele h, i = 1, …Nh, j = 1…Nh, with i≠j and Nh the number of motifs annotated to allele h) are shown in Fig G in S1 Supporting Information for both the Euclidean distance and the BLiC score [33], and indicate that for more than 99% of the pairs of motifs annotated to the same alleles, the similarity has a P-value smaller than 0.05. The few pairs with P-values larger than 0.05 come mainly from HLA-C allele, as expected since they are less expressed and more degenerate, and therefore more challenging to describe.
We further studied the impact of the threshold T and explored several values between 0.01 and 0.14. As shown in Fig H in S1 Supporting Information, smaller values of T result in fewer alleles to which motifs can be annotated, as expected. Reversely, for larger thresholds the algorithm seems to behave badly and many more motifs annotated to the same alleles are no longer statistically similar. Data in Fig H in S1 Supporting Information suggest that reasonable results can be obtained for thresholds between 0.05 and 0.09, which is compatible with our manual choice of 0.078.
In practice, HLA-A and HLA-B alleles tend to be more expressed and therefore give rise to a stronger signal in HLA peptidomics data. We therefore used first our deconvolution method [24], setting the number of motifs equal to the number of HLA-A and HLA-B alleles and identified HLA-A and HLA-B motifs with the algorithm introduced above (Step 1). We then ran our deconvolution method [24] without restricting the number of clusters (Step 2) and identified motifs corresponding to HLA-A and HLA-B alleles based on the similarity with those identified in Step 1. The remaining motifs were then analysed across all samples with the algorithm introduced above.
To expand the identification of binding motifs for alleles without known ligands, we used data from IEDB for HLA-I alleles with well-described binding motifs. In practice, for all HLA-I alleles in our samples that had not been mapped to motifs in the fully unsupervised approach and have more than twenty different ligands in IEDB, PWMs were built from IEDB data. These PWMs were used to scan the remaining motifs in each sample that contained the corresponding alleles. Motifs were mapped to HLA-I alleles if exactly one PWM obtained with the mixture model was found to be similar to the IEDB-derived motif (i.e., Euclidean distance smaller than T, as before). The unsupervised procedure described above was then applied to the remaining motifs to identify new motifs for alleles without ligands in IEDB.
To have reliable estimates of the potential technical biases due to MS, amino acid frequencies were computed at non-anchor positions (P4 to P7) for alleles in our HLA peptidomics datasets (9-mers). Alleles showing some specificity at these positions (A02:01, A02:05, A02:06, A02:20, A25:01, A26:01, A29:02, B08:01, B14:01, B14:02, C03:03, and C07:04, see Fig D in S1 Supporting Information) were excluded from this analysis. The average frequencies of amino acids across alleles were then compared against the human proteome using Pearson correlation coefficient (Fig 3). We also performed the same analysis with HLA-I ligands (9-mers) from IEDB splitting between those obtained by MS and those obtained by other assays (“non-MS data”) (see Fig J in S1 Supporting Information). To enable meaningful comparison between these datasets, only alleles present in our HLA peptidomics data, with more than 100 ligands in both IEDB MS and non-MS data were considered in this analysis (14 alleles in total, see Fig D in S1 Supporting Information).
For each HLA-I allele, PWMs were built from all peptides associated to this allele across all samples where the binding motif could be identified, using the highest responsibility values of the mixture model [24]. The frequency of each amino acid was first computed. Pseudocounts were added using the approach described in [54], based on the BLOSUM62 substitution matrix with parameter β = 200. The score of a given peptide (X1, …XN) was computed by summing the logarithm of the corresponding PWM entries, including renormalization by expected amino acid frequencies: S=1N∑i=1Nlog(pXi,iqXi). Here qA stands for frequency of amino acid A at non-anchor positions (Fig 3 and Table B in S1 Supporting Information), pA, i stands for the PWM entry corresponding to amino acid A at position i, and N stands for the length of the peptide (N = 9, 10). The final score of a peptide was taken as the maximal score across all alleles present in a given sample and a P-value estimate is computed by comparing with distribution of scores obtained from 100,000 randomly selected peptides from the human proteome, so as to have similar amino acid frequencies compared to endogenous ligands (see MixMHCpred1.0 results).
To test our ability to predict neo-antigens, we used four melanoma samples in which ten neo-antigens (9- and 10-mers) have been directly identified with in-depth immunopeptidomics analyses of the tumor samples: Mel5, Mel8, Mel15 from [17] and 12T [20]. Missense mutations (i.e. cancer specific non-synonymous point mutations) identified by exome sequencing in those four melanoma samples [17] were retrieved and a list of all possible 9- and 10-mer peptides encompassing each mutation was built (S3 Dataset). Multiple transcripts corresponding to the same genes were merged so that each mutated peptide appears only once in the list. The total number of potential neo-antigens in each sample is shown in Table 1. Predictions for each HLA-I allele of each sample were carried out with the model described above. Peptides were ranked based on the highest score over the different alleles present in their sample. In parallel, affinity predictions with NetMHC (v4.0) [8] and NetMHCpan (v3.0) [12] and stability predictions with NetMHCstabpan (v1.0) [36] were performed for the same peptides and peptides were ranked based on predicted affinity using the highest score (i.e., lowest Kd) over all alleles. Only HLA-A and HLA-B alleles were considered since HLA-C alleles are known to show much lower expression and both our predictor and NetMHC could not be run for some HLA-C alleles in these melanoma patients. Ranking of the neo-antigens compared to all possible peptides containing a missense mutation is shown in Table 1 with either our predictor (MixMHCpred) or the other tools mentioned above. Area Under the Curves were also computed (Fig 4B), as well as the fraction of neo-antigens that fell among the top 1% of the predictions (PP1%, which typically corresponds to what can be experimentally tested) (Fig 4C), to provide a graphical visualization of the results in Table 1.
The same analysis was applied to study neo-antigens (9- and 10-mers) recently identified in two lung cancer patients [37] (L011: FAFQEYDSF, GTSAPRKKK, SVTNEFCLK, RSMRTVYGLF, GPEELGLPM and L013: YSNYYCGLRY, ALQSRLQAL, KVCCCQILL) and the ranking of these epitopes with different HLA-ligand predictors with respect to the full list of potential epitopes is shown in Table C in S1 Supporting Information (see also Fig 4B and 4C).
To assess how much our improved predictions of neo-antigens in the three melanoma samples of [17] depend on HLA peptidomics data generated from these samples, we performed a careful cross-sample validation. For each of the three samples where neo-eptiopes had been identified (Mel15, Mel8, Mel5) [17], we re-run our entire pipeline (i.e., annotation of HLA-I motifs across HLA peptidomics datasets + construction of PWMs for each allele) without the HLA peptidomics data coming from this sample. The PWMs were then used to rank all possible peptides (9- and 10-mers) encompassing each mutation. Overall, the predictions changed very little (Fig K in S1 Supporting Information).
All cancer testis antigens with annotated epitopes (9- or 10-mers) and HLA restriction information were retrieved from the CTDatabase [38] for all alleles considered in our predictor (see Table C in S1 Supporting Information). All other possible peptides along these proteins (9- or 10-mers) were used as negatives when benchmarking the predictions on this dataset.
HLA-I sequences were retrieved from IMGT database [3]. All protein structures analysed in this work were downloaded from the PDB. Residues forming the P2 binding site in HLA-B14:02 (PDB: 3BVN [42]) were determined using a standard cut-off of 5Å from any heavy atoms of arginine at P2.
W97R mutation was introduced into HLA-B14:02 wt by overlap extension PCR and confirmed by DNA sequencing. BL21(DE3)pLys bacterial cells were used to produce HLA-B14:02 wt and W97R as inclusion bodies. Four peptides with histidine or arginine at P2 (A[H/R]TKPRPAL, G[H/R]YDRSKSL, A[H/R]FAKSISL, H[H/R]FEKAVTL) were synthesized at the Peptide Facility (UNIL, Lausanne) with free N and C-termini (1mg of each peptide, > 80% purity). Peptides with histidine at P2 come from our HLA peptidomics data and were assigned to HLA-B15:18 by our mixture model algorithm. Based on our analysis of HLA sequence and structure, peptides with histidine at P2 are predicted to interact with HLA-B14:02 W97R, while peptides with arginine at P2 are predicted to bind better HLA-B14:02 wt.
Synthetic peptides were incubated separately with denaturated HLA-B14:01 wt and HLA-B14:01 W97R mutant refolded by dilution in the presence of biotinylated beta-2 microglobulin proteins at temperature T = 4°C for 48 hours. The solution was then incubated at 37°C and samples were retrieved at time t = 0h, 8h, 24h, 48h and t = 72h. The known HLA-B14:02 ligand IRHENRMVL was used for positive controls. Negative controls consist of absence of peptides. koff were determined by fitting exponential curves to the light intensity values obtained by ELISA at different time points. Half-lives were computed as ln(2)/koff. Values shown in Fig 5D correspond to the average over two replicates. For two peptides showing exceptionally high binding stability, only lower bounds on half-lives could be determined (dashed lines in Fig 5D).
|
10.1371/journal.pcbi.1006604 | Learning to synchronize: How biological agents can couple neural task modules for dealing with the stability-plasticity dilemma | We provide a novel computational framework on how biological and artificial agents can learn to flexibly couple and decouple neural task modules for cognitive processing. In this way, they can address the stability-plasticity dilemma. For this purpose, we combine two prominent computational neuroscience principles, namely Binding by Synchrony and Reinforcement Learning. The model learns to synchronize task-relevant modules, while also learning to desynchronize currently task-irrelevant modules. As a result, old (but currently task-irrelevant) information is protected from overwriting (stability) while new information can be learned quickly in currently task-relevant modules (plasticity). We combine learning to synchronize with task modules that learn via one of several classical learning algorithms (Rescorla-Wagner, backpropagation, Boltzmann machines). The resulting combined model is tested on a reversal learning paradigm where it must learn to switch between three different task rules. We demonstrate that our combined model has significant computational advantages over the original network without synchrony, in terms of both stability and plasticity. Importantly, the resulting models’ processing dynamics are also consistent with empirical data and provide empirically testable hypotheses for future MEG/EEG studies.
| Artificial and biological agents alike face a critical trade-off between being sufficiently adaptive to acquiring novel information (plasticity) and retaining older information (stability); this is known as the stability-plasticity dilemma. Previous work on this dilemma has focused either on computationally efficient solutions for artificial agents or on biologically plausible frameworks for biological agents. What is lacking is a solution that is both computationally efficient and empirically testable on biological agents. Therefore, the current work proposes a computational framework on the stability-plasticity dilemma that provides empirically testable hypotheses on both neural and behavioral levels. In this framework, neural task modules can be flexibly coupled and decoupled depending on the task at hand. Testing this framework will allow us to gain more insight in how biological agents deal with the stability-plasticity dilemma.
| Humans and other primates are remarkably flexible in adapting to constantly changing environments. Additionally, they excel at integrating information in the long run to detect regularities in the environment and generalize across contexts. In contrast, artificial neural networks (ANN), despite being used as models of the primate brain, experience significant problems in these respects. In ANNs, extracting regularities requires slow, distributed learning, which does not allow strong flexibility. Moreover, fast sequential learning of different tasks typically leads to (catastrophic) forgetting of previous information (for an overview see [1]). Thus, ANNs are typically unable to find a trade-off between being sufficiently adaptive to novel information (plasticity) and retaining older information (stability), a problem known as the stability-plasticity dilemma.
In recent years, a wide variety of solutions have been provided for this stability-plasticity dilemma. These solutions can broadly be divided in two classes. The first class is based on the fact that catastrophic forgetting does not occur when tasks are intermixed. Thus, one solution is to keep on mixing old and new information [2–5]. [3] suggested that new information is temporarily retained in hippocampus. During sleep (and other offline periods), this information is gradually intermixed with old information stored in cortex. This framework inspired subsequent computational and empirical work on cortical-hippocampal interactions [6–8].
The second class of solutions is based on the protection of old information from being overwritten. Protection can occur, first, at the level of synapses. For example, [9] combined a slow and fast learning system, with slow and fast weights reflecting long- and short-time-scale contingencies, respectively. This allows the network to both extract stable regularities (slow learning system) and flexibly adapt to fast changes in the environment (fast learning system). Another recent idea is to let synapses (meta-)learn their own importance for a certain task [10], [11]. Weights that are very important for some task are not allowed to (and thus protected from) change. Hence, information encoded in those weights is preserved. Second, protection can also be implemented at the level of (neural) activation. The most straightforward approach to implement such protection is to orthogonalize input patterns for the relevant tasks [12], [13]. Another approach to achieve protection at the level of neural activation, is gating. This means that only a selected number of network nodes can be activated. Because weight change depends on co-activation of relevant neurons [14], [15], this approach protects the weights from changing. For example, [16] proposes that in each of several tasks a (randomly selected) 80% of nodes is gated out, thus effectively orthogonalizing different contexts. They showed that synaptic gating allowed a multi-layer network to deal with several computationally demanding tasks without catastrophic forgetting.
Crucially, it remains unknown how biological agents deal with this dilemma. The current study aims to provide a novel computational framework focused on biological agents that makes empirically testable predictions at MEG/EEG level. For this purpose, we combine two prominent principles of computational neuroscience, namely Binding by Synchrony [17–20]) and Reinforcement Learning (RL; [21], [22]). In BBS, neurons are flexibly bound together by synchronizing them via oscillations. This implements selective gating (e.g., [23]) in which synchronization enhances the communication between neuronal groups (gates are opened) and desynchronization disrupts the communication between neural groups (gates are closed). In sum, BBS allows the model to flexibly alter communication efficiency on a fast time scale. By using RL principles, the model can learn autonomously when neurons need to be (de)synchronized.
In the modeling framework, BBS binds relevant neural groups, called (neural task) modules, and unbinds irrelevant modules. This causes both efficient processing and learning between synchronized modules; and inefficient processing and absence of learning between desynchronized modules. The resulting model deals with the stability-plasticity dilemma by flexibly switching between task-relevant modules and by retaining information in task-irrelevant modules. An RL unit [24] uses reward prediction errors to evaluate whether the model is synchronizing the correct task modules.
In order to test the generalizability of our framework, we apply it to networks containing modules that learn via three classic synaptic learning algorithms, namely Rescorla-Wagner (RW; [15], [25]), backpropagation (BP; [26]) and Restricted Boltzmann machines (RBM; [27]). The RW algorithm [25] is one of the most well-known and basic supervised-learning algorithms in cognitive neuroscience. Here, on each trial, an error term is computed based on the discrepancy between a model-generated output pattern and some target output pattern. Learning consists of using this error term for finding a weight configuration that minimizes the average error across trials. This algorithm is typically fast and efficient for learning simple (i.e., linearly separable) input-output associations. Hence, it has no problems with plasticity. However, while learning one set of input-output associations (set B), the algorithm may unlearn another, currently irrelevant set of input-output associations (set A). Thus, when set A becomes relevant again, it will have to relearn it. In sum, the RW algorithm does suffer from a lack of stability, but due to its high plasticity it might have only minor problems with respect to the stability-plasticity dilemma, especially when the learning rate is high. In this case, it might relearn the forgotten information (set A) so fast that also the stability problem is negligible. Nevertheless, the RW algorithm suffers from some severe limitations on the complexity of problems that it can solve. It is very efficient in dealing with linearly separable input-output associations, but cannot deal with more complex, not linearly separable, problems.
This limitation of the RW algorithm is solved in BP [26]. Similar to RW, learning with BP consists of using the error term for finding a weight configuration that minimizes the average error across trials. Relative to RW, this algorithm is able to solve a much wider range of problems. In particular, it can also solve nonlinearly separable problems. It does this by adding hidden layers between input and output. For training the weights toward the hidden layers, BP propagates the error term backwards from output toward the hidden (i.e., deeper) layers in the network. Crucially, sequential learning of input-output associations poses severe computational problems on the BP algorithm. Because the number of (interdependent) weights that should be adjusted to solve a problem is much higher, the algorithm learns much slower. Hence, if the learning rate is low, new learning can be very slow, causing a lack of plasticity. If the learning rate is very fast, on the other hand, this problem is mitigated but there is no stability in the model. This is because, similar to RW, the learning algorithm will adapt all available weights and therefore overwrite previous information.
An algorithm that can also learn with hidden layers (and thus solve more complex problems) is RBM. Despite the algorithmic differences, RBM suffers from the same stability-plasticity dilemma as BP. To further illustrate the generality of our framework, S1 Text show that our framework can also be applied to networks with modules that learn via RBM. For brevity, the main text restricts attention to RW and BP.
The full model consists of three units (Fig 1A). The Processing unit contains a network consisting of a number of task-specific modules; the two learning algorithms (RW or BP) are implemented between modules of the Processing unit. In addition, RL and Control units together form an hierarchically higher network modeled after basal ganglia/primate prefrontal cortex [28]. The RL unit (modeling ventral striatum/ anterior medial frontal cortex (aMFC)) evaluates behavior. More specifically, it learns to assign a value to a specific task module (how much reward it receives by using this module) and compares this value with the externally received reward to compute prediction errors. Additionally, the RL unit has a Switch neuron (see Fig 2A). This Switch neuron computes a weighted sum of negative prediction errors across trials. When this sum reaches a threshold of .5, it signals the need for a strategy switch to the Control unit (see Methods for details). This Control unit drives neural synchronization in the Processing unit. One part of the Control unit (modeling lateral frontal cortex (LFC)) contains task neurons that point to task modules in the Processing unit [29]; another part (modeling posterior medial frontal cortex (pMFC)) synchronizes task modules based on those task neurons [30]. Crucially, LFC and pMFC both use prediction error information, but on different time scales. The pMFC uses prediction errors on a fast time scale to enhance control over the synchronization process as soon as a negative prediction error occurs. In contrast, the LFC uses prediction errors on a slow time scale to know when the task rule has changed and a switch of modules is needed.
In order to drive neural synchronization between task modules in the Processing unit, we rely on the idea of binding by random bursts [30–32]. Here, applying positively correlated noise to two oscillating signals reduces their phase difference. In addition to implementing binding by random bursts, the current work also implements unbinding by random bursts. In particular, applying negatively correlated bursts increases the phase difference between oscillating signals and thus unbinds (i.e., dephases) the two signals.
We test our model on a (cognitive control) reversal learning task. Here, each hierarchically lower algorithm (RW or BP; in the Processing unit) sequentially learns different task rules. The relevant task rule changes across task blocks (Fig 1B). The model must detect when task rules have changed, and flexibly switch between different rules without forgetting what has been learned before. We divide the task in six equally long task blocks. In the first three blocks, the model should learn three different new task rules (rule A, B and C in blocks 1, 2 and 3 respectively). In the second half, the model has to switch back to the previously learned rules (rule A, B and C in blocks 4, 5 and 6 respectively; see also Fig 1B).
For the RW network, we use a one-dimensional task. Here, on each trial one out of three stimulus features is activated. For every task rule we link a stimulus feature to a response option. More specifically, in task rule A, feature 1 (F1) is associated to response 1 (R1), feature 2 (F2) to response 2 (R2) and feature 3 (F3) response 3 (R3). In task rule B, F1 is associated to R2, F2 to R3 and F3 to R1. Task rule C associates F1 to R3, F2 to R1 and F3 to R2. For the BP network, a multi-dimensional task is used. Here, on each trial multiple stimulus features are activated. More specifically, the task utilizes four dimensions. Every dimension has three features. One of the dimensions represents a cue that indicates which out of the other three (stimulus) dimensions is relevant on the current trial. In line with the one-dimensional task, the 3 stimulus features of each dimension are within each task rule linked to one response option.
The one-dimensional task (for RW) consists of 360 trials; the multi-dimensional task (for BP) consisted of 3600 trials. For comparison, we divided each task sequence in 120 trial bins for analysis and plotting. Fig 2A illustrates the detailed model for both tasks. We compare our combined (henceforth, full) models with models that use no synchronization (i.e., only contain the Processing unit; called no-synchrony models). We evaluate plasticity as the ability to learn a new task; and stability as the interference from learning a new task toward performance on the old task (see Fig 1B and Methods).
In Fig 4A, the accuracy evolution across all task blocks is plotted for both the full and no-synchrony RW model with a slow learning rate, β = .2, for the simple (linearly separable) task. The full model is marginally slower in learning new task rules. However, when the model needs to switch back to a previously learned rule (task blocks 4–6) we observe a minor advantage for the full model in the first trials, since it does not have to relearn the task.
A very different picture emerges for the complex (nonlinearly separable) task. Fig 4D shows the accuracy of the full and no-synchrony BP model. During the first task block, the no-synchrony and full model perform similarly. When the task rule switches for the first time (i.e., after the first task block), the drop in accuracy is slightly larger for the no-synchrony model than for the full model. This is caused by the fact that the no-synchrony model has to learn task rule B with weights that were pushed in a direction opposite to those that are optimal for task rule A. Instead, the full model switches to another task module and starts learning from a random weight space. A similar phenomenon occurs after the second rule switch.
For the following task switches, the model has to switch back to rules it already learned before; it is here that the full potential of the full model emerges. The full model can switch back to a previous module, where all old information was retained. Instead, the no-synchrony model has catastrophically forgotten the first task rule and must hence relearn it.
Fig 6 shows the overall accuracy, stability and plasticity of our full model and of the no-synchrony model for the two task structures discussed in the previous section (1- and 3-dimensional tasks). In order to gain more insight in how the model performance is affected by task complexity, we also show overall results for the BP model on a 2-dimensional task. Thus, we show results for tasks of increasing complexity, namely for 1 dimension (RW model), 2 dimensions (BP model) and 3 dimensions (BP model). Results of the RBM model are discussed in S1 Text.
As a model of how the brain controls its own processing, we next aimed at describing the relation between our model and previous empirical data, and provide testable hypotheses for future empirical work.
We described a computationally efficient and empirically testable framework on how biological and artificial agents may deal with the stability-plasticity dilemma. We combined two neurocomputational frameworks, BBS [17–19] and RL [21]. BBS flexibly (un)binds (ir)relevant neural modules; RL autonomously discovers when modules need to be (un)bound. Thus, the model could flexibly switch between different tasks (plasticity) without catastrophically forgetting older information (stability). We demonstrated that the model was consistent with several behavioral and electrophysiological (e.g., MEG/EEG) data. In the remainder, we first summarize the main model components, and point to plausible neural origins of each. Second, we discuss specific empirical predictions that are made by the model. Third, we discuss limitations and possible extensions. As a fourth and last point, we describe how the current work relates to previous computational modelling work.
Plausible neural origins for all three model units are summarized in Fig 8. The Processing unit contains a task-processing network, trained by a classical learning rule (RW, BP, or RBM). Anatomically, its nodes can be localized in several posterior (neo-)cortical processing areas, depending on the task at hand (e.g., fusiform face area in a face-processing task). Its activity is strongly stimulus-dependent and synaptic strengths change slowly. The RL unit learns to attach value to specific task modules, based on prediction errors. Previous work with fMRI [24], [41] already used a probabilistic reversal learning paradigm to localize the brain areas involved in such value learning. This work localized the RL unit in MFC, which (with brainstem and striatum) is generally considered as an RL circuit [24], [42], [43]. Importantly, computations in this unit are not used for driving task-related actions, but for driving hierarchically-higher actions, namely to (de)synchronize task modules. This is in line with recent considerations of MFC as a meta-learner [44–47]. We tentatively call this unit aMFC, given this region’s prominent anatomical connectivity to autonomous regions [48]. There was also a Switch neuron in our model. Previous work on stay/switch decisions has proposed they originate from frontopolar cortex [49]. Hence, processes in the RL unit might be best explained by a neural circuit between brainstem, aMFC and frontopolar cortex.
The Control unit was adopted from [30]. Its first part contains units that point to specific posterior processing areas, indicating which neurons should be (un)bound. Thus, this area stores the task demands. We labeled this part LFC, given the prominent role of LFC in this regard [50], [51]. The second part of the Control unit sends random bursts to posterior processing areas to synchronize currently active areas. Given the prominent anatomical connectivity of pMFC to motor control and several posterior processing areas [48], we tentatively label this part pMFC. The efficiency of this controlling process is largely determined by pMFC theta power: More power leads to more and longer bursts [30]. This is consistent with empirical work linking high MFC theta power to efficient cognitive control [34], [35]. Power in the model pMFC is itself modulated by the occurrence of negative prediction errors. More specifically, when a negative prediction error occurs, the pMFC node will receive bursts, which will increase pMFC theta power. In absence of negative prediction errors, this theta power will slowly decrease across trials. This is consistent with the idea that a constant high MFC power might be computationally suboptimal and empirically implausible. For instance, MFC projects to locus coeruleus (LC;[52]); LC firing is thought to be cognitively costly, perhaps because it leads to waste product in the cortex that needs to be disposed [36]. In sum, in the Control unit, LFC and pMFC jointly align neural synchronization in modules of the Processing unit to meet current task demands [53], [54]. The LFC indicates which modules should be (de)synchronized, and the pMFC exerts control over the oscillations in the Processing unit by (de)synchronizing them via random bursts.
Crucially, both parts of the Control unit use prediction errors, but at a different time scale. More specifically, the pMFC uses an evaluation of the last prediction error to evaluate the amount of control that should be exerted (fast time scale). Hence, when an error occurs, the model will initially exert more control on the currently used task module/strategy. The LFC on the other hand, is guided by processes in the Switch neuron of the RL unit which evaluates prediction errors on a slow time scale by integrating them over multiple trials, in order to decide between staying with the current task module or switching to another. Therefore, if negative prediction errors keep on occurring after the model increased control, it will switch modules/strategies.
Importantly, our model makes several predictions for empirical data. First, it predicts significant changes in the phase coupling between different posterior neo-cortical brain areas after a task switch. Here, we suggest that desynchronization may be important to disengage from the current task. Consistently, [55] found that strong desynchronization marked the period from the moment of disambiguation of ambiguous stimuli to motor responses. Additionally, Parkinson disease patients, often characterized by extreme cognitive rigidity, show abnormally synchronized oscillatory activity [56]. Thus, we suggest that neural synchronization between task-relevant brain areas is crucial for implementing task rules. Additionally, desynchronization is necessary for disengaging from a task.
Second, we explored midfrontal theta-activation in the time-frequency domain by wavelet convolution. These analyses showed an increase of theta power after an error. This was caused by bursts that were sent from the RL unit as described in Eq (8). Hence, the model predicts an increase of theta amplitude in the MFC after negative prediction errors in tasks where these prediction errors signal the need for increased cognitive control [34], [35], [37].
Third, we connected the model to research demonstrating theta/gamma interactions where faster gamma frequencies, which implement bottom-up processes, are typically embedded in, and modulated by, slower theta-oscillations, in order to implement top-down processes [40], [57–59]. For this purpose, we considered coupling between pMFC theta phase and gamma amplitude in the Processing unit. Our model predicts a strong PAC increase in the first trial(s) after a task switch, which decays slowly after the switch. This reflects the binding by random bursts control process which is increased after task switches, and decays once a task rule is sufficiently implemented. Hence, the model predicts a strong coupling between frontal theta phase and posterior gamma amplitude when new task rules need to be implemented.
The model contained several limitations, and consequently also possibilities for future extensions. First, the RL unit currently learns to assign a value to some task module. It can determine when a task switch occurred, and then make a binary switch assessment; to switch or not to switch to another task module. Thus, when the model realizes that the current task module/strategy is incompatible with the current task/environment, it has to change its behavior. It will attempt random strategies until an appropriate one is found. Learning when to switch can be considered as a type of meta-learning. However, the full model would benefit significantly from more advanced meta-learning mechanisms. Future work will address this issue by adding second level (contextual) features which allow the LFC to (learn to) infer which of multiple task modules should be synchronized. One useful application of such second level features would be task set clustering, which allows to generalize quickly over multiple contexts. Specifically, if a novel second-level feature becomes connected to an earlier learned task set (in LFC), all the task-specific mappings of this task set would immediately generalize to the novel second-level feature. This is consistent with immediate generalization seen in humans [60–62].
Second, several parameters of the model were fixed, but might more generally be controllable (learnable) as well. For example, the time scale of the Switch neuron is controllable by the σ parameter in Eq (12). In a very stable environment, a low σ is adaptive, which slows down the time scale, decreasing the weight of more recent prediction errors. Instead, if the environment is unstable, a less conservative strategy is in order (high σ), in which case the model accumulates evidence across less trials in order to make a switch decision. Earlier models already described how switching between hypotheses could depend on environmental stability and noise [63]; such manipulation (here, of parameter σ) might be usefully implemented in future developments of the current model too.
Third, although using negative prediction errors to modulate the control amplitude of the pMFC is efficient in the current context, this might not be ideal in more complex environments. Thus, another future challenge is broadening the control signal (i.e., beyond negative prediction errors) that the model uses to optimally adapt to the environment’s reward and cost structure [45].
Fourth, the node architecture of neuronal triplets is an oversimplification of how oscillations are produced in the human brain. Several neural models propose that interacting excitatory (E) cells and inhibitory (I) cells generate oscillations [33], [64]. These oscillatory neurons are grouped with stimulus-driven neurons in cortical columns; oscillatory neurons modulate the activation of the stimulus-driven neurons [65]. In the current model, these assumptions are implemented in the simplest way, namely where each column consists of just three neurons (E, I, and x), and the oscillatory activity modulates the stimulus-driven activity. Furthermore, our implementation of processing within a neuronal triplet is perhaps biologically implausible, in the sense that the neuron that processes stimuli (x) is distinct from the neurons that generate the oscillations (E, I) which do not process any stimulus information. Future work will determine whether the current approach can be scaled to more biologically plausible architectures.
Fifth, the model ignored some aspects of oscillatory dynamics. For instance, our model only implements neural synchronization between Processing unit neurons with the same (gamma-band) frequencies. This scenario might be unrealistic in a typically noisy human brain. However, the problem of noise can be efficiently solved by employing rhythmic bursts, such as the theta-frequency we implemented here. Specifically, one-shot synchronizing bursts would cause oscillations with (slightly) different (gamma-band) frequencies to gradually drift apart after the burst. With rhythmically paced bursts, the gamma oscillations have no time to drift apart since the next burst occurs before the drift becomes substantial. In line with this idea, previous work has demonstrated how the model can deal with gamma frequency differences of at least around 2% [30]. Moreover, one might wonder if it would be optimal to send bursts at a frequency much faster than theta, thus providing no opportunity for noisy oscillations to drift apart. However, the current work showed that accuracy of the model dramatically declines if the pMFC sends bursts at a faster frequency than theta. The reason is that bursts given by the pMFC to the Processing unit introduce noise to the system. This can be clearly observed in Fig 3C, in which there are short periods of irrelevant neuronal activation during the bursts. Hence, an optimal agent would want to limit the bursts as much as possible. Since these bursts are phase locked to the pMFC oscillation and rely on its amplitude, the model performs best with slower pMFC frequencies that are rapidly attracted (high Damp) towards a small amplitude (low rmin). Again, the oscillations in the Processing unit of the current model all have the exact same frequency. When Processing unit activations do not have the same frequency, we thus conjecture that there is an optimal, intermediate (theta) bursting frequency, depending on the Processing unit (gamma) frequency. Future work should explore such an optimal balance between a Controller/ bursting (theta) frequency and a Processing (gamma) frequency in more noisy systems. Another aspect of oscillatory dynamics we ignored is that BBS may be more biologically plausible, and more efficient, with small inter-areal delays [66]. Future work will consider an additional (meta-) learning mechanism that learns to synchronize nodes with an optimal phase delay between task modules.
The current work relies heavily on previous modeling work of cognitive control processes. For instance, in the current model the LFC functions as a holder of task sets which bias lower-level processing pathways [29], [67]. It does this in cooperation with the MFC. Here, the aMFC determines when to switch between lower-level task modules. Additionally, also the amount of control/ effort that is exerted in the model is determined by the RL processes in the aMFC[44–46]. More specifically, negative prediction errors will determine the amount of control that is needed by strongly increasing the pMFC signal [42]. This is consistent with earlier work proposing a key role of MFC in effort allocation [44], [45], [68].
In the current model, the MFC, together with the LFC, functions as a hierarchically higher network that uses RL to estimate its own task-solving proficiency. Based on its estimate of the value of a module, and the reward that accumulates across trials, it evaluates whether the current task strategy is suited for the current environment. Based on this evaluation, it will decide to stay with the current strategy or switch to another. More specifically, the value learned by the RL unit acts as measure of confidence that the model has in its own accuracy. The model uses this measure of confidence to adjust future behavior, a process that has been labeled as meta-cognition [69], [70].This is in line with previous modeling work that described the prefrontal cortex as a reinforcement meta-learner [43], [46–48].
One problem we addressed in this work was the stability-plasticity dilemma. As we described before, previous work on this dilemma can broadly be divided in two classes of solutions. The first class is based on mixing old and new information [2–5]. The second class is based on protection of old information. Our solution also exploited the principle of protection. Future work must develop biologically plausible implementations of the mixing principle too, and investigate to what extent mixing and protection scale up to larger problems.
We provided a computationally efficient and empirically testable framework on how the primate brain can address the tradeoff between being sufficiently adaptive to novel information, while retaining valuable earlier regularities (stability-plasticity dilemma). We demonstrated how this problem can be solved by adding fast BBS and RL on top of a classic slow synaptic learning network. RL is used to synchronize task-relevant and desynchronize task-irrelevant modules. This allows high plasticity in task-relevant modules while retaining stability in task-irrelevant modules. Furthermore, we connected the model with empirical findings and provided predictions for future empirical work.
As mentioned before and is shown in Fig 1A, our model consists of three units. First, the Processing unit includes the task-related neural network, which is trained with a classical learning rule (RW, BP or RBM). On top of this classical network, an extra hierarchical layer is added consisting of two units [28]. The RL unit, adopted from the RVPM [24], evaluates whether the Processing unit is synchronizing the correct task modules. This evaluation is used by the Control unit [30] to drive neural synchronization in the Processing unit. Thus, this hierarchically higher network allows the models to implement BBS in an unsupervised manner.
We test our model on a reversal learning task [71], [72]. We divide the task in six equally long task blocks. In the first three blocks, the model should learn three different new task rules (rule A, B and C in blocks 1, 2 and 3 respectively). In the second half, the model has to switch back to the previously learned rules (rule A, B and C in blocks 4, 5 and 6 respectively).
For the RW network, we use a one-dimensional task. This task consisted of 360 trials. Here, on each trial one out of three stimulus features is activated. For every task rule we link a stimulus-feature to a response option. More specifically, in task rule A, feature 1 (F1) is associated to response (R1), feature 2 (F2) to response 2 (R2) and feature 3 (F3) response 3 (R3). In task rule B, F1 is associated to R2, F2 to R3 and F3 to R1. Task rule C associates F1 to R3, F2 to R1 and F3 to R2. All stimuli are presented equally often in random order.
For the BP and RBM networks, a multi-dimensional task is used consisting of 3600 trials. In order to gain insight in how the complexity of the task affects our model, we implemented a task with two stimulus dimensions (two-dimensional task) and one with three stimulus dimensions (three-dimensional task). For the RBM model, we only implemented the three-dimensional task. Every stimulus dimension has three features. In total, a task consists of N + 1 dimensions, in which N is the number of stimulus dimensions and the extra dimension is a cue dimension (with N features), indicating which of the N stimulus dimensions is relevant on the current trial. On each trial one feature of every dimension is activated. In line with the one-dimensional task, the N = 3 task features of the stimulus dimensions are, within each task rule, linked to one response option. Again, in each block, each possible stimulus is presented equally often, in a random order.
To test the generality of our findings, we varied the synaptic learning rate. This parameter was varied from 0 to 1 in 11 steps of .1. For each value, we performed 10 replications of the simulation. In every simulation, the strength of synaptic connections at trial 1 was a random number drawn from the uniform distribution, multiplied by half the bias value (and 1 for the RW based model).
The effects of other model parameters were already demonstrated in previous work [24], [30], but we again validated that the model shows qualitatively similar patterns when we varied some of the parameters. A table of all parameter values used in both the original simulations and parameter explorations is provided in the S1 Text. Specifically, we explored different frequencies (C in Eqs (1) and (2)) in the Processing unit and the pMFC. Additionally, we also explored the Damp and rmin parameters in the pMFC (again Eqs (1) and (2)). For this simulation we used the RW model with β = .2. We fully crossed all parameter values for C, Damp and rmin. and performed 5 replications. In a separate set of explorations, we varied σ and α in the RL unit (see Eqs (11) and (12)) for both the RW and BP algorithm, for both a slow and a fast-synaptic learning rate (β). Again, we performed 5 replications for each parameter combination. Results of the latter parameter exploration are described in the S1 Text.
For the purpose of comparison, we divided the trials of the task for every model into 120 bins. For the RW model, bin size equals 3 trials; for the BP and RBM models, bin size equals 30 trials. We evaluate the performance of our model on several levels. First, we evaluate overall task accuracy. Second, we evaluate plasticity. For this purpose, we explore the performance of the model during the first 5 bins of the first 3 blocks. Hence, we test how quickly a model can learn a new task rule. Third, we evaluate stability. In particular, we explore the interference of learning other task rules in between two periods of performing the same task rule. For this purpose, we compare the accuracy during the first 5 trial bins of block 4, 5 and 6 with the last 5 trial bins of block 1, 2 and 3. If the model saved what was learned, this difference should be zero. If the model displays catastrophic forgetting, it would have a negative stability score.
Importantly, we also connect with empirical data and describe testable hypotheses for future empirical work. As a measure of phase synchronization between excitatory neurons in the Processing unit, we compute the correlation at phase lag zero. A correlation of 1 indicates complete synchronization and -1 indicates complete desynchronization. Phase-amplitude coupling (PAC) is computed as the debiased phase-amplitude coupling measure (dPAC; [73]) in each trial. Here,
dPAC=|1h∑t=1hat×(eiφt−Φ−)|
(14)
in which
Φ−=1h∑t=1heiφt
(15)
In these equations, t represents one time step in a trial, h is the number of time steps in a trial, a is the amplitude, φ is the phase of a signal, and i2 = -1. In the current paper, we are interested in the coupling between the phase of the theta oscillation in the pMFC node of the Control unit and the gamma amplitude in the Processing unit. Phase was extracted by taking the analytical phase after a Hilbert transform. The gamma amplitude was derived as the mean of the excitatory phase code activation of all nodes in the Processing unit by
at=1I∑i=1I|Eit|
(16)
with I being the number of nodes in the Processing unit, t referring to time and Ei being the respective excitatory phase code neuron.
For all measures, we represent the mean value over Nrep = 10 replications and error bars or shades show the confidence interval computed by mean± 2×(SD/√Nrep).
Additionally, we evaluated the pMFC theta activation. More specifically, time–frequency signal decomposition was performed by convolving the signal of EpMFC by complex Morlet wavelets, ei2πfte−t2/(2σ2), where i2 = -1, t is time, f is frequency, ranging from 1 to 10 in 10 linearly spaced steps, and σ = 4/(2πf) is the “width” of the wavelet. Power at time step t was then computed as the squared magnitude of the complex signal at time t and frequency f. We averaged this power over all simulations and all replications of our simulations. This power was evaluated by taking the contrast between the inter-trial intervals following correct (1) and error (0) reward feedback.
Matlab codes that were used for both the model simulations and data analysis are available on GitHub (https://github.com/CogComNeuroSci/PieterV_public).
|
10.1371/journal.pntd.0004492 | Effect of Two or Six Doses 800 mg of Albendazole Every Two Months on Loa loa Microfilaraemia: A Double Blind, Randomized, Placebo-Controlled Trial | Loiasis is a parasitic infection endemic in the African rain forest caused by the filarial nematode Loa loa. Loiasis can be co-endemic with onchocerciasis and/or lymphatic filariasis. Ivermectin, the drug used in the control of these diseases, can induce serious adverse reactions in patients with high L loa microfilaraemia (LLM). A drug is needed which can lower LLM below the level that represents a risk so that ivermectin mass treatment to support onchocerciasis and lymphatic filariasis elimination can be implemented safely.
Sixty men and women from a loiasis endemic area in Cameroon were randomized after stratification by screening LLM (≤30000, 30001–50000, >50000) to three treatment arms: two doses albendazole followed by 4 doses matching placebo (n = 20), six doses albendazole (n = 20) albendazole or 6 doses matching placebo (n = 20) administered every two months. LLM was measured before each treatment and 14, 18, 21 and 24 months after the first treatment. Monitoring for adverse events occurred three and seven days as well as 2 months after each treatment.
None of the adverse events recorded were considered treatment related. The percentages of participants with ≥ 50% decrease in LLM from pre-treatment for ≥ 4 months were 53%, 17% and 11% in the 6-dose, 2-dose and placebo treatment arms, respectively. The difference between the 6-dose and the placebo arm was significant (p = 0.01). The percentages of participants with LLM < 8100 mf/ml for ≥4 months were 21%, 11% and 0% in the 6-dose, 2-dose and placebo treatment arms, respectively.
The 6-dose regimen reduced LLM significantly, but the reduction was insufficient to eliminate the risk of severe and/or serious adverse reactions during ivermectin mass drug administration in loiasis co-endemic areas.
| Loiasis is a big obstacle for the elimination of onchocerciasis and lymphatic filariasis in Central Africa in areas where loiasis is endemic. In these areas, some subjects who are heavily infected (microfilaraemia > 30 000 microfilariae/ml blood) can develop severe and serious adverse reactions to ivermectin. In rare cases, these have been fatal. A way of preventing these reactions could be to administer a treatment that decreases the microfilareamia in all subjects below the risk level before mass treatment with ivermectin. Building on results of previous studies, this randomised placebo-controlled trial evaluated the efficacy and safety of two or six doses of albendazole administered every two months on the microfilaraemia of Loa loa. Six doses led to a decrease in microfilaraemia by at least 50% for at least four months in 53% of participants. However, it did not reduce the microfilaraemia below the risk level in all participants. Therefore, this regimen has not sufficient efficacy to prevent severe adverse reactions to ivermectin.
| Loiasis is a parasitic infection endemic in the African equatorial rain forest areas, caused by the filarial nematode Loa loa. It is estimated that at least 14.4 million people live in loiasis endemic areas [1]. Clinical manifestations include chronic intense itching and transient localized edema [2]. The chronic eosinophilia observed in infected individuals has been associated with endomyocardial fibrosis and related heart failure [3, 4]. Spontaneous encephalitis has been described in some heavily infected patients [5]. As for other neglected tropical diseases, the limited geographic distribution of loiasis and the fact that it mainly affects poor rural populations have limited research on this disease [6].
There is currently no safe and effective treatment. Diethylcarbamazine is effective against the larvae and adult Loa loa [7], but can cause serious adverse reactions (SAR, for definition see Table 1 [8]), such as meningoencephalitis, which can be fatal [9, 10]. The rapid and massive effect of ivermectin on Loa loa blood-dwelling microfilariae can also lead to severe adverse drug reactions (ADRs, Table 1 [8]), including SARs such as a cerebral malaria-like encephalopathy which requires hospitalization, can lead to coma and is often fatal [11, 12]. The mechanisms of these adverse reactions are not well understood. Three mechanisms have been postulated, including (1) the obstruction of the cerebral microcirculation in consequence of massive amounts of paralyzed or dead microfilariae, (2) the penetration of live microfilariae into the brain tissue following treatment, and (3) the inflammatory processes in the brain resulting from massive release of antigen from dead microfilariae [13].
The current evidence suggests that the risk of ADRs post-ivermectin is positively correlated with Loa loa microfilaraemia (LLM). The relatively low number of participants with known pre-treatment LLM and adverse event data available from prospective studies [14, 15] together with the likely underreporting of SARs and non-serious ADRs and the lack of pre-treatment LLM values from mass treatment with ivermectin [16], do not permit a definition of the minimum pre- treatment LLM that puts subjects at risk for development of severe ADRs and/or SARs. The study by Gardon and colleagues showed that a pre-treatment LLM ≥8100mf/ml is associated with significantly increased risk of 'marked reactions' (defined by the investigators as reactions accompanied by functional impairment requiring assistance for several days) as well as 'serious reactions' (SR, including (a) non-neurological reactions associated with functional impairment which required at least a week of full-time assistance to resume normal activities and (b) reactions with objective neurological signs with hospital admission). It was estimated that participants with pre-treatment LLM ≥50000 mf/ml were 1000 times more likely to develop SR and that those with LLM >30000 were more than 200 times more likely to develop SR than non-infected participants [15].
Onchocerciasis and lymphatic filariasis (LF) are two of the 17 neglected tropical diseases according to the WHO classification (http://www.who.int/neglected_diseases/diseases/en/). The number of people at risk of O. volvulus infection was estimated to be 113.5 million, and those at risk of LF to be 1.34 billion worldwide [17]. The cornerstone of the fight against onchocerciasis and LF in highly endemic areas in Africa is mass community treatment with ivermectin (Mectizan) alone and in combination with albendazole, respectively.
In loiasis endemic areas where onchocerciasis is mesoendemic or hyperendemic (i.e. prevalence of onchocercal skin nodules in adult males aged ≥20 years higher than 20% and 40% respectively), ivermectin mass treatment is justifiable because the benefit of preventing onchocerciasis associated morbidity outweighs the risk of loiasis-related post-treatment adverse reactions [18].
In LF and loiasis co-endemic areas, the control of LF has not yet or only recently started, partly because of the risk of loiasis-related post-treatment SAEs [19]. This puts the planned elimination of LF at risk [17]. WHO has recently suggested a provisional strategy for interruption of LF transmission in loiasis-endemic areas based on biannual treatment with albendazole complemented with vector control [20, 21].
Studies in Senegal and Mali have shown that 15–17 years of annual or biannual community directed treatment with ivermectin (CDTI) can result in elimination of O. volvulus transmission [22, 23]. Furthermore, the prevalence of O. volvulus infection following long term CDTI has been reduced significantly in many other areas [24, 25]. Consequently, the objectives of the African Programme for Onchocerciasis Control were expanded from elimination of onchocerciasis as a public health problem to elimination of O. volvulus transmission where feasible [26]. This may require treatment of onchocerciasis in hypoendemic areas co-endemic with loiasis. A treatment that can safely reduce LLM below the risk threshold for severe ADRs and SARs for a time sufficiently long to implement ivermectin mass treatment, would be a major contribution to efforts to control and eliminate onchocerciasis and LF. Such a treatment may also be beneficial for patients suffering from loiasis.
Previous trials of the effect of short-term albendazole treatments (1 to 21 days) have shown that albendazole results in a slow reduction of LLM, presumably not due to a microfilaricidal effect but to an effect on the reproductive capacity and/or viability of the macrofilariae. In these studies, the reduction in LLM was either not as extensive as required and/or the study did not include significant number of participants with LLM > 30000 mf/ml [27–30]. The LLM time course in these studies and comparison of the effect of different doses evaluated led to the hypothesis that multiple exposure of the Loa loa macrofilariae to albendazole at two months intervals may result in a significant and sustained reduction in LLM.
We report findings from a double-blind, randomized, placebo-controlled trial designed to evaluate whether 2 and/or 6 doses of 800mg albendazole administered at two-months intervals can reduce Loa loa microfilaraemia in patients with pre-treatment LLM >15000 mf/ml by at least 50% or even to <8100mf/ml for at least 4 months.
http://www.controlled-trials.com/ISRCTN25831558.
This was a double-blind, randomized, placebo-controlled trial with three parallel treatment arms. Every two months (i.e. at M0, M2, M4, M6, M8, M10), participants received one oral treatment: (1) 800 mg albendazole (6x albendazole arm), (2) 800 mg albendazole at M0 and M2 and matching placebo at M4, M6, M8 and M10 (2x albendazole arm) or (3) matching placebo (placebo arm). LLM was measured during screening, before each treatment, and 14, 18, 21 and 24 months after the first treatment. Before each treatment, all participants had a general medical examination, and women up to 55 years underwent a pregnancy test. Three and seven days after each treatment, participants underwent clinical examination and questioning for any adverse events, to be followed, if clinically indicated, by a laboratory examination. Prior to the 2nd to 6th treatment and at the 14-months follow up, participants were asked about any adverse events since the last evaluation.
The study was conducted in the Mvila Division in the rain forest of the Southern Region of Cameroon in areas with high loiasis endemicity [1], but <20% prevalence of onchocerciasis (http://www.who.int/apoc/countries/cmr/en/index.html) and thus without CDTI.
Volunteers aged 18–65 years were eligible if they had a LLM >15,000 mf/ml at screening, had no plans to move out of the area over the following 2 years and had given informed consent. Individuals with past or current history of neurological or neuropsychiatric disorders, clinical or laboratory evidence of significant liver and kidney disease, anaemia, intestinal helminth infection, pregnancy, a serious medical condition or any other conditions which should exclude them from the study in the principal investigator’s (JK) opinion, treatment with benzimidazoles during the previous 12 months or with self-reported allergy to benzimidazoles were not eligible.
Participants were identified in a two-step procedure: (1) Screening for Loa loa infection: After community mobilization, screening for Loa loa infection was performed in the study area from January to March 2007 among all who had given individual written informed consent; (2) Participant selection: Individuals with LLM >15,000 mf/ml at screening and potentially willing to participate in the study, received detailed information about the study, gave informed written consent to study participation and underwent the evaluations to assess their study eligibility. Baseline LLM measurement and first treatment took place between 4 and 12 weeks after screening for loiasis.
Eligible individuals were stratified by the LLM obtained during screening: ≤30,000 mf/ml, 30,001 to 50,000 mf/ml and >50,000 mf/ml. Within each stratum, participants were assigned to one of the three treatment arms based on three randomization lists, one for each stratum, prepared by an independent statistician using a random digit table. Lists of eligible participants by stratum were provided to an independent pharmacist not otherwise involved in the study who assigned the treatment on the randomization list for that stratum which corresponded to the position of the participant on the eligible participant list. The pharmacist then prepared treatment packages with the required number of 200 mg albendazole and matching placebo tablets provided by GlaxoSmithKline (GSK).
The treatment packages were provided to the principal investigator (JK) labelled only with participant identifying information which allowed all but the pharmacist to be blinded.
Treatments were taken orally under direct observation by the investigators, 15–30 minutes after a fatty meal (fatty buns with additional ~15g butter). The first treatment occurred in March 2007.
Calibrated blood smears (CBS) to measure LLM were obtained between 11:00 and 15:00 to account for the diurnal periodicity of L. loa microfilaria in peripheral blood [31]. For each participant, blood collection was done at the same time of day ± 1 hour throughout the study. Following a finger-prick, 50μl of blood was collected using a 50μl non-heparinized capillary tube and spread on one labelled slide during screening and across two labelled slides during the study for ease and accuracy of counting. The slides were dried at room-temperature, then stained with Giemsa. All Loa loa and Mansonella perstans microfilariae were counted at 100X magnification. All slides were read by the same blinded biologist throughout the study. A second blinded reading of all slides by that biologist was performed at the end of the study, and the average of the two readings used for data analysis. Differences between the two readings did not exceed 5%.
A Reflotron Plus (Roche) was used to measure blood levels of hemoglobin, alanine aminotransferase (ALAT), aspartate aminotransferase (ASAT), and creatinine. Creatinine clearance was estimated using the Cockroft-Gault formula. Full blood counts were performed using the ABX Pentra-120 flow-cytometer. Pregnancy tests were done using AMS ßHCG urine tests.
The protocol initially planned follow-up to 18 months (M18) after the first treatment. Following review of the data after unblinding at M18 by external advisors, a protocol amendment was put in place for LLM measurements 21 and 24 months after the first dose. Participants gave written informed consent for the extended follow-up. All procedures for CBS blood collection and reading of slides remained identical.
The primary efficacy variable was the proportion of participants whose LLM was sustainably reduced by ≥50% from the pre-treatment value (value obtained before the first treatment) from any time point after the first dose onward. A sustainable reduction was defined as a reduction at each planned measurement time point over a period of at least 4 months.
Secondary efficacy variables were (1) the proportion of participants whose LLM was sustainably reduced to <8100 mf/ml by strata and by sex, (2) the percent reduction in LLM from pre-treatment at each time point, (3) time course of LLM.
Safety variables were the frequency of adverse events up to four months after the 6th treatment by type, severity, seriousness and relationship to study drug assessed by the investigator while blinded. Severity was graded as mild (event is easily tolerated by the participant, causing minimal discomfort and not interfering with everyday activities), moderate (event is sufficiently discomforting to interfere with everyday activities), severe (event prevents normal everyday activities) or not applicable (events where intensity is meaningless or impossible to determine e.g. blindness). Seriousness was determined based on the serious adverse event definition in the ICH guidelines (any untoward medical occurrence that at any dose results in death, is life-threatening, requires inpatient hospitalization or prolongation of existing hospitalization, results in persistent or significant disability/incapacity or is a congenital anomaly/birth defect and is related to any dose of a medicinal product or the doses normally used in man) [8]. Likelihood of relationship of the adverse event to study drug, i.e. presence of adverse drug reactions (defined as per ICH criteria [8] as 'any noxious and unintended responses to a medicinal product related to any dose or the doses normally used in man'), was assessed based on temporal association with drug administration and biological plausibility taking into account known adverse reactions to albendazole, the participant’s underlying clinical state and known adverse reactions to concomitant treatments.
Assuming that less than 1/1,000,000 placebo treated participants would have a 50% reduction in LLM for at least 4 months and at least 50% of participants receiving 6 albendazole doses would have such a reduction, a sample size of 16 participants per treatment provides ≥90% power to detect the treatment difference at a 2.5% two-sided significance level. The same assumptions were made regarding the effect of 2 albendazole doses. The significance level of 2.5% was chosen based on Bonferoni correction for the two planned comparisons. Assuming attrition of 20% of participants, 20 participants were recruited into each treatment group. 20 participants provide a probability of 0.87 to detect at least one adverse event with a true frequency of 10%.
All participants who received at least one dose of study drug were included in the safety analysis and analyzed as randomized.
For the efficacy analyses, participants were analysed as randomized and as part of the stratum they qualified for based on pre-treatment LLM, not the stratum they qualified for based on the screening LLM used for randomization (see Table 2).
All participants with sufficient post-treatment LLM measurements to determine whether or not they had an LLM reduction for ≥ 4 months (i.e. at least two successive measurements over a minimum of 4 months) were included in the efficacy analyses. An intent-to-treat approach was taken with participants being evaluated based on the treatment group they were randomized to, independent of whether they had received the intended number of doses.
The study protocol and protocol amendment received clearance from Cameroon’s National Ethics Committee and from the World Health Organization Ethical Review Committee. The study was granted administrative authorization by the Ministry of Public Health of Cameroon. Study participants gave written informed consent before any study procedures were conducted.
A total of 2005 people from 37 communities were included in the screening for Loa loa infection. In villages in which more than 10 participants were examined, between 10% and 49% of those screened were found to be infected. LLM was ≥15,000 mf/ml in 89 people. After screening for inclusion and exclusion criteria, 60 participants were randomized based on the LLM at screening. Screen failure reasons are shown in Fig 1.
Table 2 summarizes the LLM data obtained during screening for Loa loa infection and at the pre-treatment examination (M0, immediately before the first treatment) as well as other characteristics of the participants. The LLM pre-treatment were in some participants significantly different from the LLM at screening, with pre-treatment LLM ranging from 50% to around 500% of screening LLM. In two participants randomized to the 6x albendazole arm, the LLM dropped from 15000 mf/ml at screening to 11040 mf/ml and 12700 mf/ml, respectively, at the baseline examination. In eight participants randomized to placebo, five participants randomized to 2x albendazole and four participants randomized to 6x albendazole, the LLM measured at baseline was so different from the LLM at screening (Fig 2) that it did not fall within the stratum in which they had been randomized. This resulted in the imbalance in allocation to treatment arms in the different strata when the pre-treatment values are evaluated. There was, however, no statistically significant difference in pre-treatment LLMs between treatment arms (Table 2). As expected, ALAT and ASAT levels were highly correlated (Spearman’s correlation coefficient = 0.77), and only ALAT levels were used in the longitudinal trend analysis.
At baseline, seven participants had detectable levels of Mansonella perstans microfilariae including one in the placebo group (440 mf/ml), four in the 2x albendazole group (120 mf/ml, 440 mf/ml, 700 mf/ml, 3820 mf/ml) and two in the 6x albendazole group (200 mf/ml, 1500 mf/ml).
Up to 2 months after the last albendazole or placebo dose administered, 15, 13 and 15 participants were found to have a total of 46, 48 and 45 adverse events, respectively, in the placebo, 2x albendazole and 6x albendazole arm. None was regarded as study drug related. Across all participants, the most frequently reported AEs were different types of pain (e.g. arthralgia, myalgia, back pain, pain in extremities) and malaria.
The AEs were of mild or moderate intensity except for three severe adverse events which also met the criteria for SAEs (see Table 1). Upon unblinding, the participants with SAEs were found to have been in the placebo group. One participant sustained a severe chest trauma in a fight two days before the sixth treatment. One participant developed severe malaria and typhoid fever one week after the first treatment and was excluded from further treatment. Another participant, a 47 year old man, died 6 weeks after the 5th treatment; his death occurred at home and was preceded by a short illness including fever, cough, seizure and coma. The post-mortem was not able to establish the cause of death. The Loa loa microfilaraemia was 40,000 mf/ml at screening, 197,000 mf/ml at baseline and 440,080 mf/ml, 332,060 mf/ml, 410,200 mf/ml and 144,360 mf/ml two months after the first, second, third and fourth placebo dose, respectively. The pathology found minor brain haemorrhages, suggesting that Loa loa infection could have been a contributing factor.
During the 24 month follow up period, 11 participants who did not have detectable levels of M. perstans at baseline, had detectable levels at least once. In some participants, M. perstans levels varied significantly over time without any indication of an effect of 2x or 6x albendazole. The M. perstans microfilariae levels in all participants in whom detectable levels were detected at least once are shown in Fig 4.
As anticipated based on prior knowledge of albendazole [27–30, 33–36], no mild, moderate or severe adverse drug reactions were recorded.
This study evaluated whether a 2 dose and/or a 6 dose albendazole treatment regimen result in a ≥50% reduction in LLM from pre-treatment for at least 4 months. Depending on the extent and duration of LLM reduction, the regimen could be considered for community wide treatment in Loa loa co-endemic areas before ivermectin or ivermectin-albendazole mass treatment to reduce the risk of severe ADRs or SARs or for further improvement of the regimen.
The results show that only the 6 dose regimen had a LLM reducing effect. The time course of LLM reduction, and the absence of adverse drug reactions known to occur upon treatment of individuals with high LLM with microfilaricidal drugs support the conclusions from prior studies [28–30] that the LLM reducing effect of albendazole is likely due to microfilariae dying as they reach the end of their life span. They are not being replaced because albendazole binding to β-tubulin disrupts microtubule structure, function and formation which results in macrofilariae starvation and inhibition of reproduction [37, 38].
Around 50% of participants in the 6x albendazole arm experienced a sustained LLM decrease by ≥ 50%. None of the 9 participants with LLM >30000 mf/ml before treatment had a sustained LLM decrease to <8100 mf/ml. The LLM lowering effect of the 6x albendazole regimen is therefore insufficient to significantly reduce the population at highest risk of severe ADRs or SARs upon ivermectin mass treatment. This study thus adds to the body of data showing insufficient efficacy of different regimens of albendazole for reducing LLM. In contrast to our study, these studies included participants with pre-treatment LLM <8100 mf/ml, no or an unspecified number of participants with LLM > 30000 mf/ml and the results were presented only via summary statistics across all participants [27–30].
The cost-benefit of further efforts to improve an albendazole based treatment regimen, e.g. through sustained release formulation technology, for LLM reduction needs to be carefully considered. These considerations need to take into account the dose- and time-dependent pharmacokinetics of albendazole, including the inter-subject variability in albendazole bioavailability and conversion to the active metabolite albendazole sulfoxide, the fact that albendazole induces its own disposal during long term treatment resulting in decreasing levels of albendazole sulfoxide [39–41], the potential toxicity associated with long term exposure (http://www.accessdata.fda.gov/drugsatfda_docs/label/2015/020666s009lbl.pdf) and the time and cost to develop an affordable, safe and efficacious dose. These considerations also need to include the alternative drugs and approaches in development. An oral flubendazole formulation is now being evaluated for clinical development for onchocerciasis [42]. Prior clinical data suggest that flubendazole does not have a microfilaricidal effect, but leads to a slow reduction in microfilariae levels through an effect on the O. volvulus macrofilariae [43]. Large scale efforts to discover novel antibiotics targeting the Wolbachia endosymbionts in the filariae that cause onchocerciasis and lymphatic filariasis are under way [44]. Doxycycline treatment of O. volvulus infected individuals has provided proof-of-concept for the effect of antibiotics on the reproductive activity and viability of the macrofilariae without microfilaricidal activity [45, 46]. One study of doxycycline in O. volvulus infected people included 22 people with a pre-treatment LLM of < 8000 mf/ml. No adverse events of the type and severity observed after ivermectin treatment of people with high LLM were reported[47].
Alternate approaches to onchocerciasis and lymphatic filariasis control in Loa loa co-endemic areas are under evaluation. This includes development of diagnostics for high levels of infection with Loa loa to identify individuals at risk for severe ADRs and/or SARs to ivermectin [48]. If the ongoing field testing is successful, Loa loa infected individuals at risk of ADRs/SARs and co-infected with O. volvulus, could be treated with regimens of antibiotics already shown to be effective against O. volvulus. The implementation of this approach, including the 'cut-off' for exclusion from ivermectin treatment and the time between LLM measurements and treatment, needs to take into account the substantial intra-individual LLM variability observed in the absence of treatment in this study (Table 2, Fig 2, Fig 3). Research on LLM variability within shorter intervals than the 1–4 months in our study may be needed to inform the maximum time frame between LLM measurement and safe ivermectin treatment.
The level of variability we observed in the absence as well as during and after treatment has to our knowledge not previously been reported. It needs to be taken into account during review of the other studies which evaluated the effect of albendazole regimens on LLM based on summary statistics [27–30]. Analysis of the data from this study via geometric mean LLM (Fig 2), shows a progressive LLM decrease in the 6x albendazole arm from M2-M14 to around 50% of pretreatment levels. Only the review of the individual participant data (Fig 2, Table 3) showed that this mean decrease was driven by only a few individuals and that start time relative to treatment and the duration of LLM decrease differed between individuals (Table 4). Any treatment to ensure safe ivermectin mass treatment has, however, to reduce LLM below the level of risk in all to be treated with ivermectin and the start time relative to treatment and the duration of the LLM decrease below the risk level needs to reliable. Consequently, LLM variability needs to be taken into account in the design, analysis and reporting of all future studies on the efficacy and safety of drugs or strategies for addressing loiasis as an obstacle for onchocerciasis and lymphatic filariasis control and elimination and as a neglected disease that can negatively impact people's health, well-being and health care costs.
|
10.1371/journal.pgen.1007860 | GRAM: A GeneRAlized Model to predict the molecular effect of a non-coding variant in a cell-type specific manner | There has been much effort to prioritize genomic variants with respect to their impact on “function”. However, function is often not precisely defined: sometimes it is the disease association of a variant; on other occasions, it reflects a molecular effect on transcription or epigenetics. Here, we coupled multiple genomic predictors to build GRAM, a GeneRAlized Model, to predict a well-defined experimental target: the expression-modulating effect of a non-coding variant on its associated gene, in a transferable, cell-specific manner. Firstly, we performed feature engineering: using LASSO, a regularized linear model, we found transcription factor (TF) binding most predictive, especially for TFs that are hubs in the regulatory network; in contrast, evolutionary conservation, a popular feature in many other variant-impact predictors, has almost no contribution. Moreover, TF binding inferred from in vitro SELEX is as effective as that from in vivo ChIP-Seq. Second, we implemented GRAM integrating only SELEX features and expression profiles; thus, the program combines a universal regulatory score with an easily obtainable modifier reflecting the particular cell type. We benchmarked GRAM on large-scale MPRA datasets, achieving AUROC scores of 0.72 in GM12878 and 0.66 in a multi-cell line dataset. We then evaluated the performance of GRAM on targeted regions using luciferase assays in the MCF7 and K562 cell lines. We noted that changing the insertion position of the construct relative to the reporter gene gave very different results, highlighting the importance of carefully defining the exact prediction target of the model. Finally, we illustrated the utility of GRAM in fine-mapping causal variants and developed a practical software pipeline to carry this out. In particular, we demonstrated in specific examples how the pipeline could pinpoint variants that directly modulate gene expression within a larger linkage-disequilibrium block associated with a phenotype of interest (e.g., for an eQTL).
| With advances in sequencing technologies, a deluge of genomic data is available; however, only a fraction of non-coding genomic variants are functionally relevant. Sifting through this data to prioritize genomic variants with respect to function is an important but challenging task. In this study, we built GRAM, a GeneRAlized Model, to predict the expression-modulating effects of non-coding variants in a cell-specific manner. GRAM combines a universal regulatory score defined by transcription factor binding with an easily obtainable modifier defined by transcription factor binding and expression to reflect the particular cell type. We evaluated this framework on multiple cell lines with high performance and showed that it could be applied to any cell line or sample with gene expression data. We also integrated GRAM into a practical software pipeline to fine-map causal variants that directly modulate gene expression among a larger linkage-disequilibrium block associated with a phenotype of interest. GRAM complements other general variant effect prediction methods–which often combine disparate features–by helping to precisely define the subset of prioritized variants that directly alters gene expression.
| Advances in next-generation sequencing (NGS) technologies have enabled high-throughput whole genome and exome sequencing [1], which have led to the identification and characterization of many disease-associated mutations [2] and the vast majority of common single nucleotide variants (SNVs) in the human population [3, 4]. Genome-wide association studies (GWAS) have found that these variants mostly lie outside of protein-coding regions [5], emphasizing the functional importance of non-coding regulatory elements in the human genome. These advances have also led to an urgent need to develop high-throughput methods to sift through this deluge of sequencing data to quickly determine the functional relevance of each non-coding variant [6].
Evidence suggests that only a fraction of non-coding variants are functional, and the majority of functional variants show only modest effects [7]. Studies like GWAS [8] and expression quantitative trait eQTL [9] have evaluated the association of variants with traits of interest from a statistical perspective. In traditional GWAS and eQTL analyses, an association locus may host the tag-SNPs and a number of linked variants that may potentially account for the molecular mechanism underlying the association [10]. However, it remains difficult to distinguish those that are truly causal [11–13]. Thus, downstream analysis requires fine-mapping to identify the true causal variants by integrating the external genetic and epigenetic information [12, 14].
As association studies give little information about the mechanism of a variant’s effects, it would be helpful to directly test the molecular effects of a large numbers of variants using highly quantitative assays. Luciferase reporter assays are a common method to measure the regulatory effects of functional elements [15]. Researchers can compare the difference of luciferase expression with and without a mutation to estimate the experimental molecular effect of non-coding variants lying in a functional element. By using high-throughput microarray and NGS technology, the massively parallel reporter assay (MPRA) has extended the scales to the genome-wide level [16–21]. Recently, Tewhey and colleagues demonstrated the capability of MPRA to identify the causal variants that directly modulated gene expression [22, 23]. This study identified 842 expression-modulating variants (emVARs) showing significantly differential expression modulation effects and provided a high-quality data source for computational modeling [22, 23].
There is an increasing need for computational methods to effectively predict the molecular effects of variants and improve our understanding of the underlying biology of these effects. Several approaches have been developed to address the problem of variant prioritization from different perspectives. Based on the target of predictions, these methods roughly fall into three major categories: 1) disease-causing effect predictors (e.g. GWAVA [24], and GenoSkyline [25]), which aim to prioritize causal disease variants and distinguish them from benign ones; 2) fitness consequence prioritization tools (e.g., CADD [26], fitCons [27] and LINSIGHT [28]), which attempt to identify the variants based on evolutionary fitness; 3) comprehensive tools (e.g., DeepSEA [29], FunSeq2 [6], FUN-LDA [30]) which integrate multiple data sources for prediction of functional variants. Many of these computational methods are designed to predict and prioritize deleterious and disease-associated variants from a phenotypic perspective, but not to highlight specific molecular consequences of these variants (i.e., their effects on the activities of functional elements). Moreover, some of these tools are cell type-agnostic, and tools that are cell type-aware depend on cell type-specific data with somewhat limited availability, such as ChIP-Seq or epigenetic features. Thus, it would be helpful to build a generalized model that can be systematically specialized to any desired cell type with only a small amount of easily obtainable cell type-specific information (e.g. expression data).
In this study, we addressed the problem of molecular effect prediction of variants from a different perspective. Instead of predicting phenotypic consequences from genotypes, which is a common practice, we aimed to directly predict the expression-modulating effect of the variants from various sources of information. Our model, named GRAM (i.e., GeneRAlized Model), incorporates selected transcription factor (TF) binding information from in vitro SELEX assays, representing the general binding affinity of TFs on the variant’s location, and cell type-specific expression profiles, representing cellular contexts. Combining cell type-independent and -dependent features makes our model both flexible and specific. When we evaluated results from MPRA and luciferase assay experiments show our model achieved high predictive performance and could be easily transferred to other cell types and assay platforms. We also demonstrated the potential application of GRAM to the fine-mapping of pre-defined variants in linkage disequilibrium. As a supplement to many general variant effect prediction methods (which often combine disparate features), our model can help to precisely define the subset of prioritized variants that directly alters gene expression. For instance, after using a more general functional impact tool such as FunSeq or VEP [31, 32], one could use GRAM on the prioritized variants to identify the subset that has a direct expression modulating effect (as opposed to being prioritized for other reasons such as strong association with an organismal phenotype). Furthermore, one could use GRAM to fine-map the key causal variant modulating gene expression from the many variants in a linkage-disequilibrium block associated with gene expression in an eQTL study.
In this study, we first collected a dataset from Tewhey et al. [22] to estimate expression modulation differences between reference allele and mutants in the GM12878 cell line. This MPRA-generated dataset contains 3,222 SNVs filtered by logSkew value, which measures the log-fold change of the expression-modulating differences between reference and alternative alleles. Among them, 792 variants (named emVARs) had a significant expression-modulating effect compared with their respective reference allele, which indicates the molecular effect of the variant. Here, we treated emVARs and non-emVARs as positive and negative dataset, respectively, in our GRAM model.
As described in Fig 1, our GRAM model is implemented in three steps: (i) prediction of the universal regulatory consequences of an element with variant using the SELEX TF binding score; (ii) prediction of a cell type modifier score in a specific cellular context by combining TF binding score with cell type-specific TF expression profiles; and (iii) estimation of the expression modulating effect in a cell type-specific context by integrating outputs from the previous two steps.
We first investigated the potential of evolutionary conservation and transcription binding features as predictors. Evolutionary conservation is associated with deleterious fitness consequence and is widely used in prioritization algorithms of non-coding variants, such as PhyloP [33] and PhastCons [34] scores in LINSIGHT [28] and CADD [26], and GERP [33] score in FunSeq2 [6]. We performed comparative analyses for these three conservation features across different datasets (S1 Fig). We found that the PhastCons and PhyloP patterns of emVARs and non-emVARs are different from Human Gene Mutation Database (HGMD) [35] variants but similar to non-HGMD variants, which are thought to be benign. GERP scores show a similar pattern but have smaller variance in emVARs and non-emVARs compared to other datasets, with slightly larger values for emVARs. As we did not find differential patterns when comparing emVARs and non-emVARs, we further discovered that the correlation between logSkew and all three conservation scores was low (close to 0) by linear regression. These results suggest that the conservation scores might contribute little to the molecular effects under study that focuses on expression modulation of variants in more conserved regions with homogeneous evolutionary patterns.
TF binding can link the molecular effect of non-coding variants to a cascade of a regulatory network, which is thought to be an important contributing factor to the variants’ regulatory effects [26, 29, 36, 37]. Tewhey et al. found that the logSkew value positively associates with TF binding scores. To thoroughly evaluate the effect of TF binding, we tested TF binding peaks overlapping with the SNVs and TF motif break events in the Tewhey dataset. We annotated and analyzed the emVAR and non-emVAR variant sets with FunSeq2 [6], and found that the emVAR set had more TF binding events compared with the non-emVAR set (Fig 2A). In addition to TF binding enrichment, we examined the motif breaking scores for these TFs. After removing TFs with insufficient observations, the differences between the distributions of motif-break scores for alternative and reference alleles in emVARs are larger than those in the non-emVAR dataset (Fig 2B). According to this analysis, the emVAR set tends to have not only more TF binding events, but also larger binding alterations compared with the non-emVAR set. Our results indicate that TF binding shows high association with the expression-modulating effects of the variants and align with recent studies on the underestimated relative importance of transcription [38, 39].
We generated a candidate training feature set from the outputs of 515 DeepBind models for TF binding, inferred from both ChIP-Seq [40] and in vitro SELEX assays [41], on the adjacent sequences of the variant of interest. With a comprehensive feature selection framework for selection of impactful TF binding features, we prioritized these features across models with LASSO stability selection [42] and Random Forest (shown in Fig 3A). The 20 most important features (out of 515) with respect to the mean importance across all methods is shown in decreasing order in Fig 3A. Both ChIP-Seq and SELEX DeepBind features showed high importance, with the top two being GM12878 ChIP-Seq features (SP1 and BCL3), which are cell line specific, followed by SELEX features starting with ETP63. The top-ranked impactful TFs tend to have more protein-protein interactions than the bottom-ranked TFs, indicating that the importance of a TF reflects its role in the TF-TF cascade regulatory network (Fig 3B).
Interestingly, many SELEX features, though not cell type dependent, achieved similar predictive power as cell type-specific ChIP-Seq features. We compared the predictive performances of cell type-dependent ChIP-Seq features, cell type-independent SELEX features, and a combination of both feature sets using a LASSO regressor, support vector machine (SVM) regressor and Random Forest. Incorporating ChIP-Seq-derived features, though introducing more cell type-specificity, did not boost the accuracy significantly for any of the three models (Fig 3C and S1 Table). As the availability of ChIP-Seq data is restricted to a few cell lines (S2 Fig), we instead used SELEX features to build a more generalized model that can be easily applied to different cell types.
We then used the features generated from disease-association prediction tools (CADD [43], FunSeq2 [32], DeepSEA [44], GWAVA [45], LINSIGHT [46], and Eigen [47]) to predict the same molecular effect target. As shown in Fig 3C, this analysis indicated that the prediction of disease-associated variants is not equivalent to that of expression-modulating variants.
Using the TF binding features from DeepBind models and the MPRA dataset from Tewhey et al. [22], we implemented our multi-step model. In the first step, we predicted the universal regulatory activity of an element with or without a variant. The 10-fold cross validation demonstrated exemplary performance of the model with an area under the receiver operating characteristic curve (AUROC) of 0.938 and an area under the precision-recall curve (AUPRC) of 0.928 (Fig 4A and S3 Fig).
In the second step, we calculated a cell-type modifier score as an indicator of the experimental assay’s cell-specific nature. Briefly, we defined the prediction target using a top and bottom quantile of Vodds (S5 Fig). Vodds is the standard deviation of log odds for each variant’s read count in MPRA, which reflects the confidence interval of log odds ratio of an experiment. Vodds shows cell line-specific patterns, as the patterns of the two B-Lymphocyte cell lines (NA12878 and NA19239) are similar while distinct from HepG2 (S4 Fig) (see Methods for details). This indicates that Vodds can capture the cell type-specific information. We also found that variants with higher Vodds tend to include more non-emVARs (Chi-square test p-value: 0.0002). Hence, the cell type modifier score defined from Vodds can be used to adjust the universal regulatory effect to a cell type-specific context.
Gene expression profiles, especially TF expression profiles, are more generally available and can represent the cellular environment. We incorporated TF gene expression and TF binding scores as features to predict the cell type modifier target, and got an AUROC of 0.66 and 0.8 (Fig 4D), respectively, using Random Forest with a 10-fold cross-validation (Fig 4B and S6 Fig).
The final step is to predict the molecular effect of a variant, i.e. whether it can significantly modulate reporter gene expression. To do this, we fed the output from the first and second step into a LASSO model, with the emVAR and non-emVAR labels as targets. We found that the AUROC of a 10-fold cross-validation for the optimal model was 0.724 (Fig 4C) and the AUPRC was 0.602, both of which are higher than the state-of-the-art method (KSM) using the same dataset (AUROC: 0.684, AUPRC: 0.478) [48].
To achieve better generalizability, we built the model with SELEX features only. We performed step (i) and (ii) on the same GM12878 dataset and another multiple-cell-line dataset (MCL dataset: GM12878 plus HepG2 plus K562). The model with cell-independent features from the SELEX assay achieved comparable performance with an AUROC = 0.664 (GM12878 only) and 0.658 (MCL dataset, Fig 4D). We use the model based on the multiple-cell-line dataset in our final GRAM model for a better generalization potential.
We next evaluated performance of the model on different cell types and assay platforms. Rather than measuring read counts as in MPRA, some other assays, such as luciferase and GFP reporter assays, measure luminescence and fluorescence readouts instead. [49, 50]. To evaluate how our model, trained with multiple cell line MPRA data, can be transferred to these assay platforms we tested its performance on luciferase assay results of eight potential regulatory elements with mutations from the MCF7 cell line [51]. To predict expression-modulating effects, we defined the significant changes between alternative and reference alleles by using an absolute log2(odds ratio) cutoff. The average AUROC value was greater than 0.8 for MCF7 (Fig 5A) and 0.67 for K562 given the an absolute log2 cutoff from 0.5 to 0.8 (Fig 5B). This indicates that our model performs very well on the luciferase assay and MPRA dataset from different cell lines, even though these assays use different measurements.
In MPRA, the element is inserted upstream (5’-terminal) of the reporter gene, but for some assays, such as STARR-Seq, the element is inserted downstream (3’-terminal). Therefore, we further tested the effect of insertion location of an element in luciferase reporters in K562 cells using 14 randomly selected elements with potential regulatory activity. As shown in Fig 5C, the 5’ terminal log odds were similar to the 3’ terminal odds for region 3, 4, 5, and 13, but showed significant differences for region 6, 8, 9, 10, and 14. The prediction of GRAM for the 5’ terminal was much better than that for the 3’-terminal insertions; the AUROC was 0.25 higher for universal regulatory activity and 0.32 higher for the expression-modulating effect prediction, indicating different mechanisms for the two ends. Therefore, GRAM model is optimal for 5’ terminal assays.
As GRAM needs only gene expression and SELEX DeepBind score to predict sample-wise variants effect, it could be a flexible tool for a variety of analysis tasks. We investigated whether we could apply our GRAM model to fine-mapping of causal variants. As was described in the Methods part, we made a user-friendly pipeline GRAMMAR that could conduct the entire analysis (S9 Fig). Here we mainly focused on the task of identifying the variants that are most likely to directly modulate gene expression. For our analysis, we selected five LD blocks with known risk association with prostate cancer and high enrichment of annotated eQTL SNPs reported by Dadaev et al. [10], resulting in a set of 561 eQTL SNPs from the five LD blocks. We extracted the genotypes and gene expression data from 102 The Cancer Genome Atlas (TCGA) PRAD patients and ran GRAMMAR to get the prediction score for each allele in each patient (S4 Table).
In general, variants with high posterior probability (≥0.5, 130 variants), as a causal variant, reported by Dadaev et al. [10], generally have higher average GRAM scores as compared to those with lowest posterior probability (<0.5, 4260 variants) (p-value = 0.0545, S7 Fig). Specifically, we took a closer look at region chr6:160081543–161382029, tagged by GWAS SNP rs9364554 and enriched with 52 eQTL SNPs for genes including ACAT2, LOC729603, MRPL18, SLC22A3 and WTAP. All the FunSeq2 scores (maximum 1.40) are below 2, an empirical threshold for confident candidate causal SNVs. GRAMMAR, however, can pinpoint three SNV candidates with the highest average GRAM scores in this region (Fig 6A). Their GRAM scores differ in different patient samples, indicating different expression modulating effects of these SNVs under different personalized cellular contexts. Moreover, all three of the highest-scored variants show strong correlations between the GRAM expression modulating score and the expression of the related target gene and two of which are significant (p-value < 0.05) (Fig 6B–6D).
There has been an increasing number of computational methods that can prioritize non-coding variants. In addition, accumulating high-throughput whole-genome sequencing data have become the primary source for identifying disease-associated variants. However, we still lack an efficient prediction model for estimation of the expression-modulating effect of variants that can be universally applied to many cell lines or samples. Previous studies tend to construct one distinct model for each cell type, or predict the cell-type specificity of a variant from often very limited experimental results (e.g. ChIP-Seq) in different cell types [25, 30, 52, 53], which makes the generalization to other cell types challenging. In this study, we sought to represent the impact of cellular environments on variant function from a different perspective. We developed a multi-step generalized model called GRAM that can specifically predict the cell type-specific expression-modulating effect of a non-coding variant in the context of a particular experimental assay. Our model receives both cell type-dependent and independent input data and combines them with the same set of feature weights across different contexts, Thus, our model can be applied to any cellular context as long as cell type- or sample-specific expression data are provided.
In this study, we aim to precisely define the expression-modulating effect as a function of the predictive variables extracted from genomic data. In line with results from recent studies [38, 39], a wide array of transcription-related features demonstrated high predictive power. In contrast, three selected evolutionary features demonstrated low predictive power on used datasets. This pattern is likely due to the limited variety in evolutionary patterns in the training data and also stems from the nature of GRAM, which focuses on predicting expression-modulation effects. These effects are part of the many that are related to sequence conservation [54, 55]. In other words, the purpose of our model is to enable precise downstream analysis of molecular effects of variants in a highly conserved region, where we would not expect conservation scores to provide more additional information. We further selected a variety of TF binding features that could be useful for predicting variant effects and used direct measurements from TF binding scores and implemented a straightforward LASSO regression to assess the importance of each feature. We found that in vitro SELEX TF features (aka non-cell-specific features) achieve the highest predictive performance, a result further validated by SVM and Random Forest models trained in parallel.
We cannot ignore the cell type-specific context when predicting a variant’s effects. Usually, a model can achieve cell type-specificity in two different ways: 1) building an independent model for each cell line, or 2) building one unified model that can accept and handle specific input data from any cell lines or samples. Which strategy to use depends on the availability of the dataset and the demand for model transferability. Our model uses the second strategy, in which cell type-specific information is incorporated as an input feature and the model learns the same set of feature weights across multiple cell lines. For such a unified model, features like histone modification and TF ChIP-Seq would limit its transferability because these features may not be available for many other cell types or samples. Thus, we would prefer features that are more easily available, such as gene expression profiles. Here, we built the model using cell type-dependent gene expression and cell type-independent TF in vitro SELEX features; thus, the model can be more easily applied to various different samples and cell lines. SELEX features represent general binding strength of the TFs on the region of interest, and gene expression profiles can represent the specific cellular context.
The three-step GRAM model predicts the expression-modulating effects of variants by integrating two intermediate predictive targets: universal regulatory activity and cell type modifier score. The universal regulatory activity reflects the general regulatory effect of an element with or without a mutation in a vector-based assay without considering cell type-specific chromatin contexts or epigenomics information. Next, we modeled the cellular environment related to gene regulation with a cell type modifier score, derived from cell type-specific TF expression levels, to adjust the universal regulatory effect in the final step of the prediction model, greatly improving the performance.
GRAM performed well in validations on MPRA and luciferase assay, even across different cell types. In addition to target validations, our tool enables detailed exploration of the sensitivity of these methods and the impact of vector construct. The insertion position of the element affected the outcome of the assay, which may correspond to different types of regulatory elements. Because our model is trained on 5’-terminal insertion data, the prediction is consistent with outcomes from the same position, but not for 3’-terminal assay results. This indicates different mechanisms for two insertion positions: the assay with an element inserted upstream of a reporter gene may detect either the promoter or enhancer activity of the element. However, if the element is inserted downstream of the gene’s transcriptional start site or the 3’ terminal in the assay, the reporter readout may be specifically to the enhancer activity of the element. Large-scale experimental validation is required to further elucidate the underlying mechanisms.
Our GRAM model can be further applied to fine mapping of functional SNVs. Particularly, the prediction results of GRAM could aid in the identification of variants that are most likely to directly modulate gene expression in a fine-mapping study. In addition, the impact of variants on gene regulation could vary across different cell types or individuals depending on differential transcriptional factor activity, which is represented by the expression level of TFs in our model. Based on this consideration, our model could potentially be used to evaluate the molecular effect of variants in a sample-specific manner. Given a group of patients with paired genotype and gene expression data, we could evaluate for each patient the expression-modulating effect of the variants of interest, which can be used to: 1) evaluate the patient-specific expression modulating effect for each variant; 2) identify distinct expression modulating patterns among the patient population; and 3) evaluate the overall variant effects by integrating results from different patients. Such knowledge could potentially contribute to our understanding of the molecular mechanism underlying disease-association of variants, and guide the characterization of patient-specific candidate variants for personalized diagnosis, prognosis and medical treatments.
In summary, our GRAM model will be a useful tool for elucidating the underlying patterns of variants that modulate expression in a cell type- and tissue-specific context, and can be further applied to different samples of the same cell type or tissue. By leveraging the accumulating data generated from multiple cell lines, we can further improve for in-depth investigation in the future. We will keep abreast with the growing availability of comprehensive datasets and further expand our analyses.
We downloaded the dataset from R. Tewhey et al.’s paper [22, 23]. From about 79K tested elements, we only kept variants for which either reference or alternative allele elements show regulatory activity. This reduced the set to 3,222 SNVs in the GM12878 cell line and 1124 SNVs in the HepG2 cell line. Each SNV was extended in both directions by 74bp, for a total of 149bp. We used another dataset from Ulirsch 2016 [17], which included 2,756 variants tested in the K562 cell line.
The protein-protein interaction network used in our downstream analysis was constructed by merging all interaction pairs identified by BioGrid [56], STRING [57] and InBio Map [58].
GERP features were extracted using the FunSeq2 annotation pipeline, which averages over the whole genome-scale GERP score over the elements. We downloaded phyloP [33] and Phastcons [34] scores from the UCSC genome browser data portal (http://hgdownload-test.cse.ucsc.edu/goldenPath/hg19/).
We performed motif enrichment analysis using a hypergeometric test. To compare the motif break and gain scores, we removed the TFs that covered less than two variants for either emVARs or non-emVARs from the list of 40 TFs with the highest p-values in hypergeometric test. Then, we performed a Wilcoxon test on the motif break score.
Motif break and motif gain scores were calculated using FunSeq2. We also calculated the motif score using DeepBind [37] with both the SELEX and ChIP-Seq motif models. SELEX motif models were identified from in vitro systematic evolution of ligands by an exponential enrichment (SELEX) binding assay. ChIP-Seq models were inferred from sequences of TF binding sites from different cell lines. A total of 515 motif models were investigated (S2 Table).
To examine the importance of features, we compared different metrics learned from various models including LASSO stability selection [42] and Random Forest regression. The feature importance for each selection method was scaled to [0, 1]; we took the mean of all the selection methods to represent the overall ranking.
We compared our models’ mean standard error (MSE) with CADD, Eigen, LINSIGHT, FunSeq2, GWAVA, and DeepSea. Features from the above tools were collected and tested using both SVR and Random Forest regression with three different input feature sets: SELEX-based features, ChIP-Seq-based features, and SELEX- and ChIP-Seq-based features combined. For other variant prioritization tools, we use their outputs as features to train the SVR and Random Forest models to predict the logSkew value.
We labeled emVARs as positive and non-emVARs as negative classes following the definition of [22], where ‘expression modulating’ means having a molecular effect that significantly increases or decreases regulatory activities. We calculated the emVAR and non-emVAR for both HepG2, GM12878 and K562 cell lines from [17] [22]. For emVAR and non-emVAR, we further filtered using logSkew with an absolute value >0.5849 (skew > 1.5). In total, we used 3,222 data records, including 799 positives and 2,423 negatives.
We built a three-step GRAM model (Table 1). Step 1 predicts the universal element regulatory activity U for both reference and alternative alleles. The ground-truth of regulatory activity is determined from results of experimental assay platforms, like a luciferase assay or MPRA. In these assays, an element inserted into a plasmid, either with or without a mutation, is characterized with regulatory activity if the fold change between the vector with the inserted element and the control is larger than a statistically significant cutoff. Specifically, the predictive target is defined as follows: for the MPRA study, where expression level of the reporter gene is directly measured, a statistical test based on DESeq2 was used to indicate whether the expression change is significant; for the luciferase assay, we regarded a testing element that has a fold change of fluorescence level greater than 1.5 or 2 compared to control (like eGFP) as a regulatory element. The predictive variable is the TF binding score from reference to alternative allele, which is estimated by DeepBind. A Random Forest classifier was then trained to predict the universal regulatory activity. The predicted log odds of probability between the reference and alternative allele was calculated as log2(U(imut)1−U(imut)/U(iwt)1−U(iwt)).
Step 2 predicts the gene expression and TF binding cell type modifier scores. The cell type modifier score is defined according to the cell specificity of the experimental assay. For each variant, an MPRA experiment is performed on both the reference and alternative alleles, each paired with a null-control, resulting in a 2x2 categorical table of read counts in the MPRA experiments. The standard deviation of log(odds) of the categorical table (n1, n2, n3,n4 for the average reads count, Table 2) is calculated as 1n1+1n2+1n3+1n4. For three different cell lines, GM12878, GM19239, and HepG2, we constructed a vector of Vodds values for all the variants that are tested. By comparing principal component loading of the Vodds from three cell lines, we found that the two GM cell lines are closer to each other relative to HepG2 (S3 Fig), which indicates that Vodds could reflect cell type information. We then further compared two groups of variants above the top quartile and below the bottom quartile of Vodds in GM12878, and found that there were more non-emVAR variants in the top quantile group, which indicates that Vodds are also associated with the molecular effects of the variants. Based on these observations, we used the top and bottom quartile variants as positive and negative training sets, respectively, to predict the cell type modifier target.
The TF expression profiles were used as input features for the prediction of the cell type modifier class. For each mutation, we re-ordered the expression of TFs based on their binding scores. Given 258 TFs with a DeepBind SELEX model score S for 3,222 SNVs, the TF expression matrix for each variant was adjusted and re-ordered using the rank of SELEX binding scores of the TFs bound to these SNVs’ region. For each variant, this results in a vector reflecting the expression of TFs relative to their binding strengths. That is, the first value in the vector represents the adjusted expression of the most influential TF bound to this region, i.e. the one with highest rank of binding scores, and so forth, regardless of what the TFs actually are. We then used the TF binding score and re-ordered gene expression to predict the cell type modifier label.
The final model predicts the molecular effect of a variant using the estimated universal odds ratio and cell type modifiers from the two previous Steps. A LASSO model was used for the prediction. The LASSO model trained with L1 regularization is more robust and tolerant to noise. To achieve optimal predictive performance, we chose the regularization parameter lambda λ that gives minimal mean cross-validated error.
We hold out one-fold of same variants for all steps and perform a 10-fold cross-validation (S8 Fig). We first randomly permutate all the data by rows (variants), and split them into ten evenly distributed subsets T (1, 2…, 10). We then iteratively hold out a subset Ti (i = 1, 2…, or 10), and make sure Ti are not used for training in any steps. We trained the model using the remaining subset T−i (−i: excluding i), and predicted the results of Ti to get Ti^. Finally, we concatenated all Ti^’s and evaluated the performance using AUROC and AUPRC.
We integrated data processing pipelines and the final model into a software pipeline called GRAMMAR (S9 Fig), published on (https://github.com/gersteinlab/GRAM). The user provides the variant list and gene expression data of each sample. The sequences with and without the variants are then extracted from the hg19 genome and provided as input for DeepBind. The GRAM model receives the DeepBind results and gene expression data and assigns a score for each provided variant in each sample. Finally, the program outputs the sample-specific GRAM scores for each sample, along with heatmap for all variants and samples. If variants from multiple regions are provided, each region is plotted individually. The software is also made available as a fine-mapping module to the more generalized FunSeq tool (FunSeq.gersteinlab.org), taking in the variants prioritized by the first tool and outputting the subset of them that have a direct expression modulating function.
The work by Dadaev et al. [10] reported 75 different LD blocks characterized by a known GWAS risk association for prostate cancer. Some of the SNPs in these regions were found to be significantly co-localized with identified eQTLs, annotated as eQTL SNPs. For our analysis, we selected five regions with the largest number of eQTL SNPs, which in total contains 561 eQTL SNPs. Genotype and gene expression data for 102 TCGA PRAD patients were obtained from the TCGA data portal. These data were then provided to the GRAMMAR pipeline described above.
We plotted the estimated sample-wise GRAM scores for each region, and selected variants with the highest average GRAM scores as assumed causal variants for expression modulation. As a comparison, FunSeq [6] scores for each variant were also extracted based on position and allele. To analyze the impact of these variants on gene expression, we calculated the Pearson correlation between the sample-specific GRAM scores and expression of the target genes of each eQTL variant.
Each regulatory region (both reference and alternative alleles) was separately synthesized. Enhancer regions were designed to include 250bp upstream and 250bp downstream for each enhancer region based on the candidate SNV site. These regions were then cloned into the pGL4.23[luc2/minP] vector (Promega, Cat# E841A). Each candidate region was placed upstream of the minP promoter to determine the effect of each putative enhancer region on luciferase expression. In total, 100ng of each candidate construct and 100ng of Nano-luc control was co- transfected into MCF-7 cells (5,000 cells per well in DMEM media containing 10% FBS and 1% Penicillin-Streptomycin antibiotic) using the Lipofectamine 3000 reagent (Thermo Fisher, Cat# L3000001) according to the manufacturer’s instructions. Cells were incubated for 48 hrs before reading the luciferase signal using the Promega Nano-Glo luciferase kit (Promega, Cat# N1521) according to the manufacturer’s instructions.
|
10.1371/journal.pcbi.1005145 | Forgetting in Reinforcement Learning Links Sustained Dopamine Signals to Motivation | It has been suggested that dopamine (DA) represents reward-prediction-error (RPE) defined in reinforcement learning and therefore DA responds to unpredicted but not predicted reward. However, recent studies have found DA response sustained towards predictable reward in tasks involving self-paced behavior, and suggested that this response represents a motivational signal. We have previously shown that RPE can sustain if there is decay/forgetting of learned-values, which can be implemented as decay of synaptic strengths storing learned-values. This account, however, did not explain the suggested link between tonic/sustained DA and motivation. In the present work, we explored the motivational effects of the value-decay in self-paced approach behavior, modeled as a series of ‘Go’ or ‘No-Go’ selections towards a goal. Through simulations, we found that the value-decay can enhance motivation, specifically, facilitate fast goal-reaching, albeit counterintuitively. Mathematical analyses revealed that underlying potential mechanisms are twofold: (1) decay-induced sustained RPE creates a gradient of ‘Go’ values towards a goal, and (2) value-contrasts between ‘Go’ and ‘No-Go’ are generated because while chosen values are continually updated, unchosen values simply decay. Our model provides potential explanations for the key experimental findings that suggest DA's roles in motivation: (i) slowdown of behavior by post-training blockade of DA signaling, (ii) observations that DA blockade severely impairs effortful actions to obtain rewards while largely sparing seeking of easily obtainable rewards, and (iii) relationships between the reward amount, the level of motivation reflected in the speed of behavior, and the average level of DA. These results indicate that reinforcement learning with value-decay, or forgetting, provides a parsimonious mechanistic account for the DA's roles in value-learning and motivation. Our results also suggest that when biological systems for value-learning are active even though learning has apparently converged, the systems might be in a state of dynamic equilibrium, where learning and forgetting are balanced.
| Dopamine (DA) has been suggested to have two reward-related roles: (1) representing reward-prediction-error (RPE), and (2) providing motivational drive. Role(1) is based on the physiological results that DA responds to unpredicted but not predicted reward, whereas role(2) is supported by the pharmacological results that blockade of DA signaling causes motivational impairments such as slowdown of self-paced behavior. So far, these two roles are considered to be played by two different temporal patterns of DA signals: role(1) by phasic signals and role(2) by tonic/sustained signals. However, recent studies have found sustained DA signals with features indicative of both roles (1) and (2), complicating this picture. Meanwhile, whereas synaptic/circuit mechanisms for role(1), i.e., how RPE is calculated in the upstream of DA neurons and how RPE-dependent update of learned-values occurs through DA-dependent synaptic plasticity, have now become clarified, mechanisms for role(2) remain unclear. In this work, we modeled self-paced behavior by a series of ‘Go’ or ‘No-Go’ selections in the framework of reinforcement-learning assuming DA's role(1), and demonstrated that incorporation of decay/forgetting of learned-values, which is presumably implemented as decay of synaptic strengths storing learned-values, provides a potential unified mechanistic account for the DA's two roles, together with its various temporal patterns.
| Electrophysiological [1] and fast-scan cyclic voltammetry (FSCV) [2, 3] studies have conventionally shown that dopamine (DA) neuronal activity and transmitter release respond to unpredicted but not predicted reward, consistent with the suggestion that DA represents reward-prediction-error (RPE) [1, 4]. On the other hand, recent FSCV studies [5–8] have found DA response sustained towards presumably predictable reward, arguing that it may represent sustained motivational drive. DA's roles in motivation processes have long been suggested [9–13] primarily from pharmacological results. A key finding is that post-training blockade of DA signaling causes motivational impairments such as slowdown of behavior (e.g., [14]), and this is difficult to explain with respect to the known role of DA in RPE representation because post-training RPE should be negligible so that blockade of RPE should have little impact.
Therefore it has been considered that DA has two distinct reward-related roles, (1) representing RPE and (2) providing motivational drive, and these are played by phasic and tonic/sustained DA, respectively. Normative theories have been proposed for both the role as RPE [4] and the role as motivational drive [15, 16] in the framework of reinforcement learning (RL). On the other hand, as for the underlying synaptic/circuit mechanisms, much progress has been made for the role as RPE but not for the role as motivational drive. Specifically, how RPE is calculated in the upstream of DA neurons and how released DA implements RPE-dependent update of state/action values through synaptic plasticity have now become clarified [17–20]. In contrast, both the upstream and downstream mechanisms for DA's motivational role remain more elusive.
In fact, FSCV studies that found sustained DA signals [5, 8] have shown that those DA signals exhibited features indicative of RPE. Moreover, sustained response towards presumably predictable reward has also been found in the activity of DA neurons [21, 22], and these studies have also argued that the DA activity represents RPE. Consistent with these views, we have recently shown [23] that RPE can actually sustain after training if decay/forgetting of learned values, which can presumably be implemented as decay of plastic changes of synaptic strengths, is assumed in RL. It was further indicated that whether RPE/DA sustains or not can be coherently understood as reflecting differences in how fast learned values decay in time: faster decay causes more sustained RPE/DA. However, this account did not explain the suggested link between sustained DA and motivation. Even on the contrary, decay of learned values is apparently wasteful and could be perceived as a loss of motivational drive.
In several recent studies reporting sustained DA signals [5–8], a common feature is that self-paced actions are required, as argued in [8]. We conjectured that this feature could be critical for the putative motivational functions of sustained DA signals. However, in our previous study [23], such a feature was not incorporated: our previous model was extremely simple and assumed that the subject automatically moved to the next state at every time step. In the present work, we constructed a new model, which incorporated the requirement of self-paced approach towards a goal, represented as a series of ‘Go’ or ‘No-Go’ (or ‘Stay’) selections, into RL with decay of learned values. Using this new model, we investigated: (1) if the model (as well as the previous non-self-paced model) generates both phasic and sustained RPE/DA signals so that their mechanisms can be coherently understood, (2) if the model demonstrates any association between sustained DA signals and motivation, and (3) if the model can mechanistically account for the key experimental findings that suggest DA's roles in motivation, specifically, the (i) slowdown of self-paced behavior by post-training blockade of DA signaling [14], (ii) severe impairment of effortful actions to obtain rewards, but not of seeking of easily obtainable rewards, by DA blockade [11, 24], and (iii) relationships between the reward amount, the level of motivation reflected in the speed of behavior, and the average level of DA [7]. Through simulations and mathematical (bifurcation) analyses, we have successfully answered these questions.
We modeled a behavioral task requiring self-paced voluntary approach (whether spatially or not) towards a goal as a series of ‘Go’ or ‘Stay’ (‘No-Go’) selections as illustrated in Fig 1. We then simulated subject's behavior by a temporal-difference (TD) RL model incorporating the decay of learned values (referred to as the ‘value-decay’ below). Specifically, we assumed that at every time step the subject selects ‘Go’ or ‘Stay’ depending on their learned values, which are updated according to RPE (TD error) when the corresponding action is taken. In addition, we also assumed that the learned values of all the actions (whether selected or not) decay in time at a constant rate (see the Materials and Methods for details). RPE at each time step was assumed to be represented by the level of DA at the time step, and the value decay was assumed to be implemented as a decay of plastic changes of synaptic strengths storing learned values.
Fig 2A shows the number of time-steps needed for goal-reaching (i.e., from the start to the goal in a single trial; referred to as the ‘time needed for goal-reaching’ below) averaged over 500 trials, with the rate of the value-decay (referred to as the ‘decay rate’ below) varied. As shown in the figure, the time needed for goal-reaching is minimized in the case with a certain degree of value-decay. In other words, introduction of the value-decay can facilitate fast goal-reaching. Fig 2B shows the trial-by-trial change of the time needed for goal-reaching. Without the value-decay (Fig 2B, left), the subject initially learns to reach the goal quickly, but subsequently a significant slowdown occurs. In contrast, with the value-decay (Fig 2B, middle and right), the time needed for goal-reaching is kept small, never showing slowdown. The observed facilitation of fast goal-reaching by introduction of the value-decay (Fig 2A) is thus accompanied with such a qualitative change in the long-term dynamics.
In the same simulated task using the same model, we examined how post-training blockade of DA signaling affects the subject's speed (i.e., the time needed for goal-reaching), again varying the decay rate. Specifically, with the assumption that DA represents RPE, we simulated the post-training DA blockade by reducing the size of RPE-dependent increment of action values to zero (complete blockade) or to a quarter of the original size (partial blockade) after 250 trials were completed. Fig 2C shows the results. As shown in the left panels of Fig 2C, without the value-decay, DA blockade causes little effect on the subject's speed. In contrast, in the case with the value-decay (Fig 2C, middle and right panels), the same DA blockade rapidly causes pronounced slowdown (i.e., increase in the time needed for goal-reaching).
In order to explore mechanisms underlying the fast goal-reaching achieved with the value-decay and its impairment by DA blockade, we examined the action values of ‘Go’ and ‘Stay’ at each state. The black and gray lines in Fig 3A respectively show the action values of ‘Go’ and ‘Stay’ at the end of the 500th trial, and Fig 3B shows their trial-by-trial evolutions. Without the value-decay (left panels of Fig 3A and 3B), all the action values are eventually almost saturated to the reward amount (= 1), so that there remains little difference between the action values of ‘Stay’ and ‘Go’ at any states. As a result, subject should choose ‘Stay’ as frequently as ‘Go’. This explains the observed slowdown in the case without the value-decay (Fig 2B, left panel). In contrast, with the value-decay (Fig 3A and 3B, middle and right panels), the action values of ‘Go’ shape a sustained gradient from the start to the goal, while the actions values of ‘Stay’ remain relatively small.
Why does the value-decay create such a gradient of ‘Go’ values? Fig 3C shows examples of RPE generated during the task. In the case without the value-decay (left panel), positive RPE is generated at the beginning of each trial, but RPE is mostly nearly zero in other epochs. This is what we usually expect from TD RL models after learning [4, 25]. On the contrary, in the case with the value-decay (Fig 3C, middle and right panels), RPE remains to be positive in most of the time, indicating that decrement of action values due to the value-decay is balanced with RPE-dependent increment. Such sustained positive RPE is then considered to create the start-to-goal gradient of ‘Go’ values. This is because RPE generated when taking ‘Go’ at state Si (i = 1, …, 6) is calculated (see the Materials and Methods) as
RPE=γ⋅max{Q(‘Stay’ at Si+1), Q(‘Go’ at Si+1)}−Q(‘Go’ at Si),
(γ: time discount factor, satisfying 0 ≤ γ ≤ 1)
which is not greater than Q(‘Go’ at Si + 1) − Q(‘Go’ at Si) provided Q(‘Stay’) ≤ Q(‘Go’) (this would naturally be expected), and then "0 < RPE" ensures
0<Q(‘Go’ at Si+1)−Q(‘Go’ at Si)⇔Q(‘Go’ at Si)<Q(‘Go’ at Si+1),
which indicates a gradient towards the goal.
Looking at the pattern of RPE (Fig 3C), in the case with a relatively larger value-decay, RPE exhibits a ramp towards the goal (Fig 3C, right; notably, this decay rate does not achieve the fastest goal-reaching, but still realizes a faster goal-reaching than the case without value-decay: cf. Fig 2A). This resembles the experimentally observed ramp-like patterns of DA neuronal activity [21, 22] or striatal DA concentration [5–8] as we have previously suggested using the non-self-paced model [23]. But with a milder value-decay, RPE peaks both at the start and towards the goal, with the former more prominent (Fig 3C, middle). In this way, our model generates various patterns of RPE, from phasic to ramping, depending on the decay rate, or indeed the relative strength of the value-decay to the number of states. This could potentially be in line with the fact that the studies reporting DA ramping [5–8, 21, 22] used operant or navigation tasks in which several different states within a trial seem likely to be defined whereas the studies reporting clearly phasic DA response [1, 3] used a simple classical conditioning task where a smaller number of states might be defined.
It has been also found in other studies [5, 8] that elevations in DA levels occurred earlier in later task sessions. According to our simulation results (Fig 3C), such a change could potentially be explained in our model if the decay rate gradually decreases (i.e., from the right panel of Fig 3C to the middle panel). In our simulations, such a decrease in the decay rate is in the direction towards an optimal decay rate in terms of the time needed for goal-reaching averaged over 500 trials (Fig 2A). This suggests that the experimentally observed changes in the DA response pattern across sessions [5, 8] might be an indicative of meta-learning processes to adjust the decay rate to an optimal level. Despite these potentially successful explanations of the various DA response patterns, however, not all the patterns can be explained by our model. In particular, it has been shown that the DA concentration decreases during the reward delivery (sucrose infusion for 6 sec) [2]. Our model does not explain such a decrease of DA: to explain this, it would be necessary to extend the model to describe the actual process of reward delivery/consumption.
The reason why the blockade of DA signaling causes slowdown in the cases with the value-decay but not in the cases without the value-decay in our model (Fig 2C) can also be understood by looking at RPE. Specifically, in the cases with the value-decay, positive RPE is continued to be generated at every state (Fig 3C, middle and right), and each ‘Go’ value is kept around a certain value (Fig 3B, middle and right) because increment according to RPE and decrement due to the value-decay are balanced. Then, if DA signaling is blocked and the size of RPE-dependent increment is reduced, such a balance is perturbed and thereby ‘Go’ values decrease, resulting in the slowdown. In contrast, in the cases without the value-decay, sustained positive RPE is generated only at the beginning of each trial (Fig 3C, left), and it does not increase the value of ‘Go’ taken later in the trial. Thus, after learning has settled down, ‘Go’ values are almost frozen, and therefore blockade of DA signaling has little impact on subject behavior.
Fig 4 shows the trial-by-trial changes of the action values (the top panels of Fig 4A and 4B) and the action values at the end of the 500th trial (the bottom panels) in the simulations where the size of RPE-dependent increment of action values was reduced to zero (A) or to a quarter of the original size (B) after 250 trials were completed. As shown in these figures, the abovementioned conjectures about the effects of DA blockade on the action values were confirmed. Given that the action values are represented in the striatal neural activity, the parallel reduction in the action values and the speed for goal-reaching by DA blockade in our model can be broadly in line with a recent finding of the parallel impairment of the striatal neural representation of actions and the action vigor in DA-depleted mice [26].
Also, intriguingly, in the cases with the value-decay, after DA signaling is reduced to a quarter of the original (Fig 4B, middle and right panels), whereas the values of ‘Go’ actions distant from the goal degrade quite prominently, the values of ‘Go’ actions near the goal (i.e., A12 and A10) remain relatively large, although they are also significantly decreased from the original values. This could potentially be in line with the experimental observations that DA blockade severely impairs costly or effortful actions to obtain rewards but seeking of easily obtainable rewards are largely spared [11, 24]. In order to more directly address this issue, we simulated an experiment examining the effects of DA depletion in the nucleus accumbens in a cost-benefit decision making task in a T-maze reported in [24].
In one condition of the experiment, there was small reward in one of the two arms of the T-maze whereas there was large reward accompanied with a high cost (physical barrier) in the other arm. In the baseline period after training (exploration) of the maze, rats preferred the high-cost-high-return arm. However, DA depletion reversed the preference so that the rats switched to prefer the low-cost-low-return arm. DA depletion also increased the response latency (opening of the start door at the end of the start arm), although the latency subsequently recovered. In another condition of the experiment, the two arms contained small and large rewards as before, but neither was accompanied with a high cost. In this condition, rats preferred the large-reward arm, and DA depletion did not reverse the preference. Meanwhile, DA depletion still increased the response latency, though the latency subsequently recovered as before.
We simulated this experiment by representing a high cost as an extra state preceding the reward (State 5 in Fig 5A, right). Fig 5B and 5C show the ratio of choosing the large-reward arm (Arm 1 in Fig 5A) and the average time needed for reaching the T-junction (State 4 in Fig 5A, right), respectively, in the condition with a high cost in the large-reward arm (Fig 5A). Fig 5F and 5G show the results in the condition without a high cost (Fig 5E). As shown in these figures, the model successfully reproduces the experimental observations that DA depletion induced a preference reversal only in the condition with a high cost (Fig 5B and 5F) while increased the latency in both conditions (Fig 5C and 5G), although the subsequent recovery of the latency is not reproduced. Looking at the action values in the case with a high-cost (Fig 5D), the value of ‘Go’ to Arm 1 at the T-junction is fairly high before DA depletion. However, because this action is apart from reward, its value degrades quite prominently after DA depletion, becoming lower than the value of ‘Go’ to Arm 2, which is adjacent to reward (even though it is small reward). This explains the preference reversal (Fig 5B). In contrast, in the case without a high-cost (Fig 5H), the value of ‘Go’ to Arm 1 degrades only moderately after DA depletion, remaining higher than the value of ‘Go’ to Arm 2. In the meantime, in both conditions, initially there are value-contrasts between ‘Go’ and ‘Stay’ at States 1–3 but they degrade after DA depletion, explaining the increase in the latency (Fig 5C and 5G).
As we have shown above, the value-decay creates a gradient of ‘Go’ values towards the goal. It is known that temporal discounting of rewards also makes a gradient of values (c.f., [7]). However, we assumed no temporal discounting (i.e., time discount factor γ = 1) in the above simulations and thus the value-gradient observed in the above was caused solely by the value-decay. In order to compare the effects of the value-decay and the effects of temporal discounting, we conducted simulations of the original unbranched self-paced task (Fig 1) assuming no value-decay but instead temporal discounting (time discount factor γ = 0.8). Fig 6 shows the resulting action values (Fig 6A and 6B), RPE (Fig 6C), and the effect of DA blockade on the time needed for goal-reaching (Fig 6D). As shown in Fig 6A and 6B, a value-gradient is shaped, as expected. Contrary to the case with the value-decay, however, sustained positive RPE is generated only at the beginning of each trial (Fig 6C), and because of this, post-training blockade of DA signaling causes little effect on the subject speed (Fig 6D).
Comparing the value gradient caused by the value-decay (Fig 3A and 3B, middle/right) and the gradient caused by temporal discounting (Fig 6A and 6B), the differences of the action values between ‘Stay’ and ‘Go’ are much larger in the case with the value-decay. This is considered to be because, in the case with the value-decay, the values of unchosen actions just decay whereas those of chosen actions are kept updated according to RPE. In order to mathematically confirm this conjecture, especially, the long-term stability of such a large contrast between ‘Stay’ and ‘Go’ values, we considered a reduced dynamical system model of our original model, focusing on the last state preceding the goal (i.e., S6 in Fig 1), and conducted bifurcation analysis. Specifically, we derived a two-dimensional dynamical system that approximately describes the dynamics of the action values of A11 (‘Stay’) and A12 (‘Go’) at S6 (Fig 7A; see the Materials and Methods for details), and examined how the system's behavior qualitatively changes along with the change in the degree of the value-decay. Temporal discounting was not assumed (i.e., γ was assumed to be 1) in this reduced model so as to isolate the effect of the value-decay.
Fig 7B is the resulting bifurcation diagram showing the equilibrium action values of A11 (‘Stay’) and A12 (‘Go’) at S6 (with approximations) with the degree of the value-decay varied, and Fig 7C shows the probability of choosing A11 (‘Stay’) and A12 (‘Go’) at the equilibrium point. As shown in Fig 7B, it was revealed that as the degree of the value-decay increases, qualitative changes occur twice (in technical terms, arrangements of the nullclines shown in Fig 7E indicate that both of them are saddle-node bifurcations (c.f., [27])), and when the value-decay is larger than a critical degree (ψ ≈ 0.0559), there exists a unique stable equilibrium with a large contrast between the action values of A11 (‘Stay’) and A12 (‘Go’). It is therefore mathematically confirmed that the value-decay causes a large contrast between the steady-state action values of ‘Stay’ (A11) and ‘Go’ (A12) as conjectured in the above. Similar mechanism is considered to underlie the observed contrasts between ‘Stay’ and ‘Go’ values at the other states (Fig 3A and 3B, middle/right).
Notably, the bifurcation diagram (Fig 7B) suggests that there exists bistability when the degree of the value-decay is within a certain range. We conducted a simulation of the original model with the decay rate φ = 0.0045, and found that there indeed appears a phenomenon indicative of bistability. Specifically, the value of ‘Stay’ (A11) was shown to fluctuate between two levels in long time scales (Fig 7D). Such bistability can potentially cause a hysteresis, in a way that learned values depend on the initial condition or the learning history, although the range of the degree of the value-decay for bistability is not large. Fig 8 shows the dependence of the bifurcation diagram on the RL parameters. As shown in the figure, the existence and the range of bistability critically depend on the inverse temperature (β) (representing the sharpness of soft-max selection) and the time discount factor (γ). The figure also indicates, however, that whether bistability exists or not, as the degree of the value-decay increases, there emerges a prominent contrast between ‘Stay’ and ‘Go’ values.
Importantly, it is considered that the facilitation of fast goal-reaching by the value-decay in the simulations shown so far is actually caused by the value-contrasts between ‘Stay’ and ‘Go’ rather than the gradient of ‘Go’ values explained before, because value-based choice is made between ‘Stay’ and ‘Go’ rather than between successive ‘Go’ actions. Nevertheless, the decay-induced value-gradient can indeed cause a facilitatory effect if selection of ‘Go’ or ‘Stay’ is based on the state values rather than the action values. Specifically, if our model is modified in the way that the probability of choosing ‘Go’ or ‘Stay’ depends on the value of the current and the next state (while action values are not defined: see the Materials and Methods for details), introduction of the decay of learned (state) values can still cause facilitation of goal-reaching (Fig 9A). Since the values of ‘Go’ and ‘Stay’ are not defined and thus the "value-contrast" appeared in the original model does not exist, this facilitation is considered to come from the gradient of state values (Fig 9B). Facilitation appears to be in similar levels as the decay rate changes from 0.01 to 0.02 (Fig 9A), and it is considered to be because, while the slope near the start becomes shallower, the slope near the goal becomes steeper (Fig 9B).
We examined how the effect of the value-decay on fast goal-reaching depends on the RL parameters, specifically, the learning rate, the inverse temperature, and the time discount factor. Fig 10A shows the time needed for goal-reaching averaged over 500 trials in conditions varying one of the RL parameters and the decay rate. As shown in the figure panels, although a large inverse temperature (indicating an exploitative choice policy) realizes fast goal-reaching without the value-decay (middle panel of Fig 10A), facilitation of fast goal-reaching by introduction of the value-decay occurs within a wide range of RL parameters. Notably, the right panel of Fig 10A shows that the value-decay can realize faster goal-reaching than temporal discounting does, given that the other parameters are fixed to the values used here. This is considered to reflect that while both the value-decay and temporal discounting create a value-gradient from the start to the goal, only the value-decay additionally induces value-contrasts between ‘Stay’ and ‘Go’ as we have shown above.
In the results presented so far, we assumed in the model that RPE is calculated according to a major RL algorithm called Q-learning [28] (Eq (1) in the Materials and Methods), based on the empirical suggestions that DA neuronal activity in the rat ventral tegmental area (VTA) and DA concentration in the nucleus accumbens represent Q-learning-type RPE [21, 29]. However, there is in fact also an empirical suggestion that DA neuronal activity represents RPE calculated according to another major RL algorithm called SARSA [30] (Eq (2) in the Materials and Methods) rather than Q-learning in the monkey substantia nigra pars compacta (SNc) [31, 32]. It remains elusive whether such a difference comes from the differences in the species, regions, task paradigms or other conditions. We examined how the model's behavior changes if SARSA-type RPE is assumed instead of Q-learning type RPE. Fig 10B shows the time needed for goal-reaching averaged over 500 trials, with the RL parameters varied as before, and Fig 10C shows the learned values of each action at the end of 500 trials. As shown in the figures, it turned out that the effects of the value-decay, as well as the underlying value-gradient and value-contrast, are very similar to the cases with Q-learning type RPE.
There is, however, a prominent difference between the cases of SARSA and Q-learning. Specifically, in the case of SARSA, RPE generated upon taking ‘Go’ was much larger than RPE generated upon taking ‘Stay’ (Fig 10D, left), whereas there was no such difference in the case of Q-learning (Fig 10D, right). The difference in RPE between ‘Go’ and ‘Stay’ in the SARSA case is considered to reflect the value-contrast between the learned values of ‘Go’ and ‘Stay’ (Fig 10C). This is not the case with Q-learning because the Q-learning-type RPE calculation uses the value of the maximum-valued action candidates, which would be ‘Go’ in most cases, regardless of which action is actually selected. The SARSA-type RPE calculation, by contrast, uses the value of actually selected action (compare Eqs (1) and (2) in the Materials and Methods). The difference in RPE between ‘Go’ and ‘Stay’ in the SARSA case could potentially be related to a recent finding [33] that DA in the rat nucleus accumbens responded to a reward-predicting cue when movement was initiated but not when animal had to stay. However, our present model would be too simple to accurately represent the task used in that study and the neural circuits that are involved, and elaboration of the model is desired in the future.
We examined how the facilitatory effect of the value-decay depends on the amount of the reward obtained at the goal, which was fixed at r = 1 in the simulations so far presented (we again consider Q-learning-type RPE in the following). Fig 11A, 11B, 11C and 11D show the time needed for goal-reaching averaged over 500 trials, with the RL parameters varied as before, in the cases with reward amount 0.5, 0.75, 1.25, and 1.5, respectively. As shown in the figures, the overall tendency of the effect of the value-decay does not largely change across this threefold range of reward amount.
Meanwhile, the figures indicate that as the reward amount increases, the time needed for goal-reaching generally decreases, or in other words, the subject's speed increases. The black line in Fig 11E shows this relationship in the case with the standard RL parameters used so far and the decay rate of 0.01. As shown in this figure, there is a clear negative relationship between the reward amount and the time needed for goal-reaching. We also examined how the average RPE per time-step during 500 trials depends on the reward amount. As shown in the black line in Fig 11F, we found that there is a positive relationship between the reward amount and the average RPE. These negative and positive reward-amount-dependences of the time needed for goal-reaching and the average RPE, respectively, are in line with the experimental findings [7] that the subject's latency and the minute-by-minute DA level in the nucleus accumbens were negatively and positively related with the reward rate, respectively, given that RPE in our model is represented by DA as we assumed.
The commonality of the effect of the value-decay across the range of reward amount (Fig 11A–11D) and the positive reward-amount-dependence of the average RPE (Fig 11F, black line) are considered to appear because our model is largely scalable to (i.e., variables are scaled in proportion to) the changes in the reward amount except for the effect of the inverse temperature. The negative reward-amount-dependence of the time needed for goal-reaching (Fig 11E, black line) is considered to appear because as the reward amount increases, the overall magnitudes of learned values, and thereby also the value-contrasts between ‘Stay’ and ‘Go’, increase.
The gray lines in Fig 11E and 11F show the relationship between the reward amount and the time needed for goal-reaching (Fig 11E) or the RPE per time-step (Fig 11F) in the case without the value-decay, averaged over 500 trials. The gray circles and crosses in these figures show the averages for 1–100 trials and 401–500 trials, respectively. As shown in these, in the case without the value-decay, there are negative and positive reward-amount-dependences of the time needed for goal-reaching and the RPE per time-step in the initial phase, but such dependences gradually degrade along with trials. This is considered to be because the values of ‘Stay’ actions gradually increase toward the saturation (Fig 3B, left). In contrast, in the case with the value-decay (φ = 0.01), there are little differences in the time needed for goal-reaching and the RPE per time-step between 1–100 trials (black circles in Fig 11E and 11F) and 401–500 trials (black crosses in Fig 11E and 11F). This is reasonable given that gradual saturation of ‘Stay’ values does not occur in the case with the value-decay (Fig 3B, middle).
We further examined how the facilitatory effect of the value-decay depends on the architectures of the model, in particular, the number of states and the number of action candidates. Regarding the number of states, in the results so far shown, we assumed seven states, including the start and the goal, as shown in Fig 1. Fig 12A and 12B show the time needed for goal-reaching averaged over 500 trials in the cases with four or ten states, respectively. As shown in the figures, although the optimal decay rate that realizes fastest goal-reaching varies depending on the number of states, facilitation of fast goal-reaching by introduction of the value-decay can occur in either case.
Regarding the number of the action candidates, we have so far assumed that either of the two actions, ‘Go’ or ‘Stay’, can be taken at each state except for the goal (or the T-junction in the case of the T-maze). This can be a good model of certain types of self-paced tasks that are intrinsically unidirectional, such as pressing a lever for a fixed amount of times to get reward. However, there are also self-paced tasks that are more like bidirectional, for instance, movements in an elongated space with reward given at one of the ends. Such tasks might be better represented by adding ‘Back’ action to the action candidates at each state except for the start and the goal. Fig 12C shows the time needed for goal-reaching averaged over 500 trials in the case where the ‘Back’ action was added. As shown in this figure, while the time needed for goal-reaching is generally larger than the cases without the ‘Back’ action as naturally expected, the value-decay can facilitate fast goal-reaching in this case too.
It is also a question of how robust the effect of the value-decay is to perturbations in reward environments. In particular, given that the values of unchosen actions just decay, it is conceivable that, if small reward is given at a state between the start and the goal (e.g., S4: Fig 13A) whenever subject is located there (i.e., repeatedly at every time step if subject stays at S4), subject might learn to stay there persistently rather than to reach the goal. Denoting the size of the small reward by x (< 1, which is the amount of the reward given at the goal), if 7x < x + 1 ⇔ x < 0.166…, such a persistent stay is however inferior to the fastest repetition of goal-reaching in terms of the average reward obtained per time-step. We examined the behavior of modeled subject when small reward is given at S4 with its size x varied from 0 to 0.1, in the case with the value-decay (φ = 0.01). Fig 13B shows the resulting percentage of simulation runs (out of total 20 runs for each condition) in which subject completed 500 trials within 35000 time steps (i.e., within 70 time steps per trial on average) without settling at S4. As shown in the figure, the percentage for the completion of 500 trials is 100% when the size of the reward at S4 is ≤ 0.04, whereas the percentage then decreases as the size of the reward at S4 further increases. This indicates that a persistent stay at S4 actually occurs even if it is not advantageous: Fig 13C and 13D show such an example. Fig 13E shows the number of time steps needed for goal-reaching averaged over 500 trials, only for the simulation runs completing 500 trials in the cases where the completion rate is less than 100%. As shown in the figure, the speed of goal-reaching is kept fast, comparable to the case without reward at S4 (i.e., x = 0). These results indicate that the facilitatory effect of the value-decay on fast goal-reaching has a certain degree of tolerance to this kind of perturbation in reward environments, although it eventually fails as the perturbation becomes larger.
Nonetheless, when temporal discounting (γ = 0.9, 0.8, …) was also assumed in the model with the small reward x = 0.1 at S4, persistent stay at S4 before completing 500 trials was not observed in 20 simulation runs for each of the tested decay rates, and the value-decay could have facilitatory effects (Fig 13F). The absence of persistent stay at S4 is considered to be because the value of ‘Stay’ at S4 is bounded due to temporal discounting. For example, in the case with γ = 0.9 and no value-decay, if the subject keeps staying at S4, the value of ‘Stay’ at S4 converges to 1 (solution of the equation of V: 0 = 0.1 + 0.9V − V). This is still larger than the convergence value of ‘Go’ at S4, which is 0.92 = 0.81. However, since the growth of the ‘Stay’ value from the initial value 0 is likely to be slower than the growth of the ‘Go’ value, subject would rarely begin to settle at S4. In contrast, in the case with no temporal discounting and no value-decay, if the subject keeps staying at S4, the value of ‘Stay’ at S4 increases unboundedly, leading to a persistent stay. Actually, the value-decay also bounds the value of ‘Stay’ at S4, but its effect is weak when the decay rate is small as we have so far assumed. For example, in the case with no temporal discounting, φ = 0.01, and the learning rate α = 0.5, if the subject keeps staying at S4, the value of ‘Stay’ at S4 converges to 4.95 (solution of the equation of V: V = (1 − 0.01)(V + 0.5×0.1)), which is fairly large. In this way, temporal discounting effectively prevents the subject from settling at S4. The value-decay can then facilitate fast goal-reaching by creating the value-contrast between ‘Go’ and ‘Stay’.
So far we have assumed that subject exists in one of the discrete set of states, and selects either ‘Go’ or ‘Stay’, moving to the next state or staying at the same state. Given this simple structure, our model can potentially represent a variety of self-paced behavior, from spatial movement to more abstract Go/No-Go decision sequences. At the same time, however, our model is likely to be too simple to accurately model any specific behavior. In particular, in the case of spatial movement, subject does not really exist only in one of a small number of locations, and would not abruptly stop or literally ‘stay’ at a particular location. Meanwhile, subject should stop or slow down in the face of a physical constraint (e.g., the start, the junction, or the end of a maze) or a salient event (e.g., reward) as observed in experiments [6]. An emerging question is whether our model can be extended to reproduce these observations while preserving its main features.
In order to examine this, we developed an elaborated model of self-paced spatial movement in the T-maze. In this model, the exact one-to-one correspondence between the subject's physical location and the internal state assumed in the original model was changed into a loose coupling, in which each state corresponds to a range of physical locations (Fig 14A). Also, ‘Stay’ action in the original model was replaced with ‘Slow’ action unless there is a physical constraint (i.e., the start, the T-junction, or the end). By selecting ‘Slow’, subject moves straightforward for a time step with the "velocity" halved from the previous time step (or further decreased when there is a physical constraint). ‘Slow’ was introduced to eliminate the abrupt/complete stop appeared in the original model, and mechanistically, it can represent inertia in decision and/or motor processes [34, 35]. With these modifications, state transitions can sometimes occur even when subject chooses ‘Slow’ rather than ‘Go’ (Fig 14A, Case 2), different from the original model. At the T-junction, subject was assumed to take ‘Go’ to either of the two arms or ‘Stay’ in the same manner as in the original model. At the reward location, subject was assumed to take the consummatory action for a time step (indicated by the double-lined arrows in Fig 14B and 14F), and proceed to the end state.
Using this elaborated model (see the Materials and Methods for details), we simulated the T-maze cost-benefit decision making task with DA depletion [24] that was simulated by the original model before (Fig 5). Fig 14C and 14D show the simulation results about the ratio of choosing the large-reward arm (Arm 1) and the average time needed for reaching the T-junction in the task conditions with high cost in the large-reward arm (Fig 14B), respectively. Fig 14G and H show the results in the task conditions without high cost in the large-reward arm (Fig 14F). As shown in the figures, the experimentally observed effects of DA depletion, i.e., the severe impairment of high-cost-high-return choice but not low-cost-high-return choice (Fig 14C and 14G) and the slowdown in both conditions (Fig 14D and 14H), can be reproduced by the elaborated model, as well as by the original model (Fig 5). Simultaneously, the elaborated model can also reproduce the velocity profiles observed in a (different) T-maze task [6], specifically, the slowdown and stop at the T-junction and the end of the maze and the absence of complete stop in the other locations (Fig 14E and 14I). This exemplifies the potential of our original model to be extended to accurately represent specific self-paced behavior.
We have shown that the value-decay in RL can realize sustained fast goal-reaching in a situation requiring self-paced approach towards a goal, modeled as a series of ‘Go’ or ‘No-Go’ (or ‘Stay’) selections. The underlying potential mechanisms turned out to be twofold: (1) a value-gradient towards the goal is shaped by value-decay-induced sustained positive RPE, and (2) value-contrasts between ‘Go’ and ‘Stay’ are generated because chosen values are continually updated whereas unchosen values simply decay. We have then shown that our model with the value-decay can provide potential mechanistic explanations for the key experimental findings that suggest the DA's roles in motivation, under the parsimonious assumption that the representation of RPE is the sole reward-related role of DA. Specifically, our model explains the (i) slowdown of self-paced behavior by post-training blockade of DA signaling [14] (Fig 2C), (ii) severe impairment of effortful actions to obtain rewards, but not of seeking of easily obtainable rewards, by DA blockade [11, 24] (Figs 5 and 14), and (iii) relationships between the reward amount, the level of motivation reflected in the speed of behavior, and the average level of DA [7] (Fig 11E and 11F). Simultaneously, our model also explains the various temporal patterns of DA signals (Fig 3C), confirming and extending the suggestion previously made by the non-self-paced model [23]. Moreover, the simulation results of the SARSA-version of our model could also potentially account for the recent finding [33] that DA ramping occurred when movement was initiated but not when animal had to stay (Fig 10D).
The notion that DA represents RPE has been supported by electrophysiological [1, 4], FSCV [2, 3, 36] and neuroimaging [37–39] results. Recently, optogenetic manipulations of DA neurons causally demonstrated the DA's role in representing RPE [40, 41]. On the other hand, pharmacological blockade of DA signaling has been shown to cause motivational impairments such as slowdown of behavior [14]. Crucially, such effects have been observed even when DA signaling was blocked after animals were well trained and RPE-based learning had presumably already been completed. These motivational effects have thus been difficult to explain by the notion that DA represents RPE, unless different function of DA was also assumed [42, 43].
Given such situations, Niv and colleagues [15] proposed a hypothesis that while DA's phasic response encodes RPE, DA's tonic concentration represents the average reward rate per unit time. They argue that as the reward rate decreases, optimal action speed should also decrease because the opportunity cost for not acting becomes relatively smaller than the extra cost for quickly acting, explaining why DA blockade causes slowdown. Extending this hypothesis, Lloyd and Dayan [16] proposed that quasi-tonic DA represents the expected amount of time discount of the value of next state caused by postponing action to get to the next state. This can explain the experimentally observed ramping DA signals [5–8] as reflecting a gradient of state values created by temporal discounting (as in our Fig 6A and 6B), also consistent with the arguments by [7]. These normative hypotheses, at the Marr's levels of computation and algorithm [44, 45], provide intriguing predictions that are desired to be experimentally tested. Meanwhile, it is also important to explore the Marr's level of implementation, namely, circuit/synaptic operations, which could potentially provide inspirations for the upper levels and vice versa [45]. The abovementioned normative hypotheses highlight essential issues at the circuit/synaptic level, including how the sustained DA signals are generated in the upstream and utilized in the downstream, how the selection of action timing is implemented, and how temporal discounting is implemented.
In our model, sustained DA signals are assumed to represent RPE, and thus the upstream and downstream mechanisms of sustained DA signaling should be nothing more than the mechanisms of how RPE is calculated in the upstream of DA neurons and how RPE-dependent value-update occurs through DA-dependent synaptic plasticity. Both of these mechanisms for RPE have been extensively explored (e.g., [46, 47]) and have now become clarified [17–20]. Regarding the selection of action timing, we assumed that it consists of a series of selections of two actions, ‘Go’ and ‘Stay’. We could thus assume general mechanisms of action selection, for which implementation has been explored [48–52] with empirical supports [50, 53, 54], although this leaves an important issue regarding how time is represented. As for the implementation of temporal discounting, we will discuss it below, in relation to the value-decay that can be implemented as decay of the plastic changes of the synaptic strengths.
There exists a different model that has also tried to give a bottom-up unified explanation of both the learning and motivation roles of DA, referring to circuit architectures of the basal ganglia [55]. However, although this model captures a wide range of phenomena, there are several potential issues or limitations. Firstly, this model assumes that phasic DA represents a simple form of RPE, called the Rescorla-Wagner prediction error [56], which lacks the upcoming-value term. However, RL models of the DA system, including our present model, widely assume the more complex form of RPE called the temporal difference (TD) RPE or TD error [25] (see [57] for detailed explanation) because there is a wealth of empirical supports that DA signals represent TD-RPE [1, 20, 58]. Secondly, because this model assumes the Rescorla-Wagner, rather than TD-, RPE, this model cannot describe the learning of the values of a series of actions or states, nor the changes of RPE, within a trial. As a corollary to this, this model does not explain the experimentally observed sustained DA signals [5–8, 21, 22]. Lastly, this model assumes that the two major basal ganglia pathways, the direct and indirect pathways, are associated with positive and negative reinforcement, respectively. Although this assumption is based on several lines of empirical results, alternative possibilities [43, 46, 47, 59, 60] have also been proposed for the operations of these pathways.
Decay, or forgetting, is apparently wasteful. However, recent work [61] has suggested that decay/forgetting is in fact necessary to maximize future rewards in dynamic environments. Even in a static environment, potential benefit of decay/forgetting has been pointed out [62]. There is also a study [63] that considered decay to explain features of extinction. Forgetting for capturing extinction effects was also assumed in the model that we have discussed right above [55]. However, the authors clearly mentioned that they "assumed some forgetting" "to capture overall extinction effects" and "none of the results are qualitatively dependent on" the parameter for forgetting. Therefore, their work should not have anything to do with the effects of forgetting explored in our present work. Along with these theoretical/modeling works, it has been suggested that RL models with decay could fit the experimental data of human [64–66], monkey [67], and rat [68] choice behavior potentially better than models without decay. Moreover, existence and benefits of decay/forgetting have also been suggested in other types of learning [69, 70].
Nonetheless, decay of learned values (value-decay) is not usually considered in RL model-based accounts of the functions of DA and cortico-basal ganglia circuits. RL models typically have the time discount factor and the inverse temperature (representing choice sharpness) as major parameters [25]. Temporal discounting generates a value-gradient (Fig 6A and 6B) [7, 16], and is suggested [71] to ensure that maximizing rewards simultaneously minimizes deviations from physiologically desirable states. Gradually increasing the inverse temperature, i.e., choice sharpness, is known to be good for global optimization [72]. Possible neural implementation of these parameters have been explored [46, 73–75]. However, it is not sure whether these parameters are actually biologically implemented in their original forms. We have shown that the value-decay can generate a value-gradient, and also value-contrasts which lead to a sharp choice of ‘Go’. Choice-sharpening effect of decay is implied also in previous studies [62, 66]. These indicate a possibility that the value-decay, or its presumed biological substrate, synaptic decay, might in effect partially implement the parameters for temporal discounting and inverse temperature. In this sense, the suggestions that sustained DA represents/reflects time-discounted state values [7, 16] and our value-decay-based account are not necessarily mutually exclusive. Apart from temporal discounting and the inverse temperature, there is an additional note. There have been suggestions [34, 35] that animal's and human's decision making can be affected by the subject's own choice history, which is not included in standard RL models. The value-decay assumed in our model is expected to cause a dependency of decision making on choice history. Whether it can (partly) explain experimentally observed choice patterns would be an interesting issue to explore.
If the rate of the value-decay is always constant, after subject interrupts performing the task for a long period, learned values eventually diminish almost completely. Therefore, in order for our model to be valid, some sort of context-dependence of the value-decay needs to be assumed. There are several empirical implications. At the synaptic level, conditional synaptic decay depending on NMDA receptor-channels [76] or DA (in drosophila) [77] has been found. Behaviorally, memory decay was found to be highly context-dependent in motor learning [78]. More generally, it is widely observed that reactivation of consolidated memories makes them transiently labile [79]. With these in mind, we assume that the value-decay occurs when and only when subject is actively engaged in the relevant task/behavior. However, this issue awaits future verification.
There is also an important limitation of our present model regarding the explanatory power for the experimental observations. Specifically, as mentioned before, our model explains the increase in the latency caused by DA depletion in the cost-benefit decision making task in a T-maze [24], but does not explain the subsequent recovery of the latency. This recovery could possibly be explained if some slow compensatory mechanisms are additionally assumed in the model. It is important in future work to elaborate the model to account for this issue, as well as a diverse array of experimental observations on the DA's roles in motivation that are not dealt with in the present work.
There are also many open issues in the model, both the functional ones and the structural ones. The functional issues include how the states and the time are represented [80, 81] and how ‘Go’ and ‘Stay’ (or ‘No-Go’ or ‘Slow’) are represented. As for the latter, while ‘Go’ and ‘Stay’ might be represented as two distinct actions, ‘Stay’ could instead be represented as disengagement of working-memory/attention as proposed in a recent work [82]. The structural issues include, among others, how different parts of the cortico-basal ganglia circuits and different subpopulations of DA neurons cooperate or divide labor [83–90]. Regarding this, a recent study [91] has shown that DA axons conveying motor signals are largely different from those conveying reward signals and that the motor and reward signals are dominant in the dorsal and ventral striatum, respectively. DA in our model is assumed to represent RPE, and it should thus be released from the axons conveying reward signals that are dense in the ventral striatum. Even with this specification, the structure of our model is still quite simple, and exploring whether and to what extent the present results can be extended to models with rich dynamics at the levels of circuits (in the cortex [48, 50, 92–96], the striatum [97–103], the DAergic nuclei [104], and the entire cortico-basal ganglia system [49, 51, 105–114]), neurons [115, 116], and synapses [117–120] would be important future work.
Our model provides predictions that can be tested by various methods. First, if sustained DA signals indeed represent value-decay-induced sustained RPE, rather than being caused by other reasons [16, 121], the rate of the value-decay estimated from fitting of measured DA signals by our model should match the decay-rate estimated behaviorally. Behavioral estimation of decay-rate would be possible by preparing two choice options that are initially indifferent, manipulating the frequencies of their presentations, and then examining whether, and to what degree, less-frequently-presented option will be chosen less frequently. On the other hand, if sustained DA signals represent time-discounted state values [7, 16], time discount factor estimated from model-fitting of measured DA signals is expected to match behavioral estimation, e.g., from intertemporal choices. Note, however, that the value-decay and temporal discounting might not be completely distinct entities; the value-decay could be a partial implementation of temporal discounting (and the inverse temperature) as we discussed before.
Second, our model predicts that the strengths of cortico-striatal synapses are subject to decay in a context-dependent manner. This could be tested by measuring structural plasticity [18] during learning tasks (across several sessions and intervals). Our model further predicts that manipulations of synaptic decay affect DA dynamics and behavior in specific ways. It has been indicated that a protein kinase that is constitutively active, protein kinase Mζ (PKMζ), is necessary for maintaining various kinds of memories, including drug reward memory in the nucleus accumbens [122]. Specifically, inhibition of PKMζ in the nucleus accumbens core by injecting a selective peptide inhibitor has been shown to impair long-term drug reward memory [122]. It has also been shown that overexpression of PKMζ in the neocortex enhances long-term memory [123]. We predict that overexpression of PKMζ in the nucleus accumbens (ventral striatum) enhances reward memory, or in other words, reduces the value-decay, and thereby diminishes sustained DA signals and impairs goal-approach through the mechanisms described in the present work. Apart from PKMζ, it has also been indicated that DA is required for transforming the early phase of long-term potentiation (LTP), which generally declines, into the late phase of LTP in the hippocampus [124, 125]. Similar DAergic regulation of the stability of LTP could potentially exist in the striatum that is the target of the present work, and if so, the decay rate could be manipulated by DA receptor agonists or antagonists. In the striatal synapses, however, DA signaling would be required for the induction of potentiation before its maintenance, as we have actually assumed in our model. Therefore, it would be necessary to explore ways to specifically manipulate maintenance (decay rate) of potentiation.
The results of the present study suggest that when biological systems for value-learning are active (i.e., when subject is actively engaged in the relevant task/behavior) even though learning has apparently converged, the systems might be in a state of dynamic, rather than static, equilibrium where decay and update are balanced. As we have shown, such dynamic operation can potentially facilitate self-paced goal-reaching behavior, and this effect could be seen as a simple biologically plausible, though partial, implementation of temporal discounting and simulated annealing. It is also tempting to speculate that value-decay-induced sustained RPE might be subjectively felt as sustained motivation, considering recently suggested relationship between RPE and subjective happiness [126, 127]. This is in accordance with the suggestion that DA signals subjective reward value [128, 129], or more precisely, "utility prediction error" [130]. Despite that dynamic operation has these potential advantages, however, there can also be disadvantages. Specifically, continual decay and update of values must be costly, especially given that DA signaling is highly energy-consuming [131]. This could potentially be related to neuropsychiatric and neurological disorders, in particular, Parkinson's disease [131, 132], which is characterized by motor and motivational impairments that are suggested to be independently associated with DA [133]. Better understanding of the dynamic nature of biological value-learning systems will hopefully contribute to clinical strategies against these diseases.
We posited that behavioral task requiring self-paced voluntary approach (whether spatially or not) towards a goal can be represented as a series of ‘Go’ or ‘Stay’ (‘No-Go’) selections as illustrated in Fig 1. Discrete states (S1 ~ S7) and time steps were assumed. In each trial, subject starts from S1. At each time step, subject can take one of two actions, specifically, ‘Go’: moving to the next state or ‘Stay’: staying at the same state. Subject was assumed to learn the value of each action (‘Go’ or ‘Stay’) by a temporal-difference (TD) reinforcement learning (RL) algorithm incorporating the decay of learned values (referred to as the ‘value-decay’ below) [23], and select an action based on their learned values in a soft-max manner [134].
Specifically, at each time step (t), TD reward prediction error (RPE) δ(t) was assumed to be calculated according to the algorithm called Q-learning [28], which has been suggested to be implemented in the cortico-basal ganglia circuit [21, 43, 59], as follows:
δ(t)=R(S(t))+γmaxAcand(t){Q(Acand(t))}−Q(A(t−1)),
(1)
where S(t) represents the state where subject exists at time step t. R(S(t)) represents reward obtained at S(t), which is r (> 0) when S(t) = S7 (goal) and 0 at the other states, unless otherwise described. "Q(A)" generally represents the learned value of action A. Acand(t) represents the candidate of action that can be taken at time step t: when S(t) = Si (i = 1, 2, …, 6), Acand(t) = A2i−1(‘Stay’) or A2i(‘Go’); when S(t) = S7 (goal), candidate of action was not defined and the term γmaxAcand(t){Q(Acand(t))} was replaced with 0. A(t − 1) represents the action taken at time step t − 1; at the beginning of each trial, A(t − 1) was not defined and the term Q(A(t − 1)) was replaced with 0 so as to represent that the beginning of trial is not predictable. γ is the time discount factor (0 ≤ γ ≤ 1). In a separate set of simulations (Fig 10B, 10C and 10D, left), we also examined the case in which TD-RPE is calculated according to another RL algorithm called SARSA [30] as follows:
δ(t)=R(S(t))+γQ(A(t))−Q(A(t−1)),
(2)
where A(t) represents the action taken at time step t.
At each time step other than the beginning of a trial, the learned value of A(t − 1) was assumed to be updated as follows:
Q(A(t−1))new=Q(A(t−1))old+αδ(t),
(3)
where α is the learning rate (0 ≤ α ≤ 1). It was further assumed that the learned value of arbitrary action A decays at every time step as follows:
Q(A)new=(1−φ)Q(A)old,
(4)
where φ (0 ≤ φ ≤ 1) is a parameter referred to as the decay rate: φ = 0 corresponds to the case without value-decay. This sort of value-decay was introduced in [43] to account for the ramp-like activity of DA neurons reported in [21], and was analyzed in [23]. In the present study, the decay rate φ was varied from 0 to 0.02 by 0.002, unless otherwise described. Note that because (1 − φ) is multiplied at every time step, even if φ is very close to 0, significant decay can occur during a trial. For example, when the decay rate φ is 0.01, the action values decline to at least (1–0.01)7 (≈ 0.932)-fold of the original values during a trial. It should also be noted that the value-decay defined as above is fundamentally different from the decay of eligibility trace, which is a popular notion in the RL theory [25]: in terms of the eligibility trace, we assumed that only the value of the immediately preceding action (Q(A(t − 1))) is eligible for RPE-dependent update (Eq (3)), corresponding to the TD(0) algorithm.
At each time step other than when the goal was reached, action ‘Go’ or ‘Stay’ was assumed to be selected according to the following probabilities:
P(AGo)=exp(βQ(AGo))exp(βQ(AGo))+exp(βQ(AStay))
(5)
P(AStay)=exp(βQ(AStay))exp(βQ(AGo))+exp(βQ(AStay)) ,
(6)
where β is a parameter called the inverse temperature, which represents the sharpness of the soft-max selection [134].
A trial ended when subject reached the goal and got the reward. Subsequently the subject was assumed to be (automatically) returned to the start (S1), and the next trial began. The learning rate α, the inverse temperature β, and the time discount factor γ were set to α = 0.5, β = 5, and γ = 1 unless otherwise described. Initial values of all the action values were set to 0. The amount of reward obtained at the goal, r, was set to 1 in most simulations and analyses, but we also examined the cases with r = 0.5, 0.75, 1.25, or 1.5 (Fig 11). The magnitude of rewards can in reality vary even more drastically. However, it has been shown [135] that the gain of DA neuron's response adaptively changes according to actual reward sizes. It could thus be possible to assume that r does not vary too drastically by virtue of such adaptive mechanisms. In a separate set of simulations (Fig 13), in order to examine the robustness of the effect of the value-decay to perturbations in reward environments, we assumed that there is also small reward, with size x, at S4, which is given whenever subject is located at S4 (i.e., repeatedly at every time step if subject stays at S4).
In order to examine the dependence of the effect of the value-decay on the number of states from the start to the goal, we also conducted simulations for models that were modified to have 4 or 10 states, including the start and the goal, instead of 7 states in the original model (Fig 12A and 12B). We also examined the case where the subject is allowed to take not only ‘Go’ or ‘Stay’ but also ‘Back’ action at Si (i = 2, 3, …, 6) (for this, we again assumed 7 states), which causes a backward transition to Si−1. In this case (Fig 12C), selection of ‘Go’, ‘Stay’, and ‘Back’ at Si (i = 2, 3, …, 6) was assumed to be according to the probabilities: P(A*) = exp(βQ(A*))/Sum, where A* was either AGo, AStay, or ABack, and Sum was exp(βQ(AGo)) + exp(βQ(AStay)) + exp(βQ(ABack)). Initial values of all the action values, including the values of ‘Back’ actions, were set to 0.
Further, in a separate set of simulations (Fig 9), we considered a different model in which selection of ‘Go’ or ‘Stay’ is based on the state values rather than the action values (‘Back’ was not considered in this model). Specifically, in this model, RPE is calculated as:
δ(t)=R(S(t))+γV(S(t+1))−V(S(t)),
(7)
where V(S(t)) represents the state value of S(t); if S(t) = S7, V(S(t + 1)) is assumed to be 0. The state values are updated as follows:
V(S(t))new=V(S(t))old+αδ(t).
(8)
The learned value of arbitrary state S was assumed to decay at every time step as follows:
V(S)new=(1−φ)V(S)old.
(9)
‘Go’ is selected at Si (i = 2, 3, …, 6) with the probability exp(βV(Si+1))/{exp(βV(Si)) + exp(βV(Si+1))}, and ‘Stay’ is selected otherwise. The parameters were set to α = 0.5, β = 5, γ = 1, and φ = 0.01, and initial values of all the state values were set to 0.
For each condition with different parameter values or model architectures, 20 simulations of 500 trials with different series of pseudorandom numbers were performed, unless otherwise described. The particular number 500 was chosen because it was considered to be largely in the range of the number of trials used in experiments: e.g., in [6], rats completed ~15 or more sessions with each session containing 40 trials. 20 simulations could be interpreted to represent 20 subjects. In the figures showing the number of time steps needed for goal-reaching, we presented the mean ± standard error (SE) of the 20 simulations except for Fig 13E, where the mean ± SE for the simulation runs completing 500 trials (which could be less than 20 for several conditions) were presented. We also presented the theoretical minimum (in the model with 7 states, it is 7, including the steps at the start and the goal) and the chance level, which is calculated (in the model with 7 states) as:
7+{1⋅h(6,1)⋅12+2⋅h(6,2)⋅(12)2+3⋅h(6,3)⋅(12)3+⋯}⋅(12)6=13,
(10)
where h(6, k) represents the number of ways for a repeated (overlapping) combination of k out of 6 and is calculated as h(6, k) = (k + 5)!(k! · 5!). Simulations were performed using MATLAB (MathWorks Inc.). Program files to run simulations and make figures are available from ModelDB (https://senselab.med.yale.edu/modeldb/showModel.cshtml?model=195890) after the publication of this article.
To simulate post-training blockade of DA signaling, we replaced δ(t) in Eq (3) with 0 (complete blockade) or δ(t)/4 (partial blockade) after 250 trials (Figs 2C, 4 and 6D) or 500 trials (Figs 5 and 14) were completed. δ(t) was non-negative in those simulations because of the structure of the simulated tasks and the assumed Q-leaning-type calculation of RPE, and so the replacement of δ(t) with 0 or δ(t)/4 corresponded to that the size of an increment of action values according to non-negative RPE was reduced to zero or to a quarter of the original size. Notably, at the cellular/synaptic level, DA is known to have two major functions: (i) induce/modulate plasticity of corticostriatal synapses, and (ii) modulate responsiveness of striatal neurons [136]. Function (i) has been suggested to implement RPE-dependent update of learned values (Eq (3)) (e.g., [18]), and in the present work we incorporated the effect of DA blockade on this function into the model as described above, although function (ii) can also affect reaction time and valuation (e.g., [43]) and assuming both of (i) and (ii) might be necessary to account for a wider range of phenomena caused by DA manipulations, in particular, changes in the speed or response time of a single rapid movement (e.g., [137, 138]) rather than (or in addition to) of a series of actions.
In order to obtain qualitative understandings of how the value-decay affects the time evolution and steady-state of action values, beyond observations of simulation results, we reduced the original model (Fig 1) to a simpler model through approximations, and conducted bifurcation analysis. Specifically, we considered a reduced continuous-time dynamical system model that approximately describes the time evolution of the values of ‘Stay’ and ‘Go’ at the state preceding the goal (i.e., A11 (‘Stay’) and A12 (‘Go’) at S6 in Fig 1). The reduced model is as follows:
dq(A11)dt=yαδ˜A11−ψq(A11)
(11)
dq(A12)dt=αδ˜A12−ψq(A12),
(12)
where q(A11) and q(A12) are the continuous-time variables that approximately represent the action values of A11 (‘Stay’) and A12 (‘Go’), respectively. y approximately represents the expected value of the number of repetitions of A11 (‘Stay’) choice (i.e., how many time steps subject chooses A11 (‘Stay’) at S6) in a single trial, and it is calculated as:
y=1⋅p(A11)⋅(1−p(A11))+2⋅p(A11)2⋅(1−p(A11))+⋯=p(A11)1−p(A11),
(13)
where p(A11) represents the probability that A11 is chosen out of A11 and A12 according to Eq (6) when the values of A11 and A12 are q(A11) and q(A12), respectively:
p(A11)=exp(βq(A11))exp(βq(A11))+exp(βq(A12)),
(14)
and substituting Eq (14) into Eq (13) results in:
y=exp(β(q(A11)−q(A12))).
(15)
δ˜A11 and δ˜A12 represent TD-RPE generated when A11 or A12 with the value q(A11) or q(A12) is chosen, respectively:
δ˜A11=γmax{q(A11),q(A12)}−q(A11)
(16)
δ˜A12=r−q(A12),
(17)
where r is the reward amount (= 1). ψ is a parameter representing the degree of the value-decay in a trial, which roughly corresponds to the decay rate φ in the original model multiplied by the number of time steps needed for goal-reaching. Notably, the reduced model is a continuous-time approximation of an algorithm in which update and decay of learned values occur once per every trial in a batch-wise manner whereas the original model is described as an online algorithm where update and value-decay occur at every time step; this difference is contained in our expression "approximate" referring to the reduced model. We analyzed the two-dimensional dynamics of q(A11) and q(A12) (Eqs (11) and (12)) under the assumption that q(A11) ≤ q(A12) (i.e., max{q(A11), q(A12)} = q(A12) in Eq (16)). More specifically, we numerically solved the equations dq(A11)dt=0 and dq(A12)dt=0 to draw the nullclines (Fig 7E), and also numerically found the equilibriums and examined their stabilities to draw the bifurcation diagram (Fig 7B) and calculate p(A11) and p(A12) (Fig 7C) by using MATLAB. The result of the bifurcation analysis in the case with α = 0.5, β = 5, and γ = 1 (Fig 7B) was further confirmed by using XPP-Aut (http://www.math.pitt.edu/~bard/xpp/xpp.html).
We simulated an experiment examining the effects of DA depletion in the nucleus accumbens in a T-maze task reported in [24]. There were two conditions in the task. In the first condition, there was small reward in one of the two arms of the T-maze whereas there was large reward accompanied with high cost (physical barrier preceding the reward) in the other arm. In the second condition, the two arms contained small and large rewards as before, but neither was accompanied with high cost. We simulated this experiment by representing the high cost as an extra state preceding the reward. Specifically, we assumed a state-action diagram as shown in Fig 5A and 5E (right panels). There were two action candidates, ‘Go’ and ‘Stay’, at every state, except for the state at the T-junction (State 4) and the state at the trial end, which was reached if ‘Go’ was chosen at State 7 or 8. In State 4, there were three action candidates, ‘Choose, and Go to, one of the arm (Arm 1)’, ‘Choose, and Go to, the other arm (Arm 2)’, and ‘Stay’. In the state at the trial end (State 9, which is not depicted in Fig 5A and 5E), there was no action candidate, and subject was assumed to be automatically moved to the start state (State 1) at the next time step. In the first condition of the simulated task (Fig 5A), small reward (size 0.5) was given when subject reached State 6 for the first time (i.e., only once in a trial), whereas large reward (size 1) was given when subject reached State 7 for the first time. One extra state, i.e., State 5, preceding the state associated with large reward (State 7) was assumed to represent high cost accompanied with the large reward. In the second condition (Fig 5E), small (size 0.5) or large (size 1) reward was given when subject reached State 6 or State 5, respectively, for the first time, representing that neither reward was accompanied with high cost. Calculation of Q-learning-type RPE and RPE-dependent update of action values were assumed in the same manner as before, with the parameters α = 0.5, β = 5, and γ = 1. The value-decay was also assumed similarly, with the decay rate φ = 0.01. Initial values of all the action values were set to 0. 20 simulations of 1000 trials were conducted for each condition, and post-training DA depletion was simulated in such a way that the size of RPE-dependent increment of action values was reduced to a quarter of the original size after 500 trials were completed.
By modifying the original model described above, we developed an elaborated model of self-paced spatial movement, and simulated the cost-benefit decision making task in a T-maze mentioned above. In this elaborated model, the exact one-to-one correspondence between the subject's physical location and the internal state assumed in the original model was changed into a loose coupling, in which each state corresponds to a range of physical locations (Fig 14A). Also, ‘Stay’ action in the original model was replaced with ‘Slow’ action unless there is a physical constraint (i.e., the start, the T-junction, or the end). Specifically, it was assumed that, at each time step t, subject at a given location chooses either ‘Go’ or ‘Slow’, except that the subject is at the start, T-junction, or the reward location (in the ends of the T-maze). By selecting ‘Go’, subject moves straightforward for a time step with the "velocity" 1, meaning that the subject's physical location is displaced by 1, or moves to the T-junction or the reward location when it is within 1 from the current location. By selecting ‘Slow’, subject moves straightforward for a time step with the "velocity" halved, meaning that the subject's physical location is displaced by the half of the displacement during the previous time interval (between t − 1 and t), or moves to the T-junction or the reward location when it is within the calculated displacement from the current location. In these ways, the "velocity" in this model was defined as the displacement in a time step. At the start (State 1), subject was assumed to take ‘Go’ or ‘Stay’ as in the original model (because at the start, the previous "velocity" was not defined). At the T-junction, subject was assumed to take ‘Choose, and Go to, one of the arm (Arm 1)’, ‘Choose, and Go to, the other arm (Arm 2)’, or ‘Stay’. By selecting ‘Choose, and Go to, Arm 1 or 2’, the subject's physical location is displaced by 1 on the selected arm. By selecting ‘Stay’, subject stays at the same place (T-junction). At the reward location, subject was assumed to take the consummatory action for a time step (indicated by the double-lined arrows in Fig 14B and 14F), and proceed to the end state. Calculation of Q-learning-type TD-RPE, update of action values, and the value-decay were assumed in the same manner as in the original model.
|
10.1371/journal.pntd.0004726 | Transmission Dynamics of Zika Virus in Island Populations: A Modelling Analysis of the 2013–14 French Polynesia Outbreak | Between October 2013 and April 2014, more than 30,000 cases of Zika virus (ZIKV) disease were estimated to have attended healthcare facilities in French Polynesia. ZIKV has also been reported in Africa and Asia, and in 2015 the virus spread to South America and the Caribbean. Infection with ZIKV has been associated with neurological complications including Guillain-Barré Syndrome (GBS) and microcephaly, which led the World Health Organization to declare a Public Health Emergency of International Concern in February 2015. To better understand the transmission dynamics of ZIKV, we used a mathematical model to examine the 2013–14 outbreak on the six major archipelagos of French Polynesia. Our median estimates for the basic reproduction number ranged from 2.6–4.8, with an estimated 11.5% (95% CI: 7.32–17.9%) of total infections reported. As a result, we estimated that 94% (95% CI: 91–97%) of the total population of the six archipelagos were infected during the outbreak. Based on the demography of French Polynesia, our results imply that if ZIKV infection provides complete protection against future infection, it would take 12–20 years before there are a sufficient number of susceptible individuals for ZIKV to re-emerge, which is on the same timescale as the circulation of dengue virus serotypes in the region. Our analysis suggests that ZIKV may exhibit similar dynamics to dengue virus in island populations, with transmission characterized by large, sporadic outbreaks with a high proportion of asymptomatic or unreported cases.
| Since the first reported major outbreak of Zika virus disease in Micronesia in 2007, the virus has caused outbreaks throughout the Pacific and South America. Transmitted by the Aedes genus of mosquitoes, the virus has been linked to possible neurological complications including Guillain-Barré Syndrome and microcephaly. To improve our understanding of the transmission dynamics of Zika virus in island populations, we analysed the 2013–14 outbreak on the six major archipelagos of French Polynesia. We found evidence that Zika virus infected the majority of the population, but only around 12% of total infections on the archipelagos were reported as cases. If infection with Zika virus generates lifelong immunity, we estimate that it would take at least 12–20 years before there are enough susceptible people for the virus to re-emerge. Our results suggest that Zika virus could exhibit similar dynamics to dengue virus in the Pacific, producing large but sporadic outbreaks in small island populations.
| Originally identified in Africa [1], the first large reported outbreak of Zika virus (ZIKV) disease occurred in Yap, Micronesia during April–July 2007 [2], followed by an outbreak in French Polynesia between October 2013 and April 2014 [3], and cases in other Pacific countries [4, 5]. During 2015, local transmission was also reported in South American countries, including Brazil [6, 7] and Colombia [8].
Transmission of ZIKV is predominantly vector-borne, but can also occur via sexual contact and blood transfusions [9]. The virus is spread by the Aedes genus of mosquito [10], which is also the vector for dengue virus (DENV). ZIKV is therefore likely to be capable of sustained transmission in other tropical areas [11]. As well as causing symptoms such as fever and rash, ZIKV infection has also been linked to increased incidence of neurological sequelae, including Guillain-Barré Syndrome (GBS) [12, 13] and microcephaly in infants born to mothers who were infected with ZIKV during pregnancy [14]. On 1st February 2015, the World Health Organization declared a Public Health Emergency of International Concern in response to the clusters of microcephaly and other neurological disorders reported in Brazil, possibly linked to the recent rise in ZIKV incidence. The same phenomena were observed in French Polynesia, with 42 GBS cases reported during the outbreak [13, 15]. In addition to the GBS cluster, there were 18 fetal or newborn cases with unusual and severe neurological features reported between March 2014 and May 2015 in French Polynesia [16], including 8 cases with microcephaly [17].
Given the potential for ZIKV to spread globally, it is crucial to characterize the transmission dynamics of the infection. This includes estimates of key epidemiological parameters, such as the basic reproduction number, R0, defined as the average number of secondary cases generated by a typical infectious individual in a fully susceptible population, and how many individuals (including both symptomatic and asymptomatic) are typically infected during an outbreak. Such estimates could help assist with outbreak planning, assessment of potential countermeasures, and the design of studies to investigate putative associations between ZIKV infection and other conditions.
Islands can be useful case studies for outbreak analysis. Small, centralized populations are less likely to sustain endemic transmission than a large, heterogeneous population [18], which means outbreaks are typically self-limiting after introduction from external sources [19]. Further, if individuals are immunologically naive to a particular pathogen, it is not necessary to consider the potential effect of pre-existing immunity on transmission dynamics [20]. Using a mathematical model of vector-borne infection, we examined the transmission dynamics of ZIKV on six archipelagos in French Polynesia during the 2013–14 outbreak. We inferred the basic reproduction number and the overall size of the outbreak, and hence how many individuals would still be susceptible to infection in coming years.
We used weekly reported numbers of suspected ZIKV infections from the six main regions of French Polynesia between 11th October 2013 and 28th March 2014 (Table 1), as detailed in the Centre d’hygiène et de salubrité publique situation reports [21, 22]. Confirmed and suspected cases were reported from sentinel surveillance sites across the country; the number of such sentinel sites varied in number from 27–55 during the outbreak (raw data are provided in S1 Dataset). Clinical cases were defined as suspected cases if they presented to health practitioners with rash and/or mild fever and at least two of the following signs: conjunctivitis, arthralgia, or oedema. Suspected cases were defined as a confirmed case if they tested positive by RT-PCR on blood or saliva. In total, 8,744 suspected cases were reported from the sentinel sites. As there were 162 healthcare sites across all six regions, it has been estimated that around 30,000 suspected cases attended health facilities in total [21]. For each region, we calculated the proportion of total sites that acted as sentinels, to allow us to adjust for variation in reporting over time in the analysis. Population size data were taken from the 2012 French Polynesia Census [23]. In our analysis, the first week with at least one reported case was used as the first observation date.
We used a compartmental mathematical model to simulate vector-borne transmission [24, 25]. Both people and mosquitoes were modelled using a susceptible-exposed-infectious-removed (SEIR) framework. This model incorporated delays as a result of the intrinsic (human) and extrinsic (vector) incubation periods (Fig 1). Since there is evidence that asymptomatic DENV-infected individuals are capable of transmitting DENV to mosquitoes [26], we assumed the same for ZIKV: all people in the model transmitted the same, regardless of whether they displayed symptoms or were reported as cases.
The main vectors for ZIKV in French Polynesia are thought to be Ae. aegypti and Ae. polynesiensis [12]. In the southern islands, the extrinsic incubation period for Ae. polynesiensis is longer during the cooler period from May to September [27], which may act to reduce transmission. Moreover, temperature can also influence mosquito mortality, and hence vector infectious period [28]. However, climate data from French Polynesia [29] indicated that the ZIKV outbreaks on the six archipelagos ended before a decline in mean temperature or rainfall occurred (S1 Fig). Hence it is likely that transmission ceased as a result of depletion of susceptible humans rather than seasonal changes in vector transmission. Therefore we did not include seasonal effects in our analysis.
In the model, SH represents the number of susceptible people, EH is the number of people currently in their incubation period, IH is the number of infectious people, RH is the number of people that have recovered, C denotes the cumulative number of people infected (used to fit the model), and N is the human population size. Similarly, SV represents the proportion of mosquitoes currently susceptible, EV the proportion in their incubation period, and IV the proportion of mosquitoes currently infectious. As the mean human lifespan is much longer than the outbreak duration, we omitted human births and deaths. The full model is as follows:
d S H / d t = - β H S H I V (1) d E H / d t = β H S H I V - α E H (2) d I H / d t = α H E H - γ I H (3) d R H / d t = γ I H (4) d C / d t = α H E H (5) d S V / d t = δ - β V S V I H N - δ S V (6) d E V / d t = β V S V I H N - ( δ + α V ) E V (7) d I V / d t = α V E V - δ I V (8)
Parameter definitions and values are given in Table 2. We used weakly informative prior distributions for the human latent period, 1/αH, infectious period, 1/γ, extrinsic latent period, 1/αv, and mosquito lifespan, 1/μ. For these prior distributions, we made the assumption that human latent period was equivalent to the intrinsic incubation period, i.e. that no transmission typically occurs before symptom onset. A systematic review of the incubation period for ZIKV in humans estimated a mean value of 5.9 days [30]; the infectious period, 1/γ, lasted for 4–7 days in clinical descriptions of 297 PCR-confirmed cases in French Polynesia [22]; the extrinsic latent period has been estimated at 10.5 days [1]; and mosquito lifespan in Tahiti was estimated at 7.8 days [31]. We therefore used these values for the respective means of 1/αH, 1/γ, 1/αv and 1/δ in our prior distributions. These parameters were estimated jointly across all six regions; as mentioned above, we assumed that the parameters remained fixed over time, as temperature and rainfall levels did not change substantially during the outbreak. The rest of the parameters were estimated for each region individually; we assumed uniform prior distributions for these.
Serological analysis of samples from blood donors between July 2011 and October 2013 suggested that only 0.8% of the population of French Polynesia were seropositive to ZIKV [33]; we therefore assumed that the population was fully susceptible initially. We also assumed that the initial number of latent and infectious people were equal (i.e. E 0 H = I 0 H), and the same for mosquitoes (E 0 V = I 0 V). The basic reproduction number was equal to the product of the average number of mosquitoes infected by the typical infectious human, and vice versa [24]:
R 0 = β V γ × α V δ + α V β H δ . (9)
We fitted the model using Markov chain Monte Carlo (MCMC), where incidence in week t, denoted ct, was the difference in the cumulative proportion of cases over the previous week i.e. ct = C(t) − C(t − 1). In the model, the total number of cases included asymptomatic and subclinical cases—which would not be detected at any site—as well as those that displayed symptoms. Hence there were two sources of potential underreporting: as a result of limited sentinel sites; and as a result of cases not seeking treatment. We adjusted for the first source of underreporting by defining κt as the proportion of total sites that reported as sentinels in week t. We assumed that the population was uniformly distributed across the catchment areas of the healthcare sites. Under this assumption, the proportion of total sites that reported cases as sentinels in a particular week, κt, was equivalent to the expected fraction of new cases that would be reported in that week if the reporting proportion, r, was equal to 1. The parameter r accounted for the second source of under-reporting, and represented the proportion of cases (both symptomatic and asymptomatic) that did not seek treatment.
To calculate the likelihood of observing a particular number of cases in week t, yt, we assumed the number of confirmed and suspected cases in week t followed a negative binomial distribution with mean rκt ct and dispersion parameter ϕ, to account for potential variability in reporting over time [34]. The dispersion parameter reflected variation in the overall proportion reported, as well as potential variation in size and catchment area of sentinel sites. Hence the log-likelihood for parameter set θ given data Y = { y t } t = 1 T was L(θ|Y) = ∑t log P(yt|ct). As a sensitivity analysis (see Results), we also extended the model so the likelihood included the probability of observing 314/476 seropositive individuals in Tahiti after the outbreak, given that a proportion Z were infected in the model. Hence for Tahiti, L(θ|Y) = ∑t log P(yt|ct) + log P(X = 314), where X ∼ B(n = 476, p = Z). The joint posterior distribution of the parameters was obtained from eight replicates of 25,000 MCMC iterations, each with a burn-in period of 5,000 iterations (S2–S8 Figs). The model was implemented in R version 3.2.3 [35] using the deSolve package [36].
We implemented a simple demographic model to examine the replacement of the number of susceptible individuals over time. In 2014, French Polynesia had a birth rate of b = 15.47 births/1,000 population, a death rate of d = 4.93 deaths/1,000 population, and net migration rate of m = −0.87 migrants/1,000 [37]. The number of susceptible individuals in year τ, S(τ), and total population size, N(τ), was therefore expressed as the following discrete process:
N ( τ ) = N ( τ - 1 ) + b N ( τ - 1 ) - d N ( τ - 1 ) - m N ( τ - 1 ) (10) S ( τ ) = S ( τ - 1 ) + b N ( τ - 1 ) - d S ( τ - 1 ) - m S ( τ - 1 ) (11)
We set S(2014) as the fraction of the population remaining in the S compartment at the end of the 2013–14 ZIKV outbreak, and propagated the model forward to estimate susceptibility in future years. The effective reproduction number, Reff(τ), in year τ was the product of the estimated basic reproduction number, and the proportion of the population susceptible: Reff(τ) = R0S(τ). We sampled 5,000 values from the estimated joint posterior distributions of S(2014) and R0 to obtain the median and credible intervals.
Across the six regions, estimates for the basic reproduction number, R0, ranged from 2.6 (95% CI: 1.7–5.3) in Marquises to 4.8 (95% CI: 3.2–8.4) in Moorea (Table 3). Our results suggest that only a small proportion of ZIKV infections were reported as suspected cases: sampling from the fitted negative binomial reporting distributions for each region implied that 11.5% (95% CI: 7.32–17.9%) of infections were reported overall. Estimated dispersion in reporting was greatest for Marquises (S1 Table), reflecting the variability in the observed data (Fig 2), even after adjusting for variation in the number of sentinel sites. Dividing the 8,744 cases reported at sentinel sites by the total estimated infections, we also estimated that 3.41% (95% CI: 3.32–3.55%) of total infections were reported at the subset of health sites that acted as sentinel sites.
Our posterior estimates for the latent and infectious periods in humans and mosquitoes were consistent with the assumed prior distributions (S2 Fig), suggesting either that there was no strong evidence that these parameters had a different distribution, or that the model had limited ability to identify these parameters from the available data. As a sensitivity analysis, we therefore considered two alternative prior distributions for the incubation and infectious periods for humans and mosquitoes. First, we examined a broader prior distribution. We used the same mean values for the Gamma distributions specified in Table 2, but with σ = 2. These priors produced similar estimates for R0, proportion reported, and total number of infections (S2 Table), although the estimated parameters for humans were further from zero than in the prior distribution (S9 Fig).
As a second sensitivity analysis, we used prior distributions with mean values as given in studies of dengue fever, and σ = 0.5. As there is evidence that human-to-mosquito transmission can occur up to 2 days before symptom onset [38], and the intrinsic incubation period for DENV infection is 5.9 days [39], we assumed a mean latent period of 5.9–2 = 3.9 days. We also assumed an infectious period of 5 days [38]; an extrinsic latent period of 15 days [39]; and a longer mosquito lifespan of 14 days [28]. Again, these assumptions produced similar estimates for key epidemiological parameters (S3 Table), with posterior estimates for incubation and infectious periods tracking the prior distributions (S10 Fig).
The estimated proportion of the population that were infected during the outbreak (including both reported and unreported cases) was above 85% for all six regions (Table 3), and we estimated that 94% (95% CI: 91–97%) of the total population were infected during the outbreak. A serological survey following the French Polynesia ZIKV outbreak found 314/476 children aged 6–16 years in Tahiti were positive for ZIKV in an indirect ELISA test for IgG antibody, corresponding to an attack rate of 66% (95% CI: 62–70)[17]. To test whether this seroprevalence data could provide additional information about the model parameters, we extended the model to calculate the likelihood of observing 314/476 seropositive individuals in Tahiti after the outbreak, as well as the observed weekly case reports. We obtained a much lower R0 estimate for Tahiti, but similar results for other regions, and the median reporting rate remained unchanged for all areas (S4 Table). However, the model was unable to reproduce the Tahiti epidemic curve when the overall attack rate was constrained to be consistent with the results of the serological survey (S11 Fig).
During the 2013–14 outbreak in French Polynesia, there were 42 reported cases of GBS [13]. This corresponds to an incidence rate of 15.3 (95% binomial CI: 11.0–20.7) cases per 100,000 population, whereas the established annual rate for GBS is 1–2 cases per 100,000 [10]. In total, there were 8,744 confirmed and suspected ZIKV cases reported at sentinel sites in French Polynesia, which gives an incidence rate of 480 (95% CI: 346–648) GBS cases per 100,000 suspected Zika cases reported at these sites. However, when we calculated the GBS incidence rate per estimated total ZIKV cases, using the model estimates based on the prior distributions in Table 2, we obtained a rate of 16.4 (95% CI: 11.5–21.4) per 100,000 cases. These credible intervals overlap substantially with the above incidence rate calculated with population size as the denominator, indicating that the two rates are not significantly different.
Using a demographic model we also estimated the potential for ZIKV to cause a future outbreak in French Polynesia. We combined our estimate of the proportion of the population that remained susceptible after the 2013–14 outbreak and R0 with a birth-death-migration model to estimate the effective reproduction number, Reff, of ZIKV in future years. If Reff is greater than one, an epidemic would be possible in that location. Assuming that ZIKV infection confers lifelong immunity against infection with ZIKV, our results suggest that it would likely take 12–20 years for the susceptible pool in French Polynesia to be sufficiently replenished for another outbreak to occur (Fig 3). This is remarkably similar to the characteristic dynamics of DENV in the Pacific island countries and territories, with each of the four DENV serotypes re-emerging in sequence every 12–15 years, likely as a result of the gradual accumulation of susceptible individuals due to births [19, 40].
Using a mathematical model of ZIKV transmission, we analysed the dynamics of infection during the 2013–14 outbreak in French Polynesia. In particular, we estimated key epidemiological parameters, such as the basic reproduction number, R0, and the proportion of infections that were reported. Across the six regions, our median estimates suggest that between 7–17% of infections were reported as suspected cases. This does not necessarily mean that the non-reported cases were asymptomatic; individuals may have had mild symptoms and hence did not enter the healthcare system. For example, although the attack rate for suspected ZIKV disease cases was 2.5% in the 2007 Yap ZIKV outbreak, a household study following the outbreak found that around 19% of individuals who were seropositive to ZIKV had experienced ZIKV disease-like symptoms during the outbreak period [2].
Our median estimates for R0 ranged from 2.6–4.8 across the six main archipelagos of French Polynesia, and as a result the median estimates of the proportion of the populations that became infected in our model spanned 87–97%. This is more than the 66% (95% CI: 62–70%) of individuals who were found to be seropositive to ZIKV in a post-outbreak study in Tahiti [17]. When we constrained the model to reproduce this level of seroprevalence as well as the observed weekly reports, however, we obtained a much poorer fit to the case time series (S11 Fig). The discrepancy may be the result of population structure, which we did not include within each region; we used a homogeneous mixing model, in which all individuals had equal chance of contact. In reality, there will be spatial heterogeneity in transmission [41], potentially leading to a depletion of the susceptible human pool in some areas but not in others. Additionally, there is evidence that Ae. aegypti biting rate can vary between individual human hosts [42]. Whereas in the model everyone in regions with ZIKV infected mosquitoes had equal probability of infection, in reality there is likely to be individual-level heterogeneity in probability of infection, which could alter the proportion who seroconvert to ZIKV after the outbreak. As we used a deterministic model, differences in the estimate for the reporting dispersion parameter for different regions may to some extent reflect the limitations of the model in capturing observed transmission patterns, as well as true variability in reporting.
The ZIKV outbreak in French Polynesia coincided with a significant increase in Guillain-Barré syndrome (GBS) incidence [13]. We found that although there was a raw incidence rate of 480 (95% CI: 346–648) GBS cases per 100,000 suspected ZIKV cases reported, the majority of the population was likely to have been infected during the outbreak, and therefore the rate per infected person was similar to the overall rate per capita. This could have implications for the design of epidemiological studies to examine the association between ZIKV infection and neurological complications in island populations.
If infection with ZIKV confers lifelong immunity, we found it would take at least a decade before re-invasion were possible. In the Pacific island countries and territories, replacement of DENV serotypes occurs every 4–5 years [19, 40], and therefore each specific serotype re-emerges in a 12–15 year cycle. The similarity of this timescale to our results suggest that ZIKV may exhibit very similar dynamics to DENV in island populations, causing infrequent, explosive outbreaks with a high proportion of the population becoming infected. In September 2014, Chikungunya virus (CHIKV) caused a large outbreak in French Polynesia [43], and is another example of a self-limiting arbovirus epidemic in island populations [5]. However, it remains unclear whether ZIKV could become established as an endemic disease in larger populations, as DENV and CHIKV have.
For immunising infectious diseases, there is typically a ‘critical community size’, below which random effects frequently lead to disease extinction, and endemic transmission cannot be sustained [18, 44]. Analysis of dengue fever outbreaks in Peru from 1994–2006 found that in populations of more than 500,000 people, dengue was reported in at least 70% of weekly records [41]. Large cities could have the potential to sustain other arboviruses too, and understanding which factors—from population to climate—influence whether ZIKV transmission can become endemic will be an important topic for future research. We did not consider seasonal variation in transmission as a result of climate factors in our analysis, because all six outbreaks ended before there was a substantial seasonal change in rainfall or temperature. Such changes could influence the extrinsic incubation period and mortality of mosquitoes, and hence disease transmission. If the outbreaks had ended as a result of seasonality, rather than depletion of susceptibles, it would reduce the estimated proportion of the population infected, and shorten the time interval before ZIKV would be expected to re-emerge.
There are some additional limitations to our analysis. As we were only fitting to a single time series for each region, we also assumed prior distributions for the incubation and infectious periods in humans and mosquitoes. Sensitivity analysis on these prior distributions suggested it was not possible to fully identify these parameters from the available data. If seroprevalence data from each region were to become available in the future, it could provide an indication of how many people were infected, which may make it possible to constrain more of the model parameters, and evaluate the role of spatial heterogeneity discussed above. Such studies may require careful interpretation, though, because antibodies may cross-react between different flaviviruses [12].
Our results suggest that ZIKV transmission in island populations may follow similar patterns to DENV, generating large, sporadic outbreaks with a high degree of under-reporting. If a substantial proportion of such populations become infected during an outbreak, it may take several years for the infection to re-emerge in the same location. A high level of infection, combined with rarity of outbreaks, could also make it more challenging to investigate a potential causal link between infection and concurrent neurological complications.
|
10.1371/journal.pntd.0006737 | Spatiotemporal prediction of infectious diseases using structured Gaussian processes with application to Crimean–Congo hemorrhagic fever | Infectious diseases are one of the primary healthcare problems worldwide, leading to millions of deaths annually. To develop effective control and prevention strategies, we need reliable computational tools to understand disease dynamics and to predict future cases. These computational tools can be used by policy makers to make more informed decisions.
In this study, we developed a computational framework based on Gaussian processes to perform spatiotemporal prediction of infectious diseases and exploited the special structure of similarity matrices in our formulation to obtain a very efficient implementation. We then tested our framework on the problem of modeling Crimean–Congo hemorrhagic fever cases between years 2004 and 2015 in Turkey.
We showed that our Gaussian process formulation obtained better results than two frequently used standard machine learning algorithms (i.e., random forests and boosted regression trees) under temporal, spatial, and spatiotemporal prediction scenarios. These results showed that our framework has the potential to make an important contribution to public health policy makers.
| Infectious diseases cause important health problems worldwide and create difficult challenges for public health policy makers. That is why they need reliable computational tools to better understand disease and to predict case counts. They will benefit from such computational tools to make more informed decisions in developing control and prevention strategies. We formulated a computational framework that can be used to model spatial, temporal, or spatiotemporal dynamics of infectious diseases. We showed the utility of our framework on the problem of modeling Crimean–Congo hemorrhagic fever in Turkey.
| Infectious diseases constitute a major part of healthcare burden worldwide, leading to millions of deaths annually, which are especially seen among poor and young populations in low and middle income countries [1]. In addition to pandemic infectious diseases such as influenza and tuberculosis, there are also emerging infectious diseases such as Ebola virus disease and Zika fever, which require a worldwide effort to combat. Thus, predicting the case counts of infectious diseases is of great importance in developing control and prevention strategies. In particular, there might be spatial dependencies (e.g., humid conditions for malaria) and temporal dependencies (e.g., seasonal effects for influenza) that control the emergence and spread of such diseases [2].
To be able to develop protective measures against infectious diseases, it is very important (i) to clearly identify the disease spread and (ii) to make reliable predictions for future cases. When the disease spread is known, policy makers can develop preventive strategies against, for instance, environmental factors that promote the disease. Once we have reliable predictions for future cases, policy makers can make informed decisions on, for example, vaccine purchases, public awareness campaigns and training programs for healthcare workers.
Machine learning algorithms can contribute to the control of infectious diseases by addressing aforementioned two aims. In the literature, standard machine learning algorithms such as random forests [3] and boosted regression trees [4, 5] were frequently used in ecological and epidemiological applications [6–10]. These algorithms have been picked by the applied researchers mainly because they have a relatively simple interface for nonspecialists. However, they might fail to capture highly complex dependencies in disease modeling scenarios. Thus, we used Gaussian processes [11] to be able to identify highly nonlinear dependencies and to make more reliable predictions.
We proposed a computational framework that uses Gaussian processes as the basic building block to perform spatiotemporal prediction of infectious diseases. We first noted that the kernel matrices have a special structure owing to their dependencies on both spatial and temporal covariates and then exploited this special structure to obtain a very efficient inference algorithm. We tested our proposed framework on Turkey’s country-wide surveillance data set of a vector-borne infectious disease Crimean–Congo hemorrhagic fever, which is a widespread endemic infectious disease seen in Africa, the Balkans, the Middle East, and Asia with a case fatality rate of 5–40% [12].
We present the overview of our proposed computational framework with three possible prediction scenarios in Fig 1. We assume that the reported case counts of location and time period pairs have been recorded with additional information about their spatial and temporal properties. We first extract spatial and temporal features for each location and time period, respectively, from these properties. We then calculate two similarity matrices among locations and time periods, respectively, using the extracted features. These two similarity matrices are combined to obtain a larger similarity matrix between location and time period pairs. Using the combined similarity matrix and reported cases counts, we train a Gaussian process regression model to be able to make predictions under three different scenarios: (i) temporal prediction (i.e., predicting case counts for future time periods, leading to predicting disease prevalence for each location in the future), (ii) spatial prediction (i.e., predicting case counts for unseen locations, leading to predicting disease spread within the same time frame in other locations), which can be used to complete missing case counts for the locations that we could not obtain historical data, and (iii) spatiotemporal prediction (i.e., predicting case counts for unseen location and future time period pairs, leading to predicting disease spread to new locations in the future), which is especially important to be able to prepare against emerging infectious diseases since there will be no historical data for the locations that experience the disease for the first time.
In this study, we proposed a computational framework to perform spatiotemporal prediction of infectious diseases. To test this framework, we addressed an important public health problem in Turkey, namely, Crimean–Congo hemorrhagic fever (CCHF), which is a vector-borne infectious disease transmitted by infected tick bites and exposure to blood or bodily fluids of the infected cases.
We used an unpublished surveillance data set of 9,636 CCHF infection cases reported in Turkey between years 2004 and 2015, which was collected by the Ministry of Health of Turkey (S1 File). The reported cases were mainly because of infected tick bites, and they were diagnosed with clinical symptoms such as fever, myalgia, and bleeding from various sites. These infected cases were also confirmed with blood tests.
The Ministry of Health of Turkey provided us with spatial information (province, district, and town names) and temporal information (year and month) for each case, which made this data set suitable for studying spatiotemporal characteristics of CCHF. The data set does not include clinical covariates of infected cases, which forces our study to investigate only spatial and temporal covariates.
Infectious disease spread is usually driven by both location and time, which means nearby locations and time periods have similar characteristics. The disease spreads to adjacent province much more easily than distant provinces due to spatial dependency. Case counts in consecutive time periods or in time periods within the same season are usually heavily correlated due to temporal dependency.
We suggest using Gaussian process regression (GPR), which is suitable to capture highly complex dependencies between input and output variables thanks to its nonlinear nature brought by kernel functions. We propose a computational strategy based on GPR that enables us to perform predictions under spatial (i.e., predicting case counts for unseen locations), temporal (i.e., predicting case counts for future time periods) and spatiotemporal scenarios (i.e., predicting counts for unseen location and future time period pairs) for infectious diseases.
We first give a brief description of GPR. We then show how GPR can be modified for infectious disease modeling by introducing a structured kernel function based on two separate kernel functions over spatial and temporal covariates, respectively, and how this modified GPR formulation can be implemented very efficiently. We describe three different prediction scenarios encountered in spatiotemporal modeling of infectious diseases. We lastly discuss two baseline algorithms from the literature that will be used to benchmark against.
Table 1 reports PCC values of RFR, BRT, and GPR algorithms on our CCHF data set for three prediction scenarios. We see that GPR algorithm obtained the best PCC values by improving the results of temporal, spatial, and spatiotemporal prediction scenarios by 1.05%, 26.31%, and 16.45%, respectively. Note that RFR and BRT algorithms failed to capture the spatial spread of CCHF when predicting case counts for unseen provinces (i.e., in spatial and spatiotemporal scenarios), whereas GPR algorithm was able to capture this spread by obtaining more than 70% PCC for these two scenarios. All algorithms achieved PCC values around 75% and 85% for temporal scenario since capturing temporal dynamics is easier owing to annual periodicity of CCHF cases.
Table 2 shows NRMSE values of RFR, BRT, and GPR algorithms on our CCHF data set for temporal, spatial, and spatiotemporal prediction scenarios. We see that GPR algorithm again obtained the best NRMSE values by improving the results of temporal, spatial, and spatiotemporal prediction scenarios by 21.39%, 20.38% and 15.65%, respectively. Even though BRT algorithm obtained a PCC value comparable to that of GPR algorithm for temporal scenario, GPR algorithm obtained considerably better NRMSE values than both RFR and BRT algorithms. This shows that GPR algorithm is better than the other two algorithms in terms of capturing the range of CCHF cases in the test sets as discussed below.
Fig 5 shows the total observed and predicted case counts by RFR, BRT and GPR algorithms for years 2014 and 2015 over the five provinces with the highest case counts among 40 common test provinces of all scenarios. We see that all three algorithms captured the annual periodicity of CCHF cases, whereas GPR algorithm performed the best in terms of predicting the observed case counts. RFR algorithm was not able to predict the observed case counts owing to its lack of high order interactions between covariates, whereas BRT algorithm performed better owing to its second order interactions. The same results were also valid if we took the first 10, 15, and 20 provinces from 40 common test provinces (S14, S15 and S16 Figs).
S17 Fig gives a detailed comparison between observed and predicted case counts of RFR, BRT, and GPR algorithms for the same five provinces reported in Fig 5. We see that GPR algorithm produced predictions mostly in agreement with the range of observed CCHF case counts, whereas RFR and BRT algorithms underestimated CCHF case counts in most of the time periods. BRT algorithm obtained NRMSE value comparable to that of GPR algorithm for temporal scenario, whereas GPR algorithm reduced NRMSE values by 0.277 and 0.170 for spatial and spatiotemporal scenarios, respectively.
The results of the computational experiments reported in this study can be analyzed from different perspectives. We analyzed the results with respect to prediction scenarios, machine learning algorithms, computational complexity, dependency on training set size, and dependency on sampling over provinces.
We performed computational experiments under three different scenarios. As we can see from Tables 1, 2, Fig 5 and S17 Fig, making temporal predictions (i.e., predicting future time periods by looking at the historical data) is strikingly easier than making spatial and spatiotemporal predictions (i.e., generalizing to unseen locations). Most infectious disease outbreaks occur in cycles (i.e., ascending, plateau, and descending phases), and this structure makes temporal prediction easier. The disease we addressed is a vector-borne infectious disease mainly transmitted by infected tick bites, leading to a strong temporal dependency owing to the sleep cycles of ticks.
We used three machine learning algorithms for predicting case counts. As we discussed before, GPR algorithm was able to capture the range of CCHF case counts better than RFR and BRT algorithms. We think that this was mainly due to the capability of GPR algorithm to model highly complex dependencies between input and output covariates thanks to nonlinear kernel functions such as the Gaussian kernel we used. We also noted from Fig 5 and S17 Fig that the main improvement of GPR algorithm over the others was the ability to better capture the range of case counts in the time periods with nonzero observed case counts. In the literature, RFR and BRT algorithms were frequently used as classification algorithms to predict whether there will be cases. In terms of classification performance, we would not expect major differences between three algorithms.
Instead of using a naive version of GPR algorithm, we implemented an efficient variant that exploits the special structure of the kernel matrix to make inference very fast. We decomposed the kernel matrix into a Kronecker product of two smaller kernel matrices calculated on spatial and temporal covariates, respectively. By doing so, we were able to perform inference for our structured GPR formulation in the order of milliseconds, whereas RFR and BRT algorithms took several minutes to complete using drastically higher physical memory.
To show the dependency of GPR on training set size, we performed an additional set of experiments by changing the number of years used for training. We used CCHF case counts of the last two, four, six, eight, and ten years between 2004 and 2013, respectively. Table 3 shows PCC and NRMSE values of GPR algorithm for this new set of experiments. We can see that there was an increasing trend in predictive performance as we increased the training set size.
Up to this point, we performed our experiments on a fixed training and test set split (S13 Fig), which was designed to make training and test sets as similar as possible, to better illustrate the differences between machine learning algorithms. We also compared the predictive performances of RFR, BRT, and GPR on 100 different training and set set splits constructed by random sampling on 81 provinces. Fig 6 shows PCC and NRMSE values of the algorithms for spatial and spatiotemporal modeling scenarios. We see that our algorithm GPR was statistically significantly better (i.e., p < 0.001) than other two algorithms for both scenarios in terms of PCC values. In spatial prediction scenario, GPR achieved statistically significantly better NRMSE values than RFR (i.e., p = 0.023), but it obtained NRMSE values comparable to BRT (i.e., p = 0.052). In spatiotemporal prediction scenario, NRMSE values of GPR were statistically significantly better than those of BRT (i.e., p < 0.001), whereas NRMSE values were comparable between GPR and RFR (i.e., p = 0.932).
Infectious diseases cause important health problems worldwide and create difficult challenges for public health policy makers. To be able to make correct and effective decisions, it is quite important to understand the characteristics of each infectious disease, which includes environmental factors such as climate and animal population in addition to molecular evolution of disease sources such as bacteria and viruses. In this study, we addressed to capture the effect of environmental factors on infectious diseases by modeling their spatial and temporal dependencies on these factors.
For this purpose, several computational methods have been proposed in the literature, whereas we focused only on machine learning algorithms applied to this problem. Easy-to-use machine learning algorithms such as random forests and boosted regression trees were frequently used in infectious disease modeling studies. However, Gaussian processes might capture highly complex dependencies better than these tree-based algorithms. Thus, we formulated a computational framework based on Gaussian processes that can be used to perform spatial, temporal, or spatiotemporal prediction of infectious diseases.
We integrated spatial features (such as geographical coordinates) and temporal features (such as seasonal conditions) for location and time period pairs that were used as data instances in our Gaussian process formulation. However, a naive implementation of Gaussian processes would become computationally infeasible owing to very high numbers of pairs being modeled. We exploited the special structure (i.e., Kronecker) of similarity matrices in our formulation to obtain a very efficient implementation, which enabled us to train models for around 10,000 data instances in the order of milliseconds.
We applied our framework to the problem of predicting the case counts of a vector-borne infectious disease Crimean–Congo hemorrhagic fever using the data set of infected case counts between years 2004 and 2015 collected by the Ministry of Health of Turkey. We performed predictions under three different scenarios (Fig 1), which correspond to making predictions for unseen provinces (i.e., spatial prediction), future time periods (i.e., temporal prediction), or unseen province and time period pairs (i.e., spatiotemporal prediction) to show the suitability of our approach to distinct problems.
Predicting future cases of infectious diseases is very important for the control and prevention of the disease. The predicted case counts can be used to develop new public health policies and intervention mechanisms. It is more useful for public health policy makers to be able to predict the possible number of infected cases for a region and a time period pair rather than predicting whether there will be cases or not. Policy makers can make use of predicted number of infected cases to purchase vaccines around the right amount, to raise public awareness in the region, to educate healthcare workers, etc. From that perspective, GPR algorithm did a better job than RFR and BRT algorithms by predicting CCHF case counts more accurately (i.e., lower NRMSE values).
We tested our proposed formulation on a single disease, but the same framework can be extended towards other vector-borne infectious diseases (e.g., dengue fever, malaria, Zika fever) and as well as other infectious diseases (e.g., influenza, measles, tuberculosis). We also made the source code publicly available to enable other computational and applied researchers to make such extensions easily.
|
10.1371/journal.pcbi.1000392 | Fast Statistical Alignment | We describe a new program for the alignment of multiple biological sequences that is both statistically motivated and fast enough for problem sizes that arise in practice. Our Fast Statistical Alignment program is based on pair hidden Markov models which approximate an insertion/deletion process on a tree and uses a sequence annealing algorithm to combine the posterior probabilities estimated from these models into a multiple alignment. FSA uses its explicit statistical model to produce multiple alignments which are accompanied by estimates of the alignment accuracy and uncertainty for every column and character of the alignment—previously available only with alignment programs which use computationally-expensive Markov Chain Monte Carlo approaches—yet can align thousands of long sequences. Moreover, FSA utilizes an unsupervised query-specific learning procedure for parameter estimation which leads to improved accuracy on benchmark reference alignments in comparison to existing programs. The centroid alignment approach taken by FSA, in combination with its learning procedure, drastically reduces the amount of false-positive alignment on biological data in comparison to that given by other methods. The FSA program and a companion visualization tool for exploring uncertainty in alignments can be used via a web interface at http://orangutan.math.berkeley.edu/fsa/, and the source code is available at http://fsa.sourceforge.net/.
| Biological sequence alignment is one of the fundamental problems in comparative genomics, yet it remains unsolved. Over sixty sequence alignment programs are listed on Wikipedia, and many new programs are published every year. However, many popular programs suffer from pathologies such as aligning unrelated sequences and producing discordant alignments in protein (amino acid) and codon (nucleotide) space, casting doubt on the accuracy of the inferred alignments. Inaccurate alignments can introduce large and unknown systematic biases into downstream analyses such as phylogenetic tree reconstruction and substitution rate estimation. We describe a new program for multiple sequence alignment which can align protein, RNA and DNA sequence and improves on the accuracy of existing approaches on benchmarks of protein and RNA structural alignments and simulated mammalian and fly genomic alignments. Our approach, which seeks to find the alignment which is closest to the truth under our statistical model, leaves unrelated sequences largely unaligned and produces concordant alignments in protein and codon space. It is fast enough for difficult problems such as aligning orthologous genomic regions or aligning hundreds or thousands of proteins. It furthermore has a companion GUI for visualizing the estimated alignment reliability.
| The field of biological sequence alignment is very active, with numerous new alignment programs developed every year in response to increasing demand driven by rapidly-dropping sequencing costs. The list of approximately 60 sequence alignment programs on the wikipedia compilation provides a lower bound on the number of available tools and illustrates the confusing choice facing biologists who seek to select the “best” program for their studies. Nevertheless, the ClustalW program [1],[2], published in 1994, remains the most widely-used multiple sequence alignment program. Indeed, in a recent review of multiple sequence alignment [3], the authors remark that “to the best of our knowledge, no significant improvements have been made to the [ClustalW] algorithm since 1994 and several modern methods achieve better performance in accuracy, speed, or both.” Therefore, it is natural to ask, “Why do alignment programs continue to be developed, and why are new tools not more widely adopted by biologists?”.
A key issue in understanding the popularity of ClustalW is to recognize that it is difficult to benchmark alignment programs. Alignments represent homology relationships among the nucleotides, or amino acids, of the genomes of extant species, and it is impossible to infer the evolutionary history of genomes with absolute certainty. Comparisons of alignment programs therefore rely on databases of structural alignments for proteins and RNA or on gene loci or simulated data for DNA. Each type of benchmark is vulnerable to manipulation and furthermore may not represent the problem setups which are most relevant to biologists. The result is that biologists are confronted with many programs and publications, but it is frequently unclear which approach will give the best results for the everyday problems which they seek to address.
Adding to the difficulty of selecting software tools is the blur between programs and ideas. Developers of alignment programs make choices about the objective functions to optimize, the statistical models to use, and the parameters for these models, but the relative impact of individual choices is rarely tested [4]. Discordance among programs is frequently noted [5], but the different architectures of individual programs, and in some cases the lack of open software, makes it difficult to explore novel combinations of existing ideas for improving alignments.
In lieu of these issues, biologists have favored the conservative approach of using the tried and trusted ClustalW program, although they frequently use it in conjunction with software which allows for manual editing of alignments [6]. The rationale behind alignment-editing software is that trained experts should be able to correct alignments by visual inspection and that effort is better expended on manually correcting alignments than searching for software that is unlikely to find the “correct” alignment anyway. Although manual editing approaches may be cumbersome, they have been used for large alignments (e.g., [7]).
We therefore approached the alignment problem with the following goals in mind:
The “distance-based” approach to sequence alignment, proposed in [13] and implemented in the protein alignment program AMAP [14], offers a useful framework for these goals. Much as distance-based phylogenetic reconstruction methods like Neighbor-Joining build a phylogeny using only pairwise divergence estimates, a distance-based approach to alignment builds a multiple alignment using only pairwise estimations of homology. This is made possible by the sequence annealing technique [14] for constructing multiple alignments from pairwise comparisons.
We have implemented our approach in FSA, a new alignment program described below. We give an overview of the structure of FSA and explain the details of its components below. Text S1 contains detailed instructions for using the FSA program and webserver as well as FSA's companion programs for comparing alignments and working with whole-genome alignments.
Figure 1 shows an overview of the components of the FSA alignment algorithm, described in detail below.
The input to FSA is a set of protein, RNA or DNA sequences. These sequences are assumed to be homologous, although FSA is robust to non-homologous sequence. The output of FSA is a global alignment of the input sequences which is a (local) optima of the expected accuracy under FSA's statistical model.
FSA first performs pairwise comparisons of the input sequences to determine the posterior probabilities that individual characters are aligned (note, however, that it does not yet actually align any sequences). While the number of possible pairwise comparisons is quadratic in the number of sequences being aligned, giving unfavorable runtimes for datasets of many sequences, FSA overcomes this problem by reducing the number of pairs of sequences that are compared. This is accomplished using a randomized approach inspired by the Erdös-Rényi theory of random graphs, thereby drastically reducing the computational cost of multiple alignment.
After obtaining pairwise estimates of homology at the single-character level, FSA uses the sequence annealing technique [14] to construct a multiple alignment. This approach to alignment seeks to maximize the expected accuracy of the alignment using a steepest-ascent (greedy) algorithm.
The architecture of FSA reflects the inherent modularity of the distance-based approach to alignment. FSA's inference engine uses the flexible HMMoC code-generation tool [15] and has been parallelized with a separate module, alignments of long sequences are anchored with candidate matches found by the MUMmer suffix trie matching tool [16] or the exonerate homology-search program [17], and FSA's sequence annealing algorithm improves on the original algorithm and implementation in AMAP [14]. The stand-alone visualization tool uses statistical information produced by FSA, but otherwise is completely independent.
Each of these components can be improved independently of the others, allowing for rapid future improvements in distance-based alignment. For example, FSA's entire statistical model could easily be altered to incorporate position-specific features or completely replaced with a discriminative or hybrid generative-discriminative model.
The components described here correspond roughly to the simplest mode of operation for FSA, outlined in bold in Figure 1.
We benchmarked FSA against databases of multiple alignments compiled from reference structural alignments, including protein databases (BAliBASE 3 [24] and SABmark 1.65 [25]), small RNA databases (BRAliBase 2.1 [26]), large RNA databases (Consan mix80 [27]), and both mammalian [28] and fly [29],[30] simulated DNA alignments.
Alignment programs are commonly used to detect homology among input sequences. We conducted a series of false-positive experiments to test whether commonly-used alignment programs can reliably identify homologous and non-homologous sequence. Surprisingly, we found that for most alignment programs, aligned sequences are not necessarily homologous, indicating that biologists should use caution when interpreting the multiple alignments produced by many commonly-used tools.
We additionally performed a simple test of the consistency of common programs when aligning coding sequence: We aligned 1,502 genes orthologous across seven species of yeast in both nucleotide and protein space and compared the resulting alignments. Many programs gave surprisingly discordant results, indicating that at least one of these two alignments produced by commonly-used programs is incorrect.
Table 1 shows benchmarks of FSA and other alignment programs, including AMAP [14], ClustalW [1],[2], DIALIGN [31],[32], MAFFT [33], MUMMALS [34], MUSCLE [35], Probalign [36], ProbCons [37], T-Coffee [38], and SeqAn::T-Coffee [39], against the BAliBASE 3 [24] and SABmark 1.65 databases [25]. FSA in maximum-sensitivity mode had accuracy similar to those of the better-performing programs on BAliBASE 3 and had accuracy superior to that of any other program on SABmark 1.65 when run in default mode. FSA had higher positive predictive values than any other tested program on all datasets. Remarkably, FSA was the only tested program which achieved a mis-alignment rate <50% on the standard SABmark 1.65 datasets; all other programs produced more incorrect than correct homology statements.
In order to test the robustness of alignment programs to incomplete homology, we modified the BAliBASE 3 database such that every alignment included a single false-positive, an unrelated (random) sequence. This is a realistic setup for biologists who might want to decide whether a sequence is orthologous to a particular protein family. With the exception of FSA, the tested alignment programs suffered a false-positive rate increased by over 25% on this BAliBASE 3+fp dataset, indicating that they aligned the random sequence with the homologous set. In contrast, FSA left the random sequence unaligned and had an essentially-unchanged false-positive rate.
Table 2 shows benchmarks of FSA and the other tested alignment programs against the BRAliBase 2.1 [26] and Consan mix80 [27] databases. FSA outperformed all other programs on both datasets.
BRAliBase 2.1 was assembled from the RFAM [40] RNA database and consists of small RNAs (average length of ∼150 nt). FSA gave improved performance even on this high-identity dataset where most programs did relatively well.
The Consan mix80 dataset of alignments of Small and Large Subunit ribosomal RNAs, assembled from the European Ribosomal RNA database [41], was created for training RNA structural alignment programs and provides a test of alignment programs on difficult, large-scale alignments. The four alignments contain from 107 to 254 sequences, each 1–4 kilobases in length, with average percentage identity less than <50%. Two tested alignment programs, ProbConsRNA [42] and SeqAn::T-Coffee, were unable to align these large datasets. This dataset demonstrates that FSA's alignment speedup options, including performing inference only on a subset of all possible pairs (–fast) and anchoring alignments instead of using the full dynamic programming matrix (–anchored), are effective heuristics for large datasets.
Table 3 shows benchmarks of FSA and other genomic alignment programs, including CHAOS/DIALIGN [20], DIALIGN-TX [31],[32], MAVID [18], MLAGAN [19], Pecan [21] and TBA [28], on simulated alignments of both mammalian and Drosophila DNA sequences. FSA produced higher-accuracy alignments than the other programs on the two Drosophila datasets and only Pecan gave better alignments of the mammalian sequences.
The simulated alignments of nonfunctional DNA sequences from nine mammals (human, chimp, baboon, mouse, rat, cat, dog, cow, and pig) were created by Blanchette et al. [28]. The simulated alignments of DNA from the twelve species of Drosophila described in [43] were created with two simulation programs, DAWG [29] and simgenome [30]. As described in [30], the simulated Drosophila genomic alignments were created by parameterizing the DAWG and simgenome programs using whole-genome alignments produced by Pecan for [43]. Although two authors (RKB and IH) of this manuscript contributed to the simgenome program, simgenome was developed prior to FSA and did not influence or contribute to the methodology described here for FSA.
FSA's strong performance on all three datasets of simulated long DNA sequences indicate that it is a useful and accurate tool for genomic alignment.
In order to further test the appropriateness of using popular alignment programs to detect homology between sequences, we conducted a large-scale random-sequence experiment. We generated datasets of random sequences to simulate unrelated protein, short DNA, and genomic (long) DNA sequences. The results, shown in Table 4 and Table 5, clearly demonstrate that while for most alignment programs, aligned sequences are not necessarily homologous, FSA leaves random sequences largely unaligned.
Biologists commonly align coding regions in both amino acid and nucleotide space, but there have been few studies of the effectiveness of alignment programs across the two regimes. We tested the consistency of alignment programs on coding sequence by aligning all 1,502 genes in Saccharomyces cerevisiae identified as having orthologs in the six related yeast species S. paradoxus, S. mikatae, S. kudriavzevii, S. bayanus, S. castellii, and S. kluyveri ([44]; this dataset was also analyzed in [5]). As shown in Table 6, alignments produced by FSA had higher concordance (0.943) than those produced by any other program. If a program produces alignments with low concordance between amino acid and nucleotide alignments, then at least one of the alignments must be incorrect (and quite possibly both are).
This simple test additionally indicates the effectiveness and robustness of FSA's query-specific learning. While very different learning procedures are used for amino acid and nucleotide sequence, the concordant alignments inferred by FSA indicate that our results are robust with respect to the details of the learning procedure.
We conducted an ablation analysis of FSA's components to test the effectiveness of each component and ensure that they all contributed to the accuracy of the program. As indicated by the results in Tables 7–10, each optional component of FSA contributes to its accuracy.
The importance of each component depends strongly upon the alignment problem. The –fast heuristic for reducing the number of sequence pairs used to construct an alignment results in little loss of accuracy, at least on the benchmarks used in this paper (Tables 7 and 8). As indicated by the small and long RNA benchmarks (Table 8), iterative refinement is important for aligning many sequences and less so for small alignment problems. The anchor annealing procedure appears to be an effective heuristic for aligning long sequences. Anchoring with unique matches (MUMs) causes only a negligible loss of accuracy on the long RNA dataset (Table 8). However, results on simulated long DNA sequences (Table 9) demonstrate that inexact matches, such as those returned by exonerate, must be used during anchor annealing to obtain high sensitivity on very long or distant nonfunctional DNA sequences. Nonfunctional DNA lacks the local constraints which preserve exact matches across distant species in functional (e.g., coding) sequence. Query-specific learning is important for maintaining FSA's robustness to non-homologous sequence. While FSA aligned only 4% of random protein sequences in default mode, when run without learning it aligned 13% (Table 10), similar to the 14% aligned by AMAP (Table 4).
Biologists commonly perform alignments of hundreds or thousands of 16S ribosomal DNA sequences in order to elucidate evolutionary relationships and build phylogenetic trees. We performed alignments of prokaryotic 16S sequences to compare the speed of commonly-used programs (Table 11). MAFFT was the fastest method by an order of magnitude; MUSCLE and FSA were the next-fastest methods. Many alignment programs were unable to align these large datasets.
The results in Table 12 and Table 13 demonstrate the effectiveness of FSA's parallelization mode. Parallelizing the pairwise comparisons dramatically reduces runtime: When running in –fast mode on a small cluster with 10 processors, FSA can align 500 16S sequences in 20% of the time required without parallelization.
In the Introduction we highlighted four design criteria which we emphasized in the development of FSA. The first goal was to find alignments with high expected accuracy, where an accurate alignment has minimal distance to the truth. This objective function is markedly different from both the maximum-likelihood approaches used by programs such as ClustalW and MUSCLE and the maximum expected sensitivity approaches used by programs such as ProbCons and Pecan. Note that while the objective function used in ProbCons is called “maximum expected accuracy” in the paper [37], it is actually a maximum expected sensitivity objective function, where there is no penalty for over-aligning sequence (c.f., the results shown in Table 4). Their objective function can be recovered as a special case of our approach by ignoring the gap probabilities in FSA's objective function (Text S1, “The mathematics of distance-based alignment”). FSA's explicit search for the most accurate, rather than most likely or most sensitive, alignment is what allows FSA to so dramatically outperform most other programs on tests on unrelated sequence (Table 4).
We believe that this accuracy criterion, which gives equal weight to the correctness of all sequence positions, is a natural measure of alignment quality. Downstream analyses, such as phylogenetic reconstruction and evolutionary constraint analysis, are increasingly using indels in addition to homologous characters for more accurate estimation (e.g., [45],[46]). Thus, it is important that alignments be as “evolutionarily correct” as possible [47], which is the purpose of the accuracy criterion.
FSA's strong performance under the accuracy criterion is due to techniques such as its iterative refinement as well as its explicit attempt to maximize the expected accuracy; programs which explicitly seek to maximize an objective function of the posterior probabilities of character alignment, such as ProbCons or Probalign, could instead seek to maximize the expected accuracy described here and, as a likely result, increase their robustness to non-homologous sequence. However, while we believe that the expected accuracy is a biologically-sensible objective function, it may not be appropriate if the user desires the most sensitive alignment. While FSA can produce the most-sensitive RNA alignments, other programs can produce more sensitive alignments of proteins and genomic sequence, albeit generally at the cost of a tendency to align non-homologous sequence (Table 4).
The second goal was to create alignments which are robust to evolutionary distances and different functional constraints on patterns of molecular evolution. FSA's unsupervised query-specific learning for parameter selection frees the user from unknown systematic biases implicitly introduced by using an alignment program whose parameters were trained on a dataset whose statistics may not reflect those of the sequences to be aligned. For example, it is traditionally challenging to align sequences with unusual base composition, but FSA can easily handle such alignments by automatically learning appropriate parameters. As indicated by our ablation analysis, query-specific learning further increases FSA's robustness to non-homologous sequences beyond that offered by the minimum-distance objective function alone.
We believe that FSA's unsupervised query-specific learning is the first time a multiple alignment program has been capable of dynamically learning a complete parameterization, wherein parameters can vary for each pair of sequences to be compared, on the fly. This learning method is related to the “pre-training” option in ProbCons, which permits users to learn different models for families of homologous sequences, but does not permit parameterizations to vary between sequence pairs. We also note that the MORPH program for pairwise alignment of sequences with cis-regulatory modules learns model transition parameters from data [48]. While supervised training on curated data can give superior performance on test sets which are statistically-similar to the training data, the practical alignment problems encountered everyday by biologists do not fit into this rigid problem setup. Query-specific learning consistently gives reasonable performance.
The third and fourth goals, to develop a single, modular program which can address all typical alignment problems encountered by biologists, are naturally achieved within FSA's architecture. While almost all alignment programs are designed to either align many short sequences or a few long sequences, we have demonstrated that it is feasible to create a single program which can address both situations. This is made practical by FSA's modular nature, where the statistical model for pairwise comparisons, the anchoring scheme for finding homology between long sequences, and the sequence annealing procedure are entirely separate and can be individually modified and improved. For example, the parallelization of FSA was designed and developed with minimal changes to the rest of FSA's code base. Similarly, while FSA's basic anchoring relies only on exact matches from MUMmer, the anchoring scheme was easily extended to incorporate inexact matches from exonerate [17] and alignment constraints from Mercator [22]. In fact, this flexibility permits FSA to incorporate almost any sources of potential homology information, from predicted transcription factor binding sites to entire gene models. One natural extension of FSA's approach is to models of RNA alignment which take structure into account. The program Stemloc-AMA [49] uses a model of the pairwise evolution of RNA secondary structure in conjunction with the sequence annealing algorithm to create accurate multiple alignments of structured RNAs. By using Stemloc-AMA's probabilistic model rather than a Pair HMM and taking advantage of techniques such as query-specific learning, FSA could sum over possible pairwise structural alignments in order to get better estimates of posterior probabilities of character alignment.
FSA is a statistical alignment program insofar as it uses an explicit statistical model of alignments and a probabilistic objective function for optimization, but as discussed in “Theoretical justification of distance-based alignment” (Text S1), it also is a distance-based approximation to the “phylogenetic alignment” models of alignments on trees [8]–[11], [50]–[52]. While traditional phylogenetic alignment algorithms are currently too computationally-expensive to use on datasets of more than a few sequences, FSA's distance-based method allows biologists to use the sophisticated tools of statistical alignment algorithms on practical problems. Furthermore, while we have not addressed the phylogenetic aspects of FSA in detail in this paper, our methods can be adapted to incorporate a complete phylogenetic model (Text S1, “The mathematics of distance-based alignment”). However, we believe that FSA's current approach, which is agnostic to phylogeny, offers many practical advantages for common genomics analyses. For example, because FSA uses a sum-of-pairs objective function, there is no guide tree to potentially bias downstream phylogenetic reconstructions based on the alignment. Similarly, while most genomic alignment programs require a species tree to perform the alignment, our phylogeny-free approach will be avoid potential biases introduced by using a single species tree to align regions which may have undergone recombination.
|
10.1371/journal.pcbi.1004294 | The Opponent Channel Population Code of Sound Location Is an Efficient Representation of Natural Binaural Sounds | In mammalian auditory cortex, sound source position is represented by a population of broadly tuned neurons whose firing is modulated by sounds located at all positions surrounding the animal. Peaks of their tuning curves are concentrated at lateral position, while their slopes are steepest at the interaural midline, allowing for the maximum localization accuracy in that area. These experimental observations contradict initial assumptions that the auditory space is represented as a topographic cortical map. It has been suggested that a “panoramic” code has evolved to match specific demands of the sound localization task. This work provides evidence suggesting that properties of spatial auditory neurons identified experimentally follow from a general design principle- learning a sparse, efficient representation of natural stimuli. Natural binaural sounds were recorded and served as input to a hierarchical sparse-coding model. In the first layer, left and right ear sounds were separately encoded by a population of complex-valued basis functions which separated phase and amplitude. Both parameters are known to carry information relevant for spatial hearing. Monaural input converged in the second layer, which learned a joint representation of amplitude and interaural phase difference. Spatial selectivity of each second-layer unit was measured by exposing the model to natural sound sources recorded at different positions. Obtained tuning curves match well tuning characteristics of neurons in the mammalian auditory cortex. This study connects neuronal coding of the auditory space with natural stimulus statistics and generates new experimental predictions. Moreover, results presented here suggest that cortical regions with seemingly different functions may implement the same computational strategy-efficient coding.
| Ability to localize the position of a sound source is vital to many organisms, since audition provides information about areas which are not accessible visually. While its importance is undisputed, its neuronal mechanisms are not well understood. It has been observed in experimental studies that despite the crucial role of sound localization, single neurons in the auditory cortex of mammals carry very little information about the sound position. The joint activity of multiple neurons is required to accurately localize sound, and it is an open question how this computation is performed by auditory cortical circuits. In this work I propose a statistical model of natural stereo sounds. The model is based on the theoretical concept of sparse, efficient coding which has provided candidate explanations of how different sensory systems may work. When adapted to binaural sounds recorded in a natural environment, the model reveals properties highly similar to those of neurons in the mammalian auditory cortex, suggesting that mechanisms of neuronal auditory coding can be understood in terms of general, theoretical principles.
| The precise role played by the auditory cortex in hearing remains unclear. Before reaching cortical areas, raw sounds undergo numerous transformations in the brainstem and the thalamus. This subcortical processing seems more substantial than in other senses and is a specific property of the auditory system. What computations are performed by the cortex on the output generated by lower auditory regions is a question far from being answered.
One of the issues in functional characterization of the auditory cortex is an apparent lack of specificity. Spiking activity of cortical auditory neurons is modulated by sound features such as pitch, timbre and spatial location [1, 2]. Responses invariant to any of those features seem rare. This interdependence is especially puzzling in the context of extracting spatial information. A number of studies attempted to identify “what” and “where” streams in the auditory system (e.g. [3, 4]). Despite those efforts the existence of a sharp separation of spatial and identity information in the auditory cortex is still not definitely confirmed [5, 6].
Neurons reveal sensitivity to sound position in most parts of the mammalian auditory cortex [7]. Their spatial tuning is quite broad — neural firing can be modulated by sounds located anywhere on the azimuthal plane. While activity of single units does not carry information sufficient to accurately localize sounds, larger numbers of neurons seem to form a population code for sound location [8–11]. These observations strongly differ from assumptions made early in the field that the auditory space is represented by a topographic cortical map, where neighboring units would encode the presence of a sound source at proximal positions [12].
Results described above constitute a conceptual challenge for theoretical models of the auditory cortex and understanding its role in spatial hearing in particular. Nevertheless, a number of candidate roles for this region have been proposed. It has been suggested, for instance, that the main function of the primary auditory cortex (A1) is to process sound features extracted by subcortical structures [13] on multiple time scales. Another theory proposes that the auditory cortex in the cat represents abstract entities (such as a bird song or wind) rather than low-level spectrotemporal features of the stimulus [14]. It is also debated whether auditory cortical areas bear similarities to visual areas, and if yes, what sort of understanding can be gained by combining knowledge about those brain regions [15]. From a theoretical perspective one question seems to be particularly important — is there any general principle behind the functioning of auditory cortex, or does it carry out computations which are task- or modality-specific and therefore not performed in other parts of the brain?
A particularly succesful theoretical framework attempting to explain general mechanisms behind the functioning of the nervous system is provided by the Efficient Coding Hypothesis [16, 17]. It proposes that sensory systems maximize the amount of transmitted stimulus information. Under the additional assumption that a typical stimulus activates only a small fraction of a neuronal population, the hypothesis is known as sparse coding[18, 19]. Perhaps the strongest prediction of the efficient coding hypothesis is that the neuronal activity at consecutive stages of sensory processing should become less and less redundant, hence more independent. This prediction has been experimentally tested in the auditory system of the cat. Chechik and colleagues [20] recorded neuronal responses to natural sounds at three levels of the auditory hierarchy — Inferior Colliculus (IC), Medial Genniculate Body (MGB) and A1. They observed that spiking activity was progressively less redundant between single neurons, as quantified using information theoretic measures. This result suggests that audition can be understood using concepts provided by the efficient coding hypothesis.
In order to form an efficient stimulus representation, neuronal codes should reflect regularities present in the sensory environment [21]. This implies that by studying statistics of natural input, one should be able to predict neuronal tuning properties. In audition, this idea has been followed by a number of researchers. Starting at the lowest level of the auditory system, Lewicki and Smith [22, 23] demonstrated that learning a sparse representation of natural sound chunks reproduces shapes of cochlear filters of the cat. A recent extension of this work has suggested that while the auditory nerve may be optimally encoding all sounds it encounters, neurons in the cochlear nucleus may be tuned to efficiently represent particular sound classes [24]. Climbing higher in the auditory hierarchy — Carlson et al [25] have reproduced shapes of spectrotemporal receptive fields (STRFs) in the inferior colliculus by learning sparse codes of speech sounds. The relationship between spectrotemporal tuning of cortical neurons and sparse representation of speech spectrograms were explored by Klein, Koerding and Koenig [26, 27]. More recently, some aspects of the topographic structure of the auditory cortex were shown to emerge from statistics of speech sounds by Terashima and Okada [28]. Terashima and colleagues have also studied the connection between sparse codes of natural sound spectra and harmonic relationships revealed by receptive fields of macaque A1 neurons [29]. Maximizing coding efficiency by learning sparse codes has also lead to emergence of signal representations useful in spatial hearing tasks. Asari et al [30] learned overcomplete dictionaries of monaural spectrograms and demonstrated that this representation allows for the segregation of acoustically overlapping and yet spatially separate sound sources (the “cocktail party problem”). A recent study found that sparse coding of a spectrotemporal representation of binaural sound extracts spatial features invariant to sound identity [31].
As discussed above, a growing body of evidence seems to point to efficient coding as a computational process implemented by the auditory system. To date however, the connection between natural stimulus statistics and auditory spatial receptive fields remains unexplained. It is therefore unclear if spatial computations performed by the auditory cortex are unique to this brain area or whether they can be also predicted in a principled way from a broader theoretical perspective.
The present work attempts to connect spatial computations carried out by the auditory cortex with statistics of the natural stimulus. Here, a hierarchical statistical model of stereo sounds recorded in a real auditory environment is proposed. Based on principles of sparse coding the model learns the spectrotemporal and interaural structure of the stimulus. In the next step, it is demonstrated that when probed with spatially localized sounds, higher level units reveal spatial tuning which strongly resembles spatial tuning of neurons in the mammalian auditory cortex. Additionally, the learned code forms an interdependent representation of spatial information and spectrotemporal quality of a sound. Activity of higher units is therefore modulated by sound’s position and identity, as observed in the auditory system.
This study provides computational evidence that spatial tuning of auditory cortical neurons may be a manifestation of an underlying general design principle — efficient coding. Present results suggest that the role of the auditory cortex is to reduce redundancy of the stimulus representation preprocessed by the brainstem. Representation obtained in this way may facilitate tasks performed by higher brain areas, such as sound localization.
Binaural sound used to train the model was recorded by a human subject walking freely in a wooded area, in the presence of another speaker. The obtained recording included ambient (wind, flowing stream) and transient environmental sounds (wood cracking, steps) as well as human speech. The auditory scene changed over time due to the motion of the the recorder, the speaker, and changes in the environment. It consisted of multiple sound sources emanating from a diverse set of locations, creating together a complex, natural auditory environment. Please refer to the Methods section for details of the recording.
The present study proposes a hierarchical statistical model of binaural sounds, which captures binaural and spectrotemporal structure present in natural stimuli. The architecture of the model is shown in Fig 1. It consists of the input layer and two hidden layers. The input to the model was N sample-long epochs of binaural sound: from the left ear—xL and from the right ear—xR. The role of the first layer was to extract and separate phase and amplitude information from each ear by encoding them in an efficient manner. Monaural sounds were transformed into phase (ϕL, ϕR) and amplitude (aL, aR) vectors. This layer can be thought of as a statistical analogy to cochlear filtering. Phase vectors were further modified by computing interaural phase differences (IPDs) — a major sound localization cue [32]. This tranformation may be considered an attempt to mimic functioning of the medial superior olive (MSO) — the brainstem nucleus where phase differences are extracted [32].
The second layer of the model learned a joint sparse representation of monaural amplitudes (aL, aR) and phase differences (Δϕ). Level (amplitude) and temporal (phase) information from each ear was jointly encoded by a population of M units. Each of the units was therefore capturing higher-order spectrotemporal patterns of sound in each ear. Additionally, by combining monaural information into single units higher level representation achieved spatial tuning not present in the first layer. The second hidden layer can be interpreted as a model of auditory neurons which receive converging monaural input and jointly operate on phase and amplitude — two kinds of information known to be important for spatial hearing.
As demonstrated in previous work, filtering properties of the auditory nerve can be explained by sparse coding models of natural sounds [22]. There, short epochs of natural sounds are modelled as a linear combination of real-valued basis functions multiplied by sparse, independent coefficients (i.e. having highly curtotic marginal distributions). Adapted to sets of natural sound chunks, basis functions become localized in time and/or frequency matching properties of cochlear filters.
While being capable of capturing interesting properties of the data, real valued representations are not well suited for modeling binaural sounds. This is because binaural hearing mechanisms utilize interaural level and time differences (ILDs and ITDs respectively). In narrowband channels, differences in time correspond to phase displacements known as interaural phase differences (IPDs). Therefore a desired representation should both be adapted to the data (i.e. non-redundant) and separate amplitude from phase (where phase is understood as a temporal shift smaller than the oscillatory cycle of a particular frequency).
The present work addresses this twofold constraint with complex-valued sparse coding. Each data vector x ∈ ℝN is represented in the following way:
x t = ∑ i = 1 N R { z i * A i , t } + η (1)
where zi ∈ ℂ are complex coefficients, * denotes a complex conjugation, Ai ∈ ℂT are complex basis functions and η ∼ 𝓝(0, σ) is additive gaussian noise. Complex coefficients in Euler’s form become z i = a i e j ϕ i (where j = − 1) therefore Eq (1) can be rewritten to explicitely represent phase ϕ and amplitude a as separate variables:
x t = ∑ i = 1 N a i ( cos ϕ i A i , t R + sin ϕ i A i , t I ) + η (2)
Real and imaginary parts A i R and A i I of basis functions { Ai }i=1N span a subspace within which the position of a data sample is determined by amplitude ai and phase ϕi. Depending on number of basis functions N (each of them is formed by a pair of vectors), the representation can be complete (N/2 = T) or overcomplete (N/2 > T).
In a probabilistic formulation, Eqs (1) and (2) can be understood as a likelihood model of the data, given coefficients z and basis functions A:
p ( x | z , A ) = [ 1 σ 2 π ] T ∏ t = 1 T e - ( x t - x ^ t ) 2 2 σ 2 (3)
where x ^ t = ∑ i = 1 N R { z i * A i , t } is the reconstruction of the t−th dimension of the data vector x. A prior over complex coefficients applied here assumes independence between subspaces and promotes sparse solutions i.e. solutions with most amplitudes close to 0:
p ( z ) = 1 Z ∏ i = 1 N e - λ S ( a i ) (4)
where Z is a normalizing constant. Function S(ai) promotes sparsity by penalizing large amplitude values. Here, a Cauchy prior on amplitudes is assumed i.e. S ( a i ) = l o g ( 1 + a i 2 ). One should note however that amplitudes are always non-negative and that in general the Cauchy distribution is defined over the entire real domain. The model attempts to form a data representation keeping complex amplitudes maximally independent across subspaces, while still allowing dependence between coordinates z𝕽, z𝕴 which determine position within each subspace. Inference of coefficients z which represent data vector x in the basis A is performed by minimizing the following energy function
E 1 ( z , x , A ) ∝ 1 2 σ 2 ∑ t = 1 T ( x t ^ - x t ) 2 + λ ∑ i = 1 N S ( a i ) (5)
which corresponds to the negative log-posterior p(z∣x, A). This model was introduced in [33] and used to learn motion and form invariances from short chunks of natural movies. Assuming N = T/2 and σ = 0, it is equivalent to 2-dimensional Independent Subspace Analysis(ISA) [34].
When trained on natural image patches, real and imaginary parts of basis functions A form pairs of Gabor-like filters, which have the same frequency, position, scale and orientation. The only differing factor is phase—real and imaginary vectors are typically in a quadrature-phase relationship (shifted by π 2). By extension, one might expect that the same model trained on natural sounds should form a set of frequency localized phase-invariant subspaces, where imaginary vector is equal to the real one shifted a quarter of a cycle in time. Somewhat surprisingly, such representation does not emerge, and learned subspaces capture different aspects of the data — bandwidth, frequency or time invariance [35, 36].
In order to learn a representation from the statistics of the data that preserves a desired property such as phase invariance, one could select a parametric form of basis functions and adapt the parameter set [37]. Such a parametric approach has the disadvantage that the assumed family of solutions might not be flexible enough to efficiently span the data space. Another, more flexible alternative to learn a structured representation is to regularize basis functions by imposing temporal-coherence promoting priors [36]. This, however, requires determining the strength of regularizing priors.
To overcome these problems, a different approach was taken here. The first-layer representation was created in two steps. Firstly a real-valued sparse code was trained (see Methods). Learned basis functions were well localized in time or frequency and tiled the time-frequency plane in a uniform and non-overlapping manner (Fig 2B). They were taken as real vectors Aℜ of complex basis functions A. In the second step, imaginary parts were created by performing the Hilbert transform of real vectors. The Hilbert transform of a time varying signal y(t) is defined as follows:
H ( y ( t ) ) = 1 π p . v . ∫ - ∞ ∞ y ( τ ) t - τ d τ (6)
Where p.v. stands for Cauchy principal value. In such a way every real vector A i ℜ was paired with its Hilbert transform A i ℑ = H ( A i ℜ ) i.e. a vector which complex Fourier’s coefficients are all shifted by π 4 in phase. The obtained dictionary is adapted to the stimulus ensemble, hence providing a non-redundant data representation, yet makes phase clearly interpretable as a temporal displacement.
The model was trained using T = 128 sample-long chunks of sound sampled at 8 kHz, which corresponds to 16 ms duration. The complete representation of 128 real basis functions was trained, and each of them was paired with its Hilbert transform, resulting in the total number of 256 basis vectors. Selected basis functions are displayed in Fig 2A. Real vectors are plotted in black together with associated imaginary ones plotted in gray. Panel B of the same figure displays isoprobability contours of Wigner-Ville distributions associated with the 256 basis functions. This form of representation localizes each temporal feature on a time-frequency plane [38] (one should note that real and imaginary vectors within each pair are represented by the same contour on that plot). A clear separation into two classes is visible. Low frequency basis functions (below 1 kHz) are non-localized in time (spanning the entire 16 ms interval), while in higher frequency regions their temporal precision increases. An interesting bandwidth reversal is visible around 3 kHz, where temporal accuracy is traded for frequency precision. Interestingly, the sharp separation into frequency and time localized basis functions, which emerged in this study was not clearly visible in other studies which performed sparse coding of sound [22, 38]. Time-frequency properties observed here reflect the statistical structure of the recorded auditory scene, which mostly consisted of non-harmonic environmental sounds sparsely interspersed with human speech.
Fig 3 depicts a typical distribution of binaural phase. Phases of the same basis function in each ear reveal dependence in their difference. This means that joint probability of monaural phases depends solely on the IPD:
p ( ϕ i , L , ϕ i , R ) ∝ p ( Δ ϕ i ) (7)
where Δϕi = ϕi, L−ϕi, R is the IPD. This property is a straightforward consequence of physics of sound — sounds arrive to each ear with a varying delay giving rise to positive and negative phase shifts. From a statistical point of view this means that monaural phases become conditionally independent given their difference and a phase offset ϕi, O:
ϕ i , L ⊥ ϕ i , R | Δ ϕ i , ϕ i , O (8)
The phase offset ϕi, O is the absolute phase value — indicating the time from the beginning of the oscillatory cycle. It can be therefore said that:
ϕ i , L = ϕ i , O + Δ ϕ i 2 (9) ϕ i , R = ϕ i , O - Δ ϕ i 2 (10)
This particular statistical property allows us to understand IPDs not as an ad-hoc computed feature but as an inherent property of the probability distribution underlying the data. It is reflected in the structure of the graphical model (see Fig 1). Since the phase offset ϕi, O does not carry spatial information for the purposes of current study it is treated as an auxiliary variable and therefore marked in gray.
In an approach to model the cochlear coding of sound, monaural sound epochs xL and xR were encoded independently using the same dictionary of complex basis functions A described in the previous section. Signal from both ears converged in the second hidden layer, which role was to form a joint, higher-order representation of the entire stimulus processed by the auditory system.
The celebrated Duplex Theory of spatial hearing specifies two kinds of cues used to solve the sound-localization task: interaural level and time (or phase) differences [39]. While IPDs are supposed to be mostly used in localizing low-frequency sounds, ILDs are a cue, which (at least in the laboratory conditions) can be used to identify the position of high frequency sources. Phase and level cues are known to be computed in lateral and medial superior olive (LSO and MSO respectively) — separated anatomical regions in the brainstem [32]. However, an assumption made here was that neurons in the auditory cortex receive converging input from subcortical structures. This would enable them to form their spatial sensitivity using both fine structure phase and amplitude information. One can take also the inverse perspective: a single object (a “cause”) in the environment generates level and phase cues at the same time. Its identification therefore has to rely on observing dependencies between those features of the stimulus.
The second layer formed a joint representation of monaural amplitudes and interaural phase differences. However, not all IPDs were modelled in that stage. Humans stop utilizing fine structure IPDs in higher frequency regimes (roughly above 1.3 kHz), since this cue becomes ambiguous [32]. Aditionally, cues above around 700 Hz become ambiguous (a single cue value does not correspond to a unique source position). For those reasons, and in order to reduce the number of data dimensions, 20 out of 128 IPD values were selected. The selection criteria were the following: (i) an associated basis function should have the peak of the Fourier spectrum below 0.75 kHz (which provided the upper frequency bound), and (ii) it should have at least one full cycle (which provided the lower bound). All basis functions fulfilling these criteria were non-localized in time (they spanned entire 16 ms interval). As a result, the second layer of the model was jointly encoding T = 128 log-amplitude values from each ear and P = 20 phase differences.
Monaural log-amplitude vectors aL, aR ∈ ℝT were concatenated into a single vector a ∈ ℝ2×T, and encoded using a dictionary of amplitude basis functions B. Representation of IPDs (Δϕ) was formed using a separate feature dictionary ξ. Both — phase and amplitude basis functions (B and ξ), were coupled by associated sparse coefficients s. The overall generative model of phases and amplitudes was defined in the following way:
a n = ∑ i = 1 M s i B i , n + η (11) Δ ϕ n = | w | ∑ i = 1 M s i ξ i , n + ϵ (12)
The amplitude noise was assumed to be gaussian (η ∼ 𝓝(0, σ2)) with σ2 variance. Since phase is a circular variable its noise ε was modelled by the von Mises distribution with concentration parameter κ.
The second layer was encoding two different physical quantities — phases, which are circular values, and log-amplitudes, which are real numbers. The goal was to form a joint representation of both parameters and learn their dependencies from the data. A simple, linear sparse coding model could be in principle used to achieve this task. However, if a single set of sparse coefficients si was used to model both quantities, scaling problems could arise, namely a coefficient value which explains well the amplitude vector may be too large or too small to explain the concomittant IPD vector. For this reason an additional phase multiplier w was introduced. It enters Eq 11 as a scaling factor, which gives the model additional flexibility required to learn joint probability distribution of amplitudes and IPDs. Fig 1 depicts it in gray as an auxiliary variable. In this way, amplitude values and phase differences were modelled by variables sharing a common, sparse support (coefficients s), with a sufficient flexibility.
Pairs of basis functions Bi, ξi represent binaural spectrotemporal stimulus and IPD patterns respectively, while sparse coefficients s signal their joint presence in the encoded sound epoch. An i−th second-layer unit was activated (si ≠ 0) whenever a pattern of IPDs represented by the basis function ξi or a pattern of amplitudes represented by Bi was present in its receptive field. The activity was maximized, when both features were present at the same time. For this reason, when seeking analogies between the higher-level representation and auditory neurons, coefficients s can be interpreted as neuronal activity (e.g. firing rate) and basis function pairs Bi, ξi as receptive fields (i.e. stimulus preferred by a neuron).
The likelihood of amplitudes and phase differences defined by the second layer was given by:
p ( a , Δ ϕ | s , w , B , ξ ) = [ 1 σ 2 2 π ] 2 T ∏ n = 1 2 T e - ( a n - a ^ n ) 2 2 σ 2 2[ 1 2 π I 0 ( κ ) ] P ∏ m = 1 P e κ cos ( Δ ϕ m - Δ ϕ ^ m ) (13)
where a ^ n = ∑ i = 1 M s i B i , n, Δ ϕ ̂ m = ∣ w ∣ ∑ i = 1 M s i ξ i , m are amplitude and phase reconstructions repsectively and I0 is the modified Bessel function of order 0. The joint distribution of coefficients s was assumed to be equal to the product of marginals:
p ( s ) = 1 Z ∏ i = 1 M e - λ 2 S ( s i ) (14)
where λ2 is a sparsity controlling parameter. A Cauchy distribution was assumed as a prior over marginal coefficients (i.e. S ( s i ) = log ( 1 + s i 2 )). To prevent degenerate solutions, where sparse coefficients s are very small and the scaling coefficient w grows undbounded, a prior p(w) constraining it from above and from below was placed. A generalized Gaussian distribution of the following form was used:
p ( w ) = β 2 α Γ ( 1 β ) e - ( | w - μ | α ) β (15)Γ denotes tha gamma function, α, β and μ denote the scale, shape and location parameters respectively. When the shape parameter β is set to a large value (here β = 8), the distribution approximates a uniform distribution. Varying the scale parameter α changes the upper and the lower limit of the interval.
Taken together the negative log-posterior over the second layer coefficients was defined by the energy function:
E 2 ( s , w , B , ξ ) ∝ 1 σ 2 2 ∑ n = 1 2 × T ( a n - a ^ n ) 2 + κ ∑ m = 1 P cos ( Δ ϕ m - Δ ϕ ^ m ) + λ 2 ∑ i = 1 M S ( s i ) + λ w ( | w - μ | α ) β (16)
the λw coefficient was introduced to control the strength of the prior on the scaling coefficient w. Similarly as in the first model layer, learning of basis functions and inference of coefficients was performed using gradient descent (see Methods). The total number M of basis function pairs was set to 256.
The second layer learned cooccuring phase and amplitude patterns forming a sparse, combinatorial code of the first layer output. Fig 4 displays 10 representative examples of basis function pairs ξi and Bi, which encoded amplitudes and IPDs respectively.
Each amplitude basis function consisted of two monaural parts corresponding to the left and right ear. First-layer, temporal features were visualized using contours of Wigner-Ville distribution and colored according to the relative weight. Entries of IPD basis functions were values (marked by gray bars) modelling interaural phase disparities in each of selected 20 frequency channels.
The subset of 9 basis functions depicted in subpanels B-J of Fig 4 constitutes a good representation of the entire dictionary. Their vertical ordering corresponds to spectrotemporal properties of Bi basis functions. Amplitude features displayed in the first row (B, E, H) reveal pronounced spectral modulation, while the last row (D, G, J) are features which are strongly temporaly modulated. Columns are ordered according to the ear each basis function pair prefered. Left column (B, C, D) are left-sided basis functions. Higher amplitude values are visible in the left ear parts (although differences are rather subtle), while associated IPD features are all negative. IPDs smaller than 0 imply that the encoded waveform was delayed in the right ear, hence the sound source was closer to the left ear. The last column (H, I, J) depicts more right-sided basis functions. Features displayed in the middle column (E, F, G) weight binaural amplitudes equally, however entries of associated phase vectors are either mostly negative or mostly positive.
As Fig 4 shows, higher level representation learned spectrotemporal properties of the auditory scene, reflected in shapes of amplitude basis functions Bi. Binaural relations were captured by relative weighting of amplitudes in both ears and the shape of the IPD basis function.
To get a more detailed understanding of the spectrotemporal features captured by the representation, analysis of modulation spectra was performed. A modulation spectrum is a 2D Fourier transform of the spectrotemporal representation of a signal. It is known that modulation spectra of natural sounds posess specific structure [40]. Here, modulation spectrum was computed separately for monaural parts of amplitude basis functions Bi (see Methods). In the next step a center of mass of each of the modulation spectra was computed. Centers of mass are represented by single points in Fig 5A.
A clear tradeoff between spectral and temporal modulation was visible. Basis functions which were strongly temporally modulated revealed simultaneously weak spectral modulation (and vice versa). It is evident as a “triangular” shape of the point distribution in Fig 5A. This seems to be a robust property of natural sounds [40] and was shown to be captured by sparse coding models [25–27, 41, 42]. Interestingly, spectro-temporal receptive fields of auditory neurons share this property [43, 44]. Auditory neurons which reveal sensitivity to spectrotemporal sound patterns seem to prefer sounds which are either modulated in time or in frequency, but not both. When compared with modulation spectra of natural sound this form of tuning may be understood as an adaptation to the environmental stimulus statistics.
Average temporal modulation in the left ear is plotted against the right ear in panel B. As visible — the amplitude modulation of basis functions B varied between 0 and 40 Hz, and a general linear trend was present. A linear regression model was fitted to these data (the fit is depicted in Fig 5 as a black dashed line). The fit has revealed a relatively strong linear relationship between the temporal variation of monaural parts (with Pearson’s correlation ρ = 0.70).
Spectral amplitude modulation revealed a weaker interaural correlation (ρ = 0.36). This is visible in Fig 5C—points representing amplitude basis functions are scattered stronger than in panel B of the same figure. This property can be explained by head filtering characteristics. Acting as a low-pass filter, the head attenuates higher frequencies. For this reason, fine spectral information above 1.5 kHz was typically more pronounced in a single ear, affecting the strength of spectral modulation. This may be considered as an example of how stimulus statistics are determined not only by the environmental properties, but also by the anatomy of an organism. The majority of basis functions revealed spectral modulation smaller than 0.4 cycle per octave, and only a single one exceeded this value.
In the following analysis step, the goal was to analyze similarity in the monaural spectrotemporal patterns encoded by each second-layer unit. To this end binaural similarity index (BSI) of each amplitude basis function [43] was computed. The BSI is a correlation coefficient between the left and the right parts of a binaural, spectrotemporal feature. If the BSI was close to 0, the corresponding unit was representing different spectrotemporal patterns in each ear, while values close to 1 implied high similarity. BSIs are plotted in Fig 6A.
Clearly, an overwhelming majority of basis functions revealed high interaural similarity (BSI > 0.8, see the histogram at the inset). BSI of only one basis function was slightly below 0. If information encoded by amplitude basis functions in each ear was independent, the BSI distribution should peak at 0. This observation suggests that most of the second-layer units captured the same “cause” underlying the stimulus i.e. a binaurally redundant spectrotemporal pattern. While the BSI index measures similarity of encoded monaural sound features, it is not informative about the side-preference of each unit. To asess whether amplitude basis functions were biased more towards the left or towards the right ear, another statistic — a binaural amplitude dominance (BAD) was computed. The amplitude dominance was defined in the following way:
B A D ( B i ) = log ( ∥ exp ( B i , L ) ∥ ∥ exp ( B i , R ) ∥ ) (17)
where Bi, L = Bi, (1, …, T), Bi, R = Bi, (T+1, …,2×T) are left and right ear parts of an amplitude basis function Bi. Each of them was pointwise exponentiated to map the entries from real log-amplitude values to the positive amplitude domain. The BAD index value larger than 0 means that the left-ear amplitude vector had a larger norm (i.e., it dominated the input to the particular unit). Balanced units had a BAD value close to 0 while right-ear dominance was indicated by negative values. Two histograms of dominance scores are displayed in panel B of Fig 6. The black one is an empirical distribution of BAD values of amplitude basis functions associated with IPD features of a negative average value (left-side preferring). The gray one corresponds to amplitude features matched with right-side biased phase basis functions. Both distributions are roughly symmetric with their modes located quite close to 0. Such bimodal distribution of the amplitude dominance score implies that amplitude basis functions could be divided into two opposite populations — each preferring input from a different ear. Moreover, amplitude and phase information modelled by basis functions Bi and ξi was dependent — amplitude features dominated by information from one ear were associated with IPD features biased towards the same ear.
While amplitude representation encoded the quality of the sound together with binaural differences, the IPD dictionary was representing solely spatial aspects of the stimulus i.e. the temporal difference between the ears. In almost entire feature population, single entries of each of the phase difference basis functions ξi all had the same sign. Negative phase differences corresponded to the left-side bias (it meant that the soundwave arrived first to the left-ear generating a smaller phase value) and positive to the right-side one. These two properties allowed us to asess the spatial preference of IPD basis functions simply by computing the average of their entries. The histogram of averages of vectors ξi (normalized to have the maximal absolute value of 1) is depicted in Fig 6C. A clear bimodality is visible in the distribution. The positive peak corresponds to right-sided basis functions and the negative one to the left-sided subpopulation. Almost no balanced features (close to 0) were present in the dictionary. This dichotomy is visible also in Fig 4—binaurally balanced amplitude basis functions (middle column) were associated with phase vectors biased towards either side. This result may be related to a previous study, which showed that a representation of natural IPD distribution designed to maximize stimulus discriminability (Fisher information) also has a form of two distinct channels [45], where each of the channels preferred IPDs of an opposite sign.
The second layer of the model learned a distributed representation of sound features accesible to neurons in the auditory cortex. Assuming that the cortical auditory code indeed develops driven by principles of efficiency and sparsity, one can interpret second layer basis functions as neuronal receptive fields and sparse coefficients s as a measure of neuronal activity (e.g. firing rates). The model can be then probed using spatial auditory stimuli. If it indeed provides an approximation to real neuronal computations, its responses should be comparable with spatial tuning properties of the auditory cortex.
In order to verify whether this was true, a test recording was performed. As a test sound the hiss of two pieces of paper rubbed against each other was used. It was a broadband signal, reminiscent of white noise used in physiological experiments, yet posessing natural structure. Recording was performed in an anechoic chamber, where a person walked around the recording subject while rubbing two pieces of paper (see Methods for a detailed description). The recording was divided into 18 windows, each corresponding to a 20 degree part of a full circle. The number of windows was selected to match experimental parameters in [8, 10]. From each window 3000 epochs were drawn and each of them was encoded using the model. Computing histograms of coefficients s at each angular position θ, provided an estimate of conditional distributions p(si∣θ). Panel A in Fig 7 displays a conditional histogram of coefficient s corresponding to the basis function pair depicted in Fig 4A.
Distributions of sparse coefficients revealed a strong dependence in the position of the sound source. As visible in the figure, the conditional mean of the distribution p(si∣θ) traced by the red line varied in a pronounced way across all positions. By analogy to averaged firing rates of neurons, average unit responses at each position were further studied to understand the spatial sensitivity of basis functions. Mean vectors μi, θ were constructed for each second-layer unit by taking its average response at the sound source position θ. Each mean vector was shifted and scaled such that its minimum value was equal to 0 and the maximum to 1. Such transformation was analogical to physiological studies [8] and allowed for comparison with experimetally measured spatial tuning curves of auditory neurons, and for this reason scaled vectors μi will be referred to as model tuning curves in the remainder of the paper. In order to identify spatial tuning preferences, the population of model tuning curves was grouped into two clusters using the k-means algorithm. Obtained clusters consisted of 118 and 138 similar vectors. Tuning curves belonging to both clusters and revealing a strong correlation (∣ρ∣ > 0.75) with sound position are plotted in Fig 7C as gray lines. Cluster centroids (averages of all tuning curves belonging to a cluster) are plotted in black. Second layer units were tuned broadly—most of them were modulated by sound located at all positions surrounding the subject’s head. A clear spatial preference is visible—members of cluster 1 were most highly activated (on average) by sounds localized close to the left ear (θ ≈ −90°), while cluster 2 consisted of units tuned to the right ear (θ ≈ 90°). Very similar tuning properties of auditory neurons were identified in the cat’s auditory cortex [8]. Data from this study is plotted for comparison in the subfigure B of Fig 7. Neuronal recordings were performed in the right hemisphere and two panels depict two subpopulations of neurons. The larger contra- and the smaller ipsi-lateral one. It is important to note, that the notion of ipsi, and contra laterality is not meaningful in the proposed model, therefore one should compare shapes of model and experimental tuning curves, not the numerosity of units in each population or cluster.
Two major features of cortical auditory neurons responsive to sound position were observed experimentally: (i) tuning curve peaks were localized mostly at extremely lateral positions (opposite to each ear) and (ii) slopes of tuning curves were steepest close to the auditory midline. Both properties are visible in model tuning curves in Fig 7. However, in order to perform a more direct comparison between the model and experimental data, analysis analogous to the one described in [8] was performed. First, tuning curve centroids were computed. A centroid was defined as an average position, where the unit activation was equal to 0.75 or larger (see Methods). In the following step, the position of maximal slope towards midline was identified for each unit. This meant that for units tuned to the left hemifield (cluster 1) the position of the minimal slope value was taken, while the position of the maximal one was taken for units tuned to the right hemifield (cluster 2). In this way, the position of maximal sensitivity to changes in sound location was identified. Distributions of model centroids and maximal slope positions are depicted in Fig 8B. Centroids were distributed close to lateral positions, opposite in each cluster (−90° cluster 1, +90° cluster 2). Distribution peaks were located at positions close to each ear. No uniform tiling of the space by centroid values was present. At the same time, maximal slope values were tightly packed around the midline—peaks of their distributions were located precisely at, or very close to 0 degrees. This means that while the maximal response was on average triggered by lateral stimuli, the largest changes were triggered by sounds located close to the midline. Both properties were in good agreement with the experimental data reported in [8]. Fig 8A depicts in three panels centroid and slopes distributions measured in three different regions of cat’s auditory cortex—Primary Auditory Field (A1), Posterior Auditory Field (PAF) and Dorsal Zone (DZ). A close resemblance between the model and physiological data was visible.
It has been argued that while single neurons in the auditory cortex provide coarse spatial information, their populations form a distributed code for sound localization [8, 9, 9, 10]. Here, a decoding analysis was performed to verify whether similar statement can be made about the proposed model.
A gaussian mixture model (GMM) was utilized as a decoder. The GMM modelled the marginal distribution of sparse coefficients as a linear combination of 18 gaussian components, each corresponding to a particular position of a sound source (i.e. the θ value). In the first part of the decoding analysis, single coefficients were used to identify the sound position. The GMM was fitted using the training dataset consisting of coefficient values si and associated position labels θ. In the testing stage, position estimates θ ^ were estimated (decoded) using unlabeled coefficients from the test dataset (see Methods section for a detailed description of the decoding procedure). For each of the coefficients, a confusion matrix was computed. A confusion matrix is a two-dimensional histogram of θ and θ ^ and can be understood as an estimate of the joint probability distribution of these two variables. Using a confusion matrix, an estimate of mutual information (i.e., the number of bits shared between the position estimate θ ^ and its actual value θ) was obtained. Fig 9A depicts histograms of information carried by each coefficient si about the sound source position, estimated as described above. A general observation is that single coefficients carried very little information about the sound location. The histogram peaks at a value close to 0.1 bits. Only few units coded approximately 1 bit of positional information. Even 1 bit, however, suffices merely to identify a hemifield, not to mention the precise sound position. As can be predicted from the broad shapes of the tuning curves, single second-layer units carried little spatial information. A similar result was obtained for neurons in different areas of the cats auditory cortex [12]. The amount of information about the sound position encoded by spike count of neurons in A1 and PAF regions has a distribution closely similar to that of model units (compare with the left panel of figure 11 in [46]). Spike count (which essentially corresponds to a firing rate) is a feature of a neuronal response most directly corresponding to coefficients s in the model described here. The median of mutual information estimated from model coefficients (marked by a diamond symbol in panel A) aligns well with the same quantity estimated from neuronal data, and is close to 0.2 bits [46]. Overall, physiological measurements and the behavior of the model were highly similar.
While single neurons did not carry much spatial information, the joint population activity was sufficient to decode the sound position [8–10, 46]. Therefore in the second step of the decoding analysis, multiple coefficients s were used to train and test the GMM decoder. Results of the population decoding are plotted in Fig 9B. The decoder was trained with a progressively larger number of second-layer units (from 1 to 256) and the mutual information was estimated from obtained confusion matrices. Each line in the plot depicts the number of bits as a function of the number of units used to perform decoding. Line colors correspond to the number of samples over which the average activity was computed. Broadly speaking, larger populations of second-layer units allowed for a more precise position decoding. As in the case of single units, averages over larger amounts of samples were also more informative—population activity averaged over 32 samples saturated amount of bits required to perform errorless decoding (4.17). Two confusion matrices obtained from raw population activity and an average over 16 samples are displayed in subfigures Fig 9C and 9D. In the former case, the decoder was mostly misclassifying sound positions within each hemifield. Averaging over 16 sound samples yielded an almost diagonal (errorless) confusion matrix. The decoding analysis allowed us to draw the conclusion that while single units carried very little spatial information, their population encoded source location accurately, consistent with experimental data.
Second layer units achieved spatial tuning by assigning different weights to amplitudes in each ear, and to IPD values in different frequency channels. At the same time they encoded spectrotemporal features of sound, as depicted in Fig 4. Their activity should therefore be modulated by both sound position as well as its quality. Such comodulation is a prominent feature of the majority of cortical auditory neurons [1, 7]. In order to verify this, model spatial tuning curves were estimated with a second sound source, very different from a hiss created by rubbing paper—human speech (see Materials and Methods for details). Frequency spectra of both test stimuli are depicted in Fig 10D. Test sounds distributed their energy over non-overlapping parts of the frequency spectrum. While speech consisted mostly of harmonic peaks below 1.5 kHz, the paper sound was much more broadband and its energy was uniformely distributed between 1.5 and 4 kHz.
Panels A-C of Fig 10 depict three amplitude/IPD basis function pairs together with their spatial tuning curves estimated using different sounds. The spatial preference of depicted units (left or right hemifield) was predictable from their binaural composition. Each of them, however, was activated stronger by a stimulus, which spectrum matched better amplitude basis functions. Basis functions visible in panels A and C had a lot of energy accumulated in higher frequencies, therefore the paper sound activated them stronger (on average). Basis function B) was spectrally better corresponding to speech sounds, therefore speech was a preferred class of stimuli. This observation suggests that tuning curves i.e. position-conditional means μi, θ should be understood not as averages of coefficient ensembles conditioned only on the sound position θ but also on spectral properties of sound. When interpreting coefficients s as neuronal activity this means that spatial tuning curves would alter their shapes when the neuron is tested with two different sound sources. Taken together, one can state that the second-layer representation encoded position and identity of the stimulus in an interdependent fashion.
Previously proposed statistical models of natural acoustic stimuli focused predominantly on monaural sounds [22–25, 30, 38]. Studies modelling binaural stimuli were constrained to a limited representation—either IPDs [45] or spectrograms [31]. In contrast, the assumption behind the present work was that spatial sensitivity of cortical neurons is formed by fusing different cues. Therefore, in order to understand the role played by the auditory cortex in spatial hearing, the entire natural input processed by the auditory system was analyzed.
To this end, a novel probabilistic model of natural stereo sounds has been proposed. The model is based on principles of sparse, efficient coding—its task was to learn progressively less redundant representations of natural signal. It consisted of two hidden layers, each of them could be interpreted as an analogy to different stages of sound processing in the nervous system. The purpose of the first layer was to form a sparse, non-redundant representation of natural sound in each ear. By analogy to the cochlea, the encoding was supposed to extract and separate temporal information (i.e. phase) from the amplitude of the signal. In order to do so, a dictionary of complex-valued basis functions was adapted to short sound epochs. On top of the first model layer, which encoded sound in each ear independently, the second layer was trained. Its goal was to encode jointly amplitude and phase—two kinds of information crucial for sound localization, which may be fused together in higher stages of the auditory system. The higher-order representation captured spectrotemporal composition of the signal, by learning amplitude patterns of the first layer output as well as interaural disparities present in form of interaural phase and amplitude differences. It is important to stress that the model was fully unsupervised—at no point information about positions of sounds sources or the spatial configuration of the environment was accessible. Yet, when tested with a set of spatial sounds, activity of second layer units revealed strong dependence on sound position. Tuning curves describing relation between the sound position and model activity were in good correspondence with experimentally measured spatial tuning properties of cortical auditory neurons.
In mammals, the location of a sound is encoded by two populations of broadly tuned, spatially non-specific units [32]. This finding challenges initial expectations of finding a “labelled-line code” (i.e. a topographic map of neurons narrowly tuned to small areas of space). The “spatiotopic map” was expected by analogy to the tonotopic structure of the cortex, as well as the high localisation accuracy of humans and animals. Instead, it has been found that auditory cortical neurons within each hemisphere are predominantly tuned to far, contralateral positions. Peaks of observed tuning curves did not tile the auditory space uniformly, rather they were clustered around the two lateral positions. A prominent observed feature of cortical representation of sound location were slopes of the tuning curves. Regardless of the position of the tuning curve peak, slopes were steepest close to the interaural midline—the area where behavioral localisation acuity is highest [32]. From described observations, two prominent conclusions were drawn. Firstly, that the slope of tuning curves, not the distribution of their peaks determines spatial acuity [8, 32, 47, 48]. Secondly that sound position is encoded by distributed patterns of population activity, not single neurons [8–10]. It has been argued that these properties are a manifestation of a coding mechanism which evolved to specifically meet the demand of binaural hearing tasks [8, 32]. Here it is shown that crucial properties of cortical spatial tuning emerge in an unsupervised learning model, which learns a sparse representation of natural binaural sounds. The objective of the model was to code the stimulus efficiently (i.e. with a minimal redundancy within limits of applied transformations), while minimizing unit activity. Properties of the learned representation are therefore a reflection of stimulus statistics, not of any task-specific coding strategy (required for instance to localize sounds with the highest accuracy at the midline).
The position of the sound-generating object is a latent variable for the auditory system. It means that its value is not explicitly present in the raw stimulus—it has to be estimated. This estimation, (or inference) is a non-trivial task in the real acoustic environment, where sounds reaching ear membranes are a reflection of intricate auditory scenes. Sensory neurons perform transformations of those sound waveforms to reconstruct the spatial configuration of the scene. Therefore, in an attempt to understand cortical representation of space, it may be helpful to think what is the statistical structure of the naturally encountered binaural stimulus that the auditory system operates on. Sounds reaching the ear contain information about their generating sources, the spatial configuration of the scene, position and motion of the organism and the geometry of its head and outer ears.
Results obtained here suggest that the shapes of the model spatial tuning curves reflect regularities imposed on the sensory data by the filtering properties of the head. At lateral positions (directly next to the left or the right ear) there is no acoustic attenuation by the skull, hence sounds are loudest and least delayed. This in turn, elicits the strongest response in units preferring that side. When the sound is at a contralateral position, response is much weaker, due to the maximal head attenuation and largest delay. The curve connecting those two extrema is steepest in the transition area—at the midline. Since the auditory environment was uniformly sampled at both sides of the head, model units were clustered into two roughly equal subpopulations, basing on the shapes of their tuning curves. Clusters were symmetric with respect to each other—one tuned to to the left and the other to the right hemifield. This groupping is reminiscent of the “opponent-channel” representation of the auditory space, which has been postulated before [8, 32]. Present results provide a theoretical interpretation of this tuning pattern. They suggest that neuronal population which forms a sparse, efficient representation of natural stimuli would reveal two broadly tunned channels, when probed with sounds located at different positions.
It has been shown previously that IPD coding strategies in different species can be predicted from statistics of binaural sound [45]. Harper and McAlpine demonstrated that if the goal of the nervous system is to represent IPD values with the maximal possible accuracy (quantified by Fisher information) two populations of neurons tuned to opposite locations constitute an optimal representation of low-frequency IPDs. Their approach differs significantly from the one presented here. On the most abstract level, the authors of [45] assume that the purpose of IPD sensitive neurons is to maximize Fisher information, while here mutual information is the quantity implicitly maximized by the representation (although interesting relationships exist between those two measures [49]). Secondly, Harper and McAlpine limit their analysis to IPD statistics only—here entire binaural waveforms are modelled. Finally the current study does not assume any parameteric shape of tuning curves, nor make any other assumptions about physiology as is the case in [45]. The similarity of model responses and neuronal activity emerges from data statistics.
There is an ongoing debate about the presence (or lack of thereof) of two-separate “what” and “where” streams in the auditory cortex [5]. The streams would separate spatial information from other sound features which determine its identity. An important prediction formed by this dual-stream hypothesis is that there should exist neurons selective to sound position and invariant to other aspects in the auditory cortex. While some evidence has been found supporting this notion [3, 4] it seems that at least in vast parts of the auditory cortex neural activity can be modulated by multiple features of sound such as pitch, timbre and location [1]. Neurons are sensitive to sound position (i.e. changing position affects their firing patterns), but not selective nor invariant to it. The majority of studies analyzing spatial sensitivity in the auditory cortex use a single class of sound and the source position is the only varying parameter. Therefore, despite initial efforts, the influence jointly exerted by sound quality and position on neuronal activity is not yet well understood.
The statistical model proposed here suggests that no dissociation of spatial and non-spatial information is necessary to either reconstruct the sound source or identify its position. The learned second-layer representation carries both kinds of information—about the sound quality (contained in the spectrotemporal structure of basis functions) and about spatial aspects (contained in the binaural amplitude weighting and IPD vectors). The learned code forms a “what is where” representation of the stimulus (i.e., those two aspects are represented interdependently). A manifestation of this fact is visible in different scaling of spatial tuning curves, when probed with two different sound sources. Such comodulation of neuronal activity by sound position and quality has been observed experimentally [1], which may suggest that recorded neurons form a sparse, efficient representation of binaural sound. An advantage of an interdependent “what is where” representation is the absence of the “feature binding problem”, which has to be solved if spatial information is processed independently. After separating the location of a source from its identity they would have to be fused at processing stages beyond the auditory cortex. A code similar to the one described here does not create such a problem. This idea goes in hand with results of a recent perceptual study [50]. Parise et al. demonstrated that the perception of sound source elevation is strongly influenced by its frequency. Furthermore they show that this relationship can be explained by adaptation to the joint distribution of natural sounds’ positions and spectra. This implies that the quality of the sound source as well as its spatial position are mutually dependent, and as such should be represented jointly, if the goal of the nervous system is to increase coding efficiency.
The model proposed in this work is a statistical one—it constitutes an attempt to describe functional, not anatomical modules of the auditory system. Rather than explicitly modelling stages of the auditory pathway, its goal is to approximate the distribution of natural binaural sounds. The behaviour of units in the highest layer reveals a strong resemblance to cortical auditory neurons in an abstract, information processing domain. In the mammalian auditory system the sound is processed in at least five anatomical structures before it reaches the cortex [51]. It is therefore almost certain that the stimulus is subjected to many more complex transformations than the ones proposed here. On the other hand, the fact that similiarities between cortical and model responses emerge despite this lack of detail, imply that the model may be capturing some aspects of information processing, as it happens in the real auditory system.
The relationship between abstract computational principles such as sparse coding and neurophysiology is an area of ongoing research [52–54]. An interesting extension of the present work would attempt to increase the level of biological detail, and see whether this allows formation of more refined experimental predictions. This could be done by implementing sparse coding computations using spiking neuron models, as it has been done in studies of the visual system (e.g. [52, 54]). The match between the model and biology could be also improved by including phenomena specific to the auditory system, such as the phase locking in the auditory nerve.
This study focuses predominantly on explaining the broad spatial tuning of cortical auditory neurons estimated by the analysis of firing rates. With progressively larger amounts of biological detail added to the model, one could attempt to explain other aspects of spatial information encoding. For instance, the notion of spike timing does not exist in the approach proposed here, while temporal spike patterns of cortical neurons seem to carry relevant spatial information [9, 10, 46]. Moreover, as mentioned in the results section, the concept of contra- and ipsilaterality is spurious for high-layer model units since they are not associated with any anatomical locus (left or right hemisphere). Overrepresentation of the contralateral ear is an interesting feature of panoramic population codes [8], which is also not addressed by the present work. Further exploration of the relationship between specific biological observations and spatial information processing constitutes a possible goal for future research.
It is highly likely that the main result of this study (i.e., spatial tuning properties of the binaural sound representation) could be reproduced by replacing the first layer with a different sort of spectrotemporal signal representation. It would not necessarily have to be the sparse, efficient encoding of sound epochs. A spectrogram could be a candidate signal, although it has been demonstrated that a sparse code of relatively long binaural spectrogram chunks generates features of very different spatial tuning [31]. In this work, for the sake of theoretical consistency, both layers were learned using the same principles and statistical assumptions—sparse factorial coding.
The data used for comparisons originated from studies of cat auditory cortex ([8, 46]). Since statistics of the binaural signal are affected by the geometry of ears and the head of the organisms, one could argue that model trained on binaural recordings performed by a human should not be compared with cat physiology. As long as detailed features of neuronal tuning to a sound position may vary across those species, tuning patterns highly similar to those of the cat have been observed in the auditory cortex of primates [55, 56]. Overall, the cortical representation of sound position seems to be highly similar across mammals [32].
Finally, in the current study a binaural recording of only a single auditory scene was used to train the model. Even though the recording included many types of sound—ambient environmental noises, transient cracks and clicks and harmonic structures such as the human speech, it did not include many other possible sources (for instance animal vocalizations). The recording included also only a narrow range of other parameters which characterize natural auditory scenes, such as reverberation. Analysis of longer recordings performed in different environmental settings may generate more diverse results and additional insights. One should note however, that certain properties of the learned representation (such as the tradeoff in the spectrotemporal modulation) seem to be a general proprerty of natural sounds as such and remain invariant to a specific dataset [25, 40]. Basing on this observation one may expect that units revealing similar spatial tuning can be learned from recordings of numerous, diverse sets of natural sounds.
Taken together, this paper proposes a candidate theoretical mechanism explaining how neurons in the auditory cortex represent spatial information. This model allows us to speculate they do not have to implement any task-dependent strategy. Instead, their behavior can be explained by sparse coding—a statistical model which has succesfully predicted properties of multiple other sensory systems [18, 21]. Taking a broad perspective, (as suggested by Barlow in his later work [57, 58]) this means that redundancy reduction by sparse coding can be used by the brain to identify sensory data patterns allowing sucesful interaction with the environment.
Sound recordings received approval of the Ethics Council of the Max-Planck Society. Human participants provided a written consent to participate in recordings.
Sounds used to train and test the model were recorded using Soundman OKM-II binaural microphones placed in the ear channels of a human subject, whose head circumference was 60 cm. While recording training sounds, the subject walked freely in a wooded area accompanied by another person who spoke rarely. In this way, collected data included transient and ambient environmental sounds as well as harmonic speech. The binaural composition of sound was affected by spatial configuration of the environment and motion patterns of the recording subject. The recording used to train the model was 60 seconds long in total. Binaural recordings are availible in the supplementary material of [59].
Test recordings used to map the spatial tuning of second-layer units was performed in an anechoic chamber at the Department of Biology, University of Leipzig. The same recording subject was seated in the middle of the chamber. A female speaker walked at a constant pace following a circular path surrounding the recording subject. While walking she counted out loud. This was repeated four times. The second test recording was performed in a similar fashion, however instead of speaking the walking person rubbed two pieces of cardboard against each other, generating a broadband sound. To estimate the conditional distribution of sparse coefficients given the position and identity of the sound, test recordings were divided into 18 intervals, each corresponding to the same position on a circle.
All recordings were registered in an uncompressed wave format at 44100 Hz sampling rate. Prior to training the model, sounds were downsampled to 8000 Hz. Test recordings are availible in the supplementary material (S1, S2, S3, S4, S5, S6, S7, S8 Files).
The goal of the learning procedure was to estimate first- (A), and second- layer basis functions (B, ξ). This was done using a two-step approach. Firstly maximum a posteriori (MAP) estimates of model coefficients (z in the first layer, s and w in the second) were inferred via gradient descent [18, 33]. Secondly, a gradient update on basis functions was perormed using current coefficient estimates. Those two steps were consecutively iterated until the model converged.
A dictionary of complex-basis functions in the first layer was created by first, training a standard sparse code of sound epochs x ∈ ℝT:
x t = ∑ i = 1 T c i Θ i , t + η (18)
The negative log-posterior of this model was:
E s ( x , c , Θ ) ∝ 1 σ 2 ∑ t = 1 T ( x t - x ^ t s ) 2 + λ ∑ i = 1 T S ( c i ) (19)
where x ^ t s = ∑ i = 1 T c i Θ i , t is the reconstruction of the data vector. Corresponding gradients over linear coefficients c and basis functions Θ were given by:
∂ ∂ c i E s ∝ - 2 σ 2 ∑ j = 1 T Θ j , t ( x t - x ^ t s ) + 2 λ c i log ( 1 + c i 2 ) (20) ∂ ∂ Θ i , t E s ∝ - 2 σ 2 ∑ t = 1 T c i ( x t - x ^ t s ) (21)
Learned basis functions Θi were used as real vectors A i ℜ and extended with their Hilbert transforms. Such complex basis function dictionary was used to encode monaural sound epochs. Gradients of Eq 5 over phase ϕi and amplitudes ai of complex coefficients zi were equal to:
∂ ∂ a i E 1 ∝ - 2 σ 2 ∑ t = 1 T ( cos ϕ i A i , t ℜ + sin ϕ i A i , t ℑ ) ( x t - x ^ t ) + 2 λ a i log ( 1 + a i 2 ) (22) ∂ ∂ ϕ i E 1 ∝ - 2 σ 2 ∑ t = 1 T a i ( A i , t ℑ cos ϕ i A i , t ℑ - A i , t ℜ sin ϕ i A i , t ℜ ) ( x t - x ^ t ) (23)
The second layer of the model was trained after the first layer converged, and cofficient values z were inferred for all training data samples. The higher order encoding formed by coefficients s as well as the scaling factor w was inferred via gradient descent on function E2 (Eq 13):
∂ ∂ s i E 2 ∝ - 2 σ 2 2 ∑ n = 1 2 × T B i , n ( a n - a ^ n ) + κ | w | ∑ m = 1 P sin ( Δ ϕ m - Δ ϕ ^ m ) ξ i , m + 2 λ 2 s i log ( 1 + s i 2 ) (24) ∂ ∂ w i E 2 ∝ κ w | w | 2 ∑ m = 1 P Δ ϕ ^ m sin ( Δ ϕ m - Δ ϕ ^ m ) + λ w [ ( 1 α ) β β w | w | β - 2 ] (25)
The gradients steered sparse coefficients s to explain amplitude and phase vectors a and Δϕ while preserving maximal sparsity. Simultaneously the multiplicative factor w was adjusted to appropriately scale the estimated vector Δ ϕ ̂.
Finally, learning rules for second-layer dictionaries were given by:
∂ ∂ B i , k E 2 ∝ - 2 σ 2 2 s i ( a k - a ^ k ) (26) ∂ ∂ ξ i , k E 2 ∝ s i κ | w | sin ( Δ ϕ k - Δ ϕ ^ k ) (27)
Altogether 75000 epochs of binaural sound were used to train the model. Each of them was T = 128 samples long, which corresponded to 16 ms. Both layers were trained separately. Before training the first layer, Principal Component Analyis was perfomed and 18 out of 128 principal components were rejected, which corresponded to low pass filtering the data. Left and right ear sound epochs were shuffled together to create a 150000 sample training dataset for the first layer. The first layer sparsity coefficient λ was set to 0.2. Noise variance σ2 was equal to 2. The sparse coding algorithm converged after 200000 iterations.
A complex-valued dictionary was created by extending the real valued one with Hilbert-transformed basis functions. Amplitude and phase vectors a and ϕ were inferred for each sample using 20 gradient steps. Amplitude vectors were concatenated and transformed with a logarithmic function, and IPD vectors Δϕ were computed by substracting left ear phase vectors ϕL from right ear ones ϕR. The second layer was trained by performing 250000 gradient updates on basis functions B and ξ. The amplitude sparsity coefficient λ2 was set to 1. The λw parameter was set to 0.01 and the noise variance σ 2 2 as well as the von Mises concentration parameter κ were set to 2.
Numerical values of the prior-controlling parameters λ, λ2, λw as well as noise parameters σ, σ2, κ were set empirically in this study. By running simulations with multiple parameter settings it has been found that due to the presence of a strong environmental noise in the training recording, noise variances σ, σ2 and the von Mises concentration parameter κ should be relatively large in order to achieve convergence. Sparsity of the high layer representation was set to be larger than that of the first layer in order to mimic the biological intuition that neural responses in the ascending auditory pathway become progressively less redundant and sparser [20, 60]. It has been found however, that the exact value of sparsity paramaters did not affect the spectrotemporal properties, nor the spatial tuning of the second layer units strongly. The λw parameter which controls the strength of the prior over the multiplicative factor w was set to be relatively small. Otherwise the w prior term in the Eq 16 became too strong and dominated learning, preventing the convergence. More principled and theoretically sound ways of parameter selection are possible. One could ask what are the natural noise levels and sparsity values of the training data by specifying them as hyperparameters of the model and learning the appropriate values. Also the number of basis functions at each level could be treated as a parameter and estimated from the data, not chosen ad-hoc. After extending the model in this way, the choice of the correct parameter setting could be performed by cross-validation or Bayesian model selection (as in [61]).
Spectrograms of amplitude basis functions Bi were computed by combining spectrograms of real, first layer basis functions A n ℜ, linearly weighted by a corresponding weight exp(Bi, n). First layer spectrograms were computed using T = 29 windows, each 16 samples (0.002 second) long, with a 12 sample overlap. Altogether, F = 128 logarithmically-spaced frequencies were sampled. A two-dimensional fourier transform of each spectrogram was computed using the matlab built-in function fft2. The amplitude spectrum of obtained transform is called the Modulation Transfer Function (MTF) of each second layer feature [40]. The center of mass i.e. the point ( C S , i f , C S , i t ) of each monaural part (S ∈ {L, R}) of basis functions Bi was computed in the following way:
C S i t = ∑ t t ∑ f M T F ( B S , i ) (28) C S i f = ∑ f f ∑ t M T F ( B S , i ) (29)
where t and f are time and frequency respectively.
To estimate conditional distribution of sparse coefficients given the position and identity of the sound, test recordings of a sound source (either speech, or rubbed paper) moving around the recording subject were used. Each source circled the recording person 4 times resulting in 4 recordings. Each of them was divided into 18 intervals. Intervals corresponding to the same area on the circle were joined together across all recordings. For each out of 18 sound positions 3000 random sound chunks were drawn and encoded by the model. Position-conditional ensembles were then used to compute conditional histograms. Conditional mean vectors μi, θ were computed by averaging all values of coefficient si at position θ. Mean vectors were mapped to a [0, 1] interval by adding the absolute value of a minimal entry and dividing it by the value of the maximum. For plotting purposes in Fig 10, endings of tuning curves were connected if values at −180° and 180° were not exactly equal.
The decoding analysis was performed using K second-layer sparse coefficients s averaged over D of samples. The response vectors d ∈ ℝK were therefore formed as:
d = 1 D ∑ i = 1 D s { 1 , … , K } (30)
Such averaging procedure can be interpreted as an analogy to computation of firing rates in real neurons.
The marginal distribution of response coefficients d over all 18 sound positions θ ∈ {−180°, −160°, …,160°,180°} was equal to:
p ( d ) = ∑ θ p ( d | θ ) p ( θ ) (31)
where each conditional p(d∣θ) was a K-dimensional Gaussian distribution with class specific mean vector μθ and covariance matrix Cθ:
p ( d | θ ) = 𝓝 ( μ θ , C θ ) (32)
The prior over class labels p(θ) was uniformly distributed i.e. p ( θ i ) = 1 18 for each i.
The decoding procedure iterated over all class labels and returned the one, which maximized the likelihood of the observed data vector. Out of the entire dataset, 80% was used to train the model and remaining 20% to test and estimate the confusion matrix.
Confusion matrix M was a joint histogram of a decoded and true sound position θ ^ and θ. After normalization, it was an estimate of a joint probability mass function p ( θ ^ , θ ). Mutual information was estimated from each confusion matrix as:
M I ( θ ^ θ ) = ∑ θ ^ ∑ θ p ( θ ^ , θ ) log 2 ( p ( θ ^ , θ ) p ( θ ^ ) p ( θ ) ) (33)
|
10.1371/journal.pbio.1000352 | An Excitatory Loop with Astrocytes Contributes to Drive Neurons to Seizure Threshold | Seizures in focal epilepsies are sustained by a highly synchronous neuronal discharge that arises at restricted brain sites and subsequently spreads to large portions of the brain. Despite intense experimental research in this field, the earlier cellular events that initiate and sustain a focal seizure are still not well defined. Their identification is central to understand the pathophysiology of focal epilepsies and to develop new pharmacological therapies for drug-resistant forms of epilepsy. The prominent involvement of astrocytes in ictogenesis was recently proposed. We test here whether a cooperation between astrocytes and neurons is a prerequisite to support ictal (seizure-like) and interictal epileptiform events. Simultaneous patch-clamp recording and Ca2+ imaging techniques were performed in a new in vitro model of focal seizures induced by local applications of N-methyl-D-aspartic acid (NMDA) in rat entorhinal cortex slices. We found that a Ca2+ elevation in astrocytes correlates with both the initial development and the maintenance of a focal, seizure-like discharge. A delayed astrocyte activation during ictal discharges was also observed in other models (including the whole in vitro isolated guinea pig brain) in which the site of generation of seizure activity cannot be precisely monitored. In contrast, interictal discharges were not associated with Ca2+ changes in astrocytes. Selective inhibition or stimulation of astrocyte Ca2+ signalling blocked or enhanced, respectively, ictal discharges, but did not affect interictal discharge generation. Our data reveal that neurons engage astrocytes in a recurrent excitatory loop (possibly involving gliotransmission) that promotes seizure ignition and sustains the ictal discharge. This neuron–astrocyte interaction may represent a novel target to develop effective therapeutic strategies to control seizures.
| In focal epilepsy, seizures are generated by a localized, synchronous neuronal electrical discharge that may spread to large portions of the brain. Despite intense experimental research in this field, a key question relevant to the human epilepsy condition remains completely unanswered: what are the cellular events that lead to the onset of a seizure in the first place? In various in vitro models of seizures using rodent brain slices, we simultaneously recorded neuronal firing and Ca2+ signals both from neurons and from astrocytes, the principal population of glial cells in the brain. We found that activation of astrocytes by neuronal activity and signalling from astrocytes back to neurons contribute to the initiation of a focal seizure. This reciprocal excitatory loop between neurons and astrocytes represents a new mechanism in the pathophysiology of epilepsy that should be considered by those aiming to develop more effective therapies for epilepsies that are not controlled by currently available treatments.
| Focal epilepsies are characterized by a condition of neuronal hyperexcitability that is restricted to the epileptogenic region. Focal seizures originate at this region and secondarily spread to distant cortical areas [1]–[5]. Several factors, from ion channel mutations to brain injury, may cause neuronal hyperexcitability changes that sustain an epileptic condition [6]. Yet, the earlier cellular events that initiate a seizure in the first place are still unclear. The understanding of ictogenesis is thus central to the pathophysiology of focal epilepsies and is a requirement to develop new pharmacological therapies for drug-resistant focal epilepsies [7].
In the present study, we specifically address the hypothesis that the activation of a loop between neurons and astrocytes is an early event that contributes to focal seizure initiation. This hypothesis stems from a series of recent studies that reappraised the role of neurons in epileptogenesis and hinted at a possible, direct contribution of astrocytes to the generation of an epileptic discharge. The first clue was the observation that the release of glutamate from astrocytes, elicited by Ca2+ oscillations, promotes local synchronous activities in hippocampal neurons by acting on extrasynaptic N-methyl-D-aspartic acid (NMDA) receptors [8]. Studies performed both on brain slices and in vivo showed that during epileptiform activity, the frequency of Ca2+ oscillations in astrocytes is significantly increased [9],[10], and it is reduced by anticonvulsant drugs [9]. Moreover, the expression of metabotropic glutamate receptors (mGluRs, mediators of Ca2+ oscillations in these cells) in hippocampal astrocytes from animal models of temporal lobe epilepsy was found to be increased [11],[12]. These observations suggest that the excessive neuronal synchronization that characterizes the epileptic discharge might be sustained, at least in part, by an astrocyte hyperactivity. In support of an astrocyte role in epileptiform activities, it has been proposed that the interictal events recorded between seizures might be in some conditions tetrodotoxin (TTX)-resistant and mediated by glutamate release from astrocytes [9]. These findings fuelled a controversial debate on the role of astrocytes in focal epileptogenesis and in the generation of epileptiform discharges [13]–[15].
In the present study, we used different models of epileptic seizures, including a new model of focal seizures, to define the role of astrocytes in the generation of epileptiform activities. We performed simultaneous Ca2+ imaging and electrophysiological recordings of epileptic discharges in brain slices and in isolated intact guinea pig brains, focusing on the entorhinal cortex. This experimental approach allowed us to define the timing of astrocyte Ca2+ excitability in relation to interictal and ictal discharges. By using different pharmacological tools to affect selectively the Ca2+ signal in astrocytes, we also investigated a possible causative role of astrocyte activation in the generation of these epileptic discharges.
We demonstrate here that a recurrent excitatory loop between neurons and astrocytes involving Ca2+ elevations in a large number of glial cells is an early event that contributes to the initiation of a focal seizure-like discharge.
The activation of astrocytes by neuronal activity is mainly mediated by synaptic neurotransmitter release, such as glutamate [16],[23] and ATP [24]. We next asked whether these neuronal signals mediate Ca2+ elevations triggered in astrocytes by the ictal discharge. We found that the activation of astrocytes by the ictal discharge was significantly reduced by slice perfusion with either the antagonist of mGlu receptors 2-methyl-6-(phenylethynyl)-pyridine (MPEP), or the antagonist of purinergic (P2) receptors pyridoxal phosphate-6-azophenyl-2′,4′-disulfonic acid (PPADS, Figure 2A). MPEP/PPADS co-perfusion abolished ictal discharges, thus hampering the possibility to clarify whether glutamate and ATP can entirely account for astrocyte activation by the ictal event.
We also found that after slice perfusion with either MPEP or PPADS, the duration and frequency of ictal episodes in neurons were significantly reduced with respect to controls (Figure 2B and 2C), whereas interictal discharges were either unaffected (PPADS and MPEP/PPADS) or increased in frequency (MPEP; Figure 2D). These results clearly show that Ca2+ elevations mediated by mGlu and P2 receptors in astrocytes (and neurons) do not have a role in the generation of interictal discharges. Given that MPEP and PPADS block receptors in both neurons and astrocytes, these results also suggest that Ca2+ signals activated by these receptors, on one or both cells, may have a role in ictal discharge generation.
We next asked whether astrocyte Ca2+ elevations may have a specific role in ictal discharge generation. To investigate this hypothesis, an agonist able to selectively trigger a Ca2+ increase in astrocytes should be used. The peptide TFLLR, a PAR-1 thrombin receptor agonist, is preferentially expressed in astrocytes and is known to activate glutamate release in astrocytes [25],[26]. We found that PAR-1 immunoreactivity in the EC was largely associated with the soma and the processes of GFAP-positive astrocytes (Figure 3A). Noteworthy, GFAP-negative PAR-1 puncta appeared in continuity with distal portions of astrocyte processes, where GFAP is barely expressed [27] (Figure 3B). Following TFLLR (10 µM) bath perfusion in the presence of both TTX and D-2-amino-5-phosphonopentanoate (D-AP5), which blocks NMDAR-mediated astrocyte-to-neuron signalling [8],[28], we could not detect any Ca2+ change in EC neurons, whereas large Ca2+ elevations were observed in astrocytes (Figure 3C).
We next asked whether PAR-1 receptor activation could stimulate the release of glutamate from EC astrocytes, as previously reported for hippocampal astrocytes [25],[26]. We found that Ca2+ elevations triggered in EC astrocytes by short pressure pulses applied to a TFLLR-containing pipette (1 mM) were followed by slow inward currents (SICs) in adjacent patched neurons (Figure 3D). Most of the SICs recorded in six of 12 neurons occurred within 10 s (mean delay ± SEM, 1.3±0.3 s) after the TFLLR-induced Ca2+ elevations in astrocytes (Figure 3D and 3E). Unlike fast spontaneous synaptic currents (asterisks in Figure 3D), SICs have typical slow kinetics (rise time, 83.0±36 ms, decay time, 451±171 ms; n = 13), are insensitive to TTX, and are sensitive to the NMDAR blocker D-AP5 (Figure 3E), as demonstrated in neurons from other brain regions [8],[28]–[30].
In the picrotoxin/zero-Mg2+ entorhinal cortex slice model, we then investigated whether selective astrocyte activation enhanced ictal discharge generation. We found that Ca2+ elevations triggered in astrocytes by local TFLLR applications were sufficient to shift neurons towards the ictal discharge threshold (Figure 3F; Video S2). To demonstrate the causal link between the ictal discharge and the immediately preceding TFLLR-induced Ca2+ increase in astrocytes, we simulated the ictal occurrence by a Monte Carlo procedure. Results from this analysis revealed that in six experiments in which 30 TFLLR applications were performed, 10 of the 15 observed ictal events were correlated at the 0.05 confidence level with a preceding astrocyte Ca2+ increase (Figure S2).
These results suggest that when the level of basal excitability and the predisposition of neurons to generate epileptiform discharges is high, as in the picrotoxin/zero-Mg2+ model, activation of the NMDAR by astrocytic glutamate could trigger neuronal hyperactivity that is sufficient to generate an ictal discharge. Compelling, although indirect, support for this hypothesis derived from the observation that a short pressure-pulse application of NMDA via an NMDA-containing pipette could also evoke an ictal discharge (Figure S3).
To further investigate the possible role of astrocytes in seizure initiation, we developed a model of focal seizures alternative to the picrotoxin/zero-Mg2+ model. In this latter model, indeed, epileptic activities arise spontaneously and at unpredictable foci [31],[32], and therefore the cellular mechanism of seizure initiation cannot be analyzed accurately.
In our new model, ictal discharges are reproducibly generated at discrete sites of the EC by focal NMDA applications. Figure 4A reports schematically the positions of an NMDA-containing pipette and an OGB1-containing patch pipette in layer V of the EC. A confocal image of this region under basal conditions (t0) is also presented. Focal episodes of neuronal hyperactivity are induced in the presence of 100 µM 4-aminopyridine (4-AP) and 0.5 mM Mg2+ by short pressure-pulse applications of NMDA via the NMDA-containing pipette. The effect of the NMDA pulse ejection was monitored by simultaneously recording Ca2+ signals from neurons and the AP firing from one of the neurons close to the NMDA pipette. Notably, in the submerged chamber experiment used in our study, no spontaneous ictal discharges were observed during 4-AP slice perfusion, whereas under different experimental conditions, such as in interface chamber experiments, epileptiform activities arise spontaneously [2]. As illustrated by the fluorescence change, a single NMDA pulse stimulated a transient Ca2+ increase in a limited number of layer V-VI neurons from the region close to the pipette tip, that we termed the field A (t1, Figure 4B; Video S3). This local response is clearly illustrated by the difference image generated by subtracting the fluorescence image captured at basal conditions to that obtained after the NMDA stimulation (t1 − t0, Figure 4B). Simultaneous patch-clamp recording and Ca2+ imaging revealed that the NMDA stimulus leads to AP burst firing in the patched neuron coupled with a Ca2+ elevation in this and the other neurons from field A, but it failed to activate neurons of the surrounding region, which we termed field B (Figure 4D1).
Two-pulse NMDA stimulation with a 3-s interval evoked a stronger activation of neurons and a transient Ca2+ elevation in some of the previously unresponsive neurons from the surrounding field B (t2 and t2 − t0, Figure 4C). The response to the double NMDA pulse evolved into a sustained plateau with superimposed Ca2+ spikes correlated with AP bursts typical of an ictal discharge, i.e., the cellular equivalent of a seizure [2] (Figure 4D2, see also Figure 4C, t3). The ictal discharge was characterized by Ca2+ spikes from unpatched neurons in both field A and field B, highly synchronized with the AP bursts (Figure 4D2; Video S4). The recruitment of neurons in field B that underlines the spreading to this region of the ictal discharge is also clearly illustrated by the difference image t3 − t0 (Figure 4C). The time window between the double NMDA pulse and the Ca2+ elevation that occurs synchronously in both field A and B neurons represents a transition phase during which the ictal discharge develops in field A. In the presence of TTX, the ictal discharge in both field A and B neurons was abolished, whereas the initial response of field A neurons was unaffected (Figure 4E). The size of the cortical region occupied by neurons that respond directly with a transient Ca2+ rise to a double NMDA pulse applied in the presence of TTX was 369±17 µm. Notably, the number of neurons in this response (56.5±7.2) is underestimated since it comprises only neurons activated by NMDA in a single focal plane. These results demonstrated that i) AP-mediated events secondary to the initial activation of field A neurons are crucial for ictal discharge maturation; and ii) the activation of neurons from field B and the generation of the ictal discharge was not due to a delayed diffusion of NMDA. Paired recordings from two pyramidal neurons (one in field A and the other in field B) confirmed that similar ictal discharges were regularly evoked in field A and B by successive double NMDA pulses (Figure S4).
According to results obtained from 14 experiments, no failures were observed in a total of 101 double NMDA pulse stimulations, and the mean duration of the ictal discharge repetitively evoked by these stimulations was reproducible over long time periods (up to 60 min, Figure S4). By applying successive double NMDA pulses in the presence of TTX, no NMDA-mediated Ca2+ elevations were detected in field B neurons, whereas the number of field A neurons activated directly by NMDA and the amplitude of their Ca2+ response were found to be unchanged over the same time period (Figure S4).
Ictal discharges could be evoked also by two single NMDA pulses applied at two different sites, either simultaneously or in succession. Intervals of 3 or 5 s were successful, but not an interval of 10 s. To be effective, the two pipette tips should be positioned close enough to allow a large spatial overlapping of the two pulses. Only in this overlapping region were neurons strongly activated by the two NMDA pulses. Notably, if the distance between the two pipette tips was 172±30.2 µm (n = 5) (a value similar to the mean radius of the field A directly activated by double NMDA pulses), the two single NMDA pulses regularly evoked an ictal discharge. If the distance of the two pipette tips was 220±38.5 µm, no ictal discharges could be evoked.
Altogether, these data show that an episode of activity evoked in a group of neurons by local NMDA applications creates an initiation site for a seizure-like discharge that secondarily involves adjacent neuronal populations. They also demonstrate that our model is highly reliable since comparable ictal discharges can be evoked by repetitive stimulations applied to the same restricted site. Notably, in contrast to the picrotoxin/zero-Mg2+ model, in the 4-AP model, single NMDA pulses failed to trigger focal ictal discharges, suggesting different thresholds for seizure generation in these two models (see Discussion).
We next investigated astrocyte activities during the development of focal ictal discharges. We observed that shortly after the initial neuronal response to a double NMDA pulse, a large Ca2+ elevation occurred almost simultaneously in the large majority of field A astrocytes (Figure 5A, red traces; Video S4). Similar Ca2+ elevations in these astrocytes were never observed during the neuronal response to a single NMDA pulse. In 13 experiments, a mean of 17.4±3.5 out of 20±3.1 responsive astrocytes in field A displayed an early Ca2+ elevation during the transition phase. As a mean, astrocyte activation in field A occurred 4.8±1.1 s before field B neurons were recruited into the ictal discharge. Most of the astrocytes in field B were activated later, i.e., after the invasion of the ictal discharge into this region (Figure 5A, blue traces; Video S4). High-magnification images in Figure 5B illustrate “early” and “late” Ca2+ changes of astrocytes from field A and B, respectively. The mean percentage of astrocytes from field A and B displaying “early” and “late” responses is reported in Figure 5C. Notably, when the ictal discharge was evoked by two single NMDA pulses applied at two distinct sites (Figure 5D), most astrocytes from both the field of spatial overlapping of the two pulses and the immediately surrounding regions (fields A1 and A2) displayed a similar early Ca2+ elevation (85.6±5.4%), whereas most astrocytes from the surrounding regions (the fields B) showed a late activation (71.6±5.4%). Noteworthy is that astrocytes failed to be similarly activated by each single NMDA pulse alone (Figure 5D).
We next asked whether the initial Ca2+ elevation in astrocytes (and neurons) from field A spread to other astrocytes (and neurons) in the surrounding regions through a concentric wave of activation centred on the NMDA pipette. We found that the Ca2+ response of astrocytes as well as the recruitment of neurons into the ictal discharge is more consistent with a process of modular recruitments rather than with a propagation of a concentric wave of activity (Figure S5).
Astrocyte activation was largely due to AP-mediated neurotransmitter release since 70.4±8.3% (n = 143, 5 experiments) of the field A astrocytes, activated by a first double NMDA pulse, failed to respond to a second double NMDA pulse applied in the presence of TTX. The Ca2+ rise in still-responsive astrocytes displayed slow kinetics and were of small amplitude (ΔF/F0, 64.1±3.6 before and 29.0±2.2 after TTX; n = 41; p<0.001). This residual astrocyte response in TTX could be due either to neurotransmitter release mediated by activation of presynaptic NMDA receptors [33] or to the direct activation by NMDA of NMDA receptors that may be present on astrocytes [34],[35].
The results from these experiments indicate that the development of a focal ictal discharge is accompanied by Ca2+ elevations in astrocytes.
If this early Ca2+ elevation in astrocytes is not a mere consequence of neuronal activity and has, instead, a causative role in ictal discharge generation, its inhibition should reduce the ability of NMDA to trigger an ictal discharge. To address this hypothesis, we first bath applied MPEP and PPADS (n = 4) and found that the direct activation of neurons by a double NMDA pulse was unchanged, but early activated astrocytes were reduced to 4.6±2.6% of controls. Under these conditions, the generation of the ictal discharge in field A and the subsequent recruitment of neurons into the epileptic discharge in field B were inhibited (Figure 6A). The ictal discharge recovered after washout of the antagonists and the reappearance of the associated Ca2+ elevation in astrocytes. Interestingly, a stronger neuronal stimulation obtained by increasing the number of successive NMDA puffs evoked an ictal discharge, although of short duration, even in the presence of MPEP/PPADS and without a recovery of astrocyte Ca2+ signals (Figure 6A).
We also found that the NMDA-induced ictal discharge was blocked after inhibition of the early responsive astrocytes in field A by MPEP/PPADS applied locally to this region (Figure 6B; n = 4). Ictal discharge recovery was regularly observed 5–10 min after cessation of the MPEP/PPADS pulses. In contrast, applications of MPEP/PPADS to a limited sector of field B failed to affect the spread to field B of the ictal discharge generated in field A (n = 4). However, it is noteworthy that the Ca2+ elevations in astrocytes from this sector were poorly affected (Figure 6B).
Given that MPEP and PPADS are not selective antagonists of Ca2+ signals in astrocytes, to provide a direct evidence for a causal link between Ca2+ elevations in field A astrocytes and ictal discharge generation, we inhibited Ca2+ signals in these astrocytes selectively by introducing the Ca2+ chelator 1,2-bis(o-aminophenoxy)ethane-N,N,N',N'-tetraacetic acid (BAPTA; 50 mM) into individual astrocytes through a patch pipette [36].
First, we indirectly evaluated BAPTA spreading in the astrocyte syncytium by patching single EC astrocytes with a Texas Red-containing pipette. We counted 31±7 red-labelled astrocytes in an area of 242±50 µm in diameter (Figure 7A). This value is close to the size of the cortical region occupied by neurons that respond directly with a transient Ca2+ rise to a double NMDA pulse applied in the presence of TTX (Figure 7A). In subsequent experiments, before patching a field A astrocyte with a BAPTA-containing pipette, a double NMDA pulse was applied to trigger an ictal discharge (Figure 7B and 7C). In five out of nine BAPTA experiments, a double NMDA pulse applied 50 min after BAPTA diffusion in the astrocyte syncytium failed to activate both the Ca2+ elevations in astrocytes and the ictal discharge (Figure 7B and 7C). Notably, in these five experiments, the response of early activated field A astrocyte was strongly reduced with respect to that observed before BAPTA (Figure 7D). In these experiments, we addressed the contribution of astrocytes in the activation of neurons during the transition phase. In the presence of BAPTA, which specifically inhibited Ca2+ signals in field A astrocytes, the number of recruited neurons upon the double NMDA pulse was 33.1±3.2% lower than in controls (p<0.05). Such a reduction is unlikely due to experimental variability in the intensity of the NMDA stimulation since the number of neurons activated and the amplitude of their Ca2+ responses to successive double NMDA pulse stimulations (as measured in the presence of TTX) were unchanged over a 50-min period (Figure S4). These observations indicate that the recruitment of neurons into the ictal discharge is also mediated by the early activated astrocytes that signal back to neurons.
In the four experiments with BAPTA in which the ictal discharge was preserved, most of the astrocytes in field A still displayed an early Ca2+ response, suggesting a defective diffusion of BAPTA in the astrocyte syncytium in these experiments (Figure 7D). These data provide a plausible explanation for the lack of inhibition of the ictal discharge in these BAPTA experiments.
In a number of different control experiments, we found that i) two subsequent double NMDA pulses applied before and 50 min after patching either a neuron (n = 8) or an astrocyte (n = 4) with a pipette not containing BAPTA always evoked comparable ictal discharges, indicating that such a long time interval does not affect the ability of a double NMDA pulse to trigger an ictal discharge; ii) double NMDA pulses regularly evoked an ictal discharge even after 50 mM BAPTA was puffed directly over the neurons for 2 min via a pipette (n = 4), indicating that a leakage of BAPTA, putatively occurring during astrocyte seal formation, cannot account for the ictal discharge inhibition observed in the BAPTA experiments; iii) successive double NMDA pulses applied in the presence of TTX over a period of 50 min, while patching single astrocytes with a BAPTA-containing pipette, evoked an unchanged response in neurons (Figure 7E), demonstrating that the direct response of neurons to NMDA is not affected after BAPTA-mediated inhibition of astrocyte Ca2+ signals.
We next asked whether the late activation of astrocytes in field B contributes to the spreading of the ictal discharge. After patching individual field B astrocytes with a BAPTA-containing pipette, we observed that the ictal discharge evoked in field A by a double NMDA pulse still invaded field B and further propagated to the adjacent region, whereas the activation of field B astrocytes was drastically affected both in terms of Ca2+ signal amplitude (−56.6±2.4%, p<0.001) and kinetics (time to peak, 2.6±0.4 s and 15.2±3.3 s, before and after BAPTA, respectively; p<0.001; Figure S6).
As a further control for the specificity of the BAPTA effect, we demonstrated that the ictal discharge inhibition by BAPTA was spatially restricted. After the astrocyte syncytium in region 1 was loaded with BAPTA, a double NMDA pulse stimulation close to the BAPTA-loaded region failed to trigger an ictal discharge, whereas the same NMDA stimulation applied ∼500 µm away from region 1 readily evoked an ictal discharge (region 2, Figure 7F–7H).
The ictal discharge blocked after the BAPTA-mediated inhibition of field A astrocytes was recovered in two of three experiments by applying a stronger stimulation of neurons, such as a triple NMDA pulse (Figure 7I; white arrowheads). Notably is that under these conditions, astrocytes recovered a Ca2+ response that was, however, delayed and of reduced amplitude with respect to that without BAPTA. These results are consistent with the hypothesis that the astrocyte contribution to ictal discharge generation is not an absolute requirement and can be bypassed by a stronger stimulation of neurons, as already suggested by the results obtained in MPEP/PPADS experiments.
Taken together, the results of these series of experiments confirm the reliability of the double NMDA pulse paradigm in evoking an ictal discharge over long time periods and, on the other hand, validate the selective inhibition of astrocyte Ca2+ signals by intracellular BAPTA application.
If inhibition of Ca2+ signals in astrocytes can block the generation of a focal ictal discharge, it would be expected that direct astrocyte stimulation promotes ictal discharges. In the experiments that addressed this hypothesis, we took advantage of the finding that none of the 48 single NMDA pulses performed in the 4-AP ictogenic model could generate an ictal event. Single NMDA pulses that repetitively failed to trigger an ictal discharge became effective when they were coapplied with TFLLR (Figure 8). We found that a single NMDA pulse coupled with TFFLR, evoked an ictal event in six of nine trials from a total of three experiments. In these experiments, we also found that the number of neurons activated by the NMDA/TFLLR coapplication during the transition phase was higher than that activated by NMDA alone (mean increase, +119.3±16.3%; n = 6; p<0.001). These data confirm that the contribution of astrocytes in the recruitment of neurons can be critical for the generation of the ictal discharge.
In brain slice models of seizures, the ictal discharge is proposed to initiate at focal brain sites by asynchronous neuronal hyperactivities that progressively recruit adjacent neurons into a synchronous discharge [1],[3]–[4]. In our study, we found that neuronal hyperactivities at these restricted brain sites are accompanied by Ca2+ elevations in a large number of astrocytes that contribute to drive neurons towards seizure threshold.
The focal ictogenesis in our model is schematically illustrated in Figure 9. This process starts with an isolated episode of local neuronal hyperactivity that triggers a large and synchronous Ca2+ elevation in closely associated astrocytes (N1). These activated astrocytes signal back to neurons (A1) favouring the recruitment of neurons into a coherent activity that underlines the hypersynchronous ictal discharge. This event, in turn, triggers a second activation of astrocytes (N2). The secondary astrocyte activation may then contribute to sustain the ictal discharge (A2). This sequence of events represents a recurrent neuron–astrocyte excitatory loop that drives neurons towards the ictal discharge threshold.
Since our slice experiments were performed mainly in young animals, the role of astrocytes to seizure generation may be restricted to the immature brain. Although additional experiments are necessary to clarify this important issue, the ability of astrocytes to release glutamate and activate neuronal SICs in slices from young adult rats [8],[28],[37],[38] suggests that astrocyte-to-neuron signalling may contribute to seizure initiation also in the adult brain.
In EC slices perfused with the proconvulsant agent 4-AP in low Mg2+ conditions, we found that a synchronous Ca2+ elevation in a high number of astrocytes occurred along with the development of the ictal discharge evoked by a local NMDA application. This response was largely TTX sensitive, indicating that astrocytes were activated by AP-mediated neurotransmitter release. Most importantly, the early astrocyte activation was a crucial step in the generation of ictal discharges. Indeed, when Ca2+ elevations in field A astrocytes were inhibited by BAPTA, the episode of neuronal hyperactivity induced by NMDA failed to generate an ictal discharge. According to results obtained from different control experiments, the effect of BAPTA on ictal discharge generation was specifically linked to the inhibition of astrocyte Ca2+ signals.
The Ca2+ elevations in astrocytes are associated with the release of gliotransmitters, such as glutamate [39]–[41] and D-serine [42], that modulates neurotransmitter release [24],[43],[44], triggers AP firing in neurons [10], and promotes local neuronal synchrony [8],[28]. Ca2+-dependent release of D-serine from astrocyte activated by Schaffer collateral stimulation has been also recently shown to be crucial for the potentiation of synaptic transmission in the CA1 hippocampal region [45]. As previously reported in the hippocampus [25],[26], we show here that Ca2+ elevations stimulated in EC astrocytes by the PAR-1 receptor agonist, TFLLR, triggers glutamate release in these cells and, in turn, NMDA receptor–mediated SICs in neurons. The activation of neurons by gliotransmission can thus account for the finding that a single NMDA pulse, ineffective per se, was able to trigger the ictal discharge if coupled with the direct stimulation of a Ca2+ rise in astrocytes by TFLLR. Data analysis of these experiments revealed that the number of neurons activated after NMDA/TFLLR coapplication was higher than that activated after NMDA alone.
These results suggest that when an episode of hyperactivity in a group of neurons consistently engages nearby astrocytes, a larger population of neurons is recruited into a coherent activity. If this feedback signal operates on a brain network prone to seizures, it contributes to drive neurons towards the ictal discharge threshold. The initiation site is thus represented, not only by the neurons activated by NMDA, but also by those that are secondarily activated in a recruitment process that involves astrocytes. Consistent with this view is our finding that when a double NMDA pulse (that successfully evoked an ictal discharge) was applied either after BAPTA was introduced in the astrocyte syncytium or after local applications of MPEP/PPADS to the site of activation, astrocytes were poorly activated, fewer neurons were recruited, and no ictal discharge was evoked. Further support for this conclusion derives from the experiments with a single NMDA pulse delivered from two pipettes positioned at different distances. These experiments revealed that an ictal discharge could be evoked when astrocytes from the region of overlapping neuronal activation were activated. When the pipette tips were more distant, the overlapping region was reduced, astrocytes were poorly activated, and no ictal discharge was evoked.
Distinct subpopulations of astrocytes might differently contribute to modulate neuronal hyperactivity in the epileptogenic region, possibly by releasing in addition to glutamate, ATP, and other neuroactive signals, e.g., GABA, through a Ca2+-dependent or -independent mechanism [46],[47]. Given that inhibitory interneurons have been reported to restrain the recruitment of neurons during the development of the ictal discharge [4],[48], an opposite action of astrocytes in this process might involve a distinct inhibition of interneurons by GABA released from astrocytes. Indirect support for this possible astrocyte action derives from the observation that GABA released from astrocytes can, indeed, result in a long-lasting inhibition of inhibitory granule cell activity in the olfactory bulb [37]. Whether a similar signalling between a subpopulation of GABA-releasing astrocytes and interneurons may be involved in ictal discharge initiation in the EC represents an interesting question to be addressed in future studies.
Episodes of focal seizures can arise in a nonepileptic tissue due to genetic causes or as a consequence of various brain damage. These may lead to status epilepticus (SE), a condition of persistent seizures, and evolve into chronic epilepsy after a latent period of epileptogenesis. Our results were obtained in nonepileptic brain tissue and provide evidence for the contribution of astrocytes in the initiation of seizure during SE. Therefore, whether astrocytes contribute also to seizure initiation in chronic epilepsy is, at present, unknown and should be appropriately investigated in chronic epilepsy models. However, results from a recent in vivo study showed that astrocytes, which exhibited long-lasting Ca2+ oscillations during SE, contributed to the neuronal death that characterizes chronic epilepsy [38]. This effect was due to astrocytic glutamate that activated neuronal NMDARs, possibly favouring seizure generation. It is also worth underlining that in the epileptic brain tissue, astrocytes undergo significant changes in their physiological properties that may result in decreased glutamate uptake, altered extracellular K+ buffering capacities, and activation of inflammatory pathways [49],[50]. All these changes may contribute to the increased neuronal network excitability that characterizes the epileptic brain.
The efficacy of astrocyte stimulation in evoking an epileptic discharge was different in the two models used in the present study, probably because of differences in their intrinsic neuronal predisposition to ictal discharge generation. As suggested by the presence of recurrent spontaneous epileptic discharges, the picrotoxin/zero-Mg2+ model can be considered, indeed, a model with a low-threshold for epileptic discharges. In this model, a single NMDA pulse triggered synchronous activity in a number of neurons sufficient to reach the ictal discharge threshold, and a single stimulation of astrocytes was also sufficient to trigger an ictal discharge. As suggested by the absence of spontaneous epileptic events, the 4-AP model has a higher threshold for epileptic phenomena. In this model, seizure discharges could be triggered by a more prolonged and intense episode of neuronal activity induced by a double NMDA pulse, and not by single NMDA or TFLLR pulses. An ictal discharge could be also evoked when a single NMDA application (ineffective per se) was coupled with TFLLR-mediated astrocyte activation. Furthermore, the reduction in astrocyte Ca2+ signals blocked the ictal discharge in the 4-AP model, but not in the picrotoxin/zero-Mg2+ model. These data demonstrate that experimental manipulations of the astrocyte Ca2+ signals can influence neuronal recruitment and thus affect, in concert with the level of neuronal activity, the likelihood of ictal events.
As revealed by results from both BAPTA and MPEP/PPADS experiments, when the astrocyte contribution was reduced by inhibiting Ca2+ signals in these cells and the double NMDA pulse consequently failed to evoke an ictal discharge, we could recover an ictal discharge by applying a more intense NMDA stimulation. By activating directly a larger number of neurons, this higher stimulus evokes a level of correlated activity that is sufficient for seizure-like discharge generation, bypassing the astrocyte contribution in the recruitment process. Thus, astrocyte activation is not an absolute requirement for ictal discharge generation.
However, astrocytes respond readily to synaptic activity with Ca2+ oscillations [16],[23],[51], and the frequency of these oscillations increases in parallel with an increased neuronal activity [16]. In vivo studies also revealed that sensory stimuli can evoke distinct Ca2+ elevations in astrocytes confirming the strict association between neuron and astrocyte activities [52]–[55]. Thus, pathological hyperactivities in neurons [6] should be regularly accompanied by an increased astrocyte activity. In support of this view, studies in brain slices showed that chemically induced epileptiform activity causes a sustained increase in astrocyte Ca2+ signalling [9],[10], and in vivo studies reported a long-lasting hyperactivity of astrocytes after pilocarpine-induced SE [38]. It is conceivable that a pathological hyperexcitability that predisposes neurons to seizure discharges may originate from abnormalities in the neuron–astrocyte network activity, whatever the origin of the initial dysfunction might be. As we showed here, depending on the different level of excitability in neurons, the astrocyte contribution varies, but it can even be crucial for ictal discharge generation.
In our 4-AP slice model, a second Ca2+ elevation even of larger amplitude than that early evoked by the double NMDA pulse, occurred in astrocytes in both field A and field B. This delayed activation of astrocytes was observed also after the spontaneously occurring ictal discharges in the picrotoxin/zero-Mg2+ model in both rats and pGFAP-EGFP transgenic mice, as well as in other models such as the 4-aminopyridine/picrotoxin and high-potassium models (unpublished data). Most importantly, this observation was validated in the intact guinea pig brain preparation, a well-characterized model of EC–hippocampus focal ictogenesis [20],[21]. In this close to in vivo preparation, the development of the ictal discharge was regularly accompanied by a Ca2+ elevation in virtually all astrocytes present in the recording field, whereas large-amplitude interictal discharges were never associated with a significant Ca2+ change in astrocytes. This Ca2+ elevation and the following release of gliotransmitters may contribute to the maintenance of AP bursts and to the process of neuronal recruitment that characterize seizure discharge propagation. Our finding that the duration of the ictal discharges was significantly reduced upon inhibition of the astrocyte Ca2+ signal by bath perfusion with MPEP or PPADS is consistent with this hypothesis, which needs, however, to be specifically addressed in future experiments.
In the present study, we also addressed a possible role of the late astrocyte response in the propagation of the ictal discharge outside the focal region. After BAPTA introduction in field B astrocyte syncytium, the ictal discharge still propagated to this region and further, suggesting that Ca2+ elevations in field B astrocytes may have no role in this process. Given that initiation, propagation and cessation of the ictal discharge are likely governed by distinct mechanisms [3], it would not be surprising that astrocytes have, indeed, a role in ictal discharge initiation but not in propagation. This conclusion is, however, reasonable, but it is not proven beyond all doubt. Indeed, the inhibition by BAPTA could be exerted only in astrocytes from a small sector of the large field B, whereas astrocytes outside this sector were totally unaffected. Their activation might thus be sufficient to sustain the propagation of the ictal discharge even to the small sector where astrocytes were inhibited by BAPTA. As to MPEP/PPADS, when locally applied to field B, these competitive receptor antagonists failed to inhibit the ictal discharge propagating to this region. These results, however, do not allow us to draw any conclusions since the ictal discharge invading field B still activated a significant response in astrocytes even in the presence of MPEP/PPADS. To clarify this point, another experimental approach is thus required.
It is unclear why MPEP/PPADS failed to inhibit the Ca2+ elevation evoked by the ictal discharge in field B astrocytes. It is likely that, with respect to the NMDA pulse, the ictal discharge represents a more powerful stimulus that triggers the release of glutamate and ATP. Accordingly, the extracellular concentration of MPEP/PPADS reached after local applications might have been insufficient to inhibit the large activation of astrocyte mGlu and P2 receptors upon the ictal discharge. However, mechanisms other than mGlu and P2 receptor activation may be also involved in this astrocyte response.
Interictal discharges failed to activate significantly a Ca2+ elevation in astrocytes. Recently, it has been reported that glutamate release triggered by Ca2+ elevations in astrocytes plays a predominant, if not obligatory role in the generation of epileptic activity in the hippocampus and, in particular, in the slow depolarization shift associated with interictal discharges [9]. This conclusion is, however, at variance with a number of studies showing that both interictal and ictal seizure-like discharges from different brain regions, including the hippocampus, are strictly linked to neuronal activity being efficiently prevented or blocked, depending on the time of application, by TTX [10],[38],[56]–[58]. In the present study, we observed that i) the interictal activity was not blocked after Ca2+ elevations in astrocytes were drastically reduced; and ii) synchronous astrocyte Ca2+ elevations were never observed to accompany an interictal discharge in the different models. We thus failed to confirm a role of astrocytic glutamate in interictal discharge generation. The reasons for this discrepancy are, at present, unknown.
The present study reveals a crucial role of neuron–astrocyte interactions in sculpting activity at the epileptogenic zone. When a group of neurons is abnormally active (due to acquired or genetic causes), ictal epileptiform events may occur through the activation of astrocytes. Astrocytes can thus play a key role in seizure initiation in a nonepileptic brain tissue and, in contrast to previous observations [9], do not appear to be involved in the generation of the interictal events. This peculiarity makes the astrocyte–neuron unit a primary target for novel drug development aimed at interfering selectively with ictogenesis, without affecting the interictal activity that, by preventing seizure precipitation, may have a beneficial role in focal epilepsies [59],[60].
The high reproducibility in the generation of comparable ictal discharges represents an important advantage of our new EC slice model of ictogenesis. This model allowed us to investigate the early events that, at a restricted brain site, predispose neurons to seizure and to obtain some insights into the mechanism of focal ictal generation that involves astrocytes. Other aspects that were not addressed in the present study, such as the neuronal recruitment process during the diffusion of the ictal discharge to regions distant from the site of ictal discharge generation, could be investigated in this model. These acute experiments set the conditions for validating the mechanisms here described in future studies in chronic models of epilepsy, including genetically determined in vivo models of epilepsy, that more closely mimic the complex feature of seizures in epileptic patients. A validation of the astrocyte role in seizures generation in these models is fundamental to provide further arguments in favour of astrocytes as targets for developing new therapeutic strategies for epilepsies.
All experimental procedures were authorized by the Italian Ministry of Health.
Transverse cortico-hippocampal slices were prepared from postnatal day 14–18 Wistar rats or pGFAP-EGFP transgenic mice [61], and loaded with OGB1-AM (excited at 488 nm) or Rhod-2 (excited at 543 nm), respectively, as previously described [8]. Briefly, brain was removed and transferred to ice-cold cutting solution containing (in mM): NaCl, 120; KCl, 3.2; KH2PO4, 1; NaHCO3, 26; MgCl2, 2; CaCl2, 1; glucose, 10; Na-pyruvate, 2; and ascorbic acid, 0.6; at pH 7.4 (with 5% CO2/95% O2). Coronal slices were obtained by cutting with a Leica vibratome VT1000S in the presence of the ionotropic glutamate receptor inhibitor kynurenic acid (2 mM). Slices were recovered for 15 min at 37°C and then loaded with the Ca2+-sensitive dye OGB1-AM (Invitrogen) for 60 min at 37°C. Loading was performed in the cutting solution containing sulfinpyrazone (200 µM), pluronic (0.12%), and kynurenic acid (1 mM). After loading, slices were recovered and kept at room temperature in the presence of 200 µM sulfinpyrazone. Brains from postnatal day 14–20 guinea pigs were isolated and perfused at a rate of 5.5 ml/min through the basilar artery [19],[62] with a solution containing (in mM): NaCl, 126; KCl, 3; KH2PO4, 1.2; MgSO4, 1.3; CaCl2, 2.4; NaHCO3, 26; glucose, 15; and 3% dextran M.W. 70.000; oxygenated with a 95% O2/5% CO2 gas mixture (pH 7.3). The dye OGB1-AM (50 µg) was diluted in 5 µl of standard pluronic/DMSO solution and 75 µl of saline, and filtered through a 0.2-µm microfilter (Millipore). A patch pipette (3–4 MΩ was used to pressure inject (1–2 min at 4 PSI) the Ca2+ dye into the EC at a depth of about 200 µm via a picospritzer (NPI Electronics). Following this procedure, the Ca2+ signal from astrocytes, neurons, and neuropile was monitored. All experiments were performed at 33–35°C.
In slice experiments, we used a TCS-SP2-RS or a TCS-SP5-RS confocal microscope (Leica) equipped with a 20× objective (NA, 1.0) and a CCD camera for differential interference contrast images. For experiments on isolated guinea pig brains, we used a Fluoview 300 scanning head customized for two-photon microscopy equipped with a 5W Verdi-Mira laser (Coherent) tuned at 830 nm and external photomultipliers (Hamamatsu). Time frame acquisitions from 314 ms to 1.24 s (with one to six line averaging) were used. No background subtraction or other manipulations were applied to digitized Ca2+ signal images that are reported as raw data, with the exception of the difference images in Figure 4 that were obtained by subtracting the prestimulation image from the poststimulation image. In brain slice preparations, neurons and astrocytes were distinguished on the basis of the distinct kinetics of their Ca2+ response to a stimulation with high K+ extracellular solution (40 mM) obtained by isosmotic replacement of Na+ with K+ [16], applied at the end of the recording session in the presence of 1 µM TTX (Figure S7). Due to the lack of voltage-dependent Ca2+ channels in astrocytes, the Ca2+ elevation in these cells upon high K+ stimulation occurs with a delay of several seconds with respect to the response in neurons, and appears to be mediated by glutamate release from depolarizing neurons [17]. In the present study, the presence of TTX was necessary to block the epileptic discharges and the underlying Ca2+ changes in neurons and astrocytes that would have hampered the possibility to distinguish these cells from their different responses to high K+ stimulation. Astrocytes were identified also by their small size, low membrane potentials (−74±0.4 mV without the correction for the liquid junction potential at the pipette tip, which was 15 mV; n = 9), and passive responses to a series of depolarizing steps. In slices from pGFAP-EGFP mice, astrocytes were identified by their green GFP fluorescence. In the guinea pig brain, astrocytes were identified using the astrocyte-specific marker sulforhodamine 101 (Invitrogen) applied at 100 µM to the cortical surface [63]. The onset of the slow Ca2+ elevation in astrocytes was determined on the basis of a threshold criterion. The onset was identified by the change in ΔF/F0 that should be more than two standard deviations over the average baseline and remained above this value in the successive frames for at least 2 s (two to six frames, depending on the frame acquisition rate).
Rat brain slices in a submerged chamber (Warner Instruments) were continuously perfused at a rate of 2 ml/min with (in mM): NaCl, 120; KCl, 3.2; KH2PO4, 1; NaHCO3, 26; MgCl2, 1; CaCl2, 2; glucose, 10; sulfinpyrazone, 0.2; at pH 7.4 (with 95% O2/5% CO2). Whole-cell patch-clamp recordings in rat brain slices were performed using standard procedures and one or two Axopatch-200B amplifiers (Molecular Devices), as previously reported [8]. Typical pipette resistance was 3–4 MΩ for neurons. Data were filtered at 1 kHz and sampled at 5 kHz with a Digidata 1320 interface and pClamp8 software (Molecular Devices). Whole-cell intracellular pipette solution was (in mM): K-gluconate, 145; MgCl2, 2; EGTA, 0.5; Na2ATP, 2; Na2GTP, 0.2; HEPES, 10; to pH 7.2 with KOH, and contained a low concentration (10 µM) of OGB1 (Invitrogen); osmolarity, 305–315 mOsm. Data analysis was performed with Clampfit 8 and Origin 6.0 (Microcal Software) software. SICs with an amplitude greater than 20 pA and a rise time slower than 10 ms are classified as SICs, as described previously. SIC rise time was calculated with the 20%–80% criterion and the decay time as the time constant of a single exponential fit. The delay of each SIC activated in neurons after astrocyte stimulation with TFLLR was calculated with respect to the peak of the immediately preceding astrocyte Ca2+ increase. Interictal and ictal seizure-like events resembling those recorded at the electrographic recordings from patient's brain [21], at a cellular level manifest as intense and hypersynchronous discharges that involve large neuronal population and fundamentally differ in their duration. Despite this common characteristic, they have radically different durations. Indeed, the duration of the epileptic event was an important criterion for classifying interictal and ictal events in slice and the isolated whole guinea pig brain preparations. In Ca2+ imaging experiments, interictal events lasted less that 3 s, whereas ictal discharges were sustained for tens of seconds with a final pattern of highly synchronous afterdischarges. The duration of ictal events varied between 15 and 110 s in brain slices and between 21 and 152 s in the guinea pig brain. Postictal depression was also consistently observed after an ictal event, whereas it was not present after an interictal spike [21]. A pressure ejection unit (PDES, NPI Electronics) was used to apply pressure pulses (4–10 psi, 200–600 ms duration) to NMDA-containing pipettes. Pulse pressure (or duration) was increased until a double NMDA pulse evoked an ictal discharge. The stimulus parameters for successive stimulations remained unchanged over the entire recording session, except in the BAPTA experiments in which they were changed to increase the stimulation of neurons by NMDA and thus to recover the ictal discharge after inhibition of Ca2+ signals in astrocytes. In the double NMDA pulse, the interval between the two pulses was 3 s. NMDA pulses applied with intervals of 5 s, but not 20 s also triggered an ictal discharge (unpublished data). For BAPTA dialysis into the astrocyte syncytium, we used a patch pipette (5–6 MΩ; 310–315 mOsm) containing (in mM): K-methylsulfate, 50; ATP, 2; GTP, 0.4; HEPES, 10; BAPTA, 50. To avoid a leakage of BAPTA from the pipette during seal formation, the BAPTA solution was backfilled after loading the tip with a standard intracellular solution. Texas Red dye (excited at 543 nm) was included at 0.2 mM in a patch pipette containing standard solution and monitored 50–60 min after the whole-cell configuration. For the BAPTA and Texas Red experiments, only GluT (coupled) astrocytes were included. GluT (coupled) and GluR (uncoupled) astrocytes were distinguished according to their different responses to hyperpolarizing and depolarizing current pulses of increasing amplitude and 750 ms duration. Field potentials were recorded from the guinea pig brain with saline-filled micropipettes used to deliver OGB1-AM, via a multichannel differential amplifier (NPI Electronics). A precise allignement of Ca2+ and electrophysiological signals was achieved by acquiring with a syncronisation signal produced by the confocal microscope. Tip potential was measured against a ground reference placed in the recording chamber by means of a voltage follower coupled to an amplifier.
MPEP (50 µM), PPADS (10 µM), TTX (1 µM), D-AP5 (50 µM), 4-AP (100 µM; Ascent Scientific), TFLLR (10 µM; Tocris), and picrotoxin (50 µM; Sigma-Aldrich) were bath applied. TFLLR (1 mM) and NMDA (1 mM; Sigma-Aldrich) were pressure applied. To induce local astrocyte inhibition, we applied pressure pulses of 2 s per 5 min at a frequency of 0.1 Hz to a pipette containing MPEP (500 µM) and PPADS (5 mM). Bicuculline methiodide (50 µM; Sigma-Aldrich) was applied by arterial perfusion to guinea pig brains.
The Monte Carlo simulation was designed to test whether the observed series of stimuli and ictal episodes were compatible with a random distribution. Each simulation run generated randomly distributed stimuli and ictal events based on i) recording duration; ii) number of stimuli and ictal events; iii) minimum interval between stimuli; iv) minimum interval between two successive ictal events; and v) minimum interval between a stimulus and an ictal event (ictal events seem to be followed by at least 20 s of refractory period). These rules imply that the occurrence of ictals and pulses are not completely independent. The random generator produced 30,000 temporal series for each experimental run, using experiment-specific parameters. Figure S2A shows an experiment and three simulated runs. The distance between each stimulus and the first following ictal events were computed for each simulation. The datasets were used to compute the density probability p(t) of observing one ictal at time t after an astrocyte activation (Figure S2B). The cumulative probability CP(t) is obtained by the integration of the probability density and yields the probability of observing an ictal at a time ≤t under the hypothesis that stimuli and ictals are not causally related (Figure S2C). Each ictal event in the experiment was associated with the delay from the immediately preceding stimulus, and the probability of observing the ictal was calculated. If the cumulative probability was less than 0.05, the event was deemed as not satisfying the null hypothesis. Results are presented in Figure S2D.
Coronal slices (100-µm thick) of rat brains were cut with a VT 1000S vibratome (Leica) and directly fixed for 1 h in iced 4% paraformaldehyde in phosphate-buffered solution (PBS). Floating sections were first preincubated in a blocking solution (BS; 1% BSA, 2% horse serum, and 2% goat serum) containing 0.3% triton X-100 and subsequently incubated with the primary mouse anti-thrombin receptor PAR-1 antibody (1∶300, Zymed Laboratories, Invitrogen) and rabbit anti-GFAP (1∶500, Dako) diluted in BS. After 24 h, slices were washed at 4°C in PBS and incubated with the secondary antibodies (Cy3 conjugated donkey anti-mouse IgG, and Cy2 conjugated goat anti-rabbit F(ab')2 fragments; Chemicon International) for 2 h at room temperature. Slices were extensively washed in PBS, mounted in Elvanol, and observed with a Leica SP2 laser scanning confocal microscopy. Negative controls were performed in the absence of the primary antibodies. Images were assembled using CS Adobe Photoshop software.
The Ca2+ signal is reported as ΔF/F0, where F0 is the baseline fluorescence. Data are shown as mean ± standard error of the mean (S.E.M.). Unless stated otherwise, the Student t-test was used, with p values ≤0.05 taken as statistically significant.
|
10.1371/journal.pcbi.0030100 | Hierarchical Processing of Auditory Objects in Humans | This work examines the computational architecture used by the brain during the analysis of the spectral envelope of sounds, an important acoustic feature for defining auditory objects. Dynamic causal modelling and Bayesian model selection were used to evaluate a family of 16 network models explaining functional magnetic resonance imaging responses in the right temporal lobe during spectral envelope analysis. The models encode different hypotheses about the effective connectivity between Heschl's Gyrus (HG), containing the primary auditory cortex, planum temporale (PT), and superior temporal sulcus (STS), and the modulation of that coupling during spectral envelope analysis. In particular, we aimed to determine whether information processing during spectral envelope analysis takes place in a serial or parallel fashion. The analysis provides strong support for a serial architecture with connections from HG to PT and from PT to STS and an increase of the HG to PT connection during spectral envelope analysis. The work supports a computational model of auditory object processing, based on the abstraction of spectro-temporal “templates” in the PT before further analysis of the abstracted form in anterior temporal lobe areas.
| The past decade has seen a phenomenal rise in applications of functional magnetic resonance imaging for both research and clinical applications. Most of the applications, however, concentrate on finding the regions of the brain that mediate the processing of a cognitive/motor task without determining the interaction between the identified regions. It is, however, the interactions between the different regions that accomplish a given task. In this study, we have examined the interactions between three regions—Heshl's gyrus (HG), planum temporale (PT), and superior temporal sulcus (STS)—that have been implicated in processing the spectral envelope of sounds. The spectral envelope is one of the dimensions of timbre that determine the identity of two sounds that have the same pitch, duration, and intensity. The interaction between the regions is examined using a system-based mathematical modelling technique called dynamic causal modelling (DCM). It is found that flow of information is serial, with HG sending information to PT and then to STS with the connectivity between HG to PT being effectively increased by the extraction of spectral envelope. The study provides evidence for an earlier hypothesis that PT is a computational hub.
| The concept of an auditory object is controversial [1]. The term can be applied to a sound source like a voice, or an acoustic event generated by a source such as a vowel sound. In both cases, there are features of the object that are independent of the detailed structure of the sound: we can recognise the same vowel, or voice, regardless of the pitch. In these examples, the spectral envelope of the sound determines the particular vowel sound produced, and is, in general, one of the important acoustic features that determine its perceived timbre (Figure 1; spectrogram of the same vowel at different pitch). In this experiment we consider the “abstraction” of the spectral envelope a critical aspect of auditory cognition that defines auditory objects before semantic processing. Such analysis allows generalisation between different exemplars (e.g., the same vowel at a different pitch) in an analogous manner to the generalisation between visual objects that are seen from different perspectives.
The analysis of spectrum in the central auditory system begins in the cochlear nucleus [2], in which models specify sharpening of spectral representation by lateral inhibition [3]. Although relatively less is known about the representation of spectrum in the inferior colliculus and auditory thalamus, a general understanding is that the sharpening of spectrum representation continues in these centres [4]. At the level of the primary cortex, however, animal studies show that representation of spectrum is more complex. Specifically, a given spectrum at the cortex is represented at multiple scales in which a given spectrum has multiple representations at different levels of spectral resolution [5]. Mathematically, this representation has been called ripple analysis, where a given spectrum is decomposed into a sum of ripples of different ripple densities and velocities [6]. Neurons in the primary auditory cortex allow spectral analysis by selectively responding to a fixed ripple density and ripple velocity. The complex spectral analysis in the cortex [7] has not been demonstrated in subcortical areas to the same extent [8]. In this study, we examine the human cortical representation of the spectral envelope independently of the fine structure of the spectrum, a process for which there are a priori grounds for specifying cortical models that are based on initial complex representations in the primary auditory cortex.
We have previously demonstrated bilateral activation of the planum temporale (PT) and right-lateralised activation of the superior temporal sulcus (STS) during the analysis of the spectral envelope in a conventional analysis of functional magnetic resonance imaging (fMRI) data [9]. Such analyses identify the regional nodes of a network that are active during the task without demonstrating the pattern of connections that determines the dynamics of the system: there are multiple mechanisms by which the measured task-induced regional responses could be explained. In the current study, we go beyond classical structure–function correlations and characterise formally the functional interactions between auditory areas involved in spectral envelope analysis. This system identification approach rests on the mathematical characterization of the causal and context-dependent influences that system elements exert upon each other (i.e., effective connectivity [10–13]).
We use dynamic causal modelling (DCM) and Bayesian model selection [14] to address two fundamental questions about the biological computations that attend auditory processing. First, we assess the general structure of the HG–PT–STS network for auditory object processing. In particular, we address the critical question of whether analysis in PT and STS occurs in a serial (hierarchical) fashion, based on connections from HG to PT and from PT to STS, or whether the analysis is based on parallel processing that is mediated by connections from HG to both PT and STS. Second, we address how connection strengths between elements of this cortical network are modulated or enabled during the spectral envelope processing. The approach allows a direct test of a computational mechanism we have suggested previously [15]. This scheme is based on an initial stage of abstraction of the properties of the stimulus that occurs at the PT, before further processing of the abstracted “template” in areas that are concerned with categorical and semantic processing of auditory stimuli. The demonstration of a serial mechanism based on the PT as an intermediate stage would be consistent with such a scheme.
In brief, our results provide strong support for a serial model with increase of the connection strength of the first stage from the HG to PT during spectral envelope analysis. The results suggest a single “stream” for auditory object analysis, and are congruent with macaque models based on a predominant pattern of connectivity from core to belt to parabelt areas.
We assessed the ability of different network models to explain the variation in measured fMRI blood oxygenation level–dependent (BOLD) responses in the HG, PT, and STS in the right hemisphere during the extraction of the spectral envelope of generic sounds without any semantic association. Two broad classes of models, serial and parallel, were defined as shown in Figure 2. The serial models contain connections from HG → PT and thence from PT → STS. In contrast, parallel models postulate connections from the HG to both the PT and STS. The models within each family differ with respect to the back connections specified, and with respect to the specific site of the modulatory effect of spectral envelope analysis. The models were compared using Bayesian model selection implemented within SPM (http://www.fil.ion.ucl.ac.uk/spm/software/spm5). The selection procedure estimates the probability of each model given the data using Akaike information criterion (AIC) and Bayesian information criterion (BIC) approximations to each model's log-evidence or marginal likelihood.
Figure 3 shows the evidence for the models, determined separately using AIC and BIC, in eight participants. In this figure, we have assumed all models were equally likely a priori. This allows us to treat the normalised marginal likelihood as the conditional probability of each model. Model 1 is the optimal model over all participants, with the exception of participant 7. The parameters for this model specify a serial model with connectivity (HG → PT → STS) and modulation of connection from HG → PT during the analysis of the spectral envelope. In addition to the individual inference, Table 1 shows the group Bayes factor (GBF) for model 1 with respect to the other 15 models. Given candidate hypotheses (models) i and j, a Bayes factor of 150 corresponds to a belief of 99% in the statement that “hypothesis i is true”. Following the usual conventions in Bayesian statistics [14,16], this corresponds to “strong” evidence in favor of model i (compare Table 2). All the values of the GBF for model 1 with respect to all other models is greater than 150, corresponding to very strong evidence for the serial model number 1. Plots of measured and predicted BOLD time series for a single participant are shown in Figure S3. This figure shows that the BOLD response in all three areas, particularly in STS, is fitted well by the optimal model. This demonstrates that (1) activity in the PT can be explained as a function of the input from the HG and its modulation during spectral envelope processing, and that (2) STS activity can be explained as a function of the input from the PT (compare the structure of model 1 as shown in Figure 2).
Estimates of the interregional connection strengths and their modulation for each participant and probabilities that the coupling estimates are greater than zero are shown in Tables 3 and 4, respectively. The probabilities that the connection strengths are greater than zero are all ~1.00, with the exception of the PT → STS connection in participant 5. Furthermore, the probability that the modulation of the strength of the connection from HG → PT is greater than zero is ~1.00 in all participants except 1 and 7, where the probability is greater than 0.9. A further t-test was carried out on intrinsic and modulatory connection strengths to assess the group level connection strengths. The mean values of HG → PT and PT → STS intrinsic connection strengths are 0.37 (p < 0.01) and 0.48 (p < 0.01), respectively. The mean value of modulatory HG → PT (measured in percentage increase) in connection strength is 109.29 (p < 0.01)
Theoretically, a large number of models other than those considered above are possible. The choice of these models was motivated by preliminary analysis of the data. This analysis showed that (1) inclusion of modulation of the HG → PT pathway is critical to model performance (as evaluated using AIC and BIC), and (2) addition of a feedback path (from PT → HG) led to poorer model performance. Since AIC and BIC strike a tradeoff between predictability and cost (measured in terms of number of parameters) of the model, this implies that the feedback path does not significantly increase the predictability but adds to the cost. We have estimated 54 further models that include back connections to HG from PT and STS (in both serial and parallel models) and HG → STS → PT models. These models, which are anatomically and functionally plausible, are schematically represented in Figure S1. A plot of posterior probabilities of all the 70 models, with the first 16 as shown in Figure 1 and the next 54 as shown in Figure S1, is shown in Figure S2.
On evaluating all the 70 estimated models (Figure S2), participants 1 to 6 continued to show very strong evidence in favour of model 1. In participants 7 and 8, however, the model selection procedure failed to identify an optimal model, although for different reasons. In participant 7, there was no decisive evidence in favour of any model: the Bayes factor for comparing model 10 to model 1 (the latter being the optimum model in the first six participants) was only 1.4. This designates very little evidence in favour of model 10 and is substantially below the threshold (i.e., 3) that is commonly used in Bayesian statistics to decide between two models [16]. In contrast, in participant 8, the two approximations to the model evidence (AIC and BIC) favoured different models (21 and 19, respectively). Similarly, model 1 was superior to model 21 according to the BIC criterion, but inferior according to the AIC criterion. These contradictory constellations represent a limitation of the model selection procedure adopted here, which, in cases like this particular participant, prevents one from drawing a firm conclusion about which model is optimal [14]. Overall, therefore, six out of eight participants showed strong evidence in favour of model 1, and the remaining two participants failed to show consistent evidence in favour of any one model.
A major challenge in auditory cognition is to relate cognitive processes to dynamic interactions among cortical regions. DCM was designed specifically for functional imaging data to model and draw inferences about effective connectivity between different regions. The present study aimed to understand the systems-level organisation of the computational mechanisms in the HG, PT, and STS invoked for the analysis of the spectral envelope of sounds. Two broad categories of models, serial and parallel, were specified a priori. The data provide very strong evidence for a serial model in which analysis of the spectral envelope specifically enhances the connection from the HG to the PT.
In contrast to the visual system [17–22], the effective connectivity between auditory areas has not been studied extensively. The few previous studies used structural equation modelling (SEM; [23–25]). The present study is the first to use DCM to examine the auditory system. Effective connectivity between the HG and the PT was suggested by a previous study [23] using SEM. However, the results were inconsistent across participants and also between the group analysis and individual participant analyses. Another study using SEM [25] considered the connection between the HG and PT and frontal areas, but did not examine local connection with the STS. In the present study, we have provided evidence for a consistent model across participants based on serial analysis in the HG, PT, and STS. The group analysis also concurs with the individual participant analysis. We now consider the limitations of the model supported by our data, and its biological significance.
DCM models the causal influence of the neural activity in one area on another, where those areas have a direct or indirect anatomical connection. The connections that we have specified a priori are plausible, given available macaque data [26], but data on the interconnections between human cortical areas are limited to a small number of postmortem dye-tracing studies [27] and “opportunistic” studies of neurosurgical patients [28]. The model supported here is both consistent with the existence of direct anatomical projections between the HG and the PT and between the PT and the STS, and also provides evidence for the functional expression of these projections. The connection between the HG and the PT is supported by the neurophysiological evidence and tracing studies above, but further evidence for the anatomical connection between the PT and the STS, particularly, is required. Such evidence might accrue from further postmortem work or the application of in vivo techniques such as diffusion tensor imaging.
In this study, we have tested the simplest possible model to describe the data in individual participants. In particular, the HG volume we have used is likely to contain primary and secondary areas: we have previously argued [29] that there are three functional areas in the HG that correspond to the three macaque “core” areas A1, R, and RT. Kaas and Hacket [26] have described a macaque scheme based on a pattern of connectivity that extends from core to belt to parabelt areas. The connectivity structure of the serial model that was selected as optimal in this study is consistent with such a scheme, if PT contains the homologues of belt areas. However, the detailed pattern of interconnections within the HG could not be assessed in this dataset, as three distinct functional areas in the HG were not demonstrated in all the individual participants. The extent to which the analysis may involve several different functional areas within the HG before analysis occurring in the PT therefore cannot be determined.
Like the HG, the PT is a large anatomical area, corresponding to the cytoarchitectonic area Te 3.0 [30], within which there may be a number of functional subdivisions. Homology with the macaque becomes even more difficult than in the case of core areas. One possibility is that there may be “belt homologue” areas in the PT adjacent to the three “core” areas suggested in the HG. The connectivity between the HG and the PT identified in this analysis would then be broadly congruent with the core-to-belt projections that have been identified in the macaque [26]. Recording work in the macaque suggests that more anterior belt areas are critical for auditory object analysis, although the distinction between anterior and posterior auditory areas is not as marked as in the case of spatial analysis [31]. Connections to the STS are much more difficult to characterise in terms of homology, especially in view of the existence of three temporal gyri in the human and two in the macaque. Human functional data for the STS demonstrate complex cognitive analysis, including voice processing [32] and the integration of auditory and visual object information [33]. Whether or not the macaque homology holds, however, the other human studies suggest a role for the STS in associative analysis, whereas the serial analysis we have demonstrated is also hierarchal: perceptual analysis in the earlier areas (HG and PT) precedes more complex associative analysis in the later area (STS).
The model identified as optimal by our Bayesian selection procedure is characterised by a serial architecture in which high activity in higher auditory areas during spectral envelope extraction is explained by a modulation of the HG → PT connection. In neurophysiological terms, this means that only the HG → PT connection is dependent on the spectral envelope modulation, and that the induced context-dependent response in the PT is simply relayed on to the STS. In functional terms, this means that spectral envelope analysis is likely to be completed at the stage of the PT, and that the differential responses in the STS are a downstream reflection of this process. In contrast, if we imagine that model 6 (compare Figure 2) had been selected as optimal, the interpretation would have been that context-dependent connectivity was restricted to the STS → PT connection and that functionally, spectral envelope analysis is likely to be performed at the level of the STS and the results fed back to the PT via the STS → PT connection.
Completion of spectral envelope analysis at the PT in the absence of a task is consistent with the “obligatory” abstraction of templates before the PT that does not depend on the existence of a task. It will be of interest in future studies to see if the presence of an active task produces modulation of the second serial stage between the PT and the STS. This is also interesting in terms of the idea that the PT may be a critical computational “hub” where spectro-temporal “templates” are extracted before analysis in higher centres that assess the significance of a particular template (such as its relevance to position in space or semantic category) [15]. This abstraction is homologous to feature extraction or selection in machine learning: feature selection is a process commonly used in machine learning, in which features available from the data are selected for subsequent inference and learning. The spectral envelope corresponds to a type of template that is independent of the spectral fine structure of the sound, and is important for source identification independent of the pitch of the source or whether it was producing a harmonic sound or noise.
There are certain limitations of the models that have been tested in the present work. First, only cortical connections have been considered. The thalamic connections were not included in the models, because of (1) the absence of activation in the auditory thalamus due to our experimental manipulation and (2) the evidence from animal work that complex spectral analysis first occurs in the primary auditory cortex [8]. Also, only models of the right hemisphere have been considered here, and hemispheric interactions have been ignored. This is because the conventional fMRI analysis of spectral envelope processing has consistently demonstrated a dominant role of the right hemisphere, with substantially less involvement of the left hemisphere.
Sequences of harmonic or noise stimuli were synthesised digitally at a sampling frequency of 44.1 kHz and 16-bit resolution. The harmonic stimuli were harmonic series, whereas the noise stimuli were random-phase noise. The stimuli were synthesised in the frequency domain, allowing the same spectral envelope to be applied to either harmonic or noise sounds. The duration of stimuli was 500 ms (with a 20-ms gating window). Synthesized sounds were used to form two sets of sequences. The first set, called “all-harmonic”, consisted of harmonic sounds only; the second set, known as “alternating”, consisted of alternating harmonic and noise sounds. The “all-harmonic” set has three experimental conditions: (1) the spectral envelope and pitch (fundamental frequency) of the sounds in the sequence are fixed; (2) the spectral envelope is fixed, but the fundamental frequency of sounds in the sequence is changing; and (3) the spectral envelope is changing, but fundamental frequency is fixed. The fundamental frequencies of the sounds in this set are 120, 144, 168, or 192 Hz, either fixed or varied between successive sounds in the sequence. The “alternating” set has two conditions: (1) harmonic and noise sounds alternating with fixed envelope and (2) harmonic and noise sounds alternating with changing spectral envelope. In total, the experiment has six conditions, five as described above, and the silence condition. The conditions are schematically shown in Figure 4. Change in fundamental frequency f0 is perceived as change in pitch, whereas change in the spectral envelope is perceived as a change in the identity of the source. The critical contrast in the group analysis to assess the “extraction” of the spectral envelope is the contrast between the two alternating conditions with changing and fixed spectral envelopes when the fine spectral structure of the stimuli is continually changing. The difference between these two conditions corresponds to an alteration in the perceived source over and above the low-level analysis of the fine spectral structure. That contrast was used to define spectral envelope extraction at the group level. The total duration of each sequence was 7.5 s or 8 s.
Before carrying out the fMRI experiment, the participant's ability to perceive the change in the spectral envelope was assessed in a separate psychophysical experiment. The same elements of the sequences used in the fMRI experiment were presented to the participants in a two-interval–two-alternative forced-choice paradigm. The task was to detect change in pitch (in all harmonic sequences) or change in the spectral envelope (all-harmonic or alternate conditions). Participants were able to detect harmonic sequences with pitch change or spectral shape change with 100% accuracy. Participants were also able to detect change in the spectral envelope in (alternating condition) with 100% accuracy. The changes in spectral shape therefore could be reliably detected independent of fine spectro-temporal changes.
Data from eight healthy volunteers were used for DCM. All participants gave their informed consent, and the experiment was carried out with approval of the local ethics committee.
fMRI data were acquired from a 1.5-T Siemens SONATA system (http://www.siemens.com) using gradient echo planar imaging (echo time = 50 ms; flip angle = 90 degrees) in a sparse image acquisition protocol [34]. Stimuli were presented diotically at a fixed sound-pressure level of 80 dB during the silent phase of the protocol. A whole-brain volume of 48 slices (2-mm thickness, in plane resolution 3 × 3 mm2) was acquired every 12.5 s with a time for acquisition of 4.32 s. Participants were instructed to attend to the stimuli with their eyes closed. In a typical trial of image acquisition, stimulus is first presented for about 8 s, followed by image acquisition that lasts 4.32 s. There was no active auditory discrimination task, however; to maintain attention, participants were asked to signal the end of each sequence by pressing a button box under the right hand. The experiment was divided in two runs, with 16 scans acquired for each condition in each run. The order of conditions was fully randomised.
Images were realigned, normalized to a standard EPI template, and smoothed with a 3-D Gaussian kernel with full-width half-maximum of 8 mm. Regressors for the design matrix were created by convolving boxcar stimulus functions (representing stimulus events) with a canonical hemodynamic response function. Linear contrasts of parameter estimates were created for each participant. Finally, a random-effects group analysis was performed by comparing the participant-specific contrast images with the appropriate t-tests to produce a statistical parametric map.
The general goal of DCM is to provide mechanistic explanations, in terms of connectivity and its modulation, for local effects observed in a conventional univariate analysis. The SPM results of the present data demonstrated a neural system in which the lowest level (i.e., the primary auditory cortex in the HG) does not show any significant activity differences between experimental conditions, but is uniformly driven by auditory stimulation. In contrast, higher auditory areas (PT and STS) show higher activity when spectral envelope extraction is required. These two observations could be potentially explained by a network model (i.e., DCM) in which a “neutral” input area, perturbed by auditory stimuli per se, drives two higher auditory areas differentially (i.e., some or all of the efferent connections of the input area are modulated by spectral envelope extraction).
Our DCM included three areas (HG, PT, and STS) in the right hemisphere. These areas were identified for each participant based on the coordinates of the peak activation obtained in the group analysis. For the HG, the contrast (condition 4 + condition 5) versus silence was used to define the centre of the volume from which the time series was extracted. For the PT and the STS, the contrast between the alternating sequences with variable and fixed spectral envelopes (condition 5 versus condition 4; see Figure 4) was used to define the centres of the volumes. The centre of each volume (defined as a sphere of 4-mm radius) was located at the local maximum that was nearest to the peak coordinates in the group analysis. The selected local maximum was constrained to lie within 16 mm (twice the width of the Gaussian smoothing kernel) of the group peak coordinates and within the same anatomical gyrus/sulcus as the group activation.
The coordinates of peak activation for the three volumes in each participant are given in Table 5. A summary time series from each of the three regions was furnished by the principal eigenvariate of measurements recorded from all significant voxels located within the volume.
From a system theory point of view, the brain can be treated as a nonlinear input–output dynamic system that can be excited by controlled stimuli and which response (hemodynamic response here) can be measured. The central idea behind DCM is to estimate and draw inferences about the causal interaction between different regions of the brain by identifying a model for the system using input–output measurements.
In DCM, three different sets of parameters are used. The first set of parameters, known as intrinsic parameters, models the anatomical or hardwired connection strengths between the regions. These parameters represent the influence that one region has over the other in the absence of any external excitation of the system. The second set of parameters, known as modulatory parameters, models the change in intrinsic connection strength that is induced by the external experimental input. These parameters are therefore input-specific and are also referred to as “bilinear terms or parameters.” The third set of parameters models the direct influence of an external stimulus on a given region. The conventional general linear model analysis is based on the assumption that any external stimulus has a direct influence on a region; therefore, it is the third set of parameters on which a general linear model analysis is based exclusively. DCM, therefore, can also be regarded as more general, with the general linear model analysis being a specific situation in which the interaction parameters (first and second sets) are assumed to be zero.
DCM has several advantages over other models of effective connectivity (e.g., SEM [20], multivariate autoregression [35], or Granger causality [36]; see [13,37] for details). For example, DCM takes temporal order (and autocorrelation of the fMRI time series) into account. It further allows one to model the effects of experimentally controlled manipulations as either affecting regional activity directly (e.g., sensory inputs) or modulating the strengths of connections, and does not need to assume that the system is driven by stochastic innovations. Most important, however, DCM is currently the only model of effective connectivity that combines a neural population model with a biophysical hemodynamic forward model, and is thus able to model how system dynamics at the (hidden) neuronal level translates into measured BOLD signals.
In brief, DCM is based on a bilinear model of neural population dynamics that is combined with a hemodynamic model [38,39], describing the transformation of neural activity into predicted BOLD responses. The neural dynamics are modelled by the following bilinear differential equation:
where z is the state vector (with one state variable per region), t is continuous time, and uj is the j-th experimental input to the modelled system (i.e., some experimentally controlled manipulation). This state equation represents the strength of connections between the modelled regions (the A matrix), the modulation of these connections as a function of experimental manipulations (e.g., changes in task; the B(1)...B(m) matrices), and the strengths of direct inputs to the modelled system (e.g., sensory stimuli; the C matrix). These parameters correspond to the rate constants of the modelled neurophysiological processes. Combining the neural and hemodynamic model creates a joint forward model, which is inverted using conventional techniques (expectation maximisation) to give the posterior density of the parameters. Under Gaussian assumptions, this density can be characterised in terms of its maximum a posteriori estimate and its posterior covariance. This density obtains by optimising a free-energy bound on the models log-evidence or marginal likelihood.
DCM is a hypothesis-driven technique in which model space is specified a priori. The first objective of the present study was to test if the coupling between the HG, PT, and STS is serial or parallel. To address this, two broad categories of models, serial and parallel, were specified (Figure 2). In the serial models, auditory inputs entering the HG reach the STS via the PT, and, thus, processing in the STS depends on inputs from the PT. In contrast, in the parallel models, the HG connects to both the PT and the STS, thus enabling a parallel processing in the PT and STS. The second objective was to determine where, in the best model, task requirements (i.e., spectral envelope analysis) led to changes (i.e., modulation) in the connection strengths. The modulatory input is defined as condition 5 of the experiment. In total, 16 models (nine serial, seven parallel) were inverted and compared using their log-evidence. These models are shown in Figure 2. To rule out the possibility of other theoretically possible models, 54 additional models shown in Figure S2 were also estimated, and their log-evidence was computed.
A general problem that arises in any modelling exercise is to decide, given some data, which of several competing models is optimal. A number of criteria have been proposed in the modelling literature [40]. From a Bayesian perspective, an optimal criterion is the model evidence (i.e., the probability p(y | m) of obtaining the data y given a particular model m [16]). Critically, the model evidence not only takes into account the relative fit of competing models, but also their relative complexity (i.e., the number of free parameters). This is important because there is a tradeoff between the fit of a model and its generalizability (i.e., how well it explains different datasets generated from the same underlying process). As the number of free parameters is increased, model fit increases monotonically, whereas beyond a certain point, model generalizability decreases. The reason for this is “overfitting”: an increasingly complex model will, at some point, start to fit noise that is specific to one dataset and thus become less generalisable across multiple realizations of the same underlying generative process.
As the model evidence cannot always be derived analytically, two commonly used approximations are the AIC and the BIC [14]. These approximations, however, do not necessarily give identical results because the BIC favours simpler models, whereas the AIC is biased toward more complex models. Here, we have adopted the usual conventions (compare [14]) (1) that a conclusion can only be drawn if these two criteria agree, and (2) that the more conservative of the two estimates is chosen. Finally, the relative evidence of one model as compared with another is expressed by the so-called “Bayes factor”:
where BF12 is the Bayes factor of model 1 with respect to model 2.
Following the selection of a best model for each participant, the optimal model for a group of participants can be determined by the GBF, which is the product of the Bayes factors for each individual participant [41]. |
10.1371/journal.pgen.1000887 | Transgenic Rat Model of Neurodegeneration Caused by Mutation in the TDP Gene | TDP-43 proteinopathies have been observed in a wide range of neurodegenerative diseases. Mutations in the gene encoding TDP-43 (i.e., TDP) have been identified in amyotrophic lateral sclerosis (ALS) and in frontotemporal lobe degeneration associated with motor neuron disease. To study the consequences of TDP mutation in an intact system, we created transgenic rats expressing normal human TDP or a mutant form of human TDP with a M337V substitution. Overexpression of mutant, but not normal, TDP caused widespread neurodegeneration that predominantly affected the motor system. TDP mutation reproduced ALS phenotypes in transgenic rats, as seen by progressive degeneration of motor neurons and denervation atrophy of skeletal muscles. This robust rat model also recapitulated features of TDP-43 proteinopathies including the formation of TDP-43 inclusions, cytoplasmic localization of phosphorylated TDP-43, and fragmentation of TDP-43 protein. TDP transgenic rats will be useful for deciphering the mechanisms underlying TDP-43–related neurodegenerative diseases.
| Amyotrophic lateral sclerosis, a condition also known as Lou Gehrig's disease, is characterized by progressive degeneration of motor neurons, denervation atrophy of skeletal muscles, and eventual paralysis of affected limbs. The signature pathology of Lou Gehrig's disease is the formation of intracellular inclusions containing phosphorylated TDP-43 protein. Most cases of Lou Gehrig's disease do not have a clear cause, while only about 10% of the cases are caused by mutation of individual genes. Here, we describe a novel rat model that expresses a mutated form of the human gene encoding TDP-43 and manifests the phenotypes and pathological features observed in patients with Lou Gehrig's disease. Laboratory rats are the preferred animals for pharmacological studies. Therefore, this new rat model will be useful not only for mechanistic study of Lou Gehrig's disease, but also for the development of therapies for this devastating disease.
| TAR DNA-binding protein (TDP-43) is a highly conserved ribonucleoprotein that is encoded by the TDP gene and can bind to RNA, DNA, and proteins [1]-[3]. In mammals, the primary transcript of the TDP gene can be alternatively spliced to generate 11 mRNA molecules. The major splice variant is full-length and encodes TDP-43 [4]. While the functions of this complex molecule remain largely unknown, ubiquitinated and phosphorylated TDP-43 accumulates in the nucleus and cytoplasm of affected cells in sporadic amyotrophic lateral sclerosis (ALS) and frontotemporal lobe degeneration (FTLD) [5],[6]. TDP-43 resides predominately in the nucleus and its translocation to the cytoplasm appears to be an early event in the pathological process underlying sporadic ALS [7]. At the end-stages of sporadic ALS and FTLD, C-terminal fragments of TDP-43 are remarkably increased in the brain [5],[6], but the full-length protein remains the major species in spinal cord [8], suggesting that regional differences exist in the metabolism and pathological mechanisms of TDP-43. Although TDP-43 proteinopathies have been identified in a wide range of neurodegenerative diseases including sporadic ALS, FTLD, Alzheimer's disease, and dementia with Lewy bodies [5]–[9], TDP-43 inclusions have not been detected in familial ALS caused by mutation of the SOD1 and FUS genes [10]–[13]. These findings imply that TDP-43 proteinopathy is common to neurodegenerative diseases and that divergent pathological processes may underlie sporadic and familial cases of ALS.
Mutations in the TDP gene segregate with ALS and FTLD associated with motor neuron disease (FTLD-MND) in geographically unrelated families [14]–[18], suggesting that TDP mutation is pathogenic in a subset of neurodegenerative diseases. Transient expression of the mutant, but not the normal, human TDP gene leads to apoptotic death of spinal motor neurons in chicken embryos [15]. In Drosophila melanogaster, depletion of the TDP homolog results in deficient locomotor activity and defects at neuromuscular junctions (NMJs) [19]. Suppression of TDP gene expression induces cell death in cultured neuroblastoma cells [20]. Previous studies indicate that mutation of the TDP gene is neurotoxic and that normal TDP-43 is important to cellular function; however, how mutations in the TDP gene cause neurodegeneration remains unknown.
To study the consequences of TDP mutation in an intact system, we expressed a mutant form of the human TDP gene in rats, which were chosen over mice because they are the preferred animals for pharmacological studies. Overexpression of a mutant, but not the normal, human TDP gene caused widespread neurodegeneration, which predominantly affected the motor system. Transgenic rats that constitutively or conditionally expressed a mutant form of human TDP with a valine-to-methionine substitution at position 337 (M337V) developed similar phenotypes at early ages, the phenotypes that were characterized by motor neuron degeneration accompanied by astrocyte and microglial activation in the spinal cord.
TDP-43 is widely expressed in mammalian tissues [21]. To mimic the expression profile of the endogenous TDP gene, we extracted the minimal human TDP gene (mini TDP gene) from a BAC clone and discarded the excess flanking sequences. The mini human TDP transgene contains essential elements for regulating transgene expression but does not carry unwanted genes into transgenic rats (Figure 1A). Among all known mutations in the TDP gene, the M337V substitution is found in geographically unrelated families and thus is an excellent representative of TDP gene mutations [15],[17]. We introduced the M337V mutation into the mini TDP transgene using a recombineering technique [22]. Using pronuclear injection, we generated three transgenic founders (two males: founders 1 and 2; one female: founder 3) that robustly expressed the miniTDP43M337V transgene (Figure 1A–1C and 1F). The mutant TDP transgenic founders were indistinguishable from their nontransgenic littermates at birth; however, they soon lost mobility and died at postnatal ages. Founder 3 died at the age of 10 days. Founder 2 showed weakness in the limbs at the age of 13 days and became paralyzed by the age of 18 days. Founder 1 showed weakness in a forelimb at the age of 21 days and became paralyzed by the age of 29 days. We examined founder 1 using immunohistochemistry and observed a reduction in motor neurons in the ventral horn of the lumbar spinal cord (Figure 1F). Since none of the mutant TDP (miniTDP43M337V) transgenic rats survived to sexual maturity, mutant TDP transgenic lines could not be established. In parallel, we generated two transgenic founder rats that carried the normal human TDP transgene (miniTDP43wt), which had an identical DNA composition as miniTDP43M337V except that it lacked the M337V mutation (Figure 1A–1E). The miniTDP43wt transgenic rats expressed human TDP-43 protein at levels comparable to those detected in the miniTDP43M337V transgenic founder rats but did not develop paralysis by the age of 200 days. These findings suggest that the disease phenotypes observed in the miniTDP43M337V transgenic founder rats result from toxicity of the TDP gene mutation.
Since constitutive expression of a mutant human TDP gene caused a severe phenotype in transgenic founders, we used a tetracycline (Tet) regulatory system to express the mutant TDP transgene in a controlled manner. In this way, we could establish transgenic rat lines expressing the human TDP transgene with a pathogenic mutation. The Tet-off system is commonly used in transgenic studies and is comprised of only two elements— a Tet-controlled transactivator (tTA) and a tTA-activated promoter (TRE) [23]. Using pronuclear injection, we established two transgenic lines (line number corresponds to transgene copy) that carry 7 or 16 copies of the TRE-TDP-43M337V transgene under the control of the TRE promoter (Figure 2A). The transcriptional activator, tTA, is inactive in the presence of the Tet derivative, Doxycycline (Dox), allowing for inactivation of a TRE promoter-controlled gene through Dox administration in the bigenic rats that carry the TRE-TDP-43M337V and the tTA transgenes (Figure 2A). In the absence of Dox, tTA constantly activates the TRE-TDP-43M337V transgene, producing an expression pattern that is indistinguishable from constitutive transgene expression [24].
Constitutive expression of the miniTDP-43M337V transgene caused postnatal death in the transgenic founder rats (Figure 1), suggesting that the mutant TDP gene is highly toxic. To test whether the severe phenotype observed in the constitutive transgenic rats could be reproduced in conditional transgenic rats, we produced the TRE-TDP-43M337V and tTA double transgenic rats by crossing the TRE-TDP-43M337V transgenic lines with a tTA transgenic line that expresses the tTA transgene at levels sufficient to vigorously activate tTA reporter genes [24]. To obtain a constitutive pattern of transgene expression, we allowed the TRE-TDP-43M337V transgene to be expressed from early embryogenesis by withholding Dox treatment. Consistent with findings in constitutive transgenic rats (Figure 1), expression of the TRE-TDP-43M337V transgene from early embryonic stages caused severe phenotypes in the conditional transgenic rats of line 16 (Figure 2B). Transgenic rats of line 16 became paralyzed and died by postnatal day 20 (P20). The similarity in phenotypes between the constitutive and conditional transgenic rats indicates that the observed defects did not result from an insertional mutation.
Expression of the TDP-43M337V transgene from early embryogenesis caused early death in transgenic rats, making functional analysis of this model a challenge. To facilitate analysis of motor function, we added Dox to the drinking water (50 µg/ml) of breeding rats to suppress transgene expression during embryonic development. We then withdrew Dox at 4 days before delivery to allow for recovery of transgene expression in postnatal rats. As a result, the transgene was not expressed in newborn pups but was fully expressed in postnatal rats by P10 (Figure S1). The TRE-TDP-43M337V transgenic rats of line 16 showed a rapid progression of disease phenotypes, exhibiting limb weakness by P20 and paralysis before P35 (Figure 2E). In contrast, the TRE-TDP-43M337V transgenic rats of line 7 showed a later onset and a slower progression of similar phenotypes (Figure 2C–2E). Disease progression in line 7 could be divided into four distinct stages [25]: the nonsymptomatic stage, disease onset, the paralysis stage, and the disease end stage. Disease onset was defined as an unrecoverable reduction in running time on a rotating Rotarod. The paralysis stage was defined as visible dragging of a limb. The disease end stage was defined as paralysis in two or more limbs. Postnatal rats aged 21 days were subjected to a Rotarod test to determine disease onset (Figure 2D). Since transgenic rats of line 16 developed early paralysis and had a rapid disease progression, determining the time of disease onset for this high-copy line was technically difficult. Transgenic rats of line 16 showed limb weakness by an age of 20 days and became paralyzed in the legs by an age of 35 days, with no sexual dimorphism existing in the rate of disease progression (Figure 2E). In contrast, transgenic rats of line 7 displayed sexual dimorphism in the time of disease onset and in the rate of disease progression (Figure 2D and 2E). Sexual dimorphism in phenotypic onset has also been observed in an ALS animal model expressing mutant human SOD1 genes [26]–[29]. The disease phenotypes observed in the mutant TDP (TRE-TDP-43M337V) transgenic rats were not observed in normal TDP transgenic rats (miniTDP-43WT) by an age of 200 days, though these rats expressed the human TDP transgene at comparable levels as TRE-TDP-43M337V rats (Figure 2B–2E). An examination of TRE-TDP-43M337V transgenic offspring revealed that, consistent with findings in miniTDP-43M337V transgenic founders (Figure 1), the disease phenotypes in these animals were related to mutation of the TDP gene.
Anatomical analysis revealed that motor neurons in the spinal cord robustly expressed the human TDP transgene (Figure 3A–3C). The number of spinal motor neurons was significantly reduced in mutant TDP transgenic rats but not in normal TDP transgenic rats (Figure 3D–3F and 3I), although the mutant and normal TDP transgenic rats expressed human TDP-43 at comparable levels (Figure 2B). Large-caliber neurons were preferentially affected in mutant rats at the end stages of disease (Figure 3D–3F and 3L). During the paralysis stage, degenerating axons were clearly visible in the ventral (Figure 3G–3I and 3M) and dorsal roots (Figure S2), with motor axons of the corticospinal track also being affected (Figure S3). Confocal microscopy revealed that denervation of synaptic endplates in skeletal muscle occurred at disease onset (Figure 4B and 4D) and worsened at the end stage of disease (Figure 4C and 4D). Electron microscopy confirmed that, in the mutant transgenic rats, degeneration of motor neuron axons occurred at disease onset (Figure 4F and 4G); however, no loss of motor neurons was detected in the mutant TDP transgenic rats at this time. These findings suggest that axon terminals are the primary targets of degeneration associated with pathogenic mutation of TDP. In the mutant TDP transgenic rats, denervation of skeletal muscle fibers was confirmed by electromyography, which detected frequent fibrillation potentials—a characteristic of muscle denervation and regeneration (Figure 4E). As results of denervation, groups of skeletal muscles were atrophied (Figure 4H and 4I). These pathological changes were correlated with progressive paralysis in the mutant transgenic rats (Figure 2C–2E).
Silver staining revealed that, in mutant TDP transgenic rats, spinal motor neurons degenerated during end-stage disease (Figure 5A and 5B). A previous study showed that transient expression of the mutant, but not the normal human TDP gene, causes apoptotic death in the spinal cord of chicken embryos [15]. Consistent with the finding from this transient transfection study [15], motor neurons in the spinal cord underwent apoptosis in paralyzed transgenic rats (Figure 5C and 5D). Studies of mutant SOD1 mice suggest that glial cells play an important role in ALS pathogenesis [28], [30]–[33]. Therefore, we examined glial reactions in our paralyzed rats. We found that astrocytes and microglia were increased around the motor neurons in the spinal cord (Figure 6A–6D). The finding suggests that a glial reaction occurs in response to motor neuron degeneration.
TDP-43 inclusions are found in the brain and spinal cord of patients with sporadic ALS, FTLD, Alzheimer's disease, or dementia with Lewy bodies [5]–[9], suggesting that TDP-43 proteinopathies are common to neurodegenerative diseases. Pathogenic mutations in the TDP gene have been identified not only in ALS, but also in FTLD-MND [14]–[18]. Degeneration associated with mutations in the TDP gene may not be restricted to motor neurons. Indeed, silver staining revealed that neurodegeneration occurred in the cortex, hippocampus, and cerebellum of mutant transgenic rats with end stages of disease (Figure 7A–7F) but not in those with earlier stages of disease (Figure 2D and data not shown). Nevertheless, degenerating neurons were not detected in the substantia nigra of paralyzed rats (data not shown), despite the fact that transient overexpression of the normal human TDP gene in rats has been shown to induce a loss of dopaminergic neurons in this brain region [34]. Neuropathological findings were correlated with phenotypic expression in mutant TDP transgenic rats (Figure 2, Figure 3, Figure 4). Toxicity of the pathogenic TDP gene mutation was not restricted to motor neurons, though these neurons were affected by the mutation to a greater degree than all the other neuron types examined.
Phosphorylated TDP-43 inclusions are a signature pathological feature of sporadic ALS and FTLD [5], [6], [35]–[37]. To detect phosphorylated TDP-43 inclusions in our transgenic rats, we tested a polyclonal antibody specific to phosphorylated TDP-43 on brain sections of FTLD patients and TDP transgenic rats. This phospho-TDP-43 antibody detected cytoplasmic accumulation of phosphorylated TDP-43 in FTLD patients, but not in control subjects (Figure 8A). Similarly, phosphorylated TDP-43 was diffusely distributed in affected neurons in transgenic rats expressing the mutant or normal human TDP transgene (Figure 8D). We generated a polyclonal antibody recognizing both phosphorylated and non-phosphorylated human TDP-43 (Figure 1C–1F) and detected a robust expression of the human TDP transgene in transgenic rats (Figure 8B and 8C). TDP-43 was diffusely distributed in the nucleus and cytoplasm of cells within transgenic rats (Figure 8). However, TDP-43 inclusions were detected rarely, being present only in the cortex (Figure 8B) and not in the spinal cord (Figure 8C) of transgenic animals.
Immunohistochemistry revealed that typical ubiquitin-positive inclusions were not present in the spinal cords of normal or mutant TDP transgenic rats, though the intensity of ubiquitin immunostaining was greater in these animals than in nontransgenic rats (Figure S4). Since TDP-43 inclusions were rare in transgenic rats, even at end-stage disease, we further examined TDP-43 ubiquitination using immunoprecipitation combined with immunoblotting analysis. Ubiquitinated TDP-43 was detected in the mutant TDP transgenic rats (Figure S4). Immunoblotting revealed that a small amount of TDP-43 fragments (less than 43 kDa) was present in TDP transgenic rats (Figure 2B and Figure S5). TDP-43 fragments were detected in urea tissue extracts from rats at the paralysis stage, but not in extracts from those at disease onset (Figure S5). The finding suggests that the solubility of the small TDP-43 fragment is reduced as the disease progresses.
Expression of the human TDP gene containing a M337V substitution reproduced the phenotypes of ALS in rats. That is, these animals exhibited progressive degeneration of motor neurons and denervation atrophy of skeletal muscles. In this transgenic rat model, neurodegeneration was not restricted to motor neurons and could be seen in other types of neurons including cortical neurons, hippocampal neurons, and cerebellar neurons. However, TDP mutation affected motor neurons earlier and more severely than other neurons in the central nervous system (CNS). This robust rat model also recapitulated features of TDP-43 proteinopathies, including the formation of TDP-43 inclusions, cytoplasmic localization of phosphorylated TDP-43, and fragmentation of TDP-43.
While our transgenic rat model developed the phenotypes of ALS, it displayed degeneration of CNS neurons other than motor neurons at the end stages of the disease. Our findings in mutant TDP transgenic rats do not necessarily contradict observations in ALS patients. ALS is traditionally thought to affect only motor neurons, but recent studies showed that neurons other than motor neurons also degenerate in ALS [38]. This point is strikingly illustrated by the observation in some ALS patients who live with the disease much longer than the average disease duration [38]–[40]. Moreover, some ALS and FTLD cases share symptoms and pathological characteristics [41]. Although mutations of the TDP gene are primarily associated with ALS [14]–[17], a recent study found that a novel mutation in the TDP gene is associated with FTLD-MND [18], suggesting that the toxicity (if any) of the TDP gene mutation is not restricted to motor neurons [18]. Further studies are warranted to ascertain whether a correlation exists between the pathological changes induced by TDP mutation and TDP-43 proteinopathies observed in sporadic ALS and FTLD. The fact that TDP-43 proteinopathy is observed in a wide range of neurodegenerative diseases suggests that mutations in the TDP gene are generally neurotoxic [5], [6], [9], [42]–[45]. Neurodegenerative diseases may share common pathological mechanisms, with a certain subgroup of neurons being predominantly affected under each disease condition. Our mutant TDP transgenic rat is a robust model of neurodegeneration caused by mutation of the TDP gene.
Many features of TDP-43 proteinopathies were reproduced in our TDP transgenic rats. Redistribution, phosphorylation, and aggregation of TDP-43 are all hallmarks of sporadic FTLD and ALS [5],[44],[45]. A recent clinical study showed that TDP-43 redistribution appears to be an early event in TDP-43 proteinopathy [7], suggesting that TDP-43 redistribution underlies the pathogenesis of neurodegeneration. Our results showed that phosphorylated TDP-43 was diffusely distributed in the cytoplasm and nucleus of affected cells in paralyzed mutant TDP transgenic rats as well as in non-paralyzed, normal TDP transgenic rats. The presence of phosphorylated TDP-43 in normal TDP transgenic rats does not exclude the possibility that TDP-43 phosphorylation contributes to pathogenesis induced by TDP mutation. Specifically, TDP mutation may impart toxicity by enhancing the normal functions of the gene. For example, mutation of the LRRK2 gene causes Parkinson's disease by enhancing (at least partially) the kinase activity of LRRK2 [46],[47]. Gene mutations can be classified into three types based on their effect on protein function: gain of function, loss of function, and dominant negative effect. Pathogenic mutation of the TDP gene may cause disease through any one of these three effects on protein function. Resolving the nature of the TDP gene mutation will require a more sophisticated model such as a knockin mouse. TDP-43 inclusions and fragmentation were rarely observed and were present only at end-stage disease, suggesting that these pathologies may be consequence of, rather than a cause of, neurodegeneration in TDP transgenic rats. C-terminal truncated products of TDP-43 are thought to result from caspase cleavage of full-length TDP-43 [48]. Accordingly, C-terminal fragmentation of TDP-43 is likely a consequence, instead of a cause, of neurodegeneration because caspase activation is a terminal feature of cell death. In addition, we cannot exclude the possibility that overexpression of the TDP transgene interferes with rat development, since the mutant TDP transgenic rats died at postnatal ages.
Typical ALS has a late onset and rapidly progresses [12], [13], [17], [29], [49]–[51]. In contrast, mutant TDP transgenic rats developed paralysis at early ages, with the paralysis being similar to that seen in ALS. Early onset of disease phenotypes in our rat model likely results from toxicity of the TDP gene mutation, as evidenced by the following three findings. First, paralysis and lethality were observed in the mutant miniTDP43M337V transgenic founders, but not in the normal miniTDP43WT transgenic founders. Second, paralysis and neurodegeneration were observed in the inducible mutant TDP transgenic rats, but not in offspring of the constitutive normal TDP transgenic rats, despite the fact that both lines exhibited comparable expression of the human TDP transgenes. Third, similar phenotypes were observed in the constitutive mutant TDP transgenic founders and in the inducible mutant TDP transgenic offspring. One transgenic founder rat carried only six copies of the mini mutant TDP transgene and developed paralysis in postnatal age. The copy number of the mutant TDP transgene that is required for phenotypic expression in transgenic rats is much lower than the copy threshold of mutant SOD1 transgenes [52],[53]. To activate the inducible mutant TDP transgene, we used a low-copy tTA transgenic line that carries only two copies of the tTA transgene [24]. Therefore, expression levels of the TDP transgene in the inducible transgenic rats were comparable to those in the constitutive normal TDP transgenic rats. Transgenic rats expressing the mutant TDP gene displayed a wider range of neurodegeneration than transgenic rodents expressing mutant SOD1 genes [29], [52]–[54], with neurodegeneration predominantly affecting the motor system. Such unrestricted toxicity of the TDP gene mutation may lead to an early onset of the disease. In some aspects, phenotypes observed in our transgenic rats are similar to those detected in transgenic mice that express the human TDP gene with a A315T substitution [55]. In these rodent models, both upper and lower motor neurons are affected and TDP-43 inclusions are rare. However, our rat model developed paralysis at postnatal ages and experienced a rapid disease progression, while the mutant TDP transgenic mice develop disease phenotypes during middle age and have varying disease durations [55]. Different mutations in the TDP gene and different animal species may contribute to phenotypic variation between the rat and mouse models.
Our findings in TDP transgenic rats indicate that mutation of the TDP gene is highly toxic in rodents, though the nature of the pathogenic mutation in the TDP gene remains to be resolved. Since deletion of the TDP gene in Drosophila causes defects at NMJs [19], the possibility that the TDP gene mutation produces a dominant-negative effect cannot be excluded. Although the nature of TDP gene mutation will need to be determined using a more sophisticated model, our TDP transgenic rats will be useful for mechanistic study of TDP-43-related neurodegenerative diseases.
Animal use followed NIH guidelines. The animal use protocol was approved by the Institutional Animal Care and Use Committees (IACUC) at Thomas Jefferson University. The Committee for Oversight of Research Involving the Dead at the University of Pittsburgh School of Medicine approved the use of human tissue from the University of Pittsburgh ALS Tissue Bank. Age-matched tissue sections from two FTLD and two non-neurological disease controls were used for the study.
The 22-kb mini human TDP gene was extracted from a BAC clone (RP11-829B14), and a M337V substitution was introduced into the mini TDP gene by homologous recombination in Escherichia coli [22]. The normal and mutant mini TDP transgenes were linearized by restriction digestion, purified from agarose gels, and then used to produce transgenic rats through microinjection. To generate Tet-regulatable TDP transgenic rats, we PCR-amplified the human TDP-43 ORF from a human brain cDNA pool (Invitrogen) and generated a mutant carrying the M337V substitution using site-directed mutagenesis (Stratagene). The mutated human TDP-43 cDNA gene was inserted downstream of a tTA-dependent promoter that was constructed by fusing seven tetracycline-responsive elements (TRE) with a minimal cytomegalovirus promoter (TRE-miniCMV). To enhance gene splicing and expression, we inserted the first intron of the human ubiquitin C gene between the TRE-miniCMV promoter and the TDP-43 ORF [24].
Linearized miniTDP43 and TRE-TDP43 transgenes were injected into the pronuclei of fertilized eggs of Sprague-Dawley rats. The injected eggs were then transplanted into pseudopregnant females for embryonic development [56]. Transgenic founders carrying miniTDP-43 transgenes were analyzed for disease phenotypes. Transgenic founders carrying TRE-TDP-43 transgenes were crossed with CAG-tTA transgenic rats to produce double transgenic offspring, which were analyzed for transgene expression and disease phenotypes. The TDP transgenic rats were identified by PCR amplification of the human TDP gene using the following primer pair: 5′-TGCGGGAGTTCTTCTCTCAG (forward) and 5′-AGCCACCTGGATTACCACCA (reverse). The copy number of the transgene was determined by quantitative PCR using two primer pairs. The first primer pair was designed to amplify a DNA fragment of the same composition from both the human and the rat TDP gene: 5′-TGAGCCCATTGAAATACCATC-3′ and 5′-TACACTGAGACACTGGATTC. The second primer pair was designed to amplify the rat prolactin gene as an internal control: 5′-CCTCTATGAACGAAACCCAC-3′ and 5′-CTTCCGGCTAATCCA CAATG-3′.
A rabbit polyclonal antibody was produced by Genemed Company. Rabbits were immunized with the synthetic peptide, (N-terminal)-EDELREFFSQYGDVM. Antiserum was then affinity-purified using a peptide-conjugated column (Pierce).
Under deep anesthesia, animals were transcardially perfused with 1X PBS (pH 7.4) and then with 4% paraformaldehyde (PFA) dissolved in 1X PBS buffer. The brain, spinal cord, and gastrocnemius muscle of perfused animals were collected and further fixed in the same fixative overnight. Some tissue blocks were embedded in paraffin and sectioned into 10 µm-thick slices. Paraffin-embedded sections were treated with 10 mM sodium citrate buffer (pH 6.0) to retrieve antigens for immunostaining. Paraffin-embedded coronal sections of the brain and transverse sections of the spinal cord were deparaffinized and immunostained with human TDP-43-specific antibody (1∶1,000; made in house) or a phospho-TDP-43-specific antibody (1∶1,000; COSMO Bio Co., TIP-PTD-P02). Immunostaining was visualized using an ABC kit in combination with diaminobenzidine (Vector). The immunostained sections were lightly counterstained with hematoxylin to display nuclei. After antigen retrieval, paraffin-embedded sections of the lumbar spinal cord were immunostained for human TDP-43 (1∶300) and ChAT (goat antiserum; Millipore). Immunofluorescence staining for human TDP-43 (red) and ChAT (green) was visualized using a Nikon fluorescence microscope, and images were acquired using a Nikon digital camera. Paraffin-embedded sections of the gastrocnemius muscle were stained with hematoxylin and eosin (H&E) to visualize tissue structures.
For NMJ detection, gastrocnemius muscles were fixed in 4% PFA for 2 h and sectioned on a cryostat into 100 µm-thick sections. Serial sections of the muscles were incubated with α-bungarotoxin (Invitrogen) for 30 min, washed in PBS three times, incubated overnight with mouse monoclonal antibodies to neurofilament (Sigma) and synaptophysin (Millipore), and then incubated for 1 h with a secondary antibody (FITC goat anti-mouse IgG1; Jackson Immunology).
For detection of apoptotic cells and glial cells, 4% PFA-fixed lumbar spinal cords were cut into three sets of 10 µm-thick serial sections on a cryostat. Every first section was incubated with TUNEL staining reagent (Millipore) and goat anti-ChAT antibody. Every second section was incubated with the ChAT antibody and mouse anti-GFAP. Every third section was incubated with the ChAT antibody and mouse anti-CD68 antibody. Sections were then incubated with appropriately labeled secondary antibodies. The antibodies were purchased from Millipore. Images were captured using a Zeiss LSM510 META confocal system. The NMJ was reconstructed using z-stack projections produced from serial scanning every 1 µm.
Fresh gastrocnemius muscle was snap-frozen in liquid nitrogen and cut into 12 µm-thick sections on a cryostat. Nonspecific esterase activity was detected using the α-napthyl acetate protocol. Denervated muscle fibers were stained a red-brown color, with normal fibers displaying a yellow-to-brown color.
Degenerating neurons were visualized using the Bielschowski silver-staining method as well as the FD NeuroSilver kit (FD Neurotechnologies, Baltimore, MD). For the Bielschowski silver method, paraffin-embedded spinal cord was transversely cut into 10 µm-thick sections. For staining using the FD NeuroSilver kit, 40 µm-thick coronal sections were obtained by slicing through the forebrain and cerebellum using a cryostat and then stained per the manufacturer's instructions.
Rats were anesthetized and perfused with a mixture of 4% PFA and 2% glutaraldehyde in 0.1 M phosphate buffer (pH 7.4). The L3 and L4 ventral and dorsal roots were removed and post-fixed in the same fixative at 4°C overnight. The roots were then further fixed in 1% osmium tetroxide in 0.1 M phosphate buffer (pH 7.4) for 1 h. The well-fixed tissues were dehydrated in graded ethanol and embedded in Epon 812 (Electron Microscopic Sciences, Fort Washington, PA). Thin sections (80 nm) were then stained with uranyl acetate and lead citrate for observation under a transmission electron microscope (Hitachi H7500-I). For toluidine staining, roots were transversely cut into 1 µm-thick sections. Axons in the nerve roots were examined in the semi-thin sections under a light microscope (Olympus AX70).
A 1-mm central segment of the L3 spinal cord was cut into 30-µm thick sections using a cryostat. Every third section was stained with cresyl violet and mounted in sequential order (rostral-caudal). Neurons with a diameter larger than 25 µm were counted in the ventral horns on both sides. The number of targeted neurons was estimated using a fractionator-based unbiased stereology software program (Stereologer), which was run on a PC computer that was attached to a Nikon 80i microscope with a motorized XYZ stage (Prior). At low magnification (40x), the targeting area was outlined, and a random sampling grid was created. At 1000× magnification, an optical dissector probe was randomly generated by the program in the designated area. The presence of clearly definable neurons was noted according to defined inclusion and exclusion limits of the dissector. This process was repeated on all selected sections. The total number of defined neurons was calculated by the software based on values obtained from random counts.
Animals were anesthetized during electromyography (EMG) examination. The fibrillation potential of the gastrocnemius muscle was recorded with an EMG instrument (CMS6600; COTEC Inc.) using a 27-gauge monopolar needle electrode and a 29-gauge reference needle electrode. During recording, a sub-dermal ground electrode was placed in the forelimb. Spontaneous electrical activity of selected skeletal muscle was recorded for 2 min.
The number of defined neurons in the ventral horn was compared between groups of transgenic rats. The difference in the number of neurons was analyzed using an unpaired t test. The null hypothesis was rejected at the 0.05 level.
|
10.1371/journal.pgen.1003963 | Evidence That Masking of Synapsis Imperfections Counterbalances Quality Control to Promote Efficient Meiosis | Reduction in ploidy to generate haploid gametes during sexual reproduction is accomplished by the specialized cell division program of meiosis. Pairing between homologous chromosomes and assembly of the synaptonemal complex at their interface (synapsis) represent intermediate steps in the meiotic program that are essential to form crossover recombination-based linkages between homologs, which in turn enable segregation of the homologs to opposite poles at the meiosis I division. Here, we challenge the mechanisms of pairing and synapsis during C. elegans meiosis by disrupting the normal 1∶1 correspondence between homologs through karyotype manipulation. Using a combination of cytological tools, including S-phase labeling to specifically identify X chromosome territories in highly synchronous cohorts of nuclei and 3D rendering to visualize meiotic chromosome structures and organization, our analysis of trisomic (triplo-X) and polyploid meiosis provides insight into the principles governing pairing and synapsis and how the meiotic program is “wired” to maximize successful sexual reproduction. We show that chromosomes sort into homologous groups regardless of chromosome number, then preferentially achieve pairwise synapsis during a period of active chromosome mobilization. Further, comparisons of synapsis configurations in triplo-X germ cells that are proficient or defective for initiating recombination suggest a role for recombination in restricting chromosomal interactions to a pairwise state. Increased numbers of homologs prolong markers of the chromosome mobilization phase and/or boost germline apoptosis, consistent with triggering quality control mechanisms that promote resolution of synapsis problems and/or cull meiocytes containing synapsis defects. However, we also uncover evidence for the existence of mechanisms that “mask” defects, thus allowing resumption of prophase progression and survival of germ cells despite some asynapsis. We propose that coupling of saturable masking mechanisms with stringent quality controls maximizes meiotic success by making progression and survival dependent on achieving a level of synapsis sufficient for crossover formation without requiring perfect synapsis.
| Diploid organisms must produce haploid gametes prior to sexual reproduction in order to maintain a constant number of chromosomes from one generation to the next. Ploidy reduction is accomplished during meiosis and requires crossover recombination-based linkages between homologous chromosomes. Here, we manipulate karyotype in C. elegans to probe the mechanisms that govern stable, pairwise, homologous associations essential for crossover formation. We find that chromosomes sort into homolog groups regardless of number prior to stabilizing interactions (“synapsing”) in a preferentially pairwise manner. Increased numbers of homologs delay meiotic progression and/or boost cell death, reflecting operation of quality control mechanisms that either buy time to correct synapsis problems or eliminate defective cells. Moreover, we found evidence for mechanisms that can “mask” synapsis imperfections, thus allowing resumption of meiotic progression and survival of germ cells when synapsis is “good enough”, albeit imperfect. This strategy would maximize meiotic success by making progression and survival contingent on achieving a level of synapsis sufficient for crossover formation without imposing an onerous and unnecessary requirement for perfect synapsis. We suggest that the regulatory logic of coupling saturable masking mechanisms with stringent quality controls may be employed widely to maximize efficiency of biological circuits.
| Sexually reproducing organisms must undergo reduction in ploidy during gametogenesis in order to maintain a diploid chromosome complement from one generation to the next. Ploidy reduction is accomplished at the first division of meiosis, when homologous chromosomes segregate to opposite spindle poles. Segregation of homologs is enabled by a multi-step program of events during meiotic prophase that culminates in most organisms in the formation of recombination-based linkages (chiasmata) between each chromosome pair. Maturation of recombination intermediates into crossovers that can provide the basis of chiasmata occurs in the context of a transient, meiosis-specific structure known as the synaptonemal complex (SC). The SC is a highly ordered proteinaceous structure that assembles at the interface between paired homologs and stabilizes their alignment (reviewed in [1], [2]). While the SC usually connects homologs along their full lengths, evidence suggests that its components function locally to promote crossovers [3]. The SC then disassembles as bivalent chromosomes joined by chiasmata prepare for segregation.
Assembly of the SC (termed synapsis) between paired homologs takes place early in meiotic prophase during a period of active chromosome movement. In organisms from yeasts to mammals, chromosomes within the nucleus become attached to the cytoskeletal motility apparatus in the cytoplasm via a conserved nuclear envelope (NE)-spanning protein complex [4]. In the C. elegans system, attachment occurs at specialized domains located near one end of each chromosome called “pairing centers” (PCs) [5], [6]. PCs are bound by a family of DNA binding proteins that associate, in turn, with the NE-spanning SUN-1/ZYG-12 complex, enabling motor-driven, microtubule-dependent chromosome movements [7]–[13]. In C. elegans, this chromosome mobilization is important for efficient and timely homolog pairing and for regulation and/or efficient propagation of synapsis, but continues after both processes are largely complete [11], [14]–[16]. Further, recombination is completed only after mobilization ends at mid-prophase [11]. Thus, timely entry into and exit from the chromosome mobilization phase are both critical for forming crossovers between homologs.
Proper coordination of the events of the meiotic program is essential to a successful outcome. Recent work in C. elegans has highlighted the importance of checkpoint-like coupling mechanisms that make progression of the meiotic program contingent upon successful execution of prerequisite events. For example, licensing of SC assembly is coupled to homolog identification through a mechanism that likely acts at the level of the PCs, which locally stabilize homolog pairing independent of synapsis, and also promote synapsis between homologs and/or inhibit non-homologous synapsis [9], [10], [17]–[19]. This coupling is critical because the SC assembles in a cooperative manner, yet is structurally indifferent to homology. Another coupling mechanism makes exit from the chromosome mobilization phase contingent upon successful SC assembly [17], [19]–[21]. Exit is delayed in response to incomplete synapsis, presumably allowing more time to ensure that all chromosome pairs are synapsed, and thus competent to form crossovers, before meiocytes proceed to the next stage of meiotic prophase. Finally, several recent studies suggest that meiocytes also have the capacity to detect the presence or absence of sufficient crossover-competent recombination intermediates per se, and to respond by enabling or delaying a major transition affecting multiple distinct aspects of the meiotic program [21]–[23].
Coupling mechanisms represent a class of quality control systems that contribute to meiotic success by promoting steps essential to the formation of crossovers between all chromosomes. A second class of quality control systems is represented by checkpoint mechanisms that prevent meiotic segregation errors by eliminating defective cells prior to completion of meiosis. A recombination/DNA damage checkpoint detects unrepaired recombination intermediates in late meiotic prophase and triggers apoptosis of affected meiocytes [24]. In addition, a synapsis checkpoint can trigger apoptosis in response to synapsis failure per se [25]. Thus, coupling mechanisms and checkpoints together exert stringent quality control over the meiotic program to ensure the production of normal haploid gametes.
Our current understanding of homolog pairing and synapsis in C. elegans, as well as how these processes are coordinated and monitored by quality control systems, has been derived largely from analysis of mutants that lack components of the meiotic machinery. While this approach has been highly fruitful, we reasoned that new insights could be gained from an alternative approach: manipulation of the substrate upon which the wild-type meiotic machinery operates.
In the current work, we interrogate the processes of homolog pairing and synapsis during C. elegans meiosis by altering karyotype to disrupt the normal 1∶1 correspondence between homologous chromosomes in the context of a fully intact meiotic machinery. This analysis provides insights into the principles governing homolog pairing and synapsis, as well as the quality control systems that promote successful meiosis. Moreover, we uncover evidence for the operation of a masking mechanism that can “hide” synapsis imperfections from the monitoring machinery, thereby counterbalancing stringent quality control. As a result, a significant fraction of meiocytes containing minor synapsis imperfections may be able to continue meiotic progression, escape apoptosis, and successfully complete recombination. We propose that this strategy maximizes meiotic success by making progression and survival contingent on formation of sufficient SC to ensure crossover formation without imposing an onerous and unnecessary requirement for perfect synapsis.
To gain insight into the mechanisms governing homolog pairing and synapsis, we manipulated the karyotype of C. elegans in the context of otherwise wild-type meiotic machinery. We employed hermaphrodite karyotypes containing an increased number of chromosomes, including odd and even numbers of homologs: triplo-X trisomics carrying a third copy of the X chromosome in the presence of a diploid complement of autosomes (3X:2A), full triploids (3X:3A), and full tetraploids (4X:4A; Figure 1A).
We assayed homolog pairing status in trisomic and polyploid C. elegans by immunofluorescence (IF) for the PC binding proteins associated with the X chromosome (HIM-8 [8]) and two autosomes (ZIM-3, which binds to the PCs of both chromosomes I and IV [7]). During diploid meiosis, homologous PCs are not associated at meiotic entry; they rapidly pair in the transition zone (TZ, representing the classical leptotene/zygotene stages of meiotic prophase) and remain associated through later stages. Thus, diploid nuclei achieve a single prominent focus of HIM-8 in each nucleus, identifying the paired X-PCs, and two prominent foci of ZIM-3, identifying the separately paired chromosome I- and IV-PCs. In triplo-X, triploid, and tetraploid meioses, we also detected a single prominent HIM-8 focus and two prominent ZIM-3 foci in most nuclei starting in the TZ (Figure 1B, S1, S2) consistent with the ability of groups of two, three, or four homologs to pair at their PCs.
To analyze the extent of interactions among groups of three and four homologs at regions beyond the PC, we employed an S-phase labeling approach to identify the nuclear territories occupied by the X chromosomes in a subset of temporally synchronous nuclei. Briefly, live worms were exposed to labeled nucleotide analogs, which incorporate into replicating chromosomes in the germ line. Due to late replication of the sex chromosomes, nucleotides incorporate exclusively into the X chromosome pair in a proximally-located population of S-phase nuclei that are in late meiotic S at the time of exposure [26]. These nuclei, which represent a highly synchronous cohort despite some spatial heterogeneity, can then be analyzed at a desired time point after labeling (Figure 2A).
We used 3-D volume rendering to examine the organization of labeled X chromosome territories in late TZ nuclei, after pairing at the PCs is complete and in the context of ongoing chromosome mobilization (10 h post S-phase labeling; Figure 2B,C; S3). Synapsis is in progress in these nuclei, but is not yet complete. For all karyotypes, the major class of nuclei at this time point displayed paired PCs with X chromosome territories showing an extended appearance. Among diploid nuclei in this class, all showed a unitary labeled domain in which the paths of individual X chromosomes could not be distinguished and non-PC ends appeared to be convergent. This organization suggests close association between the two chromosomes along their lengths, consistent with recent analysis using chromosome paints [27]. Moreover, for karyotypes with three or four copies of the X chromosome, nuclei in this major class typically exhibited characteristics similar to the diploid. Therefore, in addition to pairing at the PC, three and four X chromosomes frequently appear to associate closely along their lengths at this stage, raising the possibility that they may all compete for establishing synapsis interactions.
We performed IF to evaluate synapsis patterns in trisomic and polyploid C. elegans. Staining for the meiotic chromosome axis protein HTP-3 was used to identify all chromosomes, regardless of synapsis status, and staining for central region protein SYP-1, which bridges pairs of chromosome axes in the mature SC, was used to identify synapsed regions [17], [18]. Examination of mid-pachytene region nuclei, which exhibit full synapsis in the diploid, revealed synapsis aberrations in the altered karyotypes (Figure 3A). While some triplo-X nuclei showed apparently complete synapsis, most displayed a single SYP-1-free region, consistent with an unsynapsed third X in the majority of nuclei. Most triploid nuclei, which contain six odd chromosomes, showed zero to two SYP-1-free regions, indicating that the third copies of each homolog often participate in synapsis interactions. Many tetraploid nuclei showed apparently full synapsis, but some displayed regions lacking SYP-1, suggesting that while full synapsis can be achieved in this karyotype, the process is partially compromised when four copies of each homolog are present. Taken together, these observations indicate that supernumerary chromosomes, whether odd or even in number, present challenges to achieving full synapsis.
To determine the specific synapsis configurations achieved among groups of three and four homologs, we used S-phase labeling in combination with IF and 3-D rendering to trace SC tracks associated with X chromosome territories in synchronous populations of nuclei corresponding to the mid-pachytene stage in diploids (24 h post S-phase). As expected, all diploid nuclei showed a single, robust SYP-1 track associated with the X chromosome pair identified by incorporated label and HIM-8 staining (18/18 nuclei; Figure 3B, i). Synapsis patterns among three and four homologs were more variable, consistent with supernumerary chromosomes presenting challenges to achieving full synapsis; however, analysis of triplo-X and tetraploid nuclei supports the idea that SC tends to connect pairs of chromosome axes.
In the majority of tetraploid nuclei, the X chromosome territories were marked by two SC tracks of similar length emanating from a single HIM-8 focus (24/32 nuclei). Although in most cases the two SCs were separate (Figure 3B, ii), we observed a small class in which they appeared intertwined (Figure 3B, iii); the latter could represent coiling of two separate chromosome pairs around each other or possibly switching of synapsis partners along the lengths of the chromosomes. In addition, a minority of tetraploid nuclei showed a single SC track localizing to the PC end of a domain occupied by all four chromosomes (6/32 nuclei; Figure 3B, iv), suggesting that resolution into pairwise partnerships is partly compromised.
Our analysis of triplo-X synapsis configurations further supports a preference for pairwise synapsis, extending previous EM analysis of three C. elegans triplo-X pachytene nuclei [28]. In the majority of triplo-X nuclei at 24 h post S-phase labeling, a presumed pair of X chromosomes was connected along their lengths by a single SC track, while the third X appeared to be excluded to an adjacent domain (17/22 nuclei). In some cases the excluded third X was SYP-1-negative (10 nuclei; Figure 3B, v), while in others it was SYP-1-positive (7 nuclei; Figure 3B, vi). Finally, in a small number of nuclei all three X chromosomes were found in a unitary territory that displayed a single track of SYP-1, suggesting failure to exclude the third X chromosome in a few cases (5/22 nuclei; Figure S4). All of these categories were detected at comparable frequencies in separate experiments where combined HTP-3 and SYP-1 immunostaining was assessed in mid-pachytene triplo-X nuclei (Figure 4A). These experiments confirmed that when the third X chromosome was spatially excluded, its HTP-3-positive axis either lacked SYP-1 entirely, consistent with asynapsis (Figure 4A, i; Video S1), or showed SYP-1 over part or all of its length, consistent with auto-synapsis that ranged from partial to complete (Figure 4A, ii; Video S2). Taken together our data suggest that by 24 h post-S phase, most triplo-X nuclei have transitioned from initial pairing of all three X chromosomes (characteristic of 10 h post-S phase) to pairwise synapsis accompanied by exclusion of the third X, which is itself sometimes involved in (presumably pairwise) auto-synapsis.
We also examined X chromosome organization in mid-pachytene nuclei from spo-11 mutant triplo-X worms (3X:2A spo-11), which lack the ability to form the DNA double-strand breaks (DSBs) that serve as the initiating events of meiotic recombination. In contrast to otherwise wild-type triplo-X controls, a spatially excluded third X typically was not observed, and a single SC localized to a territory apparently occupied by all three X chromosomes (55/58 mid-pachytene nuclei; Figure 4B; Video S3). Overlap between HTP-3 and SYP-1 signals suggests that this configuration may represent synapsis among three partners (either three-way synapsis throughout, or strictly pairwise synapsis combined with partner switches). Thus, although SPO-11-dependent recombination intermediates are not required in diploids to achieve synapsis between homologs [29], these data suggest that recombinational interactions between chromosomes may nevertheless contribute to maturation of the SC structure into a strictly pairwise state.
Analysis of triploid nuclei by the S-phase labeling method also revealed a defect in establishment of pairwise interactions. While a small minority of triploid nuclei showed apparent exclusion of the third X by 24 h post-S phase (5/24 nuclei, Figure S4), in most cases all three X chromosomes occupied a joint domain to which a single SC localized (19/24 nuclei; Figure 3B, vii and viii). We detected occasional trivalent chromosomes at the diakinesis stage in triploids, indicating that at least a subset of these SCs connect three partners and promote crossovers between them (Figure S5). Together, these observations suggest that the presence of a third copy of every chromosome impairs the ability to solidify pairwise relationships. In addition, we noted a shift toward exclusion of the third X chromosome and emergence of pairwise synapsis in triploids at a later time point (Figure S4). These results raised the possibility that the widespread synapsis challenges in triploids might delay progression through the meiotic program.
We assessed the impact of supernumerary chromosomes on meiotic prophase progression using IF for a phosphorylated isoform of the NE protein SUN-1 (SUN-1 S8-Pi). In diploid germ lines, SUN-1 S8-Pi is detected in a zone of nuclei extending from TZ entry until the early/mid-pachytene transition [11]. It localizes both diffusely throughout the NE and in concentrated NE-associated patches corresponding to points of attachment between the chromosomes and the cytoskeletal motility apparatus that mediates movement. Moreover, conditions that prolong the duration of chromosome mobilization in diploids extend the SUN-1 S8-Pi zone [11], and this persistent phosphorylation is required to delay meiotic progression [21]. Diffuse SUN-1 S8-Pi is also prolonged in mutants that are proficient for synapsis but impaired in the ability to form crossover recombination intermediates between homologs, but in such cases the multiple bright patches indicative of chromosome movement do not persist [21], [22]. In tetraploid germ lines, the SUN-1 S8-Pi zone occupied a proportion of the meiotic region that was not significantly different from that in diploids (Figure 5A, B; two-tailed P = 0.166, Mann-Whitney Test), suggesting that presence of an even number of supernumerary chromosomes has little impact on the relative length of the chromosome mobilization phase.
By contrast, the two “odd” karyotypes displayed extended SUN-1 S8-Pi zones (Figure 5B, two-tailed P≤0.002, Mann-Whitney Test) with multiple bright patches found in most positively-staining nuclei. Triplo-X germ lines showed a modest extension, with loss of SUN-1 S8-Pi taking place in the region corresponding approximately to the mid-pachytene to late pachytene transition in diploids. Triploid germ lines displayed a more dramatic extension than triplo-X (two-tailed P<0.0001, Mann-Whitney Test), with SUN-1 S8-Pi levels declining only toward the end of the region corresponding to the late pachytene stage in diploids, when constraints on multiple meiotic processes are known to be lifted [19], [30]. These results suggest that odd numbers of supernumerary chromosomes prolong the chromosome mobilization phase and impede meiotic progression. Further, they raise the idea that the meiotic program can ultimately accommodate problems associated with a single partnerless chromosome.
The ability to extend the SUN-1 S8-Pi positive zone in triplo-X germ lines does not require the same machinery as the “synapsis checkpoint” that triggers apoptosis in response to synapsis defects, since the extent of the SUN-1 S8-Pi positive zone is not altered in triplo-X germ lines mutant for pch-2, which encodes a conserved AAA+ ATPase required for the synapsis checkpoint (Figure 5C; [25]). Further, the length of the SUN-1 S8-Pi positive zone in triplo-X germ lines was also not altered by mutation of spo-11, indicating that that inability to form crossovers between homologs and/or unresolved multi-partner interactions (Figure 4B) do not further prolong the SUN-1 S8-Pi positive state beyond the response elicited by a single extra X chromosome.
As defective synapsis and unrepaired recombination intermediates can trigger elevated apoptosis in the late pachytene region of the C. elegans germ line through the action of meiotic checkpoints [24], [25], we assessed the impact of supernumerary chromosomes on germ cell apoptosis levels (Figure S6). The outcome of this analysis varied depending on the conditions under which apoptosis was assayed, with significantly elevated numbers of germ cell corpses detected in both triplo-X and triploid (but not tetraploid) gonads under one set of conditions, but not under another. Since the gonads of worms with altered karyotypes contain fewer meiotic nuclei overall (see Figure 5A; Table S1), we also attempted to normalize numbers of germ cell corpses relative to numbers of meiotic zone nuclei (see Figure S6 legend); when normalized values were used, all three modified karyotypes showed a significant elevation of apoptosis compared to diploids, suggesting that meiotic checkpoints may be triggered above baseline levels when supernumerary chromosomes are present. Triploids showed a larger elevation than triplo-X, suggesting a greater degree of checkpoint activation when all chromosomes are present in three copies and providing a parallel to the greater impact on meiotic progression discussed above. However, normalized apoptosis levels in tetraploids were not significantly different from those observed in tripoids despite tetraploids showing relatively normal meiotic progression as assayed by SUN-1 S8-Pi, suggesting that these two features of the meiotic program may not be inherently coupled.
In C. elegans, unsynapsed chromosomes including the single X in XO males are marked by the histone H3 dimethyl Lys9 (H3K9me2) chromatin modification during the latter half of the pachytene stage [31], [32]. During normal diploid hermaphrodite meiosis, levels of the H3K9me2 mark begin to rise at the early/mid-pachytene transition [33], and we found that this timing coincided approximately with loss of SUN-1 S8-Pi (Figure 6A). H3K9me2 signals commonly appeared faint and scattered (Figure 6B, i) until levels increased dramatically throughout the chromatin at the end of the late pachytene stage. The timing of onset of H3K9me2 staining and the very late prophase rise in H3K9me2 levels were shared by all karyotypes. However, we observed distinct nuclear patterns of H3K9me2 as well as informative relationships to the SUN-1 S8-Pi zone during trisomic and polyploid meiosis.
In triplo-X and tetraploid germ lines, acquisition of specific nuclear H3K9me2 patterns correlated with loss of SUN-1 S8-Pi staining (Figure 6A). In the triplo-X, H3K9me2 localized to a single intense domain in every nucleus by the mid-pachytene region (Figure 6A, 7B), consistent with marking of the synapsis-challenged third X chromosome in every nucleus. H3K9me2 domains corresponded to chromatin devoid of SYP-1 in 70% of nuclei (21/30; Figure 6B, ii), and to SYP-1-associated chromatin in 30% of nuclei (9/30; Figure 6B, iii), suggesting that the excluded X chromosome is marked, at least in part, whether or not it loads some SYP-1. In the tetraploid, where synapsis defects are more variable, several H3K9me2 patterns were observed (Figure 6A, iv; 6B, v): some nuclei showed faint signal comparable to the diploid; many showed small intense terminal domains colocalizing with SYP-1; and a few showed large H3K9me2 domains generally corresponding to unsynapsed regions. Together, our data suggest that in trisomic or tetraploid germ cells, chromosomes or chromosomal segments that experienced synapsis challenges are efficiently marked with H3K9me2, whether or not they ultimately acquire SC. Further, co-staining suggests a temporal correlation between appearance of the H3K9me2 mark on synapsis-challenged regions and loss of SUN-1 S8-Pi in these karyotypes.
In the triploid, H3K9me2 domains appeared with similar timing to the other karyotypes, but their appearance was not accompanied by down-regulation of SUN-1 S8-Pi (Figure 6A). Moreover, although each of the six homolog groups is challenged for achieving pairwise homologous synapsis in this karyotype, we typically observed just one or two intense H3K9me2 domains per nucleus (Figure 6A,B; 7B). Further, co-staining with SC markers revealed that while H3K9me2 localized to most regions lacking detectable SYP-1 in triploids (23/29), the H3K9me2 domains often failed to completely overlap such regions (16/29; Table S2). In addition, our analysis revealed another class of aberrant synapsis: axis segments associated with very low levels of SYP-1, which likely represent a subset of heterologously synapsed regions (Figure S7). These SYP-1-weak regions typically showed little or no H3K9me2 staining (28/37; Figure 6C, Table S2). These observations suggest that some of the many chromosomal regions experiencing synapsis problems in triploid hermaphrodites fail to acquire robust H3K9me2, and raise the possibility that incomplete marking of synapsis problems may be related to the dramatic extension of the chromosome mobilization phase in this karyotype.
We examined H3K9me2 and SUN-1 S8-Pi patterns in males with normal and altered karyotypes. Acquisition of H3K9me2 by the unsynapsed X during diploid male (1X:2A) spermatogenesis correlated temporally with loss of SUN-1 S8-Pi (Figure 7A), similar to triplo-X oogenesis. Triploid males (2X:3A) differed from triploid hermaphrodites, however, in that pachytene nuclei typically contained two to three major domains of H3K9me2 staining (Figure 7A, ii; 6B) – approximately one more than hermaphrodites despite the fact that they possess one less supernumerary chromosome. Further, while triploids of both sexes exhibited a similar range and degree of synapsis defects, males showed robust H3K9me2 marking of unsynapsed chromosome segments as well as a subset of (presumably heterologously) synapsed chromosome segments (Figure 7C). In addition, while the SUN-1 S8-Pi zone was extended relative to diploid males (Figure 7A, S8), we note that loss of SUN-1 S8-Pi took place well before the end of the pachytene region. The timing of meiotic progression differs between spermatogenesis and oogenesis, precluding a direct comparison of the two [26], [34]; however, this observation raises the possibility that triploid male germ cells resume meiotic progression more efficiently. Taken together, these data support the idea that synapsis-challenged chromosomes are more efficiently marked by the spermatogenic program than by the oogenic program.
Given the strong association between pachytene stage H3K9me2 domains and synapsis defects in meiotic mutants [32], [35] and in worms with altered karyotypes (this work), we closely examined H3K9me2 staining and synapsis status in diploid female germ cells, which display areas of H3K9me2 enrichment that are poorly understood. Rotation of 3-D images from the beginning of the late pachytene region, where specific signal is brightest, revealed concentration of H3K9me2 signal into small chromosome-associated domains in 23% of nuclei (32/142 nuclei from 5 germ lines) while the remaining nuclei showed dispersed signal (Figure 6B, i). 3-D rendering of individual diploid nuclei stained for H3K9me2 in combination with SC markers showed that these small H3K9me2 domains were typically located at the end of a SYP-1 stretch (31/35 domains from 5 germ lines) and oriented toward the nuclear interior (20/35). When both SYP-1 and HTP-3 were assessed, we detected small segments of HTP-3-positive chromosome axes lacking corresponding SYP-1 staining in 40% of H3K9me2 domain-containing nuclei (10/25 of such nuclei from 5 germ lines; Figure 8A), and all of these SYP-1-deficient segments colocalized with H3K9me2 domains. These observations indicate that small unsynapsed regions marked by H3K9me2 are present in a substantial percentage of normal diploid germ cells before large-scale desynapsis is observed.
In principle, small unsynapsed regions in approximately 9% of nuclei from the beginning of the late pachytene region could represent either unresolved synapsis defects or the onset of desynapsis. We reasoned that unresolved synapsis defects should be detected at the same or higher frequency earlier in meiotic prophase, whereas the converse should be true of desynapsing regions. Consistent with the former interpretation, examination of 3-D rendered images revealed small, internal, terminal regions of asynapsis in 16% of nuclei at the mid-pachytene stage (7/44 nuclei from 1 germ line). Although H3K9me2 staining was weaker at this stage, we detected signal at most unsynapsed regions (5/7), consistent with the idea that synapsis defects are marked by H3K9me2 around the time of exit from the chromosome mobilization phase in normal diploid meiocytes. We conclude that unresolved synapsis defects visible at the resolution of fluorescence microscopy are common during normal diploid meiosis, and are efficiently marked by H3K9me2.
The MET-2 methyltransferase is responsible for the H3K9me2 chromatin modification in C. elegans germ cells [33]. MET-2 is required for transcriptional silencing of the partnerless X and for preventing activation of the recombination checkpoint in XO germ lines, implicating H3K9me2 in these processes [36], [37]; however, the modification did not appear to mediate silencing or checkpoint shielding for a pair of unsynapsed homologs. Thus, we tested whether MET-2 function and/or H3K9me2 acquisition by imperfectly synapsed regions could play a role in promoting meiotic progression and/or protecting meiocytes from apoptosis.
We found that a met-2 mutation or depletion of MET-2 by RNAi caused loss of H3K9me2 without an accompanying extension of the SUN-1 S8-Pi zone in either diploid or triplo-X germ lines (Figure 6D and Materials and Methods). Thus, although H3K9me2 staining reveals that imperfectly synapsed regions acquire an altered character around the time of SUN-1 S8-Pi zone exit in these karyotypes, the modification itself is dispensable for exit from the mobilization phase and progression through meiosis.
However, loss of met-2 function in diploid hermaphrodites did result in a significant elevation of germ cell apoptosis (Figure 8B; two-tailed P<0.0001, Mann-Whitney Test). This effect did not require PCH-2, suggesting that elevated apoptosis caused by loss of MET-2 does not reflect activation of the synapsis checkpoint. Instead, it was suppressed by loss of spo-11 function, implying that elevated apoptosis in the met-2 mutant reflects activation of the recombination checkpoint. These findings parallel the recent report that MET-2 prevents the partnerless X chromosome in XO germ cells from activating the recombination/DNA damage checkpoint [37]. Further, they are consistent with the possibility that MET-2 and H3K9me2 may play a role in protecting normal diploid meiocytes from undergoing apoptosis by inhibiting recombination intermediates associated with unsynapsed chromosome segments from activating the recombination checkpoint. However, we cannot rule out alternative interpretations, e.g., that an altered distribution of DSBs or crossovers in a met-2 mutant might somehow increase the likelihood of activating the recombination checkpoint.
The extent to which MET-2 might influence apoptosis levels in the presence of supernumerary chromosomes is less clear. Although met-2 RNAi did result in elevated apoptosis in diploid, triplo-X, triploid and tetraploid worms compared to empty vector controls (Figure S9), these results must be interpreted with caution since a recent report indicated that RNAi treatment per se can increase germ cell apoptosis [38]. Further, triplo-X met-2(n4256) mutant germ lines exhibited a very modest, albeit significant, elevation of apoptosis when compared with otherwise wild-type triplo-X controls. Thus, while MET-2 may help to limit checkpoint activation in the presence of an extra X chromosome, it may not be a major determinant of baseline apoptosis levels in this context. However, apoptosis levels were substantially elevated in triplo-X spo-11 mutant germ lines (P<0.0001), suggesting that the atypical X-chromosome synapsis configurations observed in these worms are recognized as aberrant by the meiotic quality control machinery.
In this study, we challenged the meiotic program by manipulating karyotype in C. elegans. We began by defining pairing and synapsis phenotypes during trisomic and polyploid meiosis, taking advantage of high-resolution 3-D fluorescence imaging in the context of preserved nuclear architecture combined with an S-phase labeling method that affords both precise staging of meiocytes and delineation of the territory occupied by a specific chromosome group. While we found that synapsis was challenged in all altered karyotypes, overall we showed that tetraploids are able to achieve relatively complete, homologous synapsis, and triploids are more severely affected by synapsis failure than triplo-X worms.
This work complements classical microscopy studies of karyotype manipulation in a variety of organisms (for reviews, see [39], [40]), and highlights a set of general principles governing homolog pairing and synapsis. First, initial pairing and pre-synaptic alignment can take place among all homologous chromosomes present, even if greater than two; therefore, homolog recognition is not strictly pairwise in nature. Second, in all karyotypes, mature synapsis interactions tend to connect pairs of chromosome axes, indicating a preference for pairwise synapsis; however, supernumerary chromosomes appear to require more time to sort out pairwise interactions, and multi-partner synapsis and/or partner switches are sometimes observed. Third, synapsis failure is less prevalent than one might expect in karyotypes containing “odd” chromosomes that lack an exclusive, homologous pairing partner. Such chromosomes frequently engage in fold-back or non-homologous synapsis, indicating a drive to maximize synapsis interactions irrespective of homology [39], [41]. Fourth, our observations in triploids suggest that synapsis interactions may be remodeled over time. In combination with a number of previously documented phenomena such as SC “correction” in plants (reviewed in [42]), “synaptic adjustment” in a variety of organisms [39], [43], and incorporation of SC components throughout the pachytene stage in budding yeast [44], our work lends further credence to the view that the SC is a dynamic structure.
Unexpectedly, our work revealed a previously hidden influence of recombination on synapsis in C. elegans. Whereas recombination-based interactions are essential for normal pairing and/or synapsis in many organisms, it is well established that recombination is dispensable for homologous synapsis in C. elegans and Drosophila [29], [45], [46]. However, we found that in the context of triplo-X meiosis, spo-11 status affects the ability to solidify the relationship between one pair of homologs while excluding a third, suggesting that recombination may contribute to synapsis fidelity even in an organism where recombination-independent mechanisms predominate. Interestingly, the apparent relationship between recombination and “exclusivity” in the formation of bivalents during C. elegans meiosis is reciprocal to that observed in Bombyx mori, where recombination occurs in males but not in females: recombination in tetraploid Bombyx males appears to lock in multi-partner relationships that are resolved into bivalent relationships in tetraploid females [47], [48], whereas recombination appears to promote two-partner exclusivity in triplo-X worms.
Several quality control mechanisms have been identified that appear to contribute to meiotic success by ensuring that only meiocytes with properly synapsed chromosomes proceed to subsequent stages of meiosis. Analysis of the impact of karyotype alterations on meiotic progression and germ cell apoptosis provides insight into how these mechanisms operate. Our findings support and extend an existing model for a checkpoint-like mechanism that couples exit from the period of active chromosome movement with successful SC installation [19]. This model was initially proposed to explain the persistence of markers of chromosome mobilization into the late pachytene region (“extended TZ”) in mutants experiencing global asynapsis due to absence of SC precursors [17], and is supported by reports of similar responses to an unpaired and unsynapsed pair of chromosomes due to absence of specific PC proteins or deletion of the X-PCs [7], [8], [49]. Together, these observations imply that some aspect of unsynapsed chromosomes produces a signal that blocks exit from the chromosome mobilization phase.
Here, SUN-1 S8-Pi staining in trisomic and polyploid C. elegans indicates that odd numbers of chromosomes, which present synapsis challenges, also delay exit from the chromosome mobilization phase. While such a delay can be inferred from prior studies involving two unpaired and unsynapsed chromosomes [7], 8,49, prolonged mobilization in triplo-X germ lines clearly establishes that a single synapsis-challenged chromosome is sufficient to trigger this response. Further, the length of the delay appears to depend upon the number of chromosomes experiencing synapsis problems: triploids encountered a profound delay, whereas triplo-X germ cells resumed meiotic progression more quickly, and tetraploids were relatively unaffected. Together, these data suggest that the checkpoint-like coupling mechanism may delay exit from the mobilization phase in a manner that is sensitive to the dose of unsynapsed chromosomes.
Biological checkpoints typically function to prolong conditions necessary to resolve the lesions that trigger them. In this view, extension of the SUN-1 S8-Pi zone is a logical response to synapsis challenges because efficient SC assembly requires chromosome mobilization [9]. Although synapsis problems were enhanced in the context of karyotype alterations, we also found small unsynapsed regions in a significant fraction of wild-type pachytene meiocytes, suggesting that minor SC defects are common in diploids. In light of this finding, we propose that retention of chromosome mobilization during the early pachytene stage during normal meiosis may provide conditions necessary for resolution of synapsis problems that remain at a significant frequency after initial SC assembly.
A second class of mechanisms contributing to synapsis quality control consists of checkpoints that trigger apoptosis of late pachytene germ cells in response to asynapsis per se, or to recombination intermediates persisting at regions that are not synapsed with a homologous partner [24], [25]. Our data suggest that such checkpoints are not only relevant when meiosis is challenged by abnormal karyotype or mutation of meiotic machinery components, but also in the context of normal, diploid meiosis (see below).
During wild-type meiosis, stringent synapsis quality control can be achieved by sequential operation of these two types of mechanisms—prolonging conditions that promote resolution of synapsis problems, and then eliminating problems that fail to be corrected. We note that catastrophic synapsis problems (e.g., caused by absence of SC components [17] or triploidy (this work)) uncover limits to the capacity of synapsis quality control mechanisms: the chromosome mobilization phase can only be extended until the late pachytene region, and checkpoint-mediated apoptosis cannot cull all meiocytes. Under such circumstances, constraints on multiple meiotic processes are nevertheless ultimately lifted at the end of the pachytene region [19], [30]; this late-prophase lifting of constraints is proposed to serve as a fail-safe mechanism that safeguards the genome by facilitating DNA repair and restoration of genome integrity prior to chromosome segregation.
The purpose of pairing and synapsing homologous chromosomes during meiotic prophase is to enable the formation of crossover recombination events between homologs that underlie their ability to segregate at the meiosis I division. Given this raison d'etre, an effective scheme for promoting reproductive success would maximize assurance of crossover formation, rather than synapsis, as its outcome. Our analysis of meiosis in C. elegans with altered ploidy suggests that such a scheme does indeed operate. Specifically, our observations support a model in which the quality control mechanisms that respond to synapsis defects are counterbalanced by masking mechanisms that can “hide” limited defects (either asynapsed regions per se, or persistent recombination intermediates associated with such regions), allowing resumption of meiotic progression and evasion of checkpoint-mediated apoptosis.
Our data suggest that acquisition of the H3K9me2 chromatin modification by synapsis-challenged chromosomal regions reflects one aspect of masking, with the H3K9me2 mark identifying regions whose character has been altered by the meiotic program. We found that the unsynapsed or self-synapsed third X chromosome became marked with H3K9me2 in every triplo-X meiocyte, and this correlated temporally with loss of SUN-1 S8-Pi staining; a similar correlation was seen in tetraploids where synapsis-challenged regions also appeared to be efficiently marked. Observations in these karyotypes are therefore consistent with masking of synapsis problems relieving the block to meiotic progression.
Further, our data raise the possibility that this masking mechanism is saturable. In triploid hermaphrodite germ lines, where every homolog group experienced synapsis challenges, SUN-1 S8-Pi staining persisted until the end of the pachytene region. Our analysis indicated that some synapsis defects in triploid hermaphrodites were not marked with H3K9me2 (despite the fact that similar defects in triploid males acquired the mark robustly), suggesting that some defects remain unmasked in this karyotype. We propose that the amount of asynapsis in triploid hermaphrodites exceeds the system's capacity to “hide” problems, allowing signals that delay meiotic progression to persist. We further speculate that maximization of synapsis over time, irrespective of homology, may represent a separate means to hide chromosome regions without suitable pairing partners, thereby relieving the block to meiotic progression that unpartnered chromosome axes would otherwise present.
Taken together, our findings in trisomic and polyploid C. elegans support the existence of mechanisms that can mask synapsis problems up to a certain threshold. Importantly, we found that small synapsis defects in diploid pachytene germ cells were efficiently marked by H3K9me2, supporting the idea that these masking mechanisms also operate during the course of normal, wild-type meiosis. Although H3K9me2 serves as a useful marker for identifying masked regions with altered character, however, the modification itself participates in only a subset of the masking effects. Specifically, MET-2 and H3K9me2 do not appear to be required for the masking effect that enables meiotic progression. However, elevated apoptosis associated with loss of MET-2 function in diploids (this work) and XO germ cells [37] suggests that H3K9me2 is likely relevant for inhibiting checkpoint activation during normal diploid meiosis, presumably by preventing persistent recombination intermediates from triggering checkpoints. Together, our data suggest that the ability to limit a progression delay and the ability to prevent triggering of checkpoint-mediated apoptosis may operate at least partially through distinct mechanisms.
Our work raises the idea that “good enough”—rather than perfect—synapsis may be a preferable waypoint in the meiotic program. Partial homologous synapsis of all chromosome pairs should be sufficient to promote a single crossover on each, which is all that is required—and is in fact the normal state—in C. elegans meiosis. Examples of incomplete synapsis during normal meiosis have also been noted in other organisms (e.g., [50], [51]). Therefore, we propose that the ability to mask minor synapsis defects can provide an advantage by counterbalancing quality controls to ensure timely progression and survival of germ cells, and ultimately successful meiosis. It is well recognized that the meiotic program must employ mechanisms to accommodate asynapsis and lack of recombination for sex chromosome regions that lack a homologous partner [52]. We speculate that the saturable masking mechanism proposed here originated to handle the partnerless sex chromosome in males (1X:2A), and subsequently provided a benefit to reproductive fitness in both sexes.
Meiosis represents a multi-step biological program whose end goal is to segregate homologous chromosomes away from one another. The value of stringent quality control mechanisms monitoring intermediate steps required to achieve this goal (e.g., homolog pairing, synapsis, and crossover formation) is currently appreciated. Our work points out that such quality controls can be counterproductive, however, when perfection of an intermediate step is not an absolute requirement for achieving the desired end: “good enough” intermediates may be subjected to unnecessary delays or wasteful elimination. Therefore, by introducing the ability to hide minor defects unlikely to impact ultimate success, saturable masking mechanisms can provide a counterbalance to quality controls and improve the overall efficiency of the program. We suggest that the type of regulatory logic discussed here may be a widespread feature of biological circuits.
All C. elegans strains were cultivated at 20°C under standard conditions [53]. A mating stock of Bristol N2 provided the wild-type diploid background. Triplo-X worms are recognized by their semi-Dpy appearance [54] and were maintained by picking semi-Dpy L4s from the wild-type triplo-X “strain” AV494 at each generation. AV494 was generated by picking a spontaneous 3X:2A individual from CA257 (him-8(tm611) IV) which has a high degree of X chromosome nondisjunction, and backcrossing to N2 males to restore homozygosity for the wild-type him-8 allele (confirmed by PCR) while maintaining a 3X:2A karyotype. The wild-type tetraploid strain SP346 [55] was maintained by picking Lon L4s every 1–2 generations. Most cytological analysis of tetraploids was conducted using worms that had been maintained on plates for less than 10 generations; accurate karyotype was confirmed by cytological phenotype at diakinesis (12 bivalents). Triploids (3X:3A hermaphrodites, and 2X:3A males) were generated as needed by mating N2 males (1X:2A) to Lon SP346 hermaphrodites (4X:4A) and were invariably the only large, healthy progeny produced in this cross; cytological phenotype at diakinesis was confirmed for 3X:3A hermaphrodites (6 bivalents+6 univalents).
Triplo-X “strains” AV785 and AV784, harboring the met-2(n4256) or pch-2(tm1458) mutations, respectively, were generated by crossing met-2(n4256)/hT2[qIs48] (I;III) or pch-2(tm1458) males with 3X:2A hermaphrodites from AV494. 3X:2A worms were plated individually at the F1 and F2 generations, and F2 plates homozygous for the met-2 or pch-2 mutation were identified by PCR.
Triplo-X “strain” AV783, heterozygous for spo-11(me44) and balancer chromosome nT1[qIs51] (IV;V), was generated by crossing spo-11(me44)/nT1[qIs51] males with 3X:2A hermaphrodites from AV494. 3X:2A +/nT1[qIs51] cross progeny hermaphrodites (identified using a GFP marker associated with the balancer) were crossed with spo-11(me44)/nT1[qIs51] (IV;V) males. GFP+ 3X:2A hermaphrodites were plated individually in the next generation, and 3X:2A spo-11(me44)/nT1[qIs51] worms were identified by DAPI staining of their GFP- progeny: 3X:2A spo-11(me44) homozygotes were unambiguously identified based on the presence of 13 univalents in diakinesis-stage oocytes.
For most experiments, S-phase labeling was performed by microinjection of fluorescent nucleotides as described [3], with the following modifications: young adult worms (24 hr post L4) were microinjected in both gonad arms with 1 mM Alexa Fluor 647-OBEA-dCTP (Invitrogen) or 0.1 mM Cy3-dCTP (GE Health Sciences); worms were recovered and maintained on food at 20°C until they were dissected and fixed for IF at the indicated time points post-injection. For the images shown in Figure 2A, S-phase labeling was performed by feeding EdU-labeled bacteria prepared as in [56]. Young adult worms (24 hr post L4) were placed on plates seeded with EdU-labeled bacteria and maintained at 20°C until dissection and fixation at the indicated time point; EdU was detected using the Click-iT EdU Alexa-555 Imaging kit (Invitrogen) as in [57]. Meiotic progression of labeled nuclei was indistinguishable between the two S-phase labeling techniques. To ensure consideration of a defined, synchronous population, a field of view containing the proximal front of labeled nuclei was imaged (see [3]).
All analysis was performed on approximately 24-h post-L4 adults, with the exception of S-phase-labeling experiments as described above. Dissection of gonads, fixation, immunostaining, and DAPI staining were performed essentially as in [19]. The following primary antibodies were used: guinea pig anti-HIM-8 (1∶500 [8]); rabbit anti-ZIM-3 (1∶2,000 [7]); chicken anti-HTP-3 (1∶250 [18]); rabbit anti-SYP-1 (1∶250 [17]); guinea pig anti-SUN-1 S8-Pi (1∶800 [11]); and mouse monoclonal anti-H3K9me2 (1∶400: abcam code: ab1220; lot: 629623). Secondary antibodies were Alexa Fluor 488, 555, and/or 647-conjugated goat antibodies directed against the appropriate species (1∶400; Invitrogen).
3-D images were collected as Z-stacks (0.1 or 0.2 µm step size) using a 60× NA 1.42 objective with 1.5× optivar or a 100× NA 1.40 objective on a DeltaVision widefield deconvolution microscopy system (Applied Precision) and deconvolved (except for Figure 3A) using softWoRx software. Additional 3-D images for the experiment presented in Figure 4 were collected on an OMX microscopy system (Applied Precision) in widefield mode using a 100× NA 1.4 objective, and deconvolved and corrected for registration. Contrast adjustments, 3-D cropping and image rendering (2-D maximum intensity projections or 3-D surface opacity renderings, including videos) were performed using the Volocity 5 software package (PerkinElmer). Final assembly of high-resolution, tiled germ lines, as well as minor contrast adjustments, were performed using Adobe Photoshop. All scoring of the attributes of S-phase labeled nuclei (Figure 2B,C; 3B; S3; S4) and triplo-X synapsis configurations (Figure 4), as well as evaluation of H3K9me2 staining with respect to synapsis status (Figure 6B; 7C; 8A; Table S2), was performed within Volocity using 3-D rendered images of individually cropped nuclei. For these experiments, nuclei were scored only when staining and resolution were of sufficient quality to unambiguously trace chromosome paths or SCs in 3-D rotations; approximately 3/4 of nuclei in the gonads scored met these criteria. For quantitation of synapsis configurations, we scored all nuclei showing label incorporation on the X chromosomes (or that were located within the specified zone) that met the inclusion criteria and were fully contained within the Z stacks. Length of the SUN-1 S8-Pi zone (Figure 5B, 5C, 6D, S8), SYTO 12 staining (Figure 8B, S6, S9), and numbers of meiotic zone nuclei (Table S1) were scored by eye on an AxioVision system (Zeiss).
Germ cell corpses were scored by SYTO 12 assay essentially as in [58] with the following modifications: 24-hour post-L4 adults were stained by picking into a 35 mM dilution of SYTO 12 (Molecular Probes) in M9 and incubating in the dark for 90 min at 20°C. Worms were destained by transferring to fresh seeded plates and incubating in the dark for 45 min at 20°C. Scoring was performed within a 45 min window on worms mounted in anesthetic mix in multiwell slides as in [59].
RNAi was performed as in [3] with the following modifications: parents of the appropriate karyotype were placed on plates seeded with bacteria expressing dsRNA (empty vector L4440 or R05D3.11/met-2) from the Ahringer lab feeding RNAi library [60] at the L4 stage at 15°C and were allowed to produce progeny. For assessment of the SUN-1 S8-Pi zone, progeny were transferred to fresh RNAi plates at the L4 stage and incubated at 20°C for 24 hours prior to fixation for IF. The diploid strain AZ212 (unc-119(ed3) ruIs32 III [pie-1::GFP::H2B = unc-119(+)]), which exhibits increased sensitivity to RNAi [3], was used in the met-2 RNAi/SUN-1 S8-Pi zone experiment. Length of the SUN-1 S8-Pi zone did not differ between the control and met-2 RNAi for diploid AZ212 worms (58%+/−1.8% (n = 10 germ lines) vs. 55%+/−1.7% (n = 10) the meiotic zone, scored as in Figure 5) or triplo-X worms (72%+/−1.5% (n = 13) vs. 73%+/−1.5% (n = 16)). Efficient knockdown of met-2 was assessed by IF for H3K9me2 in all RNAi experiments.
Statistical tests were performed using GraphPad InStat or Prism software.
|
10.1371/journal.ppat.1000751 | Structural and Biochemical Characterization of SrcA, a Multi-Cargo Type III Secretion Chaperone in Salmonella Required for Pathogenic Association with a Host | Many Gram-negative bacteria colonize and exploit host niches using a protein apparatus called a type III secretion system (T3SS) that translocates bacterial effector proteins into host cells where their functions are essential for pathogenesis. A suite of T3SS-associated chaperone proteins bind cargo in the bacterial cytosol, establishing protein interaction networks needed for effector translocation into host cells. In Salmonella enterica serovar Typhimurium, a T3SS encoded in a large genomic island (SPI-2) is required for intracellular infection, but the chaperone complement required for effector translocation by this system is not known. Using a reverse genetics approach, we identified a multi-cargo secretion chaperone that is functionally integrated with the SPI-2-encoded T3SS and required for systemic infection in mice. Crystallographic analysis of SrcA at a resolution of 2.5 Å revealed a dimer similar to the CesT chaperone from enteropathogenic E. coli but lacking a 17-amino acid extension at the carboxyl terminus. Further biochemical and quantitative proteomics data revealed three protein interactions with SrcA, including two effector cargos (SseL and PipB2) and the type III-associated ATPase, SsaN, that increases the efficiency of effector translocation. Using competitive infections in mice we show that SrcA increases bacterial fitness during host infection, highlighting the in vivo importance of effector chaperones for the SPI-2 T3SS.
| Systemic typhoid fever caused by Salmonella enterica serovar Typhi leads to high mortality in the developing world and can be linked with chronic, persistent infections in survivors. To cause disease, Salmonella uses a specialized secretion device called a type III secretion system to disarm cells of the immune system and replicate within them. The assembly and function of this secretion system requires a set of chaperone proteins to direct the process, but the chaperone proteins themselves have remained elusive. Here, we found a new chaperone protein, called SrcA, which is required for proper function of the type III secretion system. Using a bacterial mutant lacking the srcA gene, we found that this chaperone was needed for Salmonella to compete against wild type cells during systemic disease because it controls secretion of at least 2 key proteins involved in immune escape and cell-to-cell transmission. This chaperone is present in all types of virulent Salmonella, but not in Salmonella that don't cause human infections, providing new insights into the pathogenic nature of this organism.
| Many Gram-negative bacteria that colonize host animals use a type III secretion system (T3SS) to deliver effector proteins directly into host cells where their interaction with host proteins and membranes contribute to pathogenesis. Comprised of over 20 proteins, T3SS are complex structures with relation to the flagellar T3SS [1],[2] and include several central features; (i) inner and outer membrane ring structures, (ii) an extracellular needle structure with pore-forming proteins at the distal tip that engage a host cell membrane, (iii) an ATPase at the base of the apparatus with energetic and chaperone-effector recruitment roles, and (iv) a suite of chaperones to coordinate the assembly and function of the apparatus during infection.
Secretion chaperones are proteins required for T3SS function with roles in apparatus assembly and effector delivery, but are not themselves subject to secretion [3]. These chaperones often have common physical features such as low molecular weight (<15 kDa), an acidic isoelectric point and a predicted amphipathic helix at the carboxyl terminus. Current literature groups secretion chaperones into three classes based on their physical interactions with cargo [3],[4]. Class I chaperones bind to translocated effector proteins at a chaperone binding domain (CBD) located in the amino terminus of the effector. Class I chaperones have a structural fold of five β-strands and three α-helices, forming homodimers that bind to the CBD in a horseshoe-like structure. These chaperones have been further sub-classified based on their substrate repertoire and location with respect to the genes encoding the T3SS [3]. Class II chaperones bind to translocon proteins that make up the secretion pore in the host target membrane and class III chaperones bind the extracellular filament proteins (or flagellin rod in the orthologous flagellar system) that polymerize into a helical structure following secretion from the bacterial cell. Secondary structure predictions suggest class III chaperones adopt an extended alpha helical structure, which was confirmed by the crystal structure of the CesA chaperone in enteropathogenic E. coli that binds the EspA filament protein [5].
Much of the virulence potential of Salmonella enterica, a group of more than 2300 serotypes, is attributed to horizontally acquired genomic islands termed Salmonella Pathogenicity Islands (SPI). SPI-1 encodes a T3SS required for host cell invasion and SPI-2 encodes a second T3SS needed for intracellular survival and immune evasion [6],[7]. To date, 13 effectors have been identified as substrates of the SPI-1 T3SS and 21 effectors for the SPI-2 T3SS, although the chaperones orchestrating the latter system have been elusive. Whereas 80% of the effectors of the SPI-1 system have defined chaperones, only two effector-chaperone interactions are known for the SPI-2 system. These include the effector-chaperone pair of SseF-SscB, and the chaperone SseA that binds translocon components SseD and SseB [8],[9],[10]. Crystal structures have been determined for three chaperones that coordinate translocation of effectors through the SPI-1 T3SS (InvB [11], SicP [12] and SigE [13]). However no structures are available for the SPI-2 T3SS chaperones whose effector repertoire seems considerably larger than that of the SPI-1 system.
In addition to maintaining a region of localized effector unfolding [12], T3SS chaperones have an emerging role as escorts that deliver their cargo to the cytoplasmic face of the inner membrane through physical interactions with an ATPase. These ATPases form a hexameric structure at the base of the T3SS [14] and are a conserved feature of both flagellar and non-flagellar type III systems to enhance secretion activity by promoting chaperone release and effector unfolding prior to secretion [15],[16],[17]. A chaperone-ATPase interaction for the SPI-2 T3SS has not been described previously and so whether this system conforms to the emerging escort paradigm is not known.
The regulation of the SPI-2 T3SS and its associated effector genes is coordinated by environmental cues signifying the intracellular environment [18]. These cues activate a two-component signaling system encoded in the SPI-2 island comprising the SsrA sensor kinase and SsrB response regulator. In addition to activating all of the T3SS structural operons, transcriptional profiling has uncovered new genes in the SsrB regulon that are required for bacterial pathogenesis including a translocated effector, SseL [19],[20], and a gene of unknown function called srfN that is common to the Salmonella genus [21]. Using a reverse genetics approach we identified an SsrB-regulated gene (STM2138) that we named srcA (SsrB-regulated chaperone A), whose gene product satisfied several a priori predictions relating to the physical properties associated with T3SS chaperones. We solved the crystal structure of SrcA and performed additional biochemical, proteomic and in vivo experiments that revealed SrcA to be a class I chaperone required for bacterial fitness in the host environment. Despite being genetically disconnected from SPI-2, SrcA is integrated functionally with this system by binding to the T3SS ATPase, SsaN, and providing chaperone activity towards two important effectors, SseL (STM2287) and PipB2 (STM2780), necessary for immune escape and cell-to-cell transmission. These data reveal structural and biochemical insight into a T3SS secretion chaperone required for intracellular pathogenesis of Salmonella.
Transcriptional profiling of SsrB-regulated genes in S. enterica serovar Typhimurium (S. Typhimurium) [22] identified a hypothetical gene, STM2138 (named srcA hereafter), that was co-regulated with genes in SPI-2 and repressed ∼20-fold in an ssrB mutant compared to wild type. This gene was also down regulated in Salmonella mutants lacking the SsrA sensor kinase [20], and was predicted to encode a possible chaperone in a bioinformatics-based screen [23]. The srcA gene is not located in the vicinity of the T3SS encoded by SPI-2 (STM1378-STM1425), but is 713 genes downstream on the chromosome (STM numbers are based on the LT2 genome and ordered sequentially on the chromosome beginning at STM0001, thrL). The predicted srcA gene product was a small protein ∼16 kDa with a pI of 4.6, similar to secretion chaperones associated with T3SS. To verify SsrB input on srcA expression, we analyzed SsrB binding in vivo at the region of DNA surrounding srcA using genome-wide ChIP-on-chip [21] (and unpublished data). This analysis revealed a strong SsrB binding site spanning 10 syntenic probes within the intergenic region (IGR) upstream of srcA, that together with the transcriptional data corroborated a direct regulatory role for SsrB on srcA expression (Fig. 1A). To determine the cellular distribution of SrcA we constructed a srcA-HA allele and expressed this gene in wild type and in ssrB mutant cells under conditions that activate the SPI-2 T3SS [24]. In whole cell lysates, SrcA protein was reduced ∼10-fold in ΔssrB cells compared to wild type (Fig. 1B) and the protein was not detected in the secreted fraction from wild type cells (Fig. 1C), consistent with the expected properties of a T3SS chaperone. As a positive control, SseC, an SsrB-regulated translocon protein of the SPI-2 T3SS was present in the secreted fraction from wild type cells but not from an ssrB mutant.
Most SsrB-regulated gene products contribute to the intracellular survival of Salmonella in a host. In comparative genomics analyses, srcA was found in all virulent strains of Salmonella enterica containing SPI-2, but was absent from the cold-blooded animal commensal, S. bongori, which lacks SPI-2 (Table S1). This suggested a co-evolution of srcA with the SPI-2 T3SS and a possible functional relationship. If so, we reasoned that SrcA should contribute to animal colonization because the SPI-2 T3SS is essential for host infection. To determine whether SrcA contributes to Salmonella fitness in a host, we created an unmarked in-frame srcA deletion in S. Typhimurium and competed this strain against wild type cells in mixed oral infections of mice [25]. After three days of infection the geometric mean competitive index (CI) for the mutant was 0.20 (95%CI 0.13–0.29) and 0.18 (95%CI 0.06–0.5) in the spleen and liver respectively (P<0.0001; Fig. 1D) indicating that bacteria lacking srcA were significantly out competed by wild type cells during systemic infection. To verify the role of srcA on this phenotype, we complemented the srcA mutant with a wild type srcA gene under the control of its endogenous promoter, which restored in vivo fitness to that of wild type (Fig. 1D). The level of attenuation of the srcA mutant was generally higher than most single effector gene mutants [26], which suggested to us that SrcA contributes to an important aspect of T3SS function in vivo.
Sequence analysis showed 59% amino acid identity between SrcA and CesT, a secretion chaperone in enteropathogenic E. coli (EPEC) (Fig. 2A). As a means to address the biological function of SrcA, we solved the crystal structure at 2.5-Å resolution (PDB 3EPU). A summary of crystallographic data collection and model refinement statistics is in Table 1. The structure was solved by molecular replacement using an initial model based on CesT (PDB 1K3E) [13]. SrcA crystallized in space group C2 with two molecules related by a 2-fold symmetry axis in each asymmetric unit (Fig. 2B). Each monomer consisted of a small and large domain. The smaller domain formed by α1 and the extended loop region preceding β1 adopts a distinct conformation in each subunit. The larger domain mediates dimerization and is comprised of a twisted anti-parallel β-sheet (β1-β2-β3-β5-β4) flanked by α-helices α2 and α3. The dimer interface formed between SrcA monomers occurs primarily through reciprocal hydrophobic interactions between α2 and α2′ with additional interface-stabilizing interactions occurring between the α2 helix of one monomer and β4 and β5 strands of the opposing monomer (Fig. 2B). The total surface area buried at the dimer interface is 1258 Å2, suggesting that SrcA would exist as a dimer in solution, which we confirmed by gel filtration analysis (see below).
A structure similarity search with SrcA revealed proteins identified as T3SS secretion chaperones. CesT and SicP were the most structurally similar to SrcA, aligning with RMSD of 1.8 Å and 2.2 Å respectively. With the exception of CesT, SrcA has weak overall sequence identity (<20%) with other T3SS chaperones. CesT, SicP and SrcA contain several clusters of highly conserved amino acids notable on primary sequence alignments (Fig. 2A). Most of these conserved sites are located in the α2-interface helix and in strands β4 and β5 that help stabilize this interface. Although the N-terminus of these proteins is conserved structurally, the tertiary structures differ for each protein. In CesT, α1 and β1 adopt an extended conformation while the equivalent domain in SicP remains closely packed against the dimerization domain [12]. In SrcA, both extended and closely packed conformations are observed in separate subunits of the same dimer within the asymmetric unit. In the extended conformation the N-terminal helix from one dimer interacts with the β4 region of an adjacent dimer, similar to a domain swap seen in CesT [13]. At this time, the possible biological relevance for such a domain swap is unclear and may reflect an artifact of crystallization as critically discussed [13].
A comparison of the SrcA dimer interface with other class I chaperone family members indicates the overall similarity of quaternary structure shared between SrcA, CesT and SicP (Fig. 3A). This is in contrast to the class II chaperone interface of Spa15, which despite having similar tertiary structure to SrcA adopts a distinct dimer interface. A structural alignment of SrcA and Spa15 generated through alignment of single monomers shows the relative difference in subunit orientation between SrcA and Spa15 reflected by the positions of each monomer in the dimer configuration. These unique orientations produce an 80° rotational offset between respective subunits and could be expected to influence the mode of effector interactions utilized by these proteins.
To evaluate the potential for an effector-binding surface on SrcA, the structure of SicP in complex with its effector SptP was aligned with SrcA and represented as a space-filling model (Fig. 3B). Binding of SptP occurs primarily in the N-terminus of SicP [12], which is similar to the effector binding surface for SrcA predicted in silico. This surface contains several conserved hydrophobic residues including L16, D24, N26, and I32 (Fig. 2A), which is consistent with SrcA using a similar mechanism for interaction with effectors.
An emerging function for T3SS chaperones is delivery of cargo to the base of the apparatus through interactions with an ATPase. This was shown for the flagellar T3SS [17] and later in the virulence-associated T3SS in E. coli [16],[27] and the SPI-1 T3SS in Salmonella [15]. However, analogous interactions have not been described for the SPI-2 T3SS. Since srcA expression was co-regulated with genes in SPI-2, we hypothesized that it had a functional role in this system. To address this biochemically we purified SrcA and the predicted ATPase for the SPI-2 T3SS, SsaN, and performed binding experiments and gel filtration chromatography of the protein mixtures. SsaN contains conserved amino acid residues characteristic of Walker-A and Walker-B motifs of P-loop nucleoside triphosphate hydrolases, as well as a number of residues shown to contribute to ATP binding or ring stacking with the adenine base of ATP in the E. coli orthologue, EscN, (Q412, E191, R366) (Fig. S1). Since SsaN had not been characterized biochemically we first verified that our purified protein had ATPase activity (Fig. S1). We then mixed SrcA and SsaN proteins and resolved the protein complexes by gel filtration chromatography. By itself, SrcA existed as a dimer in solution (Fig. 4A) with no higher oligomers present, substantiating the stoichiometry obtained from our crystal data. SsaN existed as a monomer with a minor population eluting in a volume consistent with a probable dimer (Fig. 4B). When SrcA was mixed with SsaN, a new protein complex of high molecular weight was observed, along with diminished peaks corresponding to the SrcA dimer and SsaN monomer (Fig. 4C). This new complex elutes with a Stokes radius consistent with an apparent molecular mass of ∼600 kDa. We verified the identities of protein originating from each peak by western blot (Fig. 4D) and LC-MS/MS, which showed the new complex was comprised of both SsaN and SrcA.
Since structural and biochemical data unambiguously defined SrcA as a T3SS-associated chaperone, we used two experimental approaches to identify SrcA cargo(s). First, we used stable isotope labeling of amino acids in cell culture (SILAC) [28] in conjunction with quantitative mass spectrometry-based proteomics to identify cargo immunoprecipitated with SrcA from Salmonella. For this series of experiments we constructed a mutant in which the srcA gene was replaced on the chromosome with srcA-FLAG to enable immunoprecipitation from cell lysates. Lysates prepared from wild type cells grown in 2H4-Lys and 13C6-Arg containing SILAC medium (heavy) and srcA mutant cells grown in medium containing natural amino acids of Lys and Arg (light) were mixed and subjected to an immunoprecipitation procedure with an anti-FLAG antibody followed by quantitative mass spectrometry. Peptides originating from wild type cells contained heavy atom-substituted lysine and arginine such that putative SrcA cargo proteins would generate low heavy:light SILAC peptide ratios from the complex mixtures (Fig. 5A). In these experiments the T3SS effector protein SseL was identified by quantitative SILAC mass spectrometry as a specific SrcA cargo protein (Fig. 5B). SseL was immunoprecipitated specifically along with SrcA-FLAG with a SILAC ratio of 0.08, whereas additional abundant proteins displayed SILAC ratios closer to ∼1 (OmpF is shown, Fig. 5B) (mean SILAC ratio of all other peptides identified was 0.93 (Dataset S1).
Secondly, to verify the mass spectrometry data and to identify other possible effector cargo, we examined the secretion profiles of wild type cells and an srcA mutant that each expressed HA-tagged effector genes, the products of which are secreted by the SPI-2-encoded T3SS. Using this approach SseL-HA and PipB2-HA were depleted from the secreted protein fraction of srcA mutant cells (Fig. 5C) but reached similar levels in the bacterial cytoplasm (Fig. 5D). As expected from data with the complemented mutant in vivo, expression of srcA in trans restored effector secretion in the srcA mutant (data not shown).
To further show a role for SrcA in chaperoning PipB2, we set up experiments to test whether deleting srcA would phenocopy ΔpipB2 cells for PipB2-dependent centrifugal displacement of the Salmonella containing vacuole (SCV) in epithelial cells, an event linked to cell-to-cell transfer during infection in vitro [29]. At 10 h after infection the majority of SCVs were situated near the nucleus in accordance with previous work (Fig. 6A) [29]. By 24 h after infection SCVs containing wild type bacteria were displaced centrifugally towards the cell periphery whereas SCVs containing either pipB2 or srcA mutant bacteria remained juxtaposed to the nucleus (Fig. 6A). The average distance from the nucleus of LAMP1+ SCVs containing wild type bacteria was 2.19 µm at 10h post infection and increased to 7.86 µm by 24 h after infection. Conversely, SCVs containing either ΔpipB2 cells or ΔsrcA cells were 1.38 µm and 2.09 µm at 10h but lacked centrifugal displacement at 24 h (2.23 µm and 2.85 µm, respectively) (Fig. 6B).
We used a reverse genetics approach to define a new secretion chaperone in S. Typhimurium that is integrated functionally with the T3SS encoded by SPI-2, a system well described for its role in immune subversion and intracellular infection during host colonization. Consistent with other class I secretion chaperones, SrcA has extensive electronegative charge distributed over the surface of the molecule. The exact function of this charge distribution is not known, but data from other systems suggests a docking recognition function with other components of the type III apparatus, possibly the T3SS-accociated ATPase. For instance, electronegative surface residues on the SigE chaperone in the SPI-1-encoded T3SS negatively affect cargo secretion, but not cargo stability [30]. In enteropathogenic E. coli, a surface-exposed electronegative residue in the CesT chaperone (Glu142) likewise contributes to Tir secretion but not Tir binding [16], suggesting a role in either targeting bound cargo to the T3SS or in the secretion process itself. Interestingly, SrcA lacks 17-amino acids that make up the carboxyl terminus of CesT, which includes Glu142, and yet it still retains effector binding, ATPase binding and effector secretion functionalities. Thus, it is likely that other surface charged residues of SrcA are involved in these functions or that SrcA targets effector cargo to the secretion apparatus through a mechanism distinct from CesT.
The interface for the SrcA homodimer is extensive and is more in keeping with the structural features of single-effector class IA chaperones (∼1100–1300 Å2) compared to the reduced dimer interface of Spa15, a multi-cargo class IB chaperone from Shigella [31]. Similar to CesT and SicP, the dimer interface of SrcA adopts a parallel configuration when comparing α2 helices of opposing subunits. In contrast, the subunits of Spa15 undergo a significant relative rotation (80°) about the α2-axis resulting in a different interface. These features may relate to biological function in the SPI2 T3SS and/or in vetting effector cargo amongst the >30 effectors identified in Salmonella. We found no evidence of interactions between SrcA and translocon components of the SPI-2 T3SS and so it appears as though SrcA functions specifically in effector translocation events.
The interaction between SrcA and SsaN supports an emerging paradigm whereby secretion chaperones bring effector cargo to the T3SS through physical interaction with the hexameric ATPase at the base of the apparatus [14]. This was demonstrated for chaperone-ATPase components in the flagellar type III system [17] and in non-flagellar type III systems in E. coli [27] and the SPI-1 system in Salmonella [15]. Our work shows the first chaperone-ATPase interaction for a T3SS functioning from within an intracellular vacuolar compartment and supports this interaction as a more generalize feature of type III secretion function. In our experiments, we could induce the ATPase domain of SsaN to oligomerize in the presence of SrcA, but not in its absence, which was intriguing because the purified enzyme lacked a domain at the carboxyl terminus thought to be involved in oligomer stability, at least for E. coli EscN [14]. These data suggest that type III chaperones might have an as yet undefined role in assembly of the ATPase homohexamer that gives rise to efficient effector translocation. This will be an important area for further experimentation in this and other systems.
The genes encoding the srcA chaperone and the effector cargos (pipB2 and sseL) are found in all serotypes of Salmonella enterica that contain the SPI-2-encoded T3SS. Conversely, these genes are absent from S. bongori, which lacks the SPI-2-encoded T3SS. The expression of srcA is coordinated with T3SS transcriptional activity via the SsrA-SsrB two-component regulatory system encoded in SPI-2. The direct binding of SsrB to the promoter region upstream of srcA, along with SsrB-regulation of both sseL [19],[20] and pipB2 [32] is indicative of multiple cis-regulatory mutation events that have allowed for functional coordination of the distributed secretion apparatus, chaperone and effector cargos. We recently described this type of regulatory evolution for pathogenic adaptation of Salmonella to its host [21] and srcA is consistent with regulatory evolution of chaperone-effector gene pairs that are not co-transcribed in operons.
Due to low G+C base content compared to the genome average of 52%, it's likely that srcA (32% G+C) and an adjacent gene, STM2137 (37% G+C), were acquired as a foreign islet that was retained in organisms containing the SPI-2 T3SS due to the selective advantage afforded by the new protein interactions so created. Interestingly, STM2137 (also known as SseK2) is a likely paralog of SseK1, an effector translocated by the SPI-2 T3SS [33]. SseK2 is also regulated by the SsrA-SsrB two-component system but compared to SseK1, it is translocated in much less abundance into host cells [33]. Using the methods described here, we were not able to detect SseK2 secretion or a physical interaction with SrcA, however it remains possible that SrcA also chaperones SseK2 for low-level translocation.
SrcA is unique among other multi-effector chaperones most closely related to it in that it is unlinked from the T3SS genomic island. For example, InvB and SicP (Salmonella SPI-1), CesT (enteropathogenic E. coli locus of enterocyte effacement) and Spa15 (Shigella mxi/spa virulence plasmid region) chaperones are all encoded within the T3SS structural operons, implying they have co-evolved as a single genetic entity from a common ancestor. Given its genetic neighborhood, srcA appears to be a genetic acquisition separate from SPI-2 that functionally links some effectors to the T3SS apparatus via the ATPase. The role of horizontal gene transfer and regulatory evolution in allowing for uncoupling of chaperones, effectors and the T3SS has many possible implications for T3SS function, including plasticity in chaperone-effector interaction networks, expansion of effector repertoires, and alterations to the kinetics and hierarchical delivery of effectors to a host cell. These events may improve host adaptability or even expand the host range of bacteria that acquire and integrate new functional secretion chaperones.
All experiments with animals were conducted according to guidelines set by the Canadian Council on Animal Care. The local animal ethics committee, the Animal Review Ethics Board at McMaster University, approved all protocols developed for this work.
Salmonella enterica serovar Typhimurium strain SL1344 was used as the wild type strain and all mutants were isogenic derivatives. Chromosomal replacements were done using a λ-Red-based technique described previously [34]. A synthetic minimal medium for isotopic labeling of proteins in cell culture was developed for SILAC proteomics experiments based on LPM medium that activates the SsrA-SsrB two-component regulatory system for induction of SsrB-regulated genes [24]. LPM medium was modified for compatibility with quantitative SILAC mass spectrometry by replacing casamino acids with individual l-amino acids and containing either natural l-arginine and l-lysine, or 13C-subsituted arginine (13C6-Arg) and deuterium-substituted lysine (2H4,4,5,5-Lys) (Cambridge Isotope Laboratories, Andover, MA). A full description of LPM-SILAC medium is provided in Protocol S1.
For purification of His-tagged SrcA, the srcA gene was amplified from S. Typhimurium chromosomal DNA and cloned into pET-3(a) (Novagen) as a C-terminal fusion to a 6-histidine tag. Expression plasmids were transformed into E. coli Rosetta (DE3) and cells were grown in 1-L LB broth and induced with IPTG at OD600nm 0.6 for 3 h at 37°C. Harvested cells were resuspended in 25 ml NiA buffer (20 mM Tris pH 8.5, 500 mM KCl, 20 mM imidazole, 0.03% LDAO and 10% glycerol), lysed using a French press and centrifuged at 48,383 g for 40 min. Soluble His-tagged protein was purified using nickel-chelating resin (GE Healthcare Life Sciences), followed by Mono-Q anion exchange using a 20 mM Tris pH 7.5, 500 mM KCl, 10% glycerol. Purified protein was exchanged into a final buffer of 20 mM Tris pH 7.5, 100 mM KCl, 10% glycerol and concentrated to ∼5 mg/mL. All SrcA purification steps were carried out at room temperature. For purification of His-tagged SsaN, a soluble protein form containing the ATPase domain and C-terminal domain spanning residues Q90-E433 was constructed according to previous work done on E. coli EscN [14]). SsaNΔ89 was purified from E. coli Rosetta (DE3) cells containing a pET-3(a) plasmid with the ssaNΔ89 gene. Cells were sub-cultured into 1-L of Terrific Broth (TB) and grown with shaking at 200 rpm at 20°C for 65 h for auto-induction. Cells were harvested and lysed using a French press and soluble protein was purified using nickel chromatography and ion-exchange chromatography as described above. Finally, SsaN protein was concentrated to ∼9 mg/ml. All purification steps for SsaN were carried out at 4°C.
Crystals were generated via the hanging drop method by vapor diffusion using purified protein combined with crystallization solution (100 mM Bis-Tris propane pH 7.0, 200 mM MgCl2, 35% PEG 3350, 3.95 mM FOS-choline-9, 5% Jeffamine M-600) at a 3.5∶1 ratio, and equilibrated over 500 µL of 1.7 M ammonium sulfate, at 298K. After initial crystals were formed, drops were moved over wells containing 500 µL of 3 M ammonium sulfate and further equilibrated for 2 to 3 weeks. A single native data set, collected to 2.5 Å at the National Synchrotron Light Source Beamline X12C (Brookhaven, NY) was processed using HKL2000. An initial structural solution was achieved using molecular replacement with the type III chaperone CesT (PDB ID 1K3E) as the starting search model. PHENIX was used for model building [35]. Further model building and refinement was conducted iteratively using COOT and REFMAC [36],[37]. The final structure had R and Rfree values of 21.4 and 25.3 respectively.
For competitive infections, female C57BL/6 mice (Charles River) were infected per os with a 1∶1 mixed inoculum containing srcA mutant cells and a marked wild type strain resistant to chloramphenicol as described previously [21]. Competitive index (CI) was calculated in the liver and spleen at 3 days after infection as: cfu (mutant/wild type) output/(mutant/wild type) input. For complementation experiments, srcA was cloned with its native promoter into the low-copy plasmid pWSK29 and transformed into ΔsrcA cells. The complemented mutant was competed in CI experiments against wild type cells transformed with empty pWSK29.
Co-immunoprecipitations were performed with M2-Agarose beads conjugated with anti-FLAG antibody (F-gel, Sigma, Oakville, ON). Wild type bacteria and bacteria with a srcA-FLAG allelic replacement were grown overnight in LB broth, washed in SILAC-LPM (Protocol S1), and sub-cultured 1∶50 into SILAC-LPM containing either 12C6-Arg and H4-Lys (light) or 13C6-Arg and 2H4-Lys (heavy) amino acids. Isotopic labeling of proteins was carried out until the culture reached an optical density of 0.6 at 600 nm. Cells were washed with PBS, pelleted at 3000 g for 10 minutes and resuspended in PBS containing mini-EDTA tablet (1 per 10 ml) protease inhibitors (PBS-PI) (Roche, Mississauga, ON). Cells were sonicated six times for 30 seconds each with 1 min intervals on ice (Misonix Sonicator 3000, Misonix, Farmingdale, NY). Samples were centrifuged at 3000 g for 15 minutes and the resulting supernatants from heavy and light samples were mixed. F-gel was equilibrated with PBS-PI containing 10 µg/ml BSA for 60 minutes and then lysates were immunoprecipitated with F-gel for 16 h at 4°C. F-gel was washed with PBS-PI ten times for 30 min each wash. Bound proteins were eluted with either FLAG peptide or twice with SDS-sample buffer (1 M Tris pH 8.0, 20% SDS, 0.5 M EDTA pH 8, 10% glycerol, 200 mM dithiothreitol). Final protein preparations were filter-concentrated, washed with water and diluted to a final concentration of 50 mM ammonium bicarbonate and 1% sodium deoxycholate. Proteins were digested in solution and analyzed by liquid chromatography-tandem mass spectrometry exactly as described previously [38].
Two hundred microlitres of purified SrcA protein in gel filtration buffer (20 mM Tris pH 7.5, 200 mM KCl; protein concentration, 1.1 mg/ml) was injected into a Superdex S200 10/300GL gel filtration column (Amersham Biosciences, Piscataway, NJ) at 0.2 ml/min. Elution fractions (0.5 ml) were collected at a flow rate of 0.5 ml/min. For SsaN, 40 µl (8.84 mg/ml) was diluted with 320 µl gel filtration buffer and injected into an S200 column as described. For mixing experiments, 320 µl of SrcA (0.1mg/ml) was mixed with 40.8 µl SsaN (8.84 mg/ml) at room temperature for 2 h. The mixture was centrifuged at 10,000 g for 5 minutes and the top two hundred microlitres of the supernatant was injected into an S200 column. Peak fractions were collected and protein identities in all peaks were verified by Western blot and LC-MS/MS.
Experiments to monitor secretion of type III effectors were performed according to previously published methods [19]. Wild type cells and an srcA mutant used for these experiments were transformed with low-copy plasmids expressing HA-tagged effector genes from their endogenous promoters (sifA, sopD2, gogB, pipB, sseK2) or contained allelic replacements on the chromosome to express HA-fusion proteins (pipB2, sseL). Antibodies used for Western blots were: mouse anti-HA (1∶1000), mouse anti-DnaK (1∶5000), rabbit anti-SseC (1∶20000). Secondary antibodies conjugated to horseradish peroxidase (HRP) were used at 1∶5000 and antigen-antibody complexes were detected using enhanced chemiluminescence (ECL).
ATPase activity of SsaN was measured using the pyruvate kinase-lactate dehydrogenase coupled assay that monitors NADH oxidation coupled with ATP hydrolysis [39]. Data was plotted as a decrease in absorbance at 340 nm over time.
The intracellular position of Salmonella-containing vacuoles was determined by measuring the distance of LAMP1+ SCVs to the nearest edge of the host cell nucleus (labeled by DAPI staining) in fixed HeLa cells as previously reported [29]. Measurements were made using Openlab 3.1.7 software. Experiments were done in duplicate and the resulting finalized average was calculated from two independent average values (at least 100 measurements per experiment). Average distances with average deviation are reported.
The coordinates and structure factors of SrcA have been deposited in the Protein Data Bank (accession code 3EPU).
|
10.1371/journal.pcbi.1006856 | A demonstration of modularity, reuse, reproducibility, portability and scalability for modeling and simulation of cardiac electrophysiology using Kepler Workflows | Multi-scale computational modeling is a major branch of computational biology as evidenced by the US federal interagency Multi-Scale Modeling Consortium and major international projects. It invariably involves specific and detailed sequences of data analysis and simulation, often with multiple tools and datasets, and the community recognizes improved modularity, reuse, reproducibility, portability and scalability as critical unmet needs in this area. Scientific workflows are a well-recognized strategy for addressing these needs in scientific computing. While there are good examples if the use of scientific workflows in bioinformatics, medical informatics, biomedical imaging and data analysis, there are fewer examples in multi-scale computational modeling in general and cardiac electrophysiology in particular. Cardiac electrophysiology simulation is a mature area of multi-scale computational biology that serves as an excellent use case for developing and testing new scientific workflows. In this article, we develop, describe and test a computational workflow that serves as a proof of concept of a platform for the robust integration and implementation of a reusable and reproducible multi-scale cardiac cell and tissue model that is expandable, modular and portable. The workflow described leverages Python and Kepler-Python actor for plotting and pre/post-processing. During all stages of the workflow design, we rely on freely available open-source tools, to make our workflow freely usable by scientists.
| We present a computational workflow as a proof of concept for integration and implementation of a reusable and reproducible cardiac multi-scale electrophysiology model that is expandable, modular and portable. This framework enables scientists to create intuitive, user-friendly and flexible end-to-end automated scientific workflows using a graphical user interface. Kepler is an advanced open-source platform that supports multiple models of computation. The underlying workflow engine handles scalability, provenance, reproducibility aspects of the code, performs orchestration of data flow, and automates execution on heterogeneous computing resources. One of the main advantages of workflow utilization is the integration of code written in multiple languages Standardization occurs at the interfaces of the workflow elements and allows for general applications and easy comparison and integration of code from different research groups or even multiple programmers coding in different languages for various purposes from the same group. A workflow driven problem-solving approach enables domain scientists to focus on resolving the core science questions, and delegates the computational and process management burden to the underlying Workflow. The workflow driven approach allows scaling the computational experiment with distributed data-parallel execution on multiple computing platforms, such as, HPC resources, GPU clusters, Cloud etc. The workflow framework tracks software version information along with hardware information to allow users an opportunity to trace any variation in workflow outcome to the system configurations.
| Computational modeling and simulation has proven to be a powerful approach to reveal fundamental mechanisms of the cardiac rhythm in both normal and pathological conditions. Recent studies have expanded modeling approaches to the domain of predictive pharmacology, utilizing functional in silico approaches to predict drug efficacy, screen for drug toxicity, as well as suggest disease-specific therapies [1–11]. Modeling and simulation as an approach has distinct advantages over classical experimental methods, including the potential for high throughput prediction, choice of model complexity best suited for a given problem, and investigation of a range of physiological, pathophysiological and pharmacological parameters. Furthermore, computational modeling and simulation allows for the prediction of overall emergent effects of specific parameter perturbations on the simulated system.
As computational cardiac models have become increasingly accepted as predictive tools, there has been a recent movement towards utilizing them in applied venues, especially in the domain of safety pharmacology [12, 13]. This transition has required a deep and objective assessment of the need for well-defined criteria to allow for the verification, validation, and uncertainty quantification (VVUQ) of models and model predictions [13–15]. In the VVUQ paradigm, verification ensures the computational model accurately solves the equations underlying the mathematical model, and that model reproducibility is ensured regardless of implementation environment (i.e. different computing hardware, compilers, and code libraries), validation serves as a measure of the extent, to which the model is accurate in representing the quantities of interest (that may be experimental data), and uncertainty quantification determines the extent to which the model output is sensitive (or uncertain in response) to variation, error and uncertainty in the model input. In concert with VVUQ considerations, there has been a determined effort to address the overlapping issues of reproducibility, repeatability and replicability across a variety of computational disciplines via the application of standards [16–19] [14, 15, 20, 21].
CellML and related markup languages like SBML have been utilized to provide a standard, software- and programing language-independent description of the model, which can improve consistency and reproducibility of model description and sharing [22]. No single markup language can represent a full cardiac multi-scale model, although the combination of CellML to describe the ionic model, FieldML (http://physiomeproject.org/software/fieldml/about) for describing the field equations and geometry, and SEDML (https://sed-ml.github.io) [23–26] for describing the protocol of the numerical experiment, could in principle be combined to allow a full description.
Other tools have also been developed, such as CellML API or OpenCor that can automatically implement model representations in markup languages [27, 28]. In this way, it is possible to generate whole cell ODE model equations from a language independent CellML description of the model. There are some examples of integrated frameworks (OpenCMISS [29, 30], Chaste [31–34], CARP [35]) that can solve multi-scale models that are derived from standardized model descriptions and indeed, Chaste and CARP can both be integrated and utilized in Kepler workflows [36]. Some multi-scale simulations do, however, require the use of a variety of solvers and data sets.
Moreover, reproducibility also requires development of standards for simulation and model implementation [20, 23, 25, 26, 37, 38]. SED-ML is a community effort to standardize modeling protocols, but standardized protocols that integrate or connect multiple models represented in standardized model descriptions either requires customized software or a workflow framework [24, 25, 39, 40]. To date, there are a few tools that support SED-ML (Tellurium, JWS Online, SBW Simulation Tool, CellDesigner, COPASI, iBioSim, bioUML, SED-ED) for a limited number of application domains. We tested here whether a workflow platform such as Kepler could provide a reproducible approach for integrating multi-scale models requiring more than one solver, a reproducible protocol for numerical experimentation and provenance tracking. Indeed, none of the tools described are mutually exclusive and workflows such as the one described in this study can be readily expanded to allow inclusion of code generation from CellML, FieldML and SEDML descriptions [16].
In this study, after careful analysis, we decided to utilize the Kepler scientific workflow management system. This framework enables scientists to create intuitive, user-friendly and flexible end-to-end automated scientific workflows using a graphical user interface. Kepler is an advanced open-source platform that supports multiple models of computation [41, 42]. The underlying workflow engine handles scalability, provenance, reproducibility aspects of the code, performs orchestration of data flow, and automates execution on heterogeneous computing resources. A workflow driven problem-solving approach enables domain scientists to focus on resolving the core science questions, and delegates the computational and process management burden to the underlying Kepler Workflow system [43–46]. Further, scientists can parameterize the workflow and perform large-scale search for optimal values in the parameter space. Leveraging the benefits of a workflow driven approach allows scaling the computational experiment with distributed data-parallel execution on multiple computing platforms, such as, HPC resources, GPU clusters, Cloud etc. The framework gives users flexibility to execute the workflows from command-line or GUI. Due to its large open-source developer community, Kepler has a rich library that contains over 350 ready-to-use processing components called 'actors' that can be easily customized.
There have been a number of developments aimed at solving the specific problems of reproducibility, repeatability and replicability. In the context of the work presented here, ‘reproducibility’ refers to zero-difference in outcome between two executions (say W1 and W2) of same workflow (W), when both executions of our workflow W have exact same hardware (H), same software (S), and same initial conditions (P). The Kepler workflow system captures provenance during each execution at multiple levels. The workflow records the workflow parameters, workflow outputs, intermediate data tokens and extracts the hardware system (CPU Cores, Cache, Memory etc.) profile as well. All of this information is recorded in the workflow provenance database. The information includes versions of all the programs. The key components recorded are the version information of the operating system, source code compiler, Python, Kepler, Java and associated source code. The workflow stores this information in Kepler provenance database. The detailed capture of hardware and software environment information enables users to completely reproduce and replicate the results. Our aim is to facilitate the user to setup same initial conditions and hardware environment (if required), and reproduce results in similar fashion. Notably, the definition we use for reproducibility has been described as replicability and repeatability in other descriptions, whereas reproducibility has used to describe an independent reconstruction of the model from the model equations and initial conditions [47–49]. Indeed, despite attempts to develop standard definitions, there is, as yet, no full consensus on the definitions of each term [50].
One of the main advantages of utilization of workflows is that they can integrate code written in multiple languages, allow for variation in application of compilers and can pass information from one code to another. The standardization occurs at the interfaces of the workflow elements (actors) and allows for very general applications and easy comparison and integration of code from different research groups or even multiple programmers coding in different languages for various purposes from the same group.
Kepler workflow elements can be optimized to run on different platforms and compare results (verification), switch code for different models or implementations of the same model and compare results (validation) or run code multiple times with different initial parameters and estimate variation (uncertainty quantification). Also for the reasons above, Kepler workflows are ideally suited for multi-scale modeling due to ability to integrate very different pieces of codes into a workflow and easily parsing input and output parameters between them. Another advantage is that Kepler workflows are easily accessible for non-experts in computational modeling as programming as a detailed knowledge of model inner workings are not needed to run simulations and modify parameters to suit the requirements of the end user.
Here, we present a multi-scale model of cardiac electrophysiology that is executed in the freely available Kepler scientific workflow system [41, 44]. The workflow we present here is a first required step in VVQU by ensuring reproducibility of models through inclusion of provenance information that describes the origin of the model components, referencing to the data, information about any modifications and the associated rationale, as well as the specific components and parameter settings used in each run. We implemented differential equation models of cardiac physiology that automate the execution of simulations with user defined options of outputs from a single cell (0-dimensional), 1 or 2-dimensional tissue, and a pseudo-ECG output, which can be compared to experimental or clinical data.
Many instances of models can be used with varying input parameters, and the models can be linked in the workflows in various ways. For example, single cell models can be linked to an idealized 1-dimensional fiber model, which allows us to compute a signal averaged pseudo ECG that captures temporal and spatial electrical potential gradients of a propagating wave. Another example demonstrates a thousand instances of the single cell model being linked to a 3-dimensional transmural wedge preparation for investigation of ectopic sources. In addition to these multi-modal choices, the framework can also be reused for multispecies comparisons. Users can control a wide range of input parameters from a simplified command-line, or GUI interface. The workflow is portable and scalable, having the flexibility to run on any platform a user chooses: local workstations, small clusters, or remote HPC resources.
The computational workflow we present here represents a proof of concept of a platform for the robust integration and implementation of a reusable and reproducible cardiac cell and tissue model that is expandable, modular and portable. The detailed checkpointing of version information along with hardware information gives users an opportunity to trace any variation in workflow outcome to the system configurations, when the infrastructure cannot be exactly replicated. In addition to storing in the database, the workflow generates an execution report for each workflow execution that includes the important workflow parameters, input information, software version and hardware system profile.
A cardiac ventricular electrophysiology modeling and simulation use case:
We present an automated computational workflow (Fig 1) that can perform simulations to generate user defined instances and configurations of a single-cell cardiac action potential, conduction of a cardiac action potential in a 1-dimensional (1D) or 2-dimensional (2D) tissue representation and generation of a signal average of electrical activity in time and space to represent a pseudo-ECG.
Please access all codes and associated files and attributes via the GitHub link below. The repository contains specific instructions for use of Kepler System with new source codes. The user-manual provides detailed outlines of how to install Kepler, modify workflow parameters, choose the execution platform and get results from the multi-scale cardiac workflow. The user manual can be accessed at the root of the git repository under filename: “UserManual.docx”
https://github.com/ClancyLabUCD/Workflow_Kepler
Here we demonstrate several example scenarios including: (a) Deployment of the workflow for a single-cell simulation to predict a cardiac action potential with a defined set of input parameters, (b) a configuration for a 1-dimensional cardiac tissue simulation, or (c) a 2-dimensional cardiac tissue simulation.
The model formulations for ventricular cells (the Soltis-Saucerman model [51], Morotti-Grandi model [52], or Grandi-Bers model [53] merged with the Soltis-Saucerman model) were implemented in the Kepler workflow. The source code of simulation models has been implemented in C++ and is compiled during the workflow execution using icc or gcc compiler, depending on the execution platform and compiler availability. Users can use the source code provided by us or attach their custom developed simulation models by editing the workflow parameter “sourceCode.” The workflow gives users a choice to select “compilerProgram” parameter. The workflow integrates multistep single-cell (black circle symbol), 1-dimensional (black rectangle symbol) and 2-dimensional (black square symbol) tissue model simulations in a single automated process (Fig 1). The SingleCell-Sim module includes a sub-workflow that performs single-cell simulation. Likewise, OneD-Sim and TwoD-Sim modules perform 1-dimensional and 2-dimensional tissue model simulations, respectively. The workflow includes user configuration components, simulation components with multiple execution choices and post-processing components for each model.
User configured parameter settings and initial conditions also allow the end user to control simulation constraints for single-cell, 1D and 2D modules (such as Na+-blocker concentration; rapid delayed rectifier potassium channel conductance, GKr, block ratio; ligand (β-blocker isoproterenol) concentration, CaMKII (Ca2+/calmodulin-dependent protein kinase II) activity levels; number of beats and others through workflow parameters. The simulation constraints are ported as workflow parameters, which can be modified and passed to the simulation models using the user configuration module. This workflow module is implemented using Kepler Python actor and Python libraries. Users can seamlessly configure the simulation parameters simply by changing workflow parameter values through command line or GUI as shown in Fig 1 (purple arrow). Many instances of these models can be used with varying input parameters, and the models can be linked in the workflows in various ways. The internal structure of the workflow element (actor) is shown in Fig 2. User parameter configurations can also be expanded to include more parameters by modifying a workflow actor.
The workflow incorporates flexibility for the end-user’s choice of platform depending on the use case and resource availability. Users can run the workflow on multiple computing platforms such as local, private clusters, and remote HPC clusters by configuring execution choice parameters for individual processes. Kepler allows customization of each execution instance of a workflow with user input parameters. In Fig 3, the Kepler’s Execution Choice actor was created in the Core single-cell module. The Local Execution Options and the Remote execution options are also available in the options menu at the top of the GUI. The capability of multiple execution choice on different hardware platforms is achieved by using the Kepler workflow system. By design, the Kepler framework is capable of automatically creating new jobs for execution. This functionality enables scientists to change execution platforms (local or remote) without any additional user scripting.
The post-processing module generates output data files from single-cell, 1D and 2D tissue simulation results. The workflow uses Python libraries and Kepler actors to post process the simulation results and generate plots for simulated action potentials (AP), main ionic currents (ICa, IKr, IK1, INCX, INa, Ito, IKs), intracellular (cytosolic) and sarcoplasmic reticulum concentrations of Ca2+ and Na+ in single cells, pseudo ECG in a 1D-simulation, and snapshots of AP propagation in 2D tissue.
Further, the Kepler workflow automates the provenance collection, execution report generation and reproducibility. For basic execution of the workflow in “as-is” condition, users do not need expertise in the technologies used, and can execute the workflow using GUI and command line.
All simulations of three cardiac myocyte models (the Soltis-Saucerman model [54], Morotti-Grandi model [52], or Grandi-Bers model [53] merged with the Soltis-Saucerman model) were encoded in C/C++, and run using GCC complier on Mac Pro or Linux computers.
The numerical method used for updating the voltage was forward Euler. Single cell action potentials (APs) and selected ionic currents were recorded. For higher dimension simulations, we simulated a transmural fiber composed of 165 ventricular cells (Δx = Δy = 100 μm) connected by resistances to simulate gap junctions [55]. The transmural fiber contains an endocardial region and epicardial region with a linear decreased in APD as indicated by experimental data [56, 57]. GKr was used as the index value of endocardium in the cell #1, and the index value of epicardium in cell #165. We can simulate a heterogeneous 2D cardiac tissue composed of 165 by 165 cells with Δx = Δy = 100 μm. The tissue contains an endocardial region and epicardial region with a linear decreased in APD as indicated by experimental data [56, 57]. Channel conductance and gap-junction parameters are same as in the one-dimensional simulations. Current flow is described by the following equation:
∂V(x,y,t)∂t=Dx∂2V(x,y,t)∂x2+Dy∂2V(x,y,t)∂y2−Iion−IstimCm
Where V is the membrane potential, x and y are distances in the longitudinal and transverse directions, respectively, Dx and Dy are diffusion coefficients in the x and y directions, Cm is membrane capacitance (Cm = 1). Istim is 180 mA/cm2 for the first 0.5 ms. We also incorporated anisotropic effects by setting Dx and Dy such that the ratio of conduction velocities is 1:2 [58].
One of the key added advantages of using Kepler Workflow system is the ability to deploy new source code easily. To facilitate execution of new cardiac cell models, the path to C++ source code file is parametrized in the Kepler Workflow. If a scientist wants to use their customized cardiac cell model, she/he can edit the Kepler Workflow parameter called 'sourceCode' under the category of 'SharedParameters', to point to the directory where the desired C++ source code resides. Further, the parameters unique to a given source code can be defined in a file called ‘stim_param.txt’. The last step in parametrization is to add a placeholder in the Kepler user interface using the ‘Parameter’ option under the ‘Workflow Input’ menu.
We first demonstrate the potential for the Kepler workflow environment to be used to run batch simulations for a simulated human ventricular single-cell model for varying degrees of IKr reduction (Fig 4). The workflow allows users vary rapid delayed rectifier potassium channel conductance, GKr, in the simulations. Fig 4 illustrates single-cell APs and the time-course of IKr through the Kepler workflow with varying GKr. In the top of panels A-D, various end user configurations for input parameters are shown for each simulation instance. In the middle row, simulated single-cell action potentials are shown. In the bottom row, the time-course of Ikr during the AP is shown. The Gkr was reduced via the indicated (green arrows–top panels) block ratios of 1 (used as control, panel A), 0.75 (B), 0.50 (C) and 0.25 (D), corresponding to IKr block of 0, 25, 50 and 75%, respectively.
In Fig 5, we demonstrate expansion of the workflow beyond single-cell simulation to user defined 1D and 2D-simulations. The workflow generates a single cell cardiac ventricular action potential (Fig 5A), as well as a one-dimensional simulation and a pseudo-ECG (Fig 5B) and then ingests steady-state results from 1D simulations to seed 2D simulations shown in Fig 5C. In this example, the single cell was simulated at a pacing rate of 1 Hz and 10 action potentials were generated. The last AP (10th beat) is shown in Fig 5A (bottom panel). In the tissue simulations, we simulated a heterogeneous fiber (with a linear decrease in AP duration from endocardial to epicardial region [57] (i.e. from the innermost to the outer layer of the cardiac tissue) composed of 165 ventricular cells (parameter tissue length = 165 cells in Fig 5B, top) for three beats. The pseudo-ECG is shown in Fig 5B (bottom panel). In the panel C, we demonstrated 2D AP wave propagation in response to one stimulus (a planar wave). The workflow simulated a heterogeneous 2D cardiac tissue composed of an array of 165 cells by 165 cells (1.65 cm x 1.65 cm) [57].
In the example shown in Fig 6, we tested three different species models and performed simulations to generate propagation of an action potential in one dimension using the topology shown in Fig 6 (top). Pseudo-ECGs are shown in response to seven stimuli at 1 Hz in Human (Fig 6A—orange), Rabbit (Fig 6B—purple) and Mouse (Fig 6C—green). This example demonstrates how the workflow cyberinfrastructure can also be re-used as a multi-species simulator by utilizing single cell cardiac ventricular computer models as inputs into the higher dimensional models. The cell model of choice can be linked to an idealized one-dimensional fiber model, which can be used to compute signal averaged pseudo ECG traces (Fig 6A–6C). They capture temporal and spatial gradients of electric potential during a simulation that tracks conduction and repolarization of a propagating wave.
There has been a tremendous increase in both the number of cardiac models in existence, and in model complexity over the last several decades, correlating with both an increase in computational power, and dramatically reduced computational cost. These developments have created the potential for cardiac cell models and their mathematical and/or agent-based model components to be reused and coupled with one another, creating flexible, modular, portable and potentially scalable models that can account for a range of attributes [60]. The potential for linking models together in new ways also suggests construction of multi-scale models from existing models at various temporal and spatial scales.
To ideally enable model modularity, reuse, reproducibility, portability and scalability, a model execution platform should be able to provide the reuse of code, reproduction of reported in silico predictions, as well as a way to run simulations in an efficient, expandable, modular and portable manner. Scientific workflow tools allow exactly these elements and can provide a user interface and potential for automation and optimization of software and hardware elements of the model execution. Workflows derive from the concept of directed graphs with individual nodes that represent discrete computational components that can be optimized to execute on distinct hardware architecture [61–64]. A scientific workflow is conceptualized as a set of tasks performed on a collection of datasets. The workflow-based design enables scientists to break large computational tasks into smaller manageable and reusable modules (nodes). The data flows through these modules (nodes) and gets transformed. The scientists can collaborate effectively on a large-scale problem by bringing their expertise to different modules in a workflow. Data and results flow between the individual nodes.
The computational overhead involved with workflow implementation is during the start of the Kepler Workflow Engine, and the added cost of building the workflow graph in the Kepler GUI. However, this is a one-time cost during a single execution. Once the Kepler Workflow Engine is up and running, the real advantage comes from automated execution, automated provenance collection, and parameterization driven extensibility benefits. In essence, the end user will get these benefits at least, and these can be enhanced by creation of a wrapper mechanism using the Kepler system, that enhances modularity, shareability and extensibility of their work. Wrapping with Kepler Workflow system enhances the portability of source code—it can work on local machine, or on a distributed cluster—so users are not required to modify or write any script for change in execution platforms. Moreover, diversity of parameters can be handled at two levels, when designing the wrapper workflow. The parameters common to an application area, can be abstracted away and customized at the workflow level (indicated by purple arrow in the Fig 1). The parameters unique to a given source code can be defined in a file called ‘stim_param.txt’ and the user needs to add a placeholder (Fig 1 –User configuration Parameters) in the workflow definition using the ‘Parameter’ option under the ‘Workflow Input’ menu.
The added cost of Kepler workflow system can be hedged by exploiting potential to parallelize codes across distributed systems, for problems involving large-scale computation and large datasets. The Kepler system has inbuilt mechanisms to quickly divide and conquer large computations in batch parallel computations. For cases involving specific cost benefit analysis, since the computational overhead of Kepler is dependent on each workflow, we suggest performing case-specific measurement of ‘Kepler + SourceCode + Parallelization Director’ against isolated run of ‘SourceCode’. We are happy to provide support for such efforts, using our support team for open source users of Kepler.
One critical feature of the Kepler that was decisive in selecting this engine, is the “provenance module”. This module archives workflow execution history, parameters, software and hardware signatures. Workflows Provenance can help preserve evidence and data from experiments to achieve reproducibility [43, 65, 66]. The Kepler reporting module generates informative and detailed summaries of the execution that include user configuration parameters used during the execution in various simulation steps, version of respective software tools, and system hardware information on which the workflow is executed. This “execution-signature” can drastically reduce time required to write reports or methods and material section in scientific publications, enabling domain experts to focus their energy on problem solving [43, 45, 65, 66]. Use of Kepler enabled us to delegate these critical components to the framework, and allowed us to focus on the science behind the problem.
It is important to note that workflow frameworks are not an alternative to markup languages for model description or simulation experimentation or an alternative to specialized packages that can integrate more than one kind or scale of model, but rather than efficient and reproducible approach to multi-scale modeling using multiple component models, software tools and data sets that facilitates usability, sharing and provenance tracking.
Here, we demonstrated the application of the freely-available Kepler scientific workflow system to execute a multi-scale model of cardiac electrophysiology. The workflow allows for modularity, scalability and flexibility in a deployable framework that can be configured by the end-user for maximum flexibility. Like most computational scientists, we have long shared concerns about the reproducibility and reuse of models. Versioning and provenance information can be included in Kepler workflow approach as well as the origin of the model components and user defined components and parameter settings used in each run. In this demonstration, we utilized Kepler to develop a workflow containing differential equation models of cardiac physiology that automate the execution of simulations with user defined options of outputs from a single cell (0-dimensional), 1 or 2-dimensional tissue, and a pseudo-ECG output, which can be compared to experimental or clinical data.
The workflow as presented could be readily adopted and expanded for applied use in the safety pharmacology domain. In both clinical and experimental settings, prolongation of the QT interval of the ECG and related proarrhythmia have been so strongly associated, that a prolonged QT interval is largely accepted as surrogate marker for proarrhythmia. Here we demonstrate how the workflow can be applied to an investigation of the impact of perturbation of the key repolarizing potassium current in the heart, the rapidly activating component of the delayed rectifier potassium current, IKr. Mutations in the potassium channel gene encoding IKr or drug-induced inhibition of IKr can lead to inherited or acquired long QT syndrome. The QT interval is a phase of the cardiac cycle that corresponds to action potential duration (APD) including cellular repolarization (T-wave). Our single-cell examples demonstrate that reduction of Ikr caused AP prolongation (Fig 4). In Fig 5, the workflow can be used to predict QT intervals in the setting of 1-dimensional tissue or further investigate repolarization phases on 2D AP propagation maps by modifying IKr. Finally, in Fig 6 our Kepler workflows allow to easily demonstrate that cardiac electrical signal propagation varies in different species used in experimental studies. And using this approach we can relate findings from animal model studies and correlate them to clinical human studies as well.
While considerable attention has been given to the prospects of computational modelling and simulation as a platform for prediction of cardiac drug safety, electro-toxicity and proarrhythmia risk assessment, less scrutiny over the choice of model and the impact of model choice on predicted effects has been given. Here we also show how the Kepler multi-scale workflow can be applied to multispecies to allow users to perform preliminary assessments in models for which predetermined selections of validation experiments can be performed.
The Kepler cyberinfrastructure enables biomedical scientists to (1) understand and catalog accuracy for assembly and linking of models through rigorous uncertainty quantification (UQ) and sensitivity analysis, (2) define a common practice and methodology for linking together (big) data and high-throughput, multi-spatial, multi-temporal, and complex models through reusable workflow definitions, execution, and tools, (3) develop a user interface building toolkit, and (4) develop new methods for deployment and distribution of highly scalable, portable, expandable and robust software and platforms. An additional benefit of this approach is that it allows for individual workflow elements to be optimized for hardware to maximize efficient parallel computing. Various processes of the workflow can be distributed to execute on optimized systems and then pass data though linkage between the workflow elements.
In the near future, our next steps will include the development of an online training course package with lecture material, videos and hands-on on this Multi-scale Cardiac Workflow tool on the e-learning platform called Biomedical Big Data Training and Collaborative (BBDTC) as our educational and community outreach efforts. The BBDTC (https://biobigdata.ucsd.edu) is a community-oriented platform that encourages collaborative efforts on training and education to ensure high-quality knowledge dissemination to biomedical big data scientific community. The BBDTC provides easy and intuitive interface to create, launch and share open training materials and tools for biomedical community [42, 67].
Future plans also include goals to integrate this workflow with our Machine Learning based performance prediction module to efficiently schedule different components of the workflow on available computing hardware in a way to gain performance and resource optimization [68–70]. We will couple the workflow with our provenance-based fault tolerance framework to automatically detect failure point and re-start the execution of the workflow from point of failure to save time and resources [71].
In summary, we have developed a Kepler based workflow for multi-scale cardiac electrophysiology that can be utilized and expanded for any number of predictions as defined by the end user. The approach brings us closer to the increasingly shared goal of computational scientists to enable model modularity, reuse, reproducibility, portability and scalability. The workflow concept also allows a model execution platform that allows the reuse of code, reproduction of reported in silico predictions, as well as a way to run simulations in an efficient, expandable, modular and portable manner. We have demonstrated an application of the approach by linking models together for construction of multispecies multiscale models from existing models at various temporal and spatial scales.
|
10.1371/journal.pntd.0003622 | Treatment of W. bancrofti (Wb) in HIV/Wb Coinfections in South India | The disease course of human immunodeficiency virus (HIV) is often altered by existing or newly acquired coincident infections.
To assess the influence of pre-existing Wuchereria bancrofti infection on HIV progression, we performed a case-controlled treatment study of HIV positive individuals with (FIL+) or without (FIL-) W. bancrofti infection. Twenty-eight HIV+/FIL+ and 51 matched HIV+/FIL- subjects were treated with a single dose of diethylcarbamazine and albendazole (DEC/Alb) and followed for a year at regular intervals. Sixteen of the HIV+/FIL+ subjects (54%) and 28 of the HIV+/FIL- controls (57%) were on antiretroviral therapy (ART) during the study. Following treatment, no differences were noted in clinical outcomes between the 2 groups. There also was no significant difference between the groups in the HIV viral load at 12 months as a percentage of baseline viral load (HIV+/FIL+ group had on average 0.97 times the response of the HIV+/FIL- group, 95% CI 0.88, 1.07) between the groups. Furthermore, there were no significant differences found in either the change in viral load at 1, 3, or 6 months or in the change in CD4 count at 3, 6, or 12 months between the 2 groups.
We were unable to find a significant effect of W. bancrofti infection or its treatment on HIV clinical course or surrogate markers of HIV disease progression though we recognized that our study was limited by the smaller than predicted sample size and by the use of ART in half of the patients. Treatment of W. bancrofti coinfection in HIV positive subjects (as is usual in mass drug administration campaigns) did not represent an increased risk to the subjects, and should therefore be considered for PLWHA living in W. bancrofti endemic areas.
ClinicalTrials.gov NCT00344279
| In people living with HIV infection, simultaneous infections can adversely affect HIV disease. This has been seen with bacterial (tuberculosis), viral (cytomegalovirus), and parasitic infections (toxoplasmosis). Lymphatic filariasis is caused by a thin thread-like parasite that lives in the lymph vessels of infected people. It can cause significant disability. This infection is found in much of the same areas that high levels of HIV infection. We were interested in knowing if lymphatic filariasis changed the course of HIV infection in people with both diseases. In this study, the authors enrolled people in India who were living with HIV who either had or didn’t have filarial infection. All patients were treated for filariasis with 2 drugs, and then were followed for 1 year to see how their HIV disease progressed. No difference in HIV disease progression was found between the groups that did or did not have filariasis before treatment. The patients with HIV did well with the medicine for filariasis.
| As the HIV epidemic continues in many parts of the world, more attention is being focused on strategies for prevention and management of HIV infection. In addition to highly active antiretroviral therapy (ART), the immune interactions between HIV and non-HIV co-infections have been examined. Several groups have examined the interaction of helminth infections with HIV, and many of these studies have been recently reviewed[1–3]. Some studies have shown that patients with HIV and concomitant helminth infections have higher viral loads which decrease upon anthelmintic treatment [4,5] whereas others have shown no effect of coincident helminth infections on viral load, CD4 count or HIV disease progression [6–9].
Few studies have looked at the interaction of filarial infections with HIV. In studies of patients with Onchocerca volvulus infection with and without HIV, those with HIV were found to have more significant skin disease [10] and were less likely to have antibodies to onchocercal antigens [11]. Despite these differences, HIV/Onchocerca-co-infected patients were as capable as HIV negative onchocerciasis patients of responding to treatment with ivermectin [12]. A cohort of HIV+ individuals coinfected with Wuchereria bancrofti (Wb) have been followed in Tanzania [13–15]. When treated with diethylcarbamazine (DEC), those who were circulating filarial antigen (CFA) positive (a surrogate for active infection) had a decrease in HIV viral load at 12 weeks whereas those who were CFA negative had an increase in viral load [14] following DEC administration. In a previous study, we have looked at the coinfection prevalence rates in India by surveying serum samples of HIV infected patients for filaria antigens. We found a prevalence of 5–9.5% of filarial antigenemia in HIV+ patients, similar to the prevalences found in the HIV negative population in the same region [16].
None of the studies to date has looked at the interactions between HIV and helminths outside of Africa. According to data from the Indian National AIDS Control Organization the 2011 estimated prevalence of HIV in adults in India was 0.27%. This results in about 2.1 million people infected with HIV [17]. The state of Tamil Nadu, in the south, has a higher than average adult HIV prevalence, with an estimated 133,000 infected adults [18]. Use of antiretrovirals is still rolling out in this population: as of January 2012, only 486,173 people were on ART [19].
This study was approved by the Institutional Review Board of NIAID, the Ethics Review Boards at the Tuberculosis Research Centre (TRC, now the National Institute for Research on Tuberculosis) and YRGCare as well as by the Health Ministry Screening Committee of the Government of India. All subjects provided written informed consent.
This was a prospective matched case-control study designed to compare HIV replication and progression of clinical disease between patients co-infected with HIV and Wb infection and patients that were HIV-positive but without Wb infection.
Albendazole was provided by Glaxo-Smith-Kline (GSK, Middlesex, UK) as part of the Global Program to Eliminate Lymphatic Filariasis (GPELF). Diethylcarbamazine (DEC) was from GSK-India, which is the same source that provided DEC for the GPELF. The study medication was stored at the two sites in Chennai where the study was conducted in the pharmacy under the control of a pharmacist or nurse. The study staff directly observed subjects as they took the medications.
Existing or new adult subjects ≥18 years of age cared for at clinics run by YRGCare or the TRC at their sites in Chennai, India were recruited for this study. These subjects were all HIV+ by enzyme-linked immunosorbent assay (ELISA) and Western Blot testing. About half of those enrolled were on antiretroviral medication. Subjects were eligible if they were willing to participate and could give informed consent. Pregnant or lactating women were excluded, as were subjects that were acutely ill, had a hemoglobin <9 g/dl for women and <10 g/dl for men, had AST or ALT elevations >5 times the upper limit of normal, had evidence of acute HIV infection, or had active or known untreated tuberculosis.
Subjects willing to participate were screened for filarial antigenemia using a rapid immunochromatographic test (Filariasis Now, Binax, Portland, ME). Those found to be filarial antigen positive (cases) were enrolled, and data collected included demographics and clinical details. Each “case” who fulfilled the other inclusion criteria was matched for purposes of comparison with up to two HIV+, Wb-negative patients (controls) from the screened clinic cohort matched for age (within ±5 years), gender, viral load (matched to within ± 0.5 log HIV RNA copies), CD4 counts (matched to within the same range: <50; 50–100; 101–200; 201–300; >300) and for anti-viral regimes (that were similar enough in efficacy for purposes of effective comparison).
At screening, all subjects had a thorough clinical assessment: a history and physical examination, blood for complete blood counts, liver functions (ALT, AST), CD4 count, serum storage for later analyses; stool samples were taken for examination of parasites. Female subjects underwent a urine pregnancy test. Potentially eligible subjects underwent a viral load. All enrolled subjects received a single dose of the combination of DEC (300 mg)/albendazole (400 mg). Subjects were followed at 1, 3, 6 and 12 months after receipt of the medication. At each visit, subjects were queried as to their interim history: any illnesses, new diagnoses, or changes in medications were noted. At all visits, a CBC and HIV viral load was scheduled; CD4 counts were scheduled at 3, 6 and 12 months. At the 12 month visit, all patients were retested for filarial antigenemia and, if positive, offered retreatment with DEC and albendazole.
Laboratory assays: Filarial antigen status was assessed in real time by immunochromatographic card test (Filariasis Now Binax, Portland, ME). Samples were stored, batched and grouped to run the Wuchereria bancrofti antigen ELISA (Tropbio Townsville, Australia) to quantitate levels of circulating filarial antigens. Filaria specific IgG and IgG4 levels were measured by ELISA as previously described [20].
The original endpoints of the study were the change in HIV viral load and the difference in clinical status between the cases and the controls one year after treatment with DEC and albendazole. We also examined the change in HIV viral load, CD4 counts and hemoglobin throughout the study.
Differences in groups were tested by Wilcoxon-Mann-Whitney (WMW) test (for continuous responses), Fisher’s exact test (for categorical responses) and Wilcoxon signed rank (WSR) test for paired analyses. Additionally, a version of the WMW stratified by ART status was used [21].
Viral load was analyzed on a log10 scale. According to the guidelines for HIV treatment, baseline CD4 count was put into 1 of 4 categories: <200, 200–350, 350–500 and ≥500. Patients were classified as either on or not on ART during the study. Two patients started the study not on ART and began ART during the course of the study, and those two patients are treated as on ART for the main analyses. As a sensitivity analysis, all analyses were repeated with those two patients reclassified as off ART to see if there were substantial differences in the results.
A linear model was used to assess the association of the response log percentage of baseline viral load at 1 year with filarial co-infection, adjusting for other covariates (ART status, baseline hemoglobin, baseline CD4 count and baseline viral load) by including them in the model as main effects. We report filarial coinfection effects in terms of a multiplicative factor acting on the response (percentage of baseline viral load at 1 year) in the model. Similar linear models were done at 1 month, 3 months and 6 months, as well as with CD4 counts. All tests were 2-sided and at a significance level of 0.05. All analyses were conducted using the software R (version 2.15.2).
376 subjects were screened (Fig. 1). Twenty-eight subjects were found to be filarial positive (HIV+/FIL+). These were matched with 52 HIV positive subjects who did not have Wb infection (filarial-negative) (HIV+/FIL-) as controls (one control subject had no baseline viral load). Baseline summary statistics are given in Table 1. Because the groups are matched, as expected there are no significant differences between the groups (Table 1, Fig. 2). Many subjects in both groups took co-trimoxazole and a multivitamin throughout the study. Less than 10% of subjects were positive for ova or parasites on stool exam, so this variable was not included in the analyses.
The treatment with DEC/albendazole was well tolerated in our study subjects, with no adverse events attributed to the study medication. There were several unrelated significant illnesses in the study (1 HIV+/FIL+ subject developed malaria at 3.5 months, another HIV+/FIL- subject developed tuberculous meningitis at 4 months with leg weakness as a sequelae, and a third HIV+/FIL- subject had pulmonary tuberculosis and underwent a hysterectomy for an unrelated problem. One death occurred at 4.5 months in a HIV+/FIL+ subject who had been diagnosed and treated for pulmonary tuberculosis and cryptococcal meningitis prior to starting the study, and who had also been on ART for 1 year before enrollment. His death was attributed to a recurrence of the cryptococcal meningitis and possible tuberculoma. In addition to the death, 7 HIV+FIL+ subjects were lost to follow-up as were 2 HIV+/FIL- controls (Fig. 1), and additionally some subjects missed some viral load measurements at intermediate time points. No other subjects developed significant opportunistic infections while on treatment, although several had histories of tuberculosis or other infections prior to enrollment that had been treated.
Following DEC/Alb, subjects were followed at 1, 3, 6, and 12 months following drug administration. At one year, the planned primary endpoint, we have VL measurements on 20 (HIV+/FIL+) and 49 (HIV+/FIL-) subjects. One of the HIV+/FIL+ died. There was no significant difference between the two groups (HIV+/FIL+ and HIV+/FIL-) in the percent of baseline viral load at one year (counting the subject that died as ranked highest for response, and the 2 subjects that got ART for part of the time as having ART) by stratified WMW test, stratified by ART (p = 0.41) (Fig. 2, Fig. 3). To estimate an effect of HIV+/FIL+, we ran a linear model on the log percent change. We find that, controlling for ART, on average those with HIV+/FIL+ have about 0.97 times the percent change of viral load at one year than those with HIV+/FIL- (95% CI 0.88, 1.07). Repeating the analysis at different time points post baseline, we get similar results. We get similar results in all these models if we additionally control for baseline hemoglobin, baseline CD4 count and baseline viral load, or if the 2 subjects who started ART after baseline are treated as not on ART.
When we examined the subset of subjects (n = 27) who were not on antiretrovirals (9 HIV+/FIL+ subjects, 18 HIV+/FIL- subjects), we found that at 1 month following DEC/Alb, there was a significant difference in the percent change of the HIV log10 viral load between the 2 groups with the HIV+/FIL+ subjects showing a GM increase of 5.1% [95%CI −3.7 to 13.6%] compared to the HIV+/FIL- subjects who showed a change of −2.2% [95% CI −6.3 to 2.0%]; p = 0.05 WMW. This difference in the changes in viral loads was not sustained throughout the remaining 3, 6 or 12 month time-points (at 12 months, the HIV+/FIL+ subjects not on antiretrovirals [n = 9] had a mean change of −1.0 logs [95% CI −5.2 to 3.2], and the HIV+/FIL- subjects (n = 22) had a mean change of 2.6 logs [95% CI −7.9 to 13.1] p = 0.2).
CD4 counts were assessed at 3, 6 and 12 months. Analogous to the analysis with viral load, we use the linear model with log percent change in CD4 count, and express the effects as how many times larger the average percent change is for the HIV+/FIL+ group than for the HIV+/FIL- group. We find that there are no significant differences between the groups: at 1 year (1.10; 95% CI 0.88, 1.38; p = 0.40) (Figs. 2 and 4), 3 months (0.99; 95% CI 0.82, 1.20; p = 0.92), and 6 months (1.05; 95% CI 0.89, 1.24; p = 0.55).
Hemoglobin was assessed at 1, 3, 6 and 12 months. Again we use the linear model with log percent change in hemoglobin, and express the effects as how many times larger the average percent change is for the HIV+/FIL+ group than for the HIV+/FIL- group. We find that there are significant differences between the groups at 1 year (1.08; 95% CI 1.02, 1.15; p = 0.01) (Fig. 4 C-E). We do not see those significant effects at other times: 1 month (1.00; 95% CI 0.96, 1.03; p = 0.78), 3 months (1.01; 95% CI 0.95, 1.07; p = 0.78), and 6 months (1.04; 95% CI 0.99, 1.08; p = 0.11).
To examine the efficacy of single dose DEC/Alb in HIV+/FIL+ patients, quantitation of circulating filarial antigen (CFA) levels was assessed at baseline and at 1 year in those for whom samples were available (n = 12). The GM CFA level in those who were positive at baseline was 1123 IU/ml (95% CI 590–2138); at 1 year, that level had decreased to 534 IU/ml (95% CI 235–1215) (p = 0.03 WSR) (Fig. 5). Only 2 subjects who were CFA positive at the start of the study became negative after 1 year. No CFA negative subjects became positive at the end of the study. Treated filaria positive subjects also saw a decrease in the GM filaria-specific IgG and IgG4. IgG decreased by 43% (from 40.4 μg/ml (CI 25–64) at baseline to 23.7 μg/ml (CI 17–32) at 1 year (p = 0.0008) and IgG4 decreased by 31%, although not statistically significant (p = 0.1514): from 0.65 ng/ml (CI 0.19–2.2) baseline to 0.52 ng/ml (CI 0.16–1.7) at one year (Fig. 5).
In this study, we attempted to look at the progression of HIV disease in patients with W. bancrofti/HIV co-infection compared with those with HIV alone after a one-time treatment with DEC/Alb, the standard therapy used worldwide (except in Africa) for mass drug administration programs. In addition, we also looked at biomarkers as surrogates for disease progression (CD4 count, viral load). We found no difference in the clinical outcomes of the subjects one year after treatment. We did find a transient and significant increase in viral loads in those Wb/HIV co-infected subjects not on ART after 1 month, though this difference in VLs between the Wb-infected and -uninfected HIV+ subjects equalized for the remainder of the study. When we examined CD4 counts at 3, 6 and 12 months, or viral loads and hemoglobin values at 1, 3, 6, and 12 months we found no difference between the 2 groups with the exception of higher hemoglobin values in the co-infected group at 1 year. The protective effect of filarial infection on infection-related anemia is lent support by a study of filarial/malaria co-infection in Africa [22].
In the HIV+/FIL+ subjects, circulating filaria antigen levels and filaria-specific IgG and IgG4 were lower in at 1 year than at baseline (although this did not achieve significance for the IgG4), suggesting that the treatment was effective in decreasing filarial antigen load.
The present study has several limitations, the most significant being the small sample size. We found a much lower than expected rate of W. bancrofti positivity in the HIV population in and around Chennai, although our sample size calculations were based on our 2004 prevalence data in which we found that 9.5% of HIV positive subjects were filarial also antigen positive [16]. India, like other LF endemic countries of the world, is participating in the Global Program to Eliminate Lymphatic Filariasis (GPELF), and is distributing single dose diethylcarbamazine (DEC) and albendazole on a population-wide level in some endemic regions. In the period leading up to and during the study, there was a large drop in the prevalence of W. bancrofti infection in the population with little transmission occurring in the urban areas from which the patients were recruited. It is possible that the MDA programs have lowered the prevalence and incidence of lymphatic filariasis infection to levels much lower than we had seen previously.
At the time the study was planned, India was just starting to use ART. By the time we had enrolled all of our subjects and controls, >50% of them were on ARTs. While this was a distinct benefit to our subjects, it did make it difficult for us to see a difference in viral loads between subjects and controls since many of them were already optimally virally suppressed at baseline.
Finding a difference in the course of HIV through treatment of a concomitant helminth infection is quite difficult [13,15]. However, in one randomized blinded placebo controlled crossover study in Tanzania of 34 HIV+ individuals, there was a significant salutary effect on VL at 12 week in filarial-infected individuals at 12 weeks [14]. ARVs were not used, a potential difference between this study and the present study.
Other groups have looked at the interactions with HIV and other non-filarial helminth infections with mixed findings [4,5,10,11,23–28]. Two studies have shown that anthelmintics decreased HIV viral load in helminth-infected HIV+ adults [4,14] whereas others have shown just the opposite [5]. In most other studies there has been demonstrated little to no difference in viral load after anthelmintic treatment [6–8].
The effects of helminth coinfection on CD4 counts in HIV-coinfected subjects are also conflicting. In several studies in Africa, the presence of helminth infection has been shown to adversely affect CD4 counts [26,28,29], although other studies have shown either a protective effect by helminths on CD4 counts [24,30] or no effect whatsoever [7,24,31]. A study from Uganda showed no overall effect of empiric deworming in HIV positive patients, however female patients who received deworming had a greater increase in CD4 count at 1 year that was not sustained at 2–3 years. In addition, women had an increase in hemoglobin after deworming that was not seen in men [9]. Studies of HIV and filarial co-infection in the future will be increasingly difficult as the prevalence of filarial infections decrease and more patients with HIV infection receive ART. Future studies trying to examine this issue will be limited to areas of high prevalence of both infections or as sub-studies within large cohorts in order to achieve a sample size that can address some of the remaining questions.
In conclusion, our study has clearly demonstrated that W. bancrofti infection, the most prevalent filarial infection of humans, failed to influence HIV-related clinical status, VL or CD4 counts at baseline and following treatment with DEC/ALB though there was an apparent filarial-induced increase in VL after DEC/ALB in ART-naive coinfected individuals. Our data suggest that like with vaccinations in HIV-infected individuals that the benefits of MDA in W. bancrofti and other related infections far outweigh a miniscule transient risk in parasite/HIV coinfections.
|
10.1371/journal.ppat.1007063 | Macrophages protect Talaromyces marneffei conidia from myeloperoxidase-dependent neutrophil fungicidal activity during infection establishment in vivo | Neutrophils and macrophages provide the first line of cellular defence against pathogens once physical barriers are breached, but can play very different roles for each specific pathogen. This is particularly so for fungal pathogens, which can occupy several niches in the host. We developed an infection model of talaromycosis in zebrafish embryos with the thermally-dimorphic intracellular fungal pathogen Talaromyces marneffei and used it to define different roles of neutrophils and macrophages in infection establishment. This system models opportunistic human infection prevalent in HIV-infected patients, as zebrafish embryos have intact innate immunity but, like HIV-infected talaromycosis patients, lack a functional adaptive immune system. Importantly, this new talaromycosis model permits thermal shifts not possible in mammalian models, which we show does not significantly impact on leukocyte migration, phagocytosis and function in an established Aspergillus fumigatus model. Furthermore, the optical transparency of zebrafish embryos facilitates imaging of leukocyte/pathogen interactions in vivo. Following parenteral inoculation, T. marneffei conidia were phagocytosed by both neutrophils and macrophages. Within these different leukocytes, intracellular fungal form varied, indicating that triggers in the intracellular milieu can override thermal morphological determinants. As in human talaromycosis, conidia were predominantly phagocytosed by macrophages rather than neutrophils. Macrophages provided an intracellular niche that supported yeast morphology. Despite their minor role in T. marneffei conidial phagocytosis, neutrophil numbers increased during infection from a protective CSF3-dependent granulopoietic response. By perturbing the relative abundance of neutrophils and macrophages during conidial inoculation, we demonstrate that the macrophage intracellular niche favours infection establishment by protecting conidia from a myeloperoxidase-dependent neutrophil fungicidal activity. These studies provide a new in vivo model of talaromycosis with several advantages over previous models. Our findings demonstrate that limiting T. marneffei’s opportunity for macrophage parasitism and thereby enhancing this pathogen’s exposure to effective neutrophil fungicidal mechanisms may represent a novel host-directed therapeutic opportunity.
| For people with compromised immune systems, such as those suffering from AIDS, fungal infections are difficult to treat and often deadly. Different fungal species have different ways of avoiding destruction by the immune system during infection. For Talaromyces marneffei, the ability to infect host macrophages and replicate within them as yeast is thought to be key to their survival and spread throughout the human body. Here, we use a larval zebrafish infection model to study interactions between T. marneffei and cells of the innate immune system in greater detail than previously possible. We show that during early infection, T. marneffei spores are taken up primarily by macrophages. Limiting access to this macrophage niche enhanced engulfment and destruction of spores by neutrophils, another innate immune cell type important in host defences. This neutrophil antifungal activity was reduced in animals lacking Myeloperoxidase, an abundant antimicrobial enzyme of neutrophil granules, proving that myeloperoxidase is crucial for host defence against T. marneffei. These studies suggest that blocking access of infective T. marneffei spores to their macrophage niche, thereby exposing them to neutrophil fungicidal activity, may be therapeutically effective in T. marneffei infection.
| Pathogenic fungal infections represent an important but widely overlooked human disease burden [1]. Invasive fungal infections, primarily affecting immunocompromised individuals, carry a high rate of mortality, despite the availability of antifungal drugs [2]. A uniting biological feature of a number of pathogenic fungi is dimorphism: modulation of morphological form in response to environmental cues. Most dimorphic fungal pathogens, such as Blastomyces dermatitidis, Histoplasma capsulatum, Paracoccidioides brasiliensis and Talaromyces marneffei (formerly Penicillium marneffei), exist in the environment in hyphal form, but convert to yeast growth during human infection [3–6]. Conversely, some dimorphic fungal pathogens, such as Candida albicans, exist as commensal yeasts, but under permissive conditions can cause invasive infection upon extension of germ tubes and subsequent hyphal growth [7]. Other common opportunistic fungal pathogens including Cryptococcus neoformans, which causes disseminated intracellular yeast infection leading to meningoencephalitis [8], and Aspergillus fumigatus, which causes serious pulmonary infections in immunocompromised patients [9] maintain a single morphological state despite having the capacity to change under certain circumstances, such as mating or development [10, 11].
Professional phagocytes of the innate immune system (neutrophils and macrophages) provide the first line of defence against fungal infection [12]. For T. marneffei, initial interactions are characterised by phagocytosis of conidia by leukocytes in the lung, followed by leukocyte-facilitated hematogenous dissemination [13]. For Cryptococcus, which infects as a yeast, macrophages also may play a role in pathogen dissemination [14]. Many fungal pathogens, such as H. capsulatum and T. marneffei, proliferate within macrophages as yeast [15–17], while invasive hyphae formed by A. fumigatus cannot be phagocytosed and elicit a neutrophil-dominated response, including generation of reactive oxygen species and formation of neutrophil extracellular traps (NETs)[18]. Although much has been learnt from mammalian infection models regarding disease progression, these models are inherently limited with regards to observing host-pathogen interactions in vivo and in real-time.
Zebrafish embryos and larvae provide an excellent platform for high-content imaging of early host-pathogen interactions, especially since the generation of transgenic strains labelling neutrophils [19–21] and macrophages [22–24]. The zebrafish toolbox has proven useful for modelling bacterial infection [25], particularly tuberculosis [26, 27]. Zebrafish have been utilised for modelling and imaging infections with the human fungal pathogens Candida spp. [28–31], Aspergillus spp. [32–35], Cryptococcus spp. [36–38], and Mucor spp. [39]. Such studies have provided significant new insights into the molecular and cell biology of host-pathogen interaction during these infections.
Here we present a new in vivo zebrafish model of talaromycosis (formerly called penicilliosis) that is caused by T. marneffei, a thermally dimorphic, opportunistic pathogen of humans. T. marneffei is capable of switching between a saprophytic hyphal growth form and a pathogenic yeast form in response to temperature and host cues including factors such as pH, salt concentration, calcium signalling and iron availability [40–43]. In the host, it primarily occupies the phagocyte niche but extracellular fungal cells are also evident, presumably due to host cell death.
For these studies, we have exploited the ectothermic nature of the zebrafish host. Although zebrafish are customarily held at 28°C in the laboratory, their normal development is documented for temperatures up to 33°C [44], and in the wild populations are found at temperatures up to 38.6°C [45]. We model an invasive hyphal form of T. marneffei infection unique to this zebrafish model (at 28°C) and a disseminated intracellular yeast form of infection resembling the human disease (at 33°C). These studies show that thermal dimorphism can be overridden by intracellular cues. We demonstrate a protective, G-CSFR-dependent expansion of neutrophils during protracted infection. Our studies particularly focused on the initial period of infection establishment, when fungal spores first encounter host leukocytes. These studies show that macrophages provide a protective niche for fungal conidia during infection establishment, whereas neutrophils exhibit a strongly fungicidal activity towards conidia that is myeloperoxidase dependent.
To optimise modelling of talaromycosis in zebrafish larvae, a variety of pathogen inoculation approaches and larval culture conditions were tested. At 28°C, immersion of 24 hpf embryos in T. marneffei conidia delivered within the chorion did not result in infection (0% (0/55) infected and 2% (1/56) death after 24 hours of co-incubation). However, reproducible local infection was established by intramuscular or 4th ventricle injections, while systemic infection was initiated by intravenous inoculation into the Duct of Cuvier at 52 hpf. Histology demonstrated fungal conidia in close proximity to vascular walls immediately following inoculation (Fig 1A), their immediate phagocytosis by leukocytes (Fig 1B) and their germination within 1 day post infection (dpi), including within intravascular leukocytes (Fig 1C). From these initial dose-finding studies, it was established that intravenous inoculations of 100–150 viable conidia established a systemic infection that resulted in <25% mortality during the course of the experiment, with stable fungal colony-forming unit (CFU) counts until 4 dpi, when CFU counts declined (Fig 1D and 1E). This indicates that for infective challenges in this dose range, the zebrafish innate immune system alone is capable of controlling a T. marneffei infection. As expected, considering the thermal dimorphism of T. marneffei [3], during the later days of infection extension of fungal germ tubes within phagosomes elongated some leukocytes (Fig 1C and S1D Fig). In some embryos, there was also invasive filamentous growth within various tissues (S1C and S1H Fig), including the brain (S1E Fig). By 3–4 dpi, accumulation of leukocytes at infection foci resulted in abscess and/or a tight collection of cells reminiscent of the “early granuloma” observed in larval M. marinum infection [46] (S1C and S1F–S1H Fig). Although this form of invasive filamentous infection digressed from the predominantly yeast-form infection observed in human talaromycosis, it did confirm that in vivo infection per se did not provide sufficient cues for T. marneffei to switch to its yeast form.
To model infection with yeast morphology T. marneffei, as observed in human disease [47], infection was also established at 33°C [3] (Fig 2). This is a temperature which zebrafish tolerate well [26, 29, 44, 48–51], and at which the yeast morphological switch occurs in vitro [40]. Furthermore, it reflects the temperatures of human torso and extremity skin (31–34°C) [52–54], a tissue characteristically involved in disseminated talaromycosis and where intracellular yeast forms are readily found on biopsy [55, 56].
Despite an extensive literature of zebrafish experimentation at 33°C, we specifically verified experimentally that zebrafish phagocyte function is intact at 33°C. We conducted comparative experiments at 28°C and 33°C with Aspergillus fumigatus, a fungal pathogen that is not thermally dimorphic, and has been previously studied in zebrafish at 28°C [32–35]. We compared four parameters of phagocyte function at the two temperatures: (1) the initial migratory response of neutrophils and macrophages to the site of conidial inoculation (S2A Fig); (2) the initial phagocytic response of neutrophils and macrophages to A. fumigatus conidia following their arrival at the site of conidial inoculation (S2B Fig); (3) the myelopoietic response to inoculation with A. fumigatus conidia over a 4-day period (S3 Fig); (4) the impact of morpholino-induced perturbations of leukocyte abundance on the germination of A. fumigatus 24 hpi (S4 Fig). Furthermore, for scenarios (1–3), we conducted these experiments with both live conidia and dead conidia (killed by γ-irradiation) to address the possibility that, had any variation been observed between temperatures, that this might be due to different rates of conidial germination and/or proliferation at the two temperatures. For each of these 13 experimental scenarios, no consistent significant difference was observed in any endpoint (S2–S4 Figs). 7/53 pairwise comparisons testing the null hypothesis “that there was no temperature-dependent difference in the endpoint” generated a p-value (corrected for multiple comparisons) that was <0.05. We therefore concluded that phagocyte numbers and function were intact at 33°C and that experiments conducted at 33°C would make a valid contribution to an exploration of the role of phagocyte function in the pathogenesis of T. marneffei infection establishment.
For infections that proceeded at 33°C, histology again demonstrated leukocyte conidia phagocytosis (Fig 2Ai) and T. marneffei growth within leukocytes and tissues (Fig 2Aii–2Aiv). Cytospun leukocyte preparations at 4 dpi of 33°C infections demonstrated germinated, proliferating fungal cells within leukocytes. Strikingly, at this temperature, T. marneffei assumed an elongated, septate filamentous form within all observed neutrophils, whereas within macrophages, fission yeast morphology (recognized as characteristic elongated forms with a medial septum [57]) predominated (7/9 macrophages) (Fig 2B). We confirmed that also in mammalian macrophages at 33°C, T. marneffei assumed this characteristic fission yeast morphology (S5 Fig). Collectively, these morphologically-based observations indicate that the different intracellular milieu of each phagocyte type can differentially influence T. marneffei form, and that at 33°C, the macrophage milieu favours the transition to the yeast form.
At 33°C, despite a vigorous granulopoietic response to infection (Fig 2E), fungal CFU numbers did not decrease at 4 dpi (Fig 2C), and death from infection was increased (Fig 2D), suggesting enhanced fungal viability at this temperature.
To facilitate live imaging of host-pathogen interactions, infection was performed using larvae expressing neutrophil- and/or macrophage-specific fluorescent transgenes. Quantification of leukocyte numbers over 4 days of infection at 28°C revealed a dramatic increase from 2–4 dpi in both neutrophil and macrophage numbers in response to active infection (Fig 3A–3C). By 4 dpi, macrophage and neutrophil numbers had increased to levels 150% and 260% respectively of those in uninfected controls, while pathogen numbers decreased by 65% over the same period (Fig 1D), indicating that the vigorous myelopoietic response was a component of an effective host response that controlled the infective burden. A vigorous granulopoietic response was also seen over 4 days of infection at 33°C (Fig 2E).
Amplification of leukocyte numbers during zebrafish bacterial infection has been shown to depend on signalling through the Csf3-Csf3r pathway [58], while both Csf3 and Interleukin-6 are known to be important in the response to yeast infection in mammalian systems [59, 60]. To interrogate these pathways in the context of talaromycosis, Csf3 signaling was intercepted by knocking down the single chain of the homodimeric csf3 receptor (csf3r), and interleukin-6 signalling was intercepted by knockdown of each subunit of the heterodimeric interleukin 6 receptor (il6ra and gp130). To quantify granulopoiesis, infections were performed in Tg(mpx:EGFP) larvae at 28°C and quantified as previously described [61].
Csf3r knockdown significantly reduced baseline neutrophil numbers and also significantly reduced the increase in neutrophil population size at 4 dpi (Fig 3D). The relative increase in neutrophil abundance in both control and csf3r-knockdown infected embryos was 2.2-fold. Hence the significant difference in the absolute increase is likely to have in part reflected the significantly lower basal neutrophil abundance in csf3r-knockdown embryos. This difference was not due to a different pathogen burden following infection establishment in the face of lower basal neutrophil abundance, since csf3r knockdown did not alter fungal survival/proliferation as reflected by 24 hours post infection (hpi) CFU numbers (Fig 5B). By 4 dpi, the significantly reduced granulopoietic response of infected csfr3-knockdown embryos resulted in impaired survival from infection (Fig 3E).
In contrast, il6r subunit knockdown did not impair the infection-driven granulopoietic response (Fig 3D). The effect of il6r subunit knockdown on basal neutrophil population size was also modest.
Collectively, these data indicate that the large expansion of neutrophil numbers during sustained T. marneffei infection is dependent on an interaction between pathogen and host, which mounts a cytokine-driven granulopoietic response that is in part csf3r- but not il6r-dependent. Intact basal csfr3 signalling is required for effective, protective, host-defences to talaromycosis.
Since leukocyte/conidia interactions were a prominent histological feature of the initial host response to T. marneffei conidia inoculation, they were examined in detail at 28°C (Fig 4). For these experiments, we used intramuscular inoculations, as we were not testing hypotheses about the normal route of infection, but rather about the direct interaction between conidia and leukocytes. These studies were conducted at 28°C, as the conidial state of T. marneffei at the time of inoculation is not temperature dependent, and this is the temperature at which zebrafish leukocyte function is usually characterised.
We confirmed that leukocytes actually phagocytosed fungal conidia, rather than merely associated with them. Fig 4A shows a neutrophil with a calcofluor-stained, germinated conidium, proven by z-stack analysis to be within an intracellular vacuole. Fig 4B shows a GFP-low macrophage in the Tg(mpx:EGFP) line (as previously described [62]) containing multiple germinated conidia. Using reporter lines in which red and green fluorophore expression is driven in macrophages and neutrophils respectively, active interaction of both neutrophils and macrophages with calcofluor-stained conidia was documented in the initial stages of infection (Fig 4C and 4D and S1–S3 Movies), with apparent phagocytosis. Using new reporter lines in which reporter fluorophore expression was targeted to membranes by linkage to the CAAX prenylation signal, Z-stack profiling confirmed that conidia associated with leukocytes were intracellular and within membrane-lined phagosomes, as demonstrated by cross-sectional fluorescence intensity profiles (Fig 4E–4G).
Although both neutrophils and macrophages were capable of phagocytosing T. marneffei conidia, following both somitic (Fig 4G) and intravascular inoculation (S6 Fig), T. marneffei conidia were phagocytosed almost exclusively by macrophages (Fig 4H). To quantify the various leukocyte-pathogen interactions at the inoculation site, fluorescence signal colocalization were analyzed based on the different leukocyte-associated reporter fluorophores and blue fluorescence of calcofluor-labelled conidia. Analysis of the caudal hematopoietic tissue of Tg(mpx:EGFP/mpeg1:mCherry) embryos 2 hours following intravenous calcofluor-labelled conidia inoculation showed that, while the neutrophil:macrophage voxel ratio was 1:2.6 (S6A Fig), reflecting the relative population sizes of the two phagocyte types, the ratio of conidia associated with these neutrophils and macrophages was 1:60 (S6B Fig).
In contrast, following intramuscular inoculation, although T. marneffei conidia were still preferentially phagocytosed by macrophages, neutrophils also actively engaged in conidial phagocytosis (Fig 4Hi). Hence, although both neutrophils and macrophages phagocytose T. marneffei conidia during the initial stages of infection, macrophage phagocytosis predominates.
Despite a common function as professional phagocytes, neutrophils and macrophages utilize distinctive antimicrobial arsenals [63], and hence display diverse antimicrobial properties that are often microbe-specific. Given the specific effects of the neutrophil and macrophage intracellular milieu on T. marneffei form at later stages of infection (Fig 2B), we hypothesized that neutrophils and macrophages might also play divergent roles during infection establishment.
To determine the different roles of neutrophils and macrophages during the initial stages of T. marneffei infection, CFU counts were compared at 0 and 24 hpi in embryos at 28°C that had been experimentally manipulated to modulate the relative numbers of neutrophils and macrophages (Fig 5).
Viable fungal cell number did not change between 0 and 24 hpi in leukocyte-replete control embryos (109.7% of baseline on average at 24 hpi) (Fig 5Ai and 5B), indicating either that conidia are both non-proliferative and resistant to killing, or that fungal death and proliferation were balanced in this context.
In embryos depleted of both neutrophils and macrophages by spi1+csf3r-knockdown using antisense morpholino oligonucleotides [64], 24 hpi T. marneffei CFU counts were significantly reduced to 64.4±6.7% of baseline (Fig 5Aiii and 5B). This demonstrated both that the presence of leukocytes during infection establishment enhanced conidial survival and/or fungal proliferation, and indicated the existence of a leukocyte-independent fungicidal activity.
In csf3r morphant embryos, which are depleted of neutrophils with little effect on macrophages [65, 66], 24 hpi CFU counts were 101.8±10.3% of baseline (Fig 5Aii and 5B), indicating that the presence of normal numbers of macrophages alone is sufficient to protect T. marneffei conidia from the leukocyte-independent fungicidal activity observed in spi1+csf3r morphants.
To complement the phagocyte depletion experiments, the effects of expansion of neutrophil populations on T. marneffei viability was examined in scenarios of different macrophage abundance. As previously reported [67], overexpression of csf3b from mRNA injection relatively selectively increased neutrophil numbers. In our hands, it resulted in approximately twice as many neutrophils at 2 dpf (S7Ai and S7Aii Fig) compared to controls, but only an increase of 39% in macrophages (S7Aiii and S7Aiv Fig) [68]. No significant change in fungal viability was observed over the first 24 hpi for embryos overexpressing csf3b (S7B Fig), indicating that the slightly increased macrophage population present in these embryos is sufficient to provide for conidial viability, even when neutrophil populations are expanded.
Knockdown of irf8 expands neutrophil numbers, but at the expense of macrophage numbers [61, 69]. In irf8 morphants, CFU counts at 24 hpi were 37.1±6.7% of baseline, which was significantly lower than for both control animals and spi1+csf3r morphants (Fig 5B), pointing to a potent fungicidal activity of the expanded neutrophil population when availability of the protective macrophage niche is reduced.
Collectively, all four scenarios support the hypothesis that macrophages provide a protective niche for T. marneffei conidia during infection establishment, shielding conidia from both neutrophil-dependent and leukocyte-independent fungicidal mechanisms.
In agreement with the important role played by neutrophils in controlling infection at later timepoints (Fig 3E), neutrophils were found to be strikingly fungicidal. As neutrophils were clearly the more fungicidal leukocyte during establishment of T. marneffei infection, we examined potential mechanisms that might determine neutrophil response to this pathogen and its outcome.
Myeloperoxidase is an abundant neutrophil enzyme important for generating potent antimicrobial radicals, and is critical for defence against other fungal pathogens in mammals [70–72]. We therefore hypothesized that neutrophil-dependent fungicidal activity against T. marneffei was dependent on myeloperoxidase activity. We tested the requirement for myeloperoxidase (mpx) using an mpx-deficient zebrafish mutant that is neutrophil-replete but lacks enzymatic Mpx activity [73].
During the establishment phase of infection, there was no difference in T. marneffei CFU counts at 24 hpi between WT and mpx-/- larvae for infections at both 28°C and 33°C (Fig 5B and S8A Fig) despite equivalent neutrophil populations over this period (S8B Fig). However, the enhanced fungicidal activity of the expanded neutrophil population in irf8-MO embryos, which reduced CFU counts at 24 hpi, was lost in mpx-/- embryos (Fig 5B).
During sustained infection, mpx-/- embryos mounted a vigorous granulopoietic response which is >2.5-fold higher than that of WT embryos at 3 dpi (S8B Fig). However, despite this expanded neutrophil population, mpx-/- embryos carried a fungal CFU burden similar to WT embryos (S8A Fig).
Collectively, these data indicate that the fungicidal activity of neutrophils against T. marneffei is in part myeloperoxidase-dependent, and that this component of neutrophil antifungal activity becomes most critical when the neutrophil population is expanded and macrophage abundance is depleted.
During infection establishment, our results support a model in which T. marneffei preferentially parasitizes macrophages, which provide an intracellular niche that shields conidia from both neutrophil-dependent and neutrophil-independent anti-fungal mechanisms. We therefore hypothesized that limiting access to the macrophage niche would enhance conidial clearance during infection establishment.
To test this hypothesis, macrophages were ablated prior to infection at 28°C using the metronidazole-dependent nitroreductase system [74, 75]. Metronidazole treatment of Tg(mpeg1:Gal4FF/UAS-E1b:Eco.nfsB-mCherry) embryos reduced macrophage abundance by 70% (Fig 6A, S9A Fig and S4 Movie), which resulted in a significant (35%) reduction of T. marneffei CFU counts at 24 hpi, compared to control sibling embryos that were metronidazole-treated but did not carry the UAS-E1b:Eco.nfsB-mCherry transgene (Fig 6B).
Conversely, selective ablation of neutrophils by metronidazole treatment of Tg(mpx:KalTA4/UAS-E1b:Eco.nfsB-mCherry) embryos resulted in a 20% increase in fungal viability compared to controls (Fig 6C and 6D, S9B and S9C Fig), further supporting our findings that neutrophils exhibit fungicidal properties during infection establishment.
Taken together, these results suggest that limiting access to the macrophage intracellular niche may be one targetable pathway for restricting establishment of infection for the benefit of the host.
This new model of Talaromyces marneffei infection in larval zebrafish holds many advantages over current murine in vivo models [76, 77]. It combines a replete vertebrate innate immune system with transgenically labelled lineages and optical transparency to facilitate detailed live imaging of host-pathogen interactions. Additionally, the ectothermic biology of zebrafish allows experimental separation of thermal and host-dependent influences on fungal morphology. While this can also be achieved using the recently developed Galleria mellonella and Caenorhabditis elegans models [78, 79], the invertebrate hemocyte provides only limited insight into complex innate immune responses such as those described here.
Combining an ectothermal host and a thermally dimorphic pathogen in an experimental system modelling a human infection provides both opportunity and raises some technical questions. Although zebrafish experiments are normally conducted at 28–28.5°C, zebrafish physiology is robust at 33°C. Evidence for this is: existence of wild D. rerio populations in waters up to 38.6°C [45]; normal embryological development over this temperature range [44]; experimental studies involving human xenotransplantation [49]; conditional temperature sensitive mutant alleles [26, 50, 51]; and even temperature shifts employed in previous infection modelling studies [26, 29]. We provide new experimental evidence showing that the myeloproliferative and functional phagocyte response to A. fumigatus, a non-thermally dimorphic fungal pathogen, is similar at 28°C and 33°C.
T. marneffei exhibits thermal dimorphism, but even so, the fungal to yeast shift is not absolute at 37°C. At 37°C in vitro, the rate of hyphal conversion to yeast varies from 40% to 90% between the lowest and highest tolerated pHs, and NaCl concentrations ≥4% suppress hyphal to yeast conversion almost completely [40]. Even in human infection at 37°C, extracellular forms characteristically assume an elongated shape [47, 80]. As is uniquely possible in an ectothermic model, we exploited the ectothermic biology of the zebrafish host to identify a divergence between fungal morphology and temperature, specifically in response to the leukocyte intracellular milieu. The phase transition to a yeast morphology during human T. marneffei infection at 37°C is widely regarded to be triggered primarily by the ambient temperature, with the thermal dimorphism of the pathogen seen as an evolved pathogenic trait [42]. At 33°C, we observed filamentous forms in neutrophils and tissues (Fig 2Aii, 2Aiv, 2Bi and 2Bii), and predominantly yeast forms within macrophages (Fig 2Biii and 2Biv), providing strong evidence that the intracellular milieu of the phagocytosing leukocyte type may be a determinant of T. marneffei form in vivo that can override the influence of ambient temperature. We therefore hypothesize that one factor contributing to the macrophage being the preferred infectious niche for T. marneffei is its temperature-independent support of a phase transition to the more pathogenic yeast form.
In response to prolonged infection, we observed expansion of leukocyte populations, particularly neutrophils, and demonstrated that this protective response was at least in part dependent on signalling through G-CSFR. This demand-driven hematopoietic response to infection appears to be conserved, and has been since reported by others in response to Salmonella enterica infection [58]. Overexpression of the zebrafish Csf3r receptor ligand Csf3b by mRNA overexpression phenocopied the granulopoietic response observed during T. marneffei infection, further supporting our findings.
Genetic manipulation of leukocyte lineage numbers prior to infection demonstrated that macrophages provide a protective niche for fungal conidia, while neutrophils are fungicidal. Further, using an existing mutant, we demonstrated that the fungicidal activity of neutrophils was myeloperoxidase-dependent. Myeloperoxidase deficiency is a prevalent congenital human disorder of neutrophils with an incidence of 1:2000 [81]. Epidemiological studies associate myeloperoxidase-deficiency with increased risk of fungal infection [82], myeloperoxidase-deficient mice are vulnerable to Candida infection [83], and myeloperoxidase-deficient neutrophils display reduced ability to produce fungicidal neutrophil extracellular traps (NETs) [84]. Our demonstration of myeloperoxidase deficiency as a talaromycosis disease-enhancer in zebrafish suggests it may be an unrecognized disease-modifier of human talaromycosis and invites an evaluation of myeloperoxidase status in such patients.
Selective depletion of macrophages using the metronidazole system demonstrated that removal of this protective niche exposes T. marneffei conidia to neutrophil-dependent and independent antifungal mechanisms. Temporary depletion of macrophages in patients that transiently selectively reduce access of conidia to the macrophage niche might provide a novel therapeutic strategy to restrict infection establishment. We also hypothesize that transient macrophage ablation therapy in established talaromycosis may facilitate talaromyces infection eradication by exposing the organism to fungicidal neutrophils.
One limitation of this zebrafish model is that the infection is established by inoculation, rather than via alveolar macrophages which is presumed to be that natural route of infection [85]. However, in human T. marneffei infection, fungal forms are found in tissue macrophages throughout the body, including particularly in skin [56], lymph nodes [86] and bone marrow [87], and so our observations of the behaviour of tissue macrophages provide relevant insights to understanding the pathogenesis of the human disease.
Findings from this new in vivo zebrafish T. marneffei infection model have implications not only for talaromycosis but also for the pathogenesis of infections with other fungal and dimorphic pathogens including histoplasmosis, blastomycosis and coccidioidomycosis. Furthermore, it is a unique resource for exploring in detail the in vivo cell biology of leukocyte-pathogen interactions during this infection.
Zebrafish strains were: wildtype (AB*); durif(mpx-/-)gl8 [73]; Tg(mpx:EGFP)i113 [19]; Tg(mpeg1:Gal4FF)gl25 [22]; Tg(mpeg1:mCherry)gl23 [22]; Tg(mpx:Kal4TA4)gl28 [88]; Tg(UAS-E1b:Eco.nfsB-mCherry)c264 (Zebrafish International Stock Centre, Eugene, OR). The new Tg(mpeg1:mCherryCAAX)gl26 and Tg(mpx:EGFPCAAX)gl27 lines were generated using Multisite Gateway cloning (Invitrogen) in combination with 1.87 kb of mpeg1 promoter [22] and 8.35 kb of mpx promoter [20]. Fish were held in the Walter and Eliza Hall Institute and FishCore (Monash University) aquaria using standard practices. Because zebrafish exhibit juvenile hermaphroditism, gender balance in embryonic and larval experiments was not a consideration [89]. Embryos were held in egg water (0.06 g/L salt (Red Sea, Sydney, Australia)) or E3 medium (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4, equilibrated to pH 7.0); from 12 hpf, 0.003% 1-phenyl-2-thiourea (Sigma-Aldrich) was added.
Animal experiments followed appropriate NHMRC guidelines and were conducted under protocols approved by Ethics Committees of the Walter and Eliza Hall Institute (2007.004 and 2009.031) and Monash University (MAS/2010/18). In accordance with the approved protocol requirements, all zebrafish embryos and larvae used in experiments were younger than 7 dpf (i.e. experiments were concluded on the 6th dpf which was the 4th day after infection). We performed the experiments under Institution Biosafety Committee Notifiable Low Risk Dealing (NLRD) approvals 2007.01 (Walter and Eliza Hall Institute) and PC2-N23-10 (Monash University). T. marneffei was assigned to Risk Group 2 at the time these approvals were granted. In most jurisdictions, including endemic regions, T. marneffei is a risk group 2 organism.
T. marneffei strains, derived from the FRR2161 type strain, were: acuD:RFP strain, which expresses RFP on germination and the control strain SPM4 [90]. To prepare cells for injection, T. marneffei conidia were inoculated onto Sabouraud Dextrose (SD) medium and cultured at 25°C for 10–12 days when the cultures were conidiating. Conidia were washed from the plate with 0.005% Tween 80 solution, filtered, sedimented (6000 rpm, 10 min), resuspended in dH2O and stored at 4°C. For inoculation, conidia were resedimented and resuspended in PBS. Heat-inactivation and calcofluor staining was as described previously [22]. Calcofluor staining did not affect conidial viability, as evidenced by their subsequent germination in vivo (S1A Fig).
T. marneffei colony forming unit (CFUs) numbers per embryo were determined by thorough homogenization of individual embryos in 500 μL of dH2O using a Dounce homogenizer. 250 μL of homogenate was cultured on SD medium agar with 1% ampicillin for 3–4 days at 37°C. Plates were then incubated overnight at room temperature, and colonies that underwent yeast to hyphal morphological switching were scored as T. marneffei colonies.
For inoculation, 52 hpf tricaine-anesthetized embryos were mounted on an agar mould with head/yolk within the well and tail laid flat on the agar. The T. marneffei conidial suspension was inoculated using a standard microinjection apparatus and large-bore needle via the common cardinal vein for systemic infection, or the 4th ventricle or a somite aligned to the yolk extension tip for local infection [22, 91]. Inoculated embryos were held at 28°C or 33°C according experimental design. The delivered conidial dosage was determined by immediate CFU enumeration on a group of injected embryos. Following initial dose-finding experiments that established an intravascular inoculum of 100–150 CFU/embryo achieved <25% mortality for 28°C infections (Fig 1D and 1E), this was the target inoculum dose.
Freshly-prepared A. fumigatus conidia stocks (strains CEA10 and 295) for these experiments were stored at 4°C for < 2 months. Fresh aliquots were prepared for microinjection as described for T. marneffei spores, delivering a target inoculum of 50–150 live spores, verified by back-plating as described above for T. marneffei. To prepare dead A. fumigatus conidia for microinjection, they were γ-irradiated with 10kGy (delivered over 207 hr 27 min 10 sec) from the Monash University Gammacell 40 Exactor (Theratronics) with two Caesium-137 sources (dose selected based on [92]). Irradiated spores were verified as dead by plating and incubation for 5 days: no growth occurred. The dead spores were microinjected at the same dilution of stock as used for live spores. Irradiated spores still stained well with calcofluor. For intravascular delivery experiments, microinjection of conidia utilised polydimethylsiloxane (PDMS) microstructured surface arrays [93], while imaging was performed following mounting in PDMS imaging devices, as previously described [94].
Leukocyte numbers were determined by two techniques. For direct enumeration, manual counting of neutrophil numbers was assisted by the brush tool in Paintbrush 2.1.2 (Soggy Waffles), which records clicks to avoid duplicate counting. Alternatively, “Leukocyte Units” (LUs), a surrogate parameter proportional to leukocyte numbers determined by analysis of digital images, were computed as previously described and validated [61]. LUs incorporate an internally-controlled correction for cell size, and can be independently applied to the signal from fluorescent neutrophils and macrophages. Where appropriate, LUs are called “Neutrophil Units” or “Macrophage Units”. In some cases as indicated, leukocytes were enumerated in the tail region distal to the tip of the yolk extension, in order to score cells in a representative part of the whole animal where overlap and anatomical shape did not interfere with accurate scoring. Migrating phagocyte numbers and their phagocytosis of conidia were counted manually in reconstructed 3-dimensional imaged volumes using Imaris v5 (Bitplane).
Antisense morpholino oligonucleotides (MOs) were purchased from Gene Tools, LLC (Eugene, OR) (S1 Table). New MOs were demonstrated to target their intended sequence using EGFP reporter constructs engineered to contain target sequences (S10 Fig). MO-csf3rATG specificity was controlled by MO-csf3rsplice; only MO-csf3rATG data are presented. MO-il6raATG and MO-gp130ATG served as specificity controls for each other as they targeted separate components of the same heterodimeric receptor complex.
Tg(mpeg1:Gal4FF/UAS-E1b:Eco.nfsB-mCherry) or Tg(mpx:Kal4TA4/UAS-E1b:Eco.nfsB-mCherry) embryos generated by intercrossing were treated with 10 mM metronidazole (Sigma M3761) from 28–52 hpf. The efficiency of macrophage ablation assessed using live imaging (S4 Movie) and LU at 24 h after treatment was ~70% (S9A Fig), comparable with the experience of others [25]. The efficiency of neutrophil ablation at 24 h after treatment was ~60% of Sudan Black positive neutrophils (S9B and S9C Fig). In infection experiments, metronidazole treatment was continued throughout the infection time course to restrict leukocyte recovery.
Capped csf3b mRNA was transcribed from pCS2+ plasmids containing the cDNA [67] linearized by Not1-HF using the mMessage mMACHINE kit followed by RNA cleanup using RNeasy Mini Kit (Qiagen) according to manufacturer’s instructions. 5 μL was subjected to RNA electrophoresis for a quality check and stock aliquoted and stored at -70°C. 500–1000 pg of capped RNA was microinjected directly into the cell of 1-cell embryos. Control embryos received diluent alone.
J774 murine macrophages were seeded at a concentration of 1x105 conidia/mL into a 6 well microtitre tray containing sterile coverslips and 2 mL of DMEM medium. Macrophages were incubated at 33°C or 37°C for 24 hours followed by the addition of 0.1 μg/mL lipopolysaccharide (LPS) and incubation for a further 24 hours. The cells were washed in PBS and 2 mL of complete DMEM medium containing 1x106 conidia was added. A control lacking conidia was also performed. Macrophages were incubated for 2 hours at 37°C to allow conidia to be engulfed, washed once in PBS to remove non-phagocytosed conidia and incubated a further 24 hours at 33°C or 37°C. Macrophages were fixed in 4% paraformaldehyde and stained with 1 mg/mL fluorescent brightener 28 (calcofluor) to observe fungal cell walls. Mounted coverslips were examined using differential interference contrast and epifluorescence optics for cell wall staining and imaged on an Olympus IX70 microscope.
Standard 4% paraformaldehyde-fixed, paraffin-embedded sections were stained by hematoxylin and eosin or Grocott methanamine silver stains by the Walter and Eliza Hall Institute of Medical Research Histology Department. FACS-sorted leukocytes from infected Tg(mpx:EGFP) embryos were collected based on EGFP fluorescence and cytospun preparations stained with Grocott methanamine silver / Nuclear Fast Red.
Routine brightfield and fluorescence imaging used a Zeiss Lumar V12 stereo dissecting microscope with an AxioCam MRm camera running AxioVision 4.8 software. Images were 1388x1040 pixels. Compound microscopy used an upright Nikon Optiphot-2 microscope with 40x and 100x objectives and a Zeiss AxioCam MRc5 Camera running AxioVision AC (Release 4.5) software. Images were 1292x968 pixels. Confocal microscopy used a Zeiss LSM 5 Live with a Plan-Apochromat 20x, 0.8 NA objective. Software was Zen (Version 4.0). Images were 16-bit 512 x 512 pixels. Z-depth ranged from 0–90 slices with Z-intervals optimized for 1:1:1 X:Y:Z reconstruction. Time intervals were optimized for each experiment and ranged between 30–200 s. Excitatory laser wavelengths were 405 nm for calcofluor, 489 nm for EGFP and 561 nm for mCherry. Emission detection used a BP495-555 filter for calcofluor and EGFP emission and a LP575 filter for mCherry emission.
Image processing was performed using Fiji (ImageJA 1.45b) and Imaris v5 (Bitplane). Colocalization analysis was performed using the Coloc function in Imaris, with thresholds set such that internal colocalization of each channel corresponded to the cell-specific fluorescent signal. Figures were constructed using Adobe CS5 Photoshop and Illustrator.
Descriptive and analytical statistics were prepared in Prism 5.0c (GraphPad Software Inc). Unless otherwise stated, data are mean±SEM, with p-values generated from two-tailed unpaired t-tests.
|
10.1371/journal.ppat.1006101 | The Effector Cig57 Hijacks FCHO-Mediated Vesicular Trafficking to Facilitate Intracellular Replication of Coxiella burnetii | Coxiella burnetii is an intracellular bacterial pathogen that infects alveolar macrophages and replicates within a unique lysosome-derived vacuole. When Coxiella is trafficked to a host cell lysosome the essential Dot/Icm type IV secretion system is activated allowing over 130 bacterial effector proteins to be translocated into the host cytosol. This cohort of effectors is believed to manipulate host cell functions to facilitate Coxiella-containing vacuole (CCV) biogenesis and bacterial replication. Transposon mutagenesis has demonstrated that the Dot/Icm effector Cig57 is required for CCV development and intracellular replication of Coxiella. Here, we demonstrate a role for Cig57 in subverting clathrin-mediated traffic through its interaction with FCHO2, an accessory protein of clathrin coated pits. A yeast two-hybrid screen identified FCHO2 as a binding partner of Cig57 and this interaction was confirmed during infection using immunoprecipitation experiments. The interaction between Cig57 and FCHO2 is dependent on one of three endocytic sorting motif encoded by Cig57. Importantly, complementation analysis demonstrated that this endocytic sorting motif is required for full function of Cig57. Consistent with the intracellular growth defect in cig57-disrupted Coxiella, siRNA gene silencing of FCHO2 or clathrin (CLTC) inhibits Coxiella growth and CCV biogenesis. Clathrin is recruited to the replicative CCV in a manner that is dependent on the interaction between Cig57 and FCHO2. Creation of an FCHO2 knockout cell line confirmed the importance of this protein for CCV expansion, intracellular replication of Coxiella and clathrin recruitment to the CCV. Collectively, these results reveal Cig57 to be a significant virulence factor that co-opts clathrin-mediated trafficking, via interaction with FCHO2, to facilitate the biogenesis of the fusogenic Coxiella replicative vacuole and enable intracellular success of this human pathogen.
| Human Q fever is caused by the intracellular bacterium Coxiella burnetii. Successful infection of human cells relies on a Dot/Icm secretion system and the translocation of effector proteins into the host cell cytosol. The functions of many Coxiella effector proteins, and their contribution to bacterial growth and host manipulation, remain unknown. We show that a unique effector, Cig57, has an important role in manipulation of host cellular clathrin-mediated trafficking. In particular, Cig57 binds FCHO2, a protein involved in formation of clathrin-coated vesicles, in a manner that is dependent on a tyrosine-based endocytic sorting motif. Through engaging proteins in the clathrin pathway, Cig57 facilitates expansion of the Coxiella replicative vacuole and enables the pathogen to replicate to large numbers. Thus, we identify a relationship between a host process and a key virulence protein that are required for pathogen success.
| The intracellular bacterial pathogen Coxiella burnetii is the causative agent of human Q fever, a zoonotic disease with the potential to cause life-threatening complications. Transmission to humans occurs via inhalation of contaminated aerosols. Human infection can lead to an acute, pneumonia-like illness, or proceed to a chronic disease state in which endocarditis can manifest [1]. During natural infection, Coxiella predominantly invades alveolar macrophages, and in order to replicate intracellularly, a spacious and fusogenic lysosome-derived vacuole, termed the Coxiella-containing vacuole (CCV), is established by the pathogen. After internalization, Coxiella passively traffics through the endolysosomal pathway [2, 3]. The developing vacuole obtains markers typical of early and late endosomes, such as EEA1 and Rab7, and finally matures to a lysosome [4, 5]. Here, with an internal pH of approximately 4.8, and in the presence of proteolytic and degradative enzymes, Coxiella becomes metabolically active and will direct the expansion of the CCV before replicating to large numbers [6]. The active form of Coxiella is the replicative large cell variant (LCV), distinct from the environmentally stable small cell variant (SCV) [7].
The exact requirements that render the CCV permissive for replication are unknown, however recent mutagenesis studies have demonstrated that a Dot/Icm type IVB secretion system is essential for CCV biogenesis and intracellular replication [8, 9]. This secretion system is activated by the lysosomal environment [10] and more than 130 Coxiella effector proteins are known to be translocated from the pathogen into the host cell [8, 11–18]. Multiple mutagenesis studies have identified a small cohort of Dot/Icm effectors that play important roles in CCV biogenesis and intracellular replication of Coxiella [14, 16, 19, 20]. However, the function of most of these effectors and why they are required for intracellular success of Coxiella remains to be elucidated.
Using HeLa cells as an important model for infection, various host cell vesicular trafficking pathways have been shown to facilitate CCV development and contribute to the infection cycle of Coxiella. The retromer trafficking process, required for retrograde transport from endosomes to the trans-Golgi network [21], has been shown to contribute to the maturation of the CCV with retromer subunits VPS35 and VPS29 and sorting nexins all required for expansion of the CCV [22]. In addition, v-SNAREs VAMP3, VAMP7 and VAMP8 are found on the CCV membrane and have been shown, via siRNA experiments, to aide fusion events with vesicles in the endolysosomal pathway [23]. The autophagic SNARE, syntaxin-17, is required for the homotypic fusion of CCVs [22, 23]. Indeed, perturbation of autophagy, through both enzymatic inhibition and siRNA treatment of key proteins required for autophagy, results in a multi-vacuole CCV phenotype [20]. Coxiella, does not appear to induce host cell autophagy but interaction of the CCV with autophagosomes is demonstrated by the accumulation of LC3 inside the CCV, and Rab27 on the vacuole membrane [20, 24]. This indicates that successful CCVs are most accurately described as autolysosomes [24]. Recently, host cell clathrin-mediated trafficking was also shown to be important for intracellular replication of Coxiella [25]. Clathrin is important for endocytosis as well as trafficking events within cells. Clathrin-mediated endocytosis is the process by which host cells internalize material, termed cargo, into clathrin-coated pits, to then be sorted to their subcellular compartments (For a review see [26]). Larson and colleagues showed that siRNA gene silencing of CLTC (clathrin) and AP2B1/AP2M1 (AP-2), but not AP1, significantly impeded the intracellular replication of Coxiella. The adaptor complex AP-2 acts at the plasma membrane during clathrin-mediated endocytosis, and both AP-1 and AP-3 facilitate clathrin-mediated trafficking from the trans-Golgi network [27]. These key proteins, required for endocytosis events at the plasma membrane, facilitate the normal growth of Coxiella [25]. Thus, the CCV interacts with multiple host vesicular trafficking processes for successful CCV expansion and intracellular replication.
These host pathways are likely manipulated and controlled by complex interactions with Coxiella Dot/Icm effector proteins. For example, initial studies linked the multi-vacuolar CCV phenotype seen in cig2::Tn mutants with the multi-vacuolar phenotype seen when silencing host autophagy components [20]. More recently, the effector Cig2 was found to bind phosphoinositide PI(3)P, and this was required for recruitment of autophagy machinery components to facilitate homotypic fusion of CCVs [24, 28]. Additionally, the effector CvpA was found to modulate the association of Coxiella with the clathrin trafficking pathway through interaction with AP-2. CvpA is required for intracellular replication of Coxiella and this is believed to be linked to the acquisition of endolysosomal lipids and proteins through subversion of the clathrin transport pathway.
Cig57 (CBU1751, a 48.8 kDa protein, 420 amino acids), initially identified as an effector because the promoter region contains a PmrA binding region that indicates that the gene is co-regulated with icm genes [29], is also required for intracellular replication of Coxiella [20]. A transposon-mutagenesis screen to identify novel factors that influence CCV morphology identified multiple transposon insertions, disrupting cig57, that caused a significant bacterial growth defect and small vacuole phenotype [20]. Hence this effector plays a vital role in establishing the Coxiella intracellular replicative niche. Similar to CvpA, Cig57 contains endocytic sorting motifs that mimic those recognised by adaptor protein complexes in the host cell. In Cig57 there are three endocytic sorting motifs, two dileucine motifs ([DERQ]xxxL[LI]), and one tyrosine motif (YxxΦ), where x represents any amino acid and Φ is a bulky hydrophobic residue [30, 31]. Adaptor protein complexes bind to these motifs usually found on transmembrane proteins such as the transferrin receptor (TfR), to facilitate their selection and uptake into clathrin-coated vesicles (For a review see [30]).
Herein, we examine the function of Cig57 by identifying and examining its interaction with the host protein FCHO2. FCHO2 (88.9 kDa, 810 amino acids) is part of the muniscin subfamily of the EFC domain (extended Fes-CIP4 homology) proteins. Muniscins act at the early stages of clathrin-mediated endocytosis and have been implicated in the initiation of clathrin-coated pits [32, 33]. Specifically, the N-terminal EFC domain of FCHO2 dimerizes and binds the inner plasma membrane, binding to PI(4,5)P2 enriched membranes [33, 34], to enable the curvature seen in the neck of clathrin-coated vesicles. We show FCHO2 is required for optimal CCV formation, and establish that Cig57, interacting with FCHO2 via a tyrosine-based endocytic sorting motif, subverts clathrin to the CCV to facilitate normal vacuole biogenesis and intracellular replication of Coxiella.
To identify potential host protein targets of Cig57, a yeast-two-hybrid (Y2H) assay was performed using the full length Cig57 as bait and a HeLa cDNA library as prey. Cig57 alone did not activate reporter gene expression, as Saccharomyces cerevisiae (Y2H Gold, Clontech) carrying pGBKT7-cig57 could not grow on quadruple dropout (QDO) yeast minimal media (YMM) plates (-Trp, -Leu, -His, -Ade). After screening, only one positive prey clone was identified from the cDNA library, and the insertion was sequenced. This clone encoded amino acids 1 to 433 of the human protein FER/CIP 4 homology only protein 2 (FCHO2). The interaction between Cig57 and FCHO2 1–433 was confirmed by re-transformation of both pGBKT7-cig57 and pGADT7-FCHO2 (1–433) into a different S. cerevisiae strain (AH109). S. cerevisiae harbouring both pGBKT7-cig57 and pGADT7-FCHO2 grew on both double-dropout (DDO, -Trp, -Leu) and QDO YMM agar plates (Fig 1A). No interaction was observed with either pGADT7 or pGBKT7 empty vectors. We next sought to confirm this interaction within mammalian cells using an immunoprecipitation method. Importantly, this was done in the context of infection to address whether the interaction between Cig57 and FCHO2 occurs during infection. HEK 293T cells were infected with Coxiella transposon mutant cig57::Tn expressing 3xFLAG-Cig57 from a plasmid or WT Coxiella as a negative control. These cells were then transfected to express either GFP or GFP-FCHO2. Lysates of infected and transfected cells were incubated with beads that bind GFP, and protein bound to the beads were probed with anti-GFP or anti-FLAG antibodies by Western blot. We detected FLAG-Cig57 in the immunoprecipitate of cells expressing GFP-FCHO2, but not in cells expressing GFP alone, validating the interaction between FCHO2 and Cig57 (Fig 1B). We next sought to determine the intracellular localization of Cig57. In WT infected cells, we transfected mCherry or mCherry-Cig57, and observed that while mCherry had diffuse localization in cells, mCherry-Cig57 is enriched at the CCV and in punctate structures in the cytoplasm (Fig 1C). Because of the interaction with FCHO2, we wanted to determine the relative localization of Cig57 and FCHO2 during infection. We engineered a HeLa cell line that constitutively expresses GFP-FCHO2, and transfected these cells with mCherry-Cig57. FCHO2 was not observed to be enriched at the CCV membrane, however FCHO2 and Cig57 co-localize at punctate structures in the cytoplasm (Fig 1D).
Endocytic sorting motifs are typically present on transmembrane receptor proteins, and are recognised by adaptor proteins for selection and sorting of cargo molecules for clathrin-mediated endocytosis [31, 35]. Cig57 contains three predicted endocytic sorting motifs, two of a dileucine type, and one of the tyrosine type. To evaluate the importance of these motifs in Cig57, we created an endocytic sorting motif mutant (ΔESM) construct with mutations in key residues of the predicted Cig57 endocytic sorting motifs (LI82,83AA, LL275,276AA, Y365A), (Cig57ΔESM), and evaluated whether this form of Cig57 could bind FCHO2. To investigate whether FCHO2 can recognise these motifs, we transformed yeast with pGADT7-FCHO2 (1–433) and pGBKT7-cig57 containing individual mutations in the endocytic sorting motifs, or mutations in combination with each other (Fig 2). We observed growth of all transformants on DDO plates, which indicates yeast viability and successful plasmid uptake, and found no growth on QDO where pGADT7-FCHO2 (1–433) was transformed alongside pGBKT7-cig57ΔESM, indicating that one or more of the endocytic sorting motifs are involved in the Cig57-FCHO2 interaction (Fig 2A, segment 8). Individual mutations were also assessed for binding to FCHO2, and we showed that while the dileucine motifs were not important for binding FCHO2, the tyrosine residue Y365, part of the endocytic sorting motif YRKF, is essential for binding to FCHO2 (Fig 2A, segment 2). Hence, we have been able to identify a Cig57 residue essential for this interaction and that FCHO2 might have the capacity to recognise tyrosine-based endocytic sorting motifs. Inability of the proteins to bind to each other was not due to lack of protein expression in the yeast, as all proteins were expressed from the pGAD (anti-HA) and pGBKT (anti-c myc) plasmids (Fig 2B).
Given the importance of the endocytic sorting motifs, particularly Y365, for binding FCHO2, we examined whether this mutated form of Cig57 could complement the lack of growth seen for the cig57::Tn strain. An intracellular growth curve was performed over five days (Fig 3A) which demonstrated the inability of cig57::Tn Coxiella expressing pFLAG-Cig57ΔESM or pFLAG-Cig57Y365A to grow to similar levels as WT Coxiella or cig57::Tn pFLAG-Cig57. Interestingly, the ΔESM and Y365A versions of Cig57 are not completely inactive, as growth is not decreased to the same level as the cig57::Tn mutant. This phenotype was also observed visually by quantifying and comparing the CCV areas at 72h post-infection from three independent experiments (Fig 3B). Using this measure, vacuole areas formed by the cig57::Tn pFLAG-Cig57ΔESM strain (28.5±4.6 μm2) are significantly smaller (P = 0.0001) than those produced by WT (146.8±6.7 μm2) and the complemented mutant (143.1±7.5 μm2) (P = 0.0002). Interestingly, the vacuole sizes of cig57::Tn pFLAG-Cig57Y365A (46.9±1.8 μm2) were significantly larger than cig57::Tn pFLAG-Cig57ΔESM (P = 0.020) which may suggest an additional role for the dileucine endocytic sorting motifs. Importantly, both Cig57Y365A and Cig57ΔESM shows significantly (P = 0.0001 and P = 0.02 respectively) larger vacuoles than cig57::Tn mutant (11.0±1.5 μm2), indicating that during expression of Cig57ΔESM, vacuoles are of an intermediate size (Fig 2C). Analysis of a representative experiment allowed us to visualize that indeed the distribution of vacuole sizes differed between the different strains (Fig 3D). Thus Cig57, and the Cig57 endocytic sorting motifs, particularly the Y365 residue, are important for both intracellular replication of Coxiella and expansion of the CCV. To illustrate that all of our Coxiella strains were equally as infective, we plotted raw genome equivalent (GE) values for Coxiella recovered at timepoint 0h (4 hours post infection), and show that there is no significant difference between the values obtained for each of the strains (Fig 3C).
The distribution of clathrin normally includes plasma membrane and cytosolic puncta throughout the cell [36]. During infection with Coxiella, clathrin has been observed around the CCV and this has been shown to require the AP-2 binding effector CvpA [25]. Given that Cig57 interacts with the clathrin related protein FCHO2 we sought to assess the contribution of Cig57 to the accumulation of clathrin on the CCV. HeLa cells were infected and stained for clathrin heavy chain 72 h after infection. Confocal micrographs showed an increased density of clathrin surrounding the CCV in cells infected with WT Coxiella and the cig57::Tn pFLAG-Cig57 strain (Fig 4A). Importantly, during infection with the cig57::Tn mutant strain, or the mutant strain complemented with either pFLAG-Cig57ΔESM or pFLAG-Cig57Y365A clathrin is no longer recruited to the CCV (Fig 4A). When the clathrin signal was quantified, the ratio of signal on the CCV compared to cytoplasmic signal was approximately 2 in WT (2.1±0.1) and the cig57::Tn pFLAG-Cig57 strain (2.0±0.1), which was significantly higher than during infection with the mutant (1.1±0.1), the ΔESM complemented mutant (1.3±0.05), or the Y365A complemented mutant (1.2±0.04) (Fig 4B). Datapoints from a representative experiment demonstrate the range of clathrin intensity ratios in a representative infection (Fig 4C). In order to establish whether clathrin recruitment to the CCV was linked to the size of the CCV we examined the relationship between the area of WT Coxiella CCVs and clathrin intensity around the CCV, and found no correlation (Fig 4D), with a R2 value of 0.029 and a non-significant slope (P = 0.199) deviation from zero. These data implicate Cig57, and its activity mediated by the endocytic sorting motifs, as required for clathrin recruitment to the CCV.
To further validate that clathrin is recruited to CCV membranes, we co-stained infected HeLa cells with LAMP1 and clathrin, and show that when LAMP1 signal is high on the vacuole membrane, clathrin intensity likewise increases (S1 Fig). However this is only true for WT CCVs. Clathrin on cig57::Tn vacuoles did not increase at the CCV membrane, denoted by high LAMP1 signal.
Given the CCV localization of clathrin, we next utilized our GFP-FCHO2 stable HeLa cell line to explore the localization of FCHO2. In uninfected cells, FCHO2 localized to the perinuclear region, and the plasma membrane. Likewise, in cells infected with WT Coxiella, FCHO2 localization did not change (Fig 5A). Since clathrin was discovered to be on the CCV membrane, we asked whether FCHO2 is also recruited to the CCV. As shown in Fig 5A, and as quantified in Fig 5B, we saw no significant increase in FCHO2 signal on the membrane of the WT CCV. This is despite there being an increased signal of FCHO2 within the region of the CCV, yet there was also an increased FCHO2 signal around the nucleus, thus implying that the FCHO2 signal is not specific to the CCV. To further illustrate that FCHO2 is not recruited to CCVs, we stained infected GFP-FCHO2 cells with clathrin or LAMP1, and note that while clathrin and LAMP1 intensity increases on the CCV membrane, FCHO2 signal does not (Fig 5C and 5D).
FCHO2 and clathrin are important for normal clathrin-mediated endocytosis. Without clathrin, clathrin-mediated trafficking is blocked. In the absence of FCHO2, endocytosis still progresses, yet clathrin-coated pits are abnormally arranged, with AP-2-positive structures appearing enlarged and clustered [37]. We assessed what the impact of disrupting these genes has on the intracellular replication and CCV formation of Coxiella. HeLa cells were transfected with siRNA against clathrin heavy chain (CLTC), FCHO2, or OnTarget Plus (OTP) non-targeting (OTP-NT). Cells were infected with Coxiella 2 days post-siRNA transfection (day 0 post-infection) and lysates were collected for immunoblotting to gauge the level of protein depletion at 0, 2, 4 and 6 days post infection (Fig 6A). The level of knockdown achieved was quantified by measuring band intensities compared to that of the β-actin band intensity. Coxiella replication was measured by collecting cell lysates at days 2, 4 and 6 post-infection, and performing qPCR analysis against the C. burnetii ompA gene to calculate GE. As expected, and described previously, the depletion of cellular clathrin significantly inhibits Coxiella growth (Fig 6B and [25]). We compared the level of bacterial growth during silencing of FCHO2 also, and though not statistically significant (P = 0.28), there is a trend towards less bacterial growth when the FCHO2 gene is silenced (Fig 6B). At day 4 post-infection, samples were fixed for immunofluorescence analysis (Fig 6C) to examine vacuole size during depletion of these key transcripts in clathrin-mediated endocytosis. The area of individual CCVs was quantified and plotted in Fig 6D, in which we show significantly smaller vacuoles when silencing either clathrin (43.4±3.9 μm2, P = 0.0005) or FCHO2 (154.5±22.5 μm2, P = 0.012), compared to non-targeting conditions (303.4±25.5 μm2). Individual datapoints from one representative experiment are plotted in Fig 6E, and show that during silencing of clathrin, vacuoles are uniformly small, and that there is a shift towards smaller vacuoles in FCHO2 silenced cells as noted by the population of smaller CCV areas. These data indicate that FCHO2, as well as clathrin, is required for normal CCV biogenesis. While measuring vacuole sizes, we noted a multi-vacuolar phenotype during silencing of clathrin. Approximately 50% of cells displayed more than one CCV per cell during treatment with CLTC siRNA, compared to approximately 15% in non-targeting OTP siRNA (Fig 6F).
Uptake of Coxiella is not affected by silencing the clathrin pathway [25]. To corroborate this observation, we evaluated whether the disruption of the clathrin pathway affects the early stages of Coxiella infection. HeLa cells were treated with siRNA against CLTC or FCHO2, and infected for four hours, at which time the cells were fixed and differentially stained for intracellular and extracellular bacteria. There was no difference in the number of intracellular bacteria recovered in either the OTP-NT or silencing conditions (S2 Fig). This indicates that the entry of Coxiella is not disrupted in the absence of clathrin or FCHO2.
Using siRNA, clathrin was efficiently depleted from the cells until Day 6 however silencing of FCHO2 at days 4 and 6 post infection was of poor efficiency (Fig 6). This may contribute to the lack of a significant inhibition of Coxiella replication and a partial defect in CCV expansion (Fig 6). To overcome the problem of inefficient removal of FCHO2, a HeLa cell line was created which is completely devoid of FCHO2. Using the CRISPR-Cas9 genome editing system, HeLa cells were co-transfected with constructs targeting exon 1 and exon 5 of the FCHO2 gene, resulting in stable loss of protein production and a FCHO2 knockout (KO) cell line (Fig 7A). These cells, alongside the HeLa parent cell line, were infected with Coxiella for 5 days, and the Coxiella GE were measured by ompA qPCR (Fig 7B). There is a shift towards lower Coxiella replication at day 5 post infection in the FCHO2 KO cells compared to the wild-type parent HeLa cell line. At day 3 post-infection, samples were fixed and stained for Coxiella and LAMP1 to visualize CCV size (Fig 7C). When quantified, vacuoles are significantly (P = 0.0059) smaller during infection of our FCHO2 KO cell line (73.9±3.5 μm2) than when infecting the parental HeLa cell line (251.1±32.9 μm2) (Fig 7D). We plotted individual vacuole sizes for one of the three experiments (Fig 7E), to show the distribution of CCV sizes.
We next asked the question whether clathrin recruitment to the CCV was still able to progress as previously observed in HeLa cells in our FCHO2 KO cells. Using WT Coxiella, we infected parental HeLa cells and our FCHO2 KO cell line for three days, and stained for clathrin. As expected, clathrin was found to surround WT vacuoles in HeLa cells, however we observed a lower proportion of FCHO2 KO cells harboured vacuoles that were positive for clathrin (Fig 8A). Again, we measured LAMP1 intensity at a cross section of the vacuole, and show that clathrin increases at the CCV membrane, corresponding to high LAMP1 signal, on HeLa parent CCVs, CCVs formed in FCHO2 KO cells no longer show an increased clathrin intensity at the CCV membrane, where LAMP1 signal is increased (S1 Fig). As in Fig 3, clathrin intensity was approximately double the intensity at the CCV compared to the cytoplasm of wild-type HeLa cells (ratio of 2.2±0.1), however in FCHO2 KO cells, the ratio of clathrin intensity in the CCV compared to the cytoplasm was 1.3±0.3 (Fig 8B). It is to be noted that this phenotype is not as substantial a difference as observed for the CCV clathrin intensity during infection with the cig57::Tn mutant, in which the ratio was 1.1±0.1. Indeed, over one representative experiment, we observed a greater range of CCV/cytoplasm ratios during infection of the FCHO2 KO cells (Fig 8C).
The identification and characterization of Dot/Icm effector proteins in Coxiella is an active field of research that has been significantly bolstered by the recent advances in axenic culture and genetic manipulation of Coxiella. However, ascribing functions to these unique effectors remains challenging. Here, we have identified a host factor and pathway, namely FCHO2 and clathrin-mediated endocytosis, that are targeted by the effector protein Cig57. Clathrin-mediated endocytosis plays an essential role in all nucleated cells. Uptake and recycling of a variety of molecules, from plasma membrane receptors to iron is dependent upon the clathrin pathway. Additionally, clathrin is responsible for the trafficking of early endosomal vesicles to and from the trans-Golgi network playing an important role in delivering cargo proteins to their destination organelles [38, 39].
It is known that clathrin dependent trafficking is essential for the intracellular Coxiella lifecycle, as silencing of clathrin results in diminished intracellular Coxiella replication ([25] and validated in Fig 6). CvpA, an effector required for intracellular replication of Coxiella, was shown to bind AP-2, and now we have shown that another essential effector, Cig57 interacts with a different component of clathrin-coated vesicles, FCHO2 [25]. FCHO2 belongs to the muniscin family of proteins, which also includes FCHO1 and SGIP1 [40]. FCHO1 and 2 are thought to be involved in initiating clathrin-mediated endocytosis, though this is debated in the field as clathrin-mediated endocytosis still occurs in the host in the absence of FCHO1/2 albeit with abnormal morphology [33, 37, 41]. Nevertheless, FCHO2 arrives early to the site of clathrin-mediated endocytosis and aids in sculpting the plasma membrane to form the spherical clathrin-coated vesicles [32, 34]. These proteins contain an N-terminal EFC domain responsible for membrane binding, dimerization and induction of membrane curvature, alongside a linker region, followed by a C-terminal μ-homology domain. The μ-homology domain is an interaction hub, facilitating binding to the clathrin accessory proteins EPS15 and intersectin [33]. The interaction we observed using the yeast two hybrid system with FCHO2 was restricted to the N-terminal region (amino acids 1–433), indicating that Cig57 must bind within the EFC domain or the linker region. Binding the EFC domain may mean Cig57 is altering the membrane-binding capacity of FCHO2, by either blocking its ability to dimerize or bend membranes, or possibly even post-translationally modifying the protein in this region.
We have shown that FCHO2 is required for normal CCV biogenesis during infection with Coxiella. CCVs are significantly smaller in FCHO2 KO HeLa cells and there is a trend towards reduced replication of Coxiella in these cells. Additionally, the Coxiella growth defect in the absence of FCHO2 is not as severe as the growth defect in the absence of clathrin. Taken together, this indicates a level of dependence on the clathrin-mediated pathway for growth of Coxiella, where the extent of intracellular bacterial growth may be proportional to the amount of endocytosis exhibited.
FCHO1 is a close homologue of FCHO2. They share approximately 50% homology overall at the protein level. However, the amount of FCHO1 present in HeLa cells is very low, such that endogenous FCHO1 is undetectable by Western Blot analysis [32, 37, 41]. This study has not determined whether Cig57 has the same affinity for FCHO1 as it does FCHO2, but if it does, there is a potential for there to be a greater growth defect in the absence of both FCHO1 and FCHO2. We did not pursue this in our study as it has been previously shown that a FCHO1/2 double knockout is indistinguishable from a FCHO2 knockout [37]. The formation of clustered and abnormal clathrin-coated vesicles is equivalent in both cases.
Interaction with FCHO2 occurs through the Cig57 tyrosine-based endocytic sorting motif. To our knowledge, this is the first report of the ability of FCHO2 to recognise such motifs, as they are normally recognised by adaptor protein complexes such as AP-2. Indeed, CvpA also contains endocytic sorting motifs and these mediate interaction with AP-2 [25]. Whether Cig57 is also able to simultaneously bind AP-2 will be an interesting line of further research. Complementation of cig57::Tn with cig57 encoding mutated endocytic sorting motifs leads to CCVs that phenocopy the small vacuole phenotype observed in FCHO2 KO cells (Fig 7). Hence, the inability of the Cig57ΔESM and Cig57Y365A to complement the cig57::Tn mutant is likely due to the inability of this modified effector to bind FCHO2. Importantly, these strains were not attenuated to the same levels as the cig57::Tn mutant which may indicate that Cig57 possesses other activity and perhaps the interaction between Cig57 and FCHO2 acts to facilitate Cig57 activity on other components of the clathrin machinery. Interestingly, our results show that the Y365A mutation of Cig57 is not as attenuating as the ΔESM combined mutations, particular in relation to the expansion of the CCV. This may suggest a further role for the dileucine endocytic sorting motifs in full Cig57 function.
Clathrin-coats are formed at the plasma membrane, at the trans-Golgi network or on endosomes, and associate with adaptor protein complexes for selection of protein or lipid cargo and specificity of the final destination. Vesicle budding and receptor sorting are facilitated by the formation of a clathrin coat at a membrane. During infection with Coxiella the final destination of some of the clathrin-coated vesicles appears to be the CCV, as evidenced by the accumulation of clathrin on the CCV (Fig 5 and [25]). This would offer advantages to Coxiella to sequester extra membrane from the clathrin-coated vesicles, and to enable the delivery of nutrients in the form of cargo proteins and lipids from the vesicles to the CCV. The recruitment of clathrin-coated vesicles is dependent on the interaction between Cig57 and FCHO2. Coxiella lacking Cig57 cannot recruit clathrin to the CCV and similarly, clathrin recruitment is diminished in the absence of FCHO2. Overall this leads to biogenesis of a CCV that has reduced ability to expand and support Coxiella replication.
Despite the requirement for FCHO2 to promote the recruitment of clathrin on CCV membranes, we did not observe recruitment or enrichment of FCHO2 on the CCV. This does not rule out the possibility that this host protein is dynamically cycling on and off the CCV membrane. Importantly, FCHO2 is also not observed on mature clathrin-coated vesicles as it acts early to in the initiation of endocytosis. Thus, the Cig57-FCHO2 interaction may occur at the plasma membrane or other compartments in the cell where FCHO2 has been observed, likely due to the affinity of FCHO2 to particular phospholipids [32]. We therefore cannot discount that Cig57 is taking advantage of an as yet undiscovered role that FCHO2 has in the host. Further investigation of the biochemical function of Cig57 and investigating the functional outcome of the Cig57-FCHO2 interaction will be an exciting area of future study. This line of research will likely reveal the complex mechanisms employed by Coxiella to establish a unique intracellular niche to support replication and virulence within the human host.
Bacterial strains and plasmids used in this study are listed in S1 File. Coxiella burnetii Nine Mile Phase II (NMII), strain RSA439 was cultured axenically in liquid ACCM-2 as previously described [42]. Kanamycin and/or chloramphenicol were added to ACCM-2 at 300μg/ml and 3μg/ml respectively when required. E.coli XL1-Blue or DH5α were cultured in Luria-Bertani medium. Yeast strains were grown at 30°C in YPD (yeast extract/peptone/dextrose) or YMM (yeast minimal media) supplemented with 2% glucose and amino acids including methionine (20 μg/ml), adenine (20 μg/ml), histidine (20 μg/ml), uracil (20 μg/ml) tryptophan (20 μg/ml) and leucine (30 μg/ml) when necessary. HeLa human cervical epithelial cells (CCL-2; ATCC, Manassas, VA), and human embryonic kidney (HEK) 293T cells (a gift from Elizabeth Hartland’s laboratory, University of Melbourne) were cultured in Dulbecco’s Modified Eagle’s Media (DMEM) GlutaMAX (Gibco) supplemented with 10% heat inactivated foetal bovine serum (FBS) at 37°C with 5% CO2
Plasmid DNA was isolated using the QIAprep spin miniprep kit (Qiagen), and bacterial or HeLa gDNA was isolated using the Zymo gDNA extraction kit. Oligonucleotides to amplify gene products were obtained from Sigma, and are listed in S1 File. DNA modifying enzymes were obtained from NEB and used according to standard procedures.
pSpCas9(BB)-2A-Puro (pX459) V2.0 was a gift from Feng Zhang (Addgene plasmid #62988). Guide RNA specific for regions within FCHO2 (designed using http://crispr.mit.edu/, see S1 File) were cloned into pX459 as previously described [43]. HeLa CCL2 cells were seeded at 2.5 × 105 in 6-well dishes and the following day co-transfected with a total of 2500 ng DNA specific to two exons using Lipofectamine 3000 according to the manufacturer’s protocol. Cells were selected with puromycin (5 μg/ml) and clonally selected in 96-well plates. Modification of FCHO2 resulting in no protein production was confirmed by Western Blot analysis as below.
FUGW was a gift from David Baltimore (Addgene plasmid #14883). HEK 293T cells were seeded into a 10cm dish and transfected with 7.5 μg pFUGW-GFP-FCHO2, 3.5 μg pPAX and 2.5 μg pVSV-G for 48 hours with Lipofectamine 3000. Lentiviral particles were collected, filtered through 0.45 μm and used to infect HeLa cells in 6-well plates for 48 hours as previously described [44].
For screening, Cig57 from Coxiella was used as bait. The Matchmaker pre-transformed HeLa cDNA library (Clontech) was mated with S. cerevisiae carrying pGBKT7-cig57 according to manufacturer’s protocols (Clontech PT3183-1 manual). The Y2H Gold or AH109 strains were co-transformed with the relevant pGBKT7 or pGADT7 plasmids using the lithium acetate method and plated on DDO (-Trp, -Leu) or QDO (-Trp, -Leu, -His, -Ade) YMM plates.
Samples were suspended in 4x NuPAGE LDS sample buffer (Life Technologies) containing 50μM DTT. Proteins were separated with NuPAGE 4–12% Bis-Tris gels (Life Technologies) and transferred to PVDF membranes using the iBLOT-2 (Life Technologies). Membranes were blocked in 5% skim milk in Tris buffered saline containing 0.1% Tween 20 (TBST) and antibodies were diluted in 5% BSA or skim milk in TBST.
Cells were fixed with 4% paraformaldehyde for 20 minutes at room temperature and permeabilized with 0.05% saponin and 2% BSA in PBS (blocking solution). Primary and secondary antibodies were diluted in blocking solution and DAPI was diluted in PBS before coverslips were mounted using Prolong Gold Antifade (Invitrogen). Images were acquired with a Zeiss LSM700 or LSM710 confocal laser scanning microscope and processed using Fiji [45]. To measure clathrin intensity, five measurements of intensity were taken at the CCV or in the cytoplasm and averaged. For immunofluorescence microscopy (IF) and western blotting (WB) the following antibodies were used at the designated dilutions: α-clathrin heavy chain (Abcam, WB = 1:2000, IF = 1:1000), α-FCHO2 (ThermoFisher, WB = 1:1000), α-β-actin (Perkin Elmer, WB = 1:5000), polyclonal α-Coxiella (IF = 1:10000), α-LAMP1 (DHSB, IF = 1:250), HRP-conjugated goat α-mouse (BioRad, WB = 1:3000), HRP-conjugated goat α-rabbit (Perkin Elmer, WB = 1:3000), AlexaFluor 488 (ThermoFisher, IF = 1:2000), AlexaFluor 568 (ThermoFisher, IF = 1:2000). For clathrin localization studies, HeLa cells were seeded at 2.5 × 104 per well in 24 well plates with coverslips and the following day infected at an MOI of 100. Cells were incubated for 72 hours and processed as described above.
In 24-well plates, HeLa cells were seeded at 5×104 per well and infected with the relevant Coxiella strains the following day at an MOI of 50. To calculate the MOI, Coxiella was quantified using the Quanti-iT PicoGreen dsDNA Assay kit (Life Technologies) [46]. Cells were incubated for 4 hours at 37°C before being washed once with PBS and the media replaced with DMEM containing 5% FBS. At this time (day 0), as well as 1, 3 and 5 days post infection, cells were lysed in dH2O and pelleted by centrifugation. Genomic DNA was extracted from the samples and genome equivalents (GE) were quantified by qPCR specific for the Coxiella ompA gene [47]. Samples were collected for immunofluorescence analysis at day 3 as described above.
HeLa cells were transfected with small-interfering RNA (siRNA) using siGenome SMARTpools (Dharmacon, GE Life Sciences) against human CLTC (M-004001-00) and FCHO2 (M-024508-01) or with ON-TARGETplus (OTP) Non-targeting pool (D-001810-10-05) using Dharmafect-1 (Dharmacon, GE Life Sciences). Cells were seeded at a density of 2.5×105 in 6-well plates with a final concentration of 50 nM siRNA. After a two-day incubation, cells were replated at a density of 1.0×104 in 24-well plates, and simultaneously infected with Coxiella NMII at an MOI of 10 by centrifugation at 500 × g for 30 minutes. Cells were washed once with PBS and media replaced with DMEM containing 10% FBS. This timepoint was designated day 0. Cells were processed for western blot analysis, genome quantification and immunofluorescence as described above at days 2, 4 and 6 post infection.
Persistently infected HEK 293T cells in 10 cm dishes were transfected for 48 hours with GFP constructs using FuGENE6 before lysing in lysis buffer (10mM Tris (pH 7.5), 150 mM NaCl, 0.5 mM EDTA, 0.5% NP-40) for 30 minutes on ice. Lysates were clarified by centrifugation at 17000 x g and 10% collected for input analysis. The remaining lysate was added to 20 μl of GFP beads (ChromoTek) and incubated with mixing at 4°C for four hours. Beads were washed extensively with wash buffer (10mM Tris (pH 7.5), 150 mM NaCl, 0.5 mM EDTA, 0.01% NP-40) before being resuspended in LDS sample buffer and boiled for 10 minutes.
Statistical analyses were performed with Prism (GraphPad Software, Inc.) by use of the unpaired Student’s t test. P values less than 0.05 were considered to be significant.
|
10.1371/journal.pgen.1000338 | Genome-Wide Association Study of Plasma Polyunsaturated Fatty Acids in the InCHIANTI Study | Polyunsaturated fatty acids (PUFA) have a role in many physiological processes, including energy production, modulation of inflammation, and maintenance of cell membrane integrity. High plasma PUFA concentrations have been shown to have beneficial effects on cardiovascular disease and mortality. To identify genetic contributors of plasma PUFA concentrations, we conducted a genome-wide association study of plasma levels of six omega-3 and omega-6 fatty acids in 1,075 participants in the InCHIANTI study on aging. The strongest evidence for association was observed in a region of chromosome 11 that encodes three fatty acid desaturases (FADS1, FADS2, FADS3). The SNP with the most significant association was rs174537 near FADS1 in the analysis of arachidonic acid (AA; p = 5.95×10−46). Minor allele homozygotes had lower AA compared to the major allele homozygotes and rs174537 accounted for 18.6% of the additive variance in AA concentrations. This SNP was also associated with levels of eicosadienoic acid (EDA; p = 6.78×10−9) and eicosapentanoic acid (EPA; p = 1.07×10−14). Participants carrying the allele associated with higher AA, EDA, and EPA also had higher low-density lipoprotein (LDL-C) and total cholesterol levels. Outside the FADS gene cluster, the strongest region of association mapped to chromosome 6 in the region encoding an elongase of very long fatty acids 2 (ELOVL2). In this region, association was observed with EPA (rs953413; p = 1.1×10−6). The effects of rs174537 were confirmed in an independent sample of 1,076 subjects participating in the GOLDN study. The ELOVL2 SNP was associated with docosapentanoic and DHA but not with EPA in GOLDN. These findings show that polymorphisms of genes encoding enzymes in the metabolism of PUFA contribute to plasma concentrations of fatty acids.
| Polyunsaturated fatty acids (PUFA) have a number of beneficial effects on human health. Plasma PUFA concentrations are determined by a combination of dietary intake and metabolic efficiency. To determine the genes involved in PUFA homeostasis, we scanned the genome for genetic variations associated with plasma PUFA concentrations. The fatty acid desaturase gene, studied in previous candidate gene association studies, was the strongest determinant of plasma PUFA. A second gene encoding a fatty acid elongase was associated with long chain PUFA. The results of this study contribute to our understanding of the genetics of PUFA homeostasis. These genetic markers may be useful tools to examine the inter-relationship between diet, genetics, and disease.
| Polyunsaturated fatty acids (PUFA) refer to the class of fatty acids with multiple desaturations in the aliphatic tail. Short chain PUFA (up to 16 carbons) are synthesized endogenously by fatty acid synthase. Long chain PUFA are fatty acids of 18 carbons or more in length with two or more double bonds. Depending on the position of the first double bond proximate to the methyl end, PUFA are classified as n-6 or n-3. Long chain PUFA are either directly absorbed from food or synthesized from the two essential fatty acids linoleic acid (LA; 18:2n-6) and alpha-linolenic acid (ALA; 18:3n-3) through a series of desaturation and elongation processes [1]. The initial step in PUFA biosynthesis is the desaturation of ALA and LA by the enzyme d6-desaturase (FADS2; GeneID 9415) (Figure 1). PUFA modulate inflammatory response through a number of different mechanisms including modulation of cyclooxygenase and lipoxigenase activity [2]. Cyclooxygenase and lipoxigenase are essential for production of eicosanoids and resolvins [2]–[4]. Since n-3 and n-6 fatty acids compete for the same metabolic pathway and produce eicosanoids with differing effects, it has been theorized that the balance of the two classes of PUFA may be important in the pathogenesis of inflammatory diseases.
Epidemiological studies have shown that fatty acid consumption and plasma levels, in particular of the n-3 family, are associated with reduced risk of cardiovascular disease [5]–[7], diabetes [8]–[10], depression [11],[12], and dementia [13]. However, not all studies show significant associations and there has been inconsistencies in the direction of the associations especially for the n-6 acids [14],[15]. The different methods (dietary questionnaire or biomarkers) for accessing PUFA status may contribute to discrepant results [16]–[18]. The disadvantage of using dietary PUFA intake is the evidence of inaccuracies intrinsic in any reporting methods of dietary intake that plasma levels would circumvent. In addition, direct measures of PUFA reflect the cumulative effects of intake and endogenous metabolism. Dietary fatty acids can be converted into longer chain PUFA or stored for energy thus another reason for inconsistent results may be due the general lack of control for individual differences in metabolism once fatty acids are consumed.
Previous studies have examined the association of genetic variants, especially polymorphisms in the FADS genes, on fatty acid concentrations in plasma and erythrocyte membranes [19]–[21]. There are 3 FADS (FADS1 [GeneID 3992] ,FADS2, and FADS3 [GeneID 3992]) clustered on chromosome 11. Variants in FADS1 and FADS2 have been consistently shown to be associated with PUFA concentrations. It is unknown whether other loci also determine fatty acid concentrations. To address this question, we conducted a genome-wide association study of plasma fatty acid concentration in participants in the InCHIANTI study.
Linoleic acid (LA) constituted the highest proportion of total fatty acids followed by arachidonic acid (AA) (Table 1) The narrow heritability was highest for AA (37.7%) followed by LA (35.9%), eicosadienoic acid (EDA, 33.3%), alpha-linolenic acid (ALA, 28.1%), eicosapentanoic acid (EPA, 24.4%), and docosahexanoic acid (DHA,12.0%). For EDA, AA, and EPA, genome-wide significant signals fell in the FADS1/FADS2/FADS3 region on chromosome 11 (Figure 2, Figure 3, Table S1). Of these, the most significant SNP was rs174537 for AA (P = 5.95×10−46), where the variant explained 18.6% of the additive variance of AA concentrations. This SNP was significantly associated with EDA (P = 6.78×10−9), and EPA (P = 1.04×10−14). The association with LA (P = 5.58×10−7) and ALA (P = 2.76×10−5) did not reach genome-wide significance, and there was no association with DHA (P = 0.3188). Presence of the minor allele (T) was associated with lower concentrations of longer chain fatty acids (EDA, AA, EPA), but with higher concentrations of LA and ALA (Table 2). With the exception of DHA, the SNPs exhibiting the strongest evidence of association with the fatty acids examined in this study mapped to the FADS1, FADS2, and FADS3 cluster. The most significant SNP for DHA was on chromosome 12 within the SLC26A10 gene (GeneID 65012, rs2277324; PDHA = 2.65×10−9). In all cases, inclusion of the most significant SNP as a covariate in the model resulted in attenuation of the effect of the other SNPs in the region (Figure S1). Accordingly, associated SNPs in this region were in significant linkage disequilibrium with each other in the InCHIANTI sample (Figure S2).
To investigate whether this SNP has an effect on other cardiovascular disease risk factors, we examined the association of rs174537 with plasma lipid parameters. Significant association was observed with total cholesterol (P = 0.027) and LDL-cholesterol (P = 0.011), but not with either HDL-C (P = 0.775) or triglycerides (P = 0.862; Table 2). The minor allele homozygotes (TT) had 8 mg/dL lower total cholesterol and 9 mg/dL lower LDL-C compared with GG subjects.
To identify other putative chromosome regions associated with fatty acid concentrations beyond the FADS cluster, we examined the top 3 non-redundant (r2<0.2) SNPs from the analysis of each fatty acid and selected the SNPs that mapped to candidate gene regions (defined as 20kb upstream of intron 1, or downstream of last exon) for replication in an independent study (Table S1). These included rs16940765 (HRH4 [GeneID 59340], chr18, PEDA = 2.18×10−6), rs17718324 (SPARC [GeneID 6678], chr5, PAA = 7.64×10−7) and rs953413 (ELOVL2 [GeneID 54898], chr6, PAA = 1.1×10−6). In the InCHIANTI study, these four SNPs most strongly associated with a specific fatty acid, unlike FADS cluster that was associated with multiple fatty acids. We noted, however that rs16940765 (HRH4), rs953413 (ELOVL2) and rs2277324 (SLC26A10) showed significant association at the 0.05 level for AA (P = 0.003), DHA (0.004), and EPA (P = 0.004) respectively. In addition, there were four SNPs (rs953413, rs1570069, rs3798719, rs7744440) in the ELOVL2 gene were associated with EPA with p values ranging from 9.51×10−5 to 1.10×10−6 (Figure S3). In total, 5 SNPs (rs174537, rs2277324, rs16940765, rs17718324, rs953413) were selected for replication.
In the GOLDN study, there were significant associations of FADS SNP, rs174537, with ALA, LA, AA, EPA and DHA (P<0.001) and marginal association with docosapentaenoic acid (DPA) (P = 0.068) (Table 2). As with the InCHIANTI study, presence of the T allele was associated with higher ALA and LA concentration and lower AA, EPA, DPA and DHA concentrations. Consistent with the InCHIANTI study, this SNP was associated with total cholesterol and LDL-C but not triglycerides or HDL-C. We also observed strong associations of rs953413 with docosapentanoic acid (DPA; P = 0.002) and DHA (P<0.001). The presence of the minor allele (A) was associated with lower DHA and higher DPA and higher AA compared to the minor allele carriers (Table 2). The remaining 3 SNP (rs2277324, rs16940765, rs17718324) were not associated with fatty acid concentrations in the GOLDN study (data not shown).
The genome-wide association approach enables comprehensive examination of the genome to identify novel loci contributing to PUFA homeostasis. In addition, the significance of the genes previously reported in association with PUFA can be assessed relative to other regions in the genome. Here, we demonstrated that polymorphisms in the FADS cluster are the strongest determinants of plasma and erythrocyte fatty acid concentrations, explaining up to 18.6% of the additive variance in plasma AA levels. Consistent with prior reports, the greatest evidence of association was observed in the region containing FADS1, FEN1 (flap structure specific endonuclease, GeneID 2237), two hypothetical proteins (C11orf9 [GeneID 745], C11orf10 [GeneID 746]), and the promoter region of FADS2 [20],[21]. With the exception of EDA, the direction of the association of rs174537 with plasma and erythrocyte fatty acids is consistent with previous reports. We find that there are higher levels of ALA and LA which is suggestive of an accumulation of the initial products of the PUFA metabolic pathway. The cluster of SNP ranging from rs174537 to rs509360 showed the strongest evidence for association, and contains the haplotype block previously examined in relation to plasma and erythrocyte fatty acids [20],[21]. Based on the HapMap CEU data, the r2 between rs174537 and previously reported SNPs were ≥0.8. If functional polymorphisms exist within this region, it could affect the expression of both desaturases. To this end, in a recent report of genome-wide association of global gene expression, the rs174546 that is in LD with the rs174537 (r2 = 0.99) was associated with FADS1 expression (LOD = 5, P = 1.6×10−6) but not FADS2 (LOD = 0.7, P = 0.07) in lymphoblastoid cells [22],[23]. The allele associated with higher AA showed greater expression of FADS1, consistent with our results. Since FADS1 and FADS2expression varies by tissue type, it would be of interest to examine the effect of the variant on gene expression in other cell types [24].
The T allele associated with lower AA was also associated with decreased LDL-C and total cholesterol. The association with LDL-C was also observed in a large meta-analysis of plasma lipid concentrations in ∼8500 subjects [25],[26]. In this meta-analysis, there was stronger evidence of association with high density lipoprotein (HDL-C) and triglycerides (TG), where the T allele displayed lower HDL-C and higher triglyceride concentrations. Finally, in the Welcome Trust Case-Control Consortium coronary artery disease (CAD) study, the T allele was associated with increased prevalence of CAD (P = 0.0375) [27]. The increased prevalence of CAD, low HDL-C and high TG is consistent with lower AA concentrations in the T allele carriers. Endogenous PUFA are natural ligands of peroxisome proliferator activating receptor alpha (PPARA) [28]. PPARA activation has been shown to elevate HDL-C and lower TG by inducing the expression of ApoA1, Apo-AII, lipoprotein lipase and suppressing ApoCIII [29]–[32]. Thus the low AA, EPA and EDA in the T allele carriers will results in lower PPARA activation. Under this hypothesis, we would expect the T allele carriers to display higher LDL-C since PPARA is known to enhance LDL-C clearance [33] . However, in both the InCHIANTI and GOLDN study, lower concentrations of LDL-C are observed. It is likely that there are other mechanisms by which fatty acids regulate lipoprotein homeostasis, for example through membrane fluidity. It may be possible, that the higher concentrations of linoleic and linolenic acid in the T allele carriers results in increased membrane fluidity, thus increasing LDL-receptor recycling leading to lower LDL-C.
The elongation of very long chain fatty acid (ELOVL) family genes are elongases that catalyze the rate-limiting condensation reaction resulting in the synthesis of very long chain fatty acids (VLCFA) [34]. To date, six ELOVL genes have been described. The ELOVL1, 3 and 6 are involved in synthesis of monounsaturated and saturated long chain fatty acids while ELOVL2, 4 and 5 elongate polyunsaturated fatty acids [35]. In this study, rs953413 in the ELOVL2 was the third most significant SNP in the analysis of EPA, with strong, although not genome-wide level significant association with long chain fatty acids EPA and DHA. In GOLDN, there were no significant associations of this SNP with EPA, but a significant association was observed with DPA. In mammals, two elongation steps are required for the synthesis of DHA from EPA. First, EPA is elongated to DPA, then to 24:5n-3 followed by a desaturation and retroconversion step to form DHA [1] (Figure 1). The two initial elongation steps of 20 and 22-C fatty acids are mediated by ELOVL2 [36]. The rs953413 is associated with substrate EPA (InCHIANTI), and product DPA (GOLDN) and DHA (both studies) of the EVOLV2 pathway. Plasma DPA levels were not measured in InCHIANTI, thus whether this association is also observed in this population cannot be investigated. Why the ELOVL2 SNP was not associated with EPA in GOLDN is not clear, however it may reflect the differences in fatty acid metabolism in erythrocytes versus plasma as they reflect two slightly different pools of fatty acids [37]. Plasma fatty acids reflect short term intake of fatty acids whereas erythrocyte levels reflect long term intake. Thus the different results between the plasma and erythrocyte fatty acids may reflect dietary differences between subjects in the GOLDN (USA) and the InCHIANTI study (Italy). Regardless of these differences, the results of this study suggestive of the role of ELOVL2 in the conversion of EPA to DHA in humans. The presence of the minor (A) allele was associated with higher EPA/DPA and lower DHA. If rs953413, located in intron 1, is the functional SNP (or is in LD with the functional SNP), this variant would likely be associated with lower expression of the ELOVL2 or result in a less efficient variant of the elongase resulting in decreased elongation of EPA to DHA. In lymphoblastoid cells, this SNP was not associated with ELOVL2 expression (LOD = 0.4, P = 0.2) [22],[23]. Further investigation in other cell lines and functional analysis of the different variants is warranted.
In summary, we have shown that the major loci for fatty acid concentrations in both plasma and erythrocyte membranes are in genes involved in the metabolism of PUFA. The FADS locus on chromosome 11 was the major contributor of plasma fatty acid concentrations, and thus may have implications for cardiovascular disease. In addition, we have identified a second promising locus in ELOVL2 that is involved in the homeostasis of longer chain n-3 fatty acids. Future studies should investigate the interactions between dietary intake, circulating levels of fatty acids and genetic variants on risk of diseases such as cardiovascular disease.
The InCHIANTI study is a population-based epidemiological study aimed at evaluating factors that influence mobility in the older population living in the Chianti region of Tuscany, Italy. Details of the study have been previously reported [38]. Briefly, 1616 residents were selected from the population registry of Greve in Chianti (a rural area: 11 709 residents with 19.3% of the population greater than 65 years of age) and Bagno a Ripoli (Antella village near Florence; 4704 inhabitants, with 20.3% greater than 65 years of age). The participation rate was 90% (n = 1453) and participants ranged between 21–102 years of age. Overnight fasted blood samples were collected for genomic DNA extraction and measurement of plasma fatty acids. Genotyping was completed for 1231 subjects using the Illumina Infinium HumanHap 550 genotyping chip (ver1 and ver3 chips were used). The study protocol was approved by the Italian National Institute of Research and Care of Aging Institutional Review.
There were 85 parent-offspring pairs, 6 sib-pairs and 2 half-sibling pairs documented. We investigated any further familial relationships using IBD of 10,000 random SNPs using RELPAIR and uncovered 1 parent-offspring, 79 siblings and 13 half-sibling [39]. We utilized the correct family structure inferred from genetic data for all analyses. In addition, we identified 2 duplicated samples and removed these from the study. Sample quality was assessed using the GAINQC program (http://www.sph.umich.edu/csg/abecasis/GainQC/). The average genotype completeness and heterozygosity rates were 98% and 32% respectively. We excluded subjects that had less than 97% of genotyped completeness (n = 12), heterozygosity rate of less than 30% (n = 5) and misspecified sex based on heterozygosity of the X chromosome SNPs (n = 1). The final sample size used for SNP quality control was 1210.
The confirmation study population consisted of 1120 white men and women in the United States participating in the Genetics of Lipid Lowering Drugs and Diet Network (GOLDN) Study. The majority of participants were re-recruited from the ongoing National Heart and Lung and Blood Institutes (NHLBI) Family Heart Study (FHS) [40] in two genetically homogeneous centers (Minneapolis, MN and Salt Lake City, UT). GOLDN is part of the Program for Genetic Interactions (PROGENI) Network, a group of NIH-funded intervention studies of gene-environmental interactions. The primary aim of the GOLDN study was to characterize the genetic components of triglycerides response following a high fat meal and hypolipedemic drug, fenofibrate. Detailed study design and methodology has been previously described [41],[42].
In the replication sample, we excluded persons with missing genotypes or extreme fatty acid values. The final data set consists of information on 1076 individuals. The protocol for this study was approved by the Human Studies Committee of Institutional Review Board at University of Minnesota, University of Utah and Tufts University/New England Medical Center. Written informed consent was obtained from all participants.
InCHIANTI: Plasma fatty acid measurement methods has been described previously [43]. Briefly, blood samples were collected in the morning after a 12-hr overnight fast. Aliquots of plasma were immediately obtained and stored at −80 C. Fatty acid methyl esters (FAME) were prepared through transesterification using Lepage and Roy’s method with modification [44],[45]. Separation of FAME was carried out on an HP-6890 gas chromatograph (Hewlett- Packard, Palo Alto, CA) with a 30-m fused silica column (HP-225; Hewlett-Packard). FAMEs were identified by comparison with pure standards (NU Chek Prep, Inc., Elysian, MA). For quantitative analysis of fatty acids as methyl esters, calibration curves for FAME (ranging from C14:0 to C24:1) were prepared by adding six increasing amounts of individual FAME standards to the same amount of internal standard (C17:0; 50xg). The correlation coefficients for the calibration curves of fatty acids were in all cases higher than 0.998 in the range of concentrations studied. Fatty acid concentration was expressed as a percentage of total fatty acids. The coefficient of variation for all fatty acids was on average 1.6% for intraassay and 3.3% for interassay. HDL-C, total cholesterol and triglycerides were determined using commercial enzymatic tests (Roche Diagnostics, Mannheim, Germany). Serum low-density lipoprotein cholesterol (LDL-C) was computed with the Friedewald formula (LDL-C = total cholesterol − HDL-C − triglicerides/5).
GOLDN: Fatty acids (FA) in erythrocyte membrane were measured following procedures described previously [46] Briefly, lipids were extracted from the erythrocyte membranes with a mixture of chloroform:methanol (2:1, v/v), collected in heptanes and injected onto a capillary Varian CP7420 100-m column with a Hewlett Packard 5890 gas chromatograph (GC) equipped with a HP6890A autosampler. The GC was configured for a single capillary column with a flame ionization detector and interfaced with HP chemstation software. The initial temperature of 190°C was increased to 240°C over 50 minutes. Fatty acid methylesters from 12:0 through 24:1n9 were separated, identified and expressed as percent of total fatty acid. Triglycerides were measured using a glycerol blanked enzymatic method (Trig/GB, Roche Diagnostics Corporation, Indianapolis, IN) and cholesterol was measured using a cholesterol esterase, cholesterol oxidase reaction (Chol R1, Roche Diagnostics Corporation) on the Roche/Hitachi 911 Automatic Analyzer (Roche Diagnostics Corporation). For HDL-cholesterol, the non-HDL-cholesterol was first precipitated with magnesium/dextran. LDL-cholesterol was measured by a homogeneous direct method (LDL Direct Liquid Select Cholesterol Reagent, Equal Diagnostics, Exton, PA).
In the InCHIANTI, dietary intake was assessed using a food-frequency questionnaire (FFQ) created for the European Prospective Investigation into Cancer and nutrition (EPIC) study, and has previously been validated to provide good estimates of dietary intake in this study population [47],[48]. In GOLDN, habitual dietary intake was estimated using the validated dietary history questionnaire (DHQ) developed by the National Cancer Institute [49]. We excluded subjects that reported <800 kcal and >5500 kcal in men and <600kcal and >4500kcal in women.
InCHIANTI: Genome-wide genotyping was performed using the Illumina Infinium HumanHap550 genotyping chip (chip version 1 and 3) as previously described [50]. The SNP quality control was assessed using GAINQC. The exclusion criteria for SNPs were minor allele frequency <1% (n = 25,422), genotyping completeness <99% (n = 23,610) and Hardy Weinberg-equilibrium (HWE) <0.0001 (n = 517).
GOLDN: Five SNPs were selected for replication in the GOLDN study: rs953413, rs2277324, rs16940765, rs17718324 and rs174537. One of these, rs2277324, failed genotyping and therefore another SNP in high LD, rs923838 (r2 = 0.89 in hapmap), was used as a proxy for this SNP. DNA was extracted from blood samples and purified using commercial Puregene reagents (Gentra System, Inc.) following manufacturer’s instructions. SNPs were genotyped using the 5’nuclease allelic discrimination Taqman assay with allelic specific probes on the ABI Prism 7900HT Sequence Detection System (Applies Biosystems, Foster City, Calif, USA) according to standard laboratory protocols. The primers and probes were pre-designed (the assay -on -demand) by the manufacturer (Applied Biosystem) (Assay ID: FEN_rs174537: C___2269026_10, HRH4_rs16940765: C__32711739_10, SPARC_rs17718324: C__34334455_10, ELOVL2_rs953413: C___7617198_10, rs923828: C___2022671_10).
InCHIANTI GWAS: Inverse normal transformation was applied to plasma fatty acid concentrations to avoid inflated type I error due to non-normality [51]. The genotypes were coded 0, 1 and 2 reflecting the number of copies of an allele being tested (additive genetic model). For X-chromosome analysis, the average phenotype of males hemizygous for a particular allele was treated assumed to match the average phenotype of females homozygous for the same allele. Association analysis was conducted by fitting simple regression test using the fastAssoc option in MERLIN [52]. Narrow heritability reflects the ratio of the trait’s additive variance to the total variance [51],[53]. In all the analyses, the models were adjusted for sex, age and age squared. The genomic control method was used to control for effects of population structure and cryptic relatedness [54]. An approximate genome-wide significance threshold of 1×10−7 (∼0.05/495343 SNPs) was used. For each fatty acid concentration, a second analysis included the most significant SNP from the first pass analysis as a covariate. Linkage disequilibrium coefficints within the region of interest were calculated using GOLD [55].
For the other phenotypes (total cholesterol, triglycerides, LDL-cholesterol, HDL-cholesterol and BMI), the traits were normalized either by natural log or square root transformation when necessary. Associations for each SNP were investigated using the general linear model (GLM) procedure in SAS.
GOLDN: Inverse normal transformation was applied to erythrocyte membrane fatty acid concentration to achieve approximate normality. For the additive model, genotype coding was based on the number of variant alleles at the polymorphic site. With no significant sex modification observed, men and women were analyzed together. We used the generalized estimating equation (GEE) linear regression with exchangeable correlation structure as implemented in the GENMOD procedure in SAS (Windows version 9.0, SAS Institute, Cary, NC) to adjust for correlated observations due to familial relationships. Potential confounding factors included study center, age, sex, BMI, smoking (never, former and current smoker), alcohol consumption (non-drinker and current drinker), physical activity, drugs for lowering cholesterol, diabetes and hypertension and hormones. A two-tailed P value of <0.05 was considered to be statistically significant.
|
10.1371/journal.pbio.1001225 | Mitotic Spindle Assembly around RCC1-Coated Beads in Xenopus Egg Extracts | During cell division the genetic material on chromosomes is distributed to daughter cells by a dynamic microtubule structure called the mitotic spindle. Here we establish a reconstitution system to assess the contribution of individual chromosome proteins to mitotic spindle formation around single 10 µm diameter porous glass beads in Xenopus egg extracts. We find that Regulator of Chromosome Condensation 1 (RCC1), the Guanine Nucleotide Exchange Factor (GEF) for the small GTPase Ran, can induce bipolar spindle formation. Remarkably, RCC1 beads oscillate within spindles from pole to pole, a behavior that could be converted to a more typical, stable association by the addition of a kinesin together with RCC1. These results identify two activities sufficient to mimic chromatin-mediated spindle assembly, and establish a foundation for future experiments to reconstitute spindle assembly entirely from purified components.
| The mitotic spindle is a bipolar structure that is responsible for separating the two sets of duplicated chromosomes in a dividing cell, thereby delivering one set to each of the two daughter cells. It is built from dynamic filaments called microtubules, as well as hundreds of other components that contribute to the organization and dynamics of the microtubules and to chromosome movement. To understand which proteins are essential for spindle formation and function, we would like to be able to build it from purified components. As a step towards this goal, we coupled individual proteins to inert glass beads (as a substitute for chromosomes), to determine what combination of proteins can induce spindle assembly in a complex cytoplasm derived from frog eggs. We found that a single enzyme called RCC1 is sufficient to activate a pathway that stabilizes and organizes microtubules into a bipolar structure around the bead, but that this bead then oscillated back and forth between the poles of the spindle. By coupling a microtubule-based motor protein together with RCC1 on the bead, we were able to balance the bead in the center of the spindle. Thus, two proteins immobilized on a bead can substitute for a chromosome and induce stable spindle formation.
| The spindle is a highly dynamic structure composed of microtubule polymers and hundreds of other factors including motor proteins and microtubule-associated proteins (MAPs) [1]. Its purpose is to attach to chromosomes and accurately segregate them to daughter cells. Once thought to be passive participants, chromosomes are now known to play an active role in spindle assembly, since immobilized mitotic chromatin [2]–[4], or chromosome fragments containing microtubule attachment sites (kinetochores) [5], have been shown to direct the formation of spindle structures. However, the minimal chromosome components sufficient to generate a spindle have not been defined.
One candidate enzyme associated with chromatin that could drive spindle assembly is RCC1, the guanine nucleotide exchange factor (GEF) for the small GTPase Ran. RCC1 generates a steep gradient of RanGTP near chromosomes, activating a subset of mitotic motors and MAPs that are cargoes of the importin β family of nuclear transport receptors [6]. Addition of a hydrolysis-deficient mutant of Ran bound to GTP (RanQ69L-GTP) stabilizes microtubules that are organized by motor proteins into asters and small spindle-like structures in metaphase-arrested cytoplasmic extracts prepared from Xenopus laevis eggs [7]–[10]. RanQ69L-GTP disrupts the RCC1-generated RanGTP gradient and spindle assembly [11], while flattening the gradient eliminates spindle assembly around chromatin beads [12]. These experiments demonstrate that a RanGTP gradient is required for chromatin-dependent spindle assembly in Xenopus egg extracts, but is it sufficient? We set out to test whether immobilized RCC1 in the absence of other chromatin factors can reconstitute a mitotic spindle.
To test whether a single protein factor, RCC1, is sufficient to direct spindle formation in egg extracts, we developed a novel substrate consisting of single 10 µm diameter porous NeutrAvidin beads. This approach alleviates the need for small beads to cluster or align by generating a high surface area to which biotinylated molecules can be tightly bound (Figure 1A). RCC1 (α isoform) engineered to contain a single amino-terminal biotin was coupled to the beads, which were then incubated in metaphase-arrested egg extracts containing rhodamine-labeled tubulin and observed by fluorescence microscopy. Whereas uncoupled or bovine serum albumin (BSA)-coupled NeutrAvidin beads had no activity (unpublished data), microtubule arrays formed around RCC1 beads that could be sorted into five major categories (Figure 1B). Robust bipolar structures made up greater than 30% of the arrays, with a distribution of categories similar to single chromatin-coated beads under the same reaction conditions (Figure 1C). Notably, however, RCC1 bead spindle morphology differed from that of individual chromatin beads, which induced larger spindles that contained more microtubules as determined by microtubule fluorescence intensity (Figure 1D,E). Furthermore, whereas chromatin bead spindle microtubules were most dense in the center, RCC1 bead spindle microtubules had a higher density at the poles (Figure 1D). Thus, immobilization of a single chromatin component, RCC1, is sufficient to induce bipolar spindle formation in X. laevis egg extracts, but spindle morphology is different, suggesting that the pathway is not completely active, or that other chromatin proteins contribute to microtubule stabilization and organization.
To determine whether RCC1 beads fully recapitulate RanGTP-driven mitotic cargo activation, we added the FRET probe Rango-2, which measures the level of cargo release from importin β [11],[13]. Rango gradients surrounding RCC1 beads appeared similar to those formed around chromatin beads, and both were further enhanced by addition of wild-type Ran, indicating that the RCC1 beads have similar activity compared to chromatin and that cargo gradients are limited by the amount of Ran that can be loaded with GTP (Figure 2A). Consistent with this interpretation, varying the amount of RCC1 per bead 4-fold had little effect on the distribution of microtubule structures (unpublished data), whereas addition of wild-type Ran up to three times its estimated endogenous concentration of ∼3 µM resulted in a dose-dependent decrease in monopolar and bad bipolar structures, while the percentage of multipolar structures indicative of enhanced microtubule polymerization increased (Figure S1). Further evidence of equivalent Ran pathway activity by RCC1 beads compared to chromatin was the similar localization of the cargo TPX2 to spindle microtubules, and sensitivity to spindle disruption by a truncation mutant of importin β (amino acids 71–876) that still binds cargo but does not bind RanGTP (Figure 2B; Figure S2) [14],[15]. Despite their morphological differences, chromatin and RCC1 bead spindle dynamics and organization were similar. Both spindle types displayed poleward microtubule flux at similar rates, indicating kinesin-5 activity (unpublished data) [16],[17], and spindle morphology was similarly disrupted upon inhibition of the microtubule cross-linking spindle pole protein NuMA (Figure 2C; Figure S2) [18]. Therefore, RCC1 beads appear to fully activate the RanGTP cargo release pathway and generate spindles that possess structural and dynamic features of spindles formed around sperm chromosomes or chromatin-coated beads.
One distinctive behavior observed by time-lapse fluorescence microscopy was the tendency of single RCC1 beads to oscillate within the spindle from pole to pole, whereas chromatin beads were stationary (Figure 3A,B; Video S1). We observed a range in bead oscillatory activity, which sometimes dampened over time. Interestingly, when a monopolar microtubule array formed, the RCC1 bead moved unidirectionally, appearing to be pushed along by a trail of polymerizing microtubules (Figure 3C; Video S2). This motility was reminiscent of actin polymerization-driven propulsion of the bacterium Listeria monocytogenes or beads coated with its actin nucleation promoting protein ActA, which does not require motor activity [19]. We therefore propose that the oscillatory movement of an RCC1 bead occurs because the bead does not connect to spindle microtubules and is pushed from pole to pole by polymerizing microtubule plus ends. Analogous to a bead uniformly coated with ActA that induces polarized actin assembly, a symmetry-breaking event might initiate RCC1 bead motility [20],[21]. Because of the antiparallel orientation of microtubules in bipolar structures, the RCC1 bead is driven towards the opposite spindle pole where it encounters a higher density of polarized microtubules polymerizing in the opposite direction. Such a polar ejection force of spindle microtubules has been well documented, although oscillatory chromosome movement also requires kinetochore fibers [22]. We cannot rule out that bead movement is regulated by egg extract factors, but biochemical association of specific extract proteins was not observed (unpublished data). We therefore reasoned that additional chromatin factors normally act to stabilize interaction with spindle microtubules.
Plus end-directed chromatin-associated kinesins (chromokinesins) contribute to multiple aspects of mitosis, regulating spindle microtubule dynamics and organization, as well as chromosome compaction and segregation [23], and represent excellent candidate factors for mediating chromatin-spindle interactions. Furthermore, kinesin-coated beads can move directionally along microtubules in a reconstituted system [24]. To determine whether plus end-directed microtubule-based motor activity in the absence of other chromokinesin functions was sufficient to stabilize RCC1 bead spindles, we first coupled conventional kinesin-1 motor domain (amino acids 1–560) together with RCC1 to the beads at a 1∶1 ratio of proteins. Although bipolar spindles initially formed around the hybrid beads, the poles eventually collapsed together and pushed the bead out of the spindle, generating a “push pole” morphology (Figure 4A, Video S3). These observations suggest that, like around chromatin and RCC1 beads, microtubules are nucleated in random orientations and quickly attain an antiparallel organization due to the activity of kinesin-5 and other motors [4],[25]. Once microtubule plus ends become oriented toward the bead, however, the strong processive motor activity of kinesin-1 appeared to dominate microtubule organization, clustering plus ends at the surface of the bead. Interestingly, the glass bead often cracked once the poles were pushed together and outward, indicating that the kinesin motor can generate significant force against the bead (Video S3, right panel).
In contrast to kinesin-1, chromosomal kinesins -4 and -10 have slower motility and are weakly processive [26]. We therefore mutated the motor domain of kinesin-1, changing residue 203 from arginine to lysine (R203K) to reduce its ATPase activity and motility, but preserve microtubule binding [27]. Remarkably, beads coated with a 1∶1 ratio of RCC1 and kinesin-1(R203K) induced bipolar spindle assembly but were stationary like chromatin beads (Figure 4B, Video S4). Whereas no obvious effect on spindle morphology was observed, the percentage of bipolar spindles formed after 1 h increased by approximately 30% (Figure 4C). These results show that by mediating bead-microtubule interactions and centering within the spindle, a non-motile kinesin strongly enhances the stability of bipolar microtubule arrays induced by RCC1 beads.
To investigate whether other microtubule binding activities were sufficient to stabilize RCC1 bead spindles, we substituted the microtubule plus end binding protein EB1 for kinesin-1(R203K). EB1 associates with growing microtubules and functions to recruit a large set of proteins that regulate microtubule dynamics and interactions with cellular structures including kinetochores [28]. Unlike kinesin-1(R203K), the presence of EB1 on RCC1 beads did not decrease the percentage of beads oscillating or their movement amplitudes, which appeared indistinguishable compared to RCC1 beads (Figures 4D, 3B; Video S5). Nevertheless, the percentage of bipolar spindles formed at 1 h increased, similar to the RCC1/kinesin-1(R203K) hybrid beads (Figure 4C,E). EB1 may be mediating bead-spindle microtubule interactions that promote spindle bipolarity, but through a distinct microtubule binding mechanism. Alternatively, EB1 on beads may be acting directly or indirectly to stabilize spindle microtubules, thereby facilitating the activity of kinesin-5 and other motors to sort them into bipolar arrays. These results indicate that bead oscillations per se do not impair spindle bipolarity, and differential effects on the spindle likely reflect the distinct properties of EB1 and kinesin. We hypothesize that molecular motor domains provide the microtubule interaction best suited to stabilize RCC1 bead or chromosome arm position within a spindle. Evaluation of additional microtubule binding proteins or domains, and careful analysis of bead spindle phenotypes and dynamics will be necessary to elucidate underlying mechanisms.
In summary, stable bipolar spindle assembly can be induced in the absence of chromosomes, chromatin, or kinetochores by coupling two proteins, RCC1 and kinesin-1 (R203K), to beads. These results demonstrate that the anisotropy of RanGTP distribution in the cytoplasm is sufficient to drive mitotic spindle assembly. However, chromatin spindles are still larger and more robust, indicating that other chromatin-associated factors must also contribute to normal spindle morphology. One possible activity is the Aurora B kinase, which phosphorylates and inactivates microtubule-destabilizing proteins [29],[30] and makes important contributions to kinetochore-driven spindle assembly [12]. Enrichment of mitotic kinases on beads in Xenopus egg extracts has previously been achieved by adding beads coupled with IgGs specific for Aurora A [31], or for the chromosome passenger complex protein INCENP [12], which binds Aurora B. While neither kinase was found to be sufficient for spindle assembly, they clearly play supporting roles. Crucial also may be bona fide chromokinesins such as Xklp1, which has been shown to recruit the bundling MAP PRC1 to generate the antiparallel microtubule arrays of the central spindle [32]. Our system establishes a foundation to test the roles of other chromatin and spindle factors and forms the basis for spindle reconstitution entirely from defined components.
Metaphase-arrested X. laevis egg extracts and biotinylated DNA for coupling to beads were prepared as described [4],[33]. Glass beads (GFS chemicals, #84503) were modified with silane, activated with glutaric anhydride/trimethylamine, and then linked to NeutrAvidin (Pierce), to which biotinylated proteins or plasmid DNA were coupled. Full-length human RCC1 (α isoform) containing a biotinylation tag was expressed and biotinylated in Escherichia coli, and purified by ion-exchange chromatography. Bacterially expressed 6x-histidine-tagged human kinesin motor domains purified by nickel and ion exchange chromatography were biotinylated in vitro at a single cysteine with biotin maleimide. Human EB1 with an N-terminal 6x-histidine-tag and biotinylation tag was expressed and biotinylated in E. coli and purified by nickel chromatography. Ratios of 1∶1 RCC1 to kinesin proteins, EB1, or biotinylated BSA were coupled to beads. Extract reactions containing beads and rhodamine-labeled tubulin were pre-incubated on ice and then spotted between slides and coverslips that had either been PEG-modified or treated with NaOH. After 30–60 min, reactions were viewed live by wide field fluorescence microscopy. Images and time-lapse movies were collected with a CCD camera and MicroManager software and analyzed using ImageJ. FRET images in the presence of Rango-2 were collected and processed as described [13],[14]. WT Ran (8–10 µM) was added to reactions described in Figures 2B,C, 3, and 4.
Slides were prepared in two different ways. Microscope slides and coverslips were placed in metal racks (Electron Microscopy Sciences #72239-04 and #71420-25) rinsed thoroughly with water, then dip rinsed and stored in 10 M sodium hydroxide within a glass container. The container was then bath sonicated for 30 min. The slides and coverslips were kept within the container overnight. The following morning, slides and coverslips were rinsed with water and either (1) rinsed with ethanol, blown dry, and stored under vacuum or (2) dip rinsed in 0.2 M acetic acid for 1 min. Next they were rinsed with water. Excess water was removed by blowing with clean air, and then the racks were placed in ethanol. 1.5% hydroxy(polyethyleneoxy)propyltriethoxysilane (Gelest, #SIH6188.0) in 93.5% ethanol, 5% acetic acid was added to the slides and coverslips in a new glass containers and rocked at room temperature for 2 h. The racks were then removed from the silane solution and dip rinsed in four baths of ethanol then one bath of acetone for 1 min each. After the acetone, the slides and coverslips were blown dry with clean air and baked at 107°C for 30 min. After cooling for 1 h, they were placed in a vacuum desiccator and used for the following 2 d.
Approximately 10 µg (1 µl) of beads were added to 29 µl of X. laevis egg extracts and stored on ice with gentle tapping every 10 min. After 30–60 min, 5 µl of the reactions were spotted on a slide under a 22×22 mm coverslip, put in a humidity chamber, and kept in the dark. After 30 to 60 min on the slide, the live reactions were observed by fluorescence microscopy. Beads, microtubules, and DNA were visualized by adding 2 µg/ml Streptavidin Alexa Fluor 488 conjugate (Invitrogen #S-32354), 50 µg/ml rhodamine-labeled tubulin, and 2 µg/ml Hoechst 33342 dye, respectively, to the extract.
Human full-length RCC1 was cloned into pRSETA with a biotin acceptor peptide (GLNDIFEAQKIEWHE) and a spacer (ASTPPTPSPSTPPT) on the N-terminus [34]. RCC1 and the E. coli biotin ligase (BirA) were cotransformed into BL21 DE3 competent cells. Cells were outgrown for 1.5 h, added to 1 L of LB [35] with ampicillin and chloramphenicol, and grown overnight at 37°C. The next morning, 150 ml of culture was added to 1 L of fresh LB/amp/chlor (6 L total) and grown at 37°C until OD600 = 0.4. The media was brought to 50 µM biotin and protein expression induced with 0.3 mM IPTG overnight at 16°C. The next morning, cells were pelleted, then resuspended in 50 mM Tris pH 8.0, 150 mM NaCl, 4 mM EDTA, 1 mM DTT, 1 mM PMSF, 15 mM MgCl2, DNAse I, and lysed by French press. Lysate was filtered and run on a 5 ml SP HP column (GE Biosciences) with 20 column volumes (cv) of washes and then a gradient of 10% to 100% B for 10 cv (Buffer A: 25 mM NaPO4 pH 7.8, Buffer B: Buffer A+500 mM NaCl). RCC1-containing fractions were pooled and diluted down to 100 mM total salt with cold water and filtered. The protein was then loaded onto a 1 ml Q HP column (GE Biosciences) and run with the same gradient and buffers as the SP HP column. RCC1 fractions were pooled, brought to 10% glycerol, frozen in 10 µl aliquots ∼6 mg/ml in liquid nitrogen, and stored at −80°C.
Human kinesin 1–560 wild-type and kinesin 1–560 (R203K) were purified and labeled by the same protocol. Plasmids were transformed into BL21 competent cells and grown on LB plates overnight. The next morning, several colonies were combined and shaken in 20 ml of LB for 10 min. 4 ml of the bacteria was then added to 1 L of 2xYT [35] with 0.2% dextrose (4 L total) and grown at 37°C until OD600∼0.6. Protein expression was induced overnight at 24°C with 0.1 mM IPTG. The next morning, cells were pelleted, then resuspended in 50 mM NaHPO4 pH 8.0, 250 mM NaCl, 2 mM MgCl2, 20 mM imidazole, 1 mM ATP, 0.2 mM TCEP, 1 mM PMSF, 10 µg/ml LPC, and lysed by French press. Lysate was filtered and run on a 1 ml Nickel HiTrap column (GE Biosciences) with washes of 10 cv of buffer (50 mM NaPO4 pH 7.2, 250 mM NaCl, 1 mM MgCl2, 0.1 mM ATP, 0.2 mM TCEP) containing 20 mM imidazole followed by 20 cv with 50 mM % imidazole. Protein was eluted with 300 mM imidazole. Protein-containing fractions were pooled and diluted to 100 mM total salt with Mono Q Buffer (25 mM PIPES pH 6.8, 2 mM MgCl2, 1 mM EGTA, 0.1 mM ATP, 0.2 mM TCEP). The protein was then loaded onto a 1 ml Mono Q column (GE Biosciences) and eluted with a 0.1 to 1.0 M NaCl gradient over 20 cv. The highest concentration kinesin fractions were pooled and reacted on ice for 2 h with a 40 molar excess of Biotin PEG EZ link maleimide (Pierce). The reaction was then desalted over 2 HiTrap desalting columns (GE Biosciences) into 25 mM Pipes pH 6.8, 400 mM NaCl, 2 mM MgCl2, 1 mM EGTA, 0.1 mM ATP, 0.2 mM TCEP, 20% sucrose, aliquoted, frozen in liquid nitrogen, and stored at −80°C.
Biotinylated BSA was generated by reacting 1.25 ml of Albumin Standard (Thermo Scientific) with a 20× molar excess of EZ Link Biotin-PEO12-NHS (Thermo Scientific) on ice for 1 h. The excess biotin was removed by desalting the protein over three 5 ml HiTrap Desalting columns into PBS. Fractions were pooled, aliquoted, frozen in liquid nitrogen, and stored at −80°C.
A 6xHis-tagged human Ran construct was transformed into BL21 competent cells and grown on LB plates overnight. The next morning, several colonies were combined and shaken in 150 ml of LB for 10 min. 20 ml of the bacteria was then added to 1 L of LB (6 L total) and grown at 37°C until OD600∼0.4. Protein expression was induced by 0.3 mM IPTG and grown at 25°C for 4 h. Cells were pelleted and stored at −80°C. Cells were resuspended in PBS, 1 mM MgCl2, 0.1 mM GTP, 1 mM PMSF, and lysed with a French press. Lysate was filtered and run on a 5 ml Nickel HiTrap column (GE Biosciences) with washes of 10 cv 2% B and 20 cv 10% B. Protein was eluted with 60% B (Buffer A: 50 mM NaPO4 pH 7.4, 500 mM NaCl, 1 mM MgCl2, 0.1 mM GTP, Buffer B: Buffer A+500 mM imidazole). The highest concentration fractions were pooled and desalted over 5 HiTrap desalting columns (GE Biosciences) into XB, 1 mM MgCl2, 0.1 mM GTP. Desalted fractions were aliquoted, frozen in liquid nitrogen, and stored at −80°C.
Human full-length 6x-His-tagged EB1 was cloned into the pAN6 vector (Avidity), which contains a biotin acceptor peptide (GLNDIFEAQKIEWHE) on the N-terminus [34]. EB1 and the E. coli biotin ligase (BirA) were cotransformed into BL21 DE3 competent cells. Cells were outgrown for 1.5 h, added to 1 L of LB [35] with ampicillin and chloramphenicol, and grown overnight at 37°C. The next morning, 150 ml of culture was added to 1 L of fresh LB/amp/chlor (6 L total) and grown at 37°C until OD600 = 0.4. The media was brought to 50 µM biotin and protein expression induced with 0.3 mM IPTG overnight at 16°C. The next morning, cells were pelleted, then resuspended in PBS with PMSF, and lysed by French press. Lysate was filtered and run on a 5 ml Nickel HiTrap column (GE Biosciences) according to the manufacturer's instructions. Briefly, the lysate was loaded onto the column, washed with 20 cv of 100 mM imidazole, and EB1 was eluted with 300 mM imidazole. EB1-containing fractions were pooled and desalted into PBS with 3×5 ml HiTrap deslating columns (GE Biosciences). EB1 fractions were pooled and frozen in 10 µl aliquots in liquid nitrogen, and stored at −80°C.
We empirically determined how much of each protein to add to the beads to obtain the desired ratios by combining different amounts of the proteins with streptavidin resin beads to couple them. After thorough washing, the bound proteins were eluted and analyzed by SDS-PAGE and their ratios determined.
|
10.1371/journal.pgen.1001106 | DNMT3L Modulates Significant and Distinct Flanking Sequence Preference for DNA Methylation by DNMT3A and DNMT3B In Vivo | The DNTM3A and DNMT3B de novo DNA methyltransferases (DNMTs) are responsible for setting genomic DNA methylation patterns, a key layer of epigenetic information. Here, using an in vivo episomal methylation assay and extensive bisulfite methylation sequencing, we show that human DNMT3A and DNMT3B possess significant and distinct flanking sequence preferences for target CpG sites. Selection for high or low efficiency sites is mediated by the base composition at the −2 and +2 positions flanking the CpG site for DNMT3A, and at the −1 and +1 positions for DNMT3B. This intrinsic preference reproducibly leads to the formation of specific de novo methylation patterns characterized by up to 34-fold variations in the efficiency of DNA methylation at individual sites. Furthermore, analysis of the distribution of signature methylation hotspot and coldspot motifs suggests that DNMT flanking sequence preference has contributed to shaping the composition of CpG islands in the human genome. Our results also show that the DNMT3L stimulatory factor modulates the formation of de novo methylation patterns in two ways. First, DNMT3L selectively focuses the DNA methylation machinery on properly chromatinized DNA templates. Second, DNMT3L attenuates the impact of the intrinsic DNMT flanking sequence preference by providing a much greater boost to the methylation of poorly methylated sites, thus promoting the formation of broader and more uniform methylation patterns. This study offers insights into the manner by which DNA methylation patterns are deposited and reveals a new level of interplay between members of the de novo DNMT family.
| The methylation of cytosine bases in DNA represents an extra layer of heritable biological information necessary for regulating gene expression and ensuring genomic stability in mammals. In this paper, we examine the function of the proteins responsible for laying down the initial DNA methylation patterns in the human genome. These proteins, called de novo DNA methyltransferases (DNMTs), comprise two active enzymes, DNMT3A and DNMT3B, and one stimulatory factor, DNMT3L. Our study clearly establishes that DNMT3A and DNMT3B do not methylate DNA at random but rather that they show strong and distinct preferences for their target sites in vivo. These preferences lead to the deposition of unique and reproducible patterns of methylation and may have contributed to shaping segments of the human genome. In contrast, we show that DNMT3L stimulates DNA methylation mostly at sites that are poorly methylated on their own, thus leading to patterns that are more uniform. This modulation is proposed to result from DNMT3L anchoring the DNA methylation machinery onto chromatin, the physiological form under which DNA exists in our cells. This study furthers our understanding of how genomic DNA methylation patterns are established in vivo.
| Cytosine DNA methylation, which is primarily focused at symmetrical CpG sites in mammalian cells, represents a critical epigenetic mark broadly associated with silent genomic regions. Repeated DNA elements such as dispersed transposon-derived repeats or heterochromatin-associated pericentric satellite repeats are heavily methylated, highlighting the primordial role of DNA methylation as a genome defense mechanism [1]. Cytosine DNA methylation is also essential for development [2], [3] and contributes to the regulation of gene expression through differentiation [4]–[6]. Once DNA methylation is established on both DNA strands at a CpG site, it is propagated with high fidelity at each cell division [7]. This stems directly from the fact that hemi-methylated CpG sites, a key intermediate generated by replicating through a fully methylated CpG sequence, are preferentially methylated back to a fully methylated state by the maintenance DNA methyltransferase DNMT1 [8], [9] in association with other interacting factors such as UHRF1 [10], [11]. Thus, DNA methylation profiles represent an important form of epigenetic memory.
Much progress has been made in recent years in our understanding of how cytosine DNA methylation patterns are established during development. This de novo methylation function is assigned primarily to the DNMT3 family of DNA methyltransferases (DNMTs) [9]. This family comprises the two active DNMT3A and DNMT3B enzymes, which are highly expressed at specific developmental times in germ cells and during early development, and mediate genome-wide acquisition of DNA methylation. In vivo, DNMT3A and DNMT3B possess both overlapping and specific targets. DNMT3A is particularly required for the methylation of imprinted genes and dispersed repeated elements, such as retrotransposons, while DNMT3B specializes in the methylation of pericentric satellite repeats [3], [12]–[14]. The DNMT3L protein, a non-catalytic accessory factor, also serves as an important structural and functional accessory factor for almost all types of de novo DNA methylation, particularly in germ cells [15], [16], [17].
The mechanisms by which specific DNA methylation patterns are instructed by DNMT3A, DNMT3B and DNMT3L in the mammalian genome are currently unclear. Multiple studies have indicated that chromatin composition and modification are key in setting the accessibility of certain genomic loci to the DNA methylation machinery. For instance, DNMT3L was proposed to focus DNA methylation away from CpG island promoter regions by discriminating against binding to nucleosomes marked by trimethylation at lysine 4 on histone H3 [5], [18], [19]. Likewise, recent data indicate that the presence of the H2A.Z variant is protective against DNA methylation in the model plant organism Arabidopsis thaliana [20]. Other mechanisms, such as recruitment of the de novo methylation machinery by direct association with various DNA binding proteins [21]–[23] or possibly by small non-coding RNAs [24], [25], are also likely to operate.
In this study we addressed the possibility that the human DNMT3A and DNMT3B enzymes possess an intrinsic preference for certain DNA sequences flanking their target CpG site. This notion is supported by the concept that the catalytic domains of mammalian DNMTs have evolved from bacterial methyltransferases, many of which are sequence-specific modifying enzymes [9]. Moreover, recent genome-wide bisulfite sequencing efforts have revealed clear local sequence preferences for cytosine methylation in A. thaliana, an organism that harbors two de novo DNA methyltransferases distantly related to the mammalian DNMT3A and DNMT3B enzymes [26], [27]. Finally, biochemical approaches using purified DNMTs in in vitro methylation reactions on naked DNA templates have also hinted at the fact that the mammalian DNMT3A and DNMT3B enzymes might possess an intrinsic flanking sequence preference [28], [29]. However, no specific consensus could be readily derived from these studies.
Here we have used a well-described episomal DNA methylation assay [30], [31] to determine whether the full-length human DNMT3A and DNMT3B enzymes show any flanking sequence preference in vivo. For this, we used human HEK293c18 cells, which show little to no endogenous de novo methylation activity, and co-transfected target episome DNA together with expression vectors for DNMT3A and DNMT3B in the presence or absence of the DNMT3L protein. The resulting methylation patterns were then determined at various test regions on the episome using bisulfite methylation sequencing and further validated at two additional targets. Our data clearly indicate that DNMT3A and DNMT3B show significant and distinct flanking sequence preferences and reveal a novel and unexpected role of DNMT3L in modulating DNA methylation pattern formation.
The episomal methylation assay [30] offers a powerful and versatile tool for measuring DNA methylation in human cells in culture. This assay revolves around the use of unmethylated, stably replicating minichromosomes that are transfected in HEK293c18 cells together with expression vectors for the DNMT(s) of interest. Conveniently, HEK293c18 cells show little if any endogenous de novo methylation but efficiently carry out maintenance methylation [30]. Thus, the de novo methylation patterns established on these target episomes by exogenous DNMT(s) expressed in these cells are stably maintained for prolonged periods of time. Methylation patterns can then be detected at nucleotide resolution using bisulfite methylation sequencing [32]. Here we analyzed DNA methylation at four distinct regions carried on episomal constructs. All regions fit a strict operational definition for a CpG island, namely GC content >55% and a ratio of observed versus expected CpG sites ratio >0.8 [33]. Focusing on CpG-rich regions enabled us to maximize the range of sequence flanks analyzed; altogether 271 distinct CpG sites were studied.
To validate these episomal constructs as a good tool for determining the intrinsic sequence preference of the human DNMT3A and DNMT3B enzymes, we wanted to ensure that de novo methylation could not trigger DNMT1-mediated “spreading” effects around pre-methylated sites [34]. If spreading were to occur, it would diminish our ability to discern true de novo activity by the de novo enzymes from DNMT1-mediated activity. To test this, we methylated the pFC19 episome in vitro with the HhaI DNA methyltransferase and transfected the DNA in HEK293c18 cells. After 7 days, the episomal DNA was recovered by Hirt harvest and DNA methylation patterns were determined using bisulfite methylation sequencing. The methylation pattern on the input DNA prior to transfection was also determined for comparison. The data clearly show that methylation patterns were faithfully maintained without significant modification from the initial pattern (data not shown). Thus, DNMT1 does not appear to lead to “spreading” effects in this sytem. We also transfected the pre-methylated episomes together with expression vectors for DNMT3A or DNMT3B to determine if pre-methylated sites might attract the de novo enzymes to their immediate vicinity (“seeding” or “clustering” effects). We determined the methylation patterns and compared them to the patterns obtained with unmethylated episomes and observed no significant changes (data not shown). This indicates that pre-existing CpG methylation does not stimulate DNMT3A or DNMT3B activity, in agreement with another independent study [35].
Finally, since our episomes are generated in E. coli and therefore carry Dam methylation (N6 Adenine methylation at GATC sequences) and Dcm methylation (C5 methylation at the internal cytosine in CCA/TGG sequences), we verified that such non-CpG methylation marks do not modify the CpG methylation patterns laid by DNMT3A or DNMT3B. For this, episome DNA was extracted from dam−, dcm−, and dam− dcm− E. coli strains and used for transfections together with expression vectors for DNMT3A or DNMT3B. The distribution of methylated sites, as judged by Southern blots after digestion with methyl-sensitive restriction enzymes, did not detectably vary (data not shown), indicating that pre-existing non-CpG methylation does not influence the activity of the de novo enzymes. Similar observations have been reported [35]. The presence of 5-methylcytosine at Dcm sites, however, provided us with the ability to track whether a particular DNA strand has been newly synthesized in human cells upon replication of the episome or corresponds to an “old” bacterial DNA strand that was transfected in. Indeed, non-CpG methylation is not maintained in human cells and newly synthesized DNA strands were consistently devoid of dcm methylation.
Methylation of the target minichromosome by DNMT3A was analyzed by bisulfite methylation sequencing at two test regions (Hygro and pBR – See Material and Methods) on both DNA strands. In total, 20,352 CpG sites were sequenced for the DNMT3A sample, representing over a 100-fold coverage of all available sites. For most regions, the methylation at various CpG sites was not uniform but rather showed clear evidence for high and low methylation sites, as evidenced for the pBR Top strand region (Figure 1A). While the overall methylation efficiency on that strand was 15.3%, the methylation was not evenly distributed between all 48 CpG sites, leading to the formation of a clear pattern characterized by high and low efficiency sites. Site # 3, for instance, was methylated with an efficiency of 28.5% while site # 31 was methylated with an 11-fold lower efficiency of 2.6%. The presence of hotspots and coldspots for DNA methylation and the overall pattern of DNA methylation resulting from DNMT3A activity was reproducible in a completely independent biological replicate. For example, Figure 1B shows that when CpG sites were ranked according to their individual methylation efficiencies in each sample, a clear positive correlation is observed between the two independent replicates. This reflects the fact that top methylation sites (low ranks) remain highly methylated in both samples while bottom methylation sites (high ranks) show consistently poor methylation in both samples. Indeed, 6 out of the 10 top CpG sites in sample 1 also belonged to the top 10 sites in sample 2. Conversely, 7 out of the 10 bottom CpG sites in sample 1 also belonged to the bottom 10 sites in sample 2 for that particular region. Similar observations of consistent high and low efficiency sites were made at the Hygro region with individual variation in methylation efficiencies of up to 6-fold (data not shown).
Consistent with the presence of endogenous maintenance DNA replication in HEK293 cells, the methylation patterns observed at each region were mostly symmetrical between the two DNA strands. As shown in Figure 1C, a strong correlation of methylation efficiencies is observed across the two strands of the pBR region. This observation has implications for our ability to properly identify and score the flanking sequences of high and low methylation sites. Indeed, the presence of a methylation hotspot on one strand should lead to the observation of a high methylation site on the other strand due to maintenance methylation. Therefore, some sites that appear as methylation hotspots may not directly correspond to a hotspot but may be located across a hotspot on the other strand. In contrast, this predicts that methylation coldspots should correspond to low methylation efficiency sites on both DNA strands, as observed.
To determine if the high and low methylation sites observed as a result of DNMT3A activity could be explained by a potential flanking sequence preference, we focused on the sequences flanking the 10% most methylated CpG sites and the 10% least methylated CpG sites at the pBR and Hygro regions. For this, all CpG sites within each region were ranked according to their respective methylation efficiencies and the flanking sequences extracted on each side of the target CpG site. The sequences were then systematically aligned with each other either in direct or reverse-complement orientation to identify regions of similarity in each class. In the case of DNMT3A, it rapidly became apparent that methylation hotspots were likely to share a T residue at the −2 position and a C residue at the +2 position from the target CpG sites (Figure 2A). The over-representation of the T and C residues at these positions over the average sequence composition of the regions under study was statistically significant with p-values of 5×10−11 and 2×10−3, respectively. Interestingly, low efficiency sites, which on average were methylated at a 5.3-fold reduced efficiency compared to high sites, also showed statistically significant enrichment for adenine residues at position −2 (Figure 2B), indicating that this position is particularly important for discriminating between a good and bad flank. Examination of up to 12 positions on each side of the target CpG site revealed that positions −2 and +2 were the only positions to show strong statistical significance (data not shown). Similar results were observed when the shorter DNMT3A2 isoform [36] was used as judged from the strong correlation of methylation efficiencies at individual CpG sites (Figure S1). As expected, no enrichment was observed when the entire set of CpG sites analyzed here was aligned (Figure 2C).
To determine if the flanking sequence preference observed for DNMT3A in HEK293 cells corresponds to an intrinsic enzymatic preference, we used the purified catalytic domain of DNMT3A and performed in vitro DNA methylation reactions. The resulting methylation at the pBR region was analyzed by bisulfite sequencing on both DNA strands. As observed in vivo, the methylation patterns indicated clear preference for some sites over others with a 15-fold maximal range between high and low sites (data not shown). When the sequences flanking the 15% top and bottom sites were extracted and aligned, a pattern similar to the one observed in vivo emerged. Top methylated sites (average methylation efficiency 88.2%) tended to carry a T at position −2 and a C at position +2 (Figure 2D). By contrast, the least methylated sites (average methylation efficiency 13.4%) showed an enrichment for A or G residues at the −2 position and a G at position +2 (Figure 2E). Thus, the sequence composition at the −2 and +2 positions appears critical in selecting for a good or bad flank for DNMT3A in vivo and in vitro. Moreover, since in vitro methylation patterns are not compounded with any maintenance methylation, this indicates that the DNMT3A preference was properly assigned in vivo and that this preference represents an intrinsic property of the catalytic site of the enzyme. Interestingly, nearly 25% of all methylated cytosines observed in vitro were found in CpA and CpT contexts, clearly showing that DNMT3A is capable of non-CpG methylation activity, in agreement with previous studies [37], [38]. Highly methylated non-CpG sites also showed an enrichement for a T at the −2 position (data not shown). The high preponderance of non-CpG methylation in vitro is in contrast to the situation observed in vivo as episomal substrates showed little to no non-CpG methylation (data not shown). This is likely due to the fact that methylated non-CpG sites are not maintained upon replication by DNMT1.
The DNMT3B in vivo methylation patterns were next analyzed in the same manner as DNMT3A. In total 20,203 CpG sites were sequenced for the DNMT3B1 sample, again representing over a 100-fold coverage. Detailed inspection of the patterns deposited by DNMT3B at the pBR and Hygro regions revealed that the patterns showed distinct and reproducible high and low frequency methylation sites (Figure 3A). Importantly, the DNMT3B patterns were different than those for DNMT3A with distinct CpG positions corresponding to hotspots and coldspots (compare Figure 3A to Figure 1A). Compared to DNMT3A, DNMT3B showed an even greater discrimination between high and low sites: a 34-fold difference was observed between the highest (29.3% methylation efficiency) and lowest (0.86%) sites on the pBR bottom strand. This indicates that, compared to DNMT3A, DNMT3B preferentially methylates, and avoids, different flanking sequences. As observed for DNMT3A, the overall DNA methylation patterns were symmetrical across both DNA strands (Figure 3B).
To extract the flanking sequence preference for DNMT3B, we focused on the 10% most methylated and the 10% least methylated CpG sites at the two test regions and aligned these sites with respect to the central CpG site. DNMT3B hotspots revealed an enrichment for a T residue at position −1 and a G residue at position +1 with p-values of 2×10−3 and 7×10−5, respectively. Coldspots, by contrast, showed a strong enrichment for a C residue at position +1 with a p-value of 1×10−8 (Figure 4A). Thus, unlike DNMT3A, which discriminates between flanks by the composition of the −2 and +2 flanks, DNMT3B appears to respond mostly to the sequence composition at the −1 and +1 positions.
To validate these observations, we sought to determine whether methylation by DNMT3B would be more prevalent at 5′-CCGG-3′sites, which are recognized by the methylation-sensitive HpaII enzyme, than at 5′-GCGC-3′ sites, which are recognized by the methylation-sensitive HhaI enzyme. Our observation that a G at the +1 position is a hallmark of high efficiency sites while a C in this position is highly enriched in low efficiency sites predicts this outcome. Upon cleavage of our Hirt harvest episomal vector with HpaII and HhaI, the resulting DNA fragments were separated by gel electrophoresis and the DNA methylation patterns were revealed by Southern blotting. As shown in Figure 4B, high molecular weight bands corresponding to the methylation of almost all available sites can be readily observed upon digestion with HpaII. By contrast, we reproducibly failed to detect such high mobility species upon cleavage by HhaI even in the presence of the stimulatory factor DNMT3L, indicating that HhaI sites remained available for cleavage and therefore were unmethylated. This indicates that HhaI sites, which carry a C at the +1 position, are poorly methylated in contrast to HpaII sites, which carry a G at the +1 position, in agreement with our sequencing data. As also predicted by our analysis, methylation of episomal DNA by DNMT3A did not lead to any measurable distinction in the cleavage efficiency by HpaII or HhaI (Figure 4B). These observations therefore validate our sequencing data using an independent method and indicate that most of the variation in methylation efficiencies for DNMT3B is indeed captured by the −1 and +1 positions from the target CpG site since HhaI is insensitive to all other positions. We also independently examined DNA methylation activity for the active DNMT3B2 splice isoform and observed that the flanking sequence preference for DNMT3B2 was essentially unchanged compared to the full-length DNMT3B1 protein (Figure S1).
While the test regions used so far correspond broadly to CpG islands, we wished to validate the preferences observed using sequences directly of human origin. For this, two human CpG islands were cloned into episomal constructs and used as sequence targets.
The first region analyzed corresponded to a 539 bp portion of the imprinted and maternally methylated SNRPN CpG island carrying 37 CpG sites. This region, while GC-rich overall, shows a strong strand asymmetry in the distribution of guanine and cytosines outside of CpG sites such that one strand is highly G-rich and the other highly C-rich, a property referred to as GC skew. Interestingly, the overall methylation efficiencies of the two strands were significantly different (DNMT3A: C-rich 18.5%/G-rich 31.5%; DNMT3B: C-rich 9.7%/G-rich 21.7%) despite the presence of efficient maintenance methylation at other regions tested on the same episomes. This suggests that the de novo enzymes are recruited preferentially to the G-rich strand and/or that the maintenance machinery has difficulty maintaining the methylation patterns at these regions (an intrinsic bias in our ability to detect highly methylated C-rich strands is unlikely as such molecules can be efficiently recovered upon in vitro methylation; data not shown). Inspection of the patterns deposited on both strands revealed that the G-rich strand was also more uniformly methylated (maximal fold difference between the most and least methylated sites: 2.6 and 3-fold for DNMT3A and DNMT3B, respectively) than the C-rich strand (maximal fold difference between the most and least methylated sites: 5.6 and 23-fold for DNMT3A and DNMT3B, respectively).
When focusing on the C-rich strands, which show variation in the patterns, we noted that for DNMT3A (81 independent molecules were analyzed), two of the the top three sites (average methylation efficiency 33%) displayed a T at position −2 and two displayed a C at position +2. By contrast, the bottom five sites (average methylation efficiency 8.9%) were flanked by either an A at position −2 (three cases out if five) or a G at position +2 (2 cases out of five) (Figure S2). Thus, the variation in methylation efficiencies observed at the SNRPN region for DNMT3A recapitulated the preference observed at our test regions. Likewise for DNMT3B (88 independent molecules were analyzed), six out of the seven bottom sites (average methylation efficiency 2.6%) carried a C at position +1, while only one out of the four top sites (average methylation efficiency 20.9%) carried a C at this position (Figure S2).This is again compatible with the observations reported above for test regions.
We also analyzed DNA methylation patterns at the CpG island from the TIMELESS promoter which was cloned in our episomal construct in place of the SNRPN region. In this case, the methylation of 12 CpG sites were investigated by quantitative methylation pyrosequencing. Upon expression of DNMT3A, methylation patterns were remarkably consistent over three independent experiments and showed only little variation: the maximal fold difference in methylation efficiencies between the most and least methylated sites was 1.8-fold (data not shown). Thus little information could be derived. In the case of DNMT3B, however, a highly reproducible pattern could also be detected and the most and least methylated sites showed an average 4.4-fold difference in methylation efficiencies (Figure 5). Strikingly, site #10, the most methylated site in this sequence corresponded to the predicted hotspot TCGG, while the second most methylated site (Site #1) corresponded to an HpaII site (CCGG), also predicted to represent a hotspot. Furthermore, the least methylated site, site #3, mapped onto the predicted coldspot GCGC (HhaI site), while the second least methylated site (Site #9) mapped to the palindromic CACGTG sequence. Sites with intermediate methylation efficiencies did not fit to either consensus motifs for predicted high and low sites. Altogether, this suggests that the sequence preferences derived from test regions do apply broadly to various sequences regardless of their origin.
In the euchromatic portion of the human genome, the CpG dinucleotide is mostly confined to CpG islands. Within these loci, one can distinguish between two classes of islands: “specific” CpG islands serve as promoters and are seldom methylated; “weak” CpG islands are located in the bodies of genes or in intergenic regions and are often methylated and silent [39]. It is tempting to speculate that the relative unmethylated state of “specific” CpG islands might be explained at least in part by an under-representation of DNA methylation hotspots and/or perhaps an over-representation of DNA methylation coldspots. In contrast, one might expect that weak CpG islands might show little evidence for selection of particular sequence motifs.
To evaluate this possibility, we used the R'MES statistical package, which uses Markovian models to evaluate the exceptionality of motif (or ‘word’) frequency in DNA. When looking at 6-letter words under the M1 model, which analyzes a given DNA sequence based on mono- and di-nucleotide frequencies, we first determined that words fitting the NNCGNN consensus showed a much greater range of over- or under-representation in the “specific” set compared to the “weak” set of CpG islands (data not shown). Interestingly, we determined that NGCGCN sites, which are predicted to represent DNMT3B methylation coldspots, showed significant over-representation in the set of “specific” CpG islands, but not in “weak” CpG islands in an M1 model (Figure 6). By contrast, we determined that NTCGGN sites, which are predicted to represent DNMT3B DNA methylation hotspots, tended to be under-represented in the set of “specific” islands (Figure 6). This under-representation was much less visible in the “weak” CpG islands. Interestingly, the most under-represented NNCGNN word in the set of “specific” CpG islands was 5′-TTCGGC-3′, a predicted sequence hotspot for both DNMT3A and DNMT3B. For this word, less than half of the expected occurences could be counted (249 out of 530, respectively), which is associated with a strong statistical significance (p-value<10−34). However, we could not detect further evidence for enrichment or depletion based on the −2 and +2 position of NNCGNN sites. This analysis suggests that the intrinsic flanking preference of DNMT3B, and to a lesser extent DNMT3A, might have contributed to shaping promoter CpG island sequence composition in the human genome.
Having determined the flanking sequence preference of DNMT3A and DNMT3B on their own, we then examined the effect of the DNMT3L protein on the formation of DNA methylation patterns. DNMT3L, as described previously, is a potent stimulatory factor for de novo methylation. However, at first inspection, our data revealed only a moderate 1.5 to 2-fold stimulation of DNA methylation by DNMT3L, which was somewhat lower than previously observed in vivo [40], [41]. Upon closer examination of sequencing data, we noticed a significant difference between DNA strands that had been newly synthesized inside HEK293 cells (characterized by the fact that they lost the original dcm bacterial methylation marks; these strands are thereafter referred to as dcm−) and the original DNA strands that were transfected (these strands carry the dcm bacterial marks and are referred to as dcm+). Namely, we observed that newly replicated DNA strands (dcm−) showed much stronger levels of stimulation by DNMT3L than dcm+ strands, which for most regions examined showed no significant stimulation by DNMT3L (Figure 7). Hence, the average stimulation on dcm+ strands at the pBR322 and Hygro regions for DNMT3A was 1.3-fold, while it was 3.2-fold on dcm− strands. A similar observation was made for DNMT3B. Interestingly, this distinction between dcm− and dcm+ DNA strands only applied when DNMT3L was present; the methylation efficiencies of these strands were very similar when DNMT3A or DNMT3B were considered on their own (data not shown).
This strongly suggests that the effect of DNMT3L is mostly felt on newly replicated DNA strands. There are two hypotheses that could explain this observation. One is that stimulation of DNA methylation by DNMT3L might be mechanistically coupled to DNA replication. The other is that DNMT3L might focus DNA methylation towards well-chromatinized DNA templates, thus biasing the DNA methylation machinery towards newly replicated episomal DNA molecules that have acquired chromatin with fork passage. No evidence exists so far for a coupling between de novo DNA methylation and DNA replication. On the contrary, DNMT3L clearly promotes de novo methylation in non-dividing cell types [42]. Likewise, recent evidence shows that DNMT3L physically binds to chromatin [18], [19].
To distinguish between these two hypotheses, we devised an experiment that allowed us to track DNMT3L stimulation as a function of replication on fully chromatinized episomes. Episomes were transfected in HEK293c18 cells and allowed to propagate under selective pressure through multiple rounds of replication encompassing at least 20 cell generations. At this point the episomes are expected to be fully chromatinized and episomal DNA no longer carried detectable dcm methylation (data not shown). To track the replication status of DNA strands we then transiently transfected a vector expressing a MYC-tagged version of the bacterial Dcm methyltransferase carrying a nuclear localization signal to mark DNA strands in a “pulse” of expression. Using Western blots, we determined that the Dcm methyltransferase was efficienctly expressed 24 hours after transfection and remained at high levels in the cells up until ∼5 days post transfections, at which point expression rapidly declined (data not shown).
Two days post-transfection, we extracted the genomic DNA and verified that the DNA had become marked in that it became extensively resistant to cleavage by EcoRII, a restriction enzyme sensitive to dcm methylation (data not shown). In addition, episomal DNA became re-methylated at dcm sites, as determined by bisulfite sequencing (data not shown). This transient pulse of Dcm expression therefore allowed us to re-mark endogenous DNA, enabling us to track the replication status of DNA strands. Seven days after the Dcm expression vector was first transfected, we then introduced expression vectors for DNMT3A in the presence or absence of DNMT3L. The resulting methylation patterns were determined using bisulfite methylation sequencing another seven days after transfection of DNMT vectors. Our prediction was that if DNMT3L function is coupled to DNA replication, then only the newly replicated dcm− molecules should show stimulation. By contrast, if DNMT3L function is independent from DNA replication but is sensitive to the chromatin status of its target molecules, then both dcm− and dcm+ molecules should show stimulation. Consistent with this second hypothesis, DNMT3L triggered a 6.34-fold stimulation of DNA methylation on dcm+ molecules and a 5.76-fold stimulation on dcm− molecules. This indicates that, as expected, DNMT3L stimulation is not dependent upon replication, and suggests that DNMT3L directs DNA methylation to fully chromatinized templates.
While it is clear that DNMT3L stimulates the catalytic activity of its partners, it remains to be determined if the stimulation is accompanied by any change in the flanking sequence preference of DNMT3A or DNMT3B. To determine this, we ranked CpG sites according to their individual methylation efficiencies in the presence and absence of DNMT3L and compared the ranks between the two categories. DNMT3L, while it strongly stimulated DNA methylation by DNMT3A on dcm− strands, did not alter the rankings of high and low efficiency sites (Figure 8A left). Similar results were observed for DNMT3B (Figure 8A right). In addition, DNA methylation patterns deposited in vitro by the full length DNMT3A2 enzyme in complex with DNMT3L clearly showed evidence of selection for suitable flanks at the −2 position (Figure S3).Thus, it seems likely that DNMT3L does not alter the intrinsic sequence preference of either active enzyme, consistent with the notion that DNMT3L binds to DNA poorly [43].
We noticed, however, that the spread in the individual methylation efficiencies between CpG sites was noticeably reduced in the presence of DNMT3L, thus leading to the establishment of more uniform patterns characterized by long tracks of contiguously methylated sites. For instance, the range of individual methylation efficiencies observed in the presence of DNMT3A on the pBR bottom strand was 28-fold (Figure S4). In sharp contrast, the corresponding range of methylation efficiencies at these sites was 4.3-fold in the presence of DNMT3L, which represents a significant reduction (Figures S4 and S5). Similarly, when this analysis was performed for DNMT3B at the same region, the maximal difference between individual sites shifts from 34-fold in the absence of DNMT3L to 12.5-fold in its presence, a statistically significant reduction (Figure S5). This phenomenon is explained by the fact that the sites that show the strongest stimulation by DNMT3L correspond to those that were the least methylated by DNMT3A or DNMT3B on their own. This is clearly illustrated in Figure 8B by an inverse relationship between the fold stimulation afforded by DNMT3L at each individual CpG site and individual methylation efficiencies in the presence of DNMT3A or DNMT3B alone. The sites with the lowest initial methylation showed a striking 32–35-fold increase by DNMT3L while the sites with the greatest initial methylation only showed a 2–3-fold increase. Altogether this shows that while DNMT3L does not alter the intrinsic sequence preference of DNMT3A and DNMT3B, its stimulatory effect is most felt at the sites with the initial lowest methylation efficiency. DNMT3L therefore attenuates the intrinsic sequence preference of DNMT3A and DNMT3B, resulting in the deposition of more uniform methylation patterns.
One important unanswered issue surrounding the establishment of DNA methylation patterns relates to the individual contribution of each member of the de novo DNA methyltransferase family to the actual formation of the patterns in vivo. Here, we took advantage of an episomal assay and examined de novo methylation mediated by DNMT3A or DNMT3B (with and without DNMT3L) in human cells using extensive bisulfite methylation sequencing. This assay offers several advantages. The cell line that we used, which has little or no endogenous de novo methylation activity [44], allowed us to study the activity of exogenously expressed full-length human DNMT3A and DNMT3B separately, an otherwise impossible task when analyzing genomic DNA methylation profiles. Likewise, using unmethylated episomes as substrates for DNA methylation allowed us to ensure that the patterns deposited by each enzyme were relatively unaffected by pre-existing parameters such as chromatin compaction, composition, and modification states. Such parameters are unavoidable when studying genomic methylation patterns and strongly compound the activity of de novo DNMTs. While representing a simpler substrate than genomic DNA, the episomes used here are biologically relevant in that they are self-replicating, acquire chromatin, and undergo maintenance DNA methylation and epigenetic silencing [30], [45]. In addition, episomal de novo methylation requires the SNF2-family chromatin remodeling factor Lsh and responds to the DNMT3L stimulatory factor just as observed at endogenous loci [40], [46]. Therefore, this assay offers a useful window into the intrinsic de novo activity of each enzyme in vivo.
Our data indicate that the human DNMT3A and DNMT3B enzymes instruct the deposition of unique DNA methylation patterns. These patterns are characterized by clear and reproducible high and low methylation sites distinguished by greater than 10-fold differences in individual methylation efficiencies. This indicates that DNMT3A and DNMT3B do not methylate DNA at random. For DNMT3A, the overall difference in methylation efficiency between the top 10% most methylated sites and the bottom 10% least methylated sites was 5.3-fold (Figures 1 and 2). This difference was 8.3-fold for DNMT3B, which consistently appeared more selective than DNMT3A in its DNA methylation patterns (Figure 3 and 4). The selectivity of de novo methylation measured here for human DNMT3A and DNMT3B is in agreement with data from A. thaliana for which hotspots and coldspots varied up to 13-fold in methylation efficiency depending on sequence context [26], [27]. Sequence alignments revealed that specific residues were significantly over-represented at particular positions flanking “hot” or “cold” CpG sites. Importantly, the motifs revealed by such alignments were (i) reproducible across independent biological replicates; (ii) reproducible across active isoforms of DNMT3A and DNMT3B; (iii) derived from the alignment of multiple, carefully selected, hotspots and coldspots picked from over 270 distinct CpG sites originating from several test regions of various provenance; (iv) recapitulated upon analysis of DNA methylation patterns in vitro for DNMT3A; and (v) validated by an independent enzyme-based assay for DNMT3B. Altogether, this indicates that these motifs represent high confidence assignments.
The motifs associated with these hotspots and coldspots showed a clear and consistent “mirror image” pattern of enrichment at specific positions. For instance, in the case of DNMT3A, highly methylated CpG sites showed a significant over-representation of a T at position −2, while poorly methylated sites showed enrichment for an A at this position. To a lesser degree, hotspots tended to show over-representation for a C at position +2 while coldspots tended to carry a G at this position (Figure 2). A similar observation was made for DNMT3B: hotspots were significantly enriched for a G at position +1, while coldspots displayed a C (Figure 4). The observation of such patterns of reciprocal enrichment between hotspots and coldspots strongly suggests that the positions identified in our study are key in the selection of good or bad flanks. In the case of DNMT3B, the observation that HhaI sites (GCGC) represent cold sites compared to HpaII sites (CCGG) (Figure 4), additionally suggests that most of the variation in DNA methylation efficiency was captured by the −1 and +1 positions since these restriction enzymes are not sensitive to variation outside their recognition site. Altogether, our data indicate that DNMT3A and DNMT3B discriminate between good and bad flanks by responding to the sequence composition at distinct positions around the CpG site. While DNMT3A responds to the composition at the −2 and + 2 positions, DNMT3B mediates its selection through the −1 and +1 positions. This likely reflects intrinsic differences in the catalytic properties of DNMT3A and DNMT3B and suggests that DNMT3A and DNMT3B, despite their strong amino acid conservation over the catalytic domain, contact DNA around the target CpG site differently.
Interestingly, the −2, −1, +1 and +2 positions flanking the CpG site were the only positions to show statistically significant deviations in their sequence composition. These observations are in agreement with data from A. thaliana for which selection for a good or bad flank was essentially mediated through the two positions adjacent to the target site [26], [27]. Our analysis did not uncover any significant relationship between CpG spacing and DNA methylation efficiency and, sites separated by 8–10 bp did not appear more efficiently methylated. In contrast, we observed several instances of strong methylation hotspots that were not flanked by any CpG site within 8–10 bp. Likewise, we observed in a few instances that the presence of a strong methylation hotspot did not necessarily translate into similarly high methylation efficiency for a neighboring CpG site located 8–10 bp further (data not shown). This suggests that, at least in the context of our experimental system, the proposed relationship between CpG spacing and DNA methylation efficiency [47] may not apply or may be compounded by other effects. When compared to prior studies examining possible site-preference by using purified DNMT3A or DNMT3B proteins in in vitro methylation assays, our results are in close agreement with data from Lin et al., (2002) [29]. In this study, the authors reported that the full length murine DNMT3A protein shows elevated activity at sites carrying pyrimidines at positions −2 and +1. Such preference is in remarkable agreement with our in vitro data using human DNMT3A (Figure 2 and Figure S3) and is consistent with our in vivo evidence. Our data are more difficult to reconcile with the data of Handa et al. (2002) who used murine enzymes either full-length or in a truncated C-terminal form and determined that both DNMT3A and DNMT3B share a common preference for AT-rich flanks and for certain palindromic sequences [28]. In our study, flanks such as 5′-ACGT-3′ or other combinations of A/T bases at the −1 and +1 positions, were consistently found around or slightly below the median in terms of DNA methylation efficiencies. This could represent a difference between murine and human enzymes or between the experimental systems used.
Overall, our study demonstrates that the DNMT3A and DNMT3B de novo DNMTs possess clear and distinct flanking sequence preferences in vivo. Such preferences, while clearly significant, remain sufficiently relaxed that they are compatible with the methylation of a wide variety of CpG sites, as observed in the genome. It should also be noted that only a portion of all possible flanks have been examined here and that the ultimate identities of the DNMT3A and DNMT3B signature motifs might evolve upon surveying a more complete sequence space. However, several lines of evidence suggest that the signature motifs described here may have predictive value. For instance, we showed that promoter-associated CpG islands showed a depletion for CpG sites corresponding to predicted DNMT3B hotspots and an enrichment for predicted coldspots. In contrast, a distinct set of CpG islands which tend to associate with gene bodies and tend to be methylated did not show evidence for such selection (Figure 6). This suggests that the intrinsic sequence preference of DNMT3B may have contributed to shaping the composition of CpG island promoters to favor the maintenance of a state devoid of DNA methylation. Likewise, we note that recent analysis of the complete human methylome revealed that a T at the −2 position was enriched in high methylation sites in the CHG and CHH context, a type of non-CG methylation that is only observed in pluripotent embryonic stem cells that are characterized by high de novo methyltransferase activity [48]. This enrichment is consistent with our motif assignments and with prior studies which implicated DNMT3A in non-CG methylation activity [37], [38].
In this study, we also investigated the effect of the DNMT3L protein on the deposition of DNA methylation patterns. Two main novel findings emerged. First, DNMT3L appears to direct de novo methylation towards well-chromatinized DNA templates. This was observed initially as a bias in the DNMT3L stimulatory effect in favor of newly replicated DNA strands (Figure 8). A strand-tagging experiment, however, allowed us to demonstrate that this bias was not due to a direct coupling between DNMT3L-mediated DNA methylation and DNA replication. This conclusion is expected from the fact that DNMT3L mediates the deposition of DNA methylation patterns in non-replicating cell types during germ cell development [42]. We therefore suggest that DNMT3L may require properly chromatinized DNA substrates for its function. This proposal is consistent with the fact that DNMT3L binds to histones directly [18], [19]. In that context, the “replication” bias we observed for DNMT3L likely reflected the necessity for replication-coupled nucleosome deposition to occur on newly transfected episomes. Interestingly, DNMT3A and DNMT3B on their own did not show any “replication” bias even though evidence clearly suggests that these proteins bind to nucleosomes [49], [50]. This suggests that DNMT3L imposes an even stricter requirement for well-chromatinized substrates onto the process of de novo DNA methylation. Our second finding showed that while DNMT3L does not appear to affect the intrinsic sequence preference of DNMT3A and DNMT3B, its stimulatory effect is not felt uniformly across CpG sites. On the contrary, our analysis revealed a striking inverse relationship between the stimulation afforded by DNMT3L and the initial DNA methylation efficiency by DNMT3A or DNMT3B (Figure 8). Such an effect could again result from the ability of DNMT3L complexes to associate with chromatin, thus favoring the occupancy of DNA by DNMT3A and DNMT3B. An increased DNA dwell time would greatly increase the likelihood that a methylation coldspot will become methylated without strongly affecting the outcome at a rapidly methylated hotspot. This proposal is consistent with the observation that expression of DNMT3L appears to attenuate the impact of intrinsic flanking sequence preferences of DNMT3A and DNMT3B, lowering the range of individual methylation efficiencies between CpG sites (Figure S5) and triggering the deposition of more uniform patterns characterized by longer methylation tracts (compare Figures S3 and S4). This is in agreement with the in vivo function of DNMT3L, which ensures that its multiple target regions (interspersed repeats, satellite repeats, differentially methylated imprinted regions and other chromosomal regions; [15]–[17], [51], [52]) are fully methylated over long blocks of DNA sequence. In that context, it is interesting to note that the drastic reduction of DNA methylation observed in the absence of DNMT3L at imprinting centers may reflect, at least in part, the possibility that such sequences are strongly enriched in methylation coldspots. As discussed above for the SNRPN region studied here, imprinting centers overlapping CpG islands tend to exhibit strong GC skew (P.A.G. and F.C., unpublished data). Our data suggests that the C-rich strand of such regions may be particularly difficult to methylate, thus rendering the action of DNMT3L all the more critical at these regions. Altogether, our study reports that the catalytic activity of DNMT3A and DNMT3B show significant and distinct flanking sequence preference in vivo and suggests that the ability of DNMT3L to bind to chromatin, in addition to its ability to stimulate the catalytic activity of DNMT3A and DNMT3B, are key to its biological function.
Full-length human DNMT3A, DNMT3A2, DNMT3B (the DNMT3B1 isoform was used throughout, unless indicated) and DNMT3L proteins were expressed in HEK293c18 cells using previously described vectors [41].
The E. coli dcm methyltransferease gene (GenBank accession number: YP_853012) was amplified from E. coli genomic DNA (DH10B strain) with a forward primer containing an in-frame EcoRI site (underlined) (DcmFOR: 5′-TTTTTTGAATTCATGCAGGAAAATATATCAGT-3′) and a reverse primer containing a BamHI site (underlined) located immediately after the stop codon (DcmREV: 5′-TTTTTTGGATCCTTATCGTGAACGTCGGCCAT-3′). The amplified dcm PCR fragment was then digested with EcoRI and BamHI, cloned into the corresponding sites of pcDNA3/Myc [41] and sequence verified. A nuclear localization signal (NLS) was subsequently cloned into the EcoRI site in frame using two annealed oligonucleotides: 5′-AATTCCCCAAGAAAAAGAGGAAAGTCC-3′ and 5′-GGGGTTCTTTTTCTCCTTTCAGGTTAA-3′. The resulting construct, pcDNA3/Myc-dcm, expresses an N-terminally MYC-tagged version of the E. coli Dcm methyltransferase carrying a functional NLS.
The pFC19 target episome was used as a methylation target and has been previously described [40]. pFC19 contains the EBNA1/OriP replication system derived from the Epstein-Barr virus and can be stably maintained in mammalian cells. It carries a 940 bp fragment from the differentially methylated region of the human SNRPN CpG island. The first two regions to be analyzed corresponded to sequences present on the episomal backbone, namely: (1) a ∼500 base-pair (bp) region of the pBR322 backbone carrying 48 CpG sites; (2) a ∼500 bp of the Hygromycin (Hygro) resistance gene carrying 47 CpG sites. A ∼300 bp region of the SNRPN region carrying 23 CpG sites was also analyzed. An additional ∼500 bp region from the human TIMELESS CpG island promoter was also cloned instead of the SNRPN sequence and analyzed. All sequences are available in Text S1.
The HEK293 EBNA1 cell line (293c18, ATCC) was used in all experiments and grown under standard conditions. Transfections were performed using either the calcium phosphate method or Turbofect (Fermentas). For each expression vector or episome, 500 ng of DNA was used per well of a 6-well plate. Cells were allowed to grow for 2–3 days after transfection before being transferred to a 100-mm diameter plate. Upon reaching confluence (6–7 days), cells were harvested for episomal DNA extraction according to the Hirt method [53]. No selection was applied. All experiments were conducted at least in duplicate.
For experiments involving a stably replicating pFC19, the episome was first introduced into HEK293c18 cells and the cells were kept under selective pressure (200µg/ml Hygromycin) for over 20 cell divisions. At this point, the pcDNA3/Myc-dcm expression vector was transfected and expression of Dcm methyltransferase was determined every day post-transfection by Western blot using an Anti-MYC antibody (Sigma). Dcm was clearly expressed as early as 24 hours after transfection and expression remained strong for five days, at which point it dropped rapidly (data not shown). The Dcm methyltransferase was clearly active as judged from the fact that genomic DNA extracted 5 days post-transfection was almost entirely resistant to EcoRII, an enzyme that recognizes CC(A/T)GG sites and is sensitive to dcm methylation. DNA from untransfected cells, by contrast, was extensively cleaved (data not shown). Likewise, episomal DNA harvested seven days post Dcm transfection clearly carried dcm methylation as seen by bisulfite methylation sequencing (data not shown). Seven days after transfection with pcDNA/Myc-dcm, DNMT expression vectors in appropriate combinations were introduced and the cells were allowed to grow for another 4–5 days until confluent, at which point episomal DNA was harvested.
In vitro methylation by the HhaI methyltransferase (New England Biolabs) was performed as recommended by the supplier and verified by restriction enzyme digestion. In vitro methylation by DNMT3A was performed using purified recombinant Maltose-Binding Protein (MBP)-tagged DNMT3A catalytic domain (residues 590–912 of human DNMT3A). Reactions were performed using 1µM MBP-DNMT3A and 250 ng of pFC19 DNA in the presence of 100 µM S-adenosyl-L-methionine. The reactions were incubated for 2 hours at 37°C, at which point the proteins were removed by Proteinase K treatment followed by phenol-chloroform extraction and ethanol precipitation.
Except when indicated, bisulfite methylation sequencing [32] was used systematically to determine the methylation patterns deposited by human de novo DNA methyltransferases. For this, the Hirt DNA was first digested with PstI (New England Biolabs) to linearize the DNA or, when desired, by EcoRII (Roche) to enrich for molecules with dcm methylation. Cleavage was followed by sodium bisulfite treatment as described [32] or using the EZ DNA methylation-direct kit (Zymo research). Both strands of DNA were subsequently PCR amplified from different regions of the episome (primers available upon request) and the resulting PCR fragments were cloned using the TopoTA cloning kit (Invitrogen). Single colonies carrying individual DNA molecules were then picked and plasmid DNA sequenced. The overall efficiency of bisulfite conversion in this study was 99.1%. DNA methylation was also analyzed using methylation sensitive restriction enzymes and Southern blot analysis, as described [40]. In the case of the TIMELESS and RNF168 sequences, bisulfite treatment was combined to pyrosequencing in order to extract quantitative and unbiased methylation patterns [54]. Pyrosequencing analysis was conducted by EpigenDx (Worcester, MA).
To handle and analyze the large amount of bisulfite sequencing information generated in this study, we implemented in-house software programmed in Visual Basic running under a Microsoft Excel environment. The software input consists of a typical ClustalW-type multiple alignment of trimmed sequence data (i.e.. the sequence corresponding to the region under analysis stripped of flanking vector sequence). From this, the software automatically computes the conversion efficiency for each molecule and filters any molecule below a user-defined threshold (no less than 95% conversion in all cases). It then calculates the distribution of methylated and unmethylated CpG sites and reports the data as a standard graph, as shown in Figure 2. The software also calculates the overall methylation efficiency for each DNA molecule and for each CpG site across the analyzed sample. This allows us to rank the various CpG sites according to their individual methylation efficiencies and to extract and align the sequences flanking each CpG, focusing on the top 10% most methylated sites and bottom 10% least methylated sites or any other user-defined portion of the distribution. A statistical test for the enrichment of a residue at any given position above what is expected from the average composition of the sequence being considered is also built-in using a Chi-square test. The statistical significance of enrichment is reflected by a P-value which is calculated from the distribution of Chi-square values. This software is available upon request. Enrichment plots were generated using the WebLogo application package [55].
The sequence sets used here correspond to all “specific” CpG islands on human chromosome 1 as defined by Bock and colleagues [39] and accessed from the hg18 build of the UCSC Human Genome Browser (representing a total of 1,033 islands and approximately 1 megabase of DNA sequence). The set of “weak” CpG islands was obtained from the same source but corresponded to CpG islands that have no overlap with specific or balanced CpG islands as defined by Bock and colleagues (a total of 906 islands representing approximately 0.45 megabase of DNA sequence). To evaluate the exceptionality of motif frequencies, we used the R'MES software (http://genome.jouy.inra.fr/ssb/rmes/).This software uses Markovian models to compute the expected distribution of given sequence motifs in a sequence and compares it to actual observations. The score reflects the over- or under-representation of motifs under the model being used.
|
10.1371/journal.pgen.1001315 | Correlated Evolution of Nearby Residues in Drosophilid Proteins | Here we investigate the correlations between coding sequence substitutions as a function of their separation along the protein sequence. We consider both substitutions between the reference genomes of several Drosophilids as well as polymorphisms in a population sample of Zimbabwean Drosophila melanogaster. We find that amino acid substitutions are “clustered” along the protein sequence, that is, the frequency of additional substitutions is strongly enhanced within ≈10 residues of a first such substitution. No such clustering is observed for synonymous substitutions, supporting a “correlation length” associated with selection on proteins as the causative mechanism. Clustering is stronger between substitutions that arose in the same lineage than it is between substitutions that arose in different lineages. We consider several possible origins of clustering, concluding that epistasis (interactions between amino acids within a protein that affect function) and positional heterogeneity in the strength of purifying selection are primarily responsible. The role of epistasis is directly supported by the tendency of nearby substitutions that arose on the same lineage to preserve the total charge of the residues within the correlation length and by the preferential cosegregation of neighboring derived alleles in our population sample. We interpret the observed length scale of clustering as a statistical reflection of the functional locality (or modularity) of proteins: amino acids that are near each other on the protein backbone are more likely to contribute to, and collaborate toward, a common subfunction.
| Genes are templates for proteins, yet evolutionary studies of genes and proteins often bear little resemblance. Analyses of gene evolution typically treat each codon independently, quantifying gene evolution by summing over the constituent codons. In contrast, studies of protein evolution generally incorporate protein structure and interactions between amino acids explicitly. We investigate correlations in the evolution of codons as a function of their distance from each other along the protein coding sequence. This approach is motivated by the expectation that codons near each other in sequence often encode amino acids belonging to the same functional unit. Consequently, these amino acids are more likely to interact and/or experience similar selective regimes, introducing correlation between the evolution of the underlying codons. We find codon evolution in Drosophilids to be correlated over a characteristic length scale of ≈10 codons. Specifically, the presence of a non-synonymous substitution substantially increases the probability of further such substitutions nearby, particularly within that lineage. Further analysis suggests both functional interactions between amino acids and correlation in the strength of selection contribute to this effect. These findings are relevant for understanding the relative importance of different modes of selection, and particularly the role of epistasis, in gene and protein evolution.
| There has been an ongoing debate over the past few decades about the processes underlying protein evolution [1]–[5]. The neutral theory [1] posits that protein evolution is chiefly governed by the fraction of newly arising mutations that are not detrimental enough to be removed by natural selection. However, recent population genetic analyses of closely related Drosophila species suggest that protein divergence between species is substantially in excess of the neutral model's predictions [6], [7]. Intriguingly, this protein divergence excess is consistent with an important role for positive selection in protein evolution [5], [8], [9], although the contribution of weakly deleterious mutations to this pattern is still debated [10], [11].
The dramatic shift in our view of the processes driving protein evolution in Drosophila highlights the deficiency in our understanding of the mechanisms responsible for the observed protein divergence excess. One reason for this deficiency is the explicitly sequence-based nature of the population genetic analyses used to describe the excess divergence. These methods were developed for the analysis of linear sequences of independently evolving amino acids, and quite generally ignore the fact that most proteins fold into complex three-dimensional structures, held together by interactions between amino acids and between amino acids and the surrounding medium. Protein function depends critically on this folded structure, e.g. the arrangement of specific amino acids at the active site of an enzyme [12]. This is reflected in protein evolution; both the structure and the function of homologous proteins are remarkably conserved over long times, even while primary sequences substantially diverge [13]. The maintenance of protein structure is possible because evolution preserves structurally important interactions, such as favorable biochemical interactions between amino acids in physical contact [14]. This preservation of structurally important interactions affects sequence-based analyses; the preferred state and variability of an amino acid will depend on amino acids elsewhere in the protein [15].
This study is motivated by the desire to more closely integrate protein structure and function into sequence-based inferences of selection. Correlations between substitution patterns and protein structure have yielded insights over many years, from the slower divergence of protein active sites [1], [16] to recent results indicating a correlation between estimates of positive selection and secondary structure [17]. Work demonstrating the evolutionary consequences of interactions inferred from RNA structure [18]–[20] supported the application of sequence-based inference of functional interactions to proteins, where functional interactions are difficult to identify even when structure is known [21]. Under the assumption that functionally interacting residues coevolve, interactions can be identified if enough evolutionary trajectories can be sampled. In practice this has meant multi-alignments across many species of large protein families [22]–[25], but alignments within populations of the highly mutable HIV have also been used [26], [27]. These methods have been successfully used to identify pair-wise interactions between residues that contribute to protein function. As an example, the inclusion of interactions inferred from a multi-alignment was shown sufficient to produce a stable fold [28].
Here we develop a complementary approach intended to probe the level of influence interactions have on protein evolution. Instead of focusing on a single protein and specific pairs of interacting residues, we shall aggregate evolutionary information across proteins and use the increased statistical power to look for generic patterns. Specifically, we investigate the correlations in the substitution processes at residues a given distance from each other along the protein backbone, averaged over many proteins of D. melanogaster. Our rationale is as follows: residues that are near in the primary protein sequence are also likely to be near in the folded protein (Figure 1A) and therefore more likely to interact physically and/or belong to the same protein domain. Consequently, if correlated evolution in proteins is common, it should be detectable by an increase in evolutionary correlation between residues nearby in sequence, for which physical interaction in the folded protein is more likely. While we will be unable to identify particular interactions, our approach will be informative about the overall level of influence interactions have on the evolution of proteins.
We find that amino acid substitutions cluster together on the protein sequence, i.e. amino acid substitutions are more frequent nearby other such substitutions. The strength of this effect decays exponentially with the separation between the residues along the protein sequence, with a characteristic length scale of about codons. We observe this clustering phenomenon in substitutions between D. melanogaster and several sister species (Figure 1B) as well as in polymorphisms within a Zimbabwean population sample of D. melanogaster. Clustering is absent when considering synonymous substitutions, implicating selection as the root cause. Furthermore, clustering is stronger between substitutions that arose along the same branch of the evolutionary tree than between substitutions that arose in different branches, and nearby derived alleles tend to cosegregate in our population sample. Additionally, pairs of substitutions within codons of each other that arose in the same lineage have a significant tendency to cause compensatory changes to the total charge of the protein. These lines of evidence lead us to conclude that epistasis between amino acid substitutions contributes significantly to clustering, and the substitution process as a whole.
The 12 Drosophilid genomes resource [29] serves as the primary data source in this study. We used this resource to identify protein coding sequence substitutions between D. melanogaster (Dmel) and several sister Drosophilids: D. sechellia (Dsec), D. simulans (Dsim), D. yakuba (Dyak), D. erecta (Dere), D. ananassae (Dana) and D. pseudoobscura (Dpse) available at http://rana.lbl.gov/drosophila/ (Figure 1B). Substitutions were ascertained from nucleotide alignments of the reference genomes produced by the blastz algorithm [30], and available from UCSC [31] at ftp://hgdownload.cse.ucsc.edu/goldenPath/dm3/.
Our goal here is to understand how correlation between the substitution processes at different residues is affected by the distance between those residues along the protein sequence. To this end we introduce the conditional probability function (cPDF), which we denote and define as the probability of there being a substitution of type at sequence position conditioned on the existence of a substitution of type at sequence position . To assess, for example, whether the probability of a synonymous divergence (DS) is affected by the presence of a non-synonymous divergence (DN) some distance away, we can estimate and compare it to the overall level of synonymous divergence.
cPDFs are estimated from sets of aligned coding sequences by averaging over all instances of the focal substitution in the aligned sequences (Methods). Since we are particularly interested in the functional dependence of cPDFs on the distance from the focal substitution we will normalize cPDFs by their asymptotic value (Methods). Note that we will always be measuring distance in terms of codons rather than nucleotides, as this is the natural unit of distance in a gene. Figure 2A shows three of these normalized cPDFs, , , and , estimated from the species comparison of D. melanogaster and D. yakuba.
Amino acid substitutions are not distributed uniformly along the protein sequence. The cPDF for non-synonymous substitutions, , is significantly peaked around in every species comparison we consider. This peak describes the tendency of non-synonymous substitutions to ‘clump together’ on the protein sequence, a phenomenon we call clustering. The shape of the clustering peak is well-fit by a decaying exponential with a characteristic length scale of about 10 codons. In sharp contrast, the cPDFs involving synonymous substitutions, and , have no clustering peak, indicating that synonymous substitutions are distributed uniformly along the protein sequence. The difference between non-synonymous and synonymous clustering is highly significant, the sampling p-value is essentially zero (, chi-square test).
The magnitude of clustering is large. The nearest neighbor of a codon with a non-synonymous substitution is roughly twice as likely to also have such a substitution than would otherwise be expected. The impact of clustering extends well beyond the nearest neighbor, and is appreciable out to a distance of at least codons from a focal non-synonymous substitution. We quantify the total magnitude of clustering by defining the ‘clustering count’ as the difference between the expected number of substitutions of type in the codons downstream of a focal substitution of type and the expected number in a codon sequence segment distant from the focal substitution (Methods). More plainly, is the number of extra substitutions you find in the vicinity of an substitution because substitutions cluster instead of being distributed uniformly along the sequence. Graphically, is the area under the clustering peak (and above the asymptotic value) of the normalized cPDF , multiplied by the overall density of substitutions.
We are particularly interested in , which we will simply denote . The shape of is very consistent between the different species comparisons tested, but the clustering count is not because it depends not only on , but also on the density of substitutions between the species being compared. ranges from in the D. melanogaster versus D. sechellia alignment to in the D. melanogaster versus D. ananassae alignment, as seen in Figure 3A. increases linearly with (the fraction of substituted amino acids), this is consistent with a clustering pattern that remains constant as divergence increases with time.
Clustering between nearby non-synonymous substitutions is strongly supported by the data, but it is not a priori clear whether it is the separation of amino acids along the protein backbone, or the distance in base pairs along the genome, that matters. To discriminate between these possibilities we repeated the correlation analysis including only those pairs of residues which spanned an intron. As a result the genomic separation between codons had a median increase of bp ( codons) and a minimum increase of bp ( codons), while separation between the encoded amino acids along the protein backbone was unchanged. As shown in Figure 2B, the cPDFs estimated from these intron-spanning pairs of codons correspond closely with those estimated within exons, when separation along the protein backbone (exonic distance) is used in the estimation. We conclude that the clustering length scale is set by the distance along the protein backbone, not along the genome.
Remarkably, the clustering between amino acid substitutions is not limited to substitutions between species. It is also apparent among polymorphisms within a population sample of D. melanogaster (Methods). Figure 2C shows the estimated cPDFs between synonymous and non-synonymous polymorphisms ( and ). The cPDFs estimated from polymorphisms are much noisier because our population sample sequencing spans only kb of coding sequence, as compared to Mb for the divergence data. Nevertheless, we find clustering between polymorphisms analogous to that between substitutions: non-synonymous polymorphisms cluster significantly (, chi-square test), while synonymous polymorphisms do not.
We tested for potential relationships between clustering and a number of genetic properties by estimating on subsets of the full set of coding sequences stratified by the property in question. Clustering is robust in the sense that it is not substantially affected by many of the properties we tested, such as chromosome (including autosome versus X), recombination rate and the level of gapping in the alignment (Figures S2, S3, S4). We did find a systematic relationship between the GC content of coding sequence and clustering; higher GC content correlates with stronger clustering (Figure S5).
A notable factor that influences clustering is the level of constraint under which a gene evolves, which we estimate by the fraction of substituted amino acids . Amino acid substitution are more clustered in constrained genes than they are in unconstrained genes, i.e. has a larger clustering peak when it is estimated from highly constrained (low ) coding sequences, see Figure 3B. In the inset of Figure 3B we have plotted the estimated from each subset of coding sequences against the average of that subset. It is useful to compare this plot to the one in Figure 3A, which also is a plot of versus . The difference between these plots is that in panel A effectively measures divergence time and scales linearly with , while in the inset of panel B tracks the level of constraint and is strongly sublinear in . In fact, once constraint relaxes past a certain point, becomes roughly constant. This relationship suggests that substitutions in constrained genes occur in tight clusters, and that as constraint lessens the additional substitutions which accrue do so uniformly along the sequence.
Non-selective mechanisms cannot account for both significant non-synonymous clustering and the absence of synonymous clustering. Having ruled out non-selective mechanisms, we now consider potential selective mechanisms that could cause amino acid substitutions to cluster. Perhaps the simplest explanation for clustering is that proteins have short segments, such as unstructured loops, that are under reduced purifying selection. These weakly constrained segments experience locally increased rates of amino acid substitution, which we then observe as clustering in both divergence and polymorphism data. There are also several ways in which positive selection could cause clustering. Clustering could be the result of localized ‘adaptive bursts’, i.e. functional modules in which multiple independently adaptive substitutions became available (perhaps due to a changed environment). Because amino acids close on the protein backbone are more likely to be in the same module, the resulting burst of adaptive substitutions would be clustered on the sequence. Amino acids that are close along the chain are also more likely to physically interact, even after protein folding. As a consequence, the fitness effect, and hence evolutionary fate, of nearby substitutions could be contingent on one another (i.e. epistasis). In particular we might imagine common compensatory interactions between nearby substitutions, although all synergistic interactions would contribute to clustering. Finally, another potential mechanism is hitchhiking. In this scenario mildly deleterious amino acid polymorphisms are driven to fixation by the selective sweep of a linked allele, resulting in clustered substitutions. We will now attempt to disentangle the relative contributions of these different selective scenarios.
We can polarize substitutions by the lineage on which they arose using an outgroup and then repeat our correlation analysis for pairs of substitutions which arose in the same lineage and for pairs which arose in different lineages (Methods). This allows us to begin to distinguish between potential selective mechanisms of clustering. If spatial heterogeneity in the strength of purifying selection is responsible for clustering we expect equal clustering within and between lineages, since in this case the presence of a substitution simply informs as to the level of constraint in that region of the protein sequence. In contrast, the alternative selective mechanisms (adaptive bursts, compensatory or synergistic mutations, and hitchhiking) are lineage-specific, they only apply when substitutions occur in the same lineage and therefore can only cause clustering between same-lineage substitutions.
We incorporate polarization into our analysis by extending the sequence features in our cPDFs with the specification of the species lineage on which a substitution arose, e.g. is a non-synonymous substitution in the Dmel, Dsec, Dsim, Dyak, Dere, Dana, Dpse lineage (Methods). The non-synonymous cPDFs estimated for substitutions in the same and different lineage than the focal substitution are shown in Figure 4A for each species comparison. Clustering between substitutions is always significant whether substitutions arose in the same lineage or in different lineages, but clustering between same-lineage substitutions is always significantly stronger (Table S1). We argued above that spatially heterogeneous purifying selection would cause equal clustering within and between lineages. If this is so, the excess clustering within lineages must be generated by one of the lineage-specific alternatives.
Excess lineage-specific clustering can be quantified with an extension of the clustering count . First we define the lineage-specific clustering count as an analog of with the difference that the cPDF from which derives is estimated using only substitutions in lineage . Therefore, is the increased number of -lineage DNs near a focal -lineage DN due to clustering. Next, the ‘lineage-specific excess clustering count’ is the portion of which is inconsistent with a lineage non-specific mechanism. We quantify this as the difference between the within- and between-lineage clustering over the first codons (Methods). This corresponds graphically to the area between those cPDFs (the red area in Figure 4A, ), multiplied by the density of substitutions in the lineage .
The lineage-specific excess appears to be a roughly constant fraction of the total lineage-specific clustering . The estimate of is plotted against the estimate of for both lineages of all our species comparisons in Figure 4B. This relationship is well-fit by a linear model, suggesting that approximately of clustering within a lineage is due to lineage-specific mechanisms, i.e. some combination of compensatory or synergistic mutations, adaptive bursts and hitchhiking. The D. simulans lineage is an outlier, Dsim is aberrantly high. This may be a consequence of details relating to this particular reference sequence: the D. simulans reference sequence has lower coverage and quality than the other reference sequences as well as being a ‘mosaic’ assembly constructed from multiple individuals [29]. The Dsim lineage is also picked out by the synonymous control, there is significant synonymous clustering in this lineage above that found in any other lineage we consider (Figure S10).
If compensatory mutations are contributing substantially to lineage-specific excess one might find evidence of this in a physical or biochemical quantity associated with the compensation. For example, changes in volume, hydrophobicity, charge, etc. might anti-correlate if the substitutions are compensatory. We tested several amino acid properties for such a relationship but found only one that exhibited the hypothesized behavior: nearby substitutions have a significantly increased probability to cause compensatory changes in charge, but only when they arise in the same lineage! We quantify this effect by estimating the fraction of substitutions which compensate the effect of a focal charge-altering substitution, as a function of distance from the focal substitution . In Figure 5 we see that the fraction of charge-compensating substitutions is significantly elevated near a focal charge-altering substitution, on roughly the clustering length scale of 10 codons. This compensation serves to partially conserve the total charge of the protein sequence within the clustering length scale.
Local charge compensation is significant in every species comparison we considered, all p-values , chi-square test (Table S2). A measure of the magnitude of this effect is the fraction of charge-altering substitutions that that have their charge alteration compensated for by the net change in charge caused by the other substitutions within codons. This varies by lineage, but is always significant and increases with species divergence up to for the species comparison of D. melanogaster and D. pseudoobscura. Charge compensation is a lineage-specific effect, and it is responsible for a significant fraction of the lineage-specific excess we observe, roughly depending on lineage (Table S2). The observation of substantial charge compensation, and the lack of compensation of other amino acid properties, is consistent with previous observations which suggested charge compensation to be of greater significance in protein evolution than compensation of other amino acid characteristics [22], [34]. Interestingly, while substitutions in different lineages do not exhibit the local compensation phenomenon, they do show a weaker, but statistically significant, increase in the fraction of nearby changes which alter charge in the same direction, perhaps indicating convergent evolution (Table S2).
Non-synonymous polymorphisms cluster as well, and polymorphism data provides another avenue to distinguish between the possible selective mechanisms of clustering. Under a model of bursts of independent adaptive mutations, beneficial amino acid mutations can be incorporated sequentially, and would not be expected to segregate together in the population since beneficial mutations rapidly fix after arising. In contrast, if epistatic selection is driving the observed clustering we expect that a compensatory mutation will only be found on a chromosome that already carries the first mutation, i.e. we expect the derived states of nearby polymorphic sites to cosegregate. We can quantify this expectation by estimating the average polarized linkage disequilibrium [35], [36], i.e. the frequency of the doubly derived haplotype minus the product of the frequencies of the individual derived alleles averaged over all pairs of polymorphisms a distance apart. then indicates that derived alleles occur in coupling more often than would be expected if their fitnesses were independent. Consistent with the compensatory scenario, we find when estimated from amino acid polymorphisms within codons of each other, as seen in Figure 6.
We evaluate the significance of the cosegregation of nearby derived alleles by bootstrapping: we resample polymorphic sites from the full set of polymorphic sites in our population, pair them off into a number of pairs equal to the number of pairs of polymorphisms within codons of each other, and then estimate from this resampled ensemble. Repeating this process times yields a bootstrapped probability distribution which we compare to the estimated from the data, yielding a bootstrapping p-value of of observing an equal or greater by chance from our population sample. Again, only pairs of non-synonymous polymorphisms significantly cosegregate, supporting the contention that epistasis is responsible and arguing against purely genomic explanations. Although cosegregation is statistically significant, because our polymorphism data set is limited (compared to whole-genome comparisons of divergence) there is more uncertainty about these results, and it is worth noting that cosegregation does not seem to extend beyond three codons of separation.
We have shown that the presence of an amino acid substitution substantially increases the probability of there being additional amino acid substitutions nearby in the protein sequence, with the strength of this effect decaying exponentially along the sequence with a characteristic length scale of ≈10 codons. This ‘clustering’ phenomenon is not observed for synonymous substitutions and is insensitive to the presence of intervening intronic sequence, strongly suggesting selection on proteins as the root cause. Both divergence between Drosophilids and polymorphisms within a population sample of D. melanogaster exhibit this effect. Clustering has a substantial lineage-specific component and nearby substitutions in the same lineage tend to conserve local charge, suggesting compensatory evolution plays a role.
While the results presented here are derived from Drosophila data, we expect that clustering obtains more generally. A recent study found that mutations identified as compensatory clustered near their associated deleterious mutations in eukaryotes, prokaryotes and viruses [37]. Similarly, nucleotide substitutions cluster within codons more often than expected in mammals and HIV, suggesting that two successive mutations are required for the incorporation of some fraction of amino acid substitutions [38], [39].
There are a number of selective mechanisms that could cause amino acid substitutions to cluster, and the clustering we observe most likely has multiple causes. We will now try to reconcile the various observations made above with the different mechanisms that have the potential to cause clustering, and estimate their respective contributions. Potential selective mechanisms of clustering can be grouped into two classes: (A) Heterogeneity in the strength of purifying selection acting within an ORF leads to variation in the density of substitutions and polymorphisms, resulting in clustering. (B) Novel protein variants are selected for and this adaptation leads to clusters of substitutions. The latter class of mechanisms comes in several flavors: (i) A localized adaptive burst in which several nearby substitutions independently sweep to fixation. This might be a consequence of changes in selective pressure on a protein domain that requires multiple adaptive substitutions to reach the new optimum [40]. (ii) A complex adaptation, in which several dependent substitutions are required to achieve the selected effect. This case includes scenarios of compensatory mutations, i.e. a second mutation is necessary to compensate deleterious side effects of the first [41], and evolutionary contingency, i.e. the first mutation is necessary for the second mutation to be beneficial [42]. (iii) Hitchhiking, the fixation of otherwise deleterious substitutions as a result of a selective sweep at a linked site [43], [44].
Purifying selection prunes mutations that are detrimental, perhaps because they interfere with protein structure or stability. Given that protein structure is strongly conserved across different domains of life, it is reasonable to assume that purifying selection operates in a similar fashion on homologous regions of proteins in different branches of the Drosophila phylogeny. Adaptive evolution, however, depends on the ecological niche of the species and can depend strongly on previous substitutions in that species. Adaptive evolution is therefore expected to be lineage-specific, at least moreso than purifying selection.
We observed that clustering exists between pairs of amino acid substitutions in different lineages as well as in the same lineage, the latter being consistently greater (Figure 4). Clustering across lineages implies that a substitution found in one lineage is predictive of the local substitution rate independent of lineage, which we understand as a lineage-non-specific local increase in the substitution rate. This is most consistent with a class (A) mechanism such as locally relaxed purifying selection, e.g. in an unstructured loop of a protein.
The excess clustering within lineages must be caused by a lineage-specific mechanism such as the class (B) mechanisms described above. Purifying selection can of course also vary in a lineage-specific way. If mildly-deleterious substitutions were highly clustered, and a reduced effective population size rendered them effectively neutral, this could result in excess clustering in the lower population size lineage. However, this scenario is inconsistent with the fact that we observe excess clustering within all lineages, and that it is quantitatively similar between lineage pairs diverging from a common ancestor. Locus-specific variation in purifying selection is also possible, but in most cases will affect an entire gene (e.g. via duplication or transformation into a pseudo-gene) and therefore would not lead to clustering on short length scales. Given that excess lineage-specific clustering is a substantial fraction of the total clustering in every lineage, it does not seem likely that lineage-specific variation in the strength of purifying selection can account for it.
We start by addressing the potential contribution of hitch-hiking to clustering. A selective sweep of a strongly beneficial substitution fixes a linked haplotype, converting a local snapshot of polymorphisms present in the population into substitutions. This hitch-hiking process does not affect the fixation probability of neutral (and perhaps synonymous) mutations [45], but is expected to increase the fixation probability of nearby deleterious non-synonymous substitutions. However, several observations argue against hitchhiking as the main contributor to clustering. First, hitch-hiking predicts that the length scale of clustering is given by the typical size of hitchhiked region [46]. This implies clustering dependent on separation along the DNA sequence rather than along the protein backbone, contrary to our observations (Figure 2B). Second, there is no correlation between clustering and the average recombination rate of a coding sequence, which would affect the size of hitchhiked regions (Figure S3). Finally, we can calculate a rough upper bound for the contribution of hitchhiking to lineage-specific clustering. Given a per-site heterozygosity , the expected population frequency of derived mutations per site is . Non-synonymous in D. melanogaster is per site [47]–[49] and thus per 4-fold codon (and slightly higher for 2-folds). Given this, the probability of finding a derived amino acid substitution within codons of a focal site is . This serves as a very generous upper bound on the contribution of hitch-hiking to , since only if the focal site is always adaptive and the observed variation always deleterious will this value be approached. This estimate suggests that the contribution of hitchhiking to lineage-specific clustering is minor, since this upper bound is less than the range over which we observe lineage-specific excess, from to depending on lineage (Figure 4B).
The two remaining adaptive scenarios, adaptive bursts and complex adaptations, are difficult to distinguish in part because the boundary between them is not sharply delineated. Certainly, different substitutions within codons in the same protein are never going to be completely independent. The question rather is whether one of the mutations ‘substantially’ affected the probability of the other. Do localized adaptive bursts, loosely defined as substitutions within codons which all independently improve fitness, dominate our clustering signal? Or are the interactions (epistasis) between nearby substitutions mainly responsible? We cannot fully exclude either scenario, but there is evidence that local interactions play at least a significant role. Mutations of independent beneficial effect would not be expected to compensate each others effect on total charge. This requires epistasis between the substitutions, and implies that complex adaptations are responsible for at least of lineage-specific excess. Secondly, independent beneficial mutations are expected to either fix sequentially or, if they do occur simultaneously, to generally segregate in repulsion [50]. This is inconsistent with the preferential cosegregation we observe between nearby derived alleles (Figure 6). Furthermore, charge compensation is only one of many relevant interactions, albeit the one we most readily ascertained from the primary sequence data. So the contribution of charge compensation is only a lower limit for the influence of complex adaptation on the substitution process.
While the possibility of interactions between amino acid substitutions has never been seriously questioned (and has recently been demonstrated in a number of concrete examples[42], [51]), the general importance of epistasis and compensation in evolution has been, and continues to be, controversial. We find evidence that a non-negligible fraction of substitutions are involved in patterns of adaptation suggestive of epistasis. If lineage-specific clustering is mostly due to epistasis, a scenario consistent with our results, we can use the lineage-specific excess to estimate the number of substitutions which owe their fixation to interactions with other substitutions. For example, the lineage-specific excess in the D. yakuba lineage is Dyak. If we attribute the entirety of this to epistasis we would conclude that of the substitutions on this lineage were contingent on another substitution. This estimate is clearly generous in the sense that we have not completely excluded the contribution of other processes, but it is also conservative in the sense that it only includes the effect of elevated local epistasis and excludes the contribution of long-range interactions.
To account for interactions between amino acids distant in the protein sequence but nevertheless in close vicinity in the folded protein, one would need to incorporate protein structure explicitly. However, the probability for any random pair of residues to be involved in such interaction decays rapidly with their separation along the protein backbone, likely to an asymptotic value. Hence, in our analysis we expect correlations between distant pairs to be lost in the background, with only the enriched short range interactions observable as excess clustering of substitutions. The presence of this local enrichment is the enabling factor behind our approach. In agreement with this interpretation, the inferred length scale of clustering of 10 codons is consistent with the size of secondary structure elements in proteins (e.g. 3 turns of an helix). While this manuscript was prepared for publication, another group also found clustering of positively selected amino acid substitutions [17]. Via a different approach, the authors show that the rate of evolution depends on elements of secondary structure and that nearby positively selected sites tend to cluster.
Finally, while we have focused on the mode of evolution responsible for lineage-specific excess, the clear clustering which occurs across lineages is notable in its own right. We attribute this clustering to spatially heterogeneous purifying selection. The clustering length scale is extremely consistent across all the species comparisons we considered and the polymorphism data (Figure 2). This suggests that models of protein evolution might be improved by incorporating correlation between the rate of amino acid evolution along the sequence (e.g. site-specific in PAML [52]). This is particularly true if the length scale we observe here can be shown to be consistent across phyla, demonstrating it as a generic property of proteins themselves.
We assign substitutions in coding sequence (CDS) on a codon-by-codon basis to pairwise alignments of the reference genome of D. melanogaster with the reference genomes of 6 other Drosophilids: D. sechellia, D. simulans, D. yakuba, D. erecta, D. ananassae and D. pseudoobscura. We use FlyBase release 5.26 gene models to identify the location of coding sequence in the D. melanogaster genome. Coding sequence substitutions are assigned only in the absence of gaps and ambiguous nucleotide. If the aligned codons encode different amino acids a non-synonymous substitution () is assigned, if the same amino acid, a synonymous substitution (). Substitutions are assigned in the context of an alignment between two lineages, which we can make explicit by writing where is the pairwise alignment between lineages and . We omit the alignment for notational convenience, the alignment under consideration will be clear by context, but it is worth remembering that the objects we define later depend implicitly on an alignment when they involve substitutions.
Substitutions are polarized into the lineage in which they arose by comparison to the closest available Drosophilid that is more distant from D. melanogaster than the one being aligned. Specifically, D. yakuba is used as the outgroup for the D. simulans and D. sechellia comparisons, D. ananassae for D. yakuba and D. erecta, D. pseudoobscura for D. ananassae and D. willistoni for D. pseudoobscura. Substitutions are assigned to a lineage if the assignment is unambiguous using standard parsimony criteria. A polarized into lineage is denoted . Not all substitutions can be polarized, represents the fraction of substitutions between the lineages and which are polarized. The species comparisons between D. melanogaster and either D. yakuba or D. erecta have the best properties for the analysis here: most coding sequence alignments pass quality checks and the number of substitutions, both polarizable and total, is high, as seen in Figure S1 and observed previously [53]. When we present results from just one species comparison it will be the D. melanogaster - D. yakuba comparison for this reason.
Having assigned and polarized substitutions, we study clustering between substitutions typed by synonymity, e.g. non-synonymous (DN) and synonymous (DS) divergent sites, by estimating the probability of finding a substitution (of some particular type) codons away from a focal substitution. This is formalized as a conditional probability distribution (cPDF), which we denote , where is the focal substitution type, and is the substitution whose frequency is measured at distance . and can be simply DN or DS, or in the later analysis a substitution polarized to a particular lineage.
The cPDF is calculated from a set of CDSs on which the presence/absence of substitutions and have been ascertained site-by-site. is the proportion of sites a distance downstream (coding sense) of a substitution of type , summed over all instances of in the data set. We must account for the decrease in the number of observations made as increases due to the finite length of coding sequences. To be precise, the cPDF is calculated as follows: Let us label individual CDSs with and index the codons in a CDS by , which ranges from to the length of the CDS, . We define an indicator variable for each CDS and substitution type . if codon of contains an substitution, and is otherwise. The cPDF is defined as,(1)where the Kronecker symbol if and 0 otherwise. These cPDFs generically go to an ‘asymptotic value’ , which is calculated ad-hoc by averaging over . This property allows us to separate the functional dependence of a cPDF on distance from its absolute value by introducing the ‘normalized’ cPDF .
The cPDF naturally generalizes to include polarization information. Polarized cPDFs are defined as above, with , , and an additional summation over the lineages, . A same-lineage cPDF enforces the same-lineage condition with a Kronecker delta , different-lineage with .
We define the clustering count as the sum of the difference between the cPDF and its asymptotic value over the first codons, i.e. the area between and the asymptotic value of the cPDF over . This is the difference between the expected number of substitutions within codons downstream of a focal substitution and the expected number in a codon sequence segment that is distant from the focal substitution. The choice of as the upper limit of the sum simply reflects the observation that significant clustering does not extend past this point. Additionally, this is a one-sided sum, ensuring that each pair is counted only once. We also define the lineage-specific excess clustering count for lineage , , in order to quantify the stronger clustering within a lineage. Lineage-specific excess is found by summing over the difference between the normalized same-lineage cPDF and the normalized cross-lineage cPDF and then ‘unnormalizing’,(2)(3)
The factor of , defined above, corrects for the fraction of substitutions that cannot be unambiguously polarized. Multiplying by roughly accounts for this by assuming that the polarized substitutions are representative of the unpolarized ones. When or are written without indices they should be assumed to refer to clustering, i.e. . Note that and depend on the pairwise alignment being considered via their dependence on the assignment of substitutions, as described at the beginning of the Methods.
Our definition of cPDFs implicitly involved the determination of the set of CDSs to be summed over. This set varies with the species comparison so we denote it . For the results presented here was the set of all CDSs for which the pairwise alignment between D. melanogaster and sister Drosophilid met several standards of quality: a CDS included in was required to have less than gapping in its pairwise alignment and less than amino acid substitution, the alignment could not contain out-of-frame gaps (gaps with size that is not a multiple of three), and the spliced transcript to which the CDS belongs could contain no extraneous stop codons. Furthermore, we often restrict to a specific subset in order to investigate the dependence of the clustering signal on various quantities, e.g. Figure 3B shows the cPDFs calculated using subsets of all CDSs ranked by .
Polymorphisms were identified in a Zimbabwean population sample of male D. melanogaster. We re-sequenced kb of coding sequence from genes in the highly recombining region of the X-chromosome (cytological positions 3C3 to 18F4) using standard methods reported previously [54]. Samples sizes ranged from to with a mean of . A subset of these sequences ( alleles for each of loci), were previously reported [54]. All new sequences have been submitted to GenBank, accession numbers are available in Table S3. Polymorphisms are assigned if more than one codon exists in the population sample at that site. Singletons are excluded from the analysis. A non-synonymous polymorphism is assigned if this set of codons encodes more than one amino acid, and a synonymous polymorphism if the number of codons exceeds the number of amino acids encoded. PN and PS assignment is not exclusive. Polarization into mutant/ancestral alleles is inferred by comparison to D. simulans (or D. sechellia when D. simulans is unavailable) at that site (i.e. standard parsimony criteria). cPDFs are constructed analogously to those involving substitutions.
All line plots presented are smoothed from the underlying data. We used a moving window averaging for smoothing, always with window size . The contribution from each data point to the smoothed average was weighted by the number of ‘trials’ from which the value was estimated.
We assess two ‘types’ of significance here, sampling significance and bootstrapping significance. The assessment of sampling significance is understood by recalling how cPDFs are estimated. is the mean of a set of trials which can have outcome either or (Bernoulli random variables). is the same, it is just an average over a cPDF for . Trials consist of selecting a focal substitution of type , looking away on the sequence, and recording the presence () or absence () of a substitution of type . So, assessing the significance of values of or is equivalent to assessing the significance of sums of Bernoulli random variables, for which we used chi-square tests.
Bootstrapping significance is also a measure of sampling significance, with the difference being that the effect of resampling is evaluated at the level of the largest unit in our analysis, the coding sequence. The probability distribution of a value of interest is constructed by resampling with replacement from the full set of coding sequences a ‘bootstrapped’ set of equal size, estimating the value of interest on that bootstrapped set, and repeating. Bootstrapping p-values are then determined from this estimate of the probability distribution. If the estimated distribution can be approximated as a gaussian, as is always the case here, the gaussian approximation is used to assign the p-value. A modification of this bootstrapping scheme was used for polymorphism cosegregation, and described there.
|
10.1371/journal.pntd.0005108 | Lymphatic Filariasis Elimination in American Samoa: Evaluation of Molecular Xenomonitoring as a Surveillance Tool in the Endgame | The Global Programme to Eliminate Lymphatic Filariasis has made significant progress toward interrupting transmission of lymphatic filariasis (LF) through mass drug administration (MDA). Operational challenges in defining endpoints of elimination programs include the need to determine appropriate post-MDA surveillance strategies. As humans are the only reservoirs of LF parasites, one such strategy is molecular xenomonitoring (MX), the detection of filarial DNA in mosquitoes using molecular methods (PCR), to provide an indirect indicator of infected persons nearby. MX could potentially be used to evaluate program success, provide support for decisions to stop MDA, and conduct post-MDA surveillance. American Samoa has successfully completed MDA and passed WHO recommended Transmission Assessment Surveys in 2011 and 2015, but recent studies using spatial analysis of antigen (Ag) and antibody (Ab) prevalence in adults (aged ≥18 years) and entomological surveys showed evidence of possible ongoing transmission. This study evaluated MX as a surveillance tool in American Samoa by linking village-level results of published human and mosquito studies. Of 32 villages, seropositive persons for Og4C3 Ag were identified in 11 (34.4%), for Wb123 Ab in 18 (56.3%) and for Bm14 Ab in 27 (84.4%) of villages. Village-level seroprevalence ranged from 0–33%, 0–67% and 0–100% for Og4C3 Ag, Wb123 Ab and Bm14 Ab respectively. PCR-positive Aedes polynesiensis mosquitoes were found in 15 (47%) villages, and their presence was significantly associated with seropositive persons for Og4C3 Ag (67% vs 6%, p<0.001) and Wb123 Ab (87% vs 29%, p = 0.001), but not Bm14 Ab. In villages with persons seropositive for Og4C3 Ag and Wb123 Ab, PCR-positive Ae. polynesiensis were found in 90.9% and 72.2% respectively. In villages without seropositive persons for Og4C3 Ag or Wb123 Ab, PCR-positive Ae. polynesiensis were also absent in 94.1% and 70.6% of villages respectively. Our study provides promising evidence to support the potential usefulness of MX in post-MDA surveillance in an Aedes transmission area in the Pacific Islands setting.
| Lymphatic filariasis (LF) is caused by infection with filarial worms that are transmitted by mosquito bites. Globally, 36 million are disfigured and disabled by complications such as severe swelling of the legs (elephantiasis) or scrotum (hydrocele). The Global Programme to Eliminate LF (GPELF) aims to interrupt disease transmission through mass drug administration (MDA), and to control illness and suffering in affected persons. Significant progress has been made toward eliminating LF from many parts of the world, including the Pacific Islands. Current challenges of the GPELF include identification of any residual hotspots of ongoing transmission, and effective strategies for early identification of any resurgence of infections. As humans are the only reservoirs of LF parasites and mosquitoes have short flight ranges, one such strategy is to monitor LF infection in mosquitoes as an indicator of ongoing transmission nearby. Mosquito monitoring could potentially be used to evaluate program success, provide support for decisions to stop MDA, and conduct post-MDA surveillance. Our study evaluated mosquito monitoring as a surveillance tool in American Samoa by linking village-level results of published studies of LF in humans and mosquitoes, and provides promising evidence to support the potential usefulness of mosquito monitoring in post-MDA surveillance the Pacific Islands.
| Lymphatic filariasis (LF) is a parasitic infection caused by Wuchereria or Brugia species of helminth worms, and transmitted by mosquito vectors including Aedes, Anopheles, Culex and Mansonia species. Globally, an estimated 68 million people are currently affected, including 36 million microfilaraemic persons and 36 million who are disabled or disfigured with complications such as severe lymphoedema, including elephantiasis and scrotal hydrocoeles [1]. The Global Programme to Eliminate LF (GPELF) aims to eliminate the disease as a public health problem by 2020 using two main strategies: i) to interrupt transmission through mass drug administration (MDA) and ii) to control morbidity and disability of affected persons. In the Pacific Islands, the Pacific Programme to Eliminate LF (PacELF) was formed in 1999 as part of GPELF to focus on 22 Pacific Island Countries and Territories (PICTs), which include >3000 islands and 8.6 million people [2].
The sustained success of elimination programs requires cost-effective assessment and monitoring after successful completion of MDA to determine whether there are any residual foci of infection, and to detect potential resurgence in a timely manner. The WHO currently recommends post-MDA surveillance using transmission assessment surveys (TAS), which use critical cut-off values of numbers of antigen-positive children aged 6–7 years to determine whether transmission has been interrupted in defined evaluation units [3]. In Brugia transmission areas, antibody positivity is used for TAS. Cut-off thresholds for passing TAS vary depending on population size of the target group and the local species of filarial parasites and mosquito vectors. For example, where W. bancrofti is endemic, the target cut-off value is estimated based on upper 95% confidence interval of <1% antigen prevalence if Aedes is the principal vector, or <2% antigen prevalence if Anopheles or Culex predominate. TAS typically involve school-based or community-based testing of 6–7 year old children. Community-based surveys are often logistically challenging, particularly in developing countries with limited financial and human resources. In some areas (e.g. most of the Pacific Islands), difficult access to dispersed populations in remote islands provides additional challenges. Also, TAS typically use rapid antigen detection tests (Filarial Immunochromatographic Test (ICT) cards or Filarial Test Strips [4]), which might have reduced sensitivity after many rounds of MDA [5,6]. TAS has been widely used to inform important programmatic decisions including stopping or restarting MDA, but recent studies suggest that in some settings (including American Samoa), TAS might not be sufficiently sensitive for determining whether transmission has been interrupted [7,8].
As elimination programs reach the endgame phases and antigen prevalence drops to very low levels, increasingly sensitive tools and strategies will be required to efficiently detect any evidence of ongoing transmission or resurgence. The WHO and GPELF have identified a number of operational challenges and unanswered questions for elimination programs, including the significance of residual microfilaraemia and antigenaemia in communities where the target threshold level has been achieved through MDA, identification of residual high-prevalence areas and strategies for managing them, and the need for development of cost-effective strategies for post-MDA surveillance [9]. As humans are the only reservoir for W. bancrofti parasites, one such strategy is to monitor mosquitoes for evidence of LF larval stages [10,11]. Dissection and microscopic examination of mosquitoes is time consuming and labour-intensive, and cannot be routinely recommended for post-MDA surveillance for practical reasons. With recent technological advances, LF molecular xenomonitoring (MX), the use of molecular methods (PCR) to detect filarial DNA in mosquitoes, has been explored and promising results are emerging [12,13]. PCR-positive mosquitoes provide an indirect indicator of the presence of infected humans and possible ongoing transmission [10,14–16]. For example, considering that the flight ranges of Ae. polynesiensis mosquitoes are on the order of a hundred metres [17], detection of PCR-positive mosquitoes in areas where these are the main vectors indicates that infected persons are or were recently nearby. Molecular methods are also more sensitive than manual dissection for detecting infections [18]. Studies have reported the ability of PCR to detect one microfilaria in pools of 50–100 mosquitoes [19], and for at least 2 weeks after mosquitoes (both vector and non-vector) ingest microfilaria-positive blood [20], which is close to the average life span of most mosquito species.
Molecular xenomonitoring has been found to be a potentially useful indicator of human LF infections with different species of mosquito vectors in diverse settings including American Samoa [11], French Polynesia [16], Egypt [12,21], Sri Lanka [8], Sierra Leone [22], and Ghana [23]. Molecular xenomonitoring is therefore potentially useful for evaluating the success of elimination programs, providing support for decisions to stop MDA, and conducting ongoing post-MDA surveillance [24]. Compared to TAS, MX has the advantages of being non-invasive to humans, and potentially more cost-effective in some settings. However, MX requires entomological expertise for trapping and processing mosquitoes, and laboratories capable of conducting large scale molecular diagnostics. In addition, there are currently unanswered questions about sampling strategies, limited evidence to inform the translation of MX results into operational strategies, and no clear guidelines on the thresholds of DNA prevalence that should be used to indicate likely ongoing transmission. Cut-off points of 0.25%, 0.5%, and 1% have been suggested for Culex areas [8,12,25], and 0.085% for L3 and 0.65% for any larval stage for Anopheles areas [13]. There are currently no clear recommendations for Aedes areas, but a provisional threshold of <0.1% has been suggested [26]. The lower the estimated cut-off points, the larger the sample sizes of mosquitoes that will be required for MX.
As part of the PacELF, American Samoa has made significant progress toward reducing LF infection rates. After seven rounds of MDA from 2000 to 2006, antigen prevalence in humans dropped from 16.5% (N = 3018) in the 1999 baseline assessment to 2.3% (N = 1881) in 2007 in a community cluster survey [27]. American Samoa passed TAS in 2011–2012 and again in 2015, but recently published studies using spatial analysis of antigen prevalence in adults [7] and molecular xenomonitoring [28] showed evidence of possible ongoing transmission. By linking the results of the published human and mosquito studies, we aim to evaluate MX as a surveillance tool in the post-MDA setting in American Samoa, an Aedes transmission area in the Pacific Islands.
American Samoa is a US Territory in the South Pacific, consisting of a group of small tropical islands with a total population of 56,000 [29] living in 67 villages. Over 90% of the population live in small villages on the main island of Tutuila, and the remainder on the adjacent island of Aunu’u and the remote Manu’a group of islands (Ta’u, Ofu, and Olosega). W. bancrofti is the only species of human filarial worm known to be present in American Samoa. The main vector is the highly efficient day-biting Ae. polynesiensis, and other vectors include Ae. samoanus (night-biting), Ae. tutuilae (night-biting), and Ae. upolensis (day-biting) [30–32].
Data were obtained from a published study on the seroprevalence and spatial epidemiology of lymphatic filariasis in American Samoa [7]. The study used samples from a serum bank collected from May to August 2010 for a leptospirosis study; the study design has been published previously [33,34]. Briefly, the study included 807 adults (aged 18 to 87 years, 52.4% males) from 659 households in 55 villages on all five inhabited islands of American Samoa. Sampling was designed to provide a representative sample of the adult population in American Samoa, in both age and geographic distribution. Using these 2010 samples, a seroprevalence study was conducted in 2013 [7], and found that 3.2% were seropositive for Og4C3 Ag, and 8.1% and 17.9% were seropositive for Wb123 Ab and Bm14 Ab, respectively [7]. The study also found significant spatial clustering of Ag-positive persons; average cluster size was 1,498m in diameter for those with Og4C3 Ag >32 units, and the proportion of the variation explained by geographic proximity was 62%. Higher infection rates were found in males and recent migrants to American Samoa. Antigen (Og4C3) positivity indicates the presence of adult worm antigen but does not provide information on the viability, e.g. the worm could be alive or dead, or there could be a single sex worm infection or sterile worm infection. The presence of Og4C3 Ag represents current or recent infection. The presence of antibodies represents current or past infection, possibly many years in the past.
For our study, human data were summarized by village and the following variables were generated:
A village-level summary of the human serological data is provided in S1 Appendix.
Schmaedick et al conducted a MX study in American Samoa from February to June 2011, approximately 9 months after the above human serum specimens described above were collected. Detailed descriptions of the study and its findings have been published [28], and a village-level summary is provided in S1 Appendix. Briefly, mosquitoes were collected from 34 randomly selected villages on the island of Tutuila, the only village on Aunu’u, all five villages on the Manu’a Islands, and the village of Ili’ili (on Tutuila) where two ICT-positive children were identified during the 2011 TAS. Up to 10 traps were placed in each village for 24 to 48 hours, and mosquitoes were removed from traps twice daily. The study collected a total of 22,014 female mosquitoes of Aedes and Culex genera that were sorted into 2,629 pools of ≤20 mosquitoes (range 1 to 20) for PCR testing. Real-time PCR was conducted using primers designed to amplify a fragment of W. bancrofti [35]. A positive PCR result indicates the presence of filarial worm DNA, but does not provide any information on whether the worms are alive or transmissible. Each pool included only one mosquito species, except for the Ae. (Finlaya) group of species (Ae. oceanicus, Ae. samoanus, and Ae. tutuilae) which were combined for PCR testing because of morphological similarities. The MX study calculated maximum likelihood point estimates of the prevalence of PCR-positive Ae. polynesiensis for each village or village group using PoolScreen software (version 2.0.3), which takes into account the average number of mosquitoes per pool and the proportion of pools that were PCR-positive. Point estimates of village-level prevalence of PCR-positive Ae. polynesiensis ranged from 0% to 2.8% for villages on Tutuila and Aunu’u, and was 0% for all villages in the Manu’a islands. The findings indicated widespread presence of filarial DNA in the mosquito population, suggesting persistent low-level transmission of LF on Tutuila and Aunu’u. [27]
For our study, mosquito data were summarized for each village for i) Ae. polynesiensis, and ii) other mosquito species (all species apart from Ae. polynesiensis), and iii) any mosquito species. Entomological data available by village included number of traps used; number of females and pools of each mosquito species; number of PCR-positive pools of each species; and estimated prevalence of PCR-positive Ae. polynesiensis (using PoolScreen software).
This study used de-identified data from the two previously published studies described above [7,28]. The human study only included adults, and written informed consent was obtained from each participant. The American Samoa Institutional Review Board (IRB) provided approval for the use of the human serum bank for lymphatic filariasis research.
In the MX study, some small adjacent villages were combined into groups of two to four villages for trapping and analyses, and human data were grouped accordingly to match the entomological data. Human data were not available for three of the villages included in the MX study. For this study, analyses were limited to the villages or village groups where both human data and MX data were available for 32 locations: 23 individual villages on Tutuila, 3 village groups (of two villages each) on Tutuila, the only village on Aunu’u, and all five villages on the Manu’a Islands. The 32 villages and village groups will be referred to as “villages” from here for ease of reference.
Chi-squared tests and logistic regression were used to identify associations between seropositive humans and PCR-positive mosquito pools, and answer the following operational questions:
Serological results from 376 persons residing in 32 villages were included in the analyses. The average number of persons per village was 13.9 (range 2–73) for Tutuila and Aunu’u, and 14.0 (range 11–16) for the Manu’a Islands.
Table 1 provides a summary of the number of seropositive persons and village-level seroprevalence for each serological marker in humans, and the entomological data used in this study. Of the 32 villages included in this study, 11 (34.4%) had residents who were seropositive for Og4C3 Ag, 18 (56.3%) for Wb123 Ab, and 27 (84.4%) for Bm14 Ab. On Tutuila and Aunu’u, village-level seroprevalence ranged from 0% to 33.3% for Og4C3 Ag, 0% to 66.7% for Wb123 Ab, and 0% to 100% for Bm14 Ab. In the Manu’a Islands, no individuals were seropositive for Og4C3 Ag, and village-level seroprevalence ranged from 0% to 18.8% for Wb123 Ab, and 13.3% to 27.3% for Bm14 Ab. On Tutuila and Aunu’u, the MX study identified PCR-positive pools of Ae. polynesiensis in 15 (55.6%) of the 27 villages included in this study, of other mosquito species in 7 (25.9%) villages, and of mosquitoes of any species in 17 (63.0%) of the villages. In the five villages on the Manu’a Islands, no PCR-positive pools of mosquitoes were identified during the MX study.
Associations between the presence of PCR-positive pools of mosquitoes and seropositive villages and are shown in Table 2, with analyses stratified for i) Ae. polynesiensis only, ii) for all other mosquito species and iii) for mosquitoes of any species. Chi-squared tests of association showed that PCR-positive pools of Ae. polynesiensis (p ≤ 0.001) and PCR-positive pools of any species (p = 0.002), but not pools of other species, were significantly associated with seropositive villages for Og4C3 Ag and Wb123 Ab, but not Bm14 Ab.
Fig 1 shows that the presence of at least one PCR-positive pool of Ae. polynesiensis or of any species was associated with a significantly higher probability of identifying a village with inhabitants seropositive for Og4C3 Ag (p < 0.001 and p = 0.002) and Wb123 Ab (p = 0.001 and p = 0.002). In the 15 villages with at least one PCR-positive pool of Ae. polynesiensis, 10 (67%) were seropositive for Og4C3 Ag and 13 (87%) were seropositive for Wb123 Ab, compared to 6% and 29% of villages respectively, where PCR-positive pools were not identified. Similarly, in the 17 villages where at least one PCR-positive pool of any species were identified, 11 (59%) had inhabitants who were seropositive for Og4C3 Ag and 14 (82%) with persons seropositive for Wb123 Ab, compared to 7% and 27% of villages respectively, with no PCR-positive pools. The presence of PCR-positive pools was not significantly associated with seropositivity for Bm14 Ab. PCR-positive pools of other mosquito species were not significantly associated with seropositive villages for any of the serological markers.
Table 3 provides a summary of the accuracy of PCR-positive mosquito pools for predicting seropositive villages for each antigen and antibody. PCR-positive pools of Ae. polynesiensis provide a sensitivity of 90.9% and specificity of 76.2% for identifying villages with seropositive persons for Og4C3 Ag, with a high negative predictive value of 94.1% (i.e. absence of PCR-positive pools was a good indicator of the absence of seropositive persons). PCR-positive pools of any mosquito species provide the same sensitivity (90.9%) but a lower specificity (66.7%), and a negative predictive value of 93.9%.
For Wb123 Ab, PCR-positive pools of Ae. polynesiensis provide a sensitivity of 72.2% and specificity of 85.7%, while PCR-positive pools of any mosquito species provide a sensitivity of 77.8% and specificity of 78.6% for identifying seropositive villages. For Bm14 Ab, PCR-positive pools of Ae. polynesiensis and any species had poor sensitivities (48.1% and 51.9%) and specificities (60.0% and 40.0%) for predicting seropositive villages.
PCR-positive pools of Ae. polynesiensis or any mosquito species were statistically significant predictors of villages with residents seropositive for Og4C3 Ag (odds ratios of 32.0 and 20.0) and Wb123 Ab (odds ratios of 15.6 and 12.8), but not for Bm14 Ab. The correlation between PCR-positive pools of Ae. polynesiensis and seropositive villages for Og4C3 Ag and Wb123 Ab are shown for each village in Tutuila and Aunu’u in Fig 2, and the Manu’a Islands in Fig 3.
In the MX study, the estimated prevalence of PCR-positive Ae. polynesiensis ranged from 0% (95% CI 0–0.1%) to 2.8% (0.5–7.9%) on Tutuila and Aunu’u, and was 0% for all villages in the Manu’a islands. Table 4 shows that a higher estimated prevalence of PCR-positive Ae. polynesiensis (as a continuous variable) was associated with increased odds of a seropositive village for Og4C3 Ag and Wb123 Ab, but the findings were not statistically significant with this sample size and the level of precision inherent in PoolScreen predictions based on pooled mosquito samples.
Our results show that MX is a potentially useful tool for post-MDA surveillance of lymphatic filariasis in American Samoa. The presence of PCR-positive pools of Ae. polynesiensis was found to be a good predictor of villages with persons seropositive for Og4C3 Ag and Wb123 Ab, but not Bm14 Ab. Bm14 Ab can persist for many years or decades after initial infection but does not necessarily persist for life, and antibody levels can also decline or be cleared after MDA [36,37]. Wb123 Ab can also persist for many years after initial infection, and declines after MDA [38]. Currently, there is insufficient data on the relative rates of antibody decay or clearance, but it is thought that Wb123 Ab responses decay more rapidly than Bm14 Ab. Biologically, Wb123 Ab responses might increase earlier because they are against a larval antigen and therefore also more likely to be associated with mosquito exposure. In our study, the lack of association between PCR-positive pools of mosquitoes and Bm14 Ab was therefore not unexpected, but the association with Wb123 Ab could be related to earlier appearance or faster disappearance of Wb123 Ab than Bm14 Ab during and after active infections, respectively [39–41].
In American Samoa, where ~75% of mosquitoes collected in the entomology study (using BG Sentinel traps) were Ae. polynesiensis, the presence of PCR-positive pools of either Ae. polynesiensis alone or any mosquito species provided similar predictive accuracy of identifying villages with residents seropositive for Og4C3 Ag or Wb123 Ab. Our study shows that in this setting, separation of mosquito species for MX did not improve the predictive accuracy for identifying villages with seropositive inhabitants. However, it is important to point out that our results would have been different if we had used traps with a different level of selectivity for Ae. polynesisensis. MX studies in locations with other vector species and employing different traps may require sorting of mosquito species to achieve optimal results.
The presence of PCR-positive pools of Ae. polynesiensis or mosquitoes of any species had high sensitivity and high negative predictive value (both >90%) for correctly identifying villages with antigen-positive persons. These are both important for post-MDA surveillance because tests should have a high probability of identifying residual foci of transmission (when prevalence is very low) and low probability of missing these foci. In this study, the estimated prevalence of PCR-positive Ae. polynesiensis (using PoolScreen) was no more useful than the presence/absence of PCR-positive pools, but the sampling design (small number of persons in some villages) might have limited the ability to detect significant associations.
Previous studies in three sentinel villages in American Samoa showed that MX could be a useful tool in post-MDA surveillance [11,18]; our larger study of 32 villages corroborates those conclusions. Our findings also suggest that in American Samoa, it is appropriate to conduct post-MDA surveillance at the village level. This is biologically plausible considering that the main vector, Ae. polynesiensis, has a relatively short flight range of about 100 metres [17], and village residents are generally quite mobile within their own village, e.g. visiting homes of family and friends, sharing outdoor spaces, attending school and church, and shopping at local stores. Further interventions (e.g. further targeted MDA or a test and treat approach) could also be conducted at the village or even sub-village level.
The results should be considered in light of the study’s limitations. The study was based on serological data from humans; microfilaria results were not available because the study was conducted using a pre-existing serum bank. Human serological data and entomological data were sourced from previously published studies, and there was a time lag of approximately nine months between the human and entomology studies. Sampling of the human seroprevalence study was designed to maximise spatial dispersion for the purposes of predictive risk mapping for the original leptospirosis study [7], resulting in small numbers of subjects in some villages and wide confidence intervals for the village-level LF seroprevalence estimates. The serum bank only included samples from adults (aged ≥ 18 years); a study that focused on or only included children might produce different results regarding the usefulness of serological markers, e.g. there could be significant associations between PCR-positive mosquitoes and Bm14 Ab in children. Ae. polynesiensis, the primary vector in American Samoa, is a day-biting mosquito; our human data were summarised by village of residence, and it is possible that PCR-positive mosquitoes acquired infections from residents of other villages who visited during day time.
Despite the study’s limitations, we were able to identify statistically significant associations between MX data and human seroprevalence data at the village level. Further studies specifically designed to assess the usefulness of MX in the post-MDA setting might produce results with even stronger associations. With higher resolution data, it is also potentially possible to determine thresholds for the prevalence of PCR-positive Ae. polynesiensis at which further interventions (e.g. repeating MDA or more intensive surveillance) are recommended. In American Samoa, LF transmission is dominated by Ae. polynesiensis; studies in other countries with a different mix of vector mosquitoes will be needed to determine whether separation of mosquitoes by species is necessary for MX.
This exploratory study provides promising evidence to support the potential usefulness of MX in post-MDA surveillance in an Aedes transmission area in a Pacific Island setting to predict sub-national areas where LF transmission may still be occurring. Although American Samoa has successfully completed MDA and passed two TAS of 6–7 year old children, there is evidence of ongoing low-level transmission of LF. Our findings demonstrated that in this setting, MX was useful for localising residual areas of focal transmission and could potentially be used to inform the need for additional elimination activities. Our study also highlights that assessment of antigen prevalence in adults in post-MDA surveillance could complement TAS and provide valuable information for informing programmatic decisions in the endgame.
|
10.1371/journal.pcbi.1004685 | Implementation of Complex Biological Logic Circuits Using Spatially Distributed Multicellular Consortia | Engineered synthetic biological devices have been designed to perform a variety of functions from sensing molecules and bioremediation to energy production and biomedicine. Notwithstanding, a major limitation of in vivo circuit implementation is the constraint associated to the use of standard methodologies for circuit design. Thus, future success of these devices depends on obtaining circuits with scalable complexity and reusable parts. Here we show how to build complex computational devices using multicellular consortia and space as key computational elements. This spatial modular design grants scalability since its general architecture is independent of the circuit’s complexity, minimizes wiring requirements and allows component reusability with minimal genetic engineering. The potential use of this approach is demonstrated by implementation of complex logical functions with up to six inputs, thus demonstrating the scalability and flexibility of this method. The potential implications of our results are outlined.
| Synthetic biological circuits have been built for different purposes. Nevertheless, the way these devices have been designed so far present several limitations: complex genetic engineering is required to implement complex circuits, and once the parts are built, they are not reusable. We proposed to distribute the computation in several cellular consortia that are physically separated, thus ensuring implementation of circuits independently of their complexity and using reusable components with minimal genetic engineering. This approach allows an easy implementation of multicellular computing devices for secretable inputs or biosensing purposes.
| Synthetic biological devices have been built to perform a variety of functions [1–3]. Currently, the creation of complex logic circuits capable of integrating a high number of different inputs and of performing non-trivial decision making processes is one of the major challenges of synthetic biology [4–8]. Examples of synthetic gene circuits used to perform digital computation are switches [9,10], logic gates [11,12], oscillators [13], band-pass filters [14], classifiers [15] and memory devices [16]. However, despite the enormous efforts devoted to developing such devices, the results obtained are far from the level of complexity needed for applications [17,18]. Limitations in the design of some of these devices and the lack of reusability of the genetic modules strongly constrains the degree of scalability and complexity necessary for industrial, environmental or biomedical applications [19].
In general, the implementation of biological devices that are capable of performing complex logical computations in response to a growing number of input signals involves complex genetic engineering with limited reusability. Usually, the circuits are obtained (or designed) by connecting basic logic gates following standard combinatorial logic, inspired by the circuit analogies applied to understanding genetic networks [20–24]. Dedicated efforts have been oriented towards the exploration of such combinatorial scheme within synthetic biology [25,26]. In accordance with this standard architecture, the functional complexity of a circuit will scale up with both the number of different logic gates and the number of wires that connect them (i. e. circuit connectivity). Both elements limit the scalability and complexity of these devices [19]. One of the most restrictive constraints is the so-called wiring problem. While wiring is not a major problem in standard electronics, in biological systems is a key limiting factor. This limitation arises from the fact that each connection (wire) requires a different biochemical entity and that crosstalk needs to be prevented [27].
In spite of the efforts aimed at standardization of genetic components in synthetic biology, serious limitations still exist [28]. This limits both scalability and the potential reuse of genetic components. Along with the wiring problem, novel strategies towards synthetic biological computation seem required to overcome these problems. In this context, the implementation of circuits using multicellular consortia instead of single cells allows for a reduction in the genetic engineering required in a particular cell [29,30] and the reusability of the components. In this scenario, each cell carries a particular engineered design that, when combined with other cells of the consortia, performs the final computation (hereafter distributed computation) [27]. Furthermore, when this approach is combined with distributed output production [31] or spatial segregation [32] it allows the attainment of logic circuits with a significant reduction in the number of wires and genetic manipulations required. Noteworthy, both in nature and engineering, space is used as an added dimension of information processing, such as in intracellular network computation [33], amorphous computing [34], cell-cell interaction [35], pattern formation [36–38], or in ant colonies [39,40]. Nevertheless, spatial segregation has never been fully exploited as a key computational parameter in the building of synthetic biological devices [17–19].
Here we present a novel methodology that allows designing biological devices based on the combination of three elements: multicellular consortia, distributed output production and spatial segregation. A major reason to adopt this approximation is the division of labor already present in tissues and organs, where different cell types perform different functions while communicating through signaling molecules. Such segregation of functions, combined with integration of signals is a universal design principle of multicellular systems. Our approach uses engineered cells (our cell types) implementing one-input one-output logic gates organized in several consortia, connecting cells of each consortium with a single wire, and allowing each consortium to produce the final output independently of the others. This systematic and simplified distribution of the computation, together with spatial isolation in modules of the different multicellular consortia, permits the building of complex logic circuits. Modules and cells can be reorganised to obtain different computations. Notably, only one wiring molecule and minimal genetic engineering is used. An important property of this architecture is that it does not depend on the complexity of the circuit to be implemented; thus, scalability is ensured because the required number of cell types and modules are bounded. As a proof of principle, we have built several logic circuits in eukaryotic cells with increasing complexity that respond to up to 6 inputs (such as a 4-to-1 multiplexer). Furthermore, we focused on two particular, but very relevant, types of biological devices with diverse outputs. The first type of devices produce an output with a determined function that is secreted into the medium (e.g. hormone, secretable enzyme). Such devices could be used in bioreactors for the production of enzyme, metabolites, or recombinant proteins, as well as in biomedical industry for the production of pharmaceutical products like hormones or drugs. The second type, called transducers (e.g. biosensors) transform a combination of different external inputs into a single signal that can be easily quantified with reader devices. Those are devices that could be used for example in diagnostic kits, microbiological assays or biodetectors. In both scenarios, this novel architecture allows the construction of modular biocomputers in a flexible, robust and scalable manner.
With the aim of reducing wiring requirements and minimizing in vivo genetic manipulations, we designed a new logic architecture for use in biological circuits. The basis of this architecture is the combination of multiple consortia with distributed computation [31] with the use of spatial confinement [32]. In general, the behavior of a given logic circuit responding to N inputs can be defined by a logic Boolean function described by the so-called truth table. This table defines all possible combinations of inputs and the associated outputs. According to Boolean algebra, the so-called canonical form of the Boolean function can also describe the truth table of a given logic circuit. Independently on the particular circuit analyzed, this canonical form follows the general expression:
f=∑i=1M[∏j=1Nϕij(xj)]
Here Σ represents the OR operator and Π the AND operator. The function ϕij is either a logic representation of the presence of a molecular input xj (Identity function) or of its absence (NOT function). Finally, M is the maximum number of terms present in the Boolean function, which depends on the complexity of the function, but the condition M ≤ 2N-1 is always satisfied [41]. In general, the expression of a Boolean function f can be reduced by the systematic application of standard rules of simplification, such as Karnaugh maps [42] or the Quine-McCluskey algorithm [43]. However, in the biological context easier implementations can be achieved modifying the canonical expression of the Boolean function to obtain an expression involving only OR logic (the simpler logic in a cellular implementation). To reach this goal we propose an alternative formulation of the Boolean function based on the Inverted Logic Formulation (ILF). This formulation minimizes biological constraints ensuring scalability (see S1 Text for a detailed mathematical description).
The canonical form of a general Boolean function can be rewritten applying a double negation, i.e.
f=f¯¯=∑i=1M[∏j=1Nϕij(xj)]¯¯
Applying the Morgan’s Laws [44],
{a OR b¯=a¯ AND b¯a AND b¯=a¯ OR b¯
the Boolean function can be expressed as f=f¯¯=∑i=1M[∑j=1Nθij(xj)¯] where θij(xj)=ϕij(xj)¯ (see S1 Text for a detailed mathematical description). Hence, the Boolean function results in the OR combination of several computational modules ψi, i.e.: f=∑i=1Mψi
Each module ψi is the inversion of OR combinations (symbol Σ) of inverted terms θij, i.e. ψi=∑i=1Nθij(xj)¯ where θij(xj) can be chosen among NOT or Identity functions, i.e.
θij(xj)={xj¯orxj
depending on the specific function to be implemented by the circuit.
This formalism can be easily translated into a biological implementation. Fig 1 shows a schematic representation of the proposed architecture. In the general basic design, a particular logic circuit is composed of M different multicellular consortia (Fig 1A) located in physically isolated chambers {ψ1, ψ2,… ψM}. Here, each module ψi that conform the Boolean function can be biologically implemented by a different multicellular consortia located in a physically isolated chamber. Each consortium contains two different layers of cells, namely the Input Layer (IL) and the Output Layer (OL). The Input Layer is composed of several cell types that sense single external inputs {x1, x2,…xN} and secretes a wiring molecule ω according to a particular internal logic, Identity (ID) or NOT (NOT) logic implementing the θij(xi) functions. When the wiring molecules ω secreted by each cell are mixed in the medium, the OR function (Σ) is implicitly implemented.
The OL consists in a single cell type that responds to the wiring molecule (ω) producing the output β according to a NOT logic, i.e. the output molecule (β) is produced only in the absence of the wiring molecule (ω). Of note, only two types of elementary logic responses are implemented in engineered cells (ID or NOT), yet the logic circuitry is much more sophisticated thanks to the spatial segregation of the consortia. In each consortium, the IL cells produce the same wiring molecule in a shared environment thus implementing an implicit OR logic gate. Combining this OR gate with the NOT gate of the OL cells results in a multi-input NOR gate. Remarkably, only one wire (ω) is needed. The output of the whole circuit is the OR combination of each consortium output. This OR combination can be easily implemented by connecting the chambers that contain each consortium and mixing the output (β) produced using an integrated device. Hence this architecture is optimal for systems where the output is a secreted molecule (e.g. hormone or enzyme). Because the circuit is based on distributed computation, in the presence of a given combination of inputs, the output (β) can be produced in one or more consortia at the same time. Therefore the final output concentration can be different depending on the number of consortia producing it. Despite being a digital approach, in which only the presence (logic state 1) or the absence (logic state 0) of the output molecule is relevant, in a real applied system the total amount of output production could be meaningful. Hence, the total amount of output production should not be dependent on the specific combination of inputs that induces their production. This problem can be solved introducing a buffer cell (BUF cell) that senses β secreted by the different consortia and produces the final output of the circuit according to identity logic. This buffer cell has to be designed that responds at maximum when senses the presence of the output signal (β) from a single consortium. Hence, higher levels of β will not be translated into differentiated output levels (Fig 1B, upper panel). Alternatively, in devices where β is a simple readout of the computation like “transducer circuit” (e.g. biosensors), the final OR could be assessed by quantification of a reporter (e.g. fluorescence) directly expressed in the OL cells using a reader device (e.g. FACS or microscopy). In these cases, a positive signal in any consortia can account as the final positive output of the computation bypassing the need for an integrated device and a buffer cell (Fig 1B, bottom panel).
The main feature of this architecture is its general design, i.e. for a given number (N) of inputs any arbitrary circuit can be built by using the same architecture, independently of its complexity. A simple calculation reveals that, while the upper bound size of the cell library necessary to implement circuits integrating N different inputs scales linearly as Z = 2N+1 (ID and NOT for each input plus the OL cell) and the maximum number of spatial modules (¬i = 1,…M) increases according to M = 2N-1, the number of implementable different logic functions B grows super-exponentially as B = 22N (Fig 1C). For instance, to implement all the 5-input functions (B = 4.294.967.296), only 11 different cell types and, in the most complex scenario, 16 independent modules (M = 16) are needed. Of note, we are here referring to a single output (β) circuit. In multiple-output circuits the upper bound on the number of OL cells is equal to the number of circuit’s output. When using the optional Buffer layer (BL) only one cell type and one additional module are required. Notably, the simple combination of ID and NOT logic cells, when spatially segregated, defines a functional complete set that guarantees that any logic circuit can be built by combining these elements (a formal demonstration of this design and a detailed description of a systematic methodology for logic circuit implementation are included in S1 Fig and in S1 Text. Therefore, any arbitrary logic function can be encoded in a number M of different consortia and in the particular combinations of IL cells in each of these consortia (Z) (Fig 1C). Increasing functional complexity of the logic circuits is translated into an increase in the number of consortia and the corresponding chambers, but not in the number of cell types or wires. In order to demonstrate that our architecture allows scalability together with the minimization of the genetic engineering requirements, we built several logic circuits in eukaryotic cells that respond to up to 6 inputs.
In order to implement modular biocomputing in vivo, we created a minimal library of engineered yeast cells required for the ILs and OLs of the circuits (Fig 2A and S2–S4 Figs). The IL library consists of six pairs of cells that respond to extracellular stimuli, namely: doxycycline (DOX), progesterone (PRO), aldosterone (ALD), mating α-factor from C. albicans (αCa), 17-β-estradiol (EST) and dexamethasone (DEX). The detection of those hormones is done by expressing the specific receptor in each cell type (see full description in S1 Text). Each pair of cells consists of two different types of cells that respond to the same stimulus but with a different logic, either ID or NOT logic, and that secrete a wiring molecule, the S. cerevisiae α-factor pheromone (αSc) accordingly. ID cells secrete the wiring molecule upon stimulation by expressing the MF(α)1 gene from a specific promoter that responds to a defined stimulus. The corresponding pair-wise cell with the NOT logic express the LacI repressor from the same stimulus-specific promoter. NOT cells express the MF(α)1 gene under a modified TEF1 promoter that contains LacI binding sites (PTEF1i) and thus, inhibits αSc expression in the presence of stimuli (Fig 2A and S2A Fig).
Cells in the OL respond to the αSc that is secreted by IL cells, and subsequently express, or do not express a protein (β) (the output of the module), according to NOT logic. In the general architecture, β can be a secreted molecule that performs a determined function (e.g. an enzyme or a hormone) or a fluorescent protein (i.e. GFP or mCherry). Briefly, OL able to secrete molecules (OL3) consist in a cell that constitutively expresses the C. albicans α–factor (αCa) gene, CaMF(a)1, under the TEF1i promoter. The αCa is secreted in the culture media mimicking hormones in biomedical application or protein production in bioreactors. The LacI repressor is transcribed from the FUS1 promoter. Therefore, in the presence of S. cerevisiae α-factor (i.e. the wiring molecule) the LacI repressor is produced and represses the expression of C. albicans α-factor (Fig 2A and S2C Fig). Alternatively, fluorescent OL cells (OL1 and OL2) constitutively express a modified version of a fluorescent protein (yEGFP or mCherry) fused to a degradation tag (ssrA) under the TEF1i promoter [45]. The presence of the wiring molecule induces LacI expression, which leads to down-regulation of fluorescent protein expression. Pheromone (αSc) also stimulates degradation of the fluorescent protein by induction of the ClpXP protease complex that recognizes and degrades ssrA-tagged proteins (Fig 2A and S2B Fig).
Given that the output of the circuit is distributed in different consortia the concentration of the secreted molecule (e.g. αCa) can differ according to the number of consortia simultaneously producing it. In case that the level of the secreted molecule needs to be constrained within given thresholds, we engineered a Buffer Layer cell (BUF), which is designed to produce GFP in the presence of αCa according to Identity logic. This cell contains GFP integrated into the FUS1 gene locus under its promoter (FUS1::GFP-KanMX). BL cells also express the C. albicans pheromone receptor (CaSTE2) so that they can sense the secreted pheromone (Fig 2A and S2D Fig). The buffer cell has been designed to give the maximum response when it detects αCa from a single consortium in a sharp step-like function (S7B Fig). Hence, higher levels of αCa will not be translated into different output levels. The genotype and the graphical notation of the logic function performed by each cell (Input, Output and Buffer Layers) of the library are depicted in Fig 2 and S3 and S4 Figs and in S1 Text.
Once the library of cells was built, we coupled one cell from the IL with one of the OL in the presence or in the absence of the input in order to demonstrate that the wire connection works properly. Two possible scenarios are described: computation of the IL cells in the presence (Fig 2B) or the absence (Fig 2C) of the optional BL cells. Similar output results were obtained using both strategies by measuring the fluorescence of single cells using flow cytometry. The autofluorescence and the percentage of cells able to produce a positive output signal that were fluorescent positive cells were calculated (S5 Fig). These results showed a clear separation between 0 and 1 logic states in the response of the cells.
We then extended the analysis to all of the cell types by measuring their transfer function (i.e. the relationship between different input concentrations and the corresponding output production) (S6 and S7 Figs). Briefly, the transfer function of OL cells was characterized by incubation with increasing concentrations of the input synthetic αSc and measurement of the output fluorescence by FACS (OL1 and OL2), or the output fluorescence resulting from secretion of the αCa (OL3) after incubation with BUF cells. Similarly, the transfer function of the BL cells was characterized by incubation with increasing concentration of synthetic αCa. The transfer function of IL cells was assessed by measurement of output fluorescence upon exposure to increasing levels of each stimulus in the presence of the OL1 cells. These experimental results indicate that all of the cells exhibit a proper behavior that allows definition of a clear separation between 0 and 1 logic states. Applying the same methodology used in electronics, we defined a threshold. Cells producing an output below the threshold are in the 0 logic state, whereas if the output is above this threshold is considered in the 1 logic state. This threshold is the same for all the cells and circuits analyzed.
Based on these results, we established the concentration of inputs used in the circuits (see below) so that they were clearly above the threshold in order to guarantee a correct response of the cells. Also, based on these transfer functions results, we determined a specific range of time for the response to input signals used in the circuits to ensure a robust response. When working with cellular consortia, it is critical for the system to work robustly that cells within a consortium display similar growth rates. Thus, we assessed the growth rate of each cell type and found no major differences within the entire cell library (S8 and S9 Figs) suggesting that the different consortia should not display unbalanced cell growth of any of the components.
A potential thread to implement complex biological circuits is crosstalk between cells. We therefore assessed crosstalk between the IL cells in response to different single inputs or to all of the inputs combined. Each IL cell type was mixed with OL1 cells and then treated separately with every input. The percentage of GFP positive cells was measured using FACS. Each cell type responds only to its own stimulus and secretes the pheromone only in the presence (ID) or in the absence (NOT) of the specific input. Finally, we incubated every cell type with all of the inputs to which it should not respond (ALL-I). Even in this scenario, no significant crosstalk was observed (Fig 2D). Therefore the crosstalk between the IL cells upon different inputs was not significant.
We then implemented a number of 2-input logic gates to test the combination of several cell types from the library. Here, just to test the cells we measured the output of the logic circuits as the GFP production of the OL1 cell (S10 Fig) and found that the cells computed correctly when an AND, NOR or N-IMPLIES gates were assessed.
As an example of how this modular architecture works we built the majority rule device (Fig 3), by testing it in circuits with a secretable output. This three inputs circuit is a decision-making system used in electronics as a security device against failure in redundant systems. The formal representation of the circuits is shown in S14A and S14B Fig. Using our library of cells, we implemented it as a device that detects when at least two molecules out of three are present. Determine the unbalance between molecules concentrations could be of interests in biomedical applications. The equivalent, single-cell type design of a majority rule would be very difficult to build in vivo [46]. To define the best cell combination from the library within the different modules, the design of the logic circuit is first done in silico which ensure the use of the correct combination of cells (see S1 Text for a detail description of a systematic methodology for logic circuit implementation). Following the basic architecture described above, implementation required just three different multicellular consortia, ψ1, ψ2 and ψ3, formed combining three IL cell types with the OL3 cell. Production of the secretable molecule from the independent modules was sensed by the BL cell (Fig 3A, left).
A key element in the proposed architecture is the spatial segregation of the different modules. Here, the final OR computation is done physically connecting the modules and collecting the output (i.e. a secreted molecule). To this purpose we built an open-flow computing device (Fig 3B) with physically isolated chambers and able to collect and integrate the outputs (here αCa) from the different consortia of the circuit. The different consortia were assembled in three independent computational chambers (ψ1, ψ2 and ψ3) and exposed to the same combination of the three inputs (x1 = DEX, x2 = EST, x3 = PRO). The buffer cells were incubated in the Buffer chamber (BUF). After a transitory computational time, the device is programmed to gather the fluxes of αCa produced by the independent consortia in the Buffer chamber, thereby performing the final OR computation. IL cells were prevented to enter into the OR chamber by positioning a filter before the Buffer chamber. The final output of the circuit, stored in the Buffer chamber, was quantified as % of GFP positive BL cells using both microscopy and flow cytometry (Fig 3C, grey bars). All the eight possible combinations of inputs were tested and the final outcome of the computation was as expected for a majority rule circuit: only when at least two of the inputs were present there was a positive output.
The open-flow computing device is an example of an integrated system able to implement in vivo circuits with a secreted output. The second type of devices that can be implemented with the this architecture is the transducers circuits. These circuits are devoted to translate a complex combination of multiple external inputs into a single output signal. The architecture of transducer circuits is simpler because they do not require the final output integration. These circuits can be built as an array of separated modules (chambers) that produce the same fluorescent protein as an output which, in turns, is measured by an external readout system. More specifically, we assessed the output by direct quantification of the OL1 or OL2 fluorescence using microscopy and flow cytometry as reader devices. The final output of the circuit was considered positive whenever any consortia gave a positive fluorescent signal above the 1 logic state threshold, bypassing the need of a full system integration. As a first example of a transducer, we measured the output of the same majority rule circuit using OL1 instead of OL3 (Fig 3A, right) mimicking a biosensing circuit. Fig 3C, green bars, shows that the circuit responded similar to the open-flow computing device even using a different type of OL cell. Thus, a combination of cell types with the proper design resulted in a device that was capable of implementing a majority rule circuit in vivo. The final result of the computation of the circuit is given as % of GFP cells with fluorescence below (0 state) or above (1 state) the logic threshold previously defined. Alternatively, the output could be measured in terms of total GFP fluorescence (in arbitrary units). We demonstrate that for our library of cells, both metrics are qualitatively equivalent (S12 Fig and S1 Text). A detailed description of the measurement procedures and outcome circuit production is included in the Material and Methods section and S11 Fig. To show the flexibility and robustness of the cells library, we built the same circuit using IL7 cells, which respond to DOX, instead of IL12 cells, which respond to DEX. S13 Fig shows that the logic circuit responded similarly and reliably even swapping cells from the library to respond to different inputs with the same logic, indicating the robustness of the circuit response to cell variation.
To explore the potentiality of this approach, we investigated whether more complex devices could be achieved using biosensing devices as reference, since their implementation is simpler in the laboratory when all the input combinations need to be measured. We increased circuit complexity by creating a circuit that responds to four different inputs by producing two different outputs. We chose a 2-bit magnitude comparator, which permits the comparison of two binary numbers, A and B, each having two bits (A = {a1, a0} and B = {b1, b0}). Comparators are at the heart of most central processing units (CPUs) in computers and perform a large portion of the logical operations. The circuit is able to respond to 4 inputs, upon 16 entries, and yields three different outcomes from the computation (A>B, A<B and A = B). The formal representation of the circuits is shown in S14C and S14D Fig. The implementation of such a circuit in vivo required six different consortia and different combinations of four pairs of IL cells. Cells respond to four stimuli (DOX, EST, PRO and DEX), where EST and DOX encoded A, and PRO and DEX encoded B. Of note, this circuit has an additional level of complexity because it requires two different outputs to distinguish between A<B, A>B and A = B. Therefore, we used two OL cell types, that express green (GFP, OL1) or red (mCherry, OL2) reporters (Fig 4A). The expected output would be green when A<B, red when A>B and no signal when A = B (Fig 4B). After incubation with the inputs, the fluorescence of the cell consortia was assessed and the final computation was calculated by measuring the percentage of mCherry positive cells present in the first three chambers (A>B) and the percentage of GFP positive cells in the last three chambers (A<B). All 16 combinations yielded the expected outcome, supporting the notion that multiple functions can be constructed from a small library of reusable cells.
To demonstrate the scalability of this modular approach and exploit the capability of our library of cells, we implemented a highly complex multiplexer involving 6 inputs. Of note, such computational complexity has never been reached so far in biological circuits. A multiplexer permits the sharing of one device by several signals thereby avoiding the necessity of having one device per input signal. The MUX 4-to-1 is a circuit that responds to 6 inputs. 2 of these 6 inputs are called selectors because they allow the “selection” of which one of the other 4 inputs will determine the final output. Here, PRO and DOX are the selectors (S0-S1) and ALD, αCa, EST and DEX are the inputs (I0-I3). For example, when both PRO and DOX are equal to zero (S0 = 0, S1 = 0), the selected input is ALD (I0) as indicated in the true table (Fig 5A, bottom). Thus, the circuit will produce the 0 output when ALD is equal to 0, and an output of 1 when ALD is equal to 1 (violet row in the truth table, Fig 5C, bottom). Thus, a total of 64 combinations of inputs are possible. The formal representation of the circuits is shown in S14E and S14F Fig. This circuit, which would represent an enormous effort if it was built in a single cell using standard design methods (e.g. S15 Fig), can be assembled by involving just eight IL cells and one OL cell combined in four spatially independent consortia (Fig 5B). Similarly, we directly measured the output from the modules using the fluorescent OL1 cells and microscopy and flow cytometry as reader devices. A mixture of the six inputs was simultaneously added to the four chambers and, after incubation, the fluorescence of the consortia was measured using FACS and microscopy. All the 64 combinations of inputs were tested and the final computation was assessed as before. Although the complexity of the circuit required differential outputs to 64 different input combinations, the in vivo results clearly showed the expected response (Fig 5C).
A major challenge in the field of synthetic biology is the construction of flexible, scalable and complex logic circuits using engineered cells. Many different strategies have been implemented to create logic circuits in biological systems over the last decade [6]. However, several problems, including those derived from wiring requirements, pose a serious limitation on scalability [6]. Some approaches have been advanced to overcome these obstacles, including the use of multicellular distributed computation [31] and the use of spatially restricted computational modules [32,38,47].
Here, we propose a novel alternative to the standard architecture that combines three elements to create new circuits in a strategic manner: 1) the use of multicellular consortia, 2) spatial segregation and 3) distributed output computation. On top of this, circuit design does not follow electronic standard methodology but rather we implemented a new method that permitted to obtain the maximum benefits of the combination of the three elements (i.e. inverted logic). This approach uses the simplest logic devices, i.e. one-input one-output logic gates, connects the cells of each consortium with a single wire, and allows each consortium to produce the final output independently of the others. This new architecture has several appealing properties. On one hand, using a minimal library of cells, several combinations of multicellular consortia can be assembled (modularity). Modular biocomputing profits from the enormous potential of combining a limited number of building blocks (IL and OL cells), which is comparable with the combinatory richness of standard microelectronics. Once the library of cells has been built, different combinations of the same cells can create novel circuits without additional engineering, thereby pushing the concept of reusability of parts one step further. This combination of cells allows an exponential increase in the number of different circuits available without additional engineering.
There are however many aspects that need to be taken into account when designing and implementing logic circuits with biological consortia. An extensive cell characterization is important to create proper cell libraries. It is important that the cells within the same module display similar dynamic responses that can be easily deciphered by the characterization of transfer functions. A possible alternatively, could be to introduce a certain nonlinear circuit such as a toggle in the IL cells [48]. Also, it is essential that each cell responds to only one input and thus crosstalk has to be avoided. The balance of cells within the consortia is also a key point for long term circuit responses. It is critical that no cellular imbalances occur and thus, cellular growth should be similar among cells in a module. After library construction and characterization, that serves to create cells responding to the desired input, logic design of the circuit can be established by defining the best cell combination from the library within the different modules. In general, a given circuit can be implemented by different cells and modules combination, which is optimized in silico to reduce the number of cells and modules to facilitate in vivo implementation. The reusability of the cells within a library depends on the input that needs to be sensed, however, our data indicates that thanks to the simple logic of each cell (ID and NOT), cells can be created and swapped easily if they maintain similar dynamic responses as described before.
A crucial property of this architecture is that it does not depend on the complexity of the circuit to be implemented, thereby ensuring virtual unlimited scalability, yet maintaining minimal genetic engineering requirements. For instance, a library that responds to six inputs as reported here is sufficient to create up to 1.8x1019 different circuits, with a maximum of 32 modules in the worst complex scenario. As a proof of principle, we have built several logic circuits in eukaryotic cells that respond to up to 6 inputs (such as a multiplexer 4-to-1), and that reach an increase in complexity that has not been implemented before. This design shows how a modular biocomputer can be constructed in a flexible, robust and scalable manner. In addition, computation performed by multicellular consortia opens the door to exploration of circuits obtained by combining different cellular species and the synergies that can be derived from this coexistence.
Remarkably, modular biocomputing is flexible to different types of applications; for instance, it can be use to build circuits that function as biosensors and the output of the computation is assessed through a reporter system. These are devices where the circuit outcome can be assessed by microscopy (e.g. microfluidic devices), or biodetectors, analytical and microbiological assays, and diagnostic kits. Still, there are scenarios where the output is a secreted molecule with a biological function, e.g. recombinant proteins or enzymes produced in bioreactors, chemical compounds and metabolites in industrial biotechnology or pharmaceutical products (hormones or drugs) in biomedical applications. By expanding the library with only two simple cells we showed how our design can be extended to such applications as well. Finally, we built an open-flow computing apparatus as a proof of principle of an integrated device that upgrades the potential of our architecture to circuits with a secreted output. Depending on the application, the user may require different devices with different proprieties and a diverse level of spatial isolation within each module. For example, in biotechnological applications, the production of toxic by-products by one cell type in a module may inhibit other cell type, thus affecting the computational capability of the consortia. In such cases, the implementation will require isolation of the different cell types in each module where the toxic product is trapped but the wiring molecule can flow. Also, unbalanced growth rates of different cells types calls for devices where the culture growth can be maintain constant like in a chemostat. All together these concerns are pushing the field towards a personalized device technology where users design and build their own devices specifically optimized for the desired application. Lately, many possible micro-environments, such as microfluidics devices [49–51], cell microcapsules [52], micro-fabricated implantable arrays [53] and cell culture patterns [54] have been improved. Recent advancements in photolithography, plastic molding and, recently, in 3D-printing might lead to custom-designed microdevices easily available for biomedical applications. Coupling these technologies with modular biocomputing design can provide a general and robust way of exploring the landscape of living computational devices.
Yeast W303 (ade2-1 his3-11,15 leu2-3,112 trp1-1 ura3-1 can1-100) cells were genetically modified so that they could produce αSc from an inducible promoter (IL cells), control output expression (fluorescent proteins or αCa) in response to the αSc (OL cells), or produce a fluorescent protein in response to αCa (Buffer cells). Schematic genotypic characteristics of each cell and plasmid used are summarized in S1 Text, S3 and S4 Figs and S1 and S2 Tables. The cells within a consortium can be followed by specific markers or the presence of fluorescent reporters.
Overnight cultures were diluted to OD660 nm ≈ 0.2 and were grown at 30°C in YPD or selective medium.
We followed standard electronics for defining a positive signal from a circuit as described [31]. As shown in Figs 3–5, in our biological devices the resolution of the 1 logic is more than 60% and 0 logic is less than 20% of the maximal value, indicating that these circuits are comparable with electronics in terms of resolution. However, this separation between logic states is a necessary but not a sufficient condition to guarantee that multicellular circuits can be implemented that connect different cells acting as logic blocks. A proper characterization of the library of engineered cells is necessary to analyze the so-called Transfer Function, i.e. the cellular response with respect to different input levels. An adequate Transfer Function should be characterized by several key features [55,56]: i) a step-like shape, ii) linear or higher gain ranges in order to ensure that the signal will not be degraded from input to output in a single cell, iii) the noise margins must be adequate, without overlap between the high and the low state, and iv) each cell must respond properly only to the specific inputs and must ignore the rest of inputs of the circuit. All these aspects have been experimentally addressed in the set of engineered cells of the library. S6 and S7 Figs show the full set of transfer functions for each cell. Experimental data were fitted to a Hill equation as described in S1 Text and S3 Table. All these curves exhibit the proper shape to be logic blocks for a multicellular implementation. This procedure allows characterization not only of cellular behavior but also of the wire efficiency. Cells were grown in selective media or YPD to mid exponential phase and were then diluted to OD660 nm ≈ 0.2. Input Layer (NOT) cells were washed to remove the αSc that was produced o/n and were resuspended in YPD. Each Input Layer cell was mixed with the GFP Output Layer cell (OL1) at a 4:1 ratio and the mixture was subjected to different concentrations of input (S6 Fig). OL3 cells were washed to remove the αCa that was produced o/n and were resuspended in YPD. OL3 cells were then mixed with the Buffer Layer cells (BL) at a 4:1 ratio and were subjected to different concentrations of αSc (S7A Fig (bottom)). Various concentrations of αSc were added to OL1 and OL2 cells (S7A Fig (top)) and different concentrations of αCa factor were added to BL cells (S7B Fig). Samples were incubated for 4 h at 30°C and were analyzed using flow cytometry. Data are expressed as the percentage of GFP positive cells. The transfer function represents the mean and standard deviation of three independent experiments. All of the cells exhibit a proper behavior that allows definition of a clear threshold between 0 and 1 logic states. Based on these results, we established the concentration of inputs used in the circuits (arrow in S6 and S7 Figs) to be clearly above the threshold.
Output of the circuits, transfer function and crosstalk were analyzed after 4 h incubation at 30°C with a combination of inputs unless specified differently. Samples were diluted in PBS and analyzed using flow cytometry (BD LSRFortessa). A total of 10.000 cells were collected from each sample. Constitutive fluorescence in Output Layer cells (mCherry for OL1 and YFP for OL2), was used to differentiate them from Input Layer cells (S5B and S5D Fig). In S7A Fig, bottom, constitutive fluorescence in the Buffer Layer cells (mCherry) was used to differentiate them from the OL3 cells. Specific emission in the fluorescence channel of the subsets of Output or Buffer Layer cells was measured versus autofluorescence (PerCP-Cy5-5-A channel for GFP and YFP, PerCP-Cy7 channel for mCherry). Autofluorescence in a wild type strain without carrying any reporter was measured as a reference (S5A Fig). A gate was set to subtract autofluorescence and cells inside the gate were considered as GFP positive cells. Data are expressed as percentage of fluorescent positive cells (GFP for OL1 and BL, mCherry for OL2) (S5C and S5E Fig). Also measured in a shift on total fluorescence (S5F Fig). An output expression below the 20% of GFP positive cells corresponded to the 0 logic state (low threshold) and above the 60% of GFP positive cells corresponded to the 1 logic state (high threshold). In all the circuits, we use the same low and high threshold values. Data were analyzed using FlowJo or BD FACSDiva software. A representative FACS plot of our quantification method is presented in S5 and S11 Figs. For microscopic analyses, cells were harvested and resuspended in Low Fluorescent Media. Images were collected with a Nikon Eclipse Ti Microscope using NIS elements Software (Nikon) and were analyzed using ImageJ.
Fig 2D shows the individual cellular response of each IL cell in response to the different single inputs they encounter within a circuit or to all of the inputs combined. Cells were grown in selective media or YPD to mid exponential phase (OD660 nm ≈ 0.2). Input Layer (ID) cells were mixed with the GFP Output Layer cells at a 2:1 ratio. Input Layer (NOT) cells were washed to remove the S. cerevisiae alpha factor that was produced o/n, were resuspended in YPD and were mixed with the GFP Output Layer cells at a 3:1 ratio. Each mixture was subjected to all 6 inputs individually, to all 6 inputs together (ALL) and to all inputs except for the specific input (ALL-I). Samples were incubated for 4 h at 30°C and were analyzed using flow cytometry. Data are expressed as the percentage of GFP positive cells. The experimental data shows that there is no undesired crosstalk and that each cell responds only to the expected input.
The open flow device (Fig 3B) is composed of three parts: the computational chambers (ψ1, ψ2, ψ3), the valves, and the Buffer chamber (BUF). The computational chambers are tanks with 4.5ml liquid storage and a cup allowing pneumatic actuation of the fluids (Microfluidic ChipShop). 1.6 mm tygon tubes connect the air pump (CellASIC ONIX Control System) with the tanks cup using male mini-luer connectors. The fluidic interface is realized as female luer connector. Valves (Discofix Braun) can be turned in three different positions (p1: waste, p2: closed and p3: Buffer) according to the different experimental steps. The Buffer chamber is a 4.5ml tank with a pneumatic cap carrying three male mini luers. The interconnection between the components is enables by 1.6 mm tygon tubes, male and female luers and mini luers. To prevent the cells mixture to enter the OR chamber, but still allowing the transferring of the supernatant, a 0.22 mm Millipore filter is plugged in before the OR chamber. Finally, a device carrier has been designed and built to hold the apparatus.
Fig 2B shows the single cell computation in the presence of the optional Buffer Layer cells. Input Layer cells were mixed with OL3 cells in the absence or presence of the specific input. After 4 h of computation the supernatant of the mix was added to the Buffer Layer cells, incubated for 4 h and the percentage of GFP positive BL cells was analyzed using FACS. Fig 3B and 3C show the implementation of the major rule circuit using the optional BL and the open flow device. The appropriate combinations of IL cells were mixed proportionally into the three chambers, together with the OL3 cells, and exposed to the same combination of the three inputs. After 7h of incubation stirring at RT (transitory time), we pumped into the chambers fresh media with the corresponding combination of inputs (psi: 0.5, valve: p1, minutes: 3). The valve was then turned to p2 and the cells mixture was incubated for 10 h stirring at RT (computational time). Finally, the αCa produced by the independent modules was automatically collected in the Buffer chamber (psi: 5, valve: p3, minutes: 1) and incubated with the BL cells for 4 h at 30°C. Samples were analyzed using FACS and microscopy. We repeated the same experiment in triplicate for each combination of inputs of the majority rule.
Circuits in Figs 3–5 were built mixing proportionally the appropriate combination of IL and OL cells in different tubes (i.e. the consortia). The same mixture of inputs was simultaneously added to each consortium. All the possible combinations of inputs were tested. After 4h of computation at 30°C, for each combination of inputs, the percentage of GFP positive cells in each module was analyzed using FACS and microscope. A positive signal (more than 60%) in any consortia accounts for a 1 as the final output of the circuit. When more than one consortium gave a positive fluorescent signal we choose the highest value. The same was done for negative (less than 20%) outputs (0) (S11 Fig). Data represent the mean and standard error of three independent experiments.
|
10.1371/journal.pgen.1006310 | Loss of the Greatwall Kinase Weakens the Spindle Assembly Checkpoint | The Greatwall kinase/Mastl is an essential gene that indirectly inhibits the phosphatase activity toward mitotic Cdk1 substrates. Here we show that although Mastl knockout (MastlNULL) MEFs enter mitosis, they progress through mitosis without completing cytokinesis despite the presence of misaligned chromosomes, which causes chromosome segregation defects. Furthermore, we uncover the requirement of Mastl for robust spindle assembly checkpoint (SAC) maintenance since the duration of mitotic arrest caused by microtubule poisons in MastlNULL MEFs is shortened, which correlates with premature disappearance of the essential SAC protein Mad1 at the kinetochores. Notably, MastlNULL MEFs display reduced phosphorylation of a number of proteins in mitosis, which include the essential SAC kinase MPS1. We further demonstrate that Mastl is required for multi-site phosphorylation of MPS1 as well as robust MPS1 kinase activity in mitosis. In contrast, treatment of MastlNULL cells with the phosphatase inhibitor okadaic acid (OKA) rescues the defects in MPS1 kinase activity, mislocalization of phospho-MPS1 as well as Mad1 at the kinetochore, and premature SAC silencing. Moreover, using in vitro dephosphorylation assays, we demonstrate that Mastl promotes persistent MPS1 phosphorylation by inhibiting PP2A/B55-mediated MPS1 dephosphorylation rather than affecting Cdk1 kinase activity. Our findings establish a key regulatory function of the Greatwall kinase/Mastl->PP2A/B55 pathway in preventing premature SAC silencing.
| Cdk1 phosphorylates many substrates in mitosis and simultaneoulsy reduces the activity of the corresponding phosphatase PP2A through the Greatwall kinase/Mastl. When Mastl is deleted, cells progress through mitosis with missegregated chromosomes, which become unraveled. However, the molecular mechansims by which Mastl promotes proper chromosome segregation and mitotic progression remain elusive. In this study, we show that the Cdk1->Greatwall kinase/Mastl->PP2A pathway plays a central role in regulating the spindle assembly checkpoint (SAC) by preventing premature SAC silencing. We further demonstrate that Mastl is required for multi-site phosphorylation of the essntial SAC protein MPS1 as well as robust MPS1 kinase activity in mitosis by inhibiting PP2A/B55-mediated MPS1 dephosphorylation. Our findings establish the requirement of Mastl for robust SAC maintenance.
| The activity of Cdk1/cyclin B is essential for cells to enter and complete mitosis. As recently shown in Xenopus and Drosophila, the phosphatase activity that dephosphorylates Cdk1 substrates is inhibited simultaneously with the peak of Cdk1 activity when cells enter mitosis to ensure maximal phosphorylation of Cdk1 substrates. Cdk1 phosphorylates and activates the Greatwall kinase/Mastl, which then phosphorylates Ensa or Arpp19 enabling them to bind and inhibit the phosphatase PP2A/B55 [1–4]. The Greatwall kinase is required for entry into mitosis in Xenopus [5] and similarly in human cells when Mastl was silenced completely [6], whereas mouse cells deleted for Mastl were reported to enter mitosis [7]. In contrast to mitotic entry, there is agreement that Mastl is important after nuclear envelope breakdown (NEBD) for exit from mitosis and cytokinesis [6–9]. In the context of Mastl deletion, the early mitotic defects have not been defined precisely and this is compounded by a lack of knowledge of specific PP2A substrates, which are dephosphorylated in the absence of Mastl. The only known target of Gwl/Mastl->PP2A is PRC1, an essential component assembling the central spindle during mitotic exit, with Thr481 being dephosphorylated by PP2A/B55 [8]. Therefore, identifying specific targets of the Greatwall kinase/Mastl->PP2A/B55 pathway is essential for understanding its in vivo functions.
One of the functions of Cdk1 is related to the spindle assembly checkpoint (SAC), which must be activated every time cells enter mitosis but needs to be silenced after all chromosomes have been properly attached to microtubules (for reviews see [10–12]). Nevertheless, the mechanism of silencing SAC, at a time where Cdk1 activity is still high, remains an open question. Recent studies elegantly demonstrated that besides cyclin B1 degradation, dephosphorylation of Cdk1 substrates is essential for regulation of the SAC and progression through anaphase [13–15]. Identifying the Cdk1-phosphorylated targets that need to be dephosphorylated to silence SAC is a major challenge, although the kinase MPS1 was suggested to be a potential candidate [13,16].
In this study, using conditional knockout mice for Mastl, we show the requirement of Mastl for robust spindle assembly checkpoint (SAC) maintenance. Using mass spectrometry, we have identified several mitotic targets of Cdk1 phosphorylation including MPS1 that are prematurely dephosphorylated in MastlNULL MEFs without altering the overall activity of Cdk1. Notably, we show that in MEFs lacking Mastl, mitotic multi-site phosphorylation of MPS1 as well as its kinase activity are compromised, which can be directly regulated by Cdk1 and PP2A/B55. Our findings reveal that the Cdk1->Greatwall kinase/Mastl->PP2A/B55 pathway controls mitotic phosphorylation of MPS1 and is essential for full kinase activity of MPS1 during mitosis, which may partly explain the requirement of Mastl for robust SAC maintenance.
We generated a conditional (hereafter referred to as MastlFLOX) knockout mouse model for the Mastl gene by inserting LoxP recombination sites on both sides of exon 4 in the mouse Mastl genomic locus [17]. In brief, the deletion of Mastl gene (hereafter referred to as MastlNULL) was obtained by excision of the exon 4 from the control MastlFLOX allele using a constitutively expressed (β-actin-Cre), tissue specific (Albumin-Cre), or tamoxifen inducible [Rosa26-CreERT2 or Esr1 (CreERT2)] Cre recombinase. Germ line deletion of the Mastl gene resulted in embryonic lethality (S1A Fig). In depth investigation of various developmental stages in embryos revealed that early stage embryos (E3.5 blastocysts) were viable and succeeded to implant normally (S1B Fig). However, in comparison to control embryos, their growth was arrested before E7.5 and they did not develop further (S1B and S1C Fig). To exclude that functions of Mastl are limited to early embryonic development only, we utilized conditional knockout strategies in more advanced stage embryos and in different tissues in adult mice using the inducible Rosa26-CreERT2 (S1D and S1E Fig) as well as specifically in liver using Albumin-Cre (S1E and S4 Figs). Deletion of Mastl starting from E10.5 led also to embryonic lethality resulting in reduced cellularity in all organs analysed as well as haemorrhaging in the embryos (S1D Fig). Liver specific deletion of Mastl in hepatocytes induced abnormalities in their nuclear morphology (S1Ei and S1Eii Fig). Deletion of Mastl in adult mice by tamoxifen injection resulted in lethality within 7–8 days, accompanied with severe degeneration of the crypt morphology in the intestine (S1Eiii and S1Eiv Fig). The latter results were similar to what we observed in Cdk1NULL mice [18]. These observations suggest that the Mastl kinase is essential for proliferation in adult organs and during the developmental stages we tested.
To investigate whether Mastl is required for cell proliferation in vitro, we isolated MastlFLOX/FLOX mouse embryonic fibroblasts (MEFs) carrying the Esr1 (CreERT2) transgene. Induction of Cre-mediated recombination in the Mastl locus by addition of 4-hydroxytamoxifen (4-OHT) resulted in the loss of the Mastl mRNA and protein (S2A Fig). Primary MastlNULL MEFs were unable to proliferate (Fig 1A) and displayed multi-lobular nuclear morphology (Fig 1D and 1F, S2B Fig). Further analysis of the cell cycle profile of synchronized MEFs by FACS indicated that although they progressed through S phase, MastlNULL MEFs displayed an increased G2/M population, as well as polyploidy and cell death (S2C Fig). Because of the increased G2/M population in absence of Mastl, we investigated whether these cells can enter mitosis. Unlike Cdk1NULL [18], MastlNULL MEFs did enter mitosis albeit with a delay (Fig 1B and 1C). However, time-lapse microscopy of MastlNULL MEFs stably expressing histone H2B-YFP (chromosome marker) revealed that these cells progressed through mitosis despite the presence of misaligned chromosomes (Fig 1D, S1 and S2 Movies) but never completed cytokinesis. Quantification from time-lapse images indicated that more than 90% of MastlNULL MEFs progressed through mitosis with anaphase bridges whereas ~10% of control MastlFLOX MEFs did so (Fig 1E). This mitotic phenotype with anaphase bridges resulted in binucleated cells (Fig 1D) or rupture of the nucleus bearing micronuclei in MastlNULL MEFs (Fig 1F).
Biochemical comparison of synchronized MastlFLOX and MastlNULL MEFs indicated comparable expression levels of cyclins, Cdks, and inhibitory phosphorylation of Cdk1 on Y15 (S3A Fig). Furthermore, kinase activity associated with immunoprecipitated Cdk1 and cyclin B1 was not decreased in MastlNULL compared to MastlFLOX MEFs. In contrast, Cdk2 and cyclin A2 associated kinase activity was slightly reduced in MastlNULL MEFs (S3B and S3C Fig). These results suggest that the observed accumulation in G2/M is unlikely due to a decrease in Cdk1 kinase activity.
To further confirm our results in an in vivo model of cell division, we deleted Mastl specifically in hepatocytes using Albumin-Cre and performed partial hepatectomy (PHx) in these mice (S4 Fig). Similar to Cdk1 [18] and cyclin A2 (S4D Fig), the expression of the Mastl protein increased between 24 and 72 hours after PHx in MastlFLOX mice, but not in MastlNULL liver at 48 hours after PHx, as expected (S4D and S4E Fig). Analysis of histological sections taken 48 hours post PHx displayed an elevated mitotic index as judged by phopho-histone H3 staining (S4B Fig). MastlNULL hepatocytes were unable to divide properly and displayed frequently binucleated morphology as well as anaphase bridges (S4A–S4C Fig). The phenotype of MastlNULL observed in vivo (hepatocytes) is reminiscent of MastlNULL MEFs (see Fig 1), indicating that completely different cell types react similarly to the deletion of Mastl.
Quantitatively measuring the entry and the length of mitosis using time-lapse video microscopy analysis revealed a slight but measurable delay in mitotic entry in unperturbed MastlNULL MEFs but despite this, the majority of cells entered mitosis (Fig 2A). This delay was further verified by quantification of mitotic cells by FACS analysis (Fig 2B). Moreover, in unperturbed cells, the duration of mitosis was somewhat prolonged in MastlNULL MEFs (Fig 2C) most likely due to poorly aligned chromosomes at the metaphase plate (S1 and S2 Movies). Although this delay in mitotic entry was one of three major phenotypes in MastlNULL MEFs; i) mitotic entry defects, ii) premature SAC silencing [see below], and iii) cytokinesis defects [8]; it will not be further investigated in this manuscript. Nonetheless, once MastlNULL MEFs entered mitosis, surprisingly they progressed through mitosis in the presence of chromosome segregation defects (Fig 1D–1F, S1 and S2 Movies) but never completed cytokinesis properly, which resulted in accumulation of MastlNULL MEFs with ≥4N DNA content (S2C Fig).
Under normal conditions, robust SAC signalling ensures mitotic arrest until these attachment errors are properly corrected [10]. To address whether this premature anaphase onset with misaligned chromosomes was due to weakened SAC signalling, MEFs were treated with microtubule poisons including the microtubule depolymerizer nocodazole and the percentage of cells arrested in mitosis was analysed by FACS. Under these conditions, cells are arrested in a prometaphase-like state until eventually cyclin B1 degradation and subsequent chromosome decondensation without separation takes place; a process called mitotic slippage [19]. Time-lapse microscopy analysis indicated that in nocodazole treated MastlNULL MEFs the mitotic duration was reduced compared to control MastlFLOX MEFs (Fig 2D). Furthermore, analysis of the percentage of cells remaining in mitosis measured as positive for phospho-histone H3 on Ser10 (pH3) staining using FACS confirmed that nocodazole treated MastlNULL MEFs exited mitosis significantly faster than control cells (Fig 2E, S5A Fig). Likewise, immunoblotting of whole cell lysates of these mitotic cells indicate a reduction of the pH3 level and that the degradation of cyclin B1 was accelerated in absence of Mastl in comparison to control cells (Fig 2F). This phenotype of mitotic slippage suggests that Mastl is required for robust and persistent SAC signalling.
Since numerous proteins are involved in the SAC (for a discussion see [10]), we aimed to identify specific proteins whose phosphorylation status is changed when Mastl is knocked out and therefore employed mass spectrometry-based quantitative phospho-proteomic analysis. To achieve this, we collected mitotically arrested MastlFLOX and MastlNULL MEFs cultured in SILAC media and compared their phospho-proteomes. Analysis of each sample was performed in duplicate and by reversing the isotope labelling between conditions in the second replicate (for detailed information see Materials and Methods). Our results indicated that while 136 of the identified phosphorylation sites did not change their phosphorylation level (Fig 3A, black dots), 14 phosphorylation sites corresponding to 12 different proteins were specifically decreased in MastlNULL cells with at least 1.5 fold-change (Fig 3A, blue dots and S1 Table), consistent with Mastl regulating either directly or indirectly a phosphatase activity. This “mini” screen was not saturating for technical reasons but among the proteins with decreased phosphorylation status, we identified TTK [20], a dual specificity kinase also known as MPS1 [21]. We identified serine 820 (821 in Human) of MPS1 as the site with decreased phosphorylation (Fig 3B). Since phosphorylation of MPS1 is required for SAC signalling and proper chromosome segregation and its phosphorylation on S820 peaks during mitosis [16,22–28], we decided to focus our analysis on the regulation of MPS1 activity in SAC signalling. To connect the functions of MPS1 to the observed phenotype in MastlNULL, we treated nocodazole-arrested MastlFLOX MEFs with the MPS1 inhibitor reversine [29] (S5B Fig, blue bars). This led to rapid mitotic slippage similar to what we have observed in MastlNULL MEFs, indicating that MPS1 could be responsible in principle for some of the phenotypes observed in the absence of Mastl.
The regulation of MPS1 activity and its phosphorylation is complex and not fully understood [16,22,30,31]. For example, T675 and T685 have been suggested to be autophosphorylation sites essential for MPS1 activity, while S820 is more likely phosphorylated by other kinases to regulate its localization [16,24]. To determine whether MPS1 functions are deregulated in MastlNULL MEFs, immunoblot analysis using phospho-specific antibodies against MPS1 were performed. In MastlNULL MEFs harvested by mitotic shake-off after nocodazole treatment, the levels of phosphorylated S820 was markedly reduced in comparison to MastlFLOX MEFs (Fig 3C). To determine whether S820 could be phosphorylated directly by Cdk1, an in vitro kinase assay was performed using recombinant Cdk1/cyclinB1 complexes and a GST-MPS1 fusion protein as substrate. In order to express this fusion protein in bacteria, and simultaneously avoid interfering signals from the numerous potential phosphorylation sites of MPS1 including autophosphorylation, we fused a short peptide of MPS1 (residues 811–831) containing S820 downstream of a GST tag. Cdk1/cyclin B1 phosphorylated S820 but not the non-phosphorylatable mutant S820A (Fig 3D) as expected [16]. In contrast to Cdk1, Mastl was not able to directly phosphorylate S820 in vitro (S6B Fig). Together, these results confirm that Cdk1 phosphorylates MPS1 [16] and agree with the increased phosphorylation of S820 in nocodazole treated cells (Fig 3C, lanes 1 & 4), although other kinases including MAPK may also phosphorylate MPS1 at S820 or other sites. Since there is no clear consensus on how the different phosphorylation sites in MPS1 affect its kinase activity, we aimed to measure the change in total MPS1 kinase activity directly. To achieve this, we immunoprecipitated MPS1 and determined its kinase activity towards the substrate myelin basic protein (MBP). MPS1 kinase activity was substantially reduced in mitotic extracts from MastlNULL MEFs compared to MastlFLOX MEFs (Fig 3E, compare lanes 2 & 5). Since Mastl regulates the PP2A phosphatase activity through ENSA/Arpp19 [1–3], we evaluated whether the decrease in MPS1 activity was due to the increase in PP2A activity in MastlNULL MEFs. Indeed, mitotic extracts from MastlNULL MEFs treated with OKA restored MPS1 activity similar to the levels seen in MastlFLOX MEFs (Fig 3E, lane 6 and S6A Fig, lane 6). Consistent with this result, MastlNULL and MastlFLOX MEFs when treated with OKA display a similar mitotic slippage rate (S5C Fig). Although we have already shown that Cdk1 activity was not changed in absence of Mastl (S3B and S3C Fig), we further tested whether ectopic expression of a non-degradable form of cyclin B1 (Δ85cyclin B1), which keeps Cdk1 activity high [32], would restore MPS1 activity (S6C Fig). Extracts from MastlFLOX MEFs expressing Δ85cyclin B1 displayed four-fold increased MPS1 activity due to accumulation of cells in mitosis but there was no additional increase in MastlNULL cells (S6C Fig), indicating that increasing Cdk1 activity does not elevate MPS1 activity in MastlNULL. Together, our data reveal the requirement of Mastl for full MPS1 kinase activity and persistent SAC signalling in mitosis. Furthermore, our data also suggest that the decreased activity of MPS1 is not due to a decrease in Cdk1 kinase activity, but an increase in the PP2A phosphatase activity toward MPS1 in MastlNULL MEFs.
Despite the rescue of MPS1 activity and phosphorylation in MastlNULL cells by OKA treatment, this could be an indirect consequence of PP2A inhibition. To test whether S820 MPS1 is dephosphorylated directly by PP2A as well as to identify the PP2A regulatory B subunit responsible for this dephosphorylation, we performed in vitro phosphatase assay using immunopurified PP2A/B55α or PP2A/B56α complexes. Incubation of phosphorylated pS820 MPS1 with immunoprecipitated PP2A/B55α resulted in a marked dephosphorylation of S820 MPS1 (Fig 3F, lane 4), but not in the presence of OKA (lane 5). Although both PP2A/B55 or PP2A/B56 complexes displayed phosphatase activity towards phosphorylated Rb (S6D Fig), the dephosphorylation of S820 MPS1 was more specific for B55α because immunoprecipitated PP2A/B56α was not as potent as PP2A/B55α (Fig 3F, lane 6). Of note, this difference was not simply due to formation and purification of the PP2A holocomplex as determined by immunoblot analysis, indicating similar amounts of the PP2A A and C subunits co-purified with B55α and B56α (Fig 3F, bottom panels). Furthermore, incubation with immunorecipitated PP2A complexes from increasing amounts of whole cell lysates of HA-B55/B56 overexpressing 293T cells caused a concentration-dependent dephosphorylation of S820 MPS1 (Fig 3G).
Together, these results indicate that PP2A/B55 dephosphoylates S820 MPS1 in vitro and that the Greatwall kinase/Mastl->PP2A/B55 pathway may be directly responsible for preventing premature dephosphorylation of at least S820 MPS1 in mitosis.
To exclude any effects of deletion of Mastl on the expression of PP2A, Arpp19, and ENSA, we further analysed the mRNA expression of any of the key PP2A subunits and no significant changes were observed between MastlFLOX and MastlNULL MEFs for the different isoforms of the regulatory PP2A subunits B55 and B56 (S7 Fig). While the mRNA expression level of Arpp19 was low or undetectable in primary MEFs by qPCR (Ct value 32.9), no significant change in ENSA mRNA expression level was observed between MastlNULL and MastlFLOX MEFs at each time point (S7 Fig). Nonetheless, we detected an increase of ENSA at protein level in MastlNULL MEFs in comparison to the control MastlFLOX cells (S3A Fig). Due to the lack of appropriate reagents, we were unable to determine the decrease in ENSA pS67 phosphorylation in MastlNULL MEFs similar as also in a recent study [7] on Greatwall kinase/ Mastl knockout mice.
To link the activity of MPS1 to its kinetochore function, we analyzed the localization of the phosphorylated form of MPS1 using phospho-specific antibodies [24]. We were able to detect S820-phosphorylated MPS1 at the kinetochore of MastlFLOX MEFs treated with nocodazole, whereas it was significantly decreased or absent in MastlNULL MEFs (Fig 4A and 4B). Similar results were obtained using phospho-specific antibodies for MPS1 T675 and T685 (S8 Fig, Fig 4C and 4D), suggesting that Mastl is essential for regulating multi-site phosphorylation of MPS1 in vivo. Since Mastl is known to regulate PP2A activity through phosphorylation of ENSA or Arpp19 [2–4] and inhibition of PP2A activity by OKA rescued MPS1 activity from MastlNULL MEFs (see Fig 3E), we evaluated the ability of OKA to rescue the kinetochore localization of phosphorylated MPS1. Indeed, in MastlNULL MEFs treated with OKA, MPS1 phosphorylated on S820, T675, and T685 was restored correctly at the kinetochores similar to that seen in MastlFLOX MEFs (Fig 4A–4D, S8 Fig).
The kinase activity of MPS1 is required for correct targeting and activation of a number of SAC regulators at the kinetochores [33] as cells enter mitosis. Notably, the relatively short duration of mitosis in MastlNULL MEFs upon treatment of microtubule poisons suggests SAC signalling defects (Fig 2D). However, in unperturbed mitosis, the duration of mitosis in MastlNULL MEFs was prolonged with poorly aligned chromosomes at the metaphase plate (Fig 2C, S1 and S2 Movies), suggesting that SAC signalling may be properly activated and established during mitotic entry but cannot be maintained robustly through mitosis.
To test this hypothesis, the kinetochore localization of the essential SAC protein Mad1 in early prophase (with lightly condensed chromosomes that were scattered in the cytoplasm just after NEBD) was determined by using immunofluorescence analysis using the centromeric marker ACA, as shown previously [30,31]. Notably, the kinetochore localization of Mad1 between MastlFLOX and MastlNULL MEFs displayed no significant difference in early prophase (Fig 5A (top panels) and Fig 5B), suggesting the correct establishment of proper SAC signalling in the absence of Mastl. In contrast, the kinetochore localization of Mad1 at late prometaphase-like stage (with a highly condensed chromosome mass typically caused by nocodazole treatment) was markedly defective in MastlNULL but not in MastlFLOX MEFs (Fig 5A (middle panels) and Fig 5C). Notably, OKA treatment efficiently restored the kinetochore localization of Mad1 at late prometaphase-like stage in MastlNULL MEFs similar to that seen in MastlFLOX MEFs (Fig 5A (bottom panels) and Fig 5C). Together, these results suggest that SAC signalling in MastlNULL MEFs is properly activated and established at the entry of mitosis, but it is unable to be maintained persistently even with misaligned chromosomes during progression through mitosis. Moreover, the defects in the kinetochore localization of phospho-MPS1 (T675, T685, S820) at late prometaphase-like stage in MastlNULL MEFs (see Fig 4 and S8 Fig) could not be detected in early prophase just after NEBD (S9A–S9C Fig). Thus, our results indicate that although the Cdk1->Greatwall kinase/Mastl->PP2A/B55 pathway may be responsible for MPS1 phosphorylation and full kinase activity at the kinetochores to maintain robust SAC signalling until completion of mitosis but not for the establishment and activation of SAC when cells enter mitosis. Taken together, we conclude that Mastl is important for robust SAC maintenance until the satisfaction of the checkpoint by bi-stably attaching microtubules to all kinetochores.
Here we show that the absence of the Greatwall kinase/Mastl abbrogates cell proliferation impairing embryonic development and tissue renewal in mice. Although it had been observed that Mastl deficient cells from Xenopus or human cell lines do not enter mitosis efficiently [5,6,9], our results revealed that genetic deletion of Mastl only delays mitosis entry as has been shown recently [7]. Mastl deficient unperturbed MEFs display a prolonged mitosis. However, once MastlNULL MEFs entered mitosis, a majority of cells prematurely progressed through mitosis with missegregated chromosomes, resulting in anaphase bridges and the generation of binucleated cells without completing cytokinesis. Furthermore, when treated with microtubule poisons to disrupt chromosome alignment and to activate the SAC, MastlNULL MEFs underwent mitotic slippage significantly faster than MastlFLOX MEFs. Concurrently, we observed decreases in phosphorylation levels and activity of MPS1, which coincided with the defective kinetochore localization of Mad1 and phosphorylated MPS1 in MastlNULL MEFs. Together, these results suggest defects in SAC signaling in MastlNULL MEFs. Notably, these defects occurred without decreasing the overall activity of Cdk1 in MastlNULL MEFs and they were efficiently rescued by inhbition of PP2A with OKA. Furthermore, we demonstrated that Cdk1 phospohrylation of MPS1 S820 can be reversed directly by PP2A/B55 in vitro. Collectively, our results indicate that the Cdk1->Greatwall kinase/Mastl->PP2A pathway plays a central role in the regulation of SAC, part of which is likely contributed by Mastl preventing PP2A/B55-mediated dephosphorylation of MPS1 in mitosis, although it is well possible that there are other substrates that regulate SAC since our “mini” screen was not saturating.
The role of Cdk1 in promoting SAC signalling has been well documented [13,14,34], but we are the first to report the essential role of Mastl in robust SAC maintenance. Notably, we found that neither the activity of Cdk1 nor its inhibitory phosphorylation on Y15 was affected in absence of Mastl. To confirm this, we used non-degradable cyclin B1 to ensure Cdk1 fully active in mitosis but there was no increase in MPS1 activity in MastlNULL cells. These observations make it unlikely that Mastl affects Cdk1 activity though we cannot entirely exclude this possibility. The conundrum to be resolved is that Cdk1 activates APC/CCdc20 at the same time as the SAC, which inhibits APC/CCdc20 [13]. This triangle can only be resolved when Cdk1 activity drops, which simultaneously requires active APC/CCdc20 (by degrading cyclin B) and silencing of the SAC (Fig 6). Our work suggests an additional regulatory loop, which includes Mastl, PP2A/B55, and MPS1, could contribute to the regulation of SAC silencing during mitotic progression. We propose that this regulatory loop allows SAC signalling to be robustly maintained by restraining PP2A, which removes activating phosphorylation of SAC components (e.g. MPS1 and maybe others), before all chromosomes are bi-stably attached. Although the regulation of Mastl activity has not been completely elucidated, inhibiting Mastl activity (such as by additional phosphatases) may also enable the SAC to be silenced even at a time when Cdk1 activity is still high and thereby activating APC/CCdc20 (e.g. mitotic slippage). This is an attractive hypothesis to resolve the Cdk1->APC/CCdc20->SAC conundrum.
The regulation of MPS1 including its kinetochore recruitment is complex and not fully understood [16,22,30,31]. MPS1 is phosphorylated on several sites during mitosis and in a Cdk1-dependent fashion, which may regulate its activity and/or its kinetochore localization [16,22,23,35,36]. Among these, phosphorylation sites T675 and T685 (T676 and T686 in human) regulate MPS1 activity by autophosphorylation [22–24,35]. Consistent with our observations that Cdk1/cyclin B and PP2A/B55 control the phosphorylation status of MPS1 on S820 during mitosis, previous phospho-proteome datasets of mitotic cells display a peak of phosphorylation of MPS1 on S820 during the mitosis [26–28]. However, it has been observed that phosphorylation on S820 is not essential for MPS1 activity in vitro [16,24] whereas our mass spectrometry analysis identified S820 phosphorylation as decreased in MastlNULL MEFs and therefore likely to be involved in MPS1 activity in vivo. Interestingly, a recent study by McCloy and colleagues reports that MPS1 phosphorylation on S820 as well as on S281 and S321 dramatically decrease during early mitosis in the presence of Cdk1 inhibitors [36]. Similar to the phenotype observed in MastlNULL MEFs, the decreased phosphorylation level of MPS1 leads to a drop of MPS1 activity and to the silencing of the SAC due by the dephosphorylation of MPS1 substrates [36].
Although phosphorylation of other MPS1 residues (e.g. T675, Fig 4A) may be also altered in MastlNULL MEFs since our mass spectrometry analysis was not saturating, our data indicate the potential role of S820 phosphorylation for MPS1 function. Indeed, reconstitution of a phospho-mimetic mutant for this residue in MPS1-depleted frog extracts rescued the kinetochore localization of MPS1 and SAC signalling, while non-phosphorylatable mutant S820A failed to do so [25]. We have tried to perform similar experiments in our MEFs by expressing MPS1 S820A, but these cells never entered mitosis probably because the SAC cannot be established in the absence of MPS1 activity. Nonetheless, Cdk1/cyclin B phosphorylation of MPS1 S820 may be a prerequisite for the kinetochore localization of MPS1 where it could acquire phosphorylation of other residues, resulting in enhanced activity. Of note, MastlNULL MEFs treated with microtubule poisons were transiently arrested with functional SAC signalling in early mitosis (NEBD). However, the SAC is prematurely silenced in MastlNULL MEFs. There are two possibilities; (1) Mastl does not regulate SAC and MPS1 during early mitosis, or (2) although MPS1 is hypophosphorylated and less active in MastlNULL MEFs, its remaining activity may be sufficient to initiate SAC signalling in early mitosis. The latter is in agreement with previous observations that 10% of MPS1 activity is sufficient to activate the SAC [33,37] by recruiting the Mad1-Mad2 complex to the kinetochores [29,37,38], which subsequently activates Mad2 to form the mitotic checkpoint complex (MCC) complex [39]. Our data indicate that maintenance of MPS1 activity by Mastl may be required to persistently localize Mad1 to the kinetochores, thereby sustaining SAC signalling.
We were able to detect three discernible phenotypes in MastlNULL MEFs; (i) delayed entry into mitosis possibly due to chromosome condensation defects, (ii) premature silencing of the SAC in the presence of chromosome segregation defects [this is the major focus of this manuscript], and (iii) cytokinesis defects mediated by the dephosphorylation of PRC1 [8]. While the phosphorylation status of MPS1 and the SAC are affected by Mastl deletion, this does not explain the entire phenotype observed in MastlNULL MEFs. MPS1 is likely not the only protein hypophosphorylated in the absence of Mastl. It is therefore evident that other targets are implicated in MastlNULL phenotype and that restoring MPS1 phosphorylation unlikely rescues the global defects due to Mastl loss. Nonetheless, our work has connected two important pathways, Greatwall kinase/Mastl->PP2A/B55 and Cdk1 in the regulation of the SAC, highlighting the importance of fine-tuning the signals through intricate feedback loops. Future studies will likely uncover additional mechanisms how SAC signalling is regulated in detail and new substrates of the Greatwall kinase/Mastl->PP2A/B55 pathway.
All animal studies were approved by and have be conducted in a humane manner following the rules of the Biological Resource Centre (BRC) Institutional Animal Care and Use Committee (IACUC) of A*STAR at Biopolis, Singapore (IACUC protocol #140927).
The MastlFLOX mouse strain was generated as described previously [17].
Mastl conditional knockout mice were crossed with β-actin-Flpe transgenic mice [40] (strain name: B6.Cg-Tg(ACTFLPe) 9205Dym/J; stock no.: 005703; The Jackson Laboratory) to remove the neomycin cassette, which resulted in the MastlFLOX allele. The MastlNULL allele was then generated by crossing MastlFLOX mice with β-Actin-Cre transgenic mice [41] (strain name: FVB/N-Tg(ACTB-cre)2Mrt/J; stock no.: 003376; The Jackson Laboratory), which deletes exon 4 and causes a frame shift. Liver specific Mastl knockout was accomplished by crossing MastlFLOX mice with Albumin-Cre transgenic mice [42]. Tamoxifen inducible conditional knockouts were created by crossing with either Esr1 (CreERT2) [43] (strain name: B6.Cg-Tg(CAG-cre/Esr1*)5Amc/J; stock no.: 004682; The Jackson Laboratory) or Rosa26-CreERT2 mice [44]. To induce Mastl gene deletion in adult mice (8–10 weeks old), Mastl+/FLOX or MastlFLOX/FLOX mice with two copies of the CreERT2 transgene (Rosa26-CreERT2TG/TG) were intraperitoneally injected with 1mg tamoxifen (Sigma-Aldrich, #T5648) dissolved in 50μl corn oil (Sigma-Aldrich, #C8267) for three consecutive days. For viability analysis, 4 adult mice of each genotype were used. After induction of recombination, MastlNULL/NULL mice died within 7–8 days while the Mastl+/NULL mice remained viable with no visible phenotypical abnormalities for 4 months before they were euthanized. Mice were housed under standard conditions with food and water available ad libitum and maintained on a 12 hour light/dark cycle. Mice were fed a standard chow diet containing 6% crude fat.
Primary mouse embryonic fibroblasts (MEF) of the MastlFLOX/FLOX Esr1 (CreERT2) genotype were isolated from E13.5 mouse embryos as described previously [18]. Briefly, the head and the visceral organs were removed, the embryonic tissue was chopped into fine pieces by a razor blade, trypsinized 15 minutes at 37°C, and finally tissue and cell clumps were dissociated by pipetting. Cells were plated in a 10cm culture dish (passage 0) and grown in DMEM (Invitrogen, #12701–017), supplemented with 10% fetal calf serum (Invitrogen, #26140) and 1% penicillin/streptomycin (Invitrogen, #15140–122). Primary MEFs were cultured in a humidified incubator with 5% CO2 and 3% O2.
Primary and immortalized MEFs were synchronized at the G0/G1 phase of the cell cycle by culturing at high confluence (contact inhibition) and starvation in reduced serum containing growth media (0.2% fetal calf serum) for 72 hours. Recombination at the Mastl locus was induced only during the last 24 hours when majority of the cells had already arrested at the G0/G1 phase. In paired experiments comprising control and Mastl deficient conditions, identical MEF clones treated with DMSO (Control or MastlFLOX) or 20 ng/ml 4-OHT (MastlNULL) were used. To induce synchronized entry into cell cycle, cells were trypsinized and replated. Mitotic arrest was achieved by the addition of microtubule poisons and were performed between 20–24 or 24–28 hours after release for primary cells or immortalized MEFs, respectively, and the cells were subsequently processed and analyzed as decribed below.
For Alamar Blue proliferation assays; 1500 cells were plated in 96-well plates in 5 replicates, with or without prior 4-OHT (Sigma, #H7904) treatment to induce Mastl knockout. Starting from 24 hours after seeding, cells were incubated in 150 μl of assay medium (1:9 ratio of Alamar Blue (AbD Serotec, #BUF012B) to growth medium) for 4 hours and metabolic activity was quantified by measuring the fluorescence at 590nm.
For BrdU labeling and FACS analysis: MEFs were grown to confluence in 15 cm dishes and serum starved for 72 hours in 0.2% serum containing growth medium. To induce Mastl knockout, 4-OHT was added during entire starvation period (S2C Fig). To induce synchronized entry into cell cycle, cells were trypsinized and replated in 10cm dishes in full growth medium. To monitor S phase, cells were labeled with 100 μM BrdU (BD Pharmingen, #550891) for 1 hour before collection of the cells at different time points. At the end of each time point, cells were trypsinized and fixed in -20°C cold 70% ethanol, stained with APC conjugated anti-BrdU antibodies (BD Pharmingen, #623551) and propidium iodide (Sigma, #81845). Cell cycle analysis was performed using FACSCalibur flow cytometer (BD Biosciences) and resulting data were analyzed by FlowJo 8 software.
Primary MastlFLOX/FLOX MEFs were immortalized by serial passaging for 30 times using a modified 3T3 protocol [45]. To create cell lines where Mastl gene knockout can be induced by the addition of 4-OHT, cells were infected with pBABE-CreERT2 (PKB971) or pWZL-CreERT2 (PKB931) retroviral constructs and selected with 2.5 μg/ml puromycin or 10 μg/ml blasticidin, respectively. Immortalized cells displayed essentially the same phenotype as the primary MEFs after induction of the Mastl gene knockout. However after synchronization and release into full growth medium, they were delayed for about 4 hours to enter mitosis. For this reason, mitotic arrest studies in immortalized MEFs were performed between 24–28 hours after release, instead of 20–24 hours used for primary cells. Cell lines expressing H2B-YFP (PKB1515), mAg-hGeminin1-110 (PKB1572) and HA-hMPS1 (PKB1733) were created by lentiviral transduction of these immortalized MEFs.
MEFs were transfected in 10cm Petri dishes using X-tremeGene (Roche) with 5μg of plasmid encoding Myc-MPS1 (PKB1714) or Δ85cyclinB1 (PKB1851). 293FT cells were transfected by the calcium phosphate method during 8 hours with plasmids encoding HA-B55 (PKB1848) or HA-B56 (PKB1849) and cell extracts were collected after 48 hours.
Commercially available antibodies used for immunoblot or immunofluorescence microscopy staining are: rabbit anti-Cdk1 (Santa Cruz, #SC-954), mouse anti-Cdk2 (Santa Cruz, #SC-6248), rabbit anti-phospho Cdk1 Y15 (Cell Signalling #9111), rabbit anti-cyclin A2 (Santa Cruz, #SC-596), mouse anti-cyclin B1 (Cell Signaling, #4135), rabbit anti-phospho-Histone H3 Ser10 (Cell Signaling, #9701 or Upstate, #06–570), mouse anti MPS1-N terminal (EMD Millipore, #05–682), mouse anti-MPS1-CT, clone 4-112-3 (Millipore, #05–683), mouse anti-MPS1 (Sigma, #WH0007272M1), rabbit anti-MPS1 (Sigma, #HPA016834), rabbit anti-MPS1 (Santa Cruz, #SC-540), rabbit anti-CDK phospho substrates (Cell Signaling, #9477), mouse anti-HA tag (Cell Signaling, #2367), mouse anti-Myc (Clontech, #631208), mouse anti-Hsp90 (BD Transduction Labs, #610419), goat anti-actin (Santa Cruz, #SC-1616), human anti-Centromere antibodies (Antibodies Incorporated, #15–234), rabbit anti-Aurora B (Santa Cruz, #SC-25426), mouse anti-ENSA (Santa Cruz, #SC-81883), rabbi anti-PP2A A (Cell Signaling, #2039), rabbit anti-PP2A C (Sant Cruz, #SC-14020), and rabbit anti-Mad1 (Santa Cruz, #SC-222). Rabbit polyclonal antisera against Mastl were raised using an N-terminal 6-His tagged fragment from mouse Mastl (residues 461 to 694, PKB908) as antigen using a published protocol [46]. Phospho-specific MPS1 antibodies [24] have been described before and were kindly provided by Pat Eyers (University of Liverpool).
For FACS analysis of mitotic cells, Alexa 647 conjugated antibodies against phospho-Histone H3 Ser10 (Cell Signaling, #9716) were used. For histological analysis of mitotic cells, sections were stained with rabbit anti-phospho histone H3 (Ser10) antibodies (Upstate, #06–570).
The plasmid encoding Myc tagged human MPS1 (PKB1714) was described before [22]. The plasmid encoding mAg tagged human Geminin1-110 (PKB1572) was described before [47] and was kindly provided by Atsushi Miyawaki (RIKEN). The plasmid encoding non degradable Δ85cyclinB1 human (PKB1851) was described before [32]. The plasmids encoding HA-tagged B55α and B56α (PKB1848 and PKB1849) were kindly provided by David Virshup (Duke-NUS).
Cells and tissues were lysed in EBN buffer (80 mM β-glycerophosphate pH 7.3, 20 mM EGTA, 15 mM MgCl2, 150 mM NaCl, 0.5% NP-40, 1 mM DTT, and protease inhibitors (20 μg/ml each of leupeptin, chymostatin, and pepstatin (Chemicon, EI8, EI6 and EI10)) for 20 min with constant shaking at 1200 rpm. Lysates were centrifuged for 30 min at 18,000g at 4°C and supernatants were snap frozen in liquid nitrogen and stored at -80°C. 10 μg of protein extracts were separated on 10% or 12.5% polyacrylamide gels, transferred onto polyvinylidene difluoride membranes (PVDF, Millipore, #IPVH0010) using a semi-dry system and blocked in tris-buffered saline (TBS) with 0.1% Tween20 and 4% non fat dry milk (Bio-Rad, #1706404). Blots were probed with the appropriate primary antibodies overnight at 4°C, followed by secondary goat anti-mouse (Pierce, #0031432) or anti-rabbit antibodies (Pierce, #0031462) conjugated to horseradish peroxidase and developed using enhanced chemiluminescence (PerkinElmer, #NEL105001EA).
Affinity purification/immunoprecipitations and Cdk/cyclin kinase assays were performed as described previously [46] with minor modifications. Briefly, 100–250 μg of protein extract were incubated with beads conjugated to Suc1 (Upstate, #14–132) or antibodies against Cdk1 (Santa Cruz, #SC-954AC), Cdk2 [46], cyclin A2 (Santa Cruz, #SC-751AC), cyclin B1 [46] overnight at 4°C in EBN buffer supplemented with 1 mg/ml ovalbumin (Sigma, #A5503). Antibodies were pre-coupled to protein A (Roche, #11719408001) or protein G (Roche, #11719416001) agarose beads. Following two washes in buffer EBN and one wash in buffer EB (EBN without NP-40), the precipitated proteins were used in kinase assays to determine the levels of kinase activity against the substrate histone H1 (Roche, #11004875001). Kinase assays were performed by incubating the immunoprecipitated proteins on beads in EB buffer with 10 mM DTT, 15 μM ATP, 5 μCi [ϒ-32P]ATP (PerkinElmer, #NEG502A) and 1.5 μg histone H1 for 30 min at room temperature. After inactivation with SDS-PAGE sample buffer, electrophoresis on polyacrylamide gel, fixation and staining in Bismarck Brown/Coomassie blue, quantification of incorporated radioactivity was performed with a phosphoimager (Fujifilm, FLA-7000).
For kinase assays to measure MPS1 activity, immortalized MEFs were used. Cells were synchronized and Mastl deletion was induced with 4-OHT. At the end of the serum starvation period, 2 million cells were seeded in 10 cm dishes in full growth medium while simultaneously being transfected with the Myc-MPS1 plasmid (PKB1714). 24 hours after seeding into full medium, cells were arrested in mitosis by the addition of 500 ng/ml nocodazole for 4 hours. Cells were collected and protein extracts were prepared as described above. The MPS1 kinase was immunoprecipitated from 1mg protein extracts by agarose beads conjugated to anti-Myc antibodies (Clontech, #631208). Kinase assays were performed in a 20μl reaction buffer (50 mM Tris-Cl pH7.5, 10 mM MgCl2, 100μM cold ATP, 5μCi [ϒ-32P]ATP, 1mM DTT, 4μM β-glycerophosphate, 1 mM EGTA) using 5μg of myelin basic protein (Sigma-Aldrich, #M1891) as substrate, for 30 minutes at room temperature. Kinase activity was detected as described above.
For kinase assays to test Cdk1/cyclin B1 kinase activity on MPS1 S820 residue, 5μg GST or GST-MPS1 peptide (residues 811–831; PKB1759 [S820S] and PKB1760 [S820A]) fusion protein bound to glutathione beads was used as substrates in a reaction volume of 20μl in a kinase reaction buffer comprised of 80 mM β-glycerophosphate pH 7.3, 20 mM EGTA, 15 mM MgCl2, 10 mM DTT, 100 μM cold ATP, 1 mM NaF and 5 μCi [ϒ-32P]ATP. 50 ng of purified Cdk1/cyclin B1 complexes (Cell Signaling, #7518) was added per reaction and incubated for 30 minutes at room temperature. Phosphorylation of GST-MPS1 fusion proteins was measured as above.
For phosphatase assay to test PP2A/B55 or PP2A/B56 activity on phosphorylated S820 residue, the previously 32P phosphorylated GST-MPS1 peptide fusion protein bound to gluthatione beads were washed twice in EB buffer (EBN without NP40) and twice in HEPES buffer (100 mM HEPES pH 8, 10 mM DTT, 10 mM MgCl2). Washed GST-MPS1 fusion protein was incubated in HEPES buffer complemented of protease inhibitors (20 μg/ml each of leupeptin, chymostatin, and pepstatin (Chemicon, EI8, EI6 and EI10)) for 4 hours at 30°C in the presence of immunoprecipitated HA-tagged B55 or B56 complexes from transfected 293FT cell extracts. The levels of phosphorylated GST-MPS1 was determined as above.
MEF grown on coverglass-bottom chamber slides (Lab Tek) were fixed with 4% PFA or ice-cold MeOH. The fixed cells were permeabilized with 0.5% Triton X-100 and exposed to TBS containing 0.1% Triton X-100 and 2% BSA (AbDil). Images were acquired at RT with 3D-SIM using a Super Resolution Microscope (Nikon) equipped with an iXon EM+ 885 EMCCD camera (Andor) mounted on a Nikon Eclipse Ti-E inverted microscope with a CFI Apo TIRF (100x/1.40 oil) objective and processed with the NIS-Elements AR software. For time-lapse video microscopy, immortalized MEFs expressing the fluorescent fusion proteins (as described above) were synchronized by serum starvation. Cells were released in full growth medium and plated in coverglass bottom chamber slides (Nunc, #155380 and #155383). The images were captured every 10 min upon release using live-cell fluorescent microscopy with an iXon EM+ 885 EMCCD camera (Andor), a CFI Plan Fluor (20x/N.A. 0.45) objective and a Stage Top Incubation with Digital CO2 mixer (Tokai). The images were processed using NIS-Elements AR software.
For PCR genotyping of Mastl wild type, FLOX, and NULL alleles, primers Pr1 (PKO860), Pr2 (PKO862), and Pr3 [PKO863] (S2 Table) were used at 1μM final concentration. Briefly, cells or tissue pieces to be genotyped were lysed by boiling in lysis solution (25mM NaOH pH 12, 0.2mM EDTA) for 20–30 minutes to extract genomic DNA [48]. Alkaline pH was neutralized by the addition of an equal volume of neutralization buffer (40mM Tris-HCl pH 5). 1 μl of the resultant genomic DNA solution was used as a template in a 20 μl volume of PCR reaction, using 0.5 units of MangoTaq polymerase (Bioline). 35 PCR cycles with 30 seconds denaturation at 94°C, 30 seconds annealing at 66°C, and 30 seconds extension at 72°C were performed to amplify different alleles of Mastl gene resulting in a band of 200bp (MastlWT), 304bp (MastlFLOX), or 500bp (MastlNULL). Total RNA was extracted using MN NucleoSpin RNA II kit according the manufacturer’s protocol. For each RT-PCR reaction, first strand cDNA was synthesied from 1μg total RNA using the Maxima First Strand cDNA synthesis kit (Thermo Fisher, K1642). Mastl mRNA levels at different time points after 4-OHT induction of asynchronous primary MEFs were determined by RT-PCR using primers PKO1535 and PKO1536 (S2 Table). PCR amplification was carried out using the Maxima SYBR Green qPCR Master Mix (Fermentas, K0252) and the appropriate primer pair (see S2 Table). The reactions were monitored continuously in a Rotor-Gene thermal cycler (Corbett Research) using the following program: 95°C for 10 min, followed by 40 cycles of 95°C for 15 sec, 55°C for 30 sec, and 72°C for 30 sec. All data were normalized to the expression levels of eEF2 housekeeping gene using the (2-ΔΔCt) method.
Primary MastlFLOX/FLOX Esr1 (CreERT2) MEFs were cultured in either media containing light isotopes of L-lysine-(12C614N2) [K0] and L-arginine-(12C614N4) [R0] or media containing stable heavy isotope L-lysine-(13C615N2) [K8] and L-arginine-(13C615N4) [R10] for three passages for a complete exchange of isotopes. Cells were synchronized by serum starvation and released to enter the cell cycle as usual. Cells were treated with 5μM Eg5 kinesin II inhibitor for 4 hours between 20–24 hours after serum release. Mitotic cells were isolated by pipetting and lysed in 8 M urea lysis buffer. Equal amounts of cell lysates from light (K0R0) and heavy (K8R10) cells were mixed. A total of two biological replicates were carried out, namely one Forward (light MastlFLOX MEFs versus heavy MastlNULL MEFs) and one Reverse (heavy MastlFLOX MEFs versus light MastlNULLMEFs) experiments.
To test the statistical significance of the distribution of samples displayed in Fig 2D a Student’s t-test applett available at http://www.math.kent.edu/~blewis/stat/tTest.html website was used. For Figs 4B–4D, 5B and 5C and S9B and S9C Fig, Student’s t-test was directly determined in the program Prism.
|
10.1371/journal.pntd.0006544 | Quantification of permethrin resistance and kdr alleles in Florida strains of Aedes aegypti (L.) and Aedes albopictus (Skuse) | Recent outbreaks of locally transmitted dengue and Zika viruses in Florida have placed more emphasis on integrated vector management plans for Aedes aegypti (L.) and Aedes albopictus Skuse. Adulticiding, primarily with pyrethroids, is often employed for the immediate control of potentially arbovirus-infected mosquitoes during outbreak situations. While pyrethroid resistance is common in Ae. aegypti worldwide and testing is recommended by CDC and WHO, resistance to this class of products has not been widely examined or quantified in Florida. To address this information gap, we performed the first study to quantify both pyrethroid resistance and genetic markers of pyrethroid resistance in Ae. aegypti and Ae. albopictus strains in Florida. Using direct topical application to measure intrinsic toxicity, we examined 21 Ae. aegypti strains from 9 counties and found permethrin resistance (resistance ratio (RR) = 6-61-fold) in all strains when compared to the susceptible ORL1952 control strain. Permethrin resistance in five strains of Ae. albopictus was very low (RR<1.6) even when collected from the same containers producing resistant Ae. aegypti. Characterization of two sodium channel kdr alleles associated with pyrethroid-resistance showed widespread distribution in 62 strains of Ae. aegypti. The 1534 phenylalanine to cysteine (F1534C) single nucleotide polymorphism SNP was fixed or nearly fixed in all strains regardless of RR. We observed much more variation in the 1016 valine to isoleucine (V1016I) allele and observed that an increasing frequency of the homozygous V1016I allele correlates strongly with increased RR (Pearson corr = 0.905). In agreement with previous studies, we observed a very low frequency of three kdr genotypes, IIFF, VIFF, and IIFC. In this study, we provide a statewide examination of pyrethroid resistance, and demonstrate that permethrin resistance and the genetic markers for resistance are widely present in FL Ae. aegypti. Resistance testing should be included in an effective management program.
| Aedes aegypti (Yellow-fever mosquito) and Aedes albopictus (Asian Tiger mosquito) can vector a variety of arboviruses that cause diseases and are thus a public health concern. Pyrethroid insecticide resistance is common in Ae. aegypti in many locations worldwide and can adversely affect vector control operations. However, the resistance status of these vectors in Florida is largely unreported and recent local transmission of dengue and Zika viruses has made this information critical for effective control operations. In this study, we showed that permethrin resistance and two common SNPs of the voltage gated sodium channel (V1016I and F1534C) previously associated with pyrethroid resistance were widely present in Florida Ae. aegypti strains. We also observed a strong correlation between the dilocus knock down response (kdr) genotype and resistance ratio (RR) as determined by topical application, which suggests, as have others, that kdr frequency may be a useful indicator of resistance in Aedes aegypti.
| Local vector control programs play a part in public health. In many countries, these programs serve as the primary defense against the spread of several mosquito-borne diseases. Effective Integrated Vector Management (IVM) programs rely on surveillance information coupled with multiple vector control strategies such as chemical adulticiding to reduce vector populations and arbovirus transmission. Limited recent transmission of locally acquired dengue and Zika viruses in the southeastern continental US, primarily Florida, has brought renewed attention to the importance of IVM programs where the potential vectors, Aedes aegypti (L.) and Ae. albopictus Skuse, have long been present. However, for IVM programs in Florida to effectively control Aedes vectors and to reduce dengue and Zika virus transmission during an outbreak, it is essential to know which adulticide products are effective against local Ae. aegypti and Ae. albopictus strains.
Organophosphates and pyrethroids are the only two classes of insecticides available to public health agencies for control of adult mosquitoes in the US. Compared to organophosphate insecticides, pyrethroids have higher public acceptance, rapid knockdown, relatively low costs, and are generally the product class of choice when adulticiding is required [1]. Unfortunately, years of pyrethroid insecticide use and previous DDT usage has increased the frequency of genetic resistance and enhanced enzymatic detoxification in insects like mosquitoes. Genetic target site changes of the sodium channel, known as knockdown resistance (kdr) mutations, are a relatively common insect response to selective pressure by pyrethroids [2,3]. Although a variety of other single nucleotide polymorphisms (SNPs) have been noted in Ae. aegypti, the primary two SNPs assessed to determine pyrethroid resistance are at codons 1016 and 1534 (positions according to standard M. domestica notation) [4,5]. One allele of the 1016 mutation is geographically distinct to the Western hemisphere and results in the replacement of the normal valine with an isoleucine (V1016I), while the 1534 mutation results in a phenylalanine to cysteine change (F1534C). There is currently debate about the individual toxicological effect of these two mutations, but they are consistently present in resistant strains [4]. Heterozygous and homozygous combinations of these two SNPs could result in nine possible genotypes, but strong linkage between the SNPs has been noted with only six of the genotypes observed in a study of Mexican Ae. aegypti [6].
More recent work indicates that the kdr genotype in Ae. aegypti may be more complicated than the combination of 1016I and 1534C SNPs. An additional SNP at position 410 has been identified and this valine to leucine (V410L) change, in addition to the other two SNPs, may be the strongly resistant phenotype [7]. The noted strong linkage between 410L and 1016I indicates that assessment of 1016 may still be adequate to assess the strongly resistant phenotype [7]. As expected based on the dilocus genotype distribution, several of the trilocus genotypes were also not frequently present in the wild [8].
Pyrethroid resistance and the distribution of kdr alleles have been well documented in Ae. aegypti strains from Central America, South America and the Caribbean [9–13]. Recent testing as part of the Zika emergency response in Puerto Rico has shown that isolated, early reports of resistance were indicative of widespread resistance on the island [13–16]. Studies have also shown the pyrethroid resistance is not specific to a particular pyrethroid but is generally class-wide to type I, type II, and non-ester pyrethroids [13, 17, 18]. Kdr alleles and pyrethroid resistance are widely distributed in Mexico, including Nuevo Laredo which lies just south of the US-Mexico border [11].
In contrast, little has been published about pyrethroid resistance in Aedes strains within the continental US. In the Garcia et al. [11] study mentioned above, the authors did not find kdr alleles in a Houston, TX collection from 1999. Recent reviews of the Aedes resistance literature listed no reports of resistance in continental US Ae. aegypti [4, 19], but Cornel et al. [20] recently demonstrated toxicological resistance and sodium channel mutations in invasive populations of Ae. aegypti in California. Two recent studies using CDC bottle bioassays do indicate resistance in strains from the southern US, but these studies do not provide any quantification of the strength of the resistance nor did they examine the presence of kdr alleles [21, 22]. In contrast, Ae. albopictus does not appear to frequently develop kdr resistance, and most US strains tested thus far have only shown very minimal pyrethroid resistance [4, 23, 24, 25].
Resistance surveillance is recommended by the CDC and statewide initiatives to map pyrethroid resistance have begun in Florida and California [22,26]. Resistance information is critically important for operational decision making as part of an effective IVM program. In this study, a collaborative group of government, academic, industry and vector control district stakeholders collected Ae. aegypti and Ae. albopictus adults, eggs or larvae from more than 200 locations throughout Florida to assess the extent and intensity of pyrethroid resistance. The goal of this study was to improve vector control operations by producing a resistance map and apply this information to make more effective control decisions. We determined permethrin resistance ratios (RRs) relative to the susceptible ORL1952 strain for 21 wild type strains of Ae. aegypti and 5 strains of Ae. albopictus using direct topical application, the WHO gold standard assay for determination of intrinsic toxicity [27]. While the CDC bottle bioassay is a more commonly used assay, it is a threshold assay and does not quantitate resistance levels which was one of our primary goals. Nearly 5,000 Ae. aegypti from numerous locations were genotyped by allele-specific PCR to assess the frequency of V1016I and F1534C alleles and rapidly visualize the pattern of resistance throughout the state of Florida.
The toxicological profiles and rearing procedures for the susceptible (ORL1952) and resistant (Puerto Rico–BEIResources, NR-48830) strains of Ae. aegypti used in this study have been described previously [17]. Strain information, including specific location information, collectors, and dates of collection are noted in Table 1. Field strains for toxicology testing were often collected as mixed (Ae. albopictus and Ae. aegypti) eggs laid on seed germination paper (Anchor Paper, St. Paul, MN) placed in a variety of oviposition containers including black plastic stadium cups, plastic cemetery vases, and glass containers. Field collected eggs were hatched by soaking papers for 48–72 hours at 27 oC in rearing trays of untreated well water or deionized water (diH2O). Papers were carefully removed, and larvae were reared through adult emergence with a reduced feeding regimen compared to the standard rearing protocol for the ORL1952 susceptible strain [28] due to sensitivity to overfeeding. Strains from Monroe, Seminole, Orange, Hernando and Sarasota counties were collected as larvae from a variety of manmade and natural containers (tires, plant pots, bottles, buckets, etc.), rinsed with diH2O and then placed into the standard larval rearing procedure described above. Pupae were collected from rearing trays and placed in emergence chambers or 12” x 12” screen cages (BioQuip Models 1425 & 1450B, Rancho Dominquez, CA). After emergence, wildtype mosquitoes were briefly chilled to 4 oC and sorted to species. Strains for toxicology testing were produced from locations that had more than 30 wildtype founders. If multiple nearby locations were combined to produce a strain, the GPS location in Table 1 represents the GPS centroid of the sites that contributed the mosquitoes. Individual oviposition cup locations and life stages tested are provided in S1 File. Eggs were produced using standard rearing methods [17]. Colony strains were provided 2–10 bloodmeals on weekly intervals to collect F1 eggs. Warmed bovine blood was provided to produce 2,000 or more eggs per strain. If feeding was poor, warmed blood was spiked with 1 mM ATP as a phagostimulant. Bovine blood for mosquito feeding was purchased by the USDA under contract from a local, licensed abattoir. Blood was collected during normal operations of the abattoir from the waste stream after animal slaughter. Under CFR9, Parts 1–3, tissues, including blood, collected from dead livestock intended as food are exempt from IACUC regulation.
Aedes aegypti strains from St. Augustine, Clearwater, and Vero Beach, FL were provided as F1 eggs produced at Anastasia Mosquito Control District, Pinellas County Mosquito Control, and the Florida Medical Entomology Laboratory, respectively.
To produce mosquitoes for bioassay testing, eggs from the F1 or F2 generation of field collected strains, the lab susceptible strains (ORL1952) and the pyrethroid-resistant strain were hatched and reared as described above in covered trays at a density of approximately 1,000 mosquito larvae/tray. Adult mosquitoes emerged into 12” x 12” screen cages and provided with cotton saturated with 10% sucrose in diH2O. Females used for bioassays were 3–7 days post-emergence.
The adult topical bioassay has been described previously in detail [28, 29]. For these studies, the permethrin stock solution in DMSO (Product #N-12848-250MG, Chemservice, Westchester, PA) and all dilutions in acetone were prepared gravimetrically. An initial 10-fold dilution series was prepared over a range of relevant concentrations [17]. Sub-dilutions were prepared from the 10-fold dilution series as necessary to determine the critical region of the dose curve. Three assays (n = 3) were performed for all strains with listed LD50s unless limited numbers of F1 test mosquitoes allowed only two replicates. Average mass/female was calculated for each strain by weighing a cohort of 50–100 females before each replicate.
LD50s, standard errors, and goodness of fit were determined from dose-mortality curves using SigmaPlot v13 (Systat software Inc., San Jose, CA) with data fit to a four-parameter logistic model. To provide a comparative metric between strains that may have different body sizes, doses applied were divided by the average mass of the mosquitoes before curve fitting. This results in an LD50 of ng of active ingredient per mg of mosquito. Resistance ratios for strains were calculated using the LD50 of the field strain divided by the LD50 of the susceptible ORL1952 strain included in the same assay. In this study, we use the WHO scale to define the levels of resistance [30]. When RR is less than 5, the field population is considered susceptible. When the RR is between 5 and10, the mosquitoes are considered to have moderate resistance. A RR greater than 10 indicates that mosquitoes are highly resistant [31].
Genotypes for individual mosquitoes or eggs were determined using melt curve analysis with previously described allele-specific primers for the 1016 and 1534 SNPs [32, 33]. Preliminary assays were conducted using eggs and adults from the same generation to verify that the resulting frequencies were similar (within 10%) between life stages rather than showing agenotype bias due to varied hatch rates or differing larval mortality (S2 File). Assays were performed in 96 well plates on a StepOnePlus (Applied Biosystems) or QuantStudio5 (Applied Biosystems, Thermo Fisher Scientific, Waltham, MA) using SYBR green chemistry. Plates were loaded with 8 μl of PCR master mix containing SYBR Select Master Mix (Applied Biosystems, Thermo Fisher Scientific, Waltham, MA), nuclease free water (NFW), and three primers (Table 2). Due to differences in efficiency from the added GC tails, initial primer titering was necessary to ensure that the melt curves of homozygous controls would be accurately called by the analysis software. Individual adult females were homogenized for 60 seconds at max speed in 100 μl of NFW on a bead beater (BioSpec, Bartlesville, OK). Individual eggs, randomly sampled from all available egg papers of the same generation, were similarly treated but homogenized in 50 μl of NFW. Immediately after homogenization, samples were centrifuged at 10,000 relative centrifugal force (rcf) for 60 seconds to pellet solids. Two microliters of each supernatant were added to 8 μl of PCR mix for each primer set and then subjected to standard cycling conditions (3 min @ 95 °C; 40 cycles @ 95 °C for 3 sec, 60 °C for 15 sec). Melt curve analysis followed cycling with acquisition of fluorescence data every 0.3 °C as the temperature was ramped from 60 °C to 95 °C. Characteristic melting temperature (Tm) peaks in the derivative fluorescence data indicate the presence of specific alleles [32, 33]. For the 1016 mutation, a codon for the susceptible valine has a Tm peak at 86±0.3 °C while the isoleucine codon has a Tm of 77.3±0.3 °C. The 1534 phenylalanine has a Tm of 79.8±0.3 °C while the mutant cysteine produces a peak at 84.7±0.3 °C. Homozygotes for either allele produce one peak while heterozygotes produce peaks at both Tms. All assay plates included two susceptible ORL1952 strain and two resistant PR strain samples as negative and positive controls, respectively. Most plates also contained artificial heterozygotes created by including ORL1952 and PR homogenates in the same sample. Well positions of individual mosquitoes were maintained in both plates (one plate for each locus) to allow genotyping of an individual for both SNPs. Frequencies for each of the nine genotypes (VVFF, VVFC, VVCC, VIFF, VIFC, VICC, IIFF, IIFC, IICC) were calculated by dividing the specific genotype by the total number tested from each area.
Base maps were created using ArcGIS software by Esri. ArcGIS® is the intellectual property of ESRI and is used herein under license. Copyright © ESRI. GIS data sources were ESRI and Tele Atlas. All rights reserved. (For more information about Esri® software, please visit www.esri.com). Permission to publish this content was verified from the ESRI Redistribution Rights Matrix at https://www.esri.com/~/media/Files/Pdfs/legal/pdfs/redist_rights_106.pdf?la=en
Maps were exported to GIMP 2.8 and additional layers with the kdr or resistance ratio data was added. Pie chart representations of kdr allele frequencies were created with Microsoft Excel and exported to layers added to the basemaps.
All Florida strains of Ae. aegypti were resistant to permethrin when compared to the ORL1952 strain, which has been in a continuous laboratory colony for nearly seventy years (Fig 1 and Table 3). The field strains showed varied levels of resistance, from 6-fold to 61-fold compared to the ORL1952 strain. The two least resistant Ae. aegypti strains were collected from flowerpots on Big Coppitt Key (Monroe County) and from oviposition cups in Cortez (Manatee County) at 6.0 and 6.8-fold, respectively. Ae. aegypti mosquitoes collected from a tire facility in Orlando (Orange County) were about 17.2-fold more resistant than Ae. aegypti collected from the same area generations earlier and used to produce the ORL1952 strain. While most of the FL strains were 15 to 35-fold resistant compared to the lab strains, several strains with higher RR were identified. The strain from New Port Richey in Pasco County had the highest RR at 61.3-fold, which is like the RR of the Puerto Rico-resistant reference strain [17]. Strains from Miami Beach and Fort Myers had RRs above 50.
The variability we observed in RR throughout the state was also seen at finer resolution within several, but not all, counties. In Miami-Dade County, the Miami Beach and East Wynwood strains were relatively resistant (57-fold and 33-fold), while nearby locations like South Wynwood and central Little River were much less resistant (Fig 1). Manatee County also had a range of RRs. The Anna Maria Island strain, collected from a densely populated barrier island, had a higher RR than nearby strains from Palmetto or Cortez (Fig 1 and Table 3). We also observed this same trend in Lee County but did not observe universal variation in RR.
Aedes albopictus competes with Ae. aegypti for oviposition sites, and in several locations Ae. albopictus eggs were collected in conjunction with Ae. aegypti eggs. Ae. albopictus strains were also subjected to topical application along with the Ae. aegypti from the same locations (Fig 1, in blue). In the Ae. albopictus strains we tested, only slight resistance to permethrin (<2-fold) was observed when compared to the ORL1952 strain. In Miami-Dade County, we observed large differences in RR between Ae. albopictus and Ae. aegypti even when collected from the same sites in areas such as south Wynwood and central Little River. St. Johns (St. Augustine), Orange (Orlando), and Lee (Alva) counties also had resistant Ae. aegypti and low RR Ae. albopictus.
Examination of kdr alleles in Ae. aegypti strains from 62 locations showed a range of genotypes (Figs 2 and 3). We observed that most strains were fixed or nearly fixed for the F1534C SNP. For many strains, more than 95% of the tested mosquitoes were homozygous for the 1534C (1534CC) and the remainder of the strain was made up of a few 1534 heterozygotes (1534FC). Only strains from Big Coppitt Key (13% FC), Cortez in Manatee County (38% FC), Central Wynwood (22% FC) and 2 strains from Little River (20% and 14% FC) had more than 10% of mosquitoes that were heterozygous at position 1534. Most strains had no mosquitoes without at least one copy of the 1534C allele but we did find two strains from central Wynwood (8% FF) and Cortez (11% FF) that still had appreciable numbers of susceptible alleles.
There was much more variation throughout the state at position 1016. Strains from Big Coppitt Key, Longboat Key, Orlando and Clearwater had the lowest percentages of the homozygous 1016II at 11.0, 14.1, 14.9, and 17.8%, respectively. We did observe strains with high levels of 1016II. Six of eight strains examined from Pasco County had 1016II frequencies above 75%, including two strains from New Port Richey at 100% 1016II. Due to the 2016 Zika outbreak and resulting massive public health response, we heavily sampled strains from southeast Florida for allele frequencies. Two strains from Broward and several from Miami-Dade County had 1016II frequencies above 75% including a Miami Beach strain at 91% 1016II (Fig 3). Select strains from Lee and Collier counties in southwest Florida were also above 75% 1016II (Fig 2).
As with RR, in several counties we observed large differences in allele frequency from one part of the county to another. In Lee County (Fig 2), an inland strain from Alva had a relatively low level 1016II frequency (20%) while four strains from near the Gulf Coast all had 1016II frequencies of greater than 55%. This variation was also observed at much closer scale between strains in neighboring cities. The Manatee County Anna Maria Island strain had a much higher 1016II allele frequency than nearby strains from Palmetto or Longboat Key (Fig 2). While overall IICC frequencies were high in Pasco County, we did find variation at the neighborhood level in south Pasco County, where a strain with 75% 1016II was found just blocks from a strain with 30% 1016II (Fig 2). In Miami-Dade, we were able to test four strains from Wynwood and three from Little River collected during the 2016 Zika response (Fig 3). Again, we observed variations in the levels of 1016II within each neighborhood. North Wynwood was slightly more than 25% IICC while south Wynwood was nearly 50%. In Little River, the four strains ranged from 20 to 60%. Just across the water from Wynwood, the Miami Beach strain was greater than 90% IICC.
We regularly observed only six genotypes in the field (Table 4). Only a combined seven mosquitoes of the 4,810 analyzed in this study had genotypes of IIFF, IIFC, or VIFF, which would require an allele coding for a resistant isoleucine at 1016 (1016I) and a susceptible phenylalanine at 1534 (1534F) to be contributed from at least one parent. We did commonly observe the reverse where homozygous or heterozygous susceptible 1016V alleles paired with the homozygous resistant 1534C allele (VVCC or VICC). Plotting the combined dataset for strains with both kdr genotype frequencies and resistance ratios from topical application of permethrin showed a strong correlation between increasing RR and increasing 1016II frequency (Fig 4, Pearson correlation coefficient = 0.905, P<0.00001, n = 20). Strains from Miami Beach, Fort Myers, and New Port Richey, with high frequencies of 1016II, which, based on the rarity of IIFF and IIFC mosquitoes implies the genotype IICC, were the strains with the highest resistance ratios. In contrast, the strain from Big Coppitt Key had the lowest 1016II frequency and the lowest RR even though the 1534CC frequency was relatively high.
Recent local transmission of Zika virus during 2016 and small outbreaks of dengue virus in 2010 and 2011 demonstrate that an effective IVM plan that attacks multiple life stages to reduce mosquito numbers is a necessity. However, when disease transmission is active, chemical adulticiding can be the only means to immediately reduce the population of potentially infected mosquitoes. While there is some debate as to overall efficacy of adulticiding against Ae. aegypti and Ae. albopictus, there is little question that it is a necessary part of the response that allows other slower methods of control like larviciding and source reduction to gain a foothold. Pyrethroids are the major class of chemical adulticides that Florida mosquito control programs use operationally. Thus, the efficacy of pyrethroids on Aedes vectors is a critically important part of a control program.
Direct topical application of permethrin clearly demonstrated that resistance was widespread in Ae. aegypti strains throughout Florida. Resistance ratios ranged from about 6-fold in strains from Manatee County and Big Coppitt Key to approximately 60-fold in Miami Beach, Cape Coral, and New Port Richey. Genotyping thousands of individuals indicated that common kdr alleles were also widely distributed in the state. While these findings are noteworthy as they represent the first published report of widespread pyrethroid resistance and kdr alleles in the southeast US, this result is not surprising and could be predicted based on the results from other locations including Puerto Rico, Mexico and, more recently, in invasive CA Ae. aegypti strains [11, 13, 20].
The dataset described in this study does reveal wide variations in both RR and kdr alleles within small geographic areas. Examining Miami-Dade County, the strain collected from Miami Beach was highly resistant and had high levels of kdr. However, strains from inland Wynwood and Little River were much less resistant than the Miami Beach strain. Even strains from the opposite ends of neighborhoods differed. We saw this disparate pattern in south Miami-Dade, Manatee, and Lee counties. In Pasco County, we saw very different genotypes in strains geographically close to one another, separated only by a major highway. This wide variation in resistance alleles has been observed in Mexico and has been proposed to predominate over natural gene flow [34]. Considering this variability, along with the relatively short flight ranges and limited immigration in Ae. aegypti [35], the very different allele frequencies we observed in geographically close strains would support performing testing in numerous areas of a control district to get an accurate resistance picture.
In contrast to the resistance in the Ae. aegypti strains, we observed very little permethrin resistance in Ae. albopictus statewide. This was true whether Ae. aegypti were present or absent from the same collections. Waits et al. [24] showed very low levels of resistance in St. Johns County strains collected from areas without Ae. aegypti. Miami-Dade, Orange, and Lee counties all had the same pattern of resistant Ae. aegypti and much less resistant Ae. albopictus. It has been proposed that the development of pyrethroid resistance is much more difficult in Ae. albopictus due to sequence differences in the voltage-gated sodium channel (NaV) that make kdr much less likely, although recent reports indicate it may be possible [36, 37]. Our laboratory efforts to induce permethrin resistance in wildtype Florida Ae. albopictus by gentle pressuring have failed. At this time, pyrethroid resistance in Florida Ae. albopictus does not appear to be an issue that could lead to adulticide failure.
An important observation made due to this work is the correlation between increasing RR and the frequency of the IICC genotype in Ae. aegypti, which has been anecdotally observed in other studies using the CDC bottle bioassay [6, 11, 13, 38]. The linkage between kdr mutations and a strong correlation with resistance ratios has also been observed in other dipterans [39, 40]. While more research must be done to validate the correlation, this dataset adds another 20 strains with both resistance and kdr data that support using kdr genotype as a surrogate to estimate pyrethroid resistance levels in mosquitoes [9, 41]. The use of allele frequencies has several potential benefits. Genotype data are relatively easy to collect, the results are produced in hours, and dead mosquitoes collected during standard surveillance activities can be used to provide information on resistance. With limited budgets and personnel at the operational vector control district level, predictive estimation of resistance levels could produce useful operational data from activities already being done without requiring additional efforts to collect or produce mosquitoes for bioassay testing.
Nearly a decade after Donnelly et al. [41] asked whether kdr genotypes are predictive in mosquitoes there is, to our knowledge, no published study that shows a pyrethroid-resistant strain of Ae. aegypti without also showing the presence of kdr alleles. We suggest that the major benefits to be gained from use of allele frequencies as an estimator of resistance would be improvements in coverage area (this study, 62 strains with allele frequencies vs. 26 by direct topical application) and better operational decision-making as vector control programs could access more timely information on area specific resistance levels. These challenges of getting wide coverage and providing this information on an operationally useful timeline have been present in recent response efforts to Zika virus. The efforts of CDC and vector control units in Puerto Rico and the efforts of the authors and others in Florida to develop wide-area pyrethroid resistance maps by relying strictly on bioassay data show that it is currently a slow, labor intensive process. In Florida at least, the epidemic had passed by the time more than a few strains had been tested by bioassay. Until there is reliable, published evidence to argue against the use of kdr frequencies as a predictor, it is at present the only rapid way to assess strains across a large area.
This study shows that permethrin resistance is widely present in variable intensity in Ae. aegypti throughout Florida. This variability points to the need to include resistance testing as part of an IVM plan as well as examine resistance in more than one location. But how do we use this resistance information to improve vector control? Clearly, the strain of Ae. aegypti in Miami Beach is very different from nearby Wynwood and would likely call for different treatment strategies. A treatment with permethrin would likely have much less effect in Miami Beach in comparison to other less resistant locations.
Our study also points to the value of a collaborative approach from motivated stakeholders to develop resistance information. The state of Florida began regular conferences to bring together vector control districts, researchers, and public health resources months before the first locally transmitted case of Zika was reported in 2016. Like the CDC response in Puerto Rico, development of resistance information was an early and ongoing part of this process. The dataset in this study represents the result of thousands of hours of effort from vector control districts, vector control contractors, the state of Florida, and the federal government to produce operationally useful resistance information to protect public health and improve the efficacy of control operations. However, even with these efforts, this study is very limited in scope. We examined only permethrin resistance and although the literature shows the patterns we observed would likely be applicable to other pyrethroids [13, 17], work to define statewide patterns of resistance to synergized products or organophosphates still needs to be addressed.
|
10.1371/journal.pgen.1003125 | A New Isolation with Migration Model along Complete Genomes Infers Very Different Divergence Processes among Closely Related Great Ape Species | We present a hidden Markov model (HMM) for inferring gradual isolation between two populations during speciation, modelled as a time interval with restricted gene flow. The HMM describes the history of adjacent nucleotides in two genomic sequences, such that the nucleotides can be separated by recombination, can migrate between populations, or can coalesce at variable time points, all dependent on the parameters of the model, which are the effective population sizes, splitting times, recombination rate, and migration rate. We show by extensive simulations that the HMM can accurately infer all parameters except the recombination rate, which is biased downwards. Inference is robust to variation in the mutation rate and the recombination rate over the sequence and also robust to unknown phase of genomes unless they are very closely related. We provide a test for whether divergence is gradual or instantaneous, and we apply the model to three key divergence processes in great apes: (a) the bonobo and common chimpanzee, (b) the eastern and western gorilla, and (c) the Sumatran and Bornean orang-utan. We find that the bonobo and chimpanzee appear to have undergone a clear split, whereas the divergence processes of the gorilla and orang-utan species occurred over several hundred thousands years with gene flow stopping quite recently. We also apply the model to the Homo/Pan speciation event and find that the most likely scenario involves an extended period of gene flow during speciation.
| Next-generation sequencing technology has enabled the generation of whole-genome data for many closely related species. For population genetic inference we have sequenced many loci, but only in a few individuals. We present a new method that allows inference of the divergence process based on two closely related genomes, modelled as gradual isolation in an isolation with migration model. This allows estimation of the initial time of restricted gene flow, the cessation of gene flow, as well as the population sizes, migration rates, and recombination rates. We show by simulations that the parameter estimation is accurate with genome-wide data and use the model to disentangle the divergence processes among three sets of closely related great ape species: bonobo/chimpanzee, eastern/western gorillas, and Sumatran/Bornean orang-utans. We find allopatric speciation for bonobo and chimpanzee and non-allopatric speciation for the gorillas and orang-utans. We also consider the split between humans and chimpanzees/bonobos and find evidence for non-allopatric speciation, similar to that within gorillas and orang-utans.
| The decreasing cost of genome sequencing has led to the sequencing of many species, including closely related species. With these available genomes, it becomes possible to study the speciation process in more detail. Due to the processes of recombination and coalescence, the complete genomes of two related species contain a large number of partly independent histories, with different regions having different migration histories and times to common ancestry. These histories, if inferred, can inform us about key parameters of the species divergence process. Particularly informative is the distribution of the length of genomic fragments having identical histories, since this depends directly on the time interval over which recombination could have acted and therefore on the distribution of coalescent times. Demographic parameters shape this distribution, so inference of it is informative about demographic history. However the difficulty of modelling coalescence with recombination means that most previous approaches have been unable to exploit this information, with one recent notable exception being inference of population size history from a single diploid genome [1].
Previous isolation with migration (IM) models have been designed to deal with relatively short sequences from several individuals of each species, since this was typical of data sets available until recently. Adding more individuals to a data set is often not as powerful as adding loci, since most coalescence events occur recently in the history of the samples and there are only few ancestors present at deep coalescence times. For the constant population size coalescent, the total branch length in the ancestry of a sample set grows logarithmically with the number of samples, but linearly in the number of loci, and most statistical methods for exploring the evolution of closely related species therefore employ multiple loci with small sample sizes [2]–[7]. These methods, however, typically assume loci are sufficiently short and widely separated that recombination is negligible within them and occurs freely between them. One notable exception is the MIMAR model, which does allow recombination within loci but still assumes free recombination between loci, see Becquet and Przeworski (2007) [8].
Using coalescent theory it is reasonably straightforward to compute the coalescence density under many demographic scenarios, including scenarios with and without gene flow. This follows from the Markov property of the process when viewed backwards in time. When recombination is absent, we can often derive simple maximum likelihood algorithms for inferring parameters. When recombination is present, however, the likelihood computations quickly become computationally infeasible since the process is not Markovian across loci [9]. Reasonable approximations can be made, however, by assuming the Markov property across loci [10]–[12]. In order to fully model complete recombining genomes we have developed a class of models that we term “coalescent hidden Markov models” or CoalHMMs. These models are based on the sequential Markov coalescence approach [10], [11] which models the coalescent process as a Markov process along a genome alignment. The coalescence states, however, are hidden and can only be inferred by comparing the sequences.
CoalHMMs permit recombination between any neighbouring pair of nucleotides, and represent the correlation between sites as a Markov model along the alignment [1], [13]–[16]. In Hobolth et al. (2007) [13] and Dutheil et al. (2009) [14] we analysed alignments of three genomes (human, chimpanzee and gorilla) using a Markov model with four states, in which one state corresponds to a genealogy consistent with the species phylogeny, and with the two most closely related species coalescing recently, and the other states correspond to the three possible genealogies with a deep initial coalescence (further back in time than the deepest speciation). The same model was used to analyse the human, chimpanzee and orang-utan in Hobolth et al. (2011) [17]. In Mailund et al. (2011) [15], on the other hand, the model had a variable number of states corresponding to different coalescence times in an alignment of two orang-utan genomes. In all models, speciation was modelled as a simple isolation model, with panmictic mating before the speciation event and no gene flow following speciation.
In this paper we extend the coalescent hidden Markov model of Mailund et al. (2011) [15] to an IM model, where we allow limited gene flow after an initial population split, followed by a period with no gene flow (see Figure 1). We derive the transition probabilities for the Markov model from finite state continuous time Markov chains parameterized by the split times and the rates of coalescence, recombination and migration. With this approach we exactly compute the transition probabilities between divergence times for pairs of nucleotides according to the coalescent process with recombination, and by assuming a Markov dependency along the alignment we obtain an approximation to the process that is computationally efficient for scanning whole genome data. We apply the approach to data from three pairs of recently diverged great ape species: the two orang-utan species, the eastern and western gorilla species, and chimpanzees and bonobos. We also apply it to the more ancient divergence between humans and the Pan genus (chimpanzees and bonobos).
When considering only pairs of genomes, the likelihood of a model will depend only on the divergence time at each locus [1], [3], [15]. By computing the joint coalescence time density for pairs of nucleotides we can compute a density for the coalescent time of the right nucleotide of a pair, conditional on when the left nucleotide coalesces. By assuming the Markov property along an alignment we can then efficiently compute the coalescence density for each nucleotide along a pairwise genome alignment.
In Mailund et al. (2011) [15] we used this observation to compute the joint coalescence density of pairs of nucleotides in a simple isolation model from a continuous time Markov chain (CTMC). In this paper, we take the same approach, but we now compute the joint coalescence density of pairs of nucleotides in a scenario which includes a period of restricted migration. In this setting, the state space of the CTMC explodes in the number of states. Constructing the CTMC manually thus becomes tedious and error prone, and instead we have implemented an algorithm for constructing the rate matrix (see Methods).
Once the joint coalescence densities are computed however, we can construct a hidden Markov model from following the approach of our previous work. We discretize time in a number of intervals (see Figure 1) to get a finite state space for the hidden Markov model. In all analyses we used 20 intervals, 10 in the migration period and 10 in the ancestral population; see Text S1 for results for different numbers of time intervals. From the joint coalescence densities we can compute the probability of the left nucleotide of a pair coalescing in one time interval and the right nucleotide in another, and from this obtain the transition probability matrix of the hidden Markov model. We compute the mean coalescence time in each time interval to get the emission probabilities of the hidden Markov model, assuming that coalescence occurred at that time point. Once we have computed the transition and emission probability matrices for the hidden Markov model, we can use well known hidden Markov model algorithms to compute the likelihood of a genome alignment.
The parameters of the hidden Markov model are the same as those used for the CTMC for computing the coalescence densities. The IM model introduced in this paper is parameterized by i) the coalescence rate of lineages, ii) the recombination rate between pairs of nucleotides, and iii) the rate of migration between populations. We scale all these rates by number of substitutions, such that the coalescence densities are measured in the expected number of substitutions, which simplify the likelihood computations.
We then assume three different time periods (see Figure 1). The period from the present until time allows coalescence and recombination events within populations, but no migration events between them. From time to time , migration events are allowed as well coalescence and recombination events, and further back in time than we assume a panmictic mating, where again coalescence and recombination events are allowed.
We assume that both the recombination rate and coalescence rate are constant across lineages and over time. We also assume that the migration rate is symmetric between the two populations. The mathematical framework used to construct the model does allow us to vary all rate parameters both in time and along the sequence, and allows the migration rate to be asymmetric between populations, but due to identifiability issues (see Text S1, Sections 4 and 5, and [7]) we restrict ourselves to symmetric parameters.
The coalescent HMM employs two important approximations. It assumes that the coalescent process is Markov along the sequences and that coalescent events occur at discrete time points rather than continuously in time. The Markov assumption is very difficult to relax because it enables us to reduce the problem of inference across the genome to that of the history of two adjacent nucleotides. The assumption of discrete coalescent times can be investigated by varying the number of intervals. Therefore, we have used extensive simulation studies to validate that the model can recover true parameters simulated under the coalescent with recombination process, both under ideal circumstances and under circumstances where different aspects of the model are mis-specified (see Text S1). Simulations are carried out under the more complex coalescent with recombination and migration model, whereas inference uses the assumptions of Markov property and discrete coalescence times.
For all simulations we used a coalescence rate of – corresponding to an effective population size assuming a substitution rate of substitutions per bp per year and 20 years per generation – and a recombination rate of – corresponding to 0.8 cM/Mb with the assumed mutation rate and generation time. To explore different scenarios we simulated 10 independent data sets for each combination of parameters ( and thousand years ago with the assumed mutation rate), (1 and 2 million years ago), and (). All simulation results are based on 10 Mbp of data (but see Text S1, Section 3.3 for accuracy as a function of data size). We present a broad range of analysis of the simulations in the supplement and will here only focus on a couple of key aspects.
Figure 2 shows the parameter estimation accuracy for the six different scenarios simulated. As shown, the parameters are generally well recovered except for the recombination rate that is consistently under-estimated.
The CoalHMM assumes that both the mutation rate and the recombination rate are constant in the region analysed, which in general will not be true. We explored the sensitivity to variation in rates through simulations. Figure 3 shows the effect of varying the mutation rate in segments along the alignment, in blocks of length geometrically distributed with mean either 500 or 2000 bp, and varying the mutation rate by a factor uniformly distributed in either range 0.75–1.25 or 0.5–1.5. The main effect of varying mutation rate appears to be a decrease in the estimated coalescence rate and an increase in the estimated migration rate. The decrease in coalescence rate is explained by a greater variance in estimated coalescence times when mutation rate variation is added to the variation in actual coalescence times. The model misinterpretes stretches of the genome with small divergence as recent coalescence times causes the increase in migration rate estimates.
Figure 4 shows estimation results when the recombination rate along the alignment is taken from the DeCODE recombination map [18]. Introducing variation in the recombination rate does not appear to bias the parameters, but the variance in estimates generally increases.
The model assumes that we have one haploid genome from each species. From sequencing data, however, we generally only obtain diploid genomes, and inferring the phase to split this into haploid genomes is not immediately possible with only one genome sequenced. If the species are sufficiently diverged, however, most polymorphism in the species will be local to one of the species, and which variant is considered for a heterozygotic site will not matter for the divergence to the other species. However, when species are so closely related that shared polymorphisms are common, we expect that assuming phased chromosome will make us believe that more recombination has occurred and therefore bias the inferred upwards. To test this we simulated two genomes per species and compared parameter estimates on haploid genomes and mosaic genomes constructed by taking a random allele at all positions where the two sequences in a species differed. As shown in Figure 5 and in the supplement, species have to have diverged very recently (within the last 250,000 years) for this effect to be detectable.
Recently-sequenced primate genomes allow us to apply the model on three different closely related species pairs: (a) bonobos and chimpanzees (Prüfer et al. (2012) [19]), (b) eastern and western gorillas (Scally et al. (2012) [20]), and (c) Sumatran and Bornean orang-utans (Locke et al. (2011) [21]). We analysed these species pairs in 10 Mbp intervals using both the IM model from the present study and the isolation model (I model) from Mailund et al. (2011) [15], in order to test whether a period of limited gene flow explains the data better than a clean split. Estimates for each 10 Mbp segment can be seen in Dataset S1.
Figure 6 shows the divergence times estimated under the I model (a single divergence time) and the IM model (two divergence times). In each case the I model estimates a time intermediate to the two divergence times of the IM model. The median migration rates per coalescence () are for bonobo and chimpanzee, for eastern and western gorilla and for Sumatran and Bornean orang-utan. Thus, the short migration epoch for bonobo and chimpanzee appears virtually panmictic while the epoch for the gorillas has very limited migration.
Figure 7 shows the divergence times for each chromosome and Figure 8 shows a comparison of the I and IM models for each chromosome in the three species pairs (for details of the model checking approach see Text S1, Section 8). As expected from the short time interval of gene flow and the large amount of gene flow estimated, the IM model does not provide a better fit to the data than the I model for the Pan comparison. For gorillas and orang-utans, however, the IM model is preferred, with the strongest support for the IM model in gorillas. We conclude that the Pan split is consistent with allopatry, whereas both the orang-utans and gorillas split non-allopatrically, and the split between gorillas much more recent than that between orang-utans.
Among the great ape speciation events, the human and chimpanzee speciation has received the most attention. We applied our new model to this speciation event using both the chimpanzee and bonobo genomes compared to the human genome, see Figure 9 and Figure 10. Estimating parameters using the isolation model, we see a recent speciation with a large ancestral population size (i.e. small coalescence rate), while estimating parameters using the isolation with migration model we find a relatively large interval with gene flow and a smaller ancestral effective population size, although still large compared to most extant great apes. Comparing the two models using the AIC approach, we find that the isolation with migration model is preferred, suggesting that the Homo/Pan split was not allopatric (see Figure 11).
Our study shows that detailed information on the divergence process can be gathered from just two related haploid genomes through joint inference of the length of segments with the same history and their times to coalescence. A period of time with limited migration is detectable as a period in which coalescences occur at a much lower frequency than in a panmictic population, because they are limited by the rate of migration events. Simulations under the full coalescence with recombination process show that the Markov assumption does not significantly bias the estimation of parameters, except for the recombination rate which is consistently underestimated, typically by 20–30%. The cause of this underestimation is the tendency of the real coalescent with recombination process to return to the same ancestor which implies that the average effect of recombination events is smaller than assumed by the Markov model (for an extended discussion, see Dutheil et al. (2009) [14] and Mailund et al. (2011) [15, Text S1 Section 1.4.3]).
Estimation of time and migration parameters is robust to typical violations of model assumptions such as mutation rate and recombination rate variation. A practical concern is that the model assumes haploid phased genomes whereas most genome sequences are a random mix of two haplotypes. Using a random phase should have no effect if the genomes are sufficiently diverged that they do not share any polymorphism; in this case the patterns of coalescence time will be the same for either phase at any position along the genome. If the genomes have shared polymorphism, then the phase will affect the estimated coalescence time, and with a high degree of shared polymorphism we expect that a random phase will cause the HMM to jump between states too often.
To investigate the consequences of this, we simulated diploid genomes and constructed haploid genomes by choosing a random phase, and then estimated parameters from this data (see Text S1, Section 7). We found that split times were very slightly overestimated when a random phase was used, while the recombination rate was underestimated by a somewhat greater extent and the coalescence rate could be overestimated by a factor of up to 50%. As expected however, the biases introduced by using a random phase quickly disappear when the genome divergence increases.
The demographic parameters inferred by our model are expressed in units of sequence divergence (substitutions per base pair). To translate these into units measured in years and effective population sizes measured in individuals requires both a genomic substitution rate and a generation time. A substitution rate has typically been estimated from fossil dates, with values around per base pair per year as typical for great apes.
Recent measurements of de novo mutations in modern humans, however, combined with studies of the generation time in humans and African apes, have revealed a mutation rate of around [22]–[24]. This rate is significantly lower than estimates based on fossil calibration. However there are constraints on how far back this can be extrapolated, given fossil evidence for earlier evolutionary events (for example the divergence of orang-utan from other apes seems incompatible with dates older than 15–20 Mya) [20]. It may also be that the per-generation rate differs in other apes. For these reasons, in Figure 12 we show how the estimated timescales for the speciation processes investigated depend on the mutation rate assumed. For comparison with previous analyses, we show similar plots annotated with mutation rates and time estimates in other studies in Text S2, Section 3.
Our application of the model to three closely related great ape species pairs revealed different speciation processes between chimpanzees and bonobos on one hand and the gorilla and orang-utan species on the other.
We estimate that the two gorilla species have experienced a long period of time with a very small amount of gene flow. Evidence for recent gene flow between these species was also found by Thalmann et al. (2007) [25] and by Scally et al. (2012) [20]. Thalmann et al. (2007) [25], using a mutation rate just below , estimated an initial population divergence 0.9 Mya to 1.6 Mya, with continued gene flow ceasing 80 kya to 200 kya. Under the same scaling we estimate a much more recent population split, later than 0.5 Mya, with gene flow continuing until quite recently. Scally et al. (2012) [20] also presented evidence for recent gene flow, but using a model that assumed initial divergence followed by gene flow continuing to the present day, and with found a much more recent divergence time corresponding to 300 kya with their scaling (see Text S2, Figure 8).
For the two orang-utan species we also find evidence for an extended period of limited gene flow but with a more ancient cessation of gene flow than is observed for the Gorillas. This is in agreement with the DaDi analysis presented by Locke et al. (2011) [21] which also posits a gradual divergence process. Other studies estimated much more ancient divergence times; for example Steiper (2006) [26] estimated a divergence time between 3 and 5 Mya and Becquet and Przeworski (2007) [8] a divergence around 1.4 Mya (although with a wide confidence interval that overlaps other estimates) using also (see Text S2, Figure 9). Our inference falls within these extremes, with the scaling we estimate initial divergence almost 600 kya and had an moderate level of gene flow over a period of 300 thousand years.
Finally, for the chimpanzees and bonobos we find no evidence against an allopatric separation. This is in agreement with Won and Hey (2005) [27] and Hey (2010) [28], who also used an IM model to study the separation between these two species. Indeed, the chimpanzee-bonobo speciation process has previously been suggested as an example of allopatric speciation in which the Congo River acted as a barrier to gene flow [29], since the present-day ranges of bonobos and chimpanzees are separated by that river. Fluvial drainage patterns in Central Africa may well have changed substantially in response to geological and climatological events over the last 20 million years, and could have triggered speciation. However the formation of the Congo River itself may have occurred substantially more than 2 million years ago [30], in which case it would predate most estimated divergence times for chimpanzee and bonobo, including those presented here. Prüfer er al. (2012) [19] used an isolation model and patterns of incomplete lineage sorting between the two Pan species and humans, and estimated the split time to be about 990 kya. Using the same scaling factor (mutation rate per year) we find a split time estimate of around 800 kya. Our estimate is consistent with previous estimates based on IM models from Won and Hey (2005) [27] and Hey (2010) [28], which also inferred no migration between bonobos and common chimpanzees. It is also close to the estimate of Becquet and Przeworski (2007) [8]. This study detected a weak signal of gene flow between eastern chimpanzees and bonobos, but the whole-genome analysis of Prüfer et al. (2012) [19] does not indicate gene flow between bonobos and any of the common chimpanzee sub-species. (see Text S2, Figure 7).
Thus, most estimates for the bonobo-chimpanzee separation are largely consistent with our results and there is little evidence against allopatry. If we use recent estimates of present-day human mutation rate of about instead of , the above time estimates should be multiplied by 2 (see Figure 12), putting the bonobo-chimpanzee separation around 1.5 Mya, closer to but still likely postdating the formation of the Congo River.
The human-chimpanzee speciation has been the focus of considerable attention in previous studies, most of which have assumed a simple (allopatric) speciation model. However, as shown in Figure 11, we find evidence favouring a non-allopatric model, in which the initial divergence was followed by gene flow for an extended period.
Considering the evidence reported here and previously for gene flow between species within the great ape genera [20], [21] and between modern humans and archaic humans [31]–[35], it appears that non-allopatric speciation is not uncommon within the great apes, and it is plausible that a similar scenario may have applied to the split between humans and our closest relatives. Patterson et al. (2006) [36] proposed a complex scenario involving an initial split, followed by isolation, then an admixture event and finally an isolation between the species. Several subsequent analyses concluded that an allopatric speciation could not be rejected but did not conclusively rule out more complex scenarios [5], [7], [20], [37]–[40]. Our method does not explicitly model admixture so we cannot directly test the hypothesis presented by Patterson et al. In particular the model does not distinguish between gene flow occurring as a period of limited ongoing exchange between diverging populations or in the form of one or more admixture/hybridisation events. This is a key question to explore in future extensions.
The exploitation of whole genome data in a demographic inference model is made computationally efficient by the Markov assumption underlying CoalHMMs, and should increase the statistical power for parameter estimation and for comparing different demographic scenarios. However the construction of complex demographic models with CoalHMMs is complicated by the mathematics involved in specifying transition probabilities between local genealogies.
The model we have presented in this paper is an initial attempt at building a speciation model using a CoalHMM, but the underlying framework, using a continuous time Markov chain to model transitions between genealogies, generalises straightforwardly to other demographic scenarios (see [41] for initial work in this direction). By varying the coalescence rate in different time intervals, the model captures variation in the effective population size in the past in essentially the same way as the pairwise sequential Markov coalescence (PSMC) model of Li and Durbin [1]. Varying the migration rate in a similar manner, rather than assuming a constant rate of migration during an extended speciation event, could provide information about the timing of admixture events and could also model scenarios such as a gradual speciation or the complex speciation between humans and chimpanzees suggested by Patterson et al. (2006) [36]. Adding further populations and genomes is also feasible but is limited by the state space of the continuous time Markov chain.
Using a hidden Markov model also enables us, via posterior decoding, to investigate variation in coalescent times, recombination and gene exchange along the genome. See Text S1, Section 10 and Text S2, Section 1, for initial results using posterior decoding to estimate the time to the most recent common ancestor and to detect signals of selection. Such variation is expected to be an important aspect of speciation [42], and is seen in studies of hybridisation between closely related species [43], [44]. The ability to explore it in genome-wide comparisons between populations at various stages of divergence will be important in understanding the range of evolutionary processes involved in speciation.
The crux of constructing a coalescent hidden Markov model is deriving transition and emission probability matrices from the coalescence process parameters of interest. For computing the transition probabilities we take the approach from Mailund et al. (2011) [15] and construct continuous time Markov chains (CTMCs) that explicitly track the ancestry of pairs of neighboring nucleotides. From these we can compute the the transition probabilities exactly. For emission probabilities we compute the coalescence de§nsities in the models from similar CTMCs from which we compute the mean coalescence time in each time interval. Based on the mean coalescent time, we then compute the distribution of alignment columns and use these for the emission probabilities.
The model was implemented in Python, and we used the numerical optimization functionality from the scipy optimize module, function fmin to find the maximum likelihood parameters and HMMLib [50] to compute the likelihood for the hidden Markov model. The implementation is available as Dataset S2.
For our simulation experiments, we simulated ancestral recombination graphs from the coalescent with recombination process using the CoaSim tool [51]. From this we extracted local (tree) genealogies and simulated sequences over these using the Bio++ suite [52] with the Jukes-Cantor JC69 substitution model.
Genome sequence alignments between eastern and western gorillas and between Bornean and Sumatran orang-utans were generated as follows. Illumina paired-end reads for Mukisi, an eastern lowland gorilla (Gorilla beringei graueri) [20] were aligned using BWA [53] to the gorilla reference assembly (UCSC identifier gorGor3.1), which represents the genome sequence of a western lowland individual (Gorilla gorilla gorilla). Similarly, Illumina reads from kb5404, a Bornean orang-utan (Pongo pygmaeus) [21] were mapped using Stampy version 1.0.13 [54] to the (orang-utan reference assembly (ponAbe2), which represents the Sumatran species (Pongo abelii).
In both cases mapped reads were merged using Picard (http://picard.sourceforge.net), and duplicate reads were removed and pileup information generated using Samtools [55]. Consensus sequences were called at every position on each reference based on the majority vote of aligned bases, with positions having no aligned reads represented by ‘N’. This produced two consensus genome sequences, each the same length as the corresponding reference, one representing eastern lowland gorilla and the other Bornean orang-utan. These sequences were used in subsequent analyses.
Genome alignments for the analysis of the chimpanzee-bonobo, human-chimpanzee and human-bonobo splits were produced as described in Prüfer et al. [19, Supplementary Information 3]. Briefly, pairwise lastz [56] alignments were generated from bonobo (scaffolds, i7) to human, chimpanzee (panTro2) to human and orang-utan (ponAbe2) to human. These alignments were processed using the programs of the UCSC genome browser pipeline [57], [58] and joined on the human reference using the multiz package [59]. Bonobo and chimpanzee bases with a base quality lower than 30 were masked in the resulting multiple genome alignment.
|
10.1371/journal.pcbi.1004244 | Quantitative Analysis of the Association Angle between T-cell Receptor Vα/Vβ Domains Reveals Important Features for Epitope Recognition | T-cell receptors (TCR) play an important role in the adaptive immune system as they recognize pathogen- or cancer-based epitopes and thus initiate the cell-mediated immune response. Therefore there exists a growing interest in the optimization of TCRs for medical purposes like adoptive T-cell therapy. However, the molecular mechanisms behind T-cell signaling are still predominantly unknown. For small sets of TCRs it was observed that the angle between their Vα- and Vβ-domains, which bind the epitope, can vary and might be important for epitope recognition. Here we present a comprehensive, quantitative study of the variation in the Vα/Vβ interdomain-angle and its influence on epitope recognition, performing a systematic bioinformatics analysis based on a representative set of experimental TCR structures. For this purpose we developed a new, cuboid-based superpositioning method, which allows a unique, quantitative analysis of the Vα/Vβ-angles. Angle-based clustering led to six significantly different clusters. Analysis of these clusters revealed the unexpected result that the angle is predominantly influenced by the TCR-clonotype, whereas the bound epitope has only a minor influence. Furthermore we could identify a previously unknown center of rotation (CoR), which is shared by all TCRs. All TCR geometries can be obtained by rotation around this center, rendering it a new, common TCR feature with the potential of improving the accuracy of TCR structure prediction considerably. The importance of Vα/Vβ rotation for signaling was confirmed as we observed larger variances in the Vα/Vβ-angles in unbound TCRs compared to epitope-bound TCRs. Our results strongly support a two-step mechanism for TCR-epitope: First, preformation of a flexible TCR geometry in the unbound state and second, locking of the Vα/Vβ-angle in a TCR-type specific geometry upon epitope-MHC association, the latter being driven by rotation around the unique center of rotation.
| The recognition of antigenic peptides by cytotoxic T-cells is one of the crucial steps during the adaptive immune response. Thus a detailed understanding of this process is not only important for elucidating the mechanism behind T-cell signaling, but also for various emerging new medical applications like T-cell based immunotherapies and designed bio-therapeutics. However, despite the fast growing interest in this field, the mechanistic basis of the immune response is still largely unknown. Previous qualitative studies suggested that the T-cell receptor (TCR) Vα/Vβ-interdomain angle plays a crucial role in epitope recognition as it predetermines the relative position of its antigen-recognizing CDR1-3 loops and thus TCR specificity. In the manuscript we present a systematic bioinformatic analysis of the structural characteristics of bound and unbound TCR molecules focusing on the Vα/Vβ-angle. Our results demonstrate the importance of this angle for signaling, as several distinct Vα/Vβ-angle based structural clusters could be observed and larger angle flexibilities exist for unbound TCRs than for bound TCRs, providing quantitative proof for a two-step locking mechanism upon epitope recognition. In this context, we could identify a unique rotational point, which allows a quantitative, yet intuitive description of all observed angle variations and the structural changes upon epitope binding.
| T-cells play a major role in cell-mediated adaptive immune responses necessary for the defense against foreign invaders and transformed malignant cells. Heterodimeric T-cell receptors (TCR) recognize antigenic peptides presented on the surface of cells by major histocompatibility complex (MHC) molecules. Recognition of MHC molecules presenting foreign peptides induces TCR signaling leading to T cell expansion and specific T cell functions such as elimination of virus-infected or transformed target cells. Therefore the immune system needs to balance the subtle distinction between self-restriction and self-tolerance and responses may reach extremes from multifunctional T-cell activation to tolerance induction. Due to the complexity of the signaling process, its mechanistic details are still not well understood. Several mechanisms of signal transduction have been proposed, which can be classified into (i) aggregation-, (ii) conformational change-, and (iii) segregation-models [1–3]. These three classes are not mutually exclusive. A conformational change in the TCR associated CD3 molecule was observed to be a basic early event in the signaling cascade [4]. In this context, mechanical forces applied by the TCR domains to the associated coreceptors are a suggested explanation [5]. Recent studies showed an antigen-specific conformational change of the A-B loop of the TCR constant α (Cα) domain for at least two TCR types. However, neither the structural details of this inter-subunit communication nor its initiation mechanism are yet known [4]. In order to provide the TCRs complex functions required for the signaling process, a variety of regulatory elements are involved in the process. Among those are the conformational changes within the TCR that are triggered during the early stage binding to the peptide-MHC (pMHC) complex. TCRs structurally consist of two membrane-anchored chains (α and β chain), which form two domains with an immunoglobulin-like (IG-) fold, one constant and one variable domain (Cα, Cβ, Vα, and Vβ). The variable domains of the two chains associate to the Vα:Vβ-complex, which binds to the pMHC-complexes and thus is responsible for antigen recognition. The overall structure is Fab-fragment like and each Vα and Vβ domain consists of a framework region and three antigen-MHC specific recognition loops, the CDR1 to CDR3 loops (Fig 1A and Fig 2B).
The capability of the immune system to recognize many different pMHC complexes is achieved by a vast variety of different TCRs. The T-cell repertoire was estimated to 2.5x107 for one human individual [6], and to 2x106 for mice [7]. This genetic variety together with the associated conformational differences within the TCRs seem to contribute to the structural and functional plasticity of TCRs [8]. The highly variable CDR3 loops encoded by VDJ recombinations are responsible for specific peptide recognition. Conformational changes within the CDR3 region after assembly with the pMHC complex have been demonstrated to provide an adaption to distinct peptide-MHC pairs which may additionally be influenced by CD8 co-receptor binding [9]. Moreover, the presence or absence of co-receptors as well as co-stimulatory molecules can have opposite effects on distinct TCR and T-cells suggesting an additional module for regulation [10]. Recently, it has been described that TCR Vα and Vβ domains can switch among alternate conformations when binding to MHC class I or II peptide complexes. A flex point in the FGXG motif of the J element has been proposed as swivel point for adjusting the interaction of Vα and Vβ [11]. In 1997 Li et al. proposed the capability of TCRs to increase their plasticity by rearranging the relative orientation of the Vα/Vβ domains, analogous to several known rearrangements of the VL/VH domains of antibodies [12]. Later works of Gagnon et al. reported a shift in the Vα/Vβ orientation of the A6 TCR bound to different ligands and influence of these shifts on the constant domains of the TCRs [13]. When the first structure of an A6:Tax:HLA-A2 complex was resolved, small variations in the Vα/Vβ interdomain angles could be determined [14,15]. This system was further studied with different agonistic or antagonistic peptides [16] and it was found, that different peptides induced these minor changes in the relative Vα/Vβ association geometries. Studies of the fluorination of the Tax-Peptide to increase the affinity confirmed this effect and also showed an alteration of the relative angle of the constant domains [13]. These scissoring effects were also observed for other receptors with different ligands or comparing the bound and unbound state: 2C [17], HA1.7 [18], LC13 [19], JM22 [20], DM1 [21], sc1.D9.B2 [22] and also for an invariant natural killer T cell receptor (NKT) [23].
The conformational changes were rather seen as further degree of freedom of the TCRs to adapt to the shape of their ligands [20,22]. A direct relationship between different conformational Vα/Vβ adjustments was not found [16].
In 2008, McBeth et al. systematically determined the Vα/Vβ interdomain angles for 35 TCR structures and concluded, that this angle is a general property of TCRs, which expands the repertoire of specificity [22]. Similarly, two recently published studies of Dunbar et al. investigate the interdomain geometries of antibodies [24] and compare them to the geometries of a non-redundant set of 39 structures [25]. The structure of TCRs is similar to Fab-fragments of antibodies [26], whereby the antibody VL/VH correspond to the TCR Vα/Vβ domains. In early and recent studies of antibody structures a rearrangement of the VL/VH upon ligand binding was considered and later confirmed [27–39]. Knowledge about TCR chain interactions might not only be important for the understanding of different TCR functions but may additionally provide information for reliable prediction of chain pairing. This is particularly interesting in T-cell based immunotherapy in which TCRs are considered as therapeutic tools for viral diseases and various cancers. For this purpose T cells redirected after genetic transfer of TCR chains with defined specificity are applied [40]. Understanding of TCR chain assembly is highly important in this regard as incorrect binding of introduced TCR chains with an endogenous TCR α and β chain may result in severe morbidity [41].
In the field of rational TCR engineering and optimization, homology modeling of these receptors gained in importance. Michielin et al. early created a homology model of the T1 TCR [42] using the MODELLER tool [43]. Later, other distinct TCR:pMHC models were investigated using more elaborate techniques including molecular dynamics (MD), computational alanine scan, or free energy calculations to study the influence of single mutations in the TCR or in the ligand, or to study differences of similar systems [44–54], and since recently, the automated modeling approach TCRep 3D is available to predict arbitrary TCR:pMHCI complex structures [55]. Recently Knapp et al. applied the ABangle methodology to a broad range of MD simulations of the LC13 TCR bound to 172 different ligands [24,56]. However, none of the previous modeling approaches explicitly includes any features concerning potential alterations in the Vα/Vβ interdomain angles, thus the presented structural analysis can help to improve the performance of the existing TCR modeling approaches.
In this work we perform a systematic, quantitative analysis of the Vα/Vβ interdomain angles in experimental TCR structures. For this purpose we developed a new structure-based method, which allows a systematic and very accurate quantitative comparison of the differences in the Vα/Vβ interdomain angles and introduces a new distance measure for clustering leading to a more accurate structural alignment of the TCRs than the approaches used in previous studies. The determination of TCR interdomain geometries is complicated by the fact that structural data is only available for a small subset of the vast variety of TCRs and that the TCRs for which structural data is available differ considerably in their loop structure and chain length, rendering the location of common conserved structural elements difficult. To solve these complications our method transfers all TCR variable domains into a unified geometric scaffold and performs a systematic analysis of the TCR structure geometries for 85 representative structures with respect to their Vα/Vβ interdomain geometries and interactions.
To analyze relative positions of the Vα and Vβ domains of all bound and unbound TCR structures in the dataset (Table 1, S1 Table and S2 Table) we introduced a new methodology which assigns uniquely defined cuboid-based frames to the individual Vα and Vβ domains of the TCRs and thus allows an unambiguous analysis of their relative geometries (for details see Methods). Based on this method we first examined the relative positions of the two domains with respect to each other and then performed a throughout analysis of the structural basis of the obtained observations.
For the analysis of the relative Vα and Vβ domain geometries we superposed the Vα domain of these structures and investigated the differences in the position of the corresponding Vβ domains using their assigned cuboid frames. For this purpose a conserved framework region was identified in both TCR chains and cuboids were placed around each variable domain centered on the framework region (Fig 1B–1D). Afterwards the relative Euler angles of the Vβ cuboids were measured with respect to the superposed Vα domains.
The analysis showed that the relative positions of the Vα and Vβ domains of the TCRs differ considerably with respect to each other (Fig 3), which is consistent with former qualitative studies on small subsets or individual TCRs [22]. In Fig 1B it can be observed that if the central β-sheets of the Vα domain are superposed very well, the backbone positions of the corresponding Vβ domains differ significantly featuring interdomain Euler-angle distances dE up to 30° (see Methods). Therefore, the two TCR binding domains can adopt different orientations (Fig 1B) with respect to each other. To analyze these differences in more detail we clustered all superposed (i) MHC-bound structures (Fig 3 and S1 Fig) as well as (ii) all MHC bound and unbound structures (S2 Fig and S3 Fig) according to their angular deviations in the Vβ domains with respect to the corresponding Vα domain using the Ward clustering algorithm [57]. Afterwards we performed a bootstrapping analysis (Fig 3) and identified six clusters with a significance greater than 95% [58]. Nearly all structures of TCRs of the same type from which different X-ray structures exist were placed in the same cluster (93% of the TCR types with more than one MHC bound crystal structure; except 2C TCR). This shows that the observed phenomenon is not caused by the variation of the crystallographic conditions and that the clustering is robust, describing a phenomenon which is caused by biological differences within different types of TCRs.
After the angle-based cluster analysis of the superposed structures we analyzed the structural features leading to the different interdomain geometries observed. For this purpose we used a grid-based analysis of the superimposed cuboid structures (Fig 1D, for details see Methods). This analysis showed that all TCR structures share an area, which is invariant towards rotation and translation of the TCR variable domains. At the center of this region a rotation point (Center of Rotation, CoR) can be identified, which exists in all TCR structures. The core region around this CoR is situated at the center between the two Vα and Vβ domains (Fig 1D). Notably, the average CoR position (x = 27.768Å, y = 36.783Å, z = 55.723Å) with respect to the reference coordinate system (2bnu) is located directly between or close to a twofold hydrogen bond between two conserved residues (Q for most of the structures), one from each chain (Fig 1D, magenta box). These hydrogen bonds connecting the two chains are known to be conserved through all TCRs [59]. As similar structural constraints were observed for antibodies [59,60], these features (CoR stabilized by conserved H-bonds) seem to be characteristic for Fab-fragment like domains in general.
To investigate the conservation of these two residues, we performed a sequence-based analysis with the sequences of all currently known functional variable αβ TCR gene segments as found in the database IMGT/GENE-DB [61]. In total 342 α chain and 164 β chain sequences were analyzed (six of 348 α sequences were incomplete). This analysis shows, that in contrast to antibodies, different residues can be found at the CoR position (Table 2 and S3 Table). Table 2 provides the absolute number of the observed amino acids at the CoR position separately for the known α- and β-alleles. The investigated CoR position corresponds to sequence position 44 in the IMGT unique numbering [62] scheme for both, the Vα and the Vβ domains. In case of the α chain the amino acids Q, H, R, K, L, W, and E can be observed, whereas in case of the β chain Q is overrepresented, but is occasionally replaced by R and K. The lower amount of different residues found in the β chain alleles might be a statistical artifact, since for the α chain about twice as many sequences are known than for the β chain.
Structural investigation of the interaction pattern of these alternative residues observed at position 44 showed that all of them can form strong interactions with their interacting partner residue in the complementary chain and thus compensate for the lost hydrogen bonds of the Q-Q interaction (see Fig 4): In some cases the Q residue of the α chain is replaced by an apolar W or L residue. In the case of W the Π-system of its indolyl group forms strong interactions with the Q residue from the opposite chain. In most cases Q is replaced by R or K and therefore the formation of the interchain hydrogen bonds can still be observed, as shown in Fig 4. As no structures are available for the replacement of Q by E or H no structural analysis is possible for these mutations. The same holds for the K mutant in the β chain, as in all available structures the conserved position in the β chain is occupied by Q except for the TCR KK50.4 (structure 2esv), where Q is replaced by R (Fig 4B). In this structure the side chain oxygen atom of the Q residue of the α chain forms a hydrogen bond with the guanidine group of the R residue. Compared to structures with Q-Q interactions, the Q residue is slightly displaced towards a neighboring loop, due to the size of the interacting R residue. This displacement allows a further interaction of the Q residue with a backbone carbonyl-oxygen of the neighboring loop. The α chain offers more diversity: K residues are found at the α-CoR position in the TCRs B7 (PDB ID 1bd2), 226 TCR (PDB IDs 3qiu, 3qiw), and 5c.c7 (PDB IDs 3qjh) (Fig 4A, 4D–4F). The rare W residue at the α-CoR position can be observed in the RA14 TCR (PDB ID: 3gsn, Fig 4C). In the β chain of the B7 TCR (1bd2) the CoR position is occupied by a conserved Q residue, the side chain oxygen is directed towards the side chain nitrogen atom of the K residue at the CoR position of the α chain. The distance between the two atoms is 3.59 Å. The K residue is drawn towards a neighboring loop, such that the amino group can also interact with an oxygen atom of the backbone of the loop (distance N-O: 2.89 Å). This additional interaction stabilizes the conformation of the K residue. For both “226 TCR” structures (3qui and 3qiw) very similar conformations of K- and the Q-residue can be observed: the side chain nitrogen atom of the K of the α chain is directed towards the side chain oxygen of the Q (distance in 3qui: 2.39 Å; 2.60 Å for 3qiw). In the 5c.c7 structure (3qjh) the atomic coordinates of the two observed Q and K residues are very similar compared to the two “226 TCR” structures. However, the oxygen atom and the nitrogen atom of the Q residue are swapped in one of the BUs, so that the K nitrogen is directed towards the nitrogen atom of the Q. In the RA14 TCR (PDB ID 3gsn), where W occurs at the CoR position, the conformation of the W residue is stabilized by a hydrogen bond (distance 2.15 Å) between the nitrogen NE1 and a backbone carbonyl-oxygen of the neighboring loop of the α chain. The W residue flanks the hydrophobic core of the TCR. The Q residue of the β chain pushes towards the solvent, due to the size of the W residue. For E and H we also expect the formation of hydrogen bonds with the β chain, however, no structure exist of this case yet.
Overall, the existence of the conserved CoR in such close proximity to the conserved αQ-βQ interactions confirms the hypothesis of a rotation-driven mechanism of α:β-association leading to the differences in the association angles of the Vα and Vβ domains. However, due to the low amount of mutated sequences available it was not possible to investigate the influence of the different amino acids occupying the conserved position 44 on the TCR interdomain geometry and the TCR specificity in a comprehensive manner. In general the above observations suggest an association mechanism of the Vα and Vβ domains in which the hydrogen bond interaction between the conserved residues are formed first and afterwards the domains arrange each other around this pivot point, adopting different relative association angles.
Next we performed a more detailed functional and structural analysis of the clustering results. For several different TCR types (2C, A6, 1G4, JM22, BM3.3, AHIII12.2, TK3) there exist more than one structure within the analyzed dataset in which the TCR is either bound to different MHC alleles and/or different peptides or different variants of the same TCR were crystallized.
These structures differ in several features: i) mutations in the TCR framework (S2 Table) or CDR-loop regions, ii) different presenting MHC molecules (including different alleles, single point mutations, and different MHC classes), and iii) different peptides presented to the TCR (including single point mutations). Furthermore, the data set includes the two similar TCRs “2B4” and “226”, which both share the gene loci for their variable segments of their α chains as well as their β chains, but differ in the loci for the joining segments.
Based on our cluster analysis we can distinguish between major and minor angular differences. According to our definition minor differences between two TCR structures occur if both structures can be found within the same cluster. Major angular differences between two TCR structures can be found for structures assigned to two distinct clusters.
Analyzing the clustering behavior of the different structures available for the same TCR types and their variants we observed the interesting results that for all except one TCR type (2C) all structures belonging to the same TCR type are located within one cluster (Fig 3). Thus, we investigated this phenomenon in more detail. In this section we briefly summarize the main results and refer the interested reader to a more detailed description in the supporting material (S1 Text).
Detailed analysis of the different structures available for the TCR types A6, 1G4, JM22, BM3.3, AHIII12.2, and TK3 shows that neither mutations within the TCRs nor the binding to different peptidic ligands of varying immunogenicity (e.g. A6 [13–16], S1 Table) or MHC alleles lead to major angular differences. However, minor angular differences are frequently observed within the individual sets (see Supporting Material).
In contrary, the 2C TCR can be found in two different angular clusters (Fig 3, Table 3). This is in agreement with the original publications, which show that depending on the pMHC bound, the 2C TCR can adopt two distinct docking orientations [63], but that on the other hand mutations in the CDR3 loop of the TCR do not lead to significant changes in its orientation with respect to the pMHC ligand, if the same pMHC is bound [64]. Regarding the bound pMHC alleles in our two clusters, we find that all TCRs bound to the MHC molecule H2-K1b associate in cluster 6, whereas all TCRs bound to the MHC molecule H2-Ld are located in cluster 4, except for the m67 variant, which is bound to H2-Ld, but located in cluster 6. As the reason for this unusual behavior of the m67 variant it was found that its mutation of the αCDR3 loop sterically enforces a conformation of the neighboring αCDR1 loop (binding the MHC molecule), which leads to a shift between the Vα domain and thus to different interdomain angles, closer to cluster 6 (Fig 2). The same conformation shift is also observed in the experimental publication, but as the docking orientation of the Vβ-domain on the MHC is retained and only the relative Vα loops shift, no significant changes are observed with respect to the overall docking orientation [64]. Due to this surprising result we had a closer look at the structures and discovered that actually two subtypes of the 2C TCR were crystalized: the wild type (wt) and the 2C T7 TCR, which differ in the framework region (S2 Table). The wt 2C TCRs are all bound to the MHC molecule H2-K1b and associate in cluster 6, whereas the T7 TCRs are all bound to the MHC molecule H2-Ld and belong to cluster 4, except the m67 variant. Therefore the two TCR structures compared in [63] actually belong to two different variants and thus the results and conclusions of that publication, namely that the different docking orientations are solely caused by the different pMHC ligand bound, need to be regarded with caution. Unfortunately, as for both variants only bound structures to the same pMHC are available and no “cross” TCR type ⇔ pMHC allele structures, it can not clearly be distinguished if the differences in the docking orientation and Vα/Vβ angles are caused by the framework mutations or the different pMHCs bound. However, regarding the above discussed 2C T7 m67 TCR, which is bound to H2-Ld and has the T7 framework mutations, but still adopts an angular conformation closer to the 2C-wt:H2-K1b (Table 3, underlined) structures belonging to cluster 6 (Fig 2B) [64], it seems that neither of the above features (framework mutation or pMHC allele bound) seems to induce unsurmountable restrictions on the final TCR conformation. As the m67 variant is the only T7 variant, which adopts the 2C wt Vα/Vβ angle, the induced changes in its αCDR1 conformation, which are not present in the other T7 variants (Fig 2B), seem to play a crucial role for the Vα/Vβ association angle, whereas the αCDR3 conformation influences the angle only indirectly.
These results indicate that the 2C TCR can in principle adopt two distinct conformations, which can be modulated by framework as well as the CDR mutations and presumably also its binding partner.
This is in agreement with previous qualitative observations about the overall TCR:pMHC association angle, stating that this angle is mainly dependent on the nature of the MHC allele and the TCR type rather than the antigenic peptide molecule [22,63–65]. As the CDR1 and CDR2 loops are interacting with the MHC molecule and the CDR3 loop predominantly with the bound peptide, the observed CDR1 dependent structural changes are in agreement with these former studies and might be a complementary feature to the CDR3-peptide binding in the process of TCR signaling. However, as these observations are based on one TCR only, these conclusions should be taken with caution.
To further investigate the influence of the pMHC complex on the overall TCR structure we compared the structural features of unbound and bound TCR structures of the same type (Fig 2A, S2 Fig and S3 Fig). We observed that in most of the cases the orientations of the unbound TCRs slightly differ from the bound TCRs. The seven TCR types 1G4, 2C, DM1, ELS4, JM22, 2B4, and LC13 can be found in the unbound state as well as in the MHC bound state in our data set–TCRs bound to superantigens are not considered. Only in the case of the 2B4 TCRs and the LC13 TCRs both states associate in the same clusters. In the other cases, the angular deviation of the unbound TCRs is between 5° to 11°, leading to an association to a different cluster than the bound variants. Comparing all examined structures of bound and unbound TCRs it can be observed that the differences in the Vβ domain orientations are considerably larger for the unbound TCRs (S3 Fig).
In Fig 2A the differences between the bound and the unbound structures are illustrated for several TCR types. The repertoire of analyzed 1G4 TCRs contains nine structures of wt TCRs and mutants. Two different structures are available in the unbound state: (i) The structure 2bnu is the wt and (ii) 2pyf is the variant c5c1, which differs from the wt in the αCDR3-, βCDR2-, βCDR3-loops, and in three positions of the framework region [66, 67]. The subset of bound 1G4 TCRs contains wt TCRs (2bnq and 2bnr), the variant wt-AV (2f54, contains solubility mutations in the framework region [68]), and variants, which contain mutations in the framework region and the αCDR2-, αCDR3-, βCDR2-, βCDR3 loops: c5c1 (2pye), c49c50 (2f53), c58c62 (2p5w), and c58c61 (2p5e)–S2 Table lists the mutations in detail. All ligands of the bound 1G4 TCR structures are the MHC molecule HLA-A*0201 presenting the peptide SLLMWITQC (except 2bnq: SLLMWITQV).
Notably, the two unbound orientations differ only by 2.5°, but have an average distance of 8.0° to the bound structures (Table 4). On the other hand, all bound 1G4 TCR structures are very similar (2.1°, var = 0.7°). This indicates a shift in the relative orientation of the two domains upon binding of the TCR to the peptide-MHC complex. Both unbound structures associate in cluster 2, differently to the bound 1G4 TCRs, which are found in cluster 1.
For both, the DM1 and the JM22 types, an angular deviation between the bound and the unbound state of 10.5/11° can be observed. The JM22 TCR was reported to reveal a considerably greater scissoring motion than other TCRs [20], which is consistent with our findings. The unbound JM22 TCR (2vlm) can be found in cluster 6, whereas the bound JM22 TCR structures are located in cluster 2. The bound (3dxa) or unbound (3dx9) DM1 TCRs can be found in cluster 4, respectively cluster 2. The unbound E8 (wt) structure (2ial) associates to cluster 1 and differs by 10.6° from the bound (wt) variants (2ian, 2iam), which associate to cluster 3. The bound variant of ELS4 (2nx5) is located in cluster 6 and differs to the unbound variant (2nw2, cluster 4) by 6.3°.
In the case of the 2C TCR, the unbound wt (1tcr) associates with cluster 6, which contains the bound 2C wt structures and the exceptional bound 2C T7 m67 variant (see above). The average angular distance of the unbound wt to these bound structures is 5.0° whereas it’s distance to the bound 2C T7 variants in cluster 4 is 10.5°. In contrast, for the LC13 and the 2B4 TCRs the angular difference between the bound and the unbound structure is low (3.3° and 3.9°, cluster 4).
Thus by including the unbound TCRs into the clustering process (S1 Fig), a tendency towards smaller significant clusters can be observed. This means, that the pMHC-ligand stabilizes the TCR variable domain geometries in a favored position. The TCRs’ ability to adopt multiple geometries might play an important role in the signal transduction and the loss of flexibility upon pMHC binding might induce an initial event in the signaling cascade.
Another interesting point is that structures from human and mouse are found in the same clusters, no differences were observed in their clustering behavior.
We performed a comprehensive quantitative analysis of the structural features of T-cell receptors in their bound and unbound states. For this purpose, we introduced a new cuboid-based method, which allowed us to obtain a unique quantitative measure for the Vα/Vβ association angles and thus the previously observed rotation between the two TCR domains. As our method is based on highly conserved framework residues and ignores the loop regions it can be applied to all possible chain combinations and we performed a detailed analysis based on a representative set of all currently available TCR structures in the PDB Database.
Differences in the TCR Vα/Vβ association angles were first recognized for the A6:Tax:HLA-A2 complex by Ding et al. [16]. Since then the same phenomenon was also observed for other TCR clonotypes by several groups [17–21,23]. The first more comprehensive analysis of the angular space of the TCR Vα/Vβ association was performed by McBeth et al. [22], who analyzed 38 TCR structures (biological units), including unbound TCRs, MHC I- or MHC II-bound TCRs, and three NKT TCRs. The analysis was based on three angles: two angles were defined as the pitch of a pseudodyad axis and a third angle described the rotation around this axis when superimposing the two variable domains.
The results of the study of McBeth showed that different TCRs adopt a broad range of orientations and that the orientation of TCRs of the same type in the bound and unbound states can differ. Furthermore, the authors observed angular differences between TCRs differing only in a few amino acids, concluding that the variation of the interdomain angle potentially has an effect on the TCRs specificity or polyspecificity [22].
The pseudodyad-based method used by McBeth et al. is a classical approach of crystallographers to determine the relative orientation of antibody V domains or to determine the antibody elbow angle in Fab fragments. The computation of the pseudodyad-axis is achieved by superimposing of the Vα onto the Vβ domain. The drawback of this approach is that the precision of this process depends on the similarity of the two domains and it can be expected that the cross-chain similarity of the variable domains is lower than the similarity between two variable domains of the same chain type (either Vα or Vβ). Thus, two variable domains of the same chain type can be structurally aligned more precisely than superimposing similar cross-chain domains. Due to these limitations we developed a new method for superpositioning, which allows a unique definition of the interdomain rotational angle by superimposing domains of the same type using structurally highly conserved regions for the superimpositioning process. Our method describes the orientation of the Vβ domain relative to the Vα domain by a unified cuboid instead of a pseudodyad-axis as used by McBeth. The cuboid-based description provides several benefits. First, only one angle is necessary to describe the interdomain rotation, which is not only intuitively accessible, but also forms the simplest description of the phenomenon and allows a straightforward bioinformatics structural analysis. Second, the Euler angle distance can be computed between cuboids, which can be used as a measure for clustering. Third, cuboid geometry combinations can be used as a template for an arbitrary cross type chain assembly in a modeling process.
Since 2008 the number of TCR structures available in the PDB increased considerably. Therefore, we were also able to base our analysis on a much broader data set. The data set of McBeth included 18 non-NKT and 3 NKT structures (38 BUs), whereas our set contains 37 different non-NKT TCR types (mutants not counted, 136 BUs). In both studies free, MHC I bound, and MHC II bound TCRs originating from human or mouse were studied. However, the recent data allowed us to compare additional TCRs in bound and unbound state (e.g. DM1, JM22, LC13, E8, 2C, 1G4). For other TCRs the new dataset contains structures with additional different pMHC ligands (e.g. A6, SB27, 1G4, 2C, LC13, Ob.1A1).
Recently another study was published by Dunbar et al. [25], which also analyzes the TCR Vα/Vβ interdomain angle. However, the focus of that study is on the comparison of TCR and antibody geometries, because of its importance in the field of rational design of TCR-like antibodies. This is quite different from our goal of a systematic comparison of the interdomain angle variations within different TCR structures depending on their surrounding and thus the publication of Dunbar et al. sheds light on an important, yet complementary aspect of TCR architecture. Due to its different focus, the study also differs in the methodology applied as well as the data set used and the results discussed. Regarding the data set Dunbar et al. examine a smaller set of 39 structures, which does not contain TCR type-based redundant structures to avoid statistical bias in the comparison with the antibodies. In contrast, the inclusion of different structures of the same TCR is a desired feature of our data set as our analysis focuses on the differences between the available TCR structures in dependence of their binding state and partners. However, in contrast to Dunbar’s study, we excluded NKT receptors (binding CD1d ligands) as well as structures containing superantigens as they show a different binding behavior and function and thus are not representative for TCR:pMHC complex structures.
Further, as Dunbar’s study focuses on the comparison of TCRs with antibodies, it is based on an adaption of the ABangle methodology to TCRs, which was originally developed for antibodies [24]. Although our method and ABangle have in common that they use the conserved positions in the IG-like domain for structural alignment, the ABangle method describes the rotation and translation by five angles, a distance, and a precomputed axis. A benefit of this method is the ability to inspect each component of the transformation separately and therefore it allowed identifying the main difference between the antibodies and the TCRs angular space, which lies in the HC2 (twist) angle. In contrast, the major goal of our analysis is to analyze possible orientations the two TCR variable domains can adopt depending on their type and state, functional mutations and the bound ligand. For this purpose we introduced a specialized robust method for the applied cluster analysis, which differs considerably from the method of Dunbar et al.: It reduces the variable domains to cuboids, to allow easy visualization of the transformational differences between two TCRs. Further, we describe the rotation of these cuboids by Euler angles, from which an Euclidean distance can be calculated, which is needed to obtain a robust clustering, as the commonly used RMSD-based measure was found to be too insensitive to capture the partially rather small angular differences between two TCRs. The Euler angle based measure showed a more robust performance and is, in addition, independent of protein translation. In contrary, the study of Dunbar et al. is based on the RMSD of the relative domain orientations, which is accurate enough to clearly distinguish between the two molecule classes (TCR and AB).
Finally, instead of several independent components we use one center of rotation (CoR) to describe the angular differences, which is a necessary prerequisite for the cuboid-based clustering and provides an intuitive measure for presenting and discussing our results. In addition, we use bootstrapping to confirm the significance of our clustering results [58].
As the focus of the study of Dunbar is on the comparison of the Vα/Vβ interdomain angles of TCRs and with the VH/VL angles of antibodies, the study leads to the important result that TCRs and antibodies differ significantly in their interdomain angles. However, it also demonstrates that TCR-like antibodies, which were specially designed for pMHC binding, can adopt TCR-like geometries. Thus the study provides an important contribution to a better, detailed understanding of the structural features and characteristics of immunoglobulin-like folds and should therefore be very helpful for the rational of protein-based pharmaceuticals.
In agreement with the majority of the previous, predominantly experimental studies on small sets of TCRs, our comprehensive cluster analysis of the bound structures of the TCRs showed that TCRs of the same kind normally occupy the same structural cluster. Only one exception was observed, in this case two different clusters were found for the wt and mutant form (T7) of the same TCR (2C) and a specific combination of mutations in the framework and the CDR3 loop led to a shift of one mutant structure (T7-m67) into the cluster of the corresponding wt structures. For all other TCR types only one cluster was observed for both wt and mutated MHC-bound structures. These observations indicate that the differences in the Vα and Vβ interdomain angles of the bound TCR structures are predominantly determined by the preformed chain combination and subtype dependent interdomain angles of the unbound TCR structures and that neither the type of bound peptide nor the presenting MHC molecule lead to a significant angle shift. This geometry can be altered within the range of the subtype structures of the same TCR by mutations in the CDR loops as observed for the 2C m67 TCR. However, the analysis of the 1G4 structures showed that most changes in the CDR sequences do not have a significant effect on the interdomain angles. The same holds true for the binding of different MHC-peptide complexes to the same TCR, e.g. as shown for the A6 TCRs. This is in agreement with the previous studies of these TCRs, which observed only small differences in the Vα/Vβ interdomain angles for the different variants and bound pMHC complexes studies of these TCRs [14–16].
Comparing bound and unbound structures of the same TCR, a strong shift in the interdomain angles was observed in most cases upon binding of the TCR to a pMHC complex, as the bound and unbound structures of the same TCR were observed in different structural clusters. Further analysis showed that the differences in the interdomain Euler angles between the bound and unbound structures of the same TCR were often significantly higher than the variation of these angles within the bound or unbound structure set.
As basis for the observed differences in the association angles a so-called Center of Rotation (CoR) could be identified. This CoR is situated in the vicinity of two to four conserved residues (mainly Q), which interact via hydrogen bonding or charged interactions thus stabilizing the rotation center. Sequence analysis of these conserved residues showed that in contrary to antibodies in TCRs different amino acids can occupy these positions. However, all observed side chain types share the capability to form directed interactions such as hydrogen bonds. Due to the limited amount of TCR structures featuring other residues than Q at these positions, analysis of a correlation between the occurrence of specific residues at these position and the observed interdomain angles was not possible. The observation of a CoR is in agreement with previous studies of individual TCR types, as e.g. Ishizuka et al. [20] observed that for the JM22 TCR a binding hotspot of Vα/Vβ could be a center of motion or rocking. In this study, all JM22 structures were superimposed to the Vβ domains and the hinge was located at the salt bridge Q38α-Q39β [20]. In addition, already in the first described Vα/Vβ complex structure (2C) the (i) conserved Q-Q interaction between Vα and Vβ was observed at the Vα/Vβ binding interface, as well as water mediated hydrogen bonds between conserved residues of both domains and a symmetric hydrophobic core consisting of further conserved residues [26].
These individual results are considerably substantiated by our broad analysis. Throughout our dataset only minor variations were observed in the position of the CoR, which is highly conserved. In addition, nearly no shifting motions were observed, which seem to play only a minor role in the adjustment of the variable domains compared to the angular displacement.
Next to the conserved hydrogen bonds around the CoR, the contact area between the Vα and the Vβ domain is dominated by hydrophobic residues and is shaped similar to a saddle joint. This shape allows a certain rotation and translation of the Vβ domain sliding on the Vα-interface. As found by our grid analysis, the center of rotation is located at this area, but is slightly flexible. In contrast to the constant domains, the variable domains are not bound by a rigid disulfide bridge, but are kept together more loosely at the center of rotation by conserved Q-Q H-bond interactions. Our sequence analysis showed, that the Q-Q interaction is highly conserved, but in minor cases Q can be exchanged by other H-bond donors or acceptors or charged residues. Amino acids, which neither form H-bonds nor salt-bridges occur very seldom. In the latter case, the CoR possibly is shifted to other less conserved residues in this area, such as Y. Thus, the conserved residues at the CoR area keep the CoR at a defined position, but nevertheless the nature of the interactions permits flexibility that leads to the different orientations of the variable domains.
These observations are consistent with most other studies, discussing this topic, as a twofold hydrogen bond interaction between the Q-residues of the Vα and the Vβ domain was already reported for the first resolved TCR structure (1tcr, 2C) [26] and the involved residues are highly conserved for TCRs as well as for VL/VH domains of antibodies [59]. Similarly, for antibodies it was proposed that in contrast to the constant domains the absence of a disulfide bond between the two variable domains is evolutionary preferred to allow for their rearrangement [38]. However, there exists one publication in which it was claimed that for A6 TCRs neither hydrogen bonds nor salt bridges can be observed between Vα/Vβ [13] and the authors propose that the diversity in Vα/Vβ rearrangement might be a result of the slippery hydrophobic interactions between the two variable domains. This is not only in contrast to the above discussed results from literature, but also our data, since we also observed the above described Q-Q interactions for all A6 TCRs (e.g. distances between the opposing atoms in structure 2gj6: D:Q37:NE2-E:Q37:OE1 = 3.06 Å, E:Q37:NE2-D:Q37:OE1 = 3.14 Å). Therefore, these electrostatic interactions, which are a magnitude weaker than a covalent disulfide bond, are highly conserved and most likely function as a flexible constraint, which keeps the two variable domains of the TCR in a preferred position, but at the same time allows for the necessary flexibility for their rearrangement upon binding to a specific pMHC complex.
Our analysis shows that TCRs of the same type bound to different ligands are normally found in the same clusters, whereas a significant change in the association angle can be observed upon binding of the TCR to the pMHC complex. Thus the question arises about the consequences of this behavior for the signal transduction cascade. According to our results two statements can be made. First, there seems to be a locking step upon pMHC assembly, during which the TCR is locked into a TCR clonotype specific geometry. Second, as the differences in the Vα/Vβ interdomain angles between the same TCRs bound to different binding partners, are rather small, the locking motion can be expected to be important during the signal transduction, whereas the differences in the absolute association angles are either not that significant or, assuming a signal is induced by the domain adjustment, only minor changes might be necessary. This agrees with most previous observations [22, 64], which show e.g. for the A6 TCR that peptide ligands with different affinities induce only minor changes in the relative positions of the variable domains [14–16]. In our analysis all A6 TCRs feature a very similar orientation of the Vβ domains and the structures all associate in the same cluster.
Due to its comprehensiveness our analysis puts these individual results on a common basis and provides thus a general picture of how pMHC binding influences TCR structure and function. Since many peptides with varying immunogenicity presented to the same TCR type only induce minor angular differences (see results), we conclude that signaling does not directly depend on a major change in the Vα/Vβ interdomain angle. In contrary, according to our analysis already minor changes of the Vα/Vβ interdomain geometry might have a significant influence on the triggering of the signaling cascade. These observations agree with the computational results of Knapp et al. who performed large scale MD simulations of 172 peptides of known immunogenicity presented to the LC13 TCR [56]. In that study, several features between a set of more immunogenic and less immunogenic peptides were compared, such as the Vα/Vβ geometries and the orientation of the TCR towards the pMHC, the solvent accessible surface area, the binding affinities, hydrogen bond footprints, and structural root mean square fluctuations (RMSF). The study confirms our results as the examined LC 13 TCR it was observed to adopt only a slightly more “open” binding site when recognizing more immunogenic peptides, which is consistent with the minor changes we see upon the binding of different ligands.
However, when discussing the topic of signal transduction, it needs to be stated that in our study we did not investigate positional changes of the constant domains. Such changes were observed in single studies and were postulated to have an influence on the minor changes in CD3 binding or activation [17]. On the other hand investigations of the relationships of constants domains of bound and unbound A6 TCRs showed no alteration [15]. However, possibly the conformational adjustment of the A6 constant domains might be very small. Different conformations of the Cα A-B loop dependent on the antigenic ligand were described and it was speculated that this alteration might induce the conformational changes of the CD3 molecule [69]. Possibly, the antigenic ligand induces first an adjustment of the variable domains, leading to a change of the relative positions of the constant domains and finally to the observed conformational change of the Cα A-B loop. This effect could either be achieved mechanically or by an alteration of the surrounding forces. For the JM22 TCR it was observed experimentally that the temperature factors of the constant TCR domains increase upon ligand binding [20]. This observation supports the idea of the Cα A-B loop becoming more flexible after other parts of the TCRs loose flexibility, as observed in this study through the locking motion upon pMHC binding. Regarding the structural analysis of the 2C TCR, this locking motion seems to be caused by interplay between the Vα/Vβ association angles and the bound-conformation of the MHC-binding CDR loops. MD simulations similar to the study of Knapp et al. [56] could be used to study these effects.
The structures of TCRs are generally similar to Fab-fragments of antibodies [14,26], whereby the AB VL/VH correspond to the TCR Vα/Vβ domains. In early and recent studies of AB structures a rearrangement of the VL/VH upon ligand binding was considered [25,27–39]. Computer-aided methods including MD simulations were carried out to investigate the changes of the elbow angle between the variable and the constant domains of antibodies [70–72]. TCRs feature a lower diversity in the variable loops 1 and 2 but a higher diversity in the CDR3 loop compared to ABs, resulting in a smaller diversity in the overall shape of the TCRs [14]. However, it was shown that ABs VL/VH association angles are generally incompatible to the angular space of TCRs binding to MHC molecules [25]. It remains interesting to apply our method on ABs investigating whether this molecule class also shares a CoR and if ABs of the same type adapt to similar association angles.
Since flexibility of the TCR Vα/Vβ interdomain association was considered for the first time in the end of the 90s of the last century [12], it took one decade to examine this phenomenon comparing the angular space of different TCR types due to the initial difficulties of obtaining experimental structures [22]. Now, immunologist can benefit from two independent new studies of the TCR interdomain association geometry by Dunbar et al. [25] and our present one. Both papers complement by focusing on different topics and methodologies. Whereas Dunbar et al. focus on the comparison between TCRs and antibodies, in this study we performed a systematic, exhaustive analysis of the Vα/Vβ interdomain angle for a representative set of experimental TCR structures. Our results are in agreement with the majority of previous experimental studies on small sets of TCRs. However, due to the comprehensiveness of our analysis we were able to put these individual observations on a broader, more general basis. This allowed us to deduce general features describing the relationship between TCR interdomain angle variations and pMHC binding and signaling.
First, our data clearly shows that the Vα/Vβ interdomain angle of pMHC-bound TCR structures can vary considerably, but is in most cases well conserved within the same TCR clonotype and its variants, independent of the ligand (pMHC) bound and individual mutations within the TCR. Nevertheless, there are individual exceptions like the 2C TCR, which show larger variations in their angle repertoire. Analysis of the 2C TCR structures revealed correlations between the Vα/Vβ interdomain angle, specific framework mutations, and conformational changes in the MHC-binding Vα-CDR loops due to Vα-CDR3 mutations. This is in accordance with previous experimental studies on individual TCRs, indicating that the Vα/Vβ interdomain angle is mainly influenced by the bound MHC allele and not the peptide. Unfortunately, due to the currently still sparse structural data available, no generalizable conclusions can be drawn about the dynamic mechanisms behind such Vα/Vβ angle switches.
Second, through a systematic analysis of the structural basis for the observed angular deviations we could identify a central point of rotation (CoR) common to all TCR structures independent of their state (bound or unbound) and type, which is stabilized by electrostatic and hydrogen bonding interactions. As in all previous studies the Vα/Vβ interdomain angle was described by at least three geometric quantities, the identification of one CoR, which allows a simple yet intuitive description of this functionally important, variable angle, sheds new light on the structural features and also the functional dynamics of TCRs and will also be important for the improvement of existing and for future TCR modeling approaches.
Third, analyzing bound versus unbound TCR structures, we observed that the angle variations between bound and unbound structures are more significant than between TCR structures bound to different MHC-peptide complexes or even mutated TCR structures with different specificities. This suggests that binding of the TCR to the pMHC complex is accompanied by a dynamic lock mechanism during which the two TCR variable domains are driven into a TCR-specific binding geometry leading to a stabilization of the TCR variable domain upon pMHC binding. In a previous study it was found that with a rigidification of the variable region the constant region becomes more flexible [20]. Furthermore, the influence of constant domain shifts was considered to be involved in CD3 activation [17] as well as a conformational change in the A-B loop of the Cα domain [69]. Supported by these observations we propose that locking of the variable domains upon ligand binding might enhance the motions of the constant domains in this oscillating system. The change of motion of the constant domains could then induce the conformational changes, which lead to CD3 activation and thus initiate T-cell signaling.
Based on these results the TCR/pMHC binding mechanism can be envisioned as a two-state process: First, preformation of the general α/β domain geometry in the free state and second, locking of this angle in a specific geometry upon association with the MHC-peptide complex. The last step might be an important feature during signal transduction upon binding.
A set of 85 X-ray crystal structures was acquired from the Protein Data Bank (PDB) [73]. The used structures contain bound and unbound TCRs from H. sapiens and Mus musculus. Receptors of invariant natural killer cells (iNKT) were not considered in our analysis. Although the iNKT receptors share sequential and structural similarity with other αβ TCRs, these special TCRs do not recognize pMHC ligands, but detect lipids presented by the MHC like CD1d molecule [74,75]. Thus iNKT:CD1d complexes must be treated separately. For the analysis each biological unit (BU) was treated as a separate structure, leading to a total amount of 136 different TCR complexes. For some analysis steps we computed averages over all BUs of the same crystal. This set is further referred to as S. The used structures are listed together with the names used in the literature and their bound state in Table 1. For some receptor types (e.g. A6, 1G4 etc.) several entries in the PDB are available. Furthermore, in some cases the names used in literature are ambiguous, since some of these structures bear mutations. Therefore, we compared the sequences of all structures against the wild type (wt), and we extended the TCR type names to indicate deviations. The sequence differences are presented in S1 Table and S2 Table in the supplementary material. This table also contains information about the appearing gene segments, differences in the CDR3 loops, and the bound ligands including the peptide sequence.
In a first preparatory step, all TCR structures were reduced to their variable binding domains (V). The constant domains of the TCRs were not considered for two reasons. First, in some of the TCR structures data are only available for the variable domains but not for the constant domains. Second, the constant and the variable domains are connected by a flexible loop. Superimpositioning the complete TCR chains onto their variable domains showed a visible displacement of the constant domains. Thus, including the constant domain into our superpositioning template would influence the structural alignment of the Vα domains.
Then we defined unified cuboids for each V domain of the different TCR chains. The cuboid templates (CTα and CTβ; Fig 1) comprise Vα or Vβ domain and a reduced set of the 2bnu Vα or Vβ domain framework residues. The reduced set was defined to allow for a robust superpositioning of the experimental structures during our future modeling procedure and during the geometrical measurements. For the superpositioning the tool DaliLite [76,77] was used together with the defined subset of V-framework residues as templates (Fig 1C). The Dali algorithm uses Cα-Cα distance matrices and does not depend on sequence information. Due to the high homology of the TCR framework regions, a solely structure based superpositioning method is indispensable for our needs.
To define the subset of residues contained in the superpositioning template, in a preparatory step the structures of each variable domain were superimposed separately. For this purpose all loops and turns were removed and the template residues were determined iteratively from the remaining residues, such that the set of mapped residues used as superpositioning anchors in DaliLite converged and the variance of the backbone root mean square deviation (RMSD) over all superposed structures was low. After the subsets were identified, the following procedure was used to superpose the combined variable domains and to place the cuboids. First, the Vα:Vβ-complexes were superimposed based on their Vα domains using the above defined α-subset and the tool DaliLite. All structures were superimposed to the high resolution (1.4 Å) structure with the PDB ID 2bnu. This step leads to a set of TCR-structures, which is further referred to as Sα and a corresponding set of cuboid templates (CTα), containing cuboids, which were placed around the Vα domains based on the positions of the α-subset residues. Second, the same procedure was used to place cuboid templates (CT, Fig 1C) additionally around each Vβ domain contained in the set Sα according to the relative position of the β chains towards their paired, superposed Vα domains resulting in a set of cuboid templates around the Vβ domains (CTβ). This step results in the set C consisting of the Vβ domain cuboids (CTβ) and the corresponding Vβ domain structures.
We computed the Euler angles for each cuboid-based geometry with respect to a reference coordinate system. The calculation was implemented using the GNU generic math template library; all angles were computed in xyz-order. The reference coordinate system was chosen to be the coordinate system of the 2bnu structure (Fig 1C). Since all structures of the set S were superimposed to conserved framework residues of the reference structure 2bnu, the rotation of the Vβ domain is computed relative to the orientation of the Vβ of 2bnu. The similarity between two geometries we defined as the Euclidean distance of the Euler angles [78]:
dE(i,j)=(Φi−Φj)2+(Ψi−Ψj)2+(Θi−Θj)2
(1)
The distance matrix D was clustered hierarchically, using Ward’s minimum variance method [57,79]. In the case of structures containing multiple BUs each Euler angle component Φ, Ψ, and θ was averaged, leading to an artificial unified geometry. These unified averaged geometries cluster together with other TCRs of the same type, describing a general geometrical state of a TCR. In contrast, the differing geometries found within one structure indicate, that TCRs may adopt an ensemble of different geometries. The significance of the found clusters was confirmed using a bootstrap analysis [58]. The number of bootstrap replica was set to 106.
The common center of rotation (CoR) was computed for all structures (BUs were treated independently), and subsequently the residues situated in the CoR-areas were analyzed. The conservation of these CoR-residues was confirmed by multiple sequence alignment of the TCR.
The CoR was determined for each Vβ domain of the set C. For this purpose, grids were fitted into the cuboids (as illustrated in Fig 1D) and the grid points were indexed in the same manner for each cuboid. For all grid points of the same index i, pairwise distances were computed. The variance
var(gi)=1n2−1∑k=1n∑l=1n(δS(gi,k,gi,l)−∑r=1n∑s=1nδS(gi,r,gi,s)n2)
(2)
was determined; δS(gi,x,gi,y) is the Euclidean spatial distance between a grid point with the index i of the structure x and the equivalent grid point of structure y; n denotes the number of observed structures. We define the CoR grid index point as Imin = argmin(min{var(gi)|1 ≤ i ≤ n}). The corresponding coordinate for Imin was computed according to the reference coordinate system of the reference structure 2bnu. The computations were performed on the whole set C and on subsets of this set. The first subset Cb contained only MHC bound TCRs, whereas the second subset Cu contained unbound TCRs. Furthermore, we investigated the structural environment of the location of the corresponding Imin coordinates. The grids were implemented in a cubic shape with an amount of 33,076,161 grid points and a minimum distance between each point of 0.1 Å, resulting in a grid size of 32 Å in each dimension. In contrast to the intersect method for the CoR calculation, the grid method allows highlighting invariant areas and is more robust against deviations caused by geometrical translation. Sequence similarity and conservation of amino acids located at the CoR was explored by creating multiple sequence alignments (MSAs) of all known human and murine (functional) TCR variable segment sequences. For this purpose, the tool MAFFT [80] (linsi: localpair, maxiter 1000, Blosum62 [81]) was applied on sequences obtained from the IMGT GENE-DB [61].
|
10.1371/journal.ppat.1004421 | Multivalent Adhesion Molecule 7 Clusters Act as Signaling Platform for Host Cellular GTPase Activation and Facilitate Epithelial Barrier Dysfunction | Vibrio parahaemolyticus is an emerging bacterial pathogen which colonizes the gastrointestinal tract and can cause severe enteritis and bacteraemia. During infection, V. parahaemolyticus primarily attaches to the small intestine, where it causes extensive tissue damage and compromises epithelial barrier integrity. We have previously described that Multivalent Adhesion Molecule (MAM) 7 contributes to initial attachment of V. parahaemolyticus to epithelial cells. Here we show that the bacterial adhesin, through multivalent interactions between surface-induced adhesin clusters and phosphatidic acid lipids in the host cell membrane, induces activation of the small GTPase RhoA and actin rearrangements in host cells. In infection studies with V. parahaemolyticus we further demonstrate that adhesin-triggered activation of the ROCK/LIMK signaling axis is sufficient to redistribute tight junction proteins, leading to a loss of epithelial barrier function. Taken together, these findings show an unprecedented mechanism by which an adhesin acts as assembly platform for a host cellular signaling pathway, which ultimately facilitates breaching of the epithelial barrier by a bacterial pathogen.
| Vibrio parahaemolyticus is a bacterial pathogen which occurs in marine and estuarine environments. It is a main cause of gastrointestinal illness following the consumption of raw or undercooked seafood. In immunocompromised people, the bacteria can sometimes enter the bloodstream and cause septicemia, a serious and often fatal condition. V. parahaemolyticus attaches to host tissues using adhesive proteins. Multivalent Adhesion Molecule (MAM) 7 is an adhesin which helps the bacteria to hold onto the host cells early on during infection. It does so by binding two different molecules on the host, a protein (fibronectin) and phospholipids called phosphatidic acids. We show that MAM7 does not only play a role in sticking to host cells. By forming adhesin clusters on the host surface and binding to host lipids, it triggers signaling processes in the host. These include activation of RhoA, an important mediator of cytoskeletal dynamics. By doing so, MAM7 perturbs proteins at cellular junctions, which normally maintain the cells in the gut as a tightly sealed layer protective of environmental influences. When bacteria use MAM7 to attach to the intestine, the seals between cells break, permitting bacteria to cross the barrier and cause infection of underlying tissues.
| Vibrio parahaemolyticus is an emerging food- and waterborne bacterial pathogen. It is predominantly associated with gastroenteritis but occasionally manifests as wound infection [1]–[3]. Although infections are self-limiting in immunocompetent patients, in rare cases, usually occurring in patients with an underlying primary disease, V. parahaemolyticus can rapidly disseminate into the blood stream and cause septicemia, a life-threatening condition [4], [5]. During gastrointestinal disease, the pathogen predominantly colonizes the distal small intestine, where it causes fluid accumulation, extensive tissue damage, a reduction in epithelial barrier function and inflammation [6]. V. parahaemolyticus virulence has so far mainly been attributed to secreted haemolysins (TDH and TRH) as well as a range of effector proteins secreted into the host cell cytoplasm via two type III secretion systems (T3SS1 and T3SS2) [7], [8]. V. parahaemolyticus-mediated cellular toxicity has been attributed to the effects of T3SS1 secreted proteins: VopS, VopQ and VPA0450 contribute to cell rounding, disruption of autophagic turnover and cell lysis, respectively [9]–[11]. T3SS2 has been implicated in intestinal colonization, cellular invasion and enterotoxicity [6], [12]. However, the increase in epithelial permeability seen during infection in vivo as well as in tissue culture models of infection has not been attributed to any particular virulence factor [6], [13]. Recently, we have shown that Multivalent Adhesion Molecule (MAM) 7, a constitutively expressed surface protein, contributes to pathogen attachment to host cells during the early stages of infection [14]. V. parahaemolyticus MAM7 recognizes two host surface receptors: it binds host membrane phosphatidic acid (PA) lipids with high affinity and uses the extracellular matrix protein fibronectin as a co-receptor. MAM7 contains seven mammalian cell entry (mce) domains and each individual domain is capable of binding PA, while a stretch of at least five repeats is required to interact with fibronectin. While PA binding is essential for attachment, binding to fibronectin is dispensable for the interaction but increases the on-rate of binding [15].
Both PA and fibronectin are important signaling molecules in their own right and are implicated in key cellular pathways. PAs make up an average 1–4% of a cell's total phospholipid content [16] and are important as precursors for the biogenesis of other phospholipids, in determining membrane curvature and as signaling molecules [17]–[19]. Several PA-binding proteins are known, including Raf-1, mTOR and SHP-1 [20]–[22]. As such, PAs are involved in the regulation of a diverse set of cellular functions, ranging from metabolism and trafficking to proliferation. Thus far, studies on PAs have focused on pathways involving PA localized in the inner leaflet of the plasma membrane and cellular organelles, such as the ER. Although PA can also be found in the outer leaflet of the plasma membrane, it is not characterized how this pool is generated or how it is linked to cellular functions [23], [24]. It has also been reported that PA generation in cells is localized to specific regions within the membrane, but the consequences of this compartmentalization are not well understood [25].
In this study, we found that the clustering of MAM7 molecules on the bacterial surface and subsequent binding of these clusters to phosphatidic acid lipids in the host membrane, causes downstream activation of the small GTPase RhoA. RhoA activation drives actin rearrangements which ultimately lead to redistribution of tight junction proteins and a disruption of epithelial integrity. This breach in the epithelial barrier allows V. parahaemolyticus to translocate across polarized epithelial layers. Thus, we report for the first time that a bacterial adhesin, through direct interactions with host lipid receptors, induces cellular signaling pathways facilitating epithelial barrier breaching by a bacterial pathogen.
Multivalent Adhesion Molecule (MAM) 7 present on the outer membrane of V. parahaemolyticus mediates attachment of bacteria to host cells [14]. We used V. parahaemolyticus strain CAB4 to study the infection phenotype in Hela cells. CAB4 is derived from the well characterized, pathogenic RIMD2210633 strain [26], but lacks both thermostable hemolysins (ΔtdhA ΔtdhS) and does not express the two type III secretion systems (ΔexsA ΔvtrA). Despite lacking known virulence factors, infection with V. parahaemolyticus CAB4 strain caused pronounced cytoskeletal changes, with thick strands of filamentous actin forming (Fig. 1A). The appearance of F-actin fibers was observed almost immediately upon infection and persisted throughout the course of the experiment (Fig. 1C). In contrast, no changes in the actin phenotype were observed in cells infected with CAB4Δvp1611 lacking MAM7 (Fig. 1B). As such, MAM7 is necessary to trigger the observed actin rearrangements upon infection with V. parahaemolyticus CAB4. Next, we investigated if MAM7 is sufficient to cause actin stress fiber formation in Hela cells. Heterologous surface-expression of V. parahaemolyticus MAM7 in otherwise non-adherent Escherichia coli is sufficient to mediate their attachment to a wide range of host cells [14]. Infection of cells with this recombinant, attaching E. coli strain recapitulated the same sustained actin rearrangements seen upon infection with CAB4 (Fig. 1D, F). In contrast, expression of translocation-deficient MAM7 (MAM7ΔN1–44) in E. coli lead to only low levels of attachment and did not trigger actin rearrangements (Fig. 1E). This demonstrates that V. parahaemolyticus MAM7 is necessary and sufficient to convey upon non-pathogenic bacteria the ability to attach to host cells and trigger actin rearrangements. Next, chemical cross-linking was used to directionally couple purified MAM7 protein to the surface of fluorescent polymer beads, thereby mimicking exposure of the adhesin on the bacterial surface. We used this “bacteriomimetic” system to study the effect of MAM7 on host cells independent of other bacterial molecules. Beads directionally coupled to the N-terminus of a protein containing all seven mammalian cell entry (mce) domains of V. parahaemolyticus MAM7 (GST-MAM7) attach to host cells and trigger sustained actin rearrangements, mimicking the phenotype seen upon infection with CAB4 (Fig. 1G, I). In contrast, beads coupled to GST alone did not significantly bind to host cells and caused no actin rearrangements (Fig. 1H). Beads coupled to protein containing only a single mce domain (MAM1) also failed to be recruited to the host cell surface in high numbers and did not cause changes in cytoskeletal organization (Fig. 2A, B). Free, soluble, uncoupled MAM7 or free GST also did not cause any cytoskeletal reorganization (Fig. 2C–E). The visually observed changes in actin phenotype were also recapitulated using quantitative analysis of cellular G-actin and F-actin contents by fractionation of lysates, Western Blotting and densitometry (Fig. 1J and 2F). We conclude that V. parahaemolyticus MAM7, through multivalent binding of host receptors and when clustered on the host cell surface, causes sustained rearrangements in the actin cytoskeleton, visible as bundles of F-actin.
Actin rearrangements are generally mediated by activation of small GTPases RhoA, Rac and/or Cdc42. We tested the activation levels of all three GTPases by studying the fraction of GTP-bound proteins over time, following binding of MAM7-beads to host cells (Fig. 3). We observed a sustained activation of RhoA, but not Rac or Cdc42, which persisted over several hours in the presence of cell-bound MAM7 beads (Fig. 3A–D). To analyze if actin rearrangements following MAM7 attachment would be dependent on RhoA, Rac or Cdc42, we treated cells with Clostridium difficile toxin B (TcdB) or C. botulinum C3 transferase. TcdB irreversibly deactivates Rho GTPases by glycosylation of the catalytic threonine residue. C3 selectively inactivates RhoA, B and C but not Rac or Cdc42 by ADP-ribosylation of asparagine 41 in the effector region [27]. While untreated cells displayed stress fibers when incubated with fluorescent MAM7 beads, no actin rearrangements where observed in cells pre-treated with either TcdB or C3 transferase (Fig. 3E–H). The observed change in actin phenotype was also confirmed by quantification of cellular G-actin and F-actin (Fig. 3I). We also studied the effect of MAM7 binding on cells overexpressing either dominant negative RhoA, Rac or Cdc42. Expression of RhoAT19N-GFP abolished actin rearrangements, while expression of either RacT17N-GFP or Cdc42T17N-GFP had no effect (Fig. 3J–M). We conclude that binding of multivalent, surface-coupled MAM7 to the host cell membrane specifically activates RhoA, which in turn triggers the observed actin rearrangements.
Several cellular pathways are involved in relaying signaling between activated RhoA and the actin cytoskeleton and the observed actin rearrangements could be a result of either increased stress fiber formation or a decrease in actin depolymerization [28]. We tested if the MAM-induced RhoA activation and ultimately actin rearrangements proceed via the Rho-associated serine/threonine kinase ROCK, a downstream effector of RhoA, by treating cells with the ROCK inhibitor Y-27632 [29]. Cells incubated with control beads showed no perturbation in the actin cytoskeleton, either in the presence or absence of Y-27632 (Fig. 4A, C, E). In contrast, cells incubated with bead-coupled MAM displayed stress fibers but this phenotype was almost completely abolished in Y-27632 treated cells (Fig. 4B, D, E). These findings were recapitulated when we quantified the cellular G-actin and F-actin content under identical experimental conditions (Fig. 4F).
Next, we tested whether MAM-induced ROCK activation takes place upstream or downstream of RhoA. We analyzed RhoA activation in the presence and absence of MAM beads, either on untreated or Y-27632 treated cells. These data show that when ROCK is inhibited, even though MAM-induced stress fiber formation is abolished, RhoA activation levels remain high (Fig. 4G). We thus conclude that MAM-induced ROCK activation occurs downstream of RhoA.
Next, we tested the activation of LIM kinase (LIMK) and cofilin, two key signaling proteins downstream of ROCK. A significant fraction of LIMK was phosphorylated in the presence of MAM-beads, but the p-LIMK level was much reduced if cells were pre-treated with Y-27632 prior to MAM7 bead adhesion. Incubation with either control beads alone or in combination with Y-27632 treatment did not cause significant LIMK phosphorylation (Fig. 4H). LIMK activation causes phosphorylation of cofilin and thus inhibition of its actin depolymerization activity. We observed an increase in p-cofilin in cells with attached MAM-beads, which was abolished by Y-27632 treatment prior to attachment. Incubation with either control beads alone or following Y-27632 treatment did not cause significant changes in p-cofilin levels (Fig. 4I). In addition, treatment of cells with LIMK inhibitor prior to MAM adhesion lead to a loss of the actin phenotype and concurrent loss of increased F-actin contents (Fig. 4F). We conclude that MAM-induced actin rearrangements proceed via the RhoA/ROCK/LIM-K/cofilin pathway and are the result of abrogated actin depolymerization rather than de novo polymerization.
We have previously shown that MAM7 interacts with two types of receptors in the host cell membrane. Each of the seven mce domains within MAM7 is capable of interacting with a phosphatidic acid phospholipid molecule, thereby mediating high affinity binding of bacteria to host cells. Recognition of fibronectin is achieved by a repeat of at least five mce domains and while this interaction is dispensable for attachment, it increases the on-rate of bacterial binding to host cells [15]. We asked if the observed actin rearrangements are a result of MAM binding to fibronectin or phosphatidic acid receptors on host cells, or both. We made MAM attachment to host cells independent of binding to fibronectin by blocking the MAM binding epitope on fibronectin with an antibody [15]. This way, binding of MAM7 to host cells was only mediated by phosphatidic acid receptors. Cells either pre-treated with α-Fn antibodies or non-specific control antibodies were incubated with MAM7 beads or control beads. Following incubation with MAM7 beads, stress fibers were observed in both cells treated with control antibodies (+Fn), or α-Fn antibodies (−Fn), (Fig. 5A–C). In contrast, no actin changes were observed in cells following treatment with either antibody followed by control beads (Fig. 5C). As previously described, uncoupling MAM7 binding from its co-receptor fibronectin did not change the overall number of beads bound per cell if sufficient time was allowed for attachment (Fig. 5D). The interaction between fibronectin and MAM has been mapped to the N-terminal region of fibronectin, which is an epitope commonly exploited by bacterial adhesins for binding [30]. Both Staphylococcus aureus fibronectin binding protein A (FnBPA) and Streptococcus pyogenes protein F1 bind the N-terminal part of fibronectin with high affinity [31], [32]. Thus, we tested whether portions of these two adhesins sharing the same binding epitope with MAM would cause similar actin rearrangements to those observed with MAM7. We incubated cells with beads coupled to the fibronectin-binding region of either FnBPA (FnBR1-11) or F1 (FUD), as previously described [33]. Although both preparations bound to cells with high efficiency, neither caused stress fiber formation (Fig. 5E–H). Taken together, these findings strongly suggest that fibronectin is not involved in the observed signaling pathway between MAM7, RhoA and actin. To see whether changes in the membrane lipid composition would impact MAM's ability to trigger RhoA activation, we treated cells with phospholipase C (PLC). MAM7 beads were added to cells either immediately or up to five hours following PLC treatment and subsequent enzyme removal, and levels of beads per cell as well as RhoA activation were measured. In untreated cells, approximately 23 beads were bound per cell (Fig. 5I, red bar). No bead binding was observed if cells were continuously exposed to PLC, since the interaction with fibronectin alone is insufficient to mediate binding (Fig. 5I, blue bars). If PLC was removed, the interaction between lipid receptors and MAM7, and thus bead binding, was initially completely abolished but was gradually recovered until normal binding levels were regained after four hours of recovery (Fig. 5I, black bars). A similar time course was established for RhoA activation upon MAM bead attachment, with full GTPase activation recovered four hours after removal of PLC (Fig. 5J). We conclude that the MAM7 co-receptor fibronectin is dispensable not just for MAM7 binding, but also for the subsequent activation of RhoA and actin rearrangements caused by adhesion. The observed signaling cascade thus depends on the interaction of multivalent, surface-clustered MAM7 adhesins with phosphatidic acid lipids in the host cell membrane.
Vibrio parahaemolyticus mostly causes gastroenteritis and on rare occasions it can lead to systemic disease in immunocompromised patients. To better reflect the in vivo situation, we studied the effect of MAM on polarized intestinal epithelial (Caco-2) cells. Differentiated Caco-2 monolayers are a good model of the epithelium in the small intestine, the main site of V. parahaemolyticus infection. When grown on permeable supports, Caco-2 cells form monolayers with well differentiated brush border microvilli and properties resembling those of the small intestinal epithelium [34]. First, we studied the localization of MAM7 on polarized cell layers. MAM7 exclusively bound to the apical side of the epithelial layer, with the protein being enriched at cellular junctions (Fig. 6A). No binding was observed when MAM protein was added to the basolateral side (Fig. 6B). Similar to the effects seen in Hela cells, MAM-coupled beads and V. parahaemolyticus CAB4, but not a MAM deletion strain (CAB4ΔMAM), caused a significant increase in RhoA activation compared to untreated cells (Fig. 6C). Because MAM7 was enriched at cell junctions and RhoA activation is capable of affecting the distribution of tight junction proteins, we studied the localization of tight junction markers during infection with V. parahaemolyticus. Apical infection with CAB4 caused a re-distribution of the tight junction markers occludin and zonula occludens protein 1 (ZO-1) (Fig. 6D, G). In contrast, the distribution of both tight junction proteins remained unchanged when cells were infected with CAB4 from the basolateral side (Fig. 6E, H) or apically with the MAM knockout strain CAB4ΔMAM (Fig. 6F, I).
Next, we asked if re-distribution of tight junction proteins during infection would affect epithelial barrier function. When CAB4 was added to the apical surface of a differentiated layer, a marked decrease in transepithelial electrical resistance (TER) was observed approximately three hours post infection. This change was mediated via ROCK/LIMK activation, since treatment of cells with either Y-27632 or LIMK inhibitor abolished the CAB4-mediated decrease in TER (Fig. 6J). Similarly, no significant decrease in TER was observed up to seven hours post infection with either CAB4ΔMAM added apically or CAB4 added to the basolateral side of the epithelium (Fig. 6M).
We also investigated whether the disruption of cell-cell junctions was sufficient to allow for bacterial transmigration. Polarized cells were infected with either CAB4 or CAB4ΔMAM and bacterial titers in the opposing compartment were determined either immediately or up to eight hours post infection. When either CAB4 or CAB4ΔMAM were added to the basolateral side, no bacteria were recovered on the apical side. In contrast, CAB4 was recovered from the basolateral side following infection from the apical side. Bacterial numbers on the basolateral side increased significantly 2.5 hours post infection and continued to increase until 6.5 hours post infection, reaching approximately 1% of the initial infecting population. In epithelial layers apically infected with CAB4ΔMAM, no bacteria were detected on the basolateral side (Fig. 6K). The loss of MAM could be compensated either by the expression of MAM in trans or by treatment of cells with bead-bound MAM, but not with control beads (Fig. 6L). We concluded that MAM selectively binds to the apical side of polarized intestinal epithelial cells, causing a re-distribution of tight junction proteins, disruption of barrier integrity and bacterial transmigration.
Finally, we asked if the epithelial disruption caused by MAM-mediated adhesion would contribute to infection in a virulent strain. Polarized intestinal epithelium was infected with the virulent strain POR1 or POR1ΔMAM from the apical or basolateral side. Infection with POR1 apically lead to cytotoxicity and rapid cell lysis, with almost complete cell death five hours post infection (Fig. 7A). The cytotoxicity profile was significantly delayed upon infection with POR1ΔMAM and cell death reached only approximately 70% even seven hours post infection. When cells were infected with either POR1 or POR1ΔMAM from the basolateral side, no significant increase in cytotoxicity was observed over the course of the experiment (up to seven hours post infection). POR1 contains the T3SS effector VopS, which causes RhoA inhibition by irreversible AMPylation of a threonine residue in the switch I region [9]. Thus, we investigated the contribution of MAM to the overall RhoA activation levels in polarized Caco-2 cells infected with the virulent strain. At 2 hours post infection, prior to the onset of cell lysis, RhoA activity was completely abolished in POR1 infected cells. In contrast, RhoA was highly activated in POR1ΔVopS. An intermediate level of RhoA activation was observed in cells infected with POR1ΔMAM (Fig. 7B). We also analyzed the G-actin and F-actin content of polarized Caco-2 cells 2 hours post infection. Within 2 hours, POR1 infection lead to a drop in F-actin content compared to untreated cells, which was mediated by the activity of VopS. In the absence of MAM, or in the presence of ROCK- or LIMK inhibitors, the F-actin content was higher compared to POR1 infected cells (Fig. 7C). Finally, we measured the transepithelial resistance in Caco-2 monolayers infected with the virulent strain. POR1 caused a rapid decrease of TER, which was markedly slowed by treatment of cells with Y-27632 or LIMK inhibitor. Similarly, both POR1ΔMAM and POR1ΔVopS showed a slight delay in depolarization (Fig. 7D).
Previously, we reported that V. parahaemolyticus Multivalent Adhesion Molecule (MAM) 7 and several of its homologs from other Gram-negative enteric pathogens mediate initial attachment of bacteria to host cells [14]. In this study, we demonstrated that clusters of multivalent MAM molecules, by binding to the host cell membrane, facilitate activation of the host small GTPase RhoA, which in turn leads to actin rearrangements. Clustering of MAMs is achieved by nature, through display of multiple adhesion molecules on the bacterial outer membrane [14], but can be mimicked by coupling recombinant MAM molecules to a polymer bead with roughly the same dimensions as a bacterium. Soluble MAM failed to achieve the same effect on host cell signaling.
MAMs interact with host cells via two cellular receptors, the protein fibronectin and the phosphatidic acid (PA) phospholipids. While the former is a well-characterized pathogen receptor [30], [35], [36], direct binding of a bacterial adhesin to a host cell lipid is a new paradigm of host-pathogen interaction. Over recent years, manipulation of cellular lipids by pathogens has been an emerging field of study, and it has become evident that host cellular lipids are often a primary target of bacterial virulence factors [11], [37], [38]. Herein, we showed that MAM's impact on RhoA activation is mediated through its interaction with phosphatidic acid lipids in the host membrane and that its co-receptor fibronectin is dispensable for its function as a signaling effector. Taken together, these findings suggest a mechanism whereby the interaction of clustered MAM adhesins with host membrane lipids causes rearrangements of the latter and that this acts as a signal leading to RhoA activation. However, direct observation of such hypothesized rearrangements of phosphatidic acid molecules within the host membrane on the nanoscale is not within the scope of our studies but is an intriguing possibility and something we are currently investigating.
We have elucidated the signaling pathway downstream of RhoA and show the MAM-triggered signal is relayed from activated RhoA, via the Rho-associated serine/threonine kinase ROCK and LIM kinase, to result in phosphorylation of cofilin. Cofilin is an actin-binding protein which mediates actin depolymerization [39]. Its interaction with actin and thus its depolymerization activity is disrupted by phosphorylation, resulting in a net stabilizing effect on filamentous actin and apparent increase in actin stress fibers. Although a large part of our experiments was performed on Hela cells because changes in the actin phenotype following serum starvation are visually easier discernible in this cell type, we show that the MAM-mediated effect on actin also proceeds via ROCK and LIMK activation in polarized intestinal epithelial cells, a more relevant system for studies on V. parahaemolyticus. Since we observe MAM-induced RhoA activation also in polarized epithelial cells, we hypothesize that this RhoA activation facilitates subsequent activation of the ROCK/LIMK/cofilin signaling axis, however we cannot show whether RhoA activation is required in this model, since RhoA inactivation itself leads to increased transepithelial permeability [40].
In the polarized epithelial system, MAM7 selectively attached to the apical side of the layer and attachment caused a marked redistribution of tight junction proteins. A similar phenotype has been described to occur following infection of epithelial cells with other pathogens, such as enteropathogenic E. coli (EPEC) or the protozoan parasite Giardia lamblia. With EPEC infection, paracellular permeability also resulted from a redistribution of tight junction proteins upon RhoA activation, although in that case RhoA activation has been largely attributed to the activities of type III system-secreted effectors [41], [42]. In G. lamblia, barrier failure was attributed to apoptosis of enterocyes [43]. Activation of RhoA through the establishment of a signaling complex consisting of bacterial adhesin clusters and host membrane lipids on the host cell surface is, to our knowledge, a previously unrecognized strategy to achieve epithelial barrier disruption.
We demonstrated that the action of MAM7 causes epithelial barrier disruption, as evidenced both by a decrease in transepithelial resistance and the ability of bacteria to transmigrate to the basolateral side of the epithelium. It has previously been shown that CAB4 is unable to invade epithelial cells [12], so this is likely the result of bacteria moving through compromised cell-cell junctions. It has been shown previously that epithelial integrity is compromised following V. parahaemolyticus infection, both in cultured polarized epithelial cells and in vivo. Animal infection models have shown increased transepithelial permeability using fluorescent dextran as a tracer, but the effect was not attributed to any particular virulence factor [6]. Earlier experiments on polarized Caco-2 cells demonstrated a similar effect on epithelial integrity and ruled out TDH and TRH toxins as the culprit [13]. A comparison between V. parahaemolyticus clinical isolates and environmental strains implicated T3SS2 in transepithelial permeability. However, no whole genome sequences are available for the strains used in this study and we therefore do not know if they encode for a MAM homolog and if so, to what extent it would share sequence similarity to RIMD2210633 MAM7 (vp1611) [44]. More recent studies on Caco-2 and mixed M cell-like co-cultures demonstrated that T3SS1 does not significantly contribute to translocation, while T3SS2 is dispensable but has a moderately enhancing effect on translocation in a RIMD2210633 background [45]. Herein we show that MAM7 is sufficient to cause barrier disruption in cultured polarized epithelium. In the context of a T3SS-competent, virulent strain, MAM induces transepithelial permeability and depolarization of the epithelium early during infection. Since MAM is constitutively expressed and present at the early stages of infection, its effect takes hold almost immediately and RhoA activation is detectable as early as 30 minutes post infection (the earliest time point measured here). The resulting depolarization and disruption of cell-cell junctions leads to an increase in host cell surface available for translocation of type III secreted bacterial effectors. Overall, this mechanism accelerates effector-mediated functional changes in host cells, such as VopS-mediated irreversible RhoA inactivation and concomitant actin depolymerization, thus speeding up infection. These findings strongly indicate experiments comparing the effect of wild type and MAM knockout strains in an animal model and this should be the next step to show if indeed MAM contributes to transepithelial permeability and infection in vivo.
Overall, the study we present here demonstrated that the contribution of Vibrio parahaemolyticus MAM7 to the pathogen's infection profile is not limited to its function in early bacterial attachment. By establishing signaling complexes consisting of clustered MAM adhesins and host membrane lipid receptors on the host cell surface, it additionally acts as an effector of host cellular GTPase signaling and its action culminates in breaching of the epithelial barrier. This is, to our knowledge, a previously unrecognized strategy by which a bacterial pathogen disrupts intestinal epithelial function and the detailed molecular mechanism of how this is achieved certainly deserves our further investigation.
The construction of BL21-MAM7, BL21-MAMΔN1–44, CAB4, POR1, POR1ΔMAM (POR1Δvp1611) and POR1ΔVopS has been described elsewhere [9], [12], [14]. The V. parahaemolyticus MAM deletion strain CAB4Δvp1611 was constructed using the same method and same vector construct (pDM4 containing regions 1 kb up- and downstream of vp1611) described in these references. Strains were grown on MLB (V. parahaemolyticus) or LB agar (E. coli), with 100 µg/ml of kanamycin or ampicillin added for selection where required.
HeLa and Caco-2 epithelial cell lines were cultured at 37°C and under 5% CO2 in Dulbecco's Modified Eagle Medium (DMEM) containing 10% heat-inactivated fetal bovine serum, 4500 mg/L glucose, 0.5 mM L-glutamine, 100 units/ml penicillin and 20 µg/ml streptomycin. For GTPase activation and microscopy assays, cells were serum-starved for 40 hours prior to treatment. For infection experiments, DMEM with no added antibiotics was used. For experiments on polarized Caco-2 cells, cells were seeded on polycarbonate 3.0 µm pore size transwell filters (Costar) at 200000 cells/ml. Cells reached confluency after approximately 5–6 days, at which point several transepithelial resistance (TER) measurements were taken to check the integrity of the layer and establish baseline measurements. TER measurements before and during infection experiments were taken with a Millicell-ERS resistance apparatus (Millipore).
Expression and purification procedures for recombinant proteins have been described in detail elsewhere (see [14] for GST-MAM7 [15], for GST-mce1 and [33] for GST-FnBPA FnBR1-11 and F1 FUD constructs). Purified proteins were immobilized on amine modified fluorescent blue polystyrene beads with a mean diameter of 2 µm (Sigma) using Sulfo-SMPB (sulfosuccinimidyl 4-[p-maleimidophenyl]butyrate) cross-linking under reducing conditions, as outlined in the manufacturer's protocol (Pierce). Bead-coupled proteins were added to experiments to give a final concentration of 500 nM immobilized protein and a surface density of 1.5×105 molecules per bead (giving a spacing of approximately 57 nm).
Tissue culture cells were washed with PBS (phosphate-buffered saline) prior to the addition of bacteria in tissue culture medium without antibiotics. Bacteria were added to give a multiplicity of infection (MOI) of 100, except for POR1 and derivatives, where an MOI of 10 was used. Plates were centrifuged (1000×g, 22°C, 5 minutes) prior to incubation at 37°C for 30 minutes to eight hours, depending on the experiment. To uncouple MAM binding from fibronectin or phosphatidic acid, respectively, cultured cells were incubated with anti-Fn antibody (Sigma, 50 µg/ml in PBS) or treated with 50 µg/ml phospholipase C (Sigma) in PBS for 15 min prior to infection, as previously described [15]. For enumeration of bacteria, samples were removed at time points as indicated and were serially diluted, plated on agar plates, incubated at 37°C for sixteen hours and CFU counts determined the following day. For cytoxicity assays, 200 µl of culture supernatant was removed in triplicate from each well at timepoints as indicated, centrifuged (1000×g, 22°C, 5 minutes), and 100 µl of the supernatant transferred to a fresh 96-well plate for assays. To quantitate cell lysis, the amount of lactate dehydrogenase (LDH) released into the culture medium was determined using the LDH cytotoxicity detection kit (Takara) according to the manufacturer's instructions. Results are presented in % lysis, relative to negative (uninfected) and positive (Triton X-100 lysed cells) controls.
Cells were transfected with pcDNA3 containing either EGFP, EGFP-RhoAT19N, EGFP-RacAT17N or EGFP-Cdc42T17N using Fugene HD (Roche) transfection reagent according to the manufacturer's protocol. For microscopy, cells were fixed with 3.2% formaldehyde, permeabilized with 0.1% Triton X-100 and stained with rhodamine-phalloidin to visualize F-actin and SYTO-13 to visualize DNA as indicated. For immunofluorescence microscopy, we used α-GST, α-occludin and α-ZO-1 antibodies (Sigma) diluted 1∶500, followed by FITC-labeled α-rabbit antibody (Sigma) at a 1∶1000 dilution. Images were taken either on a Zeiss LSM 510 scanning confocal microscope or a Nikon Eclipse Ti fluorescence microscope and images were prepared using ImageJ and Corel Draw X5. For quantification of the F-actin phenotype, the total number of cells as well as number of cells containing stress fibers, were enumerated. Some fields contained cells displaying cortical actin, however this phenotype was observed across experiments and was independent of MAM adhesion. Thus, these cells were not counted as positive. Data shown are means ± standard deviation from twelve images (four frames from triplicate experiments, representing at least 100 cells/experimental condition).
Proteins were separated by SDS-PAGE and transferred onto nitrocellulose membrane. Membranes were blocked with 5% BSA in TBS-T (Tris-buffered saline containing 0.05% Tween 20) for 1 hour at 22°C. Membranes were probed with primary antibodies (against LIMK, p-LIMK, cofilin, or p-cofilin, all Santa Cruz Biotechnology) diluted 1∶1000 into blocking buffer for 1 hour at 22°C. After three washes with TBS-T, membranes were incubated with anti-mouse HRP (horseradish peroxidase)-conjugated secondary antibody (GE Healthcare) diluted 1∶5000 into blocking buffer for 1 hour at 22°C. Membranes were washed three more times with TBS-T and proteins were detected using the ECL plus detection system (GE Healthcare) and a Gel Doc XR imager. Bio Rad Quantity One software was used for densitometry.
Ratios of globular (G-actin) to filamentous (F-actin) in cultured, serum-starved cells were determined using the G-actin/F-actin In Vivo Assay Kit (Cytoskeleton Inc.) as described in the manufacturer's protocol. Serum-starved, untreated cells (negative control) and cells treated with F-actin enhancing solution (positive control) were analyzed alongside experimental samples (MAM-treated and controls, as described in the figure legends). G-actin and F-actin levels were determined by Western Blotting and were quantified by densitometry. Results shown are means ± s.e.m. from two independent experiments.
Following infection or incubation with beads, cells were washed and collected by scraping into GTPase lysis buffer (20 mM Tris HCl pH 7.5, 10 mM MgCl2, 150 mM NaCl, 1% Triton X-100. Lysates were homogenized and cleared by centrifugation (13000 rpm, 20 min). 500 µg of cleared lysates were added to 30 µg of GST-PAK PBD bound to glutathione agarose beads and incubated for 1 hour at 4°C. Samples were separated by SDS-PAGE and immunoblotted with α-Cdc42 or α-Rac antibodies (Sigma) and compared to total GTPase levels detected in cell lysates. Activated RhoA was pulled down with the use of a RhoA activation kit (Cytoskeleton) according to the manufacturer's instructions. Total and GTP-bound RhoA was detected following SDS-PAGE separation and Western Blotting using α-RhoA antibody (Sigma).
To study cellular phenotypes independent of GTPase activation, cells were treated with either Clostridium difficile toxin B (TcdB) or C3 transferase to irreversibly inactivate either RhoA, Rac and Cdc42 or RhoA, respectively. Cells were treated wither with 200 ng/ml TcdB (List Biologicals) or 1 µg/ml cell-permeable C3 (Cytoskeleton) for 4 hours. Attachment experiments were carried out immediately after toxin treatment.
|
10.1371/journal.pntd.0006943 | Strongyloides stercoralis: Spatial distribution of a highly prevalent and ubiquitous soil-transmitted helminth in Cambodia | Strongyloides stercoralis is a neglected soil-transmitted helminth that occurs worldwide, though it is particularly endemic in tropical and subtropical areas. It can cause long-lasting and potentially fatal infections due to its ability to replicate within its host. S. stercoralis causes gastrointestinal and dermatological morbidity. The objective of this study was to assess the S. stercoralis infection risk and, using geostatistical models, to predict its geographical distribution in Cambodia.
A nation-wide, community-based parasitological survey was conducted among the Cambodian population, aged 6 years and older. S. stercoralis was diagnosed using a serological diagnostic test that detects IgG antibodies in urine. Data on demography, hygiene and knowledge about helminth infection were collected. S. stercoralis prevalence among 7,246 participants with a complete data record was 30.5%, ranging from 10.9% to 48.2% across provinces. The parasite was ubiquitous in Cambodia; only five south-eastern provinces had prevalence rates below 20%. Infection risk increased with age for both men and women, although girls under the age of 13 and women aged 50 years and over had lower odds of infection than their male counterparts. Open defecation was associated with higher odds of infection, while having some knowledge of the health problems caused by worms was a protective factor. Infection risk was positively associated with nighttime maximum temperature, minimum rainfall, and distance to water; it was negatively associated with land occupied by rice fields.
S. stercoralis infection is rampant in Cambodia. Control programs delivering ivermectin are needed to manage the parasite. However, the high cost of this drug in Cambodia currently precludes the implementation of control initiatives. Donations, subsidies or affordable generics are needed so that S. stercoralis, which infects almost a third of the Cambodian population, can be addressed through an adequate control program.
| The threadworm, Strongyloides stercoralis, is a highly neglected worm infection, transmitted through infective larvae in the soil. Threadworms occur worldwide, particularly in tropical climates. It may cause long-lasting and potentially fatal infections due to the parasite’s ability to replicate within its host. This study aimed to assess the risk of threadworm infection at national level in Cambodia. We conducted a nation-wide, community-based parasitological survey of the Cambodian population, aged 6 years and over. The threadworm was diagnosed using a serological diagnostic test that detects antibodies in urine. Data on demography, hygiene and knowledge about helminth infection were collected. The purpose of this study was to predict the risk of S. stercoralis infection in unsurveyed locations, assess risk factors for infection, and map its geographical distribution in Cambodia. About one third (30.5%) of the enrolled study participants (n = 7,246) were infected with threadworms. At provincial level, the lowest and highest infection rates were 10.9% and 48.2%, respectively. Prevalence rates below 20% were found in just five south-eastern provinces. The risk of a threadworm infection increased with age for both men and women. Open defecation was associated with higher risk of infection, while having some knowledge of the health problems caused by worms was a protective factor. Infection risk was positively associated with environmental factors, such as nighttime maximum temperature, minimum rainfall, and distance to water; it was negatively associated with land occupied by rice fields. Threadworm infection is highly prevalent in Cambodia and adequate control measures, including access to treatment, are warranted to address the burden of this Neglected Tropical Disease (NTD) in Cambodia.
| Strongyloides stercoralis is a highly neglected intestinal nematode, for which larvae living in soil polluted with feces infect humans transcutaneously, like hookworms. S. stercoralis occurs worldwide but thrives in warm regions with poor sanitation conditions and has been under-detected and overlooked for decades because its larvae are not uncovered by standard field diagnostic techniques [1–5]. Until recently, the only available prevalence estimates originated from a review conducted in the late 80s, which estimated some 30–100 million cases worldwide [6]. More recent estimates show prevalence rates between 10% and 40% in subtropical and tropical countries [1]. Using diagnostic approaches suitable for detecting S. stercoralis, some studies indicate that the prevalence of S. stercoralis could be half that of hookworm, i.e. 200–370 million cases worldwide [1, 7, 8].
In Cambodia, two community-based, large-scale surveys documented S. stercoralis prevalence rates of 25% and 45% in the southern province of Takeo and in the northern province of Preah Vihear, respectively [9, 10]. S. stercoralis infection is more prevalent among adults due to its unique ability among soil-transmitted helminths (STHs) to replicate within the host, which leads to infections that can last for decades in the absence of treatment [11]. In cases of immunosuppression, this auto-infection cycle accelerates and results in hyperinfection, a condition that is 100% fatal if left untreated [12–14]. Chronic infection with S. stercoralis may cause abdominal pain, nausea, vomiting, and diarrhea, as well as urticaria and larva currens [15–17]. The latter is a serpiginous intermittent moving eruption due to parasite migration under the skin. Its location on the buttocks, thighs, and trunk, together with the high speed of migration (i.e. 5 to 10 centimeters an hour), makes it a symptom highly specific to strongyloidiasis [11, 13]. Finally, although this aspect of infection needs to be confirmed, S. stercoralis infection might be associated with growth retardation in children [17]. Due to this combination of significant morbidity and high prevalence, S. stercoralis has been recognized as a public health problem in Cambodia. However, the national prevalence and the location of high-risk zones are unknown.
One of the most sensitive diagnostic approaches combines the Baermann and Koga agar plate culture techniques, but this method is costly, time consuming and requires laboratory staff specifically trained to identify S. stercoralis larvae by microscopy [10, 18, 19]. Serological diagnosis is more sensitive than most coprological approaches, but its use may be limited in endemic settings due to cross-reactions with other helminths species [20, 21]. Likewise, serology may overestimate prevalence in endemic areas as it detects parasite-specific antibodies that remain long after contact with the parasite or cure, and cannot distinguish current from past infections [20]. While this last aspect would be an issue for cure assessment, it does not affect prevalence estimates in a population naïve to treatment against the investigated parasite. A serological test was recently developed in Thailand, using an antigen from S. ratti to detect antibodies in urine [22, 23]. This technique has several strengths. First, collecting urine samples is much easier than collecting fecal samples. Second, this test has a high sensitivity of 93% when compared with coprological methods. Lastly, there is little cross-reactivity with other STH species or food-borne trematodes, including Opisthorchis viverrini [22, 23]. However, like other serological tests, it does not differentiate between active and past infections.
In the past decade, geostatistical models have increasingly been used to delineate risk zones for helminthic infections, at small and large scale, and to help target control efforts in areas with the highest need [24–31]. Based on the association between environmental variables and infection levels at survey locations, such models can be used to predict infection levels throughout entire geographical zones.
A national parasitological survey was conducted in 2016, in all provinces of Cambodia, to assess S. stercoralis prevalence based on a serological diagnostic test using S. ratti antigens [22]. Using these data, this work set out to estimate S. stercoralis prevalence in Cambodia, to assess risk factors for infection, and to predict S. stercoralis infection risk throughout the country to help guide control efforts.
The study was approved by the National Ethics Committee for Health Research, Ministry of Health, Cambodia (NECHR, reference number 188, dated 02.05.2016). Prior to enrolment, all participants received an explanation of the study goals and procedures. All participants aged 16 years and over provided written informed consent, while parents or legal guardians provided consent for participants aged 6–15 years. All S. stercoralis cases were treated with a single oral dose of ivermectin (200μg/kg BW) and all other diagnosed parasitic infections were treated according to the national guidelines [32].
Cambodia counted 15.6 million inhabitants in 2015, 79.3% of whom lived in rural areas [33]. The country has undergone rapid economic development in recent decades. With a Human Development Index ranking of 143/188 in 2016, Cambodia belonged to the group of lower middle-income countries, as per the World Bank classification [33, 34]. Poverty levels have decreased dramatically in recent years, with the proportion of the population living in extreme poverty falling to 2.2% in 2016. However, approximately one person in five (21.6%) lives on less than USD 3.1/day [33]. Adult literacy and net enrolment in primary school were 74% and 95%, respectively, in 2010–2014, while 32% of children under the age of 59 months were stunted in 2015 [34]. In 2015, 42% and 69% of the rural population had access to improved sanitation facilities and improved water sources, respectively, while those figures were 88% and 100% for the urban population, respectively [34].
A cross-sectional, community-based survey was conducted among the general population in all 25 provinces of Cambodia, between May and August 2016. In each province, 10 villages were selected from all villages using a simple random sampling procedure in STATA version 13.0 (StataCorp LP; College Station, United States of America). In each village selected, seven to eight households were randomly selected based on the list of households obtained from the village chief. Eighteen villages were originally selected and subsequently replaced because their remote locations compromised the quality of collected samples for parasitological data. In each village, households were selected using systematic proportional sampling; all household members present on the survey day were enrolled up to a maximum of 35 participants per village. All household members aged 6 years and over were eligible.
Participants were asked to provide a urine sample, from which S. stercoralis was diagnosed using an enzyme-linked immunosorbent assay (ELISA) based on S. ratti antigens [22]. After collection, urine specimens were preserved in NaN3 with a final concentration of 0.1%, and kept at 4°C until required for analysis. Samples were sent to the central laboratory of the National Centre for Parasitology, Entomology and Malaria Control (CNM) in Phnom Penh and then to Khon Kaen University, Thailand, for ELISA testing.
An individual questionnaire was administered to all study participants and covered demographics (age, sex, level of education, main occupation), the number of household members, access to sanitation (latrine availability at home, usual place of defecation) and knowledge of worm infections (transmission route of and health problems caused by helminths) (S1 Appendix, English version of questionnaire).
Environmental parameters were extracted from freely available remote sensing (RS) sources for the period September 2015 to August 2016, corresponding to one year prior to the last month of the study. Daytime and nighttime land surface temperature (LST), international geosphere biosphere programme (IGBP) type 1 land use/land cover (LULC), as well as normalized difference vegetation index (NDVI), and enhanced vegetation index (EVI) were extracted at 1 x 1 km resolution from Moderate Resolution Imaging Spectroradiometer (MODIS) Land Processes Distributed Active Archive Center (LP DAAC), U.S. Geological Survey (USGS) Earth Resources Observation and Science (EROS) Center (http://lpdaac.usgs.gov). Rainfall data was obtained from WorldClim (www.worldclim.org). Digital elevation data were retrieved from the NASA Shuttle Radar Topographic Mission (SRTM) and CGIAR-CSI database. Distance to large bodies of water was obtained from Health Mapper.
Laboratory and questionnaire data were double-entered and validated in EpiData version 3.1 (EpiData Association; Odense, Denmark). Environmental data processing, geo-referencing and maps were done in ArcGIS version 10.2.1 (ESRI; Redlands, CA, United States). LULC 18 classes were merged into four categories, according to similarity and respective frequencies. Annual and seasonal means, as well as maxima and minima of monthly EVI, LST and RFE means were calculated and standardized. Environmental data were linked to parasitological and questionnaire data, according to geo-referenced locations. Data management and non-Bayesian data analysis were done in STATA version 13.0. Bayesian geostatistical models were fitted using WinBUGS version 1.4.3 (Imperial College & Medical Research Council; London, UK). Predictions for unsurveyed locations were performed in Fortran 95 (Compaq Visual Fortran Professional version 6.6.0, Compaq Computer Corporation; Houston, United States of America). Five age groups were established as follows: (i) 6–12 years, (ii) 13–18 years, (iii) 19–30 years, (iv) 31–50 years, and (v) >50 years.
Chi-square (χ2) test was used to compare proportions. The association between infection risk and covariates was assessed using mixed non spatial bivariate logistic regressions, accounting for village clustering, i.e. with a non-spatial, village-level random effect. Covariates exhibiting an association at a significance level of at least 15%, as determined by the likelihood ratio test (LRT), were included in the multivariate logistic regression models. In the event of correlated variables, the variable resulting in the model with the smallest Akaike’s information criterion (AIC) was selected. For the risk factor analysis, variables exhibiting high Wald p-values were removed one by one and kept outside of the model if their removal resulted in a lower AIC. Summary measures of continuous environmental variables (i.e. LST day and night, rainfall, and distance to water) were standardized before inclusion in the multiple regression models. To explore the relationship between S. stercoralis infection risk and age, smoothed age-prevalence curves were produced with the “mkspline” command in STATA that regresses each outcome against a new age variable containing a restricted cubic spline of age.
For geostatistical models, a stationary isotropic process was assumed, with village-specific random effects following a normal distribution with mean zero, and a variance-covariance matrix that is an exponential function of the distance between pairs of locations. Vague prior distributions were chosen for all parameters. Further information on model specification is available in S2 Appendix. Markov chain Monte Carlo (MCMC) simulation was used to estimate model parameters [35]. Geostatistical models were run using the WinBUGS “spatial.unipred” function [36]. Convergence was assessed by examining the ergodic averages of selected parameters. For all models, a burn-in of 5,000 was followed by 30,000 iterations, after which convergence was reached. Results were withdrawn for the last 10,000 iterations of each chain, with a thinning of 10. Model fit was appraised with the Deviance Information Criterion (DIC). A lower DIC indicates a better model [37].
Three types of Bayesian mixed logistic models were run. First, models without covariates but using alternatively a geostatistical or an exchangeable random effect were run to quantify the extent of village-level spatial correlation and unexplained variance of S. stercoralis prevalence. Second, a risk factor analysis model was used to assess individual-level demographic, sanitation, and knowledge risk factors, as well as environmental covariates associated with infection risk. Third, a model including only environmental covariates was used to predict infection risk at unsurveyed locations.
To validate the model, 199 (80%) randomly selected villages were used for fitting, and the remaining 50 (20%) were used as test locations. A pair of models containing the same covariates, but including alternately a non-spatial (exchangeable) or spatial (geostatistical) random effect, was run. The predictive ability of the model was assessed by comparing the Mean Squared Error (MSE), which is obtained by squaring the average of absolute differences between predicted and observed prevalence rates at test locations.
Using the model with the best predictive ability, S. stercoralis infection risk was predicted at 68,410 pixels of 2x2 km resolution, using Bayesian Kriging [38].
Among the 8,661 participants enrolled in the study, 1,407 did not provide any urine (one entire village was excluded due to all 34 of its participants not providing urine), 338 were removed because they did not provide a stool sample (requested for other assessments not presented in this work), and eight participants did not provide questionnaire data. Overall, 7,246 participants living in 2,585 households and 249 villages were included in the analysis. The mean number of participants per village was 30.2, with an interquartile range of six, and a minimum of five. With the exception of Ou Tracheak Chet in Preah Sihanouk Province (five participants) and Kampong Chrey in Preah Vihear province (nine participants), all villages had more than 10 participants and 93.6% of villages had 20 participants or more. Table 1 shows the characteristics of participants with complete parasitological and questionnaire data.
Females (57.5%) were overrepresented in the sample, compared to their proportion in the Cambodian population (51.5%) as assessed by the 2013 inter-census population survey [39]. The age distribution of the sample was very similar to that of the total Cambodian population: children and adolescents aged 14 years and younger represented 29.95% and 29.4% of the sample and of the Cambodian population, respectively; adolescents and adults aged 15 to 64 years represented 65.6% and 64.2%; and elderly adults aged 65 and older represented 5.8% and 5.0%.
The proportion of males and females were similar in the groups excluded from and included in the analysis; children and young adults aged 6–30 years were less represented (53.0%) in the sample than among the excluded participants (64.3%). Similarly, farmers were overrepresented (53.6% of the sample vs. 41.1% of excluded participants), while scholars were underrepresented (34.3% of the sample vs. 51.6% of excluded participants) in the final sample. In terms of usual place of defecation, there was no difference between participants excluded from or included in the analysis.
Overall, S. stercoralis prevalence was 30.7% (95% confidence interval (CI): 29.7–31.8), ranging from 10.9% (95%CI: 7.4–14–4) in Prey Veng province, to 48.2% (95%CI: 42.2–54.1) in Koh Kong province. Fig 1 shows the provinces of Cambodia and Fig 2 displays provincial level prevalence rates. Prevalence was highly variable at village level. The smallest prevalence rate of 2.9% (95% CI: 0.1–14.9) was found in a village in Kandal province, where only 1 of 35 participants was infected. The highest prevalence rates were 88.9% (95%CI: 51.8–99.7) and 80% (95%CI: 63.1–91.6), observed in villages in Preah Vihear and Koh Kong province, respectively. There were, however, only nine participants in the village in Preah Vihear province, resulting in a very large confidence interval. The map in Fig 3 shows the observed S. stercoralis prevalence in each village surveyed.
Table 2 presents the model parameters of three geostatistical models, i.e. (i) model without covariates, (ii) the predictive model including only environmental variables, and (iii) the risk factor analysis model including environmental, demographic and behavioral covariates. In the absence of explanatory variables, S. stercoralis risk clustered at a distance of 85 km (range). The range dropped to 3.2 km after introducing environmental variables (predictive model).
The predictive ability of the geostatistical model (MSE = 182.9, DIC = 6894.3) including environmental covariates (predictive model) was slightly higher than that of its non-spatial counterpart (MSE = 187.7, DIC = 6894.4). Therefore, the geostatistical model was used to predict S. stercoralis risk at unsurveyed locations. The geographical distributions of the covariates used in the geostatistical predictive model, together with elevation (which was not included in the predictive model), are shown in S1 Fig. The predictive model is presented in Table 3.
The results of the non-spatial bivariate mixed regressions are presented in S1 Table. Variables that were not significant in the multivariate model and whose removal decreased the model AIC were removed from the multivariate risk factor model during the model building process. The results of the multivariate Bayesian geostatistical risk factor analysis are presented in Table 4.
Sex was an effect modifier of age. Infection risk increased with age for both sexes, but women aged 50 years and older had a lower risk of infection than males. The relationship between S. stercoralis infection risk and age is presented in Fig 4. Participants who practiced open defecation (31.5% of participants defecated either in forests, rice fields or water) had higher odds of infection, while individuals who had some knowledge about the health problems resulting from worm infection had lower odds of harboring S. stercoralis. Regarding environmental factors, S. stercoralis infection risk was positively associated with increasing nighttime land surface temperature (LST night) dry season maximum, increasing minimum annual rainfall, and increasing distance to water. Finally, the odds of S. stercoralis infection were lower among participants living in villages located in croplands (rice fields).
Figs 5 and 6 display the predicted median S. stercoralis prevalence in Cambodia and the lower and upper estimates of the predictions, respectively. Prevalence was consistently higher than 10%, except in a small area of Prey Veng province. S. stercoralis predicted risk was below 20% in just five provinces, namely, Kampong Cham, Tboung Khmum, Prey Veng, Kandal and Svay Rieng. Predicted prevalence was particularly high in the north of Preah Vihear and Stung Treng provinces, near the Lao border, as well as in the south, in parts of Kampong Speu, Koh Kong, Preah Sihanouk, and Kampot provinces.
We present the first (to our knowledge) national prevalence estimates and nation-wide infection risk map of S. stercoralis in Cambodia, where the infection is ubiquitous. Based on a sample encompassing all 25 provinces and including more than 7,200 participants, we found that prevalence rates of S. stercoralis in Cambodia are systematically higher than 10%, with a national prevalence rate of 30%.
The risk of infection was lowest in the southeast of the country, namely the provinces of Prey Veng, Kandal, and Kampong Cham, as well as in the western and southern parts of Tboung Khmum and Kampong Thom provinces, respectively. The highest provincial-level prevalence rates, above 40%, were found in Preah Vihear in the north, Kampong Chhnang in the centre and in Koh Kong and Kampong Speu in the south.
The size of S. stercoralis infection clusters was relatively small at 85 km, similar to that observed for hookworm infection risk in the country [27]. Almost all spatial correlation of S. stercoralis infection was explained by its association with environmental factors (as indicated by the dramatic drop of the range to 3.2 km, after introducing environmental covariates into the model). This result is not surprising as, in absence of treatment, the distribution of the parasite would mostly be conditioned by its biological requirements. The predicted geographical distribution of S. stercoralis risk in this study was similar to that of hookworm prevalence among school-aged children in Cambodia, as predicted by Karagiannis-Voules and colleagues, and likely due to the two nematodes’ similar transmission routes [27]. Yet, hookworm prevalence was lower over a larger area, most likely because of the impact of ongoing STH deworming programs [27].
The odds of infection increased with increasing maximum nighttime temperature and increasing minimum rainfall. S. stercoralis larvae might have the same ability as hookworm larvae to migrate into the soil, which, in the presence of sufficient humidity, confers to the parasite a tolerance for higher temperatures [40]. The positive association between temperature and risk was more surprising, although this might relate to a particularity of S. stercoralis’ life cycle. The number of females and infective larvae developing in the external environment depends on temperature, with numbers of infective larvae reaching a maximum when temperatures are 30°C and higher [11]. Hence, nighttime maximum temperatures, which range between 24°C and 32°C in Cambodia, might affect the quantity of infective larvae present in the environment.
Regarding the environmental predictors of S. stercoralis infection, neither distance to water nor the land cover category of cropland were significantly associated with infection risk in the predictive model, but they became significant in the risk factor analysis after adjusting for demographic and behavioral factors. We found a positive association between S. stercoralis infection risk and distance to water. The development and survival of S. stercoralis larvae is affected by immersion, so seasonal flooding might determine their survival in areas close to water bodies [41, 42]. Similarly, the relationship between larvae survival and water might explain the lower infection rates in areas occupied by croplands, which mostly correspond to rice fields that are regularly flooded. Yet it is also possible that distance to water captured other unmeasured features related to socio-economic factors and human activity [29]. In Cambodia, people have a clear preference for pour-flush latrines and would choose a pit latrine over a toilet, but pour-flush latrines only function with water [43]. Limited availability of water due to living farther away from permanent water bodies might result in decreased access to, or use of, sanitation facilities.
Studies that investigated risk factors for S. stercoralis infection mostly report a higher risk among men [9, 10]. This association is generally attributed to men’s extensive exposure to soil during farming activities, although the findings of our study do not support this assumption. First, in this national sample, infection risk was not associated with occupation and two-thirds of all farmers were women. Second, compared to men, only women aged 50 years and older had decreased odds of infection. The relationship between age and S. stercoralis prevalence seems to vary across settings [9, 10, 44, 45]. In this national survey of more than 7,200 individuals aged six years and older, we found that prevalence increased with age for both men and women. Previous to this national survey, in North Cambodia, prevalence was found to increase with age and reach a plateau in adulthood, while in Yunnan, China, no cases were found among individuals under the age of 15 [9, 46, 47]. Yet, no association between age and S. stercoralis infection was found in Lao PDR, South Cambodia, or Zanzibar [10, 45, 48]. Age-specific infection risk is of particular importance to target control programs and should be further documented.
Individuals who declared having some knowledge of the health problems caused by worm infections had lower odds of infection with S. stercoralis, but knowledge about the sources of infection was not associated with infection risk. While knowledge does not necessarily translate into behavior change, this result suggests that awareness of personal disease risk—which is an important driver of health promotion and increases compliance with helminth control programs—might be a better trigger of hygienic behavior than knowing exposure sources [49, 50].
The protective effect of improved sanitation against STH infection is widely acknowledged [51–55]. We found that, compared to open defecation, defecating in latrines was protective against S. stercoralis infection. This result is in line with other studies conducted in Cambodia and in Ecuador. It is also consistent with a recent meta-analysis that included nine studies investigating the impact of sanitation on S. stercoralis infection risk, and estimated a pooled OR of 0.50 (95%CI: 0.36–0.70) [9, 10, 18, 55–57]. In North Cambodia, village-level sanitation coverage was also found to reduce re-infection risk one year after treatment [47].
The present work has several limitations. First, women were overrepresented in the sample compared to the general Cambodian population; the lower prevalence among young girls and women aged 50 years and older, compared to males, might have resulted in an underestimation of the prevalence. However, our sample was representative of the 2013 Cambodian general population in terms of age [39].
Second, it was the first time that the serological diagnostic method of detecting IgG antibodies was used for a large-scale survey. This method has proven high sensitivity for S. stercoralis detection, and it does not suffer from cross-reactivity with other STHs or food-borne trematodes [22, 23]. However, validation of the method in different settings should be carried out in order to further promote its use for estimating prevalence in other settings naïve to ivermectin treatment. In a recent study using commercial ELISA kits with different types of antigens (S. ratti, S. stercoralis and rec NIE antigen) to diagnose strongyloidiasis, concordant results between urine and serum ELISA were obtained, which suggests that urine ELISA is a reliable diagnostic method [58]. Third, prevalence estimates at village level suffer from uncertainty due to the study design and should be interpreted with caution. This uncertainty might also have affected our predicted estimates, but provincial-level prevalence rates appeared to be fairly reliable and the overall sample size was reasonably large. Fourth, eight of 249 villages (7.2%) needed to be replaced after the initial selection due to their remoteness. There was reason to believe that the data from these places might be inadequate in terms of quality. Although, S. stercoralis is generally more prevalent in highly remote areas, the number of replacements were low and the geospatial modeling allowed us to predict the infection rates in these remote locations. Finally, our risk factor analysis did not adjust for socio-economic status. Although socio-economic status was found to be associated with infection risk in North Cambodia, results from the few studies that accounted for it are heterogeneous [9, 10, 47, 59, 60]. It is worth noting that socioeconomic status was not a confounder of the relationship between age or sex and S. stercoralis infection risk in North Cambodia and would probably not have substantially affected the estimates for sex and age in the present study [47, 60]. Given the strong association between poverty and other STH infections, it is likely that S. stercoralis risk distribution is also associated with socioeconomic status and future studies should account for it.
Our study represents a clear risk map of S. stercoralis in a highly endemic setting. Based on these data, the number of infected can be quantified, which allows for realistic and concrete planning of control measures. Further developing this operational approach in other settings and with other validated diagnostic approaches will result in databases for global planning. The mainstay of the WHO’s strategy to control STH is preventive chemotherapy, i.e. regular treatment of entire populations or at-risk groups with mebendazole or albendazole to prevent high intensity infections and associated morbidity [61, 62]. However, a single oral dose of either of those drugs is not efficacious against S. stercoralis, for which the drug of choice is ivermectin [63–65]. A single oral dose (200μg/kg Body Weight) of ivermectin was found to achieve a high cure rate and result in re-infection rates below 15%, one year after treatment, in a highly endemic setting in Cambodia [47, 63, 64]. As our results demonstrate, S. stercoralis is highly endemic throughout Cambodia and the inclusion of ivermectin in the control program would be required [13, 65, 66]. Yet, this drug is not subsidized in regions where onchocerciasis is absent, let alone to treat S. stercoralis. The high cost of ivermectin in Cambodia, at USD 10 per tablet (up to five tablets may be needed to treat an individual, depending on their weight) precludes the deployment of adequate control measures in the country.
In the absence of data on age-specific morbidity, the fact that individuals of any age appear to have the same risk for re-infection one year after treatment suggests a need for community-wide control [47]. Yet, a study investigating S. stercoralis-related morbidity in Cambodia found that children and adolescents with higher parasite loads had higher odds of being stunted, while S. stercoralis infection was found to be associated with anemia but not stunting in Argentina [17, 67]. The relationship between S. stercoralis parasite loads, morbidity, and transmission intensity needs to be assessed, along with age-related infection levels, using appropriately designed longitudinal studies. Cost-effectiveness studies of various control options are needed. Mathematical models could help better appraise the parasite transmission dynamics and guide control efforts, as the complex life cycle of S. stercoralis might yield transmission dynamics that differ from other STHs.
Cambodia benefits from a well-established STH control network and was among the first countries to reach the 75% national coverage target [68, 69]. STH deworming activities were recently scaled up to reach children in middle and high schools, including private schools, and women of child-bearing age, working in factories [70]. Additionally, schistosomiasis has been successfully controlled, with no severe cases recorded recently, while lymphatic filariasis has been eliminated as a public health problem and is now under surveillance for elimination [69, 71–73].
In conclusion, S. stercoralis is highly prevalent and ubiquitous in Cambodia and urgently requires control. Although Cambodia benefits from a national helminth control program that has demonstrated its capacity to efficiently address helminthic infections, the current high cost of ivermectin cannot be entirely supported by the Ministry of Health, which precludes its use for large-scale control measures. Subsidies, donations, or the production of affordable generics are necessary to start tackling this potentially dangerous parasite that infects almost a third of the Cambodian population.
|
10.1371/journal.pcbi.1003134 | Charles Bonnet Syndrome: Evidence for a Generative Model in the Cortex? | Several theories propose that the cortex implements an internal model to explain, predict, and learn about sensory data, but the nature of this model is unclear. One condition that could be highly informative here is Charles Bonnet syndrome (CBS), where loss of vision leads to complex, vivid visual hallucinations of objects, people, and whole scenes. CBS could be taken as indication that there is a generative model in the brain, specifically one that can synthesise rich, consistent visual representations even in the absence of actual visual input. The processes that lead to CBS are poorly understood. Here, we argue that a model recently introduced in machine learning, the deep Boltzmann machine (DBM), could capture the relevant aspects of (hypothetical) generative processing in the cortex. The DBM carries both the semantics of a probabilistic generative model and of a neural network. The latter allows us to model a concrete neural mechanism that could underlie CBS, namely, homeostatic regulation of neuronal activity. We show that homeostatic plasticity could serve to make the learnt internal model robust against e.g. degradation of sensory input, but overcompensate in the case of CBS, leading to hallucinations. We demonstrate how a wide range of features of CBS can be explained in the model and suggest a potential role for the neuromodulator acetylcholine. This work constitutes the first concrete computational model of CBS and the first application of the DBM as a model in computational neuroscience. Our results lend further credence to the hypothesis of a generative model in the brain.
| The cerebral cortex is central to many aspects of cognition and intelligence in humans and other mammals, but our scientific understanding of the computational principles underlying cortical processing is still limited. We might gain insights by considering visual hallucinations, specifically in a pathology known as Charles Bonnet syndrome, where patients suffering from visual impairment experience hallucinatory images that rival the vividness and complexity of normal seeing. Such generation of rich internal imagery could naturally be accounted for by theories that posit that the cortex implements an internal generative model of sensory input. Perception then could entail the synthesis of internal explanations that are evaluated by testing whether what they predict is consistent with actual sensory input. Here, we take an approach from artificial intelligence that is based on similar ideas, the deep Boltzmann machine, use it as a model of generative processing in the cortex, and examine various aspects of Charles Bonnet syndrome in computer simulations. In particular, we explain why the synthesis of internal explanations, which is so useful for perception, goes astray in the syndrome as neurons overcompensate for the lack of sensory input by increasing spontaneous activity.
| Visual hallucinations can offer fascinating insights into the mechanisms underlying perceptual processing and the generation of visual experience in the brain. A pathology known as Charles Bonnet syndrome (CBS) [1]–[4] is of particular interest, for two reasons. First, hallucinations in CBS can be very complex in the sense that they entail vivid, life-like, and elaborate imagery of objects, people, animals, or whole visual scenes. Second, the primary cause of CBS is loss of vision due to eye diseases, with no clear pathology in the brain itself and no necessary impairment to mental health other than the hallucinations. De-afferentation of the visual system and sensory deprivation thus seem to be the important factors in the development of CBS, and comparisons have been made to phantom limb phenomena. Unlike for example in the case of schizophrenia, most often accompanied by auditory hallucinations [5], in CBS there thus does not seem to be a more pervasive malfunction of the cognitive system, but rather some form of over-compensation or maladaptation of the relatively healthy brain to the lack of sensory stimulation.
From a theoretical perspective, there has been an attempt to unify complex visual hallucinations in various pathologies in a single qualitative model [6], but many argue that the underlying causal mechanisms are too varied to do so [7]–[9]. That hallucinations occur in many different circumstances however speaks to them relating to essential aspects of perceptual processing. Thus, theoretical explanations that pose that perception inherently involves some form of active synthesis of internal representations might be well positioned to shed light on the generation of spontaneous imagery in hallucinations, which occur even in CBS where there seems to be little defect in the visual system other than at the input stage. Therefore, two key questions arise here: what do complex hallucinations tell us about perceptual processing in general, and what are the mechanisms triggering CBS in particular?
The purpose of this computational study is hence threefold. First, to gain theoretical insights into important principles of cortical inference by employing the deep Boltzmann machine (DBM) as a model system which is based on such (hypothetical) principles. Second, to examine concrete causal mechanisms for CBS, we model homeostatic regulation of neuronal firing activity, elucidating on various aspects of CBS. Moreover, to examine a potential role of the neuromodulator acetylcholine, we introduce a novel model of its action as mediating the balance of feedforward and feedback processing in the cortical hierarchy. And third, with our results we aim to demonstrate the relevance of Deep Learning approaches such as the DBM as models of cortical processing. A preliminary version of the presented work has been published [10].
CBS is characterised by complex recurring visual hallucinations in people who suffer from visual impairment but no other psychological condition or hallucinations in other modalities [1]–[4]. In particular, patients generally gain insight into the unreality of their experiences. The phenomenology of CBS is multifarious, with the nature and content of hallucinatory episodes as well as the conditions favouring their occurrence varying from patient to patient or episode to episode. Common themes are the vividness and richness of detail of the hallucinations, the elaborate content often entailing images of people or animals (though often of a bizarre nature–figures in elaborate costumes, fantastic creatures, extreme colours, etc.), as well as possibly common triggers, such as being in a state of drowsiness and low arousal. Episodes can last from seconds to hours, and hallucinations can reoccur over periods lasting from days to years.
The eponym CBS itself is somewhat ambiguous or even controversial [4], [11]–[13]. Some authors put the emphasis on complex hallucinations in visually impaired but psychologically normal people, where the visual pathology can be anywhere in the visual system from the retina to cortex; others define CBS to be necessarily related to eye diseases only. Similarly, the delineation of the term ‘complex’, and whether CBS should include complex hallucinations only, appears to be not fully clear. On one end are simple or elementary hallucinations consisting of flashes, dots, amorphous shapes, etc., while on the other are fully formed objects or object parts like animals, people, and faces [4], [6]. Somewhere in between are geometric patterns (‘roadmaps’, brickwork, grids, and so forth). Some authors include the latter in CBS [13], [14]. It should be noted that simple hallucinations are actually more common in visually impaired patients than complex ones, with a prevalence of about 50% vs. about 15%, respectively [4]. Both types can occur in individual subjects, possibly with a tendency to progress from simple to complex over time.
For this modelling study, we identify the following key aspects of CBS we aim to capture and elucidate on. First, we take the common definition of hallucinations as compelling perceptual experiences in the absence of external stimuli. They are to be contrasted [4], [6] to illusions as misperceptions concerning an actual external stimulus, as well as to mental imagery. Unlike hallucinations, the latter is under complete volitional control, lacks perceptual vividness (it appears to be ‘in the mind's eye’ rather than in the world), and might also have a different neurobiological substrate [13].
Second, in the context of CBS we are interested in hallucinations that are perceptually rich in the sense that the experience is similar to that of actual seeing. Presumably, this implies that the representations instantiated in the neuronal activity patterns share significant commonalities in both seeing and hallucinating, though this requires further elaboration.
Third, we consider hallucinations on the complex end of the spectrum, i.e. objects, people, and so forth. As we currently lack good generative models of realistic images (biological or otherwise, not counting here of course purely generative algorithms from computer graphics that cannot be inverted for inference) the model we employ still relies on relatively simple binary images. However, it attempts to capture at least some aspects of how complex, object-based hallucinations might be created in the brain. For example, the content of complex hallucinations presumably cannot be accounted for by appealing to anatomical organisational properties of lower visual areas, which [14] suggested for simpler hallucinations of geometric patterns in CBS (referring to anatomical “stripes” in V2 etc.). Our model relies on distributed, high-dimensional, hierarchical representations that go beyond local low-level visual features (e.g. V1-like edge detectors). The representations are learnt and reflect structure in sensory data beyond local correlations.
Fourth, with regards to the issue of whether CBS should refer to hallucinations in the context of eye diseases only, our model is meant as a model of processing in the cortical hierarchy, and due to the level of abstraction we only require that visual input is lost somewhere at a preceding stage and do not differentiate further. We do however address the distinct roles of cortical areas within the hierarchy.
CBS is a complex phenomenon with manifold symptoms and little data beyond clinical case reports and case series. The aim of our computational model is thus to qualitatively elucidate on possible underlying mechanisms, to demonstrate how several common aspects of CBS could be explained, and to gain some potential insights into the nature of cortical inference.
The occurrence of complex visual hallucinations in various pathologies [6], [15] as well as the imagery we all experience in dreams show that the brain is capable of synthesising rich, consistent internal perceptual states even in the absence of, or in contradiction to, external stimuli. It seems natural to consider hallucinations in the context of theoretical accounts of perception that attribute an important functional role to the synthesis of internal representations in normal perception, not just in pathological conditions. In particular, one relevant notion is that of perception entailing an ‘analysis by synthesis’, which is an aspect of approaches such as predictive coding or Adaptive Resonance Theory [16]–[23]. The idea is that ambiguous sensory signals inform initial hypotheses about what is in an image in a bottom-up fashion (from low-level image features to high-level concepts, like objects and faces). These hypotheses are then made concrete in a synthesis stage that tests a hypothesis against the image (or low-level representation thereof) by making top-down predictions using a generative process.
In computational neuroscience over the last two decades, this notion of analysis by synthesis and related ones have often been framed in probabilistic or ‘Bayesian’ terms. Generally speaking, Bayesian approaches theoretically describe how inferences about aspects of the environment are to be made from observations under uncertainty (for reviews and introductions, see [24]–[26]). For hallucinations, the relevant aspect of Bayesian models could be that they offer a way of formalising notions of ‘bottom-up’ processing driven by sensory input, and internally generated, ‘top-down’ processing conveying prior expectations and more high-level learnt concepts. An imbalance of, or erroneous interaction between, such ‘bottom-up’ and ‘top-down’ information could underlie hallucinations [27]–[29]. More concretely, the mathematical entities in a Bayesian model or inference algorithm could map to neural mechanisms and processing in the cortex. For example, inference in a hierarchical model could describe hierarchical processing [21]. Top-down processing then would correspond to information flow from higher areas to lower areas, and inference would be implemented via recurrent interactions between cortical regions. Similarly, in the model of Yu and Dayan [27], a concrete biological mechanism is hypothesised to represent the uncertainty of the prior, namely the neuromodulator acetylcholine. The authors thus refer the latter's relevance in some hallucinatory pathologies as evidence, where deficient acetylcholine, corresponding to an over-emphasis of top-down information in Yu and Dayan's account, could lead to hallucinations [6], [15], [30].
As Yu and Dayan [27] state, a shortcoming of concrete Bayesian models such as theirs is that they are often formulated over very simple, low-dimensional, non-hierarchical variables. It is not clear how their treatment of priors and uncertainty translates to models that deal with high-dimensional problems like images in a biologically plausible manner. This is what we need to address if we hope to develop a computational model of CBS, and in this context we will introduce a novel model of the action of acetylcholine in similar spirit to Yu and Dayan's framework.
While hallucinations in general might relate to an imbalance of bottom-up and top-down in the cortex, the causes behind specifically CBS and the involved mechanisms are poorly understood (for discussion, see [1], [4], [12], [15]). Evidence from CBS and other pathologies suggests that an intact visual association cortex is necessary as well as sufficient for complex visual hallucinations to occur (e.g. [15]). For example, lesions to visual cortex can cause hallucinations, but only if they are localised to earlier areas and do not encompass the higher association cortex. One of the insights emerging from the debate is that the pathology in CBS appears to entail primarily a loss of input at stages prior to association cortex. In contrast, hallucinations accompanying epilepsy, for example, are thought to be caused by an irritative process that directly stimulates association cortices.
How deficient input in CBS leads to the emergence of hallucinations is unclear. Classic psychological theories suggest that the lack of input somehow ‘releases’ or dis-inhibits perceptual representations in visual association cortex. This somewhat vague notion has been made more concrete by taking neuroscientific evidence into account which shows that cortex deafferentiated from input becomes hyper-excitable and generates increased spontaneous activity. As [14] argues (also [12]), changes to neuronal excitability as a consequence of decreased presynaptic input, based on for example synaptic modifications, could thus underlie the emergence of neuronal activity which establishes hallucinatory perception in CBS.
Such adaptive changes of neuronal excitability have been studied extensively over the last two decades in experimental and theoretical work on homeostatic plasticity (see [31] for review; also [32], [33]). Rather than deeming them artifacts or epiphenomena, such changes have been attributed important physiological functions, allowing neurons to self-regulate their excitability to keep their firing rate around a fixed set-point. Homeostatic regulation is thought to stabilise activity in neuronal populations and to keep firing within the neurons' dynamic range, compensating for ongoing changes to neuronal input either due to Hebbian learning, or due to developmental alterations of the number of synapses, connectivity patterns, etc.
A neuron might track its current activity level by measuring its internal calcium levels, and several cellular mechanisms have been identified that could then implement its homeostatic adaptation. Among them is ‘synaptic scaling’, a change to synaptic efficacy that is thought to affect all synapses in a neuron together, keeping their relative strengths intact. Alternatively, the intrinsic excitability of a neuron can be regulated by changing the distribution of ion channels in the membrane. Both mechanisms have been observed experimentally, dynamically regulating neuronal firing rate over a time-span from hours to days [34] in compensation for external manipulations to activity levels–in particular, in response to an activity decrease caused by sensory deprivation.
Hence, with visual input degraded due to eye disease or other defects in the visual pathways, homeostatic over compensation is a strong contender to be the neuronal cause underlying the emergence of hallucinations in CBS. This is the mechanism we explore in our computational model.
To address CBS, we need to work towards computational models that can capture its key properties as identified earlier. Such a model should be able to internally synthesise rich representations of image content, such as objects, even in the absence of (corresponding) sensory input. We now briefly describe the deep Boltzmann machine (DBM). This being the first work that applies DBMs as models of cortical processing, we discuss its interpretation as a biological model. We also specify the parameters used in the simulation experiments. For a more extensive explanation and discussion of all aspects of the DBM framework brought up in this section, see [35].
DBMs are probabilistic, generative neural networks that learn to represent and generate data in an unsupervised fashion. They consist of several layers of neuronal units arranged in a hierarchy. The units fire stochastically, inducing a probability distribution over the network state, parametrised by the weight (and bias) parameters, i.e. the connection strengths between units. DBMs were introduced recently in machine learning by Salakhutdinov and Hinton [36]. While Deep Learning approaches such as the DBM are often taken to be inspired by the brain [37], the relevance of the DBM as a concrete model of processing in the brain has not been explored so far. We argue that DBMs are valuable as models of (hypothetical) aspects of cortical processing, as the computational principles they are based on could play an important role in cortical learning and processing as well.
A DBM is a special case of a general Boltzmann machine (BM) by virtue of its specific architecture. BMs themselves were developed in the nineteen eighties [38]. The reason that DBMs have enjoyed recent interest in machine learning is not that the basic underlying model formulation of a BM has changed; rather, recent developments in learning algorithms have made it possible to effectively train these models, taking advantage of their ‘deep’ structure to overcome earlier problems that made the application of BMs difficult.
Concretely, a DBM consists of layers of neurons (e.g. Figure 1). Usually, the states of the neurons are taken to be binary, , indicating whether a unit is ‘on’ or ‘off’, but other choices are possible, such as continuous-valued, rectified linear units [39]. The states of each layer are written as vectors, denoted by (together denoted by ). Units and in adjacent layers are connected by symmetric connections with connection weight , the latter modelling synaptic strength. For each adjacent pair of layers layers and , the weights can be combined into a weight matrix . Each unit also has a bias parameter that determines its basic activation probability by functioning as a baseline input. In a default DBM, there are no lateral connections between units within a layer.
The first layer constitutes the visible units, i.e. they represent the input data, such as the pixels of images. The higher layers contain hidden units that are not given by the data. Rather, their states form a distributed representation of the input data, the meaning of which is assumed in learning. There, the parameters (weights and biases) are adjusted to learn a good internal model of the sensory input, in a sense to be described below.
Each unit receives input from the other units it is connected to via the weights (plus the bias),(1)This input determines the probability for the unit to switch on. For binary units, it is computed using a sigmoid (logistic) activation function:(2)where denotes all unit states other than . is also called the activation (probability) of unit .
If the DBM is run over a long enough time, by stochastically activating its units, then the probability to find the network in any state asymptotically converges to an equilibrium distribution. In analogy to a system described by (classical) statistical thermodynamics (specifically, the Boltzmann machine corresponds to an Ising model), this distribution is given by the system's Boltzmann distribution (assuming a temperature of ),(3)where is called the energy of the system and is defined as(4)and is the normalisation constant.
DBMs can be understood from two perspectives. The first is to view DBMs as neural networks, simple models of neuronal processing on a comparable level of abstraction and idealisation as other connectionist-style networks used in machine learning and computational cognitive models. In particular, BMs in general can be seen as a generalisation of the Hopfield network [40], [41], which has been used as a basic model of memory storage and recall in neuronal cell assemblies [42]. BMs differ from Hopfield networks in two fundamental respects. First, in the latter, the activation rule is deterministic. Initialised in some state, a Hopfield network will converge to a state that forms a local minimum in the energy ‘landscape’. Learning aims to sets the weights such that this state corresponds to one of the input patterns to be memorised. BMs on the other hand explore the energy landscape stochastically, potentially traversing several minima in the process. The second difference is that Hopfield networks do not have hidden units. Hidden units enable BMs to learn aspects of the data that are not defined by pairwise correlations. Moreover, rather than just capturing correlations between visible units (e.g. pixels in an image) in the weights between them, hidden units can represent specific patterns or features in the visible units, and explicitly signal their presence or absence by virtue of their state. Thus, rather than just memorising patterns, BMs can learn internal representations of sensory data.
The fact that DBMs compute distributed hidden representations in several non-linear processing stages also relates them to feedforward neural networks. However, whereas the latter are usually trained by providing desired output values (i.e., in a supervised fashion), such as image labels, and tuning the weights with the backpropagation algorithm [43], DBMs learn without supervision, attempting to find an internal model from which the input data can be generated.
The second perspective on BMs (and DBMs), perhaps more in line with modern machine learning approaches, views them as probabilistic graphical models of data. In this context, a BM is an instance of a Markov random field, which is a probabilistic graphical model whose independence relationships are captured by an undirected graph (e.g. [44]). Rather than introducing BMs on the basis of the stochastic activation rule, one can instead start from the Boltzmann distribution, Eq. 3, as a definition of the model via its joint distribution over the random variables , and then derive the activation probability (Eq. 2) for each unit simply as conditional probability. ‘Running’ the BM stochastically then produces samples from the joint distribution. In fact, iteratively sampling each unit's state according to its conditional probability implements Gibbs sampling, a Markov chain Monte Carlo (MCMC) method (see e.g. [45]). MCMC and similar sampling-based methods have been suggested to relate to cortical probabilistic inference [26], [46]–[51], and it is focus of our work on modelling bistable perception within the DBM framework [35], [52].
We argue that the DBM is promising as a model of hallucinations, and other aspects of a hypothetical generative model in the cortex, because it implements a generative model that learns to synthesise representations of sensory data. A DBM can be seen as an instance of a hierarchical probabilistic model, and thus could capture the intuition of bottom-up and top-down processing in the cortex reflecting the interaction between sensory information and internal priors. An imbalance of such processing then can be seen as a cause for hallucinations to emerge. At the same time, the DBM is also a simple neural network, thus enabling us to explore concrete neural mechanisms possibly underlying CBS. Because the DBM does not just memorise given input patterns like the related Hopfield network (which itself has been used to model hallucinatory ‘memories’ in schizophrenia [53]), but rather learns internal representations of input images, it is a more concrete model of perception rather than just memory. The ‘deep’ organisation of the DBM into hierarchical layers as well as the image based representations will allow us to make concrete connections to the visual cortex.
The DBM being a generative probabilistic model of sensory data, the act of perception corresponds to inferring the hidden or latent variables that are consistent with and could have generated the observed input. We make a clarification here in light of a current debate concerned with the merit and meaning of approaches to cognition termed ‘Bayesian’ [54]–[58]. The approaches in focus there are characterised as rational, optimal, or ideal observer models. They are meant to describe specific perceptual inference problems, capturing what can in principle be inferred about a specified property of the environment from sensory data. In contrast, in case of the DBM model, the probabilistic framework is used to develop a (component) solution to perceptual tasks, perhaps capturing aspects of processing in the brain, but this solution does not need to be ‘optimal’ in any sense. In particular, the hidden variables in the DBM do not have by design a priori meaning assigned to them in terms of the environment, but rather attain any meaning due to whatever useful representations are discovered in learning. Thus models like the DBM differ conceptually from ideal observer models [35], but these different approaches can still be related to each other as they are based on the same theoretical language of probabilistic inference.
Seen as a model of aspects of cortical processing, the DBM is a rough idealisation, but comparable in that regard to other related modelling approaches [35]. As we will show, the DBM does capture several hypothetical aspects of cortical processing relevant for explaining CBS.
Developing flexible models that can learn useful representations of many kinds of sensory data is one of the key motivations behind Deep Learning approaches such as the DBM. Such versatile learning could also be what makes the cortex so flexible and powerful across many sensory modalities. The learning algorithms for BMs, and DBMs in particular, are themselves not focus of our work on hallucinations here, but we summarise the key points below (see Supplementary Text S1 and [35] for further comments on the biological relevance and plausibility).
Taking the probabilistic model interpretation of a BM, learning can be derived as likelihood optimisation of the model parameters given some sensory training data. Notably, the resulting iterative update rule for the weights of the model involves only local Hebbian learning, and an alternation between two phases where the BM either performs inference over some input or freely generates from its internal model (this second phase could possibly offer a normative explanations for dreams [59]).
There are three key aspects to why BM-based models have found renewed interest in machine learning over the recent years. First, the focus turned to BMs with simplified connectivity, in particular the Restricted BM (RBM), where neither visible units nor hidden units have connections amongst their own type (a RBM is equivalent to a 2-layer DBM). Second, making use of the simplified inference in such models, more effective approximate learning algorithms were developed, such as the Contrastive Divergence algorithm [60]. Third, RBMs were used as building blocks to train deeper, multi-layer architectures such as the DBM. Treating each pair of adjacent layers as its own RBM, the DBM is initially trained one subsequent layer at a time, with each hidden layer learning to generate the unit states in the respective layer below. Once the whole DBM is composed, further training can then be performed on the whole model.
The biological relevance of deep RBM-based models such as the DBM has been examined by matching the learnt neuronal receptive fields to those of neurons in the visual cortex [61], [62]. Our study here is the first to explore the potential of the DBM as a biological model beyond receptive field properties.
To model perceptual phenomena with the DBM, we feed sensory input to the model by clamping the visible layer to images, sampling the hidden layers, and then analyse what is represented in the states of the hidden layers during inference. In the case of hallucinations we are in particular concerned with perceptual content that is not matching the actual visual input. To decode the hidden states in terms of the sensory data they represent, we can make use of the generative nature of the DBM and ask what images would be generated from the hidden states in question. We thus take another DBM instance with the same parameters as the DBM used to model perceptual inference to implement a decoder. For any set of hidden states, the decoder is applied to obtain reconstructed input images for each hidden layer independently (Figure 1).
Specifically, given the states of any hidden layer , , at any point during perceptual inference, we set the respective hidden layer in the decoder DBM to the states to be decoded, and then perform a single deterministic top-down pass starting from there: the activations in each subsequent lower layer are computed using only the layer above as input (propagating ), until a reconstructed image is obtained in the visible layer of the decoder (taking probabilities as grey-scale values). The weights in the decoder are doubled to compensate for the lack of bottom-up input (analogously to the bottom-up initialisation used in [36]). Possible alternatives to this decoding procedure are discussed in [35].
We model CBS as resulting from homeostatic regulation of neuronal excitability in response to degrading visual input. We use DBMs that have learnt to represent images, having trained them on either of two simple data sets. We then simulate the visual impairment by using empty or corrupted input instead of the original data, and have the model perform inference over them. The change in sensory input could lead to changes in the activation levels of the model's neuronal units. To model homeostatic mechanisms, we allow the neurons to adapt their excitability in response.
As discussed earlier, homeostatic plasticity can be described as a neuron adapting its excitability to match its current average firing rate (as measured over hours or days) to a fixed set-point [33], and there are several cellular and synaptic processes making this possible. Here, for simplicity we model a single basic mechanism, namely an iterative adaptation of each neuron's intrinsic excitability. With target activity and current average activity , neuron in the DBM should become either more or less excitable according to the difference . Its bias parameter is thus iteratively incremented by(5)where is a constant parametrising the rate of adaptation. Such an adaptation of the bias has the effect of shifting the activation function of the unit, i.e. the probability for it to switch on, rendering it more or less excitable for a given amount of input (Figure 2; cf. Figure 3a in [31] on homeostatic plasticity).
To define the target activity for each neuron, we simply take the average activity of a unit during inference over the training data (after training) as the normal, ‘healthy’ level of activity for the representations learnt. An alternative would be to use the homeostatic mechanism during training itself, specifying a target activity level for the neurons. This corresponds to a regularisation that has been used in machine learning e.g. to enforce sparsity in the representations [61], [63] (weight decay during training [64] could be seen as another type of homeostatic mechanism akin to synaptic scaling). We report here results without using this mechanism in training itself, but we obtained similar results when trying the latter. Thus, what mattered here is only that the activity levels assumed during training were restored, regardless of whether these levels were originally confined to a certain regime.
We used two training data sets to explore different aspects of CBS (Figure 3). The first is a custom set of binary images containing toy shapes of various sizes at various positions. This shapes data set allowed us to examine issues related to the localisation of visual impairment, and due to its simplicity the perceptual content of the corresponding hallucinations is straightforward to analyse by directly comparing it to training images. The second data set is MNIST, which contains images of handwritten digits and is a standard benchmark used in machine learning. The advantage of MNIST is that it contains objects that, if still simple, arguably have some more interesting structure. With such kinds of data it has been shown that DBMs can learn representations that generalise to unseen instances of the data, not just in terms of classification performance but also in terms of the data they generate themselves [65]. This in particular demonstrates that learning does not simply correspond to memorising training images.
For both data sets, the employed DBMs had three layers of hidden units. The weights between layers were restricted to implement localised receptive fields so that each unit was connected only to a patch of adjacent units in the respective layer below. Receptive fields in the highest hidden layer were global. The biases of the units were initialised to negative values before training to encourage sparse representations. In particular, this lead to a breaking of symmetry between on and off states: by encouraging units to be off most of the time, they learn representations where they signal the presence of specific content in an image by switching on [35], [66]. Input degradation (which models visual impairment) then generally had the effect of decreasing neuronal activity, and in consequence homeostatic regulation would have to recover firing rates by increasing the excitability of the units. This matches the findings that cortical neurons become ‘hyper-excitable’ under sensory deprivation (as reviewed e.g. by [14]). Other than the sign of the activity changes, overall results as reported in this study did not however depend on representations being sparse.
For MNIST, the visible layer had units corresponding to the size of the images in pixels, and , , and units in the three hidden layers, from lowest to highest, respectively. Receptive field sizes were , , and . The model was trained layer-wise for 30 epochs (i.e. iterations through the training data) in each layer, using 5-step Persistent Contrastive Divergence (Supplementary Text S1). The training set contained 60,000 images, 6,000 per digit category (0 to 9). For the shapes data set, the visible layer had units and the hidden layers units each, with receptive field sizes , , . Here, layer-wise training consisted of 30 epochs of 1-step Contrastive Divergence (Supplementary Text S1). The training set again had 60,000 images in total, from six categories (squares, triangles in two orientations, all in two different sizes). It should be noted that, due the limited variability in the shapes data set, all possible image instances were covered by the training set. Hence, only the MNIST data set is suitable to test the generalisation performance of the model. Lastly, for neither MNIST nor the shapes data set were the models trained further after the layer-wise pre-training. See Supplementary Text S2 for further details on the training parameters used.
To measure the preferred activity for each hidden neuron, we averaged its activation over all training data (after learning), with one trial per input image consisting of 50 sampling cycles. Here and elsewhere, the hidden states were generally initialised to zero at the start of a trial. Similarly, to measure the current average activation during homeostatic adaptation, activities were measured over 50 cycles in 100 trials per iteration. Depending on the experiment in question, the visible units were set to a different image for each trial or remained blank (when modelling complete blindness). The adaptation rate was set to 0.1 and 0.04 for models trained on shapes or MNIST, respectively, with a lower rate for MNIST as the model was found to effectively adapt faster for this data set. For the overall results, the precise value of the rate did not matter.
To analyse the perceptual state of the model, we decoded the states of the hidden layers as described earlier, obtaining a reconstructed image for each layer at each sampling step. To evaluate the internal representations w.r.t. their possibly hallucinatory content, we analysed whether the decoded images corresponded to the kind of objects the models had learnt about in training, using the topmost hidden layer's states after 50 sampling cycles for quantitative analysis. For the shapes data set, we employed a simple template matching procedure, matching the image to the shape templates used in training by convolving the former with the latter (each image had its mean subtracted and was then normalised). The maximum value of the resulting 2D vector was taken as quantitative measure for the correspondence, termed the ‘hallucination quality’, where a perfect match corresponded to a hallucination quality of 1.
For the more varied MNIST data set, there are no fixed templates, nor do generated images necessarily match instances from the training set (which is the point of having a model that can generalise, as mentioned above). To obtain a measure of hallucination quality, we classified the decoded image as belonging to any of the digit categories, using the confidence of the classifier as a measure of the image's quality. Specifically, we used an instance of the DBM model itself (not affected by homeostasis) with a classification unit attached (see e.g. [67]). Taking the maximum of the posterior over the digit categories again yielded a measure with maximum value 1. Inspecting the generated image and resulting posterior values, we also confirmed that for images that did not look like well-defined MNIST digits, classification scores computed in this manner tended to be lower. It should be noted that the aim of our work was not achieving high classification performance, hence we did not train the full model, fine-tune the hyper-parameters, nor necessarily implement classification in an ideal fashion. Classification is merely used to analyse the quality of the internal representations. The reported error rate for MNIST (7%) is hence higher than the state of the art, the latter being around 1% for this type of model (e.g. [68]).
The hypothesis we explored is that homeostatic regulation of neuronal firing rate in response to sensory deprivation underlies the emergence of hallucinations in CBS. The possibility for synthesis of internal representations is explained by the cortex implementing a generative model of sensory input. As a first step, we aimed to demonstrate that the homeostasis mechanism as implemented in the model can actually be beneficial in this context.
In the following, we show how homeostatic adaptation could be helpful in particular for a model that implements perceptual inference by synthesising internal representations, by making the learnt representations robust against exactly the sort of visual degradation that ultimately causes CBS. To this end, we had the model (trained on either the shapes or MNIST data sets) perform inference over heavily corrupted versions of the images (Figures 4A and 4E). The latter were created by taking images from the data sets (digit instances not seen in training in the case of MNIST) and setting 65% of the pixels to black.
Degrading the input in this manner lead to profound activity changes in the neurons, which the model was then allowed to compensate for by employing homeostatic adaptation. Figure 4 shows how activity levels changed under input degradation and subsequent adaptation, plotted either against the number of preceding iterations or the total shift of the bias parameter so far (averaged over all units). For all three hidden layers, initial activities were lower when compared to normal levels. Homeostatic adaptation then led to a gradual restoration to the original values.
Importantly, this recuperation of activity levels corresponded to a restored capability of the model's internal representations to capture the underlying objects in the images. We decoded the hidden states of the top layer and classified the resulting reconstructed images using a classifier trained on the original data sets. Input degradation initially lead to a sharp drop in performance in classifying the corrupted images (Figure 4D and 4H). However, homeostatic adaptation lead to a significant improvement of classification, reaching a performance that was close to the one achieved on the decoded representations inferred from uncorrupted images.
Hence, the homeostatic mechanism as defined by Eq. 5 can be sufficient to restore the representations inferred over sensory input as to be suitable for classification. This is despite the fact that it only attempts to match the average activations, i.e. first order statistics of the inferred posteriors averaged over all input images, rather than the full distribution learnt in training, and only does so by adapting the bias parameters. Thus, homeostatic adaptation could offer a simple local neuronal mechanism that serves to make learnt representations robust for example against degradation of sensory input. It does not rely on further learning (in the sense of parameter changes that incorporate incoming sensory data), intricate synaptic changes, or network wide measurements. Rather, each neuron only needs to remember its average activity level and regulate its intrinsic excitability accordingly. However, as we will see in the following, this stabilisation of perceptual representations can be detrimental, ultimately decoupling internal representations from a further degraded sensory input, causing hallucinations.
To model more profound visual impairment or blindness, we then repeated the above experiment but with the visible units permanently clamped to completely empty input. As before, the model had initially been trained on images from either of the two data sets. With the model now performing inference over empty input, homeostatic adaptation was again allowed to take place. Any emergence of meaningful internal representations in the absence of input would correspond to hallucinations. See Figure 5 for an overview of the CBS experiment.
Before presenting the results, we should briefly comment on how the binary input images are to be interpreted so that presenting a blank image corresponds to ‘taking the input away’, i.e. blindness. After all, seeing a black image is not the same as not seeing altogether. Rather, the binary images are here to be understood as proxies of images already encoded in neuronal activity at an early stage of visual processing (e.g. primary visual cortex). We here do not model this earlier encoding for simplicity, but will consider equivalent cases later in experiments where we model loss of vision in higher stages of the hierarchy.
Figures 6A–C show the activity changes resulting from visual impairment and subsequent adaptation for a model trained on MNIST (results for the shapes set were equivalent, Supplementary Figure S1). Again we found an initial drop of activity that was subsequently fully compensated for, at least on average over each hidden layer, by the shift of the intrinsic excitability of the neurons.
What was the nature of the internal representations that allowed for a restoration of activity levels? After all, the purely local adaptation of each neuron might have recovered individual preferred firing rates on the basis of noisy firing or other activation patterns that bore no meaningful representations according to what the model had learnt about initially. Instead, when we decoded the hidden states of the model we found that the represented content after adaptation corresponded to the kind of images seen in training, whereas prior to adaptation, decoded images matched the empty input.
To quantify this, we measured hallucination quality (as defined in the Model section) over the course of homeostatic adaptation. In Figures 6D–F, each dot represents the quality of the image decoded from the topmost hidden states at the end of the 50 sampling cycles in a trial. It becomes apparent that hallucinations started to emerge only after an initial period of silence, even as excitability was already adapting. This is consistent with cases reported in CBS where loss of vision was abrupt [4]. The reported duration of this latent period, ranging from hours to days, in turn matches well the time scale over which homeostatic adaptation takes place [34].
In terms of quality, high-quality hallucinations were found soon after the point when hallucinations emerged (see Figure 7 for example decoded hallucinations). That point also marked a profound increase in the rate of activity changes. This shows that the emergence of stable internal representations is not just a epiphenomenon of underlying activity changes, but rather itself plays a key role in the system recovering normal activity levels.
Throughout the course of adaptation, we found there to be a mix of hallucinations of various qualities. Lower quality images could correspond to temporary states as the model transitioned from one relatively stable state to another. Note that within any one trial, the model never converges to a fixed internal state, as it keeps stochastically sampling from the posterior. We did observe a tendency to stay within one category of object (e.g. a specific class of digit) towards the end of a trial, but this is simply a general property of such models not specific to the hallucinations (we address this issue in [35], [52]). Similarly, hallucinations could come from various object categories (among the digit or shape classes) for an individual instance of the model. This matches reports from CBS patients, which indicate there can be a variety of hallucinatory content that varies from episode to episode for an individual subject [2], [4]. It is thus important that the model could produce varied representations rather than just a few degenerate states.
The emergence of hallucinations in the model does not require complete lack of input. We obtained similar results when performing the homeostasis experiment with images containing, for example, some noise (10% white pixels on black background randomly sampled for each image). In that case, fewer iterations and less homeostatic adaptation were needed to trigger hallucinations (Supplementary Figure S2). Hence, the nature of visual impairment can have an impact on when or whether hallucinations are occurring. This could also offer one possible explanation for why there might be a tendency for hallucinations in CBS to cease once vision is lost completely [4]. If one assumes that there are limits to how much neurons can adapt their excitability, then some remaining input, even if it is just essentially noise, might be necessary to drive cortical neurons sufficiently. On the other hand, an alternative explanation for a cessation of hallucinations might be long-term cortical reorganisation or learning (see Discussion).
Still, one potential problem with our implementation of sensory degradation so far, be it with empty input or noise, could be that it corresponds to a rather extensive damage to the visual system. Perhaps one would be inclined to interpret such input degradation as a model of complete blindness rather than a more graded visual impairment (or one that is more spatially restricted, see the next section), where in the latter case there might be some structure in the sensory data left. Moreover, in all experiments simulated so far, the emergence of hallucinations occurred due to homeostatic adaptation that compensated for a rather massive drop in activation levels caused by the lack of input. However, if the introduced homeostatic mechanism is truly effective at stabilising the distribution of learnt internal representations, one could expect that the system could be prone to hallucinate under much more general conditions than just lack of input: as long as the ongoing input does not evoke a wide variety of learnt percepts, those groups of neurons that participate in representing the lacking percepts might compensate by increasing their excitability, possibly causing corresponding hallucinations.
To address these issues, we aimed to test whether hallucinations were exclusively a consequence of compensation for overall lack of input and resulting activity decreases, or whether they could still emerge with structured input that was however highly impoverished in its variety. To this end, we simulated the homeostatic adaptation for the shapes and MNIST models, with the visible layer clamped to only a single fixed image from the respective data sets over the course of the whole experiment. To clarify, as before, this models slow neuronal changes over the course of perhaps days or longer, rather than fast neuronal adaptation during ongoing perception, with neuronal parameters being fixed during trials and only updated gradually between them.
Results are displayed in Figure 8, depicting activity changes over the three hidden layers and examples of decoded internal representations at various stages. We found that hallucinations did indeed develop: initially, the decoded internal states faithfully represented the image in the sensory input. However, as the neurons adapted over time to compensate for the impoverished input, the internal representations entailed objects not actually in the image, effectively decoupling perception from sensory input.
This result clarifies that the action of the homeostatic mechanism can be much more specific than just recovering overall activity levels. Indeed, for the fixed input images used, initial global activity levels when doing inference were actually at or above average, for the MNIST model and shapes model, respectively (as is shown in the figures). The homeostatic adaptation however acts locally for each neuron. With fixed input, one sub-population of neurons, whose activation distributedly codes for that input, will be highly active, while other groups of neurons are less active than average. Adaptation of neuronal excitability can then continuously shift the balance, even if activity averages across a layer remain similar (we examined a related functional role of neuronal adaptation on shorter time scales in [35], [52]).
As can be observed in the figures, there was an initial drop of global activity levels, especially for the shapes model. Based on the decoded representations at that point, we suggest that this results primarily from the neuronal population that represents the initial, veridical percept decreasing excitability. Then, as other neurons increase their respective excitabilities, alternative, hallucinatory internal representations take over, leading to a stabilisation of global activity levels.
The degree of decoupling of the internal percepts from the sensory input was striking. It appeared to be surprisingly robust, overcoming not just a lack of input but even contradictory input. In the case of the shapes in particular, the hallucinated objects do not even necessarily share parts with the true input. It should be recalled that the homeostatic mechanism merely adapts the local biases, and thus does not at all change the connection strengths between units or layers. Indeed, we could show that the flow of information from sensory input to the higher layers was not completely prohibited in the model after homeostatic adaptation. Running a model that currently displayed hallucinatory representations as if decoupled from input, we modestly increased the impact of feedforward processing, using a mechanism meant to model the action of acetylcholine (to be introduced below). The internal representation then reliably realigned to the actual input image.
Visual impairment leading to CBS can also be constrained to specific parts of the visual field. Although reports are conflicting [4], for some patients at least hallucinations tend to be localised to these regions. We tested whether we could reproduce this finding using the model trained on the shapes data set, in which the objects are distributed across various image positions. We simulated a more localised impairment by repeating the homeostasis experiment while blanking only half of the images (for example the top half, Figure 9A). As before, the neurons' activities dropped initially and then recovered during adaptation as hallucinations emerged (Figure 9B).
In the original homeostasis experiment, where visual impairment involved the whole visible layer, hallucinated objects were distributed across the whole visual field (Figure 9C). However, when the model where only half of the images had been blanked was tested (on blank images), hallucinated objects were restricted to the image region that had been lesioned (Figure 9D). Excitability changes due to homeostatic adaption are thus specific enough in the network to have topographic properties.
Another occasional phenomenon in CBS is that hallucinated objects appear to be “Lilliputian” or miniaturised. It has been suggested that this can be explained as resulting from a mismatch of hallucinated content and context, where hallucinations appear against real visual background that happens to be too close in relation to the size of the hallucinated objects [11]. On the basis of our simulation results, we tentatively make another prediction: if there is a propensity for hallucinatory content to consist of meaningful wholes, such as full objects or faces, then in patients where hallucinations are restricted to impaired regions of the visual field there should be a correlation between object size and the spatial extend of visual impairment. To see this in our model, consider that in our shapes data set, objects could come either in small or large versions. For models with full loss of vision, hallucinations were biased towards the larger objects (Figure 9C). Possibly, this is because larger shapes evoked higher overall activity in the model and in turn were more suitable for activity restoration (note for example in Figure 8A the transition from smaller to larger hallucinations as activity increases from point 3 to 4). On the contrary, in models with lesions restricted to the top half of the visual field, hallucinated objects were not only localised to the impaired region as reported above, but the frequency ratio was also reversed: smaller objects were much more common, and larger objects were less frequent and narrowly centred relative to the impaired region (Figure 9D). Moreover, we found that, without a single exception, all hallucinations of larger shapes happened to be of the ‘downwards-triangle’ category–the only large category where most of the object could fit into the lesioned region.
Thus, the process that generates hallucinations due to homeostatic adaptation can specifically evoke only certain types of content as determined by the nature of the visual impairment. Here, it is those objects that happen to fit within the boundaries of the lesion in the visual field.
We then turned our attention to the question of the roles of different areas in the cortical hierarchy. As described in the introduction, the complex content of hallucinations in CBS suggests the involvement of visual association cortex and other higher visual regions, and evidence implies that intact association cortex is both necessary and sufficient to develop complex hallucinations. For example, cortical lesions in early visual areas can bring about the visual impairment that causes complex hallucinations, but lesions that involve visual association cortex appear to prohibit them.
Interestingly however, a study by [69] suggests that lower areas, when at least partially intact, can still contribute to hallucinatory activity in an essential fashion. The authors examined a patient suffering from CBS due to visual impairment caused by lesions in early visual areas. Maybe contrary to expectation, applying Transcranial Magnetic Stimulation (TMS) to early areas in a way thought to cause cortical suppression lead to a temporary cessation of the hallucinations. The authors argue that their finding goes contrary to the ‘release’ theory of complex hallucinations, according to which the lack of input to higher areas from lower areas somehow disinhibits or releases perceptual representations. Under this theory, the further suppression of the already damaged early areas in the patient should only have exaggerated the hallucinations.
Using the DBM model, we examined these issues relating to the role of areas in the cortical hierarchy. The hierarchical computations in the DBM are simplistic compared to the cortical equivalent; however, we show that a generative model consisting of several subsequent processing stages differentiated at least by increasing receptive field sizes is sufficient to explain the phenomena at hand.
To begin with, we found that DBMs trained without the topmost hidden layer failed to learn generative models of the data, and thus were inevitably incapable of producing corresponding hallucinations. This mirrors visual association cortex being necessary for complex hallucinations, and can be explained in the model with lower layers being incapable of learning the full structure of objects in the images, due to their limited receptive field sizes.
What about intact higher areas being sufficient for the emergence of hallucinations, while lower ones are not necessary? To model lesions to early visual areas, we repeated the homeostasis experiment, only this time we did not blank the input but rather ‘lesioned’ the first hidden layer, i.e. we clamped units in the latter rather than the units in the visible layer to zero (thus, with the first processing stage blocked, the actual content in the visible units was rendered irrelevant). As before, hallucinations did emerge over the course of homeostatic adaptation (Supplementary Figure S3). Hence, remaining layers in the model are sufficient in principle as long as they form a network that can synthesise the relevant information about visual objects.
Finally, we modelled the suppression of early visual areas with TMS in a CBS patient as described by [69]. Unlike in the last experiment, where early areas were permanently incapacitated and higher areas adapted over time, the TMS experiment corresponded to a temporary suppression in a system that had already developed hallucinations, presumably due to prior adaptation to visual impairment. Our setup thus used a model that had undergone homeostatic adaptation in response to blank visual input but with all hidden layers intact, as in the first hallucination experiment, leading to hallucinatory activity. We then temporarily clamped the first hidden layer to zeros, modelling suppression with TMS (assuming that the cortical regions suppressed by TMS in the patient can be modelled to be downstream from the lesioned areas). This caused the hallucinations to cease. Thus, even though this ‘early area’ represented by the first hidden layer is neither sufficient nor necessary for the model to develop hallucinations in the long run (as shown earlier in this section), it can be essential for ongoing hallucinations if it was in the first place part of the system when it underwent homeostatic adaptation.
One possible interpretation of the relevance of lower areas could be that they provide higher areas with unspecific input, in the context of which the adaptation takes place. However, we suggest that the role of lower areas could be more subtle thanks to recurrent interactions with higher ones. As can be seen in the example in Figure 5, the representations assumed in lower layers during hallucinations are somewhat specific to the hallucinated object, even though those layers by themselves are incapable of synthesising it. Thus, this necessarily is a result of feedback from higher areas. It seems plausible that the lower areas could also contribute by stabilising the overall perceptual state assumed across the hierarchy. Then, any significant interference with representations in lower areas, not just suppression of activity, might impede hallucinations. Indeed, in the study of [69], even a TMS protocol used to cause not suppression but illusory flashes of light (“phosphenes”), applied to primary visual cortex of the patient, resulted in a disruption of hallucinatory content. In future work, this could be tested by trying out different forms of manipulations other than suppression in the hidden layers of the model.
Finally, one relatively common feature among CBS patients is that hallucinatory episodes are more likely to occur in states of drowsiness or low arousal. This suggests a role of cholinergic systems, which in turn are implicated in complex hallucinations in a variety of situations outside of CBS, whether drug induced or disease related [15], [30]. Indeed, in the (non-computational) model of complex hallucinations of [6], acetylcholine (ACh) dysfunction is attributed a major importance. At the same time, there is no evidence that an actual ACh dysfunction exists in CBS. Rather, in CBS the correlation with state of arousal might be effected by an interplay of hallucinations with physiologically normal fluctuations of ACh.
Making the connection between a lack of ACh and hallucinations is natural as there is experimental evidence that ACh acts specifically to emphasise sensory input over internally generated information, mediating “the switching of the cortical processing mode from an intracortical to an input-processing mode” [70]. In the computational model of [27], ACh is modelled in a Bayesian framework to modulate the interaction between bottom-up processing carrying sensory information and top-down processing conveying prior expectations. The authors noted the relation to hallucinations, but to our knowledge, there is no computational model exploring it concretely.
Here, we explore an extended interpretation of the action of ACh as mediating the balance between external and intracortical input: in the hierarchy of cortical areas, ACh could affect the balance in the integration of feedforward and feedback information at each stage of the hierarchy. At an intermediate stage, feedforward information from lower areas indirectly carries sensory input, and feedback information is more internally generated, keeping with the idea of a ACh mediated switch between external and internal inputs. However, both feedforward and feedback inputs would in this case be intracortical (perhaps with additional effects on any direct thalamic inputs).
We thus model the effect of ACh in the following way. In the DBM model, each (intermediate) hidden layer receives input from a layer below, conveying directly or indirectly sensory information, and from a layer above that has learnt to generate or predict the former layer's activity. ACh is to set the balance between feedforward and feedback flow of information. We introduce a balance factor , so that an intermediate layer is sampled as(6)given states and weights above and below (biases omitted for brevity). Hence, corresponds to increased feedforward flow of information, assumed to model increased ACh levels, and recovers the normal sampling mode for normal levels. We note that this mechanism is a heuristic in that it treats the DBM as a neural network more than a well-defined probabilistic model. In particular, for , the effective connections between layers are no longer symmetrical and thus the model no longer constitutes a Boltzmann machine (in a sense, the factor interpolates between inference in a DBM and approximate inference in a deep belief net [67], defined with the same parameters).
We modelled the emergence of complex hallucinations in CBS as a result of homeostatic regulation of neuronal firing rate in response to degradation of visual input. Our computational model thus elucidates on similar suggestions in the literature [12], [14]. The homeostasis mechanism is meant to underlie specifically CBS. Other pathologies involving complex hallucinations, such as schizophrenia or Lewy body dementia [15], might have different causes. In particular, it might not be feasible to unify complex hallucinations in a single explanatory framework (as proposed in [6]). What different conditions accompanied by complex hallucinations do have in common however is that they show that the brain can spontaneously synthesise rich representations of visual imagery, even in absence of or in contradiction to actual sensory data. Following notions of the brain implementing perception as analysis by synthesis, our study makes use of the DBM model that can learn to synthesise internal representations of images, in an unsupervised fashion, by virtue of being a generative model.
We reproduced a variety of qualitative aspects of CBS found in some patients, such as an initial latent period, a possible localisation of hallucinations to impaired parts of the visual field, and the effect of suppression of cortical activity. We predict a possible correlation between a tendency to experience miniature versions of objects and the degree to which the spatial extent of visual impairment is limited, as well as activity levels during hallucinatory episodes possibly being higher than what they had been during comparable, stimulus evoked normal perception. We introduced a novel model of the action of acetylcholine (ACh), suggesting that it could not only influence the balance between thalamic and intracortical inputs [70], but also the balance between feedforward and feedback at various stages of the cortical hierarchy. In CBS in particular, a possible lack of ACh at cortical sites, e.g. during normal fluctuations entailed in changes of state of arousal, could be conducive to the emergence of hallucinations.
We suggest that interfering with cortical homeostatic mechanisms might prevent the emergence of hallucinations in CBS. Whether such an intervention would be feasible in practice is unclear, given that the neurobiological mechanisms that underlie homeostatic plasticity are much more complex [33], [72] than our simple model of homeostatic adaptation. Alternatively, perhaps counter to intuition, it might be possible to suppress the formation of hallucinations in CBS by up regulating cortical activity in deprived areas, through pharmacological means or through methods such as TMS, as long as the externally imposed activation is too unspecific to allow for well-formed percepts to emerge.
In the model, internal representations of learnt objects were robustly recovered by the homeostatic adaptation in a variety of conditions, be it complete lack of input, noise input, or naturally structured but highly impoverished input consisting of fixed images. A key aspect of the model was that hallucinations did not consist only of stereotyped images, but rather a variety of percepts reflecting at least a part of the full distribution of objects learnt initially. Such variety across episodes is also reported in many CBS patients [2], [4]. In the model, this variability was due to different groups of neurons participating in coding for different percepts, meaning that a local homeostatic restoration of activity levels for the population required activation of a variety of percepts over time. We would predict that less variety in hallucinatory content should correlate with sensory deprivation being less extensive (e.g. only affecting colour vision, see below).
That hallucinations emerged even when normal input images were used but kept fixed over the course of homeostasis, shows that it was not so much the total lack of sensory input or global drop in evoked activity that mattered, but rather the failure of the given input to evoke a wide range of learnt percepts. Whether impoverished input can have such a powerful impact on perception in reality should be explored further. There is indeed evidence that sensory deprivation (in terms of general impoverishment, not just complete lack of sensory input) can cause hallucinations in healthy individuals [4], [29], but there seems to have been little experimental work along that direction since the nineteen sixties [73].
To our knowledge, our work constitutes the first computational model that concretely explored aspects of CBS. Other neurological pathologies have been studied before with neural network models [74]–[77]. Probably most closely related to our work, Ruppin et al. [53] modelled the emergence of hallucinatory memory patterns in schizophrenia, using a Hopfield network (a line of work initiated by [78]). The underlying mechanism, homeostatic plasticity in response to input degradation, is quite similar, and some analogous observations are made, including a beneficial role for homeostatic regulation for stabilising neuronal representations. However, in their model the hallucinatory ‘memories’, supposedly residing in prefrontal cortex, are accounted for much more abstractly, consisting of random patterns. Moreover, the retrieved patterns in a Hopfield net correspond directly to the patterns provided as input. It is thus not obvious how to relate their network and the stored patterns to specifically visual processing, which is essential for studying CBS. Our model can be seen as a significant extension of their work in that direction. It involves hierarchical, topographic representations of images, learnt in a generative model framework. In particular, the synthesised representations are interpreted to play an integral part in perception itself, not just in unspecified memory-like pattern recall. A generative model moreover relates to other approaches discussed in the context of hallucinations (Bayesian inference, predictive coding, adaptive resonance; [27]–[29], [79]).
We emphasise the distinction between the roles that homeostatic adaptation and learning play in our model and possibly the cortex. Learning is to be seen as a lasting change of circuitry that captures aspects of the sensory input in the neuronal representations, improving the network's function according to some criterion. In the generative model, that criterion would be the ability to generate or predict the input itself, but it could also be the utility of the representations towards some other goal, such as discrimination of objects. Homeostatic adaptation on the other hand could serve to stabilise neuronal representations. While such stabilisation can in turn be important during learning itself [33], we have shown in the model that it could offer a simple local mechanism to make representations more robust once they have been learnt, for instance to counteract degradation in input quality [53] –thus effectively resisting changing aspects of the input, rather than capturing them via learning.
At the point in time where we simulate homeostatic stabilisation, learning might have concluded, having taken place in earlier stages of development, or it could still occur but over longer time scales. A decoupling of the time scales of homeostatic adaptation and learning could also explain why CBS can recede over time. Hallucinations might initially be caused by the short-term homeostatic regulation of neuronal activity, but long-term cortical reorganisation could lead to their cessation [14]. In our framework, such reorganisation would correspond to learning to generate the impaired sensory input. Indeed, if we continue learning in the model as the input layer is clamped to empty or noise images, rather than just perform homeostatic adaptation, the model learns to generate and thus represent the empty input, losing the capability for hallucinations in the process.
In conditions such as schizophrenia, multiple sensory modalities are affected and hallucinations are only one symptom among many, including delusional beliefs. In contrast, hallucinations in CBS are, by definition, restricted to the visual modality and patients gain insight into the unreality of their percepts (at least upon reflection or after being corrected by others [2], [4]). These features of CBS are explained by our account: in our model, hallucinatory representations are restricted to neuronal populations most directly affected by lack of sensory drive (even respecting retinotopy). Thus, there is no reason to expect that non-visual areas should be impaired in any way, including prefrontal areas. CBS patients should hence be able to reason about their percepts being unreal.
As for the underlying mechanisms, we suggest that homeostatic compensation triggered by degrading input is key to CBS but not necessarily schizophrenia (though see [53]). Briefly, many neural network models of schizophrenia [76], [77] can be characterised as proposing that internal disruptive neural changes (such as increased noise or excessive synaptic pruning) destabilise internal representations, primarily in non-sensory areas or across cortical systems (thus affecting reasoning as well). In sensory areas deprived of sensory input, it is not clear that unspecific maladaptive changes such as increased noise alone could generate the lasting, complex, coherent, and varying hallucinations of CBS. Instead our proposal is that in CBS, it is in a sense a stabilisation of internal representations, in response to external disruptions in the sensory periphery, that causes hallucinations.
It should be noted that neurobiological changes such as increased noise or synaptic pruning could also be explored in the DBM. However, if non-sensory areas such as prefrontal cortex are the subject of inquiry, then the DBM and the hierarchical generative model it embodies might not be the most appropriate framework.
Our study can also be compared to recently proposed Bayesian accounts of schizophrenia [29], [77]. Hallucinations in CBS could on a high level be described as internal priors being too strong. Bayesian accounts of schizophrenia, however, involve more complex hypotheses about the role of feed-forward and feed-back processing (e.g. in the context of predictive coding [29]) that are not the focus of our study.
One of the issues we have not addressed is what limits the incidence of complex hallucinations and CBS to about 11% to 15% of patients suffering from visual impairment [4]. Our modelling results suggest however that a variety of parameters can influence whether and when hallucinations occur. In the model, the nature and degree of visual impairment as well the effect and variability of other interacting factors, such as ACh levels, determine how much homeostatic adaptation is necessary to push cortical activity into the hallucinating regime. Limits on how much cortical neurons can adapt their excitability therefore would restrict hallucinations to only certain cases, and there might be variability in such parameters of homeostasis across the population as well. Thus, that only some patients with visual impairment develop hallucinations could simply reflect the variance of the underlying relevant parameters. Similar reasoning might explain the diversity of symptoms among CBS patients.
Differences in hallucinatory content, e.g. whether it does or does not involve movement, faces, strong colours, etc., likely relate to the specialisation of different cortical areas [3], [71], and potentially to their selective sensory deprivation (such as more extensive impairment of colour vision possibly predisposing patients with senile macular degeneration to experience coloured hallucinations [3]). A specialisation of different areas to different aspects of the sensory data was not a feature of our model. However, it seems reasonable to extrapolate from our results to a model extended in that regard. In our simulations, restricting sensory input by either removing only parts of the images or by just fixing input to a single image led to hallucinations that reflected the specific lack in the input (namely hallucinations in the deprived part of the visual field, or of object types not present in the fixed input image, respectively). If different parts of the model were to distinctly represent properties of visual input in analogy to for example cortical areas V4 for colour and MT for motion, we would expect a specific deprivation of that input property to lead to corresponding hallucinatory representations.
An open question in CBS is also in how far hallucinated content reflects visual memories of some sort [4], although the elaborate and occasionally bizarre nature of the images might speak against this (see [2], [12] for examples). In this context it is relevant that the DBM has been shown to be capable of synthesising images that generalise beyond what it has been trained on [65]. Moreover, in light of the bizarre or unusual hallucinatory imagery in CBS, some hallucinations with low quality in our simulations (as measured relative to training images) could possibly be interpreted as such unnatural imagery (see e.g. Figure 8b (3); Ruppin et al. [53] made a similar observation in their model).
The key for a model of CBS is to account for the ability of the brain to synthesise rich internal representations of images even without visual input, representations that possibly generalise over earlier experienced inputs (as argued above). This does not necessarily imply that the brain implements a generative model, in the sense captured by the DBM. However, the strength of such generative frameworks is that they account for these aspects naturally, at least in principle.
For comparison, a perceptual Bayesian model defined over a single low-dimensional variable can be sufficient to account for perceptual illusions concerning a property of an object (e.g. due to a prior for slow speeds [80]), but it is far-off from actually generating a full visual representation of the object itself. Similarly, the necessity for synthesis without input implies that a model computing a rich code of a given image is on its own not sufficient either. For example, the predictive coding model of [18] and the sparse coding model of [81] are both formulated as generative models that learn representations from images. Given an input image, they can infer a code that is rich enough in information to reconstruct the former. However, neither model can, when run purely generatively, synthesise structured images or anything akin to objects (although [82] demonstrate that memorised images can be recalled). In particular, sparse coding trained on images tends to discover localised patches of edges as independent ‘causes’. Thus, without an extension to higher level causes, a generated image will be a random superposition of such edges.
Similarly, neural networks like (deep) auto-encoders learn internal representations by reconstructing input. Using bottlenecks in the hidden layers, sparsity, input reconstruction from noise-corrupted input and other techniques [37], they also learn about the underlying structure in images, enabling them to reconstruct from corrupted input, perform dimensionality reduction, or even learn transformations of the content [83]. However, there is no way of generating from these models in the absence of input (but see the recent work of [84]). Hence, again such an approach might be used to model illusions, but not hallucinations.
Clearly, while our model, the DBM, is a generative model, its capability to generate ‘images’ still leaves much to be desired when it comes to matching the perceptual richness attributed to real images (although the DBM and closely related models have shown more potential in that regard than what is demonstrated here, see [36], [85], [86]). As model of cortical representations and processing, it also makes several simplifying abstractions, such as lumping together the highly differentiated feedforward and feedback connections in the cortex (e.g. [87]) into simple symmetrical connections. Of particular interest are thus recent extensions that could enhance the generative performance of DBM-like approaches while at the same time having biological relevance as well, such as including lateral connections [88] or complex cell like pooling [89], [90].
However, our work here demonstrates that the DBM does in principle capture several aspects important for explaining CBS, idealisations notwithstanding. It is not meant as definitive model of generative processing in the brain, but rather serves as a simple idealised model system just complex enough to convey the points in question. Among the relevant aspects it captures is, first, the aforementioned capability to synthesise representations of input. Second, its hierarchical and topographic representations allowed us to model localised impairment and a role for ACh. Third, the nature of the DBM as a neural network made it possible to model concrete cellular homeostatic mechanisms. Fourth, unlike for example the earlier Helmholtz machine model [91], the DBM uses top-down interactions also during inference, not just learning, another requirement for modelling the role of hierarchical bottom-up and top-down processing for hallucinations. There are other aspects of cortical processing that are not part of the DBM framework but were not essential to the questions we sought to address in this work. The DBM would be less suitable if, for example, one were to hypothesise that some features of CBS relate specifically to the anatomical or functional asymmetry of cortical feedforward and feedback connections.
Our model of the action of ACh is closely related in spirit to that of Yu and Dayan [27]. In a sense we addressed some of the issues they identified with their own approach, namely only dealing with a localist representation of a low-dimensional variable, and only with a shallow hierarchy where the interaction of bottom-up and top-down is confined to a single stage. As they write, “it would be more biologically realistic to consider distributed representations at each of many levels in a hierarchy”, which might be closer to what our model implements.
In Yu and Dayan's model, the ACh mechanism implements an approximation to exact inference: only a single hypothesis is maintained at any point in time by the top-down part of the system, with ACh controlling the impact of that hypothesis on perceptual inference. This is comparable to the action of ACh on the influence of higher layers on lower layers in our model. However, the functional role of ACh was not the main focus of our work, and in some ways their model is significantly more sophisticated than ours in that regard. In particular, in their model the ACh level is itself controlled by the system dynamically during ongoing inference, whereas we merely manipulated ACh manually to explore its impact on emerging hallucinations. Whether such an internal control of the ACh parameter could be implemented in the DBM framework, in particular in a principled fashion, is open.
Another issue is in how far the role of ACh, and the interaction of top-down and bottom-up in hallucinations in general, is necessarily to be interpreted in ‘Bayesian’ or probabilistic terms. In Yu and Dayan's model, ACh represents the uncertainty associated with the current top-down hypothesis, and this uncertainty is itself subject to ongoing probabilistic inference. Because a mechanism for inferring this uncertainty is lacking in our model, we would be more cautious to necessarily frame the interaction of bottom-up and top-down as ‘Bayesian’ here. For our approach here, the probabilistic nature of the DBM only comes into play in so far as it allows for a means of formulating and deriving a generative model of sensory data (we emphasise the probabilistic aspect of the DBM model elsewhere [35], [52]).
A subtle issue is how much information needs to be synthesised in the brain, and in what form, to generate the visual experience of hallucinations. Mostly avoiding the difficult question of the neural correlates of consciousness here (e.g. [92]), we can at least pose necessary, though not sufficient, conditions for the generated neuronal representations to evoke complex visual hallucinations: they somehow must entail the information content that is implied in the percepts (assuming CBS patients are not just confabulating). For example, both seeing and hallucinating a dog entails much more than just being aware (and able to report) that the object in question is indeed a dog, i.e. some sort of category label. Rather, it involves perceiving the shape, contours, texture, colours, and so forth. Thus, internal activation of an abstract, low-dimensional representation of the concept of a dog would not be sufficient.
For instance, consider a simple perceptual model consisting of a neural network classifier such as a perceptron, which has learnt to classify images of dogs against other images, using a single binary output ‘neuron’. Internal activation of this unit alone cannot possibly be accompanied by the visual experience of seeing a dog, as the single bit of information conveyed by its state cannot possibly be used to differentiate among the various possible instantiations of dogs (a dalmatian in a specific pose rather than a poodle in another, etc.) [93].
The synthesis of rich internal representations of data, and how this capability is acquired through learning in the first place, is naturally explained in strong generative models such as the DBM. In the cortical hierarchy, a top-down generative component could also offer a mechanism to recover more detailed low-level representations from more high-level abstract representations, details that might be discarded during bottom-up or feedforward processing to obtain invariant representations (e.g. [94]). Alternatively, such detailed information might still be present at the high-level, but be only implicit and not easy to access by the rest of the brain. Top-down processing could then serve to transform such information into a more explicit (for the rest of the brain) representation. Either could explain why the generation of conscious experience might be related to re-entrant top-down processing [92].
We have demonstrated how the DBM as a generative neural network can provide potential insights into the mechanisms underlying complex visual hallucinations in CBS. Our results here, together with other work [35], [52], [66], offer a novel perspective on perceptual phenomena by relating them to inference in a generative model in the cortex.
|
10.1371/journal.pcbi.1005358 | How evolution learns to generalise: Using the principles of learning theory to understand the evolution of developmental organisation | One of the most intriguing questions in evolution is how organisms exhibit suitable phenotypic variation to rapidly adapt in novel selective environments. Such variability is crucial for evolvability, but poorly understood. In particular, how can natural selection favour developmental organisations that facilitate adaptive evolution in previously unseen environments? Such a capacity suggests foresight that is incompatible with the short-sighted concept of natural selection. A potential resolution is provided by the idea that evolution may discover and exploit information not only about the particular phenotypes selected in the past, but their underlying structural regularities: new phenotypes, with the same underlying regularities, but novel particulars, may then be useful in new environments. If true, we still need to understand the conditions in which natural selection will discover such deep regularities rather than exploiting ‘quick fixes’ (i.e., fixes that provide adaptive phenotypes in the short term, but limit future evolvability). Here we argue that the ability of evolution to discover such regularities is formally analogous to learning principles, familiar in humans and machines, that enable generalisation from past experience. Conversely, natural selection that fails to enhance evolvability is directly analogous to the learning problem of over-fitting and the subsequent failure to generalise. We support the conclusion that evolving systems and learning systems are different instantiations of the same algorithmic principles by showing that existing results from the learning domain can be transferred to the evolution domain. Specifically, we show that conditions that alleviate over-fitting in learning systems successfully predict which biological conditions (e.g., environmental variation, regularity, noise or a pressure for developmental simplicity) enhance evolvability. This equivalence provides access to a well-developed theoretical framework from learning theory that enables a characterisation of the general conditions for the evolution of evolvability.
| A striking feature of evolving organisms is their ability to acquire novel characteristics that help them adapt in new environments. The origin and the conditions of such ability remain elusive and is a long-standing question in evolutionary biology. Recent theory suggests that organisms can evolve designs that help them generate novel features that are more likely to be beneficial. Specifically, this is possible when the environments that organisms are exposed to share common regularities. However, the organisms develop robust designs that tend to produce what had been selected in the past and might be inflexible for future environments. The resolution comes from a recent theory introduced by Watson and Szathmáry that suggests a deep analogy between learning and evolution. Accordingly, here we utilise learning theory to explain the conditions that lead to more evolvable designs. We successfully demonstrate this by equating evolvability to the way humans and machines generalise to previously-unseen situations. Specifically, we show that the same conditions that enhance generalisation in learning systems have biological analogues and help us understand why environmental noise and the reproductive and maintenance costs of gene-regulatory connections can lead to more evolvable designs.
| Explaining how organisms adapt in novel selective environments is central to evolutionary biology [1–5]. Living organisms are both robust and capable of change. The former property allows for stability and reliable functionality against genetic and environmental perturbations, while the latter provides flexibility allowing for the evolutionary acquisition of new potentially adaptive traits [5–9]. This capacity of an organism to produce suitable phenotypic variation to adapt to new environments is often identified as a prerequisite for evolvability, i.e., the capacity for adaptive evolution [7, 10, 11]. It is thus important to understand the underlying variational mechanisms that enable the production of adaptive phenotypic variation [6, 7, 12–18].
Phenotypic variations are heavily determined by intrinsic tendencies imposed by the genetic and the developmental architecture [18–21]. For instance, developmental biases may permit high variability for a particular phenotypic trait and limited variability for another, or cause certain phenotypic traits to co-vary [6, 15, 22–26]. Developmental processes are themselves also shaped by previous selection. As a result, we may expect that past evolution could adapt the distribution of phenotypes explored by future natural selection to amplify promising variations and avoid less useful ones by evolving developmental architectures that are predisposed to exhibit effective adaptation [10, 13]. Selection though cannot favour traits for benefits that have not yet been realised. Moreover, in situations when selection can control phenotypic variation, it nearly always reduces such variation because it favours canalisation over flexibility [23, 27–29].
Developmental canalisation may seem to be intrinsically opposed to an increase in phenotypic variability. Some, however, view these notions as two sides of the same coin, i.e., a predisposition to evolve some phenotypes more readily goes hand in hand with a decrease in the propensity to produce other phenotypes [8, 30, 31]. Kirschner and Gerhart integrated findings that support these ideas under the unified framework of facilitated variation [8, 32]. Similar ideas and concepts include the variational properties of the organisms [13], the self-facilitation of evolution [20] and evolution as tinkering [33] and related notions [6, 7, 10, 12]. In facilitated variation, the key observation is that the intrinsic developmental structure of the organisms biases both the amount and the direction of the phenotypic variation. Recent work in the area of facilitated variation has shown that multiple selective environments were necessary to evolve evolvable structures [25, 27, 34–36]. When selective environments contain underlying structural regularities, it is possible that evolution learns to limit the phenotypic space to regions that are evolutionarily more advantageous, promoting the discovery of useful phenotypes in a single or a few mutations [35, 36]. But, as we will show, these conditions do not necessarily enhance evolvability in novel environments. Thus the general conditions which favour the emergence of adaptive developmental constraints that enhance evolvability are not well-understood.
To address this we study the conditions where evolution by natural selection can find developmental organisations that produce what we refer to here as generalised phenotypic distributions—i.e., not only are these distributions capable of producing multiple distinct phenotypes that have been selected in the past, but they can also produce novel phenotypes from the same family. Parter et al. have already shown that this is possible in specific cases studying models of RNA structures and logic gates [34]. Here we wish to understand more general conditions under which, and to what extent, natural selection can enhance the capacity of developmental structures to produce suitable variation for selection in the future. We follow previous work on the evolution of development [25] through computer simulations based in gene-regulatory network (GRN) models. Many authors have noted that GRNs share common functionality to artificial neural networks [25, 37–40]. Watson et al. demonstrated a further result, more important to our purposes here; that the way regulatory interactions evolve under natural selection is mathematically equivalent to the way neural networks learn [25]. During evolution a GRN is capable of learning a memory of multiple phenotypes that were fit in multiple past selective environments by internalising their statistical correlation structure into its ontogenetic interactions, in the same way that learning neural networks store and recall training patterns. Phenotypes that were fit in the past can then be recreated by the network spontaneously (under genetic drift without selection) in the future or as a response to new selective environments that are partially similar to past environments [25]. An important aspect of the evolved systems mentioned above is modularity. Modularity has been a key feature of work on evolvability [6, 29, 41, 42] aiming to facilitate variability that respects the natural decomposable structure of the selective environment, i.e., keep the things together that need to be kept together and separate the things that are independent [6, 12, 20, 41]. Accordingly, the system can perform a simple form of generalisation by separating knowledge from the context in which it was originally observed and re-deploying it in new situations.
Here we show that this functional equivalence between learning and evolution predicts the evolutionary conditions that enable the evolution of generalised phenotypic distributions. We test this analogy between learning and evolution by testing its predictions. Specifically, we resolve the tension between canalisation of phenotypes that have been successful in past environments and anticipation of phenotypes that are fit in future environments by recognising that this is equivalent to prediction in learning systems. Such predictive ability follows simply from the ability to represent structural regularities in previously seen observations (i.e., the training set) that are also true in the yet-unseen ones (i.e., the test set). In learning systems, such generalization is commonplace and not considered mysterious. But it is also understood that successful generalisation in learning systems is not for granted and requires certain well-understood conditions. We argue here that understanding the evolution of development is formally analogous to model learning and can provide useful insights and testable hypotheses about the conditions that enhance the evolution of evolvability under natural selection [42, 43]. Thus, in recognising that learning systems do not really ‘see into the future’ but can nonetheless make useful predictions by generalising past experience, we demystify the notion that short-sighted natural selection can produce novel phenotypes that are fit for previously-unseen selective environments and, more importantly, we can predict the general conditions where this is possible. This functional equivalence between learning and evolution produces many interesting, testable predictions (Table 1).
In particular, the following experiments show that techniques that enhance generalisation in machine learning correspond to evolutionary conditions that facilitate generalised phenotypic distributions and hence increased evolvability. Specifically, we describe how well-known machine learning techniques, such as learning with noise and penalising model complexity, that improve the generalisation ability of learning models have biological analogues and can help us understand how noisy selective environments and the direct selection pressure on the reproduction cost of the gene regulatory interactions can enhance evolvability in gene regulation networks. This is a much more sophisticated and powerful form of generalisation than previous notions that simply extrapolate previous experience. The system does not merely extend its learned behaviour outside its past ‘known’ domain. Instead, we are interested in situations where the system can create new knowledge by discovering and systematising emerging patterns from past experience, and more notably, how the system separates that knowledge from the context in which it was originally observed, so that it can be re-deployed in new situations.
Some evolutionary mechanisms and conditions have been proposed as important factors for improved evolvability. Some concern the modification of genetic variability (e.g., [36, 44, 45] and [46]), while others concern the nature of selective environments and the organisation of development including multiple selective environments [36], sparsity [47], the direct selective pressure on the cost of connections (which can induce modularity [27, 44] and hierarchy [48]), low developmental biases and constraints [49] and stochasticity in GRNs [50]. In this paper, we focus on mechanisms and conditions that can be unified and better understood in machine learning terms, and more notably, how we can utilise well-established theory in learning to characterise general conditions under which evolvability is enhanced. We thus provide the first theory to characterise the general conditions that enhance the evolution of developmental organisations that generalise information gained from past selection, as required to enhance evolvability in novel environments.
The main experimental setup involves a non-linear recurrent GRN which develops an embryonic phenotypic pattern, G, into an adult phenotype, Pa, upon which selection can act [25]. An adult phenotype represents the gene expression profile that results from the dynamics of the GRN. Those dynamics are determined by the gene regulatory interactions of the network, B [38, 39, 47, 53, 54] (see Developmental Model in S1 Appendix). We evaluate the fitness of a given genetic structure based on how close the developed phenotype is to the target phenotypic pattern, S. S characterises the direction of selection for each phenotypic trait, i.e., element of gene expression profile, in the current environment. The dynamics of selective environments are modelled by switching from one target phenotype to another every K generations. K is chosen to be considerably smaller than the overall number of generations simulated. Below, we measure evolutionary time in epochs, where each epoch denotes NT × K generations and NT corresponds to the number of target phenotypes. (Note that epoch here is a term we are borrowing from machine learning and does not represent geological timescale.)
In the following experiments all phenotypic targets are chosen from the same class (as in [25, 34]). This class consists of 8 different modular patterns that correspond to different combinations of sub-patterns. Each sub-pattern serves as a different function as pictorialised in Fig 1. This modular structure ensures that the environments (and thus the phenotypes that are fittest in those environments) share common regularities, i.e., they are all built from different combinations from the same set of modules. We can then examine whether the system can actually ‘learn’ these systematicities from a limited set of examples and thereby generalise from these to produce novel phenotypes within the same class. Our experiments are carried out as follows. The population is evolved by exposure to a limited number of selective environments (training). We then analyse conditions under which new phenotypes from the same family are produced (test). As an exemplary problem we choose a training set comprised of three phenotypic patterns from the class (see Fig 2a).
One way to evaluate the generalisation ability of developmental organisations is to evolve a population to new selective environments and evaluate the evolved predisposition of the development system to produce suitable phenotypes for those environments (as per [34]). We do this at the end of experimental section. We also use a more stringent test and examine the spontaneous production of such phenotypes induced by development from random genetic variation. Specifically, we examine what phenotypes the evolved developmental constraints and biases B are predisposed to create starting from random initial gene expression levels, G. For this purpose, we perform a post-hoc analysis. First, we estimate the phenotypic distributions induced by the evolved developmental architecture under drift. Since mutation on the direct effects on the embryonic phenotypes (G) in this model is much greater than mutation on regulatory interactions (B) (see Methods), we estimate drift with a uniformly random distribution over G (keeping B constant). Then we assess how successful the evolved system is at producing high-fitness phenotypes, by seeing if the phenotypes produced by the evolved correlations, B, tend to be members of the general class (see Methods).
In this section, we focus on the conditions that promote the evolution of adaptive developmental biases that facilitate generalised variational structures. To address this, we examine the distributions of potential phenotypic variants induced by the evolved developmental structure in a series of different evolutionary scenarios: 1) different time-scales of environmental switching, 2) environmental noise and 3) direct selection pressure for simple developmental processes applied via a the cost of ontogenetic interactions favouring i) weak and ii) sparse connectivity.
We next asked why costly interactions and noisy environments facilitate generalised developmental organisations. To understand this, we monitor the match between the phenotypic distribution induced by the evolved developmental process and the ones that describe the past selective environments (training set) and all potential selective environments (test set) respectively over evolutionary time in each evolutionary setting (see Methods). Following conventions in learning theory, we term the first measure ‘training error’ and the second ‘test error’. This demonstrates predictions c, e and f from Table 1.
The dependence of the respective errors on evolutionary time are shown in Fig 3. For the control scenario (panel A) we observe the following trend. Natural selection initially improved the fit of the phenotypic distributions to both distributions of past and future selective environments. Then, while the fit to past selective environments continued improving over evolutionary time, the fit to potential, but yet-unseen, environments started to deteriorate (see also Fig B in Supporting Figures in S1 Appendix). The evolving organisms tended to accurately memorise the idiosyncrasies of their past environments, at the cost of losing their ability to retain appropriate flexibility for the future, i.e., over-fitting. The dashed-line in Fig 3A indicates when the problem of over-fitting begins, i.e., when the test error first increases. We see that canalisation can be opposed to the evolution of generalised phenotypic distributions in the same way over-fitting is opposed to generalisation. Then, we expect that preventing the canalisation of past targets can enhance the generalisation performance of the evolved developmental structure. Indeed, Fig 3B, 3C and 3D confirm this hypothesis (predictions a-c from Table 1).
In the presence of environmental noise, the generalisation performance of the developmental structure was improved by discovering a set of regulatory interactions that corresponds to the minimum of the generalisation error curve of 0.34 (Fig 3B). However, natural selection in noisy environments was only able to postpone canalisation of past targets and was unable to avoid it in the long term. Consequently, stochasticity improved evolvability by decreasing the speed at which over-fitting occurs, allowing for the developmental system to spend more time at a state which was characterised by high generalisation ability (see also Fig A in The Structure of Developmental Organisation in S1 Appendix). On the other hand, under the parsimony pressure for weak connectivity, the evolving developmental system maintained the same generalisation performance over evolutionary time. The canalisation of the selected phenotypes was thus prevented by preventing further limitation of the system’s phenotypic variability. Note that the outcome of these two methods (Fig 3B and 3C) resembles in many ways the outcome as if we stopped at the moment when the generalisation error was minimum, i.e., early stopping; an ad-hoc solution to preventing over-fitting [51]. Accordingly, learning is stopped before the problem of over-fitting begins (see also Fig A in The Structure of Developmental Organisation in S1 Appendix). Under parsimony pressure for sparse connectivity, we observe that the generalisation error of the evolving developmental system reached zero (Fig 3D). Accordingly, natural selection successfully exploited the time-invariant regularities of the environment properly representing the entire class (Fig 2h). Additionally, Fig D in Supporting Figures in S1 Appendix shows that the entropy of the phenotypic distribution reduces as expected over evolutionary time as the developmental process increasingly canalises the training set phenotypes. In the case of perfect generalisation to the class (sparse connectivity), this convergence reduces from 16 bits (the original phenotype space) to four bits, corresponding to four degrees of freedom where each of the four modules vary independently. In the other cases, overfitting is indicated by reducing to less than four bits.
As seen so far, the generalisation ability of development can be enhanced under the direct selective pressure for both sparse and weak connectivity and the presence of noise in the selective environment, when the strength of parsimony pressure and the level of noise were properly tuned. Different values of λ and κ denote different evolutionary contexts, where λ determines the relative burden placed on the fitness of the developmental system due to reproduction and maintenance of its elements, or other physical constraints and limitations, and κ determines the amount of extrinsic noise found in the selective environments (see Evaluation of fitness).
In the following, we analyse the impact of the strength of parsimony pressure and the level of environmental noise on the evolution of generalised developmental organisations. Simulations were run for various values of parameters λ and κ. Then, the training and generalisation error were evaluated and recorded (Fig 4). This demonstrates prediction (g) from Table 1.
We find that in the extremes, low and high levels of parsimony pressures, or noise, gave rise to situations of over-fitting and under-fitting respectively (Fig 4). Very small values of λ, or κ, were insufficient at finding good regulatory interactions to facilitate high evolvability to yet-unseen environments, resulting in the canalisation of past targets, i.e., over-fitting. On the other hand, very large values of λ over-constrained the search process hindering the acquisition of any useful information regarding environment’s causal structure, i.e., under-fitting. Specifically, with a small amount of L1-regularisation, the generalisation error is dropped to zero. This outcome holds for a wide spectrum of the regularisation parameter ln(λ) ∈ [0.15, 0.35]. However, when λ is very high (here λ = 0.4), the selective pressure on the cost of connection was too large; this resulted in the training and the generalisation errors corresponds to the original ‘no model’ situation (Fig 4C). Similarly, with a small amount of L2-regularisation, the generalisation error quickly drops. In the range [10, 38] the process became less sensitive to changes in λ, resulting in one optimum at λ = 38 (Fig 4B). Similar results were also obtained for jittering (Fig 4A). But the generalisation performance of the developmental process changes ‘smoothly’ with κ, resulting in one optimum at κ = 35 × 10−4 (Fig 4A). Inductive biases need to be appropriate for a given problem, but in many cases a moderate bias favouring simple models is sufficient for non-trivial generalisation.
Lastly we examine whether generalised phenotypic distributions can actually facilitate evolvability. For this purpose, we consider the rate of adaptation to each of all potential selective environments as the number of generations needed for the evolving entities to reach the respective target phenotype.
To evaluate the propensity of the organisms to reach a target phenotype as a systemic property of its developmental architecture, the regulatory interactions were kept fixed, while the direct effects on the embryonic phenotype were free to evolve for 2500 generations, which was empirically found to be sufficient for the organisms to find a phenotypic target in each selective environment (when that was allowed by the developmental structure). In each run, the initial gene expression levels were uniformly chosen at random. The results here were averaged over 1000 independent runs, for each selective environment and for each of the four different evolutionary scenarios (as described in the previous sections). Then, counts of the average number of generations to reach the target phenotype of the corresponding selective environment were taken. This was evaluated by measuring the first time the developmental system achieved maximum fitness possible. If the target was not reached, the maximum number of generations 2500 was assigned.
We find that organisms with developmental organisations evolved in noisy environments or the parsimony pressure on the cost of connections adapted faster than the ones in the control scenario (Fig 5). The outliers in the evolutionary settings of moderate environmental switching, noisy environments and favouring weak connectivity, indicate the inability of the developmental system to express the target phenotypic pattern for that selective environment due to the strong developmental constraints that evolved in those conditions. This corresponds to the missing phenotype from the class we saw above in the evolved phenotypic distributions induced by development (Fig 2e, 2f and 2g). In all these three cases development allowed for the production of the same set of phenotypic patterns. Yet, developmental structures evolved in the presence of environmental noise or under the pressure for weak connectivity exhibited higher adaptability due to their higher propensity to produce other phenotypes of the structural family. In particular, we see that for the developmental process evolved under the pressure for sparsity, the rate of adaptation of the organisms was significantly improved. The variability structure evolved under sparsity to perfectly represent the functional dependencies between phenotypic traits. Thus, it provided a selective advantage guiding phenotypic variation in more promising directions.
The above experiments demonstrated the transfer of predictions from learning models into evolution, by specifically showing that: a) the evolution of generalised phenotypic distributions is dependent on the time-scale of environmental switching, in the same way that generalisation in online learning algorithms is learning-rate dependent, b) the presence of environmental noise can be beneficial for the evolution of generalised phenotypic distributions in the same way training with corrupted data can improve the generalisation performance of learning systems with the same limitations, c) direct selection pressure for weak connectivity can enhance the evolution of generalised phenotypic distributions in the same way L2-regularisation can improve the generalisation performance in learning systems, d) noisy environments result in similar behaviour as favouring weak connectivity, in the same way that Jittering can have similar effects to L2-regularisation in learning systems, e) direct selection pressure for sparse connectivity can enhance the evolution of generalised phenotypic distributions in the same way that L1-regularisation can improve the generalisation performance in learning systems, f) favouring weak connectivity (i.e., L2-regularisation) results in similar behaviour to early stopping, g) the evolution of generalised phenotypic distributions is dependent on the strength of selection pressure on the cost of connections and the level of environmental noise, in the same way generalisation is dependent on the level of inductive biases and h) in simple modularly varying environments with independent modules, sparse connectivity enhances the generalisation of phenotypic distributions better than weak connectivity, in the same way that in problems with independent features, L1-regularisation results in better generalisation than L2-regularisation.
Learning is generally contextual; it gradually builds upon what concepts are already known. Here these concepts correspond to the repeated modular sub-patterns persisting over all observations in the training set which become encoded in the modular components of the evolved network. The inter-module connections determine which combinations of (sub-)attractors in each module are compatible and which are not. Therefore, the evolved network representation can be seen as dictating a higher-order conceptual (combinatorial) space based on previous experience. This enables the evolved developmental system to explore permitted combinations of features constrained by past selection. Novel phenotypes can thus arise through new combinations of previously selected phenotypic features explicitly embedded in the developmental architecture of the system [25]. Indeed, under the selective pressure for sparse connectivity, we observe that the phenotypic patterns generated by the evolved developmental process consisted of combinations of features from past selected phenotypic patterns. Thus, we see that the ‘developmental memories’ are stored and recalled in combinatorial fashion allowing generalisation.
We see that noisy environments and the parsimony pressure on the cost of connections led to more evolvable genotypes by internalising more general models of the environment into their developmental organisation. The evolved developmental systems did not solely capture and represent the specific idiosyncrasies of past selective environments, but internalised the regularities that remained time-invariant in all environments of the given class. This enabled natural selection to ‘anticipate’ novel situations by accumulating information about and exploiting the tendencies in that class of environments defined by the regularities. Peculiarities of past targets were generally represented by weak correlations between phenotypic characters as these structural regularities were not typically present in all of the previously-seen selective environments. Parsimony pressures and noise then provided the necessary selective pressure to neglect or de-emphasise such spurious correlations and maintain only the strong ones which tended to correspond to the underlying problem structure (in this case, the intra-module correlations only, allowing all combinations of fit modules). More notably, we see that the parsimony pressure for sparsity favoured more evolvable developmental organisations that allowed for the production of a novel and otherwise inaccessible phenotype. Enhancing evolvability by means of inductive biases is not for granted in evolutionary systems any more than such methods have guarantees in learning systems. The quality of the method depends on information about past targets and the strength of the parsimony pressure. Inductive biases can however constrain phenotypic evolution into more promising directions and exploit systematicities in the environment when opportunities arise.
In this study we demonstrated that canalisation can be opposed to evolvability in biological systems the same way under- or over-fitting can be opposed to generalisation in learning systems. We showed that conditions that are known to alleviate over-fitting in learning are directly analogous to the conditions that enhance the evolution of evolvability under natural selection. Specifically, we described how well-known techniques, such as learning with noise and penalising model complexity, that improve the generalisation ability of learning models can help us understand how noisy selective environments and the direct selection pressure on the reproduction cost of the gene regulatory interactions can enhance context-specific evolvability in gene regulation networks. This opens-up a well-established theoretical framework, enabling it to be exploited in evolutionary theory. This equivalence demystifies the basic idea of the evolution of evolvability by equating it with generalisation in learning systems. This framework predicts the conditions that will enhance generalised phenotypic distributions and evolvability in natural systems.
We model the evolution of a population of GRNs under strong selection and weak mutation where each new mutation is either fixed or lost before the next arises. This emphasises that the effects we demonstrate do not require lineage-level selection [61–63]—i.e., they do not require multiple genetic lineages to coexist long enough for their mutational distributions to be visible to selection. Accordingly a simple hill-climbing model of evolution is sufficient [25, 36].
The population is represented by a single genotype [G, B] (the direct effects and the regulatory interactions respectively) corresponding to the average genotype of the population. Similarly, mutations in G and B indicate slight variations in population means. Consider that G′ and B′ denote the respective mutants. Then the adult mutant phenotype, P a ′, is the result of the developmental process, which is characterised by the interaction B′, given the direct effects G′. Subsequently, the fitness of Pa and P a ′ are calculated for the current selective environment, S. If f S ( P a ′ ) > f S ( P a ), the mutation is beneficial and therefore adopted, i.e., Gt+1 = G′ and Bt+1 = B′. On the other hand, when a mutation is deleterious, G and B remain unchanged.
The variation on the direct effects, G, occurs by applying a simple point mutation operator. At each evolutionary time step, t, an amount of μ1 mutation, drawn from [−0.1, 0.1] is added to a single gene i. Note that we enforce all gi ∈ [−1, 1] and hence the direct effects are hard bounded, i.e., gi = min{max{gi + μ1, −1}, 1}. For a developmental architecture to have a meaningful effect on the phenotypic variation, the developmental constraints should evolve considerably slower than the phenotypic variation they control. We model this by setting the rate of change of B to lower values as that for G. More specifically, at each evolutionary time step, t, mutation occurs on the matrix with probability 1/15. The magnitude μ2 is drawn from [−0.1/(15N2), 0.1/(15N2)] for each element bij independently, where N corresponds to the number of phenotypic traits.
Following the framework used in [64], we define the fitness of the developmental system as a benefit minus cost function.
The benefit of a given genetic structure, b, is evaluated based on how close the developed adult phenotype is to the target phenotype of a given selective environment. The target phenotype characterises a favourable direction for each phenotypic trait and is described by a binary vector, S = 〈s1, …, sN〉, where si ∈ {−1, 1}, ∀i. For a certain selective environment, S, the selective benefit of an adult phenotype, Pa, is given by (modified from [25]):
b = w ( P a , S ) = 1 2 1 + P a · S N , (1)
where the term Pa ⋅ S indicates the inner product between the two respective vectors. The adult phenotype is normalised in [−1, 1] by Pa ← Pa/(τ1/τ2), i.e., b ∈ [0, 1].
The cost term, c, is related to the values of the regulatory coefficients, bij ∈ B [65]. The cost represents how fitness is reduced as a result of the system’s effort to maintain and reproduce its elements, e.g., in E. coli it corresponds to the cost of regulatory protein production. The cost of connection has biological significance [27, 64–67], such as being related to the number of different transcription factors or the strength of the regulatory influence. We consider two cost functions proportional to i) the sum of the absolute magnitudes of the interactions, c = ∥ B ∥ 1 = ∑ i = 1 N 2 | b i j | / N 2, and ii) the sum of the squares of the magnitudes of the interactions, c = ∥ B ∥ 2 2 = ∑ i = 1 N 2 b i j 2 / N 2, which put a direct selection pressure on the weights of connections, favouring sparse (L1-regularisation) and weak connectivity (L2-regularisation) respectively [68].
Then, the overall fitness of Pa for a certain selective environment S is given by:
f S ( P a ) = b - λ c , (2)
where parameter λ indicates the relative importance between b and c. Note that the selective advantage of structure B is solely determined by its immediate fitness benefits on the current selective environment.
The χ2 measure is used to quantify the lack of fit of the evolved phenotypic distribution P t ^ ( s i ) against the distribution of the previously experienced target phenotypes Pt(si) and/or the one of all potential target phenotypes of the same family P(si). Consider two discrete distribution profiles, the observed frequencies O(si) and the expected frequencies E(si), si ∈ S, ∀i = 1, …, k. Then, the chi square error between distribution O and E is given by:
χ 2 ( O , E ) = ∑ i ( O ( s i ) - E ( s i ) ) 2 E ( s i ) (3) S corresponds to the training set and the test set when the training and the generalisation error are respectively estimated. Each si ∈ S indicates a phenotypic pattern and P(si) denotes the probability of this phenotype pattern to arise.
The samples, over which the distribution profiles are estimated, are uniformly drawn at random (see Estimating the empirical distributions). This guarantees that the sample is not biased and the observations under consideration are independent. Although the phenotypic profiles here are continuous variables, they are classified into binned categories (discrete phenotypic patterns). These categories are mutually exclusive and the sum of all individual counts in the empirical distribution is equal to the total number of observations. This indicates that no observation is considered twice, and also that the categories include all observations in the sample. Lastly, the sample size is large enough to ensure large expected frequencies, given the small number of expected categories.
For the estimation of the empirical (sample) probability distribution of the phenotypic variants over the genotypic space, we follow the Classify and Count (CC) approach [69]. Accordingly, 5000 embryonic phenotypes, P(0) = G, are uniformly generated at random in the hypercube [−1, 1]N. Next, each of these phenotypes is developed into an adult phenotype and the produced phenotypes are categorised by their closeness to target patterns to take counts. Note that the development of each embryonic pattern in the sample is unaffected by development of other embryonic patterns in the sample. Also, the empirical distributions are estimated over all possible combinations of phenotypic traits, and thus each developed phenotype in the sample falls into exactly one of those categories. Finally, low discrepancy quasi-random sequences (Sobol sequences; [70]) with Matousek’s linear random scramble [71] were used to reduce the stochastic effects of the sampling process, by generating more homogeneous fillings over the genotypic space.
|
10.1371/journal.pgen.1007717 | Quantifying how constraints limit the diversity of viable routes to adaptation | Convergent adaptation occurs at the genome scale when independently evolving lineages use the same genes to respond to similar selection pressures. These patterns of genetic repeatability provide insights into the factors that facilitate or constrain the diversity of genetic responses that contribute to adaptive evolution. A first step in studying such factors is to quantify the observed amount of repeatability relative to expectations under a null hypothesis. Here, we formulate a novel index to quantify the constraints driving the observed amount of repeated adaptation in pairwise contrasts based on the hypergeometric distribution, and then generalize this for simultaneous analysis of multiple lineages. This index is explicitly based on the probability of observing a given amount of repeatability by chance under a given null hypothesis and is readily compared among different species and types of trait. We also formulate an index to quantify the effective proportion of genes in the genome that have the potential to contribute to adaptation. As an example of how these indices can be used to draw inferences, we assess the amount of repeatability observed in existing datasets on adaptation to stress in yeast and climate in conifers. This approach provides a method to test a wide range of hypotheses about how different kinds of factors can facilitate or constrain the diversity of genetic responses observed during adaptive evolution.
| How many ways can evolution solve the same adaptive problem? While convergent adaptation is evident in many organisms at the phenotypic level, we are only beginning to understand how commonly this convergence extends to the genome scale. Quantifying the repeatability of adaptation at the genome scale is therefore critical for assessing how constraints affect the diversity of viable genetic responses. Here, we develop probability-based indices to quantify the deviation between observed repeatability and expectations under a range of null hypotheses, and an estimator of the proportion of loci in the genome that can contribute to adaptation. We demonstrate the usage of these indices with individual-based simulations and example datasets from yeast and conifers and discuss how they differ from previously developed approaches to studying repeatability. Because these indices are unitless, they provide a general approach to quantifying and comparing how constraints drive convergence at the genome scale across a wide range of traits and taxa.
| If different species encounter the same selection pressure, will adaptive responses occur via homologous genes or follow distinct genetic routes to the same phenotype? What factors limit the diversity of viable genetic routes to adaptation and how does variation translate into evolution? Empirical studies have identified different amounts of convergent adaptation at the genome scale across a range of species, traits, timescales, and levels of developmental-genetic hierarchy [1–3]. When evolution uses the same genes repeatedly to generate a given trait value, is this because of constraints acting on the genetic and developmental pathways limiting the production of variation (i.e., there is only a limited number of ways to generate a given trait value), or because of fitness constraints acting on genotypes that yield the same trait value (i.e., only some genotypes are selectively optimal)? Note that “constraint” in the evolutionary literature is commonly invoked to refer to factors that limit an adaptive phenotypic response in general (e.g., [4,5]). Here, we use it to refer to the factors that limit the diversity of genes used in independent bouts of adaptation and use the term “diversity constraint” hereafter for clarity.
As a case study to examine the differing types of diversity constraint, Mc1r provides perhaps the most well-known example of convergent local adaptation at the gene scale, and has been implicated in driving colour pattern variation in mice, lizards, mammoths, fish, and a range of other organisms [6–10]. Extensive studies in mice have revealed that over 50 genes can be mutated to give rise to variation in colour pattern [11], yet Mc1r consistently tends be one of the main contributors to locally adapted colour polymorphisms. Mc1r has minimal pleiotropic side effects [8,11] and it can mutate to a similar trait value through numerous different changes in its protein sequence [10,11] and therefore may have a higher rate of mutation to beneficial alleles than other genes. As such, it seems to be driven by a combination of both types of constraint: more ways to mutate via Mc1r implies that developmental-genetic constraints limit the contribution of other genes, while limited pleiotropy in Mc1r implies that fitness costs constrain which genes can yield mutations that provide a viable route to adaptation.
Are the diversity constraints acting on melanism representative of the kinds of constraints that shape patterns of genome-scale convergence and non-convergence across the tree of life? To answer this question, it is necessary to quantify the extent of genome-scale convergence in a wide variety of organisms and traits and ascertain what kind of diversity constraints are operating, which requires the development of an appropriate statistical framework. To this end, it is helpful to frame the above questions based on the flexibility of the mapping from genotype to trait to fitness: does repeatability occur because of low redundancy in the mapping of genotype to trait (hereafter low GT-redundancy [4]: only a few ways to make the same trait value; Fig 1A), or because of low redundancy in the mapping of genotype to fitness? (hereafter low GF-redundancy: only a subset of the genotypes yielding the same trait value are optimal; Fig 1B).
GT-redundancy is determined by two factors: 1) the difference between the number of genes that need to mutate to yield a given trait value and the number of genes that could mutate to give rise to variation in the trait, and 2) the extent to which different genes have interchangeable vs. uniquely important effects on the phenotype. High GT-redundancy means that many different combinations of alleles can yield the same trait value, so if all else is equal, then independent bouts of adaptation are likely to occur via different sets of mutations and repeatability will be low [12,13]. The standard quantitative genetic model implicitly assumes complete GT-redundancy with fully interchangeable allelic effects, while the recently proposed omnigenic model assumes high but incomplete redundancy, with “core” vs. “peripheral” genes having different potential to affect variation [14].
GF-redundancy is determined by differences in fitness among genotypes that produce the same trait value and can increase the diversity constraints driving repeatability above the level incurred by GT-redundancy. Such differences in fitness can occur when mutations cause correlated effects on other traits that also affect fitness, such that not all mutations that are equally suitable for adaptation (i.e., pleiotropy), or when interactions among particular mutations have negative fitness effects, such that only particular combinations of mutations tend to contribute to adaptation (i.e., epistasis). It is also possible for effects on fitness to arise independent of a phenotypic effect, because architectures with different allele effect sizes and linkage relationships can have different fitness depending on the interaction between migration, selection, and drift. For example, if a given phenotype is coded by many small unlinked alleles, this architecture would be less fit than a similar phenotype coded by a few large or tightly linked alleles, in the context of migration-selection balance [15] or negative frequency dependence [16,17]. Similarly, the increased drift that occurs in small populations may prevent alleles of small effect from responding to natural selection [18,19], resulting in such genotypes being effectively neutral and therefore lower in realized fitness than those made up of large-effect alleles. For example, polygenic models of directional selection (e.g., [20]) assume no GT- and GF-redundancy, while traditional quantitative genetic models of Gaussian stabilizing selection assume high GT- and GF-redundancy (e.g., [21]).
In addition to GT- and GF-redundancy, other factors also impact signatures of convergence, such as differences among genes in recombination rate or propensity to retain standing variation. There has been considerable discussion in the literature about the effects of these and other factors on convergence [3,23–30], and various indices have been previously used to quantify repeatability in empirical contexts (e.g. Jaccard index, Proportional Similarity; [1,2]). These existing indices provide a useful description of how often the same gene is used in adaptation, but as we will show below, they are not well-suited for testing of hypotheses to discriminate between these different kinds of constraint. They do not incorporate information about the genes that could contribute to adaptation but don’t, which is necessary to evaluate what kinds of diversity constraints are operating, and they are not explicitly tied to the probability of repeatability occurring under a null model.
Here, we develop novel statistical approaches for quantifying the diversity constraints that drive repeatability in genomic data from studies of local adaptation and experimental evolution. To study these constraints, we formulate an explicit probability-based representation of the deviation of observed repeatability from expectations under different null hypotheses. This approach can be used after standard tests have been applied to identify the putative genes driving adaptation and uses as input either binary categorization of genes as “adapted” or “non-adapted” or any continuous index representing the relative amount of evidence for a given gene being involved in adaptation (e.g. FST, p-values, Bayes factors). We begin by formulating an analytical model for a contrast of two lineages with binary data, and then generalize this model for contrasts of multiple lineages using either binary or continuous data. We also propose a novel index estimating the proportion of genes in the genome that can potentially give rise to adaptation. In all cases, these models can be used to successively test null hypotheses that incorporate different amounts of information about the constraints that could shape repeatability. The simplest null hypothesis is that there are no constraints and all genes have equal probability of contributing to adaptation. If more repeatability is observed than expected under this null model, then two inferences can be made: natural selection is driving patterns of convergence (and that observed signatures are not all false positives), and some diversity constraints are operating to increase the repeatability of adaptation. We then consider how other null hypotheses can be formulated to represent the various kinds of constraints discussed above. We focus mainly on the effect of low GT-redundancy, where the number of genes that could potentially contribute to adaptation is much smaller than the total number of genes in the genome, but also discuss how constraints arising from GF-redundancy, standing variation, or mutation rate could be modeled. Because this method quantifies repeatability in terms of probability-scaled deviations from expectations, it can be applied across any trait or species of interest, allowing contrasts to be made on the same scale of measurement.
Suppose there are two lineages, x and y, that have recently undergone adaptation to a given selection pressure, resulting in convergent evolution of the same trait value within each lineage. This adaptation could be global, with new mutations fixed within lineages (e.g., in experimental evolution studies with multiple replicate populations), or local, with mutations contributing to divergence among populations within each lineage (e.g., in observational studies of natural adaptation to environmental gradients). In either case, we assume that adaptation can be reduced to a binary categorization of genes as “adapted” or “non-adapted” to represent which genes contribute to fitness differences (either relative to an ancestor or a differently-adapted population). We use the following notation to represent different properties of the genomic basis of trait variation: the number of loci in the genome of each species is nx, and ny, with the number of orthologous loci shared by both species being ns; the adaptive trait is controlled by gx and gy loci in each species, with gs shared loci (i.e. the loci in which mutations will give rise to phenotypic variation in the trait, hereafter the “mutational target”); of the g loci that give rise to variation, only a subset have the potential to contribute to adaptation due to the combined effect of all constraints, represented by gax and gay, with gas shared loci (the “effective adaptive target”); in a given bout of adaptation, the number of loci that contribute to adaptation in each lineage is ax and ay, with as orthologous loci contributing in both lineages. For simplicity, we assume that there is complete overlap in the genomes (ns = nx = ny), mutational targets (gs = gx = gy), and loci potentially contributing to adaptation (gas = gax = gay) in both species (see supplementary materials and S1 Fig for set notation). These assumptions are most appropriate for lineages that are relatively recently diverged, where most orthologous genes are retained at the same copy number and the developmental-genetic program is relatively conserved, so that the same genes potentially give rise to variation in both lineages. Lineages separated by greater amounts of time would be expected to have reduced ns due to gene deletion, duplication, and pseudogenization in either lineage, and reduced gs and gas due to evolution and divergence of the developmental-genetic program, through sub- and neo-functionalization, and divergence in regulatory networks.
Under the assumption that all gas genes have equal probability of contributing to adaptation (i.e., no diversity constraints are operating), the amount of overlap in the complement of genes that are adapted in both lineages (as) is described by a hypergeometric distribution where the expected amount of overlap is ās = axay/gas (e.g. [31]). In practice, we typically have little prior knowledge about which genes have the potential to contribute to either adaptation (gas) or standing variation in the trait (gs), but we can draw inferences about how these parameters constrain the diversity of adaptive responses by testing hypotheses and comparing the observed amount of overlap (as) to the amount expected under a given null hypothesis (ās). To test different hypotheses about how diversity constraints give rise to repeated adaptation, we represent the total number of genes included in the test set as g0. The simplest null hypothesis is that there are no diversity constraints and all genes potentially give rise to variation and contribute to adaptation (g0 = gas = gs = ns), so by rejecting this null, we can infer that gas < ns, and calculate an effect size that represents the magnitude by which all types of constraints contribute to repeatability (see Eq 1, below; note that it is also possible that the null hypothesis could be falsified in the opposite direction, with less overlap in the loci contributing than expected under the null, which might occur if evolution had occurred towards a different optimum in each lineage). Without independent lines of evidence about which genes potentially contribute to variation in the trait (gs), it is not possible to evaluate the relative importance of GT- vs. GF-redundancy using the framework here. In model systems where independent information is available for the magnitude of gs (based on mutation accumulation or GWAS; see Discussion), then a more refined null hypothesis can be tested, where g0 = gs, allowing some inferences to be made about the relative importance of GT- and GF-redundancy (Table 1). By rejecting this null, we can infer that gas < gs, which could occur due to low GF-redundancy or differences among genes in mutation rate or standing variation. Alternatively, if we fail to reject this null hypothesis, then it suggests that gs ≅ gas, which would imply that GF-redundancy doesn’t make any additional contribution to repeatability beyond the contribution of GT-redundancy. We can also reverse the direction of inquiry and estimate gas directly from the data by calculating ga^s=axay/as, such that an index representing the effective proportion of the genome that can potentially contribute to adaptation can be calculated as P^a,hyper=axay/(asns).
For any value of g0, an effect size representing the excess in overlap due to convergence relative to the null hypothesis can be expressed by standardizing the observed overlap by subtracting the mean (ās = axay/g0) and dividing by the standard deviation of the hypergeometric distribution:
Chyper=(as−(axayg0))/(axay)(g0−ax)(g0−ay)/(g02(g0−1)).
(1)
This index provides a quantitative representation of how much more overlap occurs than expected under the null hypothesis, scaled according to how much a given bout of evolution would deviate from this expectation if the null hypothesis were true. Similarly, the exact probability of observing as or more shared loci contributing to adaptation can also be calculated using the hypergeometric probability (see Supplementary Information for sample R-script), which provides a p-value.
While pairwise contrasts are most straightforward statistically, they have considerably lower power than comparisons among multiple lineages. If one gene (such as Mc1r) tends to drive adaptation repeatedly in a large number of lineages, this may go undetected in an approach using multiple pairwise comparisons but would be readily detected in a simultaneous comparison of multiple lineages. Unfortunately, while the hypergeometric distribution provides an exact analytical prediction for the amount of overlap in a pairwise comparison, which can be used to calculate a p-value and the probability-based effect size (Chyper), it cannot be easily generalized to simultaneously analyze multiple lineages. While it is possible to conduct pairwise analysis and average the results across multiple comparisons, p-values from this approach might fail to detect cases where a single gene contributes repeatedly to adaptation in more than two lineages, as information does not transfer among the pairwise comparisons. We now develop an alternate, approximate approach to assess repeatability in multiple lineages by calculating Pearson’s χ2 goodness of fit statistic and comparing this to a null distribution of χ2 statistics simulated under the null hypothesis to calculate a p-value as the proportion of replicates in the null that exceed the observed test statistic. The p-value obtained by this approach represents the probability of observing a test statistic as extreme or more extreme under the null hypothesis, considering all lineages simultaneously. While the p-value is calculated from simultaneous analysis of all lineages, the effect size is instead calculated as an average across all pairwise comparisons among the k replicate lineages, because this represents the increase in repeatability relative to expectations under the null for a given bout of adaptation in any single lineage. This difference is important because the effect size should not depend on sampling effort in terms of the number of lineages, while the p-value should reflect the statistical power gained from multiple lineages.
Consider the case where g0 genes can potentially contribute to adaptation in the given trait and each lineage has some complement of genes that have mutated to drive adaptation, with αi,j representing the binary score for gene i in lineage j (1 = adapted, 0 = non-adapted). The summation for gene i across all lineages provides the observed counts (oi = Σjαi,j) while the expected counts (ei) can be set based on the null hypothesis being tested. Under null hypotheses where all genes in g0 have equal probability of contributing to adaptation, the expected counts are equal to the mean of the observed counts (e = Σioi/g0), and Pearson’s χ2 statistic can be calculated by the usual approach: χ2 = Σ(o − e)2/e. Under ideal conditions, Pearson’s χ2 would approximate the analytical χ2 distribution with its mean and standard deviation equal to the degrees of freedom (df) and 2df, respectively. While this could be used to make an analytical hypothesis test (as above), in practice there will often be large deviations between Pearson’s χ2 and the analytical distribution, due to violation of the assumptions when expected counts are low (See Supplementary Materials, S2 Fig). Instead, we simulate a null distribution of χsim2 values under the null hypothesis by using permutation within each lineage and recalculating χsim2 for each replicate. The p-value is then equal to the proportion of the χsim2 values that exceed the observed χ2 (using all lineages simultaneously), while the effect size is calculated as the mean C-score across all pairwise contrasts (simulating χsim2 for each pairwise contrast):
Cchisq=χ2−mean(χsim2)sd(χsim2).
(2)
The magnitude of Cchisq therefore represents deviation between the observed amount of repeatability and that expected under the null hypothesis, which will vary as a function of the diversity constraints affecting the trait evolution, but not the number of lineages being compared. While Cchisq relies upon simulation of a null distribution, it can be calculated relatively quickly. Importantly, the magnitude of Cchisq varies linearly with Chyper (Fig 2A & 2B), showing that it represents the extent of diversity constraints in the same way as the analytically precise Chyper. While this approach provides a more accurate p-value for comparisons of multiple lineages, there is no particular reason to use Cchisq rather than Chyper for binary input data, as both effect sizes are calculated on a pairwise basis. The main reason that we develop this approach is to extend it to continuously distributed data, which can allow greater sensitivity and avoid arbitrary choices necessary to categorize the commonly used indices of local adaptation (e.g. FST or p-values) into “adapted” or “non-adapted”.
In many empirical contexts, genome scans for selection yield continuously distributed scores representing the strength of evidence for each locus contributing to adaptation (e.g., FST, p-values, Bayes factors). Using the same notation as above, but with αi,j representing the continuous score for the ith gene in the jth lineage, the total score for each gene can be calculated as a sum across lineages, αi¯=∑jkαi,j, while the mean score over all genes and lineages is α¯¯=∑ig0α¯i/g0. A statistic analogous to the above χ2 can then be calculated as χ2=∑(α¯i−α¯¯)2/α¯¯, and the same approach for calculating the null distribution of this statistic can then be used to calculate Cchisq according to Eq 2.
With continuous data, there are additional complexities that arise depending on the distribution of the particular dataset being used and how its magnitude represents evidence for a gene’s involvement adaptation. One approach, which we used in all examples here, is to transform data so that values scale positively and approximately linearly with the weight of evidence for adaptation, by standardizing data within each lineage by subtracting their observed within-lineage minimum and dividing by their observed within-lineage maximum, such that the values within each lineage are bounded from 0 to 1. This reduces differences among lineages in the absolute magnitude of indices representing adaptation, which can be desirable when they vary across many orders of magnitude (e.g. p-values from GWAS of 10−10 and 10−20 both provide strong evidence of adaptation). However, if some lineages actually have stronger signatures of adaptation at more loci, then this kind of standardization should not be used, as it would obscure these true differences among lineages. In this case, it would be preferable to use the same standardization across all lineages by subtracting the minimum and dividing by the maximum values observed across all lineages. While Pearson’s χ2 statistic was designed for discrete data, the above approach using continuous data represents the variability among lineages in the same way, as a variance among genes in the sum of their scores representing putative adaptation. The Cchisq statistic on continuous data behaves similarly to the Chyper statistic across wide ranges of parameter space, as both are formulated in terms of deviations from the null distribution (see below).
While the number of genes that potentially contribute to adaptation (gas) can be estimated using the hypergeometric equation, ga^s=axay/as, it is difficult to apply this to comparisons of multiple lineages, as some pairwise contrasts may have no overlap in the genes contributing to adaptation (as = 0), making the equation undefined. To estimate ga^s from all lineages simultaneously, we can instead formulate a likelihood-based approach where the probability that we observe locus i adapted in oi lineages is:
ζi=Pa*Bin(k,o¯,oi)+(1−Pa)Bin(k,0,oi),
(3)
where Bin(n,y,x) is the probability under the binomial distribution of getting x successes in n trials, each with probability y. As above, oi is the number of adapted genes in k lineages (with oi = Σjαi,j), Pa is the proportion of g0 that can actually contribute to adaptation (Pa = gas/ns), and ō is the probability of each gene contributing to adaptation ō = Σoi/(gas k). The estimated value of ga^s is then the value at which the log-likelihood function:
L(gas)=∑logζi
(4)
is maximized. Once the maximum-likelihood value of ga^s is estimated, this can be expressed either as an absolute number representing the effective number of genes that can contribute to adaptation or as a proportion of the total number of shared genes in the genome: P^a,lik=ga^s/ns. This approach implicitly assumes that all genes that have the potential to contribute to adaptation (gas) have approximately equal probabilities of actually contributing to adaptation. In very extreme cases, such where one gene is very highly repeatable while other genes only contribute to adaptation in a single lineage, ga^s will tend to represent the contribution of the repeatable genes and discount the contribution of the idiosyncratic genes (see Supplementary Materials). Multi-class models could be developed to estimate gas for different classes of genes in such scenarios by accounting for their different probabilities of contributing to adaptation (See https://github.com/samyeaman/dgconstraint for scripts containing functions for the above calculations).
The Chyper, Cchisq, and P^a,lik estimators capture different aspects of the biology underlying convergence than other previously used estimators of repeatability. To estimate the repeatability of evolution, Conte et al. [1] used the additive and multiplicative Proportional Similarity (PSadd and PSmult) indices of [32] in a meta-analysis of QTL and candidate gene studies, while Bailey et al. [2] used the Jaccard Index to quantify patterns in bacterial evolution experiments. The PS indices are defined as PSadd = Σ min(αix, αiy) and PSmult=∑(αixαiy)/∑(αix)2(αiy)2, where αix and αiy are the relative contribution of gene i to adaptation in lineages x and y [33], while the Jaccard index is defined as J = (Ax ∩ Ay)/(Ax ∪ Ay), where Ax and Ay are the sets of adapted genes in each lineage [2]. Both of these indices are based on standardizing the number of overlapping adapted loci by the total number of adapted loci, and neither includes information about non-adapted genes that potentially could have contributed to adaptation.
To illustrate the differences between these various indices of convergence, we generated four example datasets showing either randomly drawn complements of genes with adapted mutations (Fig 3A) or highly convergent datasets drawn from a smaller (Fig 3B) or larger (Fig 3C & 3D) pool of genes that potentially contribute to trait variation (gs), with differing numbers of loci contributing to adaptation. Scenario C is the most constrained, as it exhibits the same amount of overlap as B, but this overlap is drawn from a larger pool of genes so it is less likely to occur by chance. While neither the Jaccard index nor the PS indices distinguish between the B, C, and D scenarios (as the same proportions of genes are being used for adaptation, so repeatability is the same), both the Cchisq and Chyper indices show the highest scores for scenario C, because it has the smallest probability of occurring by chance if all genes had equal probabilities of contributing to adaptation. The P^a,lik index also identifies scenario C as most constrained in terms of the smallest proportion genes potentially contributing to adaptation. The P^a,lik index also shows that this proportion is equal for scenarios B & D, despite differences in the probability of the observed repeatabilities occurring by chance (as per the C-scores). More generally, while P^a,lik tends to decrease with increasing C-score, these indices differ in magnitude (Fig 2C & 2D), as they represent different aspects of diversity constraints. In summary, the Jaccard and PS indices quantify the proportion of genes used for adaptation that are used repeatedly, the C-score indices are inversely proportional to the probability of the observed repeatability occurring if there were no constraints, and P^a,lik represents the proportion of genes in the genome that are available for adaptation, given the existing diversity constraints (also see S3 Fig for further comparisons).
To further explore the effect of population genetic parameters on the behaviour of the above indices of repeatability and constraint, we used Nemo (v2.3.45; [34]) to simulate two scenarios of two-patches under migration-selection balance: (i) constant size of mutational target with variable proportions of small- and large-effect loci; and (ii) constant number of large-effect loci and variable number of small effect loci, resulting in a variable size of mutational target. For scenario (i), simulations had n = gs = 100 loci, of which u loci had alleles of size +/- 0.1, while (100 − u) loci had alleles of size +/- 0.01 (with subsequent mutations causing the allele sign to flip from positive to negative or the reverse). For scenario (ii), simulations had 10 large-effect loci with alleles of size +/- 0.1 and v small-effect loci with alleles of size +/- 0.01, resulting in a variable size of mutation target. In all simulations, migration rate was set to 0.005 and the strength of quadratic phenotypic selection was 0.5, so that an individual perfectly adapted to one patch would suffer a fitness cost of 0.5 in the other patch (patch optima were +/- 1; similar to [13]). Simulations were run for 50,000 generations and censused every 100 generations. For binary categorization of the input data, loci were considered to be “adapted” if FST > 0.1 for >80% of the last 25 census points (these cut-offs are somewhat arbitrary, but qualitative patterns were comparable under different cut-offs); for continuous input data, raw FST values were used. Results are averaged across 20 runs, each with 20 replicates, with Cchisq calculated across the 20 replicates within each run.
These scenarios further illustrate the difference between the Jaccard and PSadd indices of repeatability and the C-score and P^a,lik indices of constraint. In both scenarios, the small effect loci do not tend to contribute much to adaptation because large effect loci are more strongly favoured under migration-selection balance [35], which results in low GF-redundancy. In scenario (i), all indices show qualitatively similar patterns, with decreasing repeatability occurring as a result of the decreasing constraints that occur as the number of large-effect loci increases, increasing the GT- and GF-redundancy (Fig 4A). By contrast, in scenario (ii), the Jaccard and PSadd indices indicate that roughly the same amount of repeatability is occurring regardless of the number of small effect loci and total size of mutational target (Fig 4B). However, over this same range of parameter space, the C -score indices show that constraint increases as the total mutational target is increasing. This occurs because while a larger number of potential routes to an adaptive phenotype are available with increasing number of small effect loci, only the same small number of loci are actually being involved in adaptation (i.e. the large effect loci), which is illustrated by the decrease in the P^a,lik index. While there are many potential genetic routes to adaptation that could involve these small effect loci (high GT-redundancy), the large effect loci tend to be favoured and repeatedly involved in adaptation (low GF-redundancy). Thus, when the size of the mutational target increases in scenario (ii), the repeatability tends to stay about the same (Jaccard and PSadd) but the amount of constraint is higher (C-scores), because a smaller proportion of the available routes to adaptation are being used (P^a,lik). The continuous and binary Cchisq indices are broadly similar across these parameters because there is very little variation in FST among loci within the same size class (see Supplementary Materials for additional simulations under varying allele effect sizes).
The amount of constraint quantified by the C-score will depend upon the proportion of the mutational target (gs) that is sampled by the sequencing approach, which should be proportional to the sampling of the total number of genes in the genome (ns). Some approaches, such as targeted sequence capture, will sample only a subset of the total number of genes in the genome, which will therefore cause a bias in the estimation of constraint due to incomplete sampling, even if the genes included are a random subset of gs. This can be most clearly seen in the calculation of Chyper, where multiplying all the variables in Eq 1 by a given factor will cause a change in the magnitude of the effect size. By contrast, the Jaccard and PS measures of repeatability are not affected by incomplete sampling. If binary input data are being used and the proportion of gs that has been sampled can be accurately estimated (q), then the calculation of Chyper can be corrected by dividing all input variables by q prior to calculation, yielding a corrected score Chyper-adj. If continuously distributed input data are being used, then the dataset can be adjusted by adding g0 (1 − q) new entries to the dataset by randomly sampling genes with replacement from the existing dataset, and then applying Eq 2 to this extended set.
To explore the effect of incomplete sampling of the genome on the calculation of C-scores and the impact of these types of correction, we constructed a test dataset by concatenating 5 replicates from the simulations in Fig 4A with 10 large effect loci, yielding a dataset with 500 loci in total and a high amount of repeatability. We then sampled a proportion q of this total dataset to simulate incomplete representation of the genome and used the above approach calculate uncorrected and corrected C-scores. While incomplete sampling can cause considerable bias in C-scores, as long as q is not too small, these approaches yield relatively accurate corrections of these estimates (Fig 5). At very low values of q, the variance in estimation among replicate subsets increases as a result of sampling effects when only a small number of adapted loci are included, but on average the magnitude of the corrected C-score is independent of q.
Experimental evolution studies provide a controlled framework to test theories on the genetic basis of adaptation under a diversity of scenarios. Gerstein et al. (2012) previously conducted an experiment to examine the diversity of first-step adaptive mutations that arose in different lines initiated with the same genotypes in response to the antifungal drug nystatin [36] and in response to copper [37]. The design allowed them to directly test how many different first-step solutions were accessible to evolution when the same genetic background adapted to the same environmental stressor. In the nystatin-evolved lines they identified 20 unique and independently evolved mutations in only four different genes that act in the nystatin biosynthesis pathway: 11 unique mutations in ERG3, seven unique mutations in ERG6, and one unique mutation in each of ERG5 and ERG7 [36]. The genotypic basis of copper adaptation was broader, and there were both genomic (SNPs, small indels) and karyotypic (aneuploidy) mutations identified. If we consider just the genomic mutations, mutations were found in 28 different genes, with multiple mutations identified in four genes (12 unique mutations in VTC4, four unique mutations in PMA1, and three unique mutations in MAM3 and VTC1). If we assume that all genes in the genome could potentially contribute to adaptation (i.e. g0 = 6604), then Chyper-nystatin = 32.5, while Chyper-copper = 12.3, and p < 0.00001 in both cases.
If we assume much lower GT-redundancy and that only the observed genes could possibly contribute to the trait (i.e. g0-nystatin = 4, g0-copper = 28), we can test whether the mutations are still more clustered than expected within these sets. Using the methods outlined above, we find Chyper-nystatin = 0.35, p = 0.002, and Chyper-copper = 0.43, p < 0.0001, indicating that even under the severe developmental-genetic constraints to diversity represented by this model, these data are slightly more overlapping than expected at random, likely due low GF-redundancy and potentially gene-specific differences in mutation rate (because these experiments were initiated using isogenic strains, standing variation was precluded).
Experimental evolution studies lend themselves nicely to future hypothesis testing about the impact of constraint on the genetic basis of adaptation and provide us with hypotheses about differences between the genes that were and were not observed in the screen. For example, we parsed the Saccharomyces Genome Database (http://www.yeastgenome.org) for genes that have been annotated as “resistance to nystatin: increased”, where this phenotype is conferred by the null mutation. This should be a conservative dataset, as we also expect there could be mutations in additional genes that do not result in a loss-of-function phenotype that could also confer tolerance to nystatin (although we expect that the mutations we recovered in ERG3, ERG5 and ERG6 are all similar to loss of function mutations, ERG7 is inviable when null [36]). This identified an additional five genes (KES1, OSH2, SLK19, VHR2, YEH2). We can test whether the five genes without an observed mutation have a negative pleiotropic effect when null or are in areas of the genome with a lower mutation rate compared to the ERG genes (particularly compared to ERG3 and ERG6). Similar experiments could also be conducted with different Saccharomyces cerevisiae genetic backgrounds, with closely related species, or under slightly different environmental conditions (e.g., increased or decreased concentration of stress) to directly examine how different aspects of the genomic and ecological environments influence the observed level of constraints acting on adaptation.
Lodgepole pine and interior spruce both inhabit large ranges of western North America and display extensive local adaptation, with large differences in cold tolerance between northern and southern populations in each species. Recent work studied the strength of correlations between population allele frequencies and a number of environmental variables and phenotypes in each species [38]. Taking one representative environmental variable as an example, a total of 50 and 121 single-copy orthologs showed strong signatures of association to Mean Coldest Month Temperature (MCMT) in pine and spruce, respectively, with 5 of these genes overlapping (based on binary categorization using the binomial cutoff “top candidate” method, as per [38]). This study included a total of 9891 one-to-one orthologs with sufficient data in both species (i.e. at least 5 SNPs per gene), so observing 5 genes overlapping corresponds to Chyper = 5.6 and p = 0.00034 under the null hypothesis that all genes had equal potential to contribute to adaptation. Alternatively, it is also possible to estimate Cchisq on continuously distributed data by calculating top candidate scores for each gene using the binomial probability of seeing u outliers when there are v SNPs in a given gene, with an overall rate of w outliers per SNP (as per [38], this yields an index rather than an exact probability, due to linkage among SNPs). This approach is more sensitive to weak signatures of adaptation that occur below the binary categorization cutoff, yielding Cchisq = 5.1 and p < 0.00001. Assuming that the 9891 studied genes represent a random sample from approximately 23,000 genes in the whole genome and ignoring divergence in gene content between species (ns = nx = ny), the adjusted C-scores are Chyper-adj = 8.6 and Cchisq-adj = 7.8 (with resampling of 50 replicates and 10,000 permutations per replicate), providing a very rough estimate of the total diversity constraints driving repeatability.
As it is possible that some factors such as conservation of the genomic landscape of cold- and hot-spots of recombination could spuriously drive signatures of convergence (see Discussion), it is possible to perform a basic control by comparing the above results for pine-MCMT vs. spruce-MCMT to the overlap between top candidates for different environments in each species, where convergence would not be expected. Examining the variable least strongly correlated to MCMT (annual heat-moisture index; AHM), we find 23 top candidates in spruce with one overlap with pine-MCMT top candidates and 25 in pine with no overlaps with spruce-MCMT, which correspond to p = 0.11 and p = 1, respectively. Thus, there was no significant increase in overlap among the top-candidates for different variables, where signatures of convergence are unexpected but could have been generated spuriously by some combination of demography and low recombination in some regions of the genome. However, as this is a negative control, the lack of a significant result does not prove that such effects are absent, so caution is necessary when drawing inferences from these data.
These diversity constraints correspond to an effective adaptive target of gas = 1462 genes that could potentially contribute to adaptation in these species (out of 9891), which yields P^a,lik=0.15. However, a large number of the 50 and 121 genes identified using their “top candidate test” were likely false positives, because there were no controls for population structure during the association test, as this was subsequently accounted for by the among-species comparison. Thus, if we assume a 50% false positive rate for ax and ay, then gas declines to 370 genes, with P^a,lik=0.037. In their analysis, Yeaman et al. [38] used another more sensitive test (null-W) to identify loci with signatures of convergence that were not detected based on overlap in the top candidates lists, which suggests the true amount of repeatability may be higher than inferred here. This example illustrates how these kinds of statistics may be used to make inferences about constraints, but also highlights the sensitivity of the results to small changes in parameters.
The methods developed here provide a way to estimate the effective number of loci that can potentially contribute to adaptation and an index to quantify the total amount that all constraints contribute to repeatability relative to the null hypothesis of no constraints (“C-score”). Importantly, these statistics can be used to contrast the total constraints affecting convergence in highly divergent traits and species. The comparison between antimicrobial and copper resistance in yeast vs. cold tolerance in conifers suggests that the latter trait is less constrained than the former (Cantimicrobial = 32.5; Ccopper = 12.3; Cclimate = 7.8), but that in all cases considerably more repeatability in adaptation is seen than under a model with no constraints. As C-scores are scaled by the standard deviation of the probability distribution, their magnitude scales linearly with decreasing probability of the observed repeatability occurring by chance under the null model. In conjunction with other information about number of loci that could potentially give rise to standing variation (gs), this method can be used to test hypotheses about whether observed repeatability is due to GT- or GF-redundancy, or other confounding factors. We now review how these methods can be used to draw inferences and discuss potential problems that should be considered in their implementation.
Under the simplest null hypothesis that there are no diversity constraints, all genes can give rise to potentially adaptive variation in the trait (g0 = gs = gas = ns). While simplistic, this approach provides an intuitive method to assess whether the amount of convergence observed is more than expected due to pure randomness. But what do we learn if we reject such a simple null hypothesis? Two inferences can be drawn in this case: many of the genes flagged by our tests for selection are likely evolving by natural selection (i.e., they are not all false positives) and some kind of constraint is involved in shaping this adaptation. The former inference means that analyzing comparative data for convergence can provide a powerful tool for identifying the genes involved in local adaptation, as this is often a significant methodological hurdle in evolutionary biology (e.g., [38]). The latter inference may seem a straw-man, as few molecular biologists would advocate a model where every gene can mutate to give rise to adaptively useful variation in a given trait. However, different forms of the “universal pleiotropy” model have been assumed in theoretical quantitative genetics [39], and the recently proposed “omnigenic model” advocates extensive pleiotropy [14]. Regardless of the true number of genes involved, this null hypothesis provides a benchmark against which we can quantify how all factors constraining the diversity of forms combine to drive repeatability, which is useful for interpreting patterns of repeatability among species and traits.
In order to make inferences about the potential importance of different kinds of diversity constraints driving repeatability, it is necessary to specify more realistic models for the evolution of local adaptation that incorporate different assumptions about size of the mutational target of the trait, extent of shared standing variation, differences in mutation rate among genes, distribution of mutation effect sizes, and species demography. The simplest modification to the above null model is to represent the extent of GT-redundancy by specifying the number of loci that potentially contribute to trait variation as a subset of the total number of loci in the genome (g0 = gs < ns). In the context of the Chyper index (Eq 1), reducing g0 increases both the mean and standard deviation of the hypergeometric distribution and therefore decreases Chyper and the inferred level of residual (unexplained) constraints. If empirical estimates of gs result in Chyper ~ 0, then it is reasonable to conclude that low GT-redundancy is mainly responsible for the observed amount of convergent adaptation. This would not discount the importance of natural selection overall, as selection on the phenotype is still responsible for adaptation but would suggest that gs individual loci are more or less interchangeable and GF-redundancy contributes no additional constraints above those imposed by GT-redundancy (Table 1). However, as we have few (if any) conclusive estimates of gs in highly polygenic traits [40,41], the extent of constraint arising through low GT-redundancy will be difficult to assess without further directed study. Although they are by no means simple experiments to conduct, it should be possible to estimate gs from QTLs identified in multiple mutation accumulation experiments, as the number of loci detected across all experiments should asymptote towards gs, and rarefaction designs could be used to estimate gs based on the overlap between QTLs detected in two experiments (although this would likely still be biased by failing to detect loci of small effect). A similar approach to the study of repeatability in adaptive loci taken here could be applied to multiple GWAS results on standing variation for a given trait conducted independently in different species to assess the proportion of shared loci. However, it should be noted that in this case, the loci that contribute to standing variation could be shaped by previous selection and might therefore be more convergent than those identified using mutation accumulation, especially if long-term balancing selection is operating (e.g. [42]).
In order to draw inferences about the importance of these types of redundancy, it is critical to account for other factors unrelated to GT- and GF-redundancy that might drive repeatability, mainly through differences among genes in mutation rate or standing variation. The simplest approach to control these factors is to design studies that preclude shared standing variation, either through experiments founded from isogenic strains (e.g., [2,36]) or comparisons of distantly related lineages (divergence time >> 4Ne) where lineage sorting has been completed (as per [38]). While repeatability could still be driven by differences among genes in mutation rate, this can be seen as a component of GT-redundancy and therefore as a factor that can also constrain diversity. By contrast, the existence of shared standing variation occurs mainly due to historical contingency and is therefore a bias affecting estimation of C-scores, rather than a constraint. As such, parsing the contribution of mutation rate to C-scores and P^a is less critical than parsing the contribution of standing variation when using these as overall indices of constraint. Unfortunately, in studies of recently diverged natural populations, it is not possible to preclude shared standing variation, so C-scores and P^a could be strongly driven by this factor and therefore not particularly representative of diversity constraints. The recently developed likelihood-based method for discriminating between convergence via de novo mutation, migration, or shared standing variation ([24]) may provide a means to parse these contributions to repeatability and refine the inference of constraint. While testing the null hypothesis of no constraints is relatively straightforward, discriminating among other potential factors constraining diversity is much more complicated. Although it is possible to make very intricate models with variable mutation rates, selection coefficients, indices of pleiotropy, shared standing variation, and/or other factors determining the likelihood of each gene contributing to adaptation [3,25,28,43], it may be very difficult to actually confidently discriminate between such models.
The accuracy of the indices developed here will critically depend on the correct identification of the genes contributing to adaptation. Studies of local adaptation are particularly prone to false positives when population structure is oriented on the same axis as adaptive divergence, and it is unclear how extensively methods that correct for population structure induce false negatives or fail to accurately control for false positives [44,45]. Assuming false positives are distributed randomly throughout the genome in each lineage, failure to remove them will cause the C-scores derived here to be biased downwards. Failure to identify true positives (i.e. false negatives) will impair the accuracy of Cchisq, with the direction dependent upon the underlying biology. Assuming false negatives are randomly distributed in the genome, they could also reduce the magnitude of C-scores due to lower information content. On the other hand, if large-effect loci are more likely than small-effect loci to be both detected and convergent, false negatives will tend to bias C-scores upwards. As it is typically necessary to set arbitrary cutoffs for statistical significance to identify putatively adapted loci, we might expect Cchisq to increase with increasing stringency of these cutoffs, as this would be expected to reduce false positives (but also increase false negatives). However, as there are many potential contingencies and interactions between the factors that affect these two types of error, there is a clear need for both theoretical studies on how the repeatability of local adaptation is affected by the interplay between demography and selection (e.g. [28]), and refinement of these methods to derive confidence intervals taking into account likely error rates.
A particularly important problem to address in implementing this method is that false positives may be non-randomly distributed throughout the genome in a similar way in different lineages. As local variations in the rate of mutation or recombination can drive genome-wide patterns in some indices used to identify selection and adaptation [46–48], this could lead to signatures of convergence among distantly-related species if such patterns are conserved over long periods of evolutionary time. For example, genome-wide patterns of variation in nucleotide diversity, FST, and dxy were all significantly correlated across three distantly related bird species, likely driven in part by conservation of local recombination rate coupled with linked selection [49]. The extent of convergence of local recombination rates appears to vary considerably among species [50–53], so it will be important to consider this factor as a potential driver of similarity in the genomic signatures used to identify selection. Methods for identifying signatures of adaptation that are explicitly linked to a phenotype or environment of interest across multiple pairs of populations may be less likely to be affected by such factors, as recombination and linked selection are unlikely to drive a pattern of repeated correlation between allele frequency and phenotype/environment. However, such methods are still vulnerable to potential biases that arise from the complex interplay between genomic landscape, selection, and recombination, and further study in both theoretical and empirical contexts will be important to test the robustness of different methods to this important source of bias. One potential approach to estimate the contribution of such confounding factors would be to compare signatures of the repeatability for adaptation in two different traits (which are not phenotypically convergent) to those for a phenotypically convergent trait.
While studying adaptation across multiple pairs of populations can greatly increase the power to detect signatures of selection when all populations are adapting via the same loci, such methods are inherently unable to detect idiosyncratic patterns where different populations of a given species are adapting via different loci. By its very nature, it may be very difficult, if not impossible to detect local adaptation in traits with high GT- or GF-redundancy, as each pair of populations may be differentiated via a different set of loci [13]. If local adaptation is much more readily detected when it arises repeatedly within a lineage, then it will be difficult to identify conclusive cases with low C-scores, causing an overestimation of the prevalence of highly repeated adaptation.
If patterns of genomic convergence are compared among multiple differentially-related lineages, it is important to consider their phylogeny when testing the importance of phylogenetic sharing of different factors affecting the propensity for gene reuse [54]. The ability to resolve orthology relationships also decreases with increasing phylogenetic distance, which can affect the estimation of ns. Similarly, the set of genes in a trait’s mutational target (gi) is expected to evolve over time, so the set of shared genes should decrease with phylogenetic distance (so that gi − gs increases with divergence time), leading to decreased repeatability over time [33]. When studies include multiple differentially-related lineages, it is probably useful to estimate C-scores on both a pairwise and mean-across-all-lineages basis to more clearly describe cases where convergence is high within pairs of closely related lineages but low among more distantly related lineages.
Finally, physical linkage is a factor that could critically affect the measurement of repeatability, as neutral alleles in other genes linked to a causal allele will tend to respond to indirect selection, causing spurious signatures of selection/local adaptation. If the same causal gene is driving adaptation in two lineages, this will tend to overestimate repeatability on a gene-by-gene basis, whereas the opposite will occur if different causal genes are driving adaptation. Yeaman et al. [38] found significantly elevated levels of linkage disequilibrium (LD) among candidate genes for local adaptation, which may have arisen due to physical linkage (with or without selection on multiple causal loci) or statistical associations driven by selection among physically unlinked loci. In this case, the fragmented genome and lack of suitable genetic map precluded a comprehensive analysis of the impact of LD. If genome/genetic map resources permit, it may be possible to analyse repeatability on haplotype blocks rather than individual genes, which could minimize the biases due to physical linkage.
A large number of indices have been developed to characterize similarity among ecological communities, which can be broadly grouped based on binary vs. quantitative input data and whether they account for joint absence of a given type (reviewed in [55]). In most cases, these indices are not derived from a probability-based representation of expectations, though Raup and Crick [56] quantified an index of similarity based on the p-value of a hypergeometric test (see also [57]). The Chyper index that we have developed here uses the same underlying logic as the Raup-Crick index but quantifies the effect size as a deviation from the expectation under the null hypothesis in units of the standard deviation of the null distribution. The C-score and P^a indices developed here provide a complement to indices of repeatability that have been used in previous studies of convergence at the genome scale (e.g. [2,33]). Whereas the Jaccard, PSadd, and other similar indices represent how commonly a given gene tends to be used in adaptation, the C-score indices quantify how much constraint is involved in driving this observed repeatability, whereas P^a quantifies the proportion of the genome that is effectively available for adaptation. In some cases, these indices will be qualitatively similar in quantifying patterns of convergence (e.g. Fig 4A), but in other cases they will diverge considerably, because the C-scores are explicitly aimed at representing the importance of genes that could contribute to adaptation but do not.
The repeated observation of convergent adaptation at the genome scale violates a fundamental assumption of the infinitesimal model of quantitative genetics: that the mutations responsible for adaptation have small effects and are essentially interchangeable [58]. If there are infinitely many interchangeable loci that could contribute to adaptation, the chance of the same gene playing a causal role in independent bouts should be vanishingly small. Molecular-genetic studies show that some traits are causally generated by only a few genes in a specific pathway, presumably limiting the mutational target and increasing the potential for convergence. For example, only the small number of genes that are directly involved in terpene production [59] would likely contribute to large variations in the amount of terpene produced by a plant. However, a second category of mutations in other non-pathway genes could also indirectly contribute to variation in terpene production through perturbations of regulatory networks. The recently proposed “omnigenic model” posits that genes can be categorized into “core” vs. “peripheral” function for a given trait, as a way to distinguish between those with larger direct effects vs. smaller indirect effects [14], although this model has also been criticized [60]. The majority of evidence that has been considered in the context of the omnigenic model has come from Genome-Wide Association Studies (GWAS) of standing variation, but it is unclear whether this represents the “stuff” of long-term adaptation. Indeed, it appears that in humans there are pronounced differences in the distributions of alleles that contribute to standing vs. adaptive genetic variation, as GWAS studies of standing variation find mainly small-effect variants [61,62], whereas studies of local adaptation have found a number of large effect loci (e.g, for lactase persistence [63], diving [64], and high altitude [65]). If GT-redundancy is typically high and GF-redundancy commonly low, then there will be little correlation between the loci that can give rise to standing variation and the smaller subset of those responsible for long-term adaptation. Studying whether adaptation is commonly repeatable at the genome scale will therefore make an important complement to GWAS studies of standing variation, providing a window into the factors that constrain the diversity of viable routes to adaptation, and informing our broader understanding of how variation translates into evolution.
We present a method to quantify the constraints that drive genomic repeatability of adaptation, to enable testing of hypotheses about the nature of these constraints. Contrasting the repeatability of adaptation with studies of standing variation will deepen our understanding of evolution and the factors that affect how it gives rise to diversity. Comparative approaches examining C-scores and the proportion of adaptation-effective loci (P^a,lik) for the same trait across different branches of the phylogeny may allow us to infer rates of evolution in constraints and potential differences between rapidly vs. slowly radiating lineages, and study whether adaptation drives such changes. Comparisons across traits within lineages will illuminate how different kinds of traits are constrained, and whether low GT- and GF-redundancy constitute important constraints at different levels of biological organization. Similarly, this approach could be used to examine whether the types of constraint that predominate depend upon critical population genetic parameters such as effective population size, which affects the long term efficiency of selection on the developmental-genetic program. While we have focused on repeatability at the gene level, this framework could be applied at other levels of organization, such as gene network, protein domain, or individual nucleotide (reviewed by [3]), and could include the contribution of intergenic regulatory regions if it is possible to identify orthology. These methods therefore provide a first step towards comprehensive quantification and understanding of evolutionary constraints and the role that different factors play in the rise of diversity during adaptation.
|
10.1371/journal.pgen.1002934 | Citrullination of Histone H3 Interferes with HP1-Mediated Transcriptional Repression | Multiple Sclerosis (MS) is an autoimmune disease associated with abnormal expression of a subset of cytokines, resulting in inappropriate T-lymphocyte activation and uncontrolled immune response. A key issue in the field is the need to understand why these cytokines are transcriptionally activated in the patients. Here, we have examined several transcription units subject to pathological reactivation in MS, including the TNFα and IL8 cytokine genes and also several Human Endogenous RetroViruses (HERVs). We find that both the immune genes and the HERVs require the heterochromatin protein HP1α for their transcriptional repression. We further show that the Peptidylarginine Deiminase 4 (PADI4), an enzyme with a suspected role in MS, weakens the binding of HP1α to tri-methylated histone H3 lysine 9 by citrullinating histone H3 arginine 8. The resulting de-repression of both cytokines and HERVs can be reversed with the PADI-inhibitor Cl-amidine. Finally, we show that in peripheral blood mononuclear cells (PBMCs) from MS patients, the promoters of TNFα, and several HERVs share a deficit in HP1α recruitment and an augmented accumulation of histone H3 with a double citrulline 8 tri-methyl lysine 9 modifications. Thus, our study provides compelling evidence that HP1α and PADI4 are regulators of both immune genes and HERVs, and that multiple events of transcriptional reactivation in MS patients can be explained by the deficiency of a single mechanism of gene silencing.
| In patients suffering from Multiple Sclerosis (MS), T lymphocytes express abnormally high levels of a subset of cytokines. The same cells also transcribe a series of vestigial retroviral sequences normally silenced by chromatin factors. In this study, we have searched for regulatory mechanisms shared between the cytokines and the retroviral sequences. We find that the repressor protein HP1α is present on the promoter of both types of transcription units in normal cells and that the recruitment of this protein to these promoters is decreased in MS patients. Furthermore, we show that the delocalization of HP1α from these promoters can be caused by citrullination of histone H3, and we provide evidence indicating that levels of this histone modification is augmented in MS patients. Together our data provide a possible explanation for the simultaneously increased transcriptional activity of cytokines and endogenous retroviruses in MS-patient T lymphocytes and suggest that inhibitors of the enzyme responsible for the increased citrullination of histone H3 could help restore normal levels of cytokine activity in the patients.
| Multiple Sclerosis (MS) is a progressive inflammatory disease of the central nervous system in which leukocytes and antibodies attack myelin sheaths, resulting in demyelination and ultimately destruction of the axons [1]. Many lines of evidence point at inappropriate activation of T cells as an initiating event of the pathological process, although the mechanism at the root of this T cell activation is still poorly defined (for a recent review see [2]).
In MS patients, activation of the T cell population is associated with increased expression of a series of cytokines [3], [4]. The abnormally abundant expression of the genes encoding these regulators of the immune system may be a consequence, but also possibly a cause of the activation of the T cells. It is therefore essential to explore the mechanisms that keep these genes in check in normal cells and that may be defective in MS patients.
Interestingly, in MS and other autoimmune diseases including Rheumatoid Arthritis and Systemic Lupus Erythematosus, transcription of Human Endogenous RetroViruses (HERVs) is also increased in T cells [5], [6], [7]. HERVs are abundant vestigial retroviral sequences that in healthy cells are largely silenced by the epigenetic mechanisms repressing most repeated DNA sequences. These mechanisms include DNA methylation and histone H3 lysine 9 (H3K9) methylation. DNA methylation favors chromatin compaction by promoting recruitment of histone deacetylases (HDACs) or alternatively by directly reducing the affinity of transcription factors to their cognate DNA binding sites [8]. Consistent with this, deletion of the de novo DNA methyltransferase Dnmt1 results in massive re-expression of ERVs in the mouse embryo [9]. DNA methylation at ERV promoters is particularly high in differentiated mouse cells [10], while it may be partially dispensable in mouse embryonic stem (ES) cells [11].
In these cells, the major source of silencing appears to be H3K9 methylation [12], [13], [14]. This histone modification is recognized by a number of proteins containing either chromo- [15], [16], [17], [18], [19], [20], MBT- [21], PHD- [22], or Tudor- [21], [23] domains. Mainly HP1 proteins have been detected on mouse ERV promoter sequences [24], [25], although their role in the repression of these sequences in mouse ES cells is still at debate [26].
HP1 proteins are particularly interesting in the context of MS because in addition to their possible function in the silencing of repeated DNA [27], [28], they are present on the promoters of a number of genes involved in immune defense, including the immunomodulatory cytokine TNFα [29], the interleukins IL1β [30], [31], IL6 [32], and IL8 [33], and several interferon-inducible genes [34]. They also participate in the regulation of the HIV1 long terminal repeat (LTR) that shares several regulatory mechanisms with immune genes [35], [36], [37].
Consistent with their role in the transcriptional control of inducible genes, the binding of HP1 proteins to chromatin is subject to regulation. In particular, the methylation mark on histone H3 can be removed by histone demethylases [38]. A more transient regulation of HP1 binding may occur by modification of residues neighboring H3K9, including phosphorylation of serine 10 and acetylation of lysine 14 [39], [40], [41]. The histone H3 arginine 8 (H3R8) located immediately upstream of H3K9, is also subject to modifications [42] that theoretically could interfere with HP1 binding, although this has never been investigated. This arginine can be either methylated [43] or converted into the non-coding amino acid citrulline [44], [45], [46], [47], [48].
Citrullination of H3R8 is catalyzed by the calcium-dependent peptidylarginine deiminase PADI4. This enzyme, that is the only member of its family to enter the nucleus, also citrullinates histone H3 on arginines 2, 17, and 26, as well as histones H2A and H4 on their respective arginine 3 [44], [45], [46], [49]. Many reports describe PADI4 as a regulator of transcription. On p53 targets [50], [51], and estrogen-regulated genes, including pS2 [44], [45], [52], it functions as a repressor either by interfering with activating arginine-methylation events [44], or by favoring recruitment of HDACs [52]. Inversely, PADI4 also associates with a number of transcriptionally active promoters and functions as an activator of c-Fos via a mechanism that involves facilitated phosphorylation of the ETS-domain protein Elk-1 [53]. In Asians, hyperactivity of PADI4 has been associated with Rheumatoid Arthritis [54], [55], [56], and contributes to the generation of antibodies directed against citrullinated proteins during the development of this disease [57]. An earlier study has also reported increased nuclear localization of this enzyme in white brain matter of MS patients [58]
The transcriptional deregulation of both cytokines and HERVs in MS patient T cells prompted us to investigate a possibly co-regulation of these two types of transcription units by HP1 proteins. At cytokine genes and HERVs we examined in tissue culture cells, transcriptional repression required HP1α, while PADI4 functioned as an activator by destroying the HP1α binding site on the tail of histone H3. Consistent with this, we observed that in circulating blood cells from MS patients, recruitment of HP1α to the promoter of the master cytokine TNFα and to HERV sequences is significantly reduced, while citrullination of H3R8 at these positions is increased. Taken together, our data strongly suggest that increased citrullination of histone H3 can antagonize gene-specific chromatin-mediated silencing in T cells and thereby participate in increased cytokine expression during the normal inflammatory response and in MS patients.
While several reports describe an implication of HP1 proteins in the regulation of genes involved in immune defense [29], [30], [31], [32], [33], [34], the role of these proteins in the silencing of HERVs in human cells needed to be clarified. We carried out these experiments in MCF7 cells, a breast tumor-derived cell line frequently used to examine expression of HERVs (see for example [59]). Chromatin Immunoprecipitation (ChIP) assays demonstrated that in these cells, HP1α accumulates on HERV-K, HERV-W, and HERV-H promoters at levels similar to those observed on Satellite-2 sequences (Figure 1A). As expected, HP1α was also detected on the promoters of cytokines TNFα and IL8. Consistent with this, depletion of HP1α with small interfering RNAs (siRNAs) resulted in increased expression of the HERVH/env62, HERVH/env59, HERVH/env60, ERVWE1, HERVK/env102, TNFα, IL8, and IL16, a cytokine also relevant for MS [60] (Figure 1B, Figure S1A and S1B). In these experiments, expression of IL23 [61] was unaffected, while the control estrogen-responsive pS2 gene was repressed rather than activated. Reactivation of HERVs and TNFα was also observed upon depletion of HP1β and HP1γ, two other members of the HP1 family (Figure 1C).
HP1α binds tri-methylated histone H3 lysine 9 (H3K9me3). The neighboring H3R8 residue is one of the 3 arginines recognized by the anti-H3cit (2, 8, 17) antibody used to show increased histone H3 citrullination in MS patients [58]. This raised a possibility of interference between citrullination of H3R8 and HP1α binding to histone H3. We therefore explored whether citrullination of histone H3R8 occurs in vivo on histone H3 tails already tri-methylated on K9. To this end, we generated an antibody recognizing the double citrullination-methylation modification H3cit8K9me3 (Figure 2A).
The specificity of this antibody was verified by dot blots using synthetic peptides mimicking modified histone tails (Figure 2B) and by testing the ability of the same peptides to compete with the binding of the antibody to cellular targets in fixed breast cancer-derived MCF7 cells (Figure 2C). In the later assay, the anti-H3cit8K9me3 antibody yielded an immunofluorescent staining very similar to that obtained with anti-H3cit(2,8,17) antibody (Figure 2D).
We next focused our attention on PADI4, the partially nuclear peptidylarginine deiminase responsible for the citrullination of histone H3. To determine whether this enzyme can citrullinate H3R8 when H3K9 is methylated, we generated HEK293-derived cell lines expressing either wild-type (WT), hyperactive [62], or hypoactive [63] versions of PADI4 under the control of a ponasterone-inducible promoter (Figure 2E). Induction of PADI4 synthesis and activity with ponasterone and the ionophore A23187, respectively, allowed detection of the H3cit8K9me3 double modification when the cells were expressing WT or hyperactive versions of PADI4. Under these conditions, we also observed a general increase in histone H3 citrullination, but no change in the levels of H3K9me3, indicating that H3R8 citrullination and H3K9 tri-methylation are not antagonistic.
Taken together, these experiments demonstrate that the H3cit8K9me3 double modification exists in vivo and that its formation is favored by increased PADI4 activity.
We next investigate the impact of the H3R8 citrullination on the binding of HP1α to histone H3 tails tri-methylated on K9. When histone H3 peptides were spotted on membrane, HP1α bound a peptide carrying the single K9me3 modification, but not a peptide with a double cit8K9me3 modification (Figure 3A). We also tested the ability of these peptides to interfere with the binding of HP1α to its endogenous target sites. For this, we took advantage of the fact that recombinant GST-HP1α protein incubated on fixed permeabilized cells distributes in a pattern indistinguishable from that of the endogenous protein [64]. In this assay, while H3R8K9me3 peptide competed with the cellular sites for GST-HP1α binding, H3cit8K9me3 peptide did not (Figure 3B). As expected, H3cit8K9 and H3R8K9 peptides also failed to compete for GST-HP1α binding.
Surface plasmon resonance allowed us to quantify the effect of H3R8 citrullination and indicated a more than 200-fold decrease in the affinity of HP1α for H3cit8K9me3 compared to H3R8K9me3 (Kd 313±28 µM and 1.39±0.06 µM, respectively; Figure 3C–3D). We noted also that citrullination of H3R8 exerted a 10-fold higher effect on HP1α binding than did its methylation (Kd 29.6±1.3 µM).
To document that citrullination compromises transcriptional repression of HERVs and cytokines, we finally used siRNAs against PADI4 in the MCF7 cells known to express relatively high levels of PADI4 [65]. Depletion of this protein had an effect inverse to that of HP1α depletion and resulted in decreased levels of HERV transcripts (Figure 3E, Figure S1A and S1C). Levels of Satellite 2 transcripts (but not α-Satellite transcripts) were also decreased, suggesting a broad yet selective effect of citrullination on the silencing of repeats. PADI4 depletion also decreased expression of the immune genes TNFα, IL16, and IL8, as well as IL23 and IL1A, but not TGFß1 (Figure 3F). The control pS2 gene known to be negatively regulated by PADI4 [44], [45], was, as expected, moderately stimulated.
MCF7 cells are estrogen-responsive and an estradiol (E2) treatment combined with an ionophore increases total levels of both PADI4 (Figure S2A and [65]) and H3cit8K9me3 modification (Figure 4A, compare lanes 1 and 2). In contrast, reduced PADI activity can be obtained with the specific inhibitor Cl-amidine [66]. This drug affects the nuclear PADI4 as illustrated by the decreased levels of H3Cit8K9me3 in MCF7 cells (Figure 4A; compare lanes 1 and 3). Thus, treatment with either E2/A23187 or Cl-amidine allowed us to control endogenous nuclear PADI activity at will.
As in the PADI4-depletion experiments, treatment of the MCF7 cells with Cl-amidine reduced expression of HER V-H/env62, ERVWE1, and the selected cytokine genes, but not α-Satellite and TGFβ1 (Figure 4B, black bars). Inversely, augmenting PADI4 activity by treating the MCF7 cells with E2 and ionophore resulted in a Cl-amidine-sensitive increase in expression of the same HERVs and cytokines (Figure 4B, grey and white bars & S2B).
Finally, we performed ChIP to follow the impact of endogenous nuclear PADI activity on the citrullination of histone H3 and the recruitment of HP1α to the LTRs of HERV-H and ERVWE1 (HERV-W/LTR) and the promoter of TNFα. These assays confirmed that stimulation of PADI4 activity with E2 and ionophore locally increases levels of citrullinated histone H3, as detected with either anti-H3cit (2, 8, 17) or anti-H3cit8K9me3 antibody (Figure 4C–4D and Figure S2C), while the levels of HP1α recruitment were markedly decreased (Figure 4E, black and grey bars). These levels of HP1α occupancy were partially restored by further treating the cells with Cl-amidine (Figure 4E, white bars), illustrating that this inhibitor can overcome the detrimental effect of excessive PADI activity. The level of HP1α recruitment at the pS2 control promoter region was low and was not substantially affected by changes in citrullination levels (Figure 4C–4E).
Taken together, these results demonstrate that HP1α and citrullination antagonistically regulate several immune genes and HERVs, and that this regulation is druggable.
An inflammatory response can be induced in Jurkat T cells stimulated with an ionophore and the phorbol ester PMA. The stimulation of the Jurkat cells correlates with an eviction of HP1α from the promoter region of TNFα and IL8 (Figure 5A), and also from HERV-H/LTR62 and HERV-W/LTR (Figure S3A). The treatment also results in an approx. 6-fold increase of PADI4 accumulation (Figure 5B) and is expected to increase PADI activity as a consequence of the ionophore-induced calcium influx. We therefore used this system to determine whether increased PADI activity is associated with normal transcriptional activation of immune genes. Stimulation of the Jurkat cells resulted in a rapid and very transient recruitment of PADI4 to the promoters of TNFα and IL8, and at HERV-H/LTR62 and HERV-W/LTR (Figure S3B). This recruitment correlated with increase levels of H3Cit8K9me3 at these positions (Figure 5C and Figure S3C). Finally, inhibition of PADI activity with Cl-amidine reduced the kinetic and the abundance of TNFα and IL8 mRNA accumulation in Jurkat cells stimulated by ionomycin and PMA (Figure 5D and ). Together, these observations suggest that PADI activity participates in the modification of the epigenetic landscape at the promoter of immune genes upon stimulation of T cells, and thereby facilitate the transcriptional activation of these genes.
To investigate whether levels of HP1α recruitment and H3cit8K9me3 double modification at HERV and cytokine promoters were affected in MS, we collected PBMCs from 18 families, each family consisting of one MS patient and a genetically related healthy control (Table S1). The patients suffered from either relapsing-remitting (n = 10), or secondary progressive (n = 8) MS. As PBMCs yield only minute amounts of chromatin, our analysis was restricted to the LTRs of the unique HERV-H locus LTR59 [67] and the unique HERV-W locus ERVWE1 [68], and to the promoter of the cytokine TNFα. Consistent with previous observations [4], [5], [6], [7], transcription of these loci was significantly augmented in the MS patients (Figure 6A–6B).
ChIP assays revealed that recruitment of HP1α to the TNFα and the examined HERV promoters was significantly reduced in the MS patients compared to their genetically related healthy controls, while recruitment to a control promoter (RPLP0) was unchanged (Figure 6C). We also observed a significant correlation between HP1α levels on the TNFα promoter and HP1α levels on the LTRs of the HERVs, further suggesting that a single pathway regulates HP1α binding to both types of transcription units (Figure S4).
In ChIP assays, the anti-H3cit8K9me3 antibody was functional, but with a relatively poor sensitivity. Therefore, our analysis was restricted to the 9 families (9 patients and their respective healthy relatives) from whom we had the most abundant material. In these samples, levels of H3cit8K9me3 at the promoters of HERV-W/ERVWE1 and TNFα were significantly increased in the patients when compared to the genetically related healthy controls (Figure 6D–6E). On HERV-H/LTR59, levels of H3cit8K9me3 also appeared increased in the patients, but the p value associated with this data (0.06) is above the significance level of 0.05.
We finally questioned whether PADI4 could be linked to the increased levels of H3cit8K9me3. To this end, we examined PBMCs collected from the 18 families described above. Analysis of these samples by RT-PCR showed that PADI4 mRNA levels were significantly elevated in MS patients compared to the genetically related healthy controls (approx. 1.5-fold, Figure 6F).
Altogether, these experiments showed that in the patients, increased expression HERV-W/ERVWE1 and TNFα transcripts and decreased recruitment of HP1α at their promoter region is accompanied by a local increase in H3R8 citrullination and a moderate up-regulation of PADI4 expression.
In this report, we show that dependence on HP1α-mediated silencing is a common denominator between cytokines and HERVs, both expressed at abnormally high levels in T cells from MS patients, and we suggest that a decreased efficiency of the HP1-mediated silencing may participate in the pathological deregulation of these transcription units.
In this context, we find that one source of defective HP1-mediated silencing is citrullination of H3R8. This histone modification reduces the affinity of the chromo domain of the HP1 proteins to the methylated histone H3K9 residue and thereby defines a novel mechanism regulating HP1-binding to chromatin.
Using an antibody specifically recognizing H3cit8K9me3, we show that this double modification is induced in the presence of elevated levels of PADI4, the only known nuclear peptidylarginine deiminase. Interestingly, when an inflammatory response is induced in Jurkat T cells, expression of PADI4 is increased and levels of H3cit8K9me3 rise at the promoters of the immune genes IL8 and TNFα. Under these conditions, inhibiting PADI activity with the chemical inhibitor Cl-amidine results in reduced kinetic and amplitude in the activation of the two immune genes. These observations show that PADI activity and citrullination of histone H3 are required for normal activation of immune genes and define PADI4 as a novel regulator of cytokine expression. We speculate that H3cit8, together with other histone modifications such as H3S10 and H3S28 phosphorylation participate in creating an epigenetic landscape favorable for the transcriptional activation of a subset of immune genes.
PADI4 activity could also be artificially increased in MCF7 cells treated with estradiol and an ionophore. This allowed us to show that abnormally elevated levels of PADI activity result in transcriptional stimulation of several immune genes. Consistent with this, PBMCs collected from MS patients (and compared to healthy relatives) showed in average increased expression of TNFα, increased levels of H3cit8K9me3 at the promoter of this gene, and increased expression of PADI4. These observations strongly suggest that inappropriate activity of PADI4 can participate in the deregulation of immune genes relevant for MS (see model Figure 7). We here note that the estrogen/ionophore treatment inducing PADI4 expression in MCF7 cells also stimulated production of this enzyme in Jurkat T cells (data not shown). We therefore speculate that PADI4 could be involved in the activating effect of estrogen on TNFα expression observed in T cells under some conditions ([69] and references therein) and could thereby play a role in the higher incidence of MS in females [70].
Other enzymes affecting the affinity of HP1 proteins for chromatin may also be good candidates for an implication in MS. For example, Jak2 that is expressed at increased levels in MS Th17 cells [71] also cause exclusion of HP1α from chromatin by phosphorylating H3Y41, a residue contacted by the C-terminal region of the HP1 proteins [72]. Along the same lines, we note that levels of arginine methylation of myelin basic protein MBP is increased in MS patients [73], while we find that methylation of H3R8 reduces affinity of HP1α for the neighboring methylated H3K9 approximately 10-fold. Possibly, the same arginine methylases may be involved in the modification of both MBP and histones. In addition, PADI4 has earlier been described as involved in arginine demethylation, although methylated arginines are rather poor substrates for this enzyme in vitro [44], [45], [47]. Methylation and citrullination may therefore allow for a gradual activation of HP1α target genes in response to external stimuli.
The fact that PADI4 is a regulator of cytokines that can be either positively regulated by cellular stimuli or negatively regulated by specific inhibitors provides yet unexplored avenues to the control of inflammation, and in the case of MS, molecules such as Cl-amidine may potentially allow restoring chromatin-mediated repression of over-activated cytokine genes.
While HP1 proteins are best described as heterochromatic silencers and suppressors of variegation, our observations confirm that these proteins are also highly relevant for the transcriptional control of inducible genes that require a transient phase of silencing. The sharing of regulatory mechanisms between euchromatic cytokine genes and repeated sequences such as HERVs suggests that many bridges may exist between active and inactive chromatin, and that there is a continuum and not a clear-cut boarder between euchromatin and heterochromatin. Therefore, probing the status of heterochromatic silencing as well as its defects may provide much new insight on the transcriptional programs in which cells are engaged.
Blood samples were collected from each participant after informed consent as approved by the local Danish ethical committee.
The study population consisted of 36 subjects, encompassing 18 MS patients clinically diagnosed for MS and fulfilling the diagnostic criteria of Poser et al., 1993 [75] and 18 unaffected (healthy) first or second degree relatives, one for each of the MS patients. The participants were from a homogenous population (Caucasian, Northern European descent). Venous blood was drawn and processed on the same day in our laboratory. The clinical and demographic data of each participant are summarized in Table S1. The mean age of both the MS patients and their unaffected relatives was 52 years. The gender ratios for MS patients (11 female/7 male) and unaffected relatives (9 female/9 male) were also comparable. Peripheral blood mononuclear cells (PBMCs) were prepared by standard Isopaque-Ficoll centrifugation. The separated cells were cryo-preserved in RPMI with addition of 20% inactivated human serum (HS) and 10% DMSO, at −135°C until use. For the assays, PBMCs were thawed and cultured for 24 h in RPMI-1640 with 10% inactivated human serum, and 100 Uml-1 penicillin-streptomycin at 37°C in a 5% CO2 incubator prior to use. For each family, PBMCs from the patient and the control individual were analyzed at the same time.The data were analyzed by using the software XLSTAT (version 2010.5.06, www.xlstat.com). When indicated in the text, Wilcoxon signed rank test [76] was used to determine whether a significant (p<0.05) difference.
Anti-H3cit8K9me3 antibody was produced in rabbits using a peptide coupled to KLH with the following sequence: ARTKQTA (cit)(Kme3)STGGKAPRC. Anti-PADI4 (ChIP: P4749; immunoblots: ab50332), anti-H3cit(2,8,17) (ab5103), anti-H3(ab1791), and anti-H3K9me3 (ab8898) antibodies were purchased from Abcam. Anti-HP1α (ChIP: 1H5; immunoblots: 2G9) and anti-Brg1 (2E12) were from Euromedex. Calcium ionophore A23187 (used on HEK293 and MCF7 cells) and ionomycin (only ionophore tolerated by Jurkat cells), and 17-ß-estradiol (E2758) were purchased from Sigma. DNA was labeled with 4′,6-diamidino-2-phenylindole (DAPI; Invitrogen) at a concentration of 150 ng.ml−1. The peptide ARTKQTARKSTGGKAPRC was used for competition, overlay, and surface plasmon experiments, either unmodified or with K9me3, R8meK9me3, cit8K9me3, or cit8 modifications. Peptides were carefully quantified by amino acid analysis, and the presence of the modifications was confirmed by mass spectrometry.
ChIP was carried out essentially as previously described [74], with minor alterations. MCF7 or Jurkat cells, or PBMCs were cross-linked in phosphate-buffered saline (PBS) containing 1% formaldehyde (Sigma) for 10 min at room temperature. The crosslinking reaction was quenched with PBS containing 125 mM glycine, followed by three washes with ice-cold PBS. The chromatin was fragmented by sonication to produce average DNA lengths of 0.5 kb. After ChIP, the eluted DNAs were detected by quantitative PCR using the primers listed in Table S2. Levels of histone modifications are expressed as % of H3, and levels of HP1α are expressed relatively to the signal obtained for ChIP using non-immune IgGs. Values are averaged from 3 independent experiments.
Real-time SPR assays were performed at 25°C in PBS. GST-HP1α was covalently coupled to a CM5 sensor chip, using a Biacore 2000 instrument and an Amine Coupling Kit (GE Healthcare), achieving three different immobilization densities (Rimmo) of 3500, 6500, and 15000 resonance units (RU; 1RU ≈1 pg.mm−2). On the remaining flow cell, 5700 RU of GST were immobilized to prepare a reference surface. A series of 10 concentrations of peptides (50 nM–25 µM for H3R8K9me3, 200 nM–100 µM for the H3cit8K9me3, H3cit8K9, and unmodified peptides) were injected for 2 min over the GST-HP1α and GST surfaces at a flow rate of 50 µL.min−1. After following the dissociation for 5 min, the surfaces were regenerated with a 3-min wash of 2 M NaCl, and two 15-sec washes with 10 mM glycine-HCl (pH 1.5) and 0.05% SDS.
The association and dissociation profiles were double-referenced using the Scrubber 2.0 software (BioLogic Software) (i.e. both the signals from the reference GST surface and from blank experiments using PBS instead of peptide were subtracted). The steady-state SPR responses (Req) were plotted against the peptide concentration (C) and fitted according to the following equation:(1)where Kd is the equilibrium dissociation constant of the peptide/GST-HP1α interaction and Rmax the maximal binding capacity of GST-HP1α.
The percentage of bound HP1α sites was determined as follows:(2)
MCF-7 and HEK293 cells were cultured in Dulbecco's modified Eagle's medium (DMEM, Gibco BRL), and Jurkat cells were cultured in RPMI-1640, all with 10% decomplemented fetal bovine serum (FBS) and 100 U.ml−1 penicillin-streptomycin at 37°C in a 5% CO2 incubator. MCF7 cells were treated with 200 nM of estradiol (E2) for 24 h, washed three times with 1× PBS and then incubated for 30 minutes in Locke's solution (10 mM HEPES.HCl, pH 7.3, 150 mM NaCl, 5 mM KCl, 2 mM CaCl2, and 0.1% glucose) supplemented with 5 µM A23187 (C7522, Sigma). Jurkat cells were treated with phorbol myristate acetate (PMA) at 40 nM. Ionophores were used at 1 µM. PADI4 inhibitor, Cl-amidine (from Bertin Pharma) was dissolved in 1× PBS as a 50 mM stock solution. MCF7 and Jurkat cells were treated with 200 µM Cl-amidine in complete cell culture medium for exactly 72 h.
The cDNAs coding for either WT PADI4, hyperactive PADI4 (Mut+) [62], or hypoactive PADI4 (Mut−) [63] were inserted into the pi_tk_hygro vector for retroviral delivery [77]. Two days after transfection using FuGENE (Roche), the medium of packaging HEK293 cells was filtered on a 0.45 µm filter (Millipore) and supplemented with polybrene (AL-118, Sigma) at 100 µg.ml−1. This medium was used to infect host HEK293 cells (3 consecutive infections) followed by selection with hygromycin (H3274, Sigma) at a final concentration of 100 µg.ml−1.
HP1 small interfering RNAs (siRNAs) were described previously [35]. PADI4 (J-012471-05) and Glyceraldehyde-3-phosphate dehydrogenase control siRNAs were purchased from Dharmacon. Cells were harvested 72 h after transfection with DharmaFECT 1 (T 2001-03). Total RNA from PBMCs and MCF7 cells was extracted with RNeasy (Qiagen) and quantified with an ND-1000 (Nanodrop). After DNase treatment (Roche), reverse transcription was performed using SuperScript III (Invitrogen) and random hexanucleotides according to the manufacturer's instructions. Complementary DNA was quantified by RT-qPCR as previously described [35]. PCR primers are listed in Table S2. Proteins were extracted as described previously [35] and detected by Western blotting. Acid extraction of histones was conducted as described by Shechter, et al., [78].
Immunofluorescent labeling was performed in MCF7 cells after sequential treatment with 200 nM of estradiol (E2) for 24 h in culture media followed by 5 µM A23187 for 30 min in Locke's solution. Cells were permeabilized in ice cold CSK (20 mM PIPES pH 6.8, 200 mM NaCl, 600 mM sucrose, 6 mM MgCl2, and 2 mM EGTA) 0.5% Triton X-100 (v/v), 0.1 mM PMSF for 30 sec. Cells were fixed with CSK-3.7% paraformaldehyde at room temperature (RT) for 10 min, then blocked in PBS-0.05% Tween-20 (v/v), 10% (v/v) FBS for 30 min at RT. In peptide competition experiments, primary antibodies were pre-incubated with 1 µg of indicated histone H3 peptides or without peptide for 30 min. Cells were incubated with antibodies at 4°C overnight and then coverslips were washed three times in PBS-BSA 0.5% (w/v). Coverslips were incubated with secondary antibodies for 1 h at room temperature protected from light, before being washed three times in PBS-BSA 0.5% (w/v), and once in PBS, before final staining with DAPI. Imaging was conducted on an Axiovert 200 M microscope (Zeiss) coupled with an Apotome with Axiovision 4.7 (Zeiss).
|
10.1371/journal.pgen.1005963 | De Novo and Rare Variants at Multiple Loci Support the Oligogenic Origins of Atrioventricular Septal Heart Defects | Congenital heart disease (CHD) has a complex genetic etiology, and recent studies suggest that high penetrance de novo mutations may account for only a small fraction of disease. In a multi-institutional cohort surveyed by exome sequencing, combining analysis of 987 individuals (discovery cohort of 59 affected trios and 59 control trios, and a replication cohort of 100 affected singletons and 533 unaffected singletons) we observe variation at novel and known loci related to a specific cardiac malformation the atrioventricular septal defect (AVSD). In a primary analysis, by combining developmental coexpression networks with inheritance modeling, we identify a de novo mutation in the DNA binding domain of NR1D2 (p.R175W). We show that p.R175W changes the transcriptional activity of Nr1d2 using an in vitro transactivation model in HUVEC cells. Finally, we demonstrate previously unrecognized cardiovascular malformations in the Nr1d2tm1-Dgen knockout mouse. In secondary analyses we map genetic variation to protein-interaction networks suggesting a role for two collagen genes in AVSD, which we corroborate by burden testing in a second replication cohort of 100 AVSDs and 533 controls (p = 8.37e-08). Finally, we apply a rare-disease inheritance model to identify variation in genes previously associated with CHD (ZFPM2, NSD1, NOTCH1, VCAN, and MYH6), cardiac malformations in mouse models (ADAM17, CHRD, IFT140, PTPRJ, RYR1 and ATE1), and hypomorphic alleles of genes causing syndromic CHD (EHMT1, SRCAP, BBS2, NOTCH2, and KMT2D) in 14 of 59 trios, greatly exceeding variation in control trios without CHD (p = 9.60e-06). In total, 32% of trios carried at least one putatively disease-associated variant across 19 loci,suggesting that inherited and de novo variation across a heterogeneous group of loci may contribute to disease risk.
| Congenital heart disease (CHD) is a leading cause of childhood morbidity in the developed world. There are few prevalent clinical risk factors and though it is possible that up to 90% of risk for CHD may be genetic, the number of genes clinically associated with disease is small. Rather than grouping disparate CHD phenotypes as other studies have done, we studied a single specific malformation- the atrioventricular septal defect (AVSD). Instead of recurrent variation in a handful of genes, we observed de novo and inherited variation in 19 genes associated with human disease, syndromic loci, and genes implicated in cardiac development by mouse knockout. The number of loci identified support the longstanding hypothesis of a complex oligogenic inheritance for a single malformation and suggest that analyses of CHD data to include inherited variation may uncover additional loci contributing risk for cardiac malformations.
| Congenital heart disease (CHD) is the most common congenital malformation and the most common cause of mortality during the first year of life in the United States [1,2]. Most cases occur sporadically without a strong family history or identifiable genetic syndrome, and the primary heritable basis of most non-syndromic congenital heart disease has yet to be identified [3,4]. Studies of affected kindreds and syndromic disease have revealed high-penetrance mutations at a small number of key loci [5]. Exome sequencing and studies of structural variation of mixed cardiac phenotypes focusing on de novo events have identified novel disease loci in 4–10% of participants [6,7]. However the remaining majority of non-syndromic subjects in exome and CNV studies are without an identified genetic cause.
Atrioventricular septal defects (AVSD) are a rare cardiac malformation associated to date with a handful of canonical genes in cardiac development (NKX2-5, GATA4, GATA6, CRELD1) and may co-occur with certain rare syndromes. A recent study discovered causal variation in the nuclear receptor NR2F2 in 4% of 125 subjects with AVSD, pinpointing a single additional disease-associated gene [7]. However the 4–10% discovery rate in studies of CHD highlight the observation that for any individual gene, highly penetrant de novo coding mutations may only account for a small portion of disease incidence, a phenomenon similar to the sporadic occurrence and complex genetics of neurodevelopmental disorders [8]. Therefore, expanding the scope of analysis in studies of CHD to include both inherited and de novo variation in multiple genes could increase the sensitivity of genetic studies of this heterogeneous group of oligogenic diseases [9–13].
To this end we assembled a multi-institutional cohort combining a discovery cohort of 59 trios with non-syndromic AVSD and a replication cohort of 100 single affected individuals and performed a genetic survey by exome sequencing and array-CGH. In a primary analysis we identified a novel candidate gene for AVSDs using inheritance modeling and prior knowledge of early cardiac gene expression (Fig 1). In secondary analyses, we searched protein interaction networks to identify the contribution of additional loci to this rare cardiovascular malformation. Finally we explored the contribution of rare inherited variation in genes related to other types of human and mouse cardiac malformations to AVSD.
We determined the sensitivity of our informatics approach using a recently described consensus standard dataset of exonic 24,734 variants from NA12878 [14]. Raw exome data from a well-characterized individual was analyzed with BWA/GATK 3.2 best practices and the RTG version 3.3.2 software pipeline (Real Time Genomics Inc., Hamilton, New Zealand). RTG displayed greater sensitivity detecting 84.5% consensus standard variants compared to 79.9% for BWA/GATK (S1 Table). Using the RTG pipeline we analyzed exome sequencing data on 159 individuals with AVSD but without a syndrome (a discovery cohort of 59 trios, and a replication cohort of 100 singletons derived largely from a published study of AVSDs (S2 Table), along with 710 controls without congenital heart disease (59 trios, 533 singletons). The affected patients were situs solitus with a simple AVSD. Patients with other cardiac defects, heterotaxy, anatomical malformations, or developmental delay were excluded. Across all individuals, called variants displayed a median Ti/Tv ratio of 3.10, and a median of 89.6% phased genotypes within trio probands, which suggested an overall highly sensitive and accurate variant call set (S3 Table). All de novo variants and insertions/deletions of interest were confirmed by Sanger sequencing.
Within the 59 AVSD trios and 59 control trios we analyzed variants with minor allele frequency of 0.03 or less in the 61,468 multiethnic individuals in the EXaC consortium [15]. With the remaining protein-altering variants, we applied a rare-disease inheritance model to select only variants displaying classical inheritance patterns associated with sporadic presentation of a rare disease (de novo, homozygous, compound heterozygous) (Fig 1) [16]; this filtering process yielded a list of 710 variants in 399 genes in the 59 AVSD trios (S4 Table).
To identify novel genes involved in cardiac development and disease among the 399 loci, we reanalyzed 72 digital gene expression datasets derived from 22 tissues during mouse embryonic development (www.mouseatlas.org) [17] (S1 Fig). The tissue types included the AV-canal along with 5 other cardiac tissues, and 16 other tissues from other organs or structures. After constructing co-expression modules using unsupervised weighted-gene coexpression network analysis, we observed that one of the co-expression modules expressed in mouse AV-canal tissue included four of six genes known to cause AVSDs (GATA4, GATA6, NKX2-5, and CRELD1, p = 6.57e-05, one-tail hypergeometric test) along with 69 of 756 genes related to other human or mouse cardiac malformations (S5 Table) (p = 7.81e-08, one-tail hypergeometric test). These observations suggested the discovery of a co-expression module highly enriched for genes related to heart development and cardiac malformations.
Intersecting the genes in this unique coexpression module (S6 Table) with the 399 genes identified by the rare-disease inheritance model, two probands from the 59 AVSD trios displayed de novo mutations. One proband had a missense mutation in a non-conserved residue of KCNJ3 and a second proband had a missense variant in NR1D2. The KCNJ3 variant has been observed in low frequencies in European and African populations in the ExAC database and the available literature on Kcnj3 knockout animals did not suggest occult cardiovascular malformations [18,19]. By contrast, the NR1D2 variant causes an arginine to tryptophan mutation at position 175 (p.R175W) in a highly conserved DNA binding domain (Fig 2a and 2b). NR1D2 is a transcriptional co-repressor and modulator without a described role in cardiac development [20]. De novo mutations in NR1D2 or any gene in the cardiac malformation module were absent from 59 control trios without congenital heart disease. Among the 61,468 putatively healthy individuals cataloged in the ExAC database there was only a single non-synonymous mutation in the surrounding five protein residues surrounding the p.R175W allele. Overall the data were suggestive that this de novo allele might impact the function of NR1D2.
To characterize the transcriptional activity of the p.R175W mutation we developed an in vitro transactivation assay performed in HUVEC cells, which employed a murine Nr1d2 expression vector co-transfected with an Nr1d2 response element (RE) consisting of 5 tandem repeats of a conserved binding site upstream of a minimal CMV promoter driving GFP. As NR1D2 is thought to act as a transcriptional co-repressor, a positive change in transcriptional activity of the RE vector may represent a decrease in the transcriptional co-repressor activity of Nr1d2. In this in vitro assay, the p.R175W mutation displayed an increased transcriptional activity relative to the wild-type allele (Fig 2c), which suggested that the p.R175W mutation in a conserved region of the DNA binding domain might functionally impair the native co-repressor function of NR1D2.
Though reports of a previous characterization of an Nr1d2 knockout allele did not show cardiovascular malformations [20], a percentage of homozygous knockout animals have been reported to die within hours of birth consistent with the presence of hemodynamically significant cardiac malformations [21]. Two pairs of heterozygous founder animals were bred, yielding 17 pups (2 wild type, 7 Nr1d2tm1-Dgen +/-, and 8 Nr1d2tm1-Dgen -/-) which did not deviate obviously from expected Mendelian allelic ratios. Upon careful histological examination of two spontaneously deceased Nr1d2tm1-Dgen -/- pups at P0 we detected a previously undescribed AVSD phenocopy (Fig 2d). We performed additional matings of +/- and -/- animals, and sacrificed mothers to obtain embryos at e16.5 and e17.5. A single -/- animal at e16.5 displayed an AVSD and a single -/- animal at e17.5 displayed an inlet ventricular septal defect which is closely related to AVSDs (Fig 2d). In total, 4 out of 15 -/- hearts assayed displayed a cardiac defect. Thus by combining inheritance modeling with a gene-coexpression network enriched for genes causing CHD, we identified a variant in NR1D2 which impacts transcriptional activity in vitro, and uncovered previously unrecognized cardiovascular malformations in an Nr1d2 knockout allele. Together these data suggest a new role for the transcriptional repressor NR1D2 in cardiac development and human disease.
The relatively low discovery rate in rare-variant association studies of CHD suggests that alternative analytical approaches may be necessary to distinguish the contribution of novel loci to disease risk [6,7]. Protein interaction networks have successfully integrated known disease genes to discover the impact of novel loci in neurodevelopmental disorders and cancer, thus in a secondary analysis we employed an algorithm which searches protein-interaction data for over-representation of genetic variation within interacting proteins in the AVSD trios [22] (Fig 3).
Including protein altering single nucleotide mutations and CNVs derived from the discovery cohort of 59 AVSD trios, the algorithm identified 86 enriched subnetworks of interacting proteins containing 2 to 7 genes. By comparison, applying the algorithm to 59 control-trios identified 26 subnetworks of interacting proteins containing only 2 to 4 genes (Fig 4a). Using a procedure where the protein interaction network is randomly permuted, the genes within the AVSD-trio subnetworks were found to be enriched for true protein-protein interactions (median p = 0.01, network permutation procedure), while true protein-protein interactions were not observed within the control trio subnetworks (median p = 1.0, network permutation procedure) [23] (Fig 4a).
To further characterize the protein interaction networks detected, we compared the subnetworks to gene expression in mouse cardiac development (S8 Table). Genes within the AVSD-trio subnetworks were strongly overrepresented during mouse heart development (p = 9e-09, one-tailed hypergeometric test) while genes within the control-trio subnetworks were not (p = 0.34, one-tailed hypergeometric test) (Fig 4b). Thus, mapping genetic variation in the AVSD trios to protein interaction networks identifies 86 enriched subnetworks with deleterious variation in 231 genes that are preferentially expressed during cardiac development, a phenomenon not seen in the control trios without CHD.
To validate the genetic associations suggested by the discovered AVSD-trio subnetworks, we assembled a separate replication cohort of 100 singleton individuals (S2 Table) and performed burden testing for each of the 86 protein networks (Fig 3). After Bonferroni correction for multiple hypothesis testing, a single subnetwork from the AVSD trios composed of a pair of interacting collagen genes (COL2A1, COL9A1) (Fig 4a and 4c) displayed an elevated burden of rare coding variation in 100 affected individuals with AVSD compared to 533 controls without congenital heart disease (p = 8.37e-08, SKAT linear weighted test) (Fig 4c). Interestingly, the two genes COL2A1 and COL9A1 have evolutionarily conserved roles in cardiac development in both zebrafish and mouse [24–26]. One mouse knockout allele of Col2a1 displays cardiac valve abnormalities [27], and mutations in both genes are associated with Stickler syndrome where 46% of patients are affected with congenital dysfunction of the mitral valve [28]. Together the functional data on COL2A1 and COL9A1 accompanied by the identification of these genes with two separate methodologies in two separate cohorts of AVSD patients, support a potential association with other congenital structural malformations of cardiac valve tissue such as AVSDs.
Outside of novel genes identified by developmental coexpression and protein interaction analyses, we wished to examine the impact of genes known to play a role in cardiovascular development or CHD in our cohort. Interestingly, de novo single nucleotide mutations in genes previously associated with AVSD (NKX2-5, GATA4, GATA6, EVC, CRELD1, NR2F2) were absent and we did not detect genes with recurrent de novo mutations.
In an effort to categorize and catalog variation at known CHD loci within the 710 variants in 399 genes identified by the rare-disease inheritance model in the 59 AVSD-trios (S4 Table), we assembled a predefined group of 756 loci associated with any human or mouse cardiac malformation (S5 Table). Among the genes identified by the rare-disease inheritance model in the 59-AVSD probands (S4 Table), we observed inherited and de novo variation in 16 genes (Table 1 and S7 Table) [29]. Four of the affected probands displayed variation in more than one gene. In a set of 59 control trios without congenital heart disease we applied the identical variant calling pipeline and rare-disease inheritance model. Comparing the number of AVSD probands with inherited mutations in the identified 16 genes to unaffected controls, we observed 16 mutations in the AVSD-cases and only a single variant in controls (p = 9.60e-06, Fisher's exact test), and additional simulations confirmed an unusual distribution of mutations in the AVSD-cases compared to controls was unlikely to be a chance occurrence (p = 1.23e-06, Monte Carlo simulation). As an additional negative control we compared mutations in a list of 43 genes associated with congenital ocular malformations between the 59 AVSD cases and 59 controls; there was only a single inherited variant among the AVSD trios and none within the control trios (p = 1.0, Fisher’s exact test). Together these results suggest a preponderance of de novo and inherited variation in genes associated with human or mouse cardiac malformations detected in the AVSD trios which greatly exceeded similar variation in control trios.
Within the discovery cohort of 59 trios we also assessed structural variation by array CGH and read-depth analysis from exome studies. One patient was observed to have a 3.7 Mb de novo deletion at 8p23.1 encompassing 43 genes including GATA4 (Table 2a). An additional paternally inherited duplication at chr22:21,989,140–23,627,391 partially overlapping the congenital heart disease-associated 22q11.2 duplication syndrome region was also detected [30]. Additional CNVs with a previous association to CHD were identified in the singleton subjects (Table 2b) [31]. CNVs in these regions were absent from the trio and singleton controls. Thus, including inherited variants, de novo mutations, and structural variation, rare deleterious variants or CNVs in genes with strong prior evidence for association with congenital heart disease were observed in 14 of 59 or 23% of affected trios.
In this study combining both exome-sequencing and array-CGH for a single specific cardiac malformation, we observed de novo and inherited variation in 19 genes associated with human disease, syndromic loci, and genes implicated in cardiac development by mouse knockout. In the absence of recurrent de novo events in a moderately sized cohort of 159 affected individuals, we applied an array of analytical techniques to look for both de novo and inherited variation associated with AVSDs.
When combined with inheritance analysis, a gene-coexpression network derived from mouse development allowed us to identify a previously unrecognized role for the transcriptional repressor NR1D2 in cardiac development and human disease. Our experimental studies suggested that the observed p.R175W mutation impacts the transcriptional activity of murine Nr1d2, and we observed previously unrecognized cardiac malformations in the Nr1d2tm1-Dgen knockout mouse. Yet within a cohort of 159 affected individuals, there was only a single patient with a de novo mutation in NR1D2, which highlights the underlying genetic heterogeneity of CHD and the utility of applying orthogonal datasets to pinpoint causal variation.
The -/- animals for the Nr1d2tm1-Dgen allele display incomplete penetrance; a majority of animals do not display cardiovascular malformations and develop normally to adulthood displaying phenotypes related to circadian rhythm and abnormal lipid metabolism [20]. Nr1d2 may retain multiple roles in modulating transcription but is most clearly described as a transcriptional repressor, therefore a mutation in the DNA binding domain might impact Nr1d2 to binding to target sequences resulting in a “de-repression” transcriptional targets of Nr1d2. A knockout allele such as Nr1d2tm1-Dgen might similarly “de-repress” targets of Nr1d2. From the standpoint of transcriptional repression, increased transcription of the p.R175W mutant observed in vitro may represent a decrease in transcriptional repression relative to the wild-type protein, and as such could be entirely consistent with the phenotype of a mouse knockout allele. Interestingly, NR1D2 is a well-characterized component of the molecular clock, and further studies would be necessary to investigate NR1D2 as a link between the molecular clock and timing of cardiac development.
Chromatin remodeling factors have recently been implicated as primary and secondary causal factors in CHD [6,32], and both newly discovered factors NR1D2 and NR2F2 may play integrated roles in chromatin remodeling during cardiac development. The key histone deacetylase HDAC1 is directly activated by NR1D2 binding and indirectly activated by NR2F2 via PROX1 [33–35]. Additionally NR1D2 may function upstream of NR2F2, modulating the auto-regulatory activity of NR2F2 via HDAC1 and the glucocorticoid receptor GR complex [36,37]. Further experiments are necessary to delineate the tissue localization, timing, expression, and functional roles of these two transcription factors and their role in chromatin modulation and transcriptional regulation during cardiac development.
Within a complex cellular or tissue signaling pathway, capturing the genetic variation in one or more interacting proteins has yielded novel candidate genes in cancer and neurodevelopmental disorders [22,38]. Adapting a tool designed to search protein interaction networks in cancer, we identified a small number of variants in genes within the AVSD trios, two of which (COL2A1 and COL9A1) were subsequently validated in burden testing of a separate replication cohort of 100 individuals at a statistically significant threshold. Independent experimental data implicates these genes in the development of the cardiac valve structures, and links these genes to a genetic syndrome that includes abnormalities of the cardiac valves among a host of other phenotypes. Though our observations are not firmly conclusive of a causal role for COL2A1 and COL9A1 in the pathogenesis of AVSDs, they are supportive of such a role, and we believe, consistent with the idea that network-based approaches may be fruitfully applied to gene discovery in CHD phenotypes.
Surprisingly, with the exception of a deletion encompassing GATA4 seen in one trio subject and one singleton subject (Table 2), we did not discover de novo coding mutations or gene dosage alterations within the 59 trios in canonical AVSD genes (NKX2-5, EVC, CRELD1, GATA6) or at the newly discovered CHD risk locus NR2F2. This finding is consistent with studies of CHD examining candidate genes [39] and exome sequencing where protein-altering variants in any single gene are reported in no more than 1–4% of patients [7]. The absence of recurrent de novo variants in a cohort of 59 patients with AVSD is in striking contrast with other cardiac conditions such as long QT syndrome where pathogenic coding variation in only 5 genes accounts for disease in 70% patients [40]. We hypothesize that the absence of de novo variation observed at canonical loci in a cohort of this size reflects the complexity of CHD genetics and highlights the utility of considering alternative inheritance patterns to detect disease-associated variation.
Among a list of 756 genes with either a clinical or experimental association to cardiac malformations (S5 Table) we observed rare inherited variants in the AVSD trios that were not seen in control trios [29]. In genes clinically associated with CHD we observed compound heterozygous variants inherited in trans in ZFPM2, NSD1, NOTCH1, VCAN, and MYH6 and rare homozygous variants in MYH6. Variation was observed in 9 additional genes including compound heterozygous variants in ADAM17, CHRD, IFT140, PTPRJ, and RYR1 and rare homozygous variants in ATE1, and the presence of heart defects in mouse knockout models for these loci supports their association with human cardiac malformations.
In an independent forward mutagenesis screen Ift140 causes AVSD among a variety of congenital malformations [41]. The calcium channel RYR1 is associated with skeletal myopathies and malignant hyperthermia, but primum atrial septal defects in one mouse allele suggest a role in early cardiac development [42]. The metalloproteinase ADAM17 may link NOTCH1 signaling in cardiac valve development to the left-right patterning of the heart [43,44]. Individual knockout alleles of CHRD, PTPRJ, and ATE1 each show defects in heart development recapitulating different human malformations [45–47]. Importantly, we observed rare inherited variation in genes with experimental or clinical evidence for a role in cardiac development and CHD within the AVSD trios, but rare inherited variation in these same genes was largely absent from the control-trios.
Despite excluding syndromic features and developmental delay from our patients at the time of recruitment, we observed inherited and de novo variation in genes causing syndromic disease that include heart malformations. A de novo mutation was detected in an unknown protein domain of EHMT1 the causal gene in Kleefstra syndrome, compound heterozygous variants inherited in trans were observed in SRCAP which was recently associated with Floating-Harbor syndrome, two individuals showed compound heterozygous mutations in BBS2 which causes Bardet-Biedel syndrome, a fifth individual displayed compound heterozygous mutations in NOTCH2 which causes Alagille syndrome, and a sixth individual displayed a compound heterozygous mutation in KMT2D the locus implicated in Kabuki syndrome. Each of these multi-organ syndromes is frequently accompanied by AVSD or another related form of congenital heart disease [48–52].
Within a single locus associated with a genetic syndrome, different alleles may vary in their expressivity. We hypothesize that these variants represent hypomorphic alleles of syndromic genes, where the patients affected present only one aspect of the phenotype associated with the syndrome, in this case a phenocopy of a syndromic associated cardiovascular malformation [53]. Indeed on secondary followup, none of the included probands with EHMT1, KMT2D, SRCAP, or NOTCH2 variants displayed other characteristics of their associated syndromes, while the patients with BBS2 mutations were not available for review (additional phenotypic information on patients carrying syndromic alleles is detailed in the S1 Text). Supporting the possibility of hypomorphic alleles, there was a striking absence of de novo or inherited variants in these syndromic genes within the control trios suggestive that the discovered variants may confer risk for AVSD.
These findings have limitations. Although we excluded patients with a family history of cardiac malformations, in an earlier era of surgical care the parents of the study participants would have been less likely to survive to reproductive age [54]. In our study the parents received only a questionnaire and did not receive screening echocardiogram, thus we cannot rule out that a parent in an included trio may have a forme fruste of an AVSD-related malformation such as a cleft mitral valve or ostium primum ASD. Additionally, there is emerging evidence that maternal risk factors (both genetic and environmental) which confer risk for CHD that were not considered in our study design [55–57].
Genetic studies of CHD are challenged by the fact that specific individual malformations are quite rare (AVSD is approximately 0.3 per 10,000 live births), and that any substantial group of patients with a single malformation will contain some population stratification. The unexplored role of non-coding gene regulatory variation in congenital heart disease is not surveyed by our exome-sequencing approach [58,59]. The power of SKAT tests are likely limited by a small cohort size, the heterogeneous genetic backgrounds of the case and control populations, the differences in exome sequence capture and sequencing chemistries employed, the absence of an inheritance model, and most importantly the underlying complex oligogenic architecture of cardiac malformations [12,60,61]. Finally there are no statistical models that account for ethnicity in models of rare-variant transmission, therefore the influence of population stratification or ethnicity upon our rare-inheritance model of disease is not known.
Overall our analysis suggests locus heterogeneity in the pathophysiology of a single cardiac developmental malformation. We observed recurrent variation within three genes (GATA4, MYH6, and BBS2) in only 6 of 59 trios. Including inherited, de novo, and discovered loci, 32% of trios displayed one or more putatively contributory mutations in the 19 genes identified. Supported by both experimental and clinical evidence, we suggest that inherited rare variants with a moderate effect size across multiple loci may impact the risk of congenital heart disease in addition to de novo variation. Taken together our catalog of 19 loci with experimental evidence for disease among 159 patients is consistent with the long-hypothesized oligogenic inheritance of congenital heart disease [12].
The guidelines of the Declaration of Helsinki were followed and the study was approved by the institutional review board of Stanford University (IRB-23637, IRB-23572) along with each institution from which participants were recruited. Written informed consent was obtained from each participant. Trios or single patients with an AVSD and situs solitus were recruited. We excluded patients with other major congenital malformations, developmental delay, or other types of CHD (excluding patent ductus arteriosus or secundum atrial septal defect). All participants were assessed clinically by pediatric cardiologists or clinical geneticists to exclude other syndromes associated with AVSD or CHD. Participants were obtained from Seattle Children’s Hospital [62], the Pediatric Cardiac Genomics Consortium, the University of Iowa, and reanalyzed from a published study of AVSDs drawn from patients at the University of Toronto [7] (S2 Table).
Exome sequences for 531 control subjects of Caucasian and African American descent without CHD were derived from the atherosclerosis risk in communities (ARIC) consortium warehoused at dbGAP, and from local recruitment efforts. An additional 59 control trios (healthy parents with a healthy child) were obtained from the Simons Foundation. Raw sequence data in the form of bam or fastq files was re-aligned and re-analyzed with the below-described bioinformatic pipelines.
DNA was isolated by standard techniques from either whole blood, saliva samples, or immortalized lymphoblasts. Exome sequencing was performed for all complete trios and single affected individuals at two academic centers (University of Washington or Yale University) and two commercial sequencing providers with the SeqCap EZ Human Exome Library v2.0 (Roche NimbleGen, Madison, Wisconsin, USA), SureSelectXT Human All Exon V4 (Agilent Technologies Inc., Santa Clara, California, USA), or a proprietary capture library based on the Agilent SureSelectXT Human All Exon V5 platform (Personalis, Corp, Menlo Park, California, USA). Paired-end sequencing was performed on Illumina HiSeq 2500 machines with 75-, 100-, or 150-bp read lengths in all but two trios, which were sequenced with 33 bp paired end reads. For all included samples Median Ts/Tv was 3.10 while coverage depth was 43.8x (S3 Table). Exome sequencing on single individuals from the Toronto cohort was performed as described [7], and for subset of unrelated individuals raw fastq files were obtained and reanalyzed via the below described bioinformatics platform for the purposes of reanalysis.
AVSD trios and the singleton/replication patients not originating from Toronto were assayed for CNVs by array comparative genomic hybridization (CGH) by using a custom chip described previously or the SurePrint G3 Human CGH Microarray Kit, 8x60K (Agilent Technologies Inc., Santa Clara, California, USA) with described protocols [31,62]. When a CNV was detected in an affected participant, a dye-swap experiment was repeated to ensure reproducibility and the parents were then assayed if available to determine inheritance status. For detection of smaller variation, two exome CNV detection assays using read depth data were also employed for all participants [63,64]. When there was agreement in a structural variant call between at least two calling methods for any variant of interest, we employed an orthogonal CNV genotyping assay when DNA was available (for all of the AVSD trios and the singleton/replication patients not originating from Toronto). The CNV genotyping assay was carried out with 10 ng of DNA per manufacturer instructions against the RNaseP CNV reference assay (Life Technologies, Carlsbad, California, USA). The assay was run on a ViiA 7 Real Time PCR System (Life Technologies, Carlsbad, California, USA) for 40 cycles under standard reaction conditions, and CNV genotypes were called with the copycaller software [65].
Two genotyping pipelines were tested. A rapid and sensitive commercial software package, rtg-core version 3.3.2 was applied to the raw exome sequence data for mapping, pedigree-aware variant calling, and genotype filtration (Real Time Genomics Inc., Hamilton, New Zealand) [66] to the UC Santa Cruz human genome reference sequence (hg19) (S1 Script). A second pipeline based on the HugeSeq BWA/GATK HaplotypeCaller pipeline was also employed for purposes of comparison.
To determine the accuracy of our genotyping pipeline we re-genotyped available exome data from NA12878 for comparison to a recently described gold standard dataset of 24,734 variants from this individual [14]. This set of variants from the consensus standard dataset was limited to the exome capture region of the nextera kit all human exome v2 (Illumina Corp, San Diego, California, USA) and regions of suspicious variant quality [67] were excluded to yield a true positive dataset of 24,734 true positive variants from NA12878. The vcfeval tool from RTG was used for all comparisons of vcf files as it robustly handles the different possible textual representations of insertions, deletions, and substitutions that may be produced by different genotyping algorithms. Using raw fastq files generated by the Garvan Institute, comparing the unfiltered output of both pipelines the two algorithms both called 19,566 variants (79.9%) of the NA12878 true positive dataset (TP) in common. RTG called an additional 1,469 TP, compared to BWA/GATK which only called an additional 198 unique TP variants. Overall RTG displayed a greater unfiltered sensitivity at 84.5% compared to 79.9% for BWA/GATK; therefore we selected the RTG pipeline for further analysis of the cohort (S1 Table).
To classify variants we selected an AVR score of 0.5 to balance a sensitivity of 99.4% and positive predictive value of 90.3% for the purposes of variant discovery. Alignments for all variants of interest were manually inspected with the IGV software, and all de novo coding variants and any insertion/deletion of interest was confirmed by direct Sanger sequencing of PCR amplicons, or alternately clonal Sanger sequencing of 12 colonies from cloned PCR amplicons. For burden testing, pooled simultaneous variant calling for all included cases and controls was performed with the RTG population caller and variants filtered for a read depth of 8 and AVR score of 0.5. Analyses were limited to the regions of intersection of the bed files of the exome capture kits obtained from the respective manufacturers. Known artifactual variants arising from exome sequencing were removed at the time of variant filtration [67,68].
All statistical analyses employed the R language for statistical computing version 3.1 unless otherwise specified. For population analyses we selected 8,940 snps from the Affymetrix Genome-Wide Human SNP Array 6.0 (Affymetrix, Santa Clara, California, USA) common to the five exome capture protocols employed in the study. All included probands were re-genotyped for this limited set of variants, and combined with individual level data from 1032 diverse samples of known ethnicity from the 1000 genomes project. The MDS and kinship modules from the KING software were used to estimate ethnic background and five individuals displaying cryptic familial relationships from the Toronto cohort among the singletons were identified and excluded from further analysis [69]. A multinomial linear model was built for each population and the presence of admixture from the populations of the 1032 known samples, and used to infer ethnic background and the presence of admixture in the included probands. Self-reported ethnicity was available for 547 individuals, which was 97.6% concordant with predicted ethnicity.
Protein altering variants were sorted for minor allele frequency less than 0.03 and prioritized by inheritance patterns consistent with rare disease (de novo, rare homozygous, and compound heterozygous) using the trioTools module from the STMP package [16]. Protein-altering variants from the 59 trios were selected, haplotypes constructed, and variants phased with the PLINKseq software package [70]. Imputation was not performed thus all variants analyzed originated from the primary genotyping pipeline.
After sorting for inheritance consistent with rare disease (de novo, rare homozygous, and compound heterozygous), variants were collated and assembled into lists. A blacklist of genes with low prior likelihood of causality was excluded from further analysis, which included genes with a residual variation intolerance score greater than 90 and genes with copy number polymorphisms [68,71,72].
For statistical comparisons between the groups of trios, the number of individuals with rare inheritance (de novo, rare homozygous, or compound heterozygous) in these genes in the 59 trios and 59 controls was counted and compared with a Fisher’s Exact test. As the underlying distribution combining de novo, rare homozygous, and compound heterozygous inheritance into a unified “rare inheritance” model is not readily estimated from available genotyping data, and because variant detection (particularly de novo variation) may vary systematically with the technical aspects of sequencing (variant calling algorithm, read depth, exome capture platform, and sequencing chemistry) we estimated empiric p-values by Monte Carlo simulation. To simulate an underlying distribution we performed permutations drawing 16 genes randomly from a list of 18,495 protein-coding genes, the rare inheritance events for each random list counted within 59 cases and 59 controls by individual, and a Fisher’s exact test applied.
The p-value was estimated with the formula (r+1)/(n+1) where n is the number of simulated replicate samples and r is the number of test statistics exceeding the calculated test statistic from the observed data (p = 9.60e-06). This simulation procedure suggested an empiric p-value of 1.23e-06, a similar order of magnitude as the p-value calculated from the observed data.
Gene modules or subnetworks identified by protein interaction networks in the HotNet2 algorithm (see below), were subjected to burden testing with the SKAT linear weight test and a Bonferroni correction for multiple hypothesis testing employed [73,74]. The first four principal components from a principal components analysis of all variants were used as covariates for burden testing. Within the SKAT algorithm variant weighting was derived from a beta density function (pi, 1, 25) where pi is the minor allele frequency. For variant weighting, minor allele frequencies were derived from the 61,428 individuals in the multiethnic EXaC dataset (version 0.2 http://exac.broadinstitute.org/), while for variants not observed or reported in the EXaC dataset minor allele frequencies were calculated from frequency observed among the genotyped individuals. The quantile-quantile plots for the SKAT linear weight test of 86 subnetworks (inclusive of combinations of 231 single genes—see below) suggest that the test-statistics derived are controlled for ethnicity or other systematic differences between cases and controls such as coverage (S2 Fig).
Unique tag-count data from 74 SAGE libraries representing 22 tissues constructed as a part of the mouse atlas of gene expression project were downloaded from www.mouseatlas.org [17]. SAGE tags were mapped with the Burrows-Wheeler aligner to the 115,746 unique mouse RefSeq transcripts downloaded from UC Santa Cruz website (http://genome.ucsc.edu) and tag data converted to a digital gene expression format constituting a tag counts per transcript using custom Perl scripts. Tags matching pseudogenes were removed. Using the R environment for statistical computing, tag counts by transcript were normalized by library size with the EdgeR package and collapsed to gene by connectivity with the WGCNA package [75,76]. A standard WGCNA workflow for digital gene expression was applied to the normalized and collapsed data for coexpression module construction followed by correlation to the tissue of origin and annotation with the human orthologous gene name when available (S2 Script). To independently assess the predictive capacities of our unsupervised network building procedure, we observed that the developmental expression of 13 well characterized congenital heart disease genes are accurately localized by the network model to their appropriate cardiac tissue, suggesting excellent specificity for detecting developmental cardiac related gene expression (S9 Table). Comparison of variant and gene lists for under- and over-representation were performed with a one-tailed hypergeometric test. Gene expression networks were visualized with the Gephi software package [77]. Because the DMP is a key structure in development of the atrioventricular septum [78] and was not explicitly included in the original set of micro-dissected tissues from the mouse atlas of organ development (www.mouseatlas.org), we added the DMP gene expression data to the cardiac development gene expression set. Among genes expressed in the DMP we selected the most highly expressed 1000 genes across 6 datasets generated from the posterior second heart field including the dorsal mesenchymal protrusion. The microdissection of the posterior heart field was performed in the laboratory of Dr. Moskowitz at embryonic mouse tissue at E9.5, subject to reverse-transcription, amplification, and sequencing by The University of Chicago Genomics Core (GSE75077).
Variant data annotated with the STMP package [16] including SNVs and CNVs from the 59 AVSD trio probands and 59 control trio probands was converted and formatted with custom python scripts, and included all protein altering variants with a minor allele frequency in EXaC of 0.02 or less were included in the analysis using four included protein interaction networks. For each of three protein interaction networks (Multinet, IrefIndex9, and HINT [22]) four delta values were derived with the network permutation procedure and applied to identify subnetworks. The resulting subnetworks identified by the HotNet2 algorithm were manually inspected for validity, and the subnetworks from the MultiNet protein interaction networks were chosen for further analysis. For each set of subnetwork sizes ranging from 2 to 10, the HotNet2 algorithm derives a p-value from the hypergeometric distribution comparing the observed number of subnetworks of size n within in a dataset compared to an expected number of subnetworks. The expected number of subnetworks is derived from a computationally intensive network permutation procedure; in this case the MultiNet protein interaction network was subject to 100 permutations limiting the range of p-values to a minimum p-value to 0.01 and maximum p-value to 1. For the AVSD-trio subnetworks the median p-value was 0.01 across the nine subnetwork sizes and four derived delta values, in comparison to the control-trio subnetworks where the median p-value was 1.0; this suggested an enrichment in variants occurring in genes with true protein-protein interactions within the AVSD-trios (rather than randomly occurring simulated protein-protein interactions in the permuted networks). A single output run utilizing the MultiNet protein interaction network applying a delta value of 0.00126036397514 yielded 86 subnetworks containing 231 genes in the AVSD trios, and the 86 subnetworks were subject to burden testing in the singleton cohort (see above). Protein interaction data were processed with custom python and shell scripts for visualization using the Gephi software tool [77].
Live breeding pairs of the B6;129P2-Nr1d2tm1Dgen/H mouse line (hereafter referred to as Nr1d2tm1-Dgen) were obtained from the European Mouse Mutant Archive (Munich, Germany). Animals were in housed and cared for in AAALAC accredited facilities under standard conditions with oversight and approval from the Stanford University APLAC committee (protocol APLAC-11334). Euthanasia was carried out under anesthesia with isofluorane following APLAC and AAALAC guidelines using carbon dioxide followed by cervical dislocation. Genotyping was performed by PCR of toe or tail clippings with gel electrophoresis using standard methodology with two primer pairs (CAAGTAACAAGCCTGGGACATAAAG and CTTCGTAGAGGGAGTAATATGACAC yield a 517 bp PCR product from the WT allele; CAAGTAACAAGCCTGGGACATAAAG and GACGAGTTCTTCTGAGGGGATCGATC yield a 757 bp product from the knockout allele). Two pairs of heterozygous founder animals were bred, yielding 17 pups (2 wild type, 7 Nr1d2tm1-Dgen +/-, and 8 Nr1d2tm1-Dgen -/-) which did not deviate obviously from expected Mendelian allelic ratios. Spontaneous death within hours after birth occurred in a single +/- and two -/- animals. The thoracic and abdominal cavities of spontaneously deceased homozygous knockout animals were visually inspected to examine the great vessels and visceral situs which were normal, and hearts dissected out and subject to sectioning and H & E staining by standard techniques. Sectioning of tissue samples was performed following either embedding of frozen sections followed by dehydration and fixation in 10% neutral buffered formalin or alternately fixation in 3% paraformaldehyde followed by alcohol dehydration and paraffin embedding. We performed additional matings of +/- and -/- animals, and sacrificed mothers to obtain embryos at e16.5 and e17.5. Visualization was performed by brightfield microscopy on a Nikon 90i Eclipse upright with a DS Fi1 camera with a 20x objective. Both P0 Nr1d2tm1-Dgen -/- animals displayed atrioventricular septal defects, a single -/- animal at e17.5 displayed an inlet ventricular septal defect, and single -/- animal at e16.5 displayed an AVSD. In total, 4 out of 15 -/- hearts assayed displayed a cardiac defect. Spontaneously deceased animals were analyzed, and additionally euthanasia of pregnant female mice was performed to obtain embryonic animals.
The crystal structure of the DNA binding domain of NR1D1 (RCSB 1A6Y) was visualized with the PyMOL Molecular Graphics System, Version 1.7.4 Schrödinger, LLC (New York, New York, USA). A construct containing wild-type murine Nr1d2 cDNA construct under control of a CMV promoter in the pCS6 expression vector was obtained from Transomic Technologies Inc. (Huntsville, Alabama, USA), and subject to a site-directed mutagenesis yielding a codon switch at p.R175W which corresponds to the de novo mutation observed in the subject with AVSD (S4a Fig). An Nr1d2 response element vector was constructed cloning 5 tandem repeats of an evolutionarily conserved NR1D2 binding site REV-DR2 [80,81] upstream of a minimal CMV promoter driving GFP expression using the pSF-MinCMV-daGFP vector (Sigma-Aldrich Inc, St. Louis, Missouri, USA) (S4b Fig). Site-directed mutagenesis and cloning were performed by a commercial provider GENEWIZ Inc. (South Plainfield, New Jersey, USA), and sequence verified in our own laboratory. The three vectors were subject to routine endotoxin free preparation. Commercially available primary HUVEC cells obtained from Cell Applications (San Diego, California, USA) were seeded at 80% confluency in black transparent-flat bottom 96 well plates (Greiner, North Carolina, USA) and transfected the following day with 100ng of each vector (response-element, wild-type or p.R175W) using Lipofectamine 3000 (Life Technologies, Grand Island, New York, USA). Twenty-four hours after transfection, GFP fluorescence was measured on a Tecan Infinite M1000-multimode plate reader (Tecan Group Ltd, Mannendorf, Switzerland). We performed three transfection conditions including response-element+Nr1d2-WT, response-element+Nr1d2-P403W and response-element alone in 24 technical replicates. Autofluoresence of untransfected wells were averaged, and subtracted from the response-element alone and the two experimental conditions. The highest and lowest value from each condition were excluded from analysis yielding 22 technical replicates per experimental condition, and 4 technical replicates for the response element alone. Statistical analysis and graphing of the transfection experiments was performed in Prism 6 Graphpad Software (La Jolla, California, USA).
|
10.1371/journal.pgen.1005923 | NCP1/AtMOB1A Plays Key Roles in Auxin-Mediated Arabidopsis Development | MOB1 protein is a core component of the Hippo signaling pathway in animals where it is involved in controlling tissue growth and tumor suppression. Plant MOB1 proteins display high sequence homology to animal MOB1 proteins, but little is known regarding their role in plant growth and development. Herein we report the critical roles of Arabidopsis MOB1 (AtMOB1A) in auxin-mediated development in Arabidopsis. We found that loss-of-function mutations in AtMOB1A completely eliminated the formation of cotyledons when combined with mutations in PINOID (PID), which encodes a Ser/Thr protein kinase that participates in auxin signaling and transport. We showed that atmob1a was fully rescued by its Drosophila counterpart, suggesting functional conservation. The atmob1a pid double mutants phenocopied several well-characterized mutant combinations that are defective in auxin biosynthesis or transport. Moreover, we demonstrated that atmob1a greatly enhanced several other known auxin mutants, suggesting that AtMOB1A plays a key role in auxin-mediated plant development. The atmob1a single mutant displayed defects in early embryogenesis and had shorter root and smaller flowers than wild type plants. AtMOB1A is uniformly expressed in embryos and suspensor cells during embryogenesis, consistent with its role in embryo development. AtMOB1A protein is localized to nucleus, cytoplasm, and associated to plasma membrane, suggesting that it plays roles in these subcellular localizations. Furthermore, we showed that disruption of AtMOB1A led to a reduced sensitivity to exogenous auxin. Our results demonstrated that AtMOB1A plays an important role in Arabidopsis development by promoting auxin signaling.
| MOB1 protein is a key component of the Hippo signaling pathway in animals, and it plays critical roles in organ size control. The plant hormone auxin regulates many aspects of plant growth and development including organogenesis. In this work, we showed that AtMOB1A, which is highly homologous to animal MOB1 proteins, plays an important role in plant organogenesis. Furthermore, we demonstrated that AtMOB1A synergistically interacts with auxin biosynthesis, transport, and signaling pathways to regulate Arabidopsis development. We further showed that AtMOB1A likely controls plant development by promoting auxin signaling. This work identified a new player in auxin-mediated plant development and lays a foundation for further dissection of the mechanisms by which auxin regulates organogenesis.
| In recent years, the Hippo signaling pathway has emerged as a very important pathway for animal development [1]. This highly conserved pathway was initially identified in Drosophila as a key pathway controlling organ size, and later was shown to play a role in controlling cell fate and pattern formation in mammals [2–5]. The core part of the pathway is a phosphorylation cascade composed of four key components in mammals and Drosophila: a Ste20-like Ser/Thr protein kinase Mst1/2 [Hippo (Hpo) in Drosophila] [6,7], an NDR-family protein kinase Lats1/2 [Warts (Wts) in Drosophila] [8,9], and two kinase regulatory components, Sav and MOB1 (Sav and Mats in Drosophila) [10,11] (S1 Fig). Mst1/2 phosphorylates MOB1 and Lats1/2, and activates Lats1/2. MOB1 can bind to Lats1/2 and potentiate its intrinsic kinase activity. The activated Lats1/2 in turn phosphorylates and inactivates a transcriptional co-activator YAP/TAZ (Yorkie in Drosophila) [12]. YAP/TAZ is an effector of the Hippo pathway. Phosphorylation of YAP/TAZ results in its cytoplasmic retention, largely by facilitating its interaction with 14-3-3 proteins. Dephosphorylation of YAP/TAZ promotes its nuclear localization where it interacts with transcription factors and regulates gene expression. Drosophila mutants of core components in this pathway, such as hpo, wts, mats, sav, showed larger organs. In mammals, Hippo signaling controls patterning and differentiation of airway epithelial progenitors, mammary gland differentiation, intestinal fate, cardiovascular, liver, pancreas, central nervous system, and lymphocyte development [2]. It also regulates stem cell self-renewal and cell polarity in animals [2,13,14]. Recently, it was reported that the Arabidopsis thaliana MOB1A gene is required for tissue patterning of the root tip [15] and the development of both sporophyte and gametophyte [16]. MOB1 proteins in plants and animals share high sequence homology [11]. It is tempting to hypothesize that the Hippo pathway may also function in plants. However, very little is known regarding how the hypothesized Hippo pathway may regulate plant growth and development.
The plant hormone auxin plays critical roles in plant growth and development. Local auxin biosynthesis, polar transport, and auxin signaling all contribute to proper plant growth and development. The best characterized tryptophan-dependent auxin biosynthesis pathway is the indole-3-pyruvate pathway, in which tryptophan is converted into indole-3-pyruvate by TAA/TAR family of amino transferases. Indole-3-pyruvate is then converted into IAA by YUC family of flavin-containing monooxygenases [17–21]. Auxin biosynthesis is temporally and spatially regulated [22,23]. Auxin transport is carried out by auxin influx carriers AUX1/LAXs, auxin efflux carriers PINs, and ABCB transporters [24]. Both local auxin biosynthesis and polar transport are important for generating auxin gradients and maxima, which are perceived by auxin receptors. The best characterized auxin receptor is TIR1/AFBs and Aux/IAA co-receptor complexes [25,26]. Disruption of auxin biosynthesis, polar transport or signal transduction pathways leads to defects in almost every aspect of developmental processes, such as flower, embryo, root, and leaf development [22,27,28]. For example, auxin biosynthetic mutants yuc1/4/10/11 quadruple mutants are defective in embryogenesis, and auxin signaling mutants such as mp fail to develop normal hypocotyls and roots [23]. Auxin transport mutant pin1 develops pin-like inflorescences, which was also observed in auxin signaling mutant mp and npy mutants [29,30]. Although it has been well documented that auxin plays essential roles in plant development, little is actually understood regarding how auxin gradients are translated into guiding proper developmental events.
In this paper, we provide evidence that links AtMOB1A, which is homologous to a key component of the animal Hippo pathway, to auxin-mediated plant organogenesis and development. We conducted a genetic screen for mutants that could enhance the phenotypes of pid, which is defective in auxin signaling and transport [31,32]. One of the pid enhancers, ncp1 (no-cotyledon in pid 1) failed to develop cotyledons in pid background. We further showed that ncp1 single mutant displays strong developmental defects in early embryos, seedlings, and in adult plants. NCP1 encodes a protein with significant homology to the animal MOB1s, a core component of the Hippo pathway. We showed that NCP1/AtMOB1A probably has biochemical activities similar to those of animal MOB1, because the Drosophila MOB1 (Mats) can fully rescue the developmental defects of ncp1/atmob1a. The atmob1a mutant showed synergistic genetic interactions with known auxin biosynthetic, transport, and signaling mutants, suggesting that AtMOB1A functions in parallel to auxin pathways or affecting some aspects of auxin biology. Furthermore, disruption of AtMOB1A led to a decrease in sensitivity to auxin treatments and down-regulation of auxin reporters including DR5-GFP, ProARF7:GUS, and ProARF19:GUS. Our findings demonstrate that AtMOB1A likely promotes auxin signaling, thus impacting various Arabidopsis developmental processes.
Genetic enhancement has been widely used to identify components in signaling and metabolic pathways. We previously identified npy1 as a genetic enhancer of yuc1 yuc4, which are defective in auxin biosynthesis. NPY1 is a key component of a signaling pathway responsible for auxin-mediated organogenesis [29]. Previous studies have shown that several Arabidopsis auxin mutants/mutant combinations, including npy1, yuc1 yuc4, wag1 wag2, pin1, and wei8 tar2, have no cotyledons when combined with pid, which encodes a protein kinase important for auxin signaling and transport [20,29,30,33]. Therefore, pid provides a sensitized background, and cotyledon formation serves as an easy phenotypic readout for us to genetically identify additional components in auxin-mediated plant development. We conducted a genetic screen for enhancers of pid and isolated a new mutant that lacked cotyledons. We name the mutant as ncp1 (no-cotyledon in pid 1). At seedling stage, ncp1 pid failed to develop cotyledons, but they appeared to have normal hypocotyls and roots (Fig 1A and 1C). The no-cotyledon phenotype of ncp1 pid was highly penetrant: the majority (90%) of the mutants completely lacked both cotyledons, while some plants occasionally developed one cotyledon (Table 1). Interestingly, ncp1 pid plants could develop true leaves, however, they were abnormal in morphology and vascular development (S2 Fig). The ncp1 pid plants were able to transition from vegetative growth to reproductive development, but their inflorescences were all pin-like and failed to produce fertile flowers (Fig 1B). The no-cotyledon phenotype in seedlings of ncp1 pid was caused by defects occurred during embryogenesis. In mature embryos, the cotyledon formation was abolished in ncp1 pid, while two cotyledons in WT and two or three cotyledons developed in pid (Fig 1D).
The observed no-cotyledon phenotype was dependent on the presence of the pid mutation. We genotyped 48 individual plants that showed the no-cotyledon phenotype and found out that they were all pid homozygous, suggesting that the phenotype was dependent on the presence of the pid mutation. We further analyzed the progenies from a single ncp1+/- pid+/- plant, 22 of 427 seedlings (about 1/20) showed the no-cotyledon phenotype, indicating that the phenotype was caused by two un-linked recessive mutations, i.e. pid and ncp1.
We crossed ncp1+/- pid+/- to Arabidopsis Landsberg ecotype and allowed the F1 plants to self-fertilize to generate a mapping population. In the F2 mapping population, we isolated 1325 seedlings that failed to develop cotyledons from about 26,000 F2 individuals. We found that the no-cotyledon phenotype was linked to two genetic loci: one on the bottom arm of chromosome II and the other on chromosome V. The Chromosome II locus is pid, further supporting that the no-cotyledon phenotype was dependent on pid. We narrowed the mapping interval on Chromosome V down to about 140 kb, between the two genetic markers on K9E15 and MRA19 (Fig 1E). We sequenced all of the open reading frames (ORFs) in the mapping interval and identified a G to A conversion at the splicing junction of the second intron and the third exon of the gene At5g45550. Further analysis of At5g45550 cDNA from the ncp1 mutant plants revealed that the mutation caused a single base-pair shift of the splicing acceptor of the second intron and the deletion of the first G of the third exon. The mutation led to a frame shift after the Lys24, and introduced a premature stop codon (Fig 1F). Therefore, this mutant is likely a null allele.
To further confirm that the identified mutation in At5g45550 was responsible for the observed no-cotyledon phenotype in pid background, we transformed a genomic fragment containing the coding region and its up- and down-stream regulatory sequences of At5g45550 into ncp1-/- pid+/- plants. All of the T1 transgenic plants (341 in total) had two or three cotyledons. We genotyped the T1 plants and found that 86 of them were double mutants, indicating that wild type (WT) copy of At5g45550 complemented the phenotype (Fig 1G).
We also identified a T-DNA insertion allele of ncp1 (GK_719G04) from the NASC stock center, and named it ncp1-2. We generated double mutants ncp1-2 pid and ncp1-2 pid-714 (SAIL_770_E05). Both of the double mutants displayed the same no-cotyledon phenotype as ncp1-1 pid (S3 Fig). Therefore, we conclude that At5g45550 is NCP1 and the identified mutations in At5g45550 are responsible for the no-cotyledon phenotype in pid backgrounds. We used ncp1-1 allele for further detailed analysis and genetic interaction studies in the paper.
NCP1 was identified in the pid mutant background. We segregated out pid and investigated whether ncp1-1 mutation alone caused any developmental defects. At seedling stage, ncp1-1 had shorter root meristems zones when compared to WT. The root phenotypes of ncp1-1 were caused by decreased cell numbers in its root meristem (S4E and S4F Fig). Compared to WT plants, ncp1-1 single mutant plants were slightly taller with shorter siliques and smaller flowers (S4A–S4C Fig). The mutant was much less fertile. Our observed phenotypes of roots, siliques and flowers were consistent with previous findings from the analyses of the T-DNA allele of AtMOB1A, GK_719G04 (ncp1-2) [15]. The shorter root phenotype and the decrease in cell numbers in the root meristems of ncp1 and ncp1 pid may be caused by defects in cell division. To test this hypothesis, we investigated cell division activities in ncp1 and ncp1 pid mutants. CycB1;1:GUS is a widely used marker for the G2/M phase of the cell cycle [34]. The GUS staining domains were dramatically decreased in both ncp1 and ncp1 pid mutants, indicating that cell division activities were decreased in these mutants. These findings could partially account for the observed short root phenotypes of ncp1 and ncp1 pid (S4I and S4J Fig).
Because the no-cotyledon phenotype of ncp1 pid was caused by defects occurred during embryogenesis, we also analyzed whether disruption of AtMOB1A alone is sufficient to affect embryogenesis. Another indication that AtMOB1A is important for embryogenesis is that about 66.6% of the embryos (n = 785) were aborted in the ncp1-1 siliques (S4D Fig). We carefully analyzed various stages of embryogenesis of the ncp1-1 mutants and discovered that AtMOB1A plays an important role in early embryogenesis. The cell division in some mutant embryos was disturbed as early as 8-cell embryo stage. The upmost suspensor cell divided longitudinally in some ncp1-1 embryos whereas the cell in WT divides horizontally (Fig 2C). At 16-cell stage, both the pro-embryos and suspensor cells were abnormal in ncp1-1 (Fig 2D). At globular stage, the upmost suspensor cell became the hypophysis and remained as a lens-shape cell in WT [35]. But in ncp1-1 mutant, it was no longer lens-shaped and was divided into 2 cells (Fig 2E). The observed defects in the embryogenesis of ncp1-1 mutant would severely affect its embryo development, indicating that NCP1 is important for embryogenesis.
The predicted NCP1 protein contains 215 amino acid residues. It shares high sequence homology (63% identity) to the Drosophila MOB1 (Mats) (S5 Fig). MOB1 was first identified in yeast as Mps One Binder 1, an essential protein required for the completion of mitosis and maintenance of ploidy [36]. It has been shown that MOB1 is a key component of the Hippo signaling pathway [11]. In the Arabidopsis genome, there are four MOB1-like genes, At5g45550, At4g19045, At5g20430, and At5g20440. They have been renamed as AtMOB1A, AtMOB1B [15], AtMOB1C, and AtMOB1D herein, respectively.
MOB1 is highly conserved in plant species. For example, the MOB1As of Brassica rapa and Arabis alpina are almost identical to AtMOB1A: they differ from AtMOB1A in only one and three amino acid residues out of 215, respectively. Other putative plant MOB1A proteins and AtMOB1A share more than 90% identities (S5 Fig). It is clear from our phylogenetic analysis that animal MOB1 proteins and plant MOB1s belong to different clades (S6 Fig). The MOB1 genes have duplicated in the most recent common ancestor of land plants (Embryophyte) during evolution, and evolved with frequent duplication or deletion in the derived lineages of land plants. Selaginella moellendorffii only has one MOB1 gene whereas Physcomitrella patens has two. Monocots and dicots often have two to four copies of MOB1 genes (S6 Fig).
To further demonstrate that NCP1 is functionally related to MOB1 homologs from other organisms, we put the Drosophila MOB1 (Mats) under the control of the NCP1 promoter and transformed the construct into ncp1-1. We confirmed that all of the 88 transgenic ncp1-1 seedlings contained the Mats gene. The adult plants of ncp1-1 mutant showed severe defects in fertility, but the Mats transgenic ncp1-1 mutant plants were able to produce siliques like WT (Fig 3). Our results indicated that the Drosophila Mats gene complemented the defects caused by ncp1-1 mutation, and the function of MOB1/Mats is conserved from plants to Drosophila.
To investigate the expression pattern of NCP1/AtMOB1A, we generated a construct containing the NCP1 genomic DNA including its regulatory and coding sequences, with the GFP gene inserted immediately before the stop codon. We transformed ncp1 and ncp1 pid+/- mutants with this construct and found that the construct complemented both ncp1 and ncp1 pid, indicating that the NCP1-GFP fusion protein was fully functional. AtMOB1A is uniformly expressed in embryonic and suspensor cells from one-cell to mature embryo stages. The expression patterns of AtMOB1A are consistent with its role in embryo development. AtMOB1A protein is localized to nucleus, cytoplasm and associated to plasma membrane (Fig 4). The observed nuclear localization was consistent with previously findings [16,37].
The ncp1-1 mutant was isolated as an enhancer of pid, which is a well-known auxin mutant. We further analyzed whether ncp1 could genetically interact with other known auxin mutants. We tested three groups of auxin mutants that are defective in either auxin biosynthesis, or transport, or signaling.
It has been shown that YUC flavin-containing monooxygenases and TAA1/TAR tryptophan amino transferases define a main auxin biosynthetic pathway in Arabidopsis [19,20]. Both YUCs and TAAs play essential roles in all of the major developmental processes including embryogenesis and flower development in Arabidopsis [18,22,23]. When we disrupted NCP1 in yuc1 yuc4 background, the resulting triple mutants developed pin-like inflorescence whereas yuc1 yuc4 never form pins, demonstrating that ncp1 greatly enhanced the phenotypes of auxin biosynthetic mutants (Fig 5A and 5C and 5D).
We previously reported that NPY1 is involved in auxin-mediated organogenesis. The pid npy1 double mutants had no cotyledons, and npy1 yuc1 yuc4 triple mutants developed pin-like inflorescences. NPY1 is proposed to play a role in auxin transport and signaling [29,30]. When we introduced ncp1 into npy1 background, the double mutants produced pin-like structures whereas either single mutant did not form any pins (Fig 5B and 5C and 5D).
We next tested if ncp1 could synergistically interact with auxin signaling mutants. TIR1/AFBs are the best characterized auxin receptors responsible for regulating expression of auxin inducible genes. We crossed ncp1 to tir1-1 afb2-1 afb3-1 [27] and obtained various combinations of ncp1 and tir1 afb mutants from the F2 populations. The phenotypic analysis was performed in F4 generation. The single mutants of tir1-1, afb2-1, afb3-1, and combinations of their double mutants did not display dramatic developmental defects under normal growth conditions [27]. Interestingly, the ncp1 tir1-1 double mutants showed severe reduction in fertility, which was caused mainly by the defects in gynoecium patterning. The defects were further enhanced in ncp1 tir1-1 afb2-1 afb3-1 and led to complete sterility (Fig 6A and 6B). The adult plants of tir1-1 afb2-1 double mutants showed a reduction in rosette leaf size and inflorescence height, but their seedlings were similar to WT (Fig 6C) [27]. However, the ncp1 tir1-1 afb2-1 triple mutants exhibited strong developmental defects. Six of 34 (18%) triple homozygous seedlings of ncp1 tir1-1 afb2-1 mutants had no roots, whereas the tir1-1 afb2-1 double mutants never displayed such phenotypes. The observed no-root phenotypes closely resembled those of bdl/iaa12 or mp/arf5 mutants. The tir1-1 afb2-1 afb3-1 and tir1-1 afb1-1 afb2-1 afb3-1 mutants also showed the mp-like rootless seedling phenotypes at frequency of 36% and 49%, respectively (Fig 6C) [27]. Our results indicated that ncp1 genetically interacts with auxin signaling pathway.
In Arabidopsis, PID has three close homologs: WAG1, WAG2, and PID2, which redundantly control cotyledon development [30]. The ncp1 pid double mutants phenocopied the pid wag1 wag2 pid2 quadruple mutants (S7A and S7B Fig). We further tested if ncp1 could enhance pid wag1 wag2 pid2 phenotypes. The double or higher orders of mutant combinations between ncp1 and wag1 wag2 pid2 did not show obvious phenotypic enhancement, suggesting that PID played a more predominant role in regulating cotyledon development than WAG1, WAG2, and PID2. The ncp1 pid wag1 wag2 pid2 quintuple mutants displayed no-cotyledon phenotype similar to that of pid wag1 wag2 pid2. However, the quintuple mutants showed strong developmental defects in true leaves (S7 Fig). In dark grown seedlings, the initiation of true leaves was delayed in the quintuple mutants, compared to ncp1 pid (S7C Fig). In 14-day-old light grown seedlings, the quintuple mutants developed single or two leaves, and occasionally developed a pin-like true leaf (S7D and S7E Fig). In 36-day-old plants, the quintuple mutants showed two types of phenotypes. The type I plants (44%, n = 61) developed one to three true leaves and a pin-like inflorescence, and were arrested at this developmental stage. The type II plants (56%, n = 61) could produce more than three true leaves, and continued to grow with the phenotypes similar to those of ncp1 pid (S7F Fig).
We showed that NCP1 genetically interacted with PID to control cotyledon development in Arabidopsis. In animals, MOB1 physically interacts with and activates NDR/LATS through recruitment to the plasma membrane [38,39]. Because both PID and NDR/LATS are AGC kinases, we hypothesized that AtMOB1A may use a mechanism analogous to that of animal MOB1. PID may play a role equivalent to that of NDR/LATS. To test this hypothesis, we conducted both pull-down and Co-IP assays to determine whether AtMOB1A physically interacts with PID/WAGs. However, we did not detect direct physical interactions between NCP1/AtMOB1A and PID, or WAG1/2 in our experiments (S8 Fig). These results suggested that there may not be direct interactions between NCP1 and PID/WAGs, or the interactions are transient and difficult to be detected under our assay conditions. There are at least 39 AGCs in Arabidopsis. The observed genetic synergism of AtMOB1A with PID may suggest that AtMOB1A is necessary for the function of other AGCs that have overlapping functions with PID/WAG1/WAG2.
To assess the role of NCP1 in auxin response, we introduced the auxin reporter DR5-GFP into ncp1, pid, and ncp1 pid mutants background. At heart and torpedo stages of embryogenesis, strong DR5-GFP signals were observed at the cotyledon primordia and hypophysis in WT. In pid, the DR5-GFP signals remained similar to WT. In contrast, the GFP signals were significantly decreased at the cotyledon primordia in ncp1 single and ncp1 pid double mutants. It is worth noting that the auxin responses at hypophysis seemed not changed in ncp1 single and ncp1 pid double mutants (S9A Fig). These observations suggested that NCP1 might be involved in auxin signaling.
It is known that auxin is required for the initiation and growth of lateral root (LR) and root hairs, and exogenous auxin can stimulate these developmental processes [40]. To further investigate the roles of NCP1 in auxin responses, we examined the response of ncp1 mutant to exogenous auxin treatment. Four-day-old seedlings of WT, pid, ncp1, and ncp1 pid germinated on 1/2 strength of Murashige and Skoog medium (MS) plates were transferred and grew on 1/2 MS plates containing 50 nM 2,4-D, a synthetic auxin. It is obvious that the lengths of root hairs and density of LR/LR primordium were dramatically increased in WT, pid and ncp1, however the effects of exogenous auxin on ncp1 pid were much weaker (S9B–S9F Fig). This suggested that ncp1 pid double mutants are partially resistant to auxin in terms of root hair and LR initiation and growth. The pericycle cells in ncp1 and ncp1 pid were similar to WT, suggesting that the LR defects in ncp1 pid was likely due to slow LR primordium growth and the failure to emerge from the epidermis of the primary root. It could also be a defect in pre-branch site formation, which is not morphologically distinct.
ARF7 and ARF19 redundantly control LR development, and they are expressed in lateral and/or primary roots [41,42]. We analyzed the expression of ARF7 and ARF19 in seedlings of ncp1 pid mutants by using ProARF7:GUS and ProARF19:GUS reporter lines [41,42]. The expression levels of ProARF7:GUS and ProARF19:GUS were dramatically decreased in the LR primordium of ncp1, pid, and ncp1 pid mutants, compared to WT. The expression levels of ProARF19:GUS were also reduced in primary roots of ncp1, pid, and ncp1 pid mutants (S10 Fig). These findings suggested that the LR defects in ncp1 pid were partially caused by down-regulation of ARF7 and ARF19.
PIN1 plays an important role during embryogenesis [28]. It is reported that PIN1-GFP is asymmetrically localized on plasma membrane [43,44]. We introduced the PIN1-GFP marker into ncp1, pid, and ncp1 pid mutants, and carefully checked the subcellular localization of PIN1-GFP from transition to torpedo stage of embryogenesis. No obvious alteration of the subcellular localization of PIN1-GFP was observed in these stages. However, we found that the expression levels of PIN1-GFP were altered in ncp1 mutants compared to WT. At these stages, PIN1-GFP was mainly expressed at the cotyledon primordia and ground tissue, which formed a Y-shape pattern. At transition stage, the expression pattern of PIN1-GFP in ncp1 mutants was similar to that of WT. However, in ncp1 pid double mutants, PIN1-GFP was found to be mainly expressed at the epidermal cell layer of apical part of embryos and ground tissue, which were barely connected by weak PIN1-GFP-expressing cells (S11 Fig). This result suggested that NCP1 plays a role in controlling the expression pattern of PIN1. It was previously reported that the localization pattern of PIN1 appeared normal in roots of Mob1A RNAi seedlings [15]. The discrepancy between our findings and those of the previous study might be because of the tissue specificity.
The Hippo signaling pathway has been shown to play a critical role in organ size control and morphogenesis in animals, but it is still an open question whether the Hippo pathway exists in plants. Because MOB1 proteins share high sequence homology in animals and plants, it is tempting to hypothesize that the Hippo pathway may also exist and play a role in plant growth and development. Here we show that AtMOB1A is functionally conserved with the Drosophila protein because atmob1a was fully rescued by its Drosophila counterpart, suggesting that at least part of the Hippo pathway is functional in plants. NCP1/AtMOB1A synergistically interacts with key genes in auxin biosynthesis, transport, and signal transduction pathways to regulate Arabidopsis development. The observed synergistic genetic interactions and the decreased auxin responses in various ncp1 and auxin mutant combinations suggest that there is an intrinsic link between auxin pathway and the hypothesized Hippo pathway in plants. Our finding that the expression levels of ProARF7:GUS and ProARF19:GUS were dramatically decreased in ncp1 pid further supports the notion that AtMOB1A is important for auxin-mediated developmental processes. This work provides a genetic framework for the Hippo pathway in auxin-mediated plant development.
It was reported that about 2% of the progeny of AtMob1A RNAi silenced plants were tetraploid [16], which is a result of cell division defects. Auxin is also known to control plant development by regulating cell division and expansion. Therefore, AtMOB1A may be involved in auxin-controlled cell division. The mutants in animal Hippo pathway display defects in organ overgrowth [1], due to a loss of control of cell proliferation. In ncp1 mutant, the length and the cell number of root meristem were decreased compared to WT (S4 Fig). The different developmental outcomes between animals and plant mob1 mutants suggest that Hippo pathway/MOB1 protein may play different roles in plants and animals regarding cell proliferation. Recently, the Hippo pathway has been shown to control cell fate in animals. For example, the Hippo pathway activity is essential for the maintenance of the differentiated hepatocyte state. Acute inactivation of the Hippo signaling in vivo is sufficient to dedifferentiate adult hepatocytes into cells bearing progenitor characteristics [4]. In Arabidopsis, ncp1 yuc yuc4 and ncp1 npy1 mutants failed to develop flowers (Fig 5). The cotyledons were also eliminated in ncp1 pid, and the hypophysis was lost in ncp1 tir1-1 afb2-1 during embryogenesis (Fig 2 and Fig 6). The observed defects in organ and embryo development in these mutants indicated that the Hippo pathway also plays a critical role in determining cell fate in plants.
It has been shown that the Hippo pathway is highly conserved in mammals and insects. A human MOB1 gene rescued the developmental defects of the Drosophila MOB1 mutant mats [11]. We show that the Drosophila Mats fully rescued developmental defects of the Arabidopsis ncp1 mutant (Fig 3), indicating that at least some of the components of the Hippo pathway are conserved between plants and animals. This functional conservation of MOB1 proteins is consistent with the high similarities of their amino acid sequences (S5 Fig). It has been shown that MOB1 is a phospho-protein in animal systems. Phosphorylation of Thr12 and Thr35 of hMOB1 by MST1 or MST2 is required for the interaction of hMOB1 with NDR/LATS kinases in human [45,46]. Thr12 and Thr35 are absolutely conserved in MOB1s of plants and animals (S5 Fig). Both AtMOB1A and AtMOB1B were identified as phospho-proteins in a proteomic study [47], suggesting AtMOB1A/B is also phosphorylated by some kinase(s). AtMOB1A may also interact with Arabidopsis NDR/LATS kinases. In line with this hypothesis, there are eight NDR-like kinase genes in Arabidopsis [48], and they share high similarities with their human counterparts (S12 Fig).
It is well known that auxin promotes root hair and LR formation [40]. Gain-of-function mutant msg2 of Aux/IAA19 had severely reduced LR and LR formation was not normally induced by exogenous auxin [49]. Root hair and LR formation are also inhibited in arf7 arf19 double mutants [41,42]. pid did not show obvious defects in root development [31]. However, NCP1 and PID synergistically control LR formation and root hair growth in seedlings (S9 Fig). ncp1 pid also displayed strong defects in LR development in response to exogenous auxin treatment (S9 Fig). Expression levels of ProARF7:GUS and ProARF19:GUS were decreased in ncp1 pid. Moreover, ncp1 enhanced tir1-1 afb2-1 mutants’ phenotypes (Fig 6). These findings suggested that NCP1/AtMOB1A plays a positive role in promoting auxin signaling.
In the Hippo pathway, MOB1 binds and activates the AGC kinase NDR/LATS1/2 [38,39]. In Arabidopsis, there are 39 AGC kinases [48]. Some of them have been demonstrated to be involved to auxin pathways, such as PID/WAGs and D6PKs [31,50,51], which phosphorylate PIN1 at different phosphosites with different preference [52]. d6pk0123 quadruple mutants showed somewhat pin-like axillary shoots [51]. pid wag1 wag2 mutants phenocopied ncp1 pid [30]. Because NCP1/AtMOB1A is functionally conserved MOB1 in Arabidopsis, it is possible that PID/WAGs/D6PKs function as a plant counterpart of LATS1/2. It would be interesting to test if PID/WAGs/D6PKs can rescue Drosophila lats mutant phenotypes. AtMOB1A may associate with PID/WAGs/D6PKs and regulates its kinase activity, which subsequently modifies activities of PIN1. In human and Drosophila, MOB1 can activate LATS/NDRs when targeted to the plasma membrane [39,53]. AtMOB1A is localized to nucleus [16,37] and also associated with plasma membrane (Fig 4). PID, WAGs and D6PKs are also associated with plasma membrane [54,55], making it possible for AtMOB1A to activate PID/WAGs/D6PKs. However, we did not detect direct physical interactions between AtMOB1A and PID/WAGs by using pull-down and Co-IP assays. The negative results do not rule out the possibility that AtMOB1A is in a complex with AGC kinases. On the other hand, ncp1 pid wag1 wag2 pid2 showed no-cotyledon phenotypes similar to those of pid wag1 wag2 pid2. But the quintuple mutants displayed enhanced developmental defects in true leaves. These findings support the hypothesis that AtMOB1A may function with PID/WAGs. Alternatively, AtMOB1A and PID/WAGs/D6PKs may regulate transcription levels of auxin related genes. Indeed, we observed the alteration of expression pattern of PIN1-GFP and down-regulation of ARF7:GUS and ARF19:GUS in ncp1 pid double mutants (S10 Fig and S11 Fig). This finding is consistent with the mechanism that the animal Hippo pathway functions through regulating expression of downstream genes via a common growth regulatory effector, the transcriptional co-activator YAP/TAZ [1]. Another possibility is that the Hippo pathway functions in parallel to auxin pathway, yet they crosstalk to control plant development. This would be similar to the crosstalk between Wnt/β-catenin pathway and Hippo pathway to regulate animal development and tumorigenesis. It has been shown that cytoplasmic TAZ of the Hippo pathway can bind to DVL of the Wnt/β-catenin pathway and negatively regulate the Wnt/β-catenin pathway [56].
In conclusion, we demonstrate that AtMOB1A, a key component of the Hippo pathway, plays critical roles in auxin-mediated development in Arabidopsis. AtMOB1A synergistically interacts with auxin biosynthesis, transport, and signaling pathways to regulate Arabidopsis development. MOB1 is a regulator of AGC kinases in animal systems. PID/WAGs, D6PKs are AGC kinases, suggesting that NCP1/AtMOB1A may also regulate kinase activities of PID/WAGs and D6PKs, and possibly other AGC kinases in Arabidopsis. The fact that auxin responses and expression of auxin related genes such as ARF7 and ARF19 were down-regulated in ncp1 pid mutants suggests that NCP1/AtMOB1A may promote auxin signaling. This provides another layer of regulation of plant development by auxin. Further identification of other components of the Hippo pathway in Arabidopsis will help elucidate the mechanisms.
Plants were grown under 16-h light/8-h dark cycle at 22℃. The T-DNA insertion lines were obtained from NASC. The mutants used in this work were: pid (SALK_049736), pid-714 (SAIL_770_E05), ncp1-2 (GK_719G04). T-DNA insertion sites were determined by sequencing. Genotyping primers for pid (SALK_049736) and pid-714 (SAIL_770_E05) are: 5’-CCTCAGATTTCGCTTACGCAG-3’, and 5’- GCGAGACGAGTGAATCGTCG-3’, combined with JMLB1 and SAIL-LB1, respectively. For genotyping ncp1-2 (GK_719G04), 5’-ATGGATTCGTGTGGCTTTC-3’, 5’-TGTTTACAGCAAGCCATTC-3’, and PGABI1: 5’-ATATTGACCATCATACTCATTGC-3’ were used. To genotype ncp1-1, 5’-TGACCGTCTTCTTCCTAT-3’ and 5’-TGTTTACAGCAAGCCATTC-3’ were used and the PCR products were digested with MseI. npy1-2, yuc1, yuc4, tir1-1, afb2-1, afb3-1 were previously described [22,27,29]. All T-DNA insertion lines were genotyped as described previously [57–59].
For complementation of ncp1 pid mutants, a genomic DNA fragment containing the coding region as well as up- and down-stream regulatory sequences of At5g45550 was amplified by PCR using the following primers: 5’-CCCCCCGGGGAAACGGTGACCAAAATGCT-3’ and 5’-GCTCTAGAAGACGAGGCTCCAACACG-3’. The PCR product was digested with BamHI and XbaI and subcloned into pPZP211 vector [60] to generate pPZP211-NCP1gDNA. The plasmid was transformed into ncp1 pid+/- mutants via Agrobacterium strain GV3101 using floral dipping method [61]. The transgenic seedlings were selected on 1/2 MS plates containing 50 μg/mL kanamycin.
For expression of the Drosophila Mats under the control of NCP1 promoter, the Mats cDNA was amplified with PCR using the primers: 5’-ACTCCCGGGATGGACTTCTTGTTCGGTTC-3’, and 5’- GCTCTAGACTATATCTGCCGCTCATCCT-3’. The NCP1 promoter was amplified with primers: 5’-ACTGTCGACCTGCCCAATCAGCAAGAA-3’ and 5’-ACTCCCGGGGGCGACAAAAAGCAAGCGAG-3’. The PCR products were digested with SalI, XmaI and XbaI and subcloned into pCambia-1300 to generate pCambia-1300-NCP1p:Mats.
For expression pattern and subcellular analysis of NCP1, the pPZP211-NCP1gDNA construct was modified. The GFP gene was inserted immediate before the stop codon of NCP1 gene with restriction site of ApaI.
SEM samples were prepared as described previously [62], and analyzed using a HITACHI S-4800 FESEM microscope. For whole-mount analysis of vascular structures and embryos, samples were prepared as previously described [63], and photographed under differential interference contrast (DIC) field or dark field on Leica DM 4500 and Leica S8AP0 microscopes. DR5-GFP and PIN1-GFP signals in embryos were viewed on Olympus FV1000MPE following the manufacturer’s instructions.
Sequences were aligned using Clustal X version 1.81 [64], then refined manually. Maximum Likelihood method was used to reconstruct the phylogenetic tree using Mega5 [65]. Topological robustness of the phylogenetic tree was assessed by bootstrapping with 1000 replicates [66].
For the pull-down assay, cDNA of PID and NCP1 was cloned into pGEX-4T-1 and pET30a vectors to generate the expression constructs. The His-tagged and GST-tagged proteins were expressed in E. coli strain BL21. The subsequent protein purification and pull-down assay with Glutathione Sepharose 4B (GE) or His beads (Bio-Rad Ni-NTA Agarose) were carried out following the manufacturers’ manuals. The bound proteins were eluted and analyzed with anti-GST and anti-HIS antibodies (CWBIO).
To perform Co-IP assay of NCP1 and PID/WAGs, we constructed pEarleyGate104-35S:YFP-NCP1, pSuper1300:PID-Myc, pSuper1300:WAG1-Myc, pSuper1300:WAG2-Myc. YFP-NCP1 and PID-Myc or WAGs-Myc constructs were transformed into tobacco (Nicotiana Benthamiana) by injection. Leaves were grounded into fine powder in liquid nitrogen. Proteins were extracted with the extraction buffer [100 mM HEPES (pH 7.5), 5 mM EDTA, 5 mM EGTA, 10 mM NaF, 5% Glycerol, 10 mM Na3VO4, 10 mM DTT, 1 mM PMSF, 0.1% Triton X-100, 10 μg/mL Aprotinin, 10 μg/mL Leupeptin, 10 μg/mL Antipain]. The protein extracts were spun twice for 30 min at 14,000 g at 4℃. The supernatant was incubated for 3 hr with anti-Myc-tag mAb-agarose (MBL) in IP buffer [20 mM Tris-HCl (pH 7.5), 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1 mM Na3VO4, 1 mM NaF, 10 mM glycerophosphate, 0.1% Triton X-100, 1 μg/mL Aprotinin, 1 μg/mL Leupeptin, 1 μg/mL Antipain]. The agarose was washed for three times with 1 ml of PBS. Proteins were then released and collected by boiling in 2×SDS loading buffer for 5 min. IP products were detected by SDS-PAGE and immunoblot analysis using anti-Myc or anti-GFP antibodies (CWBIO). These experiments were repeated at least three times.
|
10.1371/journal.pgen.1004176 | Noise Genetics: Inferring Protein Function by Correlating Phenotype with Protein Levels and Localization in Individual Human Cells | To understand gene function, genetic analysis uses large perturbations such as gene deletion, knockdown or over-expression. Large perturbations have drawbacks: they move the cell far from its normal working point, and can thus be masked by off-target effects or compensation by other genes. Here, we offer a complementary approach, called noise genetics. We use natural cell-cell variations in protein level and localization, and correlate them to the natural variations of the phenotype of the same cells. Observing these variations is made possible by recent advances in dynamic proteomics that allow measuring proteins over time in individual living cells. Using motility of human cancer cells as a model system, and time-lapse microscopy on 566 fluorescently tagged proteins, we found 74 candidate motility genes whose level or localization strongly correlate with motility in individual cells. We recovered 30 known motility genes, and validated several novel ones by mild knockdown experiments. Noise genetics can complement standard genetics for a variety of phenotypes.
| Inferring the function of proteins and the role they play in cellular processes is essential for our understanding of cell biology, genetics and biology in general. Standard genetic approaches use large perturbations to cells such as gene knockout, knockdown or over expression of genes. Such methods are powerful, but have the drawback of taking the cell far from its normal working point. Here, we provide a new and much milder approach, which uses the natural cell-cell variation in protein level and expression pattern as a source of mild perturbation. We monitor individual live cancer cells under the microscope and correlate their protein levels and localization with phenotype in the same cells. We use the motility of human cancer cells as a model system that is highly important for understanding metastasis in cancer. We find that our approach uncovers most of the known motility proteins, as well as new ones which we validate using knockdown experiments. Our novel approach is widely applicable to any phenotype that can be visualized in individual cells, and for any organism for which one can measure proteins in individual cells.
| To understand which proteins contribute to a biological phenomenon, current approaches use perturbations such as gene knockdown, over-expression or knockout. These approaches have provided the basis for much of what we know about cell biology. However, such perturbations also have drawbacks. Perturbations currently used are usually large - a protein expression is either markedly reduced or increased, and the measurement is therefore far from the cells normal working condition. This can lead to artificial off-target effects or to masking of the perturbation by changes in the cell that compensate for the loss of a protein. It is thus possible that some of the information about protein function has remained hidden due to these features of current methods.
To offer a complementary way to understand protein function, we present an approach called noise genetics. Noise genetics uses the natural cell-cell variation in protein levels and localization [1]–[8] as a source of mild perturbations to reveal protein function. Since natural fluctuations are mild, the risk of compensation is reduced. The idea is to correlate the protein levels and localization in individual cells to the phenotype in the same cells. Notably, cell-cell variation in protein level changes slowly over time: cells keep their individual levels for about a cell generation [7]. Thus, the noise we use is a type of cell individuality (Figure S1). Cells have individual character in many of their phenotypes as well, that also last for about a cell generation [9], [10].
Previous studies used noise for understanding regulatory interactions between a few proteins in bacteria [1], [11], whereas here we screen hundreds of proteins. As a model system, we use the motility phenotype of human cancer cells. Motility of cancer cells is of general interest both as a well-studied biological phenotype [12], and as a feature of normal physiology and cancer metastasis [13]. Wide scale genetic screens, including siRNA knockdowns, have revealed numerous genes involved in motility [14]–[16]. Moreover, natural phenotypic variability and fluorescent microscopy were used to study the shape of motile cells [10] and the cytoskeleton dynamics [17].We use a library of human cancer cell clones each with a different protein fluorescently tagged at its endogenous chromosomal locus [18], [19] to follow the natural variability of proteins, and the natural variability of motility in the same cell. Proteins whose level or localization correlate with motility are identified as candidate motility proteins.
We find that about 15% of the 566 highly expressed proteins that we tested exhibit a significantly high correlation between their protein features and motility in individual cells. This correlation can suggest that the protein has a role in cell motility. About half of these candidate proteins were previously known to be involved in cell motility. We validated a sample of these candidates using mild siRNA knockdown.
To study natural variability between individual cells, we used the LARC library of human clones with tagged proteins [18]–[20]. The library is made of clones of a parental human lung cancer cell line, H1299. In each clone, a full-length protein is fluorescently tagged with YFP as an internal exon (Figure 1A). The protein is tagged at its endogenous chromosomal locus, preserving the natural promoter and regulatory sequences (Figure 1B). Previous studies suggest that most (70–80%) of the tagged proteins preserve their wild-type dynamics and localization [18], [20].
The parental clone also expresses proteins tagged with red florescence using mCherry. This red fluorescence is used for image analysis of time lapse movies (Figure 1C), allowing automated segmentation and tracking of the nucleus and cytoplasm in all clones (Figure 1D). The tagged proteins are XRCC5 and DAP1, both not known to be involved in motility.
A previous study employed this library to follow 1260 clones with different tagged proteins as they responded to an anti-cancer drug using time-lapse movies [18]. Here, we re-analyzed these movies, that also included the 24 h period before drug addition, together with movies from a recent study on protein half-lives using the same library and microscopy system (Eden et al, 2011) and chose 704 unique proteins with high quality movies for further analysis (movies chosen had 4 fields of view totaling at least 20 cells at each time-point). Of these, we chose only known proteins (as opposed to ESTs) with subcellular localization matching the literature. This results in a final set of 566 different proteins. These proteins have diverse cellular localizations and functions (Supplementary File S1).
We tracked the protein level and localization in each cell, and also the motility of the same cells. Protein level is given by the summed YFP fluorescence of all pixels in the cell. Most proteins in our dataset did not show large translocation events between cell compartments such as nucleus and cytoplasm. To parameterize protein localization, we therefore characterized the spatial distribution inside the cell, using two well-known measures from image analysis: contrast (the existence of sharp changes in intensity) and texture (also called texture correlation, the linear dependency of grey levels on those of neighboring pixels) [21]–[23].
Cell motility was quantified in terms of speed and angle change of the motion (Figure 2A). Cell velocity was measured as the change in cell center of mass between frames (20 min). The angle change of cell movement in frame i was calculated based on the deviation of the cell in frame i+1 from its movement between frame i-1 and frame i. Persistent motion results in low values of angle change (Figure 2B). We found that angle change was negatively correlated with cell velocity (Figure S2).
If a protein is involved in motility, we expect a significant positive or negative correlation between at least one of the protein parameters and one of the motility parameters. For example, if cells that express protein X at high levels move faster, whereas cells that express it at lower levels move slower, we predict that protein X is a candidate motility gene (Figure 1E). A similar conclusion is reached if contrast or texture correlate with motility. For example, if cells that show a homogenous spatial distribution of protein Y move slower than cells that express protein Y in more punctuate manner (high contrast of fluorescence across the cell), one may predict that Y is a candidate motility gene. The contrast and texture differences we observe between cells are subtle, not gross changes such as transitions between organelles. Images of cells are provided in S3–S5.
We compared the three protein properties (protein level, contrast and texture) to the two motility parameters (speed, angle change) in each cell and each time-point using Spearman and Pearson correlations (Figure 2C, Figure S6). The observed correlation values are centered at around zero and range from −0.4 to 0.4.
In order to test the significance of the calculated correlation values, we compared them to correlations in randomized data. To make a stringent comparison, we note that data from different time-points of the same cell are not independent. Furthermore, cells from the same field of view are potentially more dependent than cells in other fields of view, due to possible systematic effects in the experiment. We thus constructed the randomized dataset by associating the motility parameters for cell i at all time points with the protein properties from a different cell j at the same time points in the same field of view, with i and j randomly chosen. This provides a randomized control of the same size as the original data.
The permuted datasets showed correlations between protein and motility parameters mainly (90%) in the range −0.1 to 0.1 (gray bars in Figure 2D). Comparing the randomized distribution to the measured data (black bars in Figure 2D) shows proteins with correlations higher or lower than expected by chance. In this study, we define a candidate motility gene if its absolute correlation coefficient |R| exceeds 0.15 in both Pearson and Spearman correlations. Based on the comparison to randomized data, the rate of false positives is expected to be 15% for comparisons of protein level to cell velocity (Figure S7), 23% for contrast versus velocity, 30% for texture versus velocity. To test this, we conducted a new set of time-lapse microscopy experiments on a random sample of 19 candidate motility proteins, and found that 16/19 showed the same above-threshold correlations as in the original movie dataset, consistent with a false-detection rate of about 15% (see Supplementary File S3).
Similar false discovery rates were obtained for the comparison between angle change and the protein properties (Figure S7). Several examples of proteins with positive and negative high correlations are shown (Figure 2E) along with examples of proteins that did not show a significant correlation between the protein level and the motility in individual cells (Figure 2F).
We found 74 candidate motility genes (Supplementary File S1, S2). Of these, 31 (41%) correlate by protein level. The rest of the proteins correlate with motility by contrast (25%) or texture (23%), and 15% of the candidate proteins correlate by more than one measure.
The candidate genes are highly enriched in genes previously known to play a role in motility: 41% (30/74) were previously characterized as motility genes (hypergeometric p = 0.0009) according to the Genecards database (see Methods). Among the candidate genes are actin regulators in the ARP complex (ARPC3 [24]), two actin related proteins (ACTR2 and ACTR1A), RAC1 [25] that is essential for cell migration and WASF2 [26] that is part of the WAVE complex that regulates lamellipodia formation [12]. (Other examples are described in Figure 3A). Some of the candidate genes have no known role in motility (Figure 3B).
The subcellular localization of the candidate genes is enriched in the cytoskeleton (hypergeometric p-value = 0.0008), the plasma membrane (p = 0.02) as well as the ER/golgi (p = 0.06) (Figure 3D).
In an attempt to estimate the false-negative rate of this assay, we considered the 13 genes out of the 566 in this study that were listed as motility genes in a recent review of motility (Ridley, 2011). Among these 13 relatively well-characterized genes, 7 genes were not identified as candidate motility genes in this study (Figure 3C). This suggests a false-negative rate on the order of 50%. However, some of these false negative genes showed a relatively high correlation in one of the examined comparisons and would be scored positive in a less stringent threshold choice. A more extensive false-negative test compared the present assay to all 110 genes in our set of 566 that are listed as motility related in the Genecards database. Of these, only 29 are found in our assay. Our analysis is not expected to find all motility genes due to several limitations that will be addressed below.
In order to test the involvement of the candidate genes in the motility process, we used siRNA directed against the YFP tag to lower the expression level of the tagged proteins. Since we introduced YFP as an exon to all clones, the anti-YFP siRNA can be used to knockdown expression in any clone from the library. The knockdown is mild (at most half-knockdown) because only one allele is tagged with YFP. Thus, a 50% reduction in YFP corresponds to a 25% reduction in total protein due to the expression from the untagged allele. We chose 11 candidate genes that had positive correlation between protein level and motility or a significant correlation between a protein feature and motility. Of these candidate genes, 4 are previously known motility genes and 7 are novel, randomly chosen from the candidate list. We also tested 4 control genes that did not correlate with motility. We took time-lapse movies for 48 hours starting 24 hours after the siRNA infection (Figure 4A).
We used the YFP fluorescence to observe the extent of knockdown (Figure 4B). The mCherry labeling of the cells was not influenced by the siRNA (Figure S8). 10 out of the 11 candidate motility gene clones showed significantly lowered velocity (reduction of 15–35%), whereas 3 of the control genes (GAPDH and TOP1 and one gene of unknown function) showed no measurable reduction in velocity (Figure 4C).
One control gene – the ribosomal gene RPS3, not known to affect motility, showed a motility defect upon mild knockdown. It was not picked up as a candidate motility gene in the noise genetics assay based on cell-cell variations. This may point to a difference between knockdown that affects only a single gene in a module such a ribosome, and noise genetics, where fluctuations in all genes in a module (e.g. all ribosomal genes) are expected to be correlated [7], [8].
In total, candidate motility genes, both previously known and novel were validated by mild knockdown at a level of about 90%. We further tested proteins with a negative correlation between their protein level and motility and the results are summarized in the supplementary information (see Supplementary File S4).
In this study, we presented ‘noise genetics’ - an approach to assign function to proteins that uses the natural noise in protein level and localization and correlates it to the variation in phenotype of the same individual cells. We demonstrated this using the motility phenotype of cancer cells. Noise genetics recovers 30 of the known motility genes in our clone library and also 43 novel motility candidates, of which 10 were validated by siRNA knockdown (Figure 4).
Noise genetics can complement standard genetic perturbations. Among its advantages are the non-invasive and mild nature of the natural variations used, which keep the cell near its normal working point. Proteins whose knockdown is lethal are hard to evaluate using standard genetics, but can potentially be picked up by the assay. Similarly, proteins whose knockdown effects are masked by compensation from other proteins in the same module may be picked up by noise genetics, because one expects the entire module to show correlated noise [7], [8]. Noise genetics compares individual cells from the same field of view, and therefore contains a type of internal control for systematic errors. In standard perturbation assays, one needs to compare perturbed cell populations to a separate experiment with unperturbed cell populations in order to control for experimental systematic errors.
Among the limitations of noise genetics as implemented here are the need for a fluorescent cell library or other means of observing cell-cell variation in both protein and phenotype. Such libraries exist, for example, in S. cerevisae [27], E. coli [28], [29], C. elegans [30] and Zebrafish [31]. If the phenotype of interest can be observed in fixed cells, individual cell imaging of proteins [32] or mRNA [33]–[35] might be used for noise genetics. Such an approach has been used with pre-selected genes, for example, to explore the effect of protein variability on stem-cell differentiation [36], sporulation timing in Bacillus subtilis [37] and meiosis timing in Saccharomyces cerevisiae [38]. We currently tested only linear correlation; more elaborate time-series analysis methods or non-linear correlation analysis may be able to improve the resolution of this approach [39].
The cells in this study are diploid (or multiploid), but only one copy of each gene was labeled with YFP. Therefore, the fluorescence measurement does not necessarily reflect the total protein level or distribution. Previous work with the present cell system showed that there is high correlation in the expression of two alleles of the same ribosomal gene [7]. Future work is needed to test the present approach with all alleles tagged.
Noise genetics can miss proteins whose effect on phenotype is small in the working point of the cell. Such proteins can be picked up by standard genetics which makes large perturbations. Similarly, if natural variations in proteins or phenotypes are very small, noise genetics may not be applicable. Importantly, noise genetics on its own can only detect correlations, and additional experiments such as the mild knockdown performed here, are needed to gain evidence for causality.
Noise genetics uses natural cell-cell variation in proteins to discover links with phenotypes. Additional phenotypes that can be readily studied include cell size and shape, and any other phenotype measurable by time-lapse microscopy.
Movies from previous studies [18], [20] were used for this analysis, as well as new movies on 19 clones. In the previous studies, 4 movies (fields of view) were taken for each of the 1,000 clones totaling about 4,000 movies. In each movie, 10–20 cells were tracked over 24 hours at least, every 20 minutes. Some of these clones were filmed more than once and some clones represent the same protein. For the present analysis, we combined all data for the same protein from all relevant movies. Each time point included transmitted light image (phase contrast) and two fluorescent channels (red and yellow). Of the original movie sets, we chose 566 unique protein clones as described in the text. The 566 known proteins that were used in our analysis tend to have high expression levels (so that they are picked up in the LARC library construction which used FACS to select for fluorescent clones).
We used the image analysis software described in [18] with minor modifications. The main steps in this software include background correction (flat field and background subtraction), segmentation, cell tracking, and automated identification of cell phenotypes (mitosis and cell death). Cell and nuclei segmentation was based on the red fluorescent images of the two red tagged proteins found in all clones, localized to the cytoplasm (DAP1) and nucleus (XRCC5), with intensity which is very uniform across cells and clones. Segmentation used global image threshold and seeded watershed segmentation. The cell-tracking procedure maps each cell to the appropriate cells in the preceding and following frames as described [40]. Texture parameters (contrast and correlation) of the proteins were measured for each cell in each time point based on the YFP image of the tagged protein as described below.
In our previous studies with the same movies, we also analyzed the protein concentration by taking the average or median fluorescent intensity inside the cell, as opposed to the total intensity. We find this measure to be more sensitive to image outlier pixels - even when using the median pixel intensity [23]; we therefore use total fluorescent intensity in the present study.
To calculate texture and contrast, we first evaluated a gray-level (fluorescence intensity) co-occurrence matrix (GLCM) from each fluorescent image of the cells [41]. Each element (i, j) in GCLM specifies the number of times that the pixel with gray-level i occurred horizontally adjacent to a pixel with gray-level j.
From the matrix one can compute the various texture features. For example, contrast is giving a value of 0 for a constant intensity image and high values when adjacent pixels have different intensity. Our second measure, ‘texture’ or ‘correlation texture’ is . The correlation texture measures the linear dependency of grey levels on those of neighboring pixels. The two measures are weakly anti-correlated (Fig. 2C). We verified that the texture and contrast values that were calculated after removing the background and after rescaling to 64 gray levels have a low sensitivity to rotations (Figure S5).
For each protein, we collected all cells in all fields of view at all time frames. For each cell, 3 protein parameters (total protein, contrast and texture) and 2 motion parameters (velocity and angle change) (See Figure 2A) were calculated. Then, 6 correlation values (between all pairs) were calculated using Pearson and Spearman correlation. No binning was used to compute these correlations. The correlation values are summarized in Supplementary File S1. We tested other texture features as well (homogeneity and energy), but these did do not add known genes to the list of candidates (data not shown). Next, we collected only cells that were tracked from the beginning of the movie until its end and repeated the same calculation.
In order to establish a threshold that minimizes the false positive rate, we generated permuted datasets. In each permuted dataset, for each of the protein parameter and for each of the motion parameter, the protein values from one cell time trace were correlated to the motion values in another cell time trace in the same field of view. We repeated this permutation 10 times for each of the 6 comparisons (and both for Pearson and Spearman calculations) and generated 6 correlation distributions, for all proteins. These correlation values are summarized in the Supplementary File S3. We chose to use both Spearman and Pearson correlations since using only Spearman correlation results in marginally higher false positives (Data not shown). A correlation value of R = 0.15 (or similar) was chosen in order to minimized false positives, while maximizing potential true candidates (See Figure S7). We further tested for all the proteins with significant correlation values in which the correlation still holds even when calculating it only in one field of view (FOV) for most of its FOVs. The final candidate list includes only proteins that passed this test.
The shape of the cell is known to affect its motility [10]; in order to estimate this effect in our dataset, we calculated the correlations between the aspect ratio (major axis/minor axis) of each cell and the cell velocity and found no significant correlation in the examined clones (Figure S9).
We conducted new experiments by performing time-lapse movies for 19 of the candidate clones. Of the 19, 13 proteins were chosen randomly out of the list of novel candidate proteins, and the other were candidates that were also known motility genes. 16 out of the 19 showed a correlation that is similar to the previous calculated correlation. (Results are shown in Supplementary File S3).
Several databases were used to annotate whether genes are known to be involved in motility (Supplementary File S2). We used GeneCards (http://www.genecards.org/) to download all genes with the keywords “motility” or “migration”. We further considered genes that were identified as part of the “adhesome” (http://www.adhesome.org/). Finally, well-studied characterized genes listed in the Review by [12] were also considered as motility genes.
We used subcellular localization according to GeneCards (http://www.genecards.org/) and other databases as provided in the LARC database (http://www.weizmann.ac.il/mcb/UriAlon/DynamProt/). All clones used here have localization in the experiment that agrees with the previously known localization. When more than one localization was assigned to a protein, the first localization in the list was used for the category assignment. The subcellular information and categories are list in the “Candidate_genes.xls” file. Hypergeometric p-value [42] was used to calculate the enrichment of specific subcellular localization categories in the candidate genes group over all the 566 genes used in this analysis.
To knockdown the expression of the tagged protein in clones from our library, we used siRNA against GFP (QIAGEN, 1022064) transfection using lipofectamin (Invitrogen) as described in their protocol. As a control siRNA, we used the non-targeting siRNA (Dharmacon, D-001810-10-05). No significant difference between siRNA used from QIAGEN or Dharmacon was detected in our system (Figure S10).
Briefly, 2×104 cells were grown on 12-well glass bottom MatTek plates. The next day, siRNA transfection was performed. We used 24 pmol of si-RNA for each well and 0.8 ul lipofectamin and incubated it for 6 hours, then we replaced with fresh media and let cells grow overnight. The next day, we took a time-lapse movie of the plate for 48 hours. We took fields of view from each well. We used the same exposure time for the well with the non-targeting siRNA (where no decrease in expression is expected) and the well with the si-GFP (which showed a decrease in the YFP fluorescent and not in the mCherry fluorescent).
|
10.1371/journal.pntd.0005747 | Analysing published global Ebola Virus Disease research using social network analysis | The 2014/2015 West African Ebola Virus Disease (EVD) outbreak attracted global attention. Numerous opinions claimed that the global response was impaired, in part because, the EVD research was neglected, although quantitative or qualitative studies did not exist. Our objective was to analyse how the EVD research landscape evolved by exploring the existing research network and its communities before and during the outbreak in West Africa.
Social network analysis (SNA) was used to analyse collaborations between institutions named by co-authors as affiliations in publications on EVD. Bibliometric data of publications on EVD between 1976 and 2015 was collected from Thomson Reuters’ Web of Science Core Collection (WoS). Freely available software was used for network analysis at a global-level and for 10-year periods. The networks are presented as undirected-weighted graphs. Rankings by degree and betweenness were calculated to identify central and powerful network positions; modularity function was used to identify research communities. Overall 4,587 publications were identified, of which 2,528 were original research articles. Those yielded 1,644 authors’ affiliated institutions and 9,907 connections for co-authorship network construction. The majority of institutions were from the USA, Canada and Europe. Collaborations with research partners on the African continent did exist, but less frequently. Around six highly connected organisations in the network were identified with powerful and broker positions. Network characteristics varied widely among the 10-year periods and evolved from 30 to 1,489 institutions and 60 to 9,176 connections respectively. Most influential actors are from public or governmental institutions whereas private sector actors, in particular the pharmaceutical industry, are largely absent.
Research output on EVD has increased over time and surged during the 2014/2015 outbreak. The overall EVD research network is organised around a few key actors, signalling a concentration of expertise but leaving room for increased cooperation with other institutions especially from affected countries. Finding innovative ways to maintain support for these pivotal actors while steering the global EVD research network towards an agenda driven by agreed, prioritized needs and finding ways to better integrate currently peripheral and newer expertise may accelerate the translation of research into the development of necessary live saving products for EVD ahead of the next outbreak.
| Ebola Virus Disease (EVD) research publications were used to analyse and visualise collaborations between institutions jointly publishing research results, using freely available social network analysis tools. Constructed co-authorship networks between author affiliated institutions showed EVD research publications increased and networks evolved over time. The global network is organised around a few co-authoring, mostly publicly financed key actors, highly connected with powerful and broker positions. The results present an extensive narrative how modern empirical scientific methods for data processing and translation can supplement evidence-based arguments for public discussion on the status and focus of global EVD research. Based on the network characteristics or concentration of expertise, we recommend a globally agreed and prioritized EVD research agenda may facilitate the translation of this research into new EVD tools. Also, to analyse research networks regularly to enable public discussion on the direction in which research could be organized and optimised. We would like to encourage others to utilize our methods with open access tools to enhance new methods to the field of NTD R&D.
| The 2014/2015 West African Ebola Virus Disease (EVD) outbreak with more than 28,000 cases and 11,000 deaths, was a public health emergency of international concern [1,2]. Although EVD was discovered in the former Zaire (now: Democratic Republic of Congo) more than 40 years ago, the absence of treatment generated global alarm and raised questions on the state of EVD research. Studies analysing EVD transmission and clinical trials testing EVD treatments or vaccines have been difficult due to the small number of infected cases in previous outbreaks [3,4]. Moreover, the pharmaceutical industry has been criticized for neglecting EVD research because it is not profitable enough as EVD occurred rarely and mostly in impoverished African communities [3,5–7]. EVD outbreaks have attracted general public attention since the mid-90s, benefitting science funding, leading to increased publications, but EVD research funding is mostly spent outside of affected African countries and research capacity building there was neglected [8].
The World Health Organization (WHO) called for greater transparency and better sharing of results from clinical trials as being a necessary contribution to facilitate research and development (R&D) for the benefit of science and patients [9] and published a research priority agenda [10]. The necessity for increased transparency also applies to any existing EVD research and expertise to improve the value and efficiency of research efforts.
In order to enhance the understanding of on-going EVD research activities and its communities, social network analysis (SNA) of bibliometric data of EVD related scientific publications can be used. Since co-authorships are the most visible and accessible indicator for collaborations, co-authorship-based SNA studies can be used to measure the presence of research collaborations and their evolution over time [11–13]. SNA metrics can reveal network patterns and identify its most central and influential actors [14–16].
The volume of publications, in combination with results from a co-authorship network analysis, can serve as a proxy indicator for R&D. Besides mapping the research landscape [17], especially co-authorship network analysis can provide insight into the degree of research governance and be relevant for strategic research planning [18,19]. Moreover, information from collaboration networks can be used to identify potential collaborations in order to improve research communication and therefore maybe also influence research outcomes [12,20].
The aim of this study is to identify EVD research activities and to analyse the structure of the evolving EVD research community network over time to map existing research collaborations and influential actors based on centrality network metrics.
Based on bibliometric data we analysed the development of EVD research in two steps. Firstly, we measured the annual EVD research publications amongst all published materials. Secondly, we conducted a co-authorship network analysis at institutional level based on original research publications between 1976 and 2015. Additionally, network analyses were conducted for 10-years’ time periods in order to assess temporal network dynamics.
Bibliometrics of 2,528 articles resulting for our WoS search were exported as tab-delimited data and imported into MS Excel as one bibliometric data set (Fig 2). In the raw data set each entry referred to one publication. We included data on title, authors, address of authors’ affiliated institution, publication year, source, language, document type, cited references, funding agency, publisher and subject category in further analysis. Other columns were deleted from the data set. Information on addresses of author’s affiliated institution, e.g. institution name, sub-departments and institution address including city and country, were split into separate columns.
Data processing and further cleaning was performed using the software AppleScript [22] and OpenRefine [23]. Name disambiguation, e.g. Centers for Disease Control and Prevention was abbreviated as CDC, Ctr Dis Contr and Centers Dis Cont, orders within names, e.g. Univ Washington and Washington Univ or name spellings, e.g., Univ Georgia, UNIV GEORGIA were identified and harmonised using OpenRefine algorithms or manually. Missing data, e.g. missing country information of an affiliated institution, were substituted by manual web search.
If an institution name appeared with addresses in different locations in the data set, e.g. WHO with location Switzerland and location Copenhagen e.g. due to different regional offices, different locations were considered for construction of the network to account for institutions international representations. Institutions duplicates originating from publications with multiple co-authors affiliated with the same institutions were eliminated to ensure a single weighting of institutions.
The free online application Table2net was used to extract network information from the refined data set to construct a Gephi readable file [24]. Network nodes (i.e. actors) are institutions named as authors’ affiliations in original research publications. Network edges are titles of joint publications from authors’ affiliated institutions.
The free software Gephi was used to calculate network metrics and visualise the networks [25].
Network analysis provides various tools and metrics in order to assess different notions of importance of individual nodes and node groups. As the simplest metric of centrality we calculated each node's degree, as the sum of direct links to other nodes. Nodes with more direct connections are considered more central. The average node degree captures the number of actors that each actor is connected with on average. The average weighted node degree also takes the weight of a connection between a pair of nodes into account [26,27].
Betweenness centrality measures the frequency with which a particular node lays on the shortest paths between all other node pairs. Therefore, nodes with a high betweenness are considered to have a broker position as they connect many other nodes and thus have a large influence on the transfer of items through the network, under the assumption that item transfer follows the shortest paths [26,28]. We used a betweenness calculation algorithm for weighted graphs as developed by Opsahl [29].
Besides positional properties of the nodes within the network, metrics are capturing topological aspects of the network as a whole. This information can provide an insight on the evolution at network level. Density measures were calculated to assess the connectivity of the network. The density of a network is defined as the total number of existing edges divided by the total number of possible connections. If edges exist between all nodes (density = 1) a network is considered completely dense [26,28]. Since density captures the probable feasible number of connections in a network, it is an indicator for possible community building [30] or innovation flow within a network [15].
Communities within the network were detected using Gephi’s modularity algorithm. Modularity measures the degree of separation of a network into modules or clusters (communities). While a modularity value of 1 indicates that the actors separate perfectly into self-contained clusters, a value of -.5 suggest the opposite, a homogeneously connected network [27,31]. Networks with a high modularity score employ dense connections between nodes within the modules but sparse connections between nodes from different modules.
For visual presentation of network metric calculations we used Gephi's Force Atlas II algorithm in log-linear mode optimized towards hub dissuasion [32].
Systematic search in WoS for publications containing “Ebola*” yielded a total of 4,587 publications between 1976 and 2015, including original articles (2,531), editorial material (659), news items (437), reviews (415), letters (325), meeting abstracts (157), corrections (36), notes (14), reprints (7), biographical items (4) and book reviews (2). Amongst the 2,531 original articles were 75 article proceedings and five article book chapters. Three of those publications appeared with anonymous authors and were therefore deleted for social network analysis (Figs 1 & 2).
The first EVD research article was published in 1977, shortly after the first noted EVD outbreak in 1976. Only few EVD publications were visible until the early nineties, whereas from 1994 onwards the number of yearly EVD publications increased continuously (Fig 3). Since 1994 a higher frequency of EVD outbreaks were recorded and more EVD cases were being detected in almost every year. Several localised EVD outbreaks in Africa have occurred with up to several hundred cases. The initial EVD outbreak in 1976, with a relatively high number of reported cases (>600), was followed by only a small number of publications on EVD research. No EVD outbreaks were reported between 1979 and 1994 and hardly any publications were published on the topic. The number of publications increased gradually and continuously after the second outbreak in 1994, although compared to the 1976 outbreak only about one-tenth of cases were reported (Fig 4).
A substantial increase in EVD research publications occurred during the 2014/2015 West African outbreak. An almost 10-fold increase from 2013 (171), 2014 (772) to 2015 (1,621) was visible for almost all document types, but it was most pronounced for editorials (5, 220, 343), letters (1, 75, 213), news items (4, 190, 118) and meeting abstracts (9, 5, 66) respectively. An increase in reprints, notes, biographical items and book reviews was not detected.
Bibliometrics of 2,528 original research articles were used for social network analysis. Based on their co-authors’ affiliated institutions a global network including institutions from 101 different countries with 704 connections was constructed (Figs 5 & 6).
Research institutions in the United States (US) are among the most highly connected institutions in EVD research (degree (d) = 80). They are mostly connected to institutions in Canada (d = 40) with an edge weight (ew) of 130 and Europe, especially Germany (d = 53, ew = 110), the United Kingdom (UK) (d = 60, ew = 90) and France (d = 57, ew = 51), but also to Japan (d = 32, ew = 99). Connections between US institutions and institutions in EVD affected African countries are less frequent (e.g. Guinea-USA ew = 14, Sierra Leone-USA ew = 32, Liberia-USA ew = 30). However, institutions in Sierra Leone and Guinea (both d = 32) and other African countries, especially Nigeria, Uganda and Ghana, are embedded in the global research network with connections to UK, Germany, France and Switzerland. The overall density of the global country-level EVD research network measures 0.15, with an average degree of 14.65 and an average weighted degree of 61.01.
Amongst all collaborations on country-level, nine research communities were identified using modularity-based community detection and visualised by different colours (Fig 6). The largest community (red) is centred around the US with strong collaborations to Canada, Germany and the UK, representing 59.41% of the co-authorships collaborations (weighted edges). Another large community is a (mostly francophone) European–African community (blue) representing 31.68% of all co-authorships connections.
Among all published original research articles between 1976 and 2015 a total of 1,644 co-author's affiliated institutions were named, which yielded 9,907 co-authorship connections in the overall research network (Fig 7). The main actors according to degree are the US government (CDC USA, d = 353; NIH, d = 315; USAMRIID, d = 283) and WHO (d = 256). Other prominent actors are from the US and European countries. Most central institutions are publicly funded (e.g. CDC USA, USAMRIID), government research institutions (e.g. BNI, ISERM), (mostly public) universities (e.g. Uni London, Univ Marburg) or international institutions (e.g. WHO) or non-governmental institutions (NGOs) (e.g. MSF).
Modularity analysis reveals 166 communities within the network (Fig 7), whereas the largest community (blue) represents 17.33% of the total network nodes and the second largest (green) represents 14.44% of the network nodes. Numerous smaller and less connected communities exist in the periphery, with some being entirely disconnected from the main network.
The temporal development of the research network is visualised over four 10-year time periods (Figs 8, 9, 10 & 11).
In the first decade 1976–1985, (Fig 8) the network consists of only a few actors, with one large central cluster surrounded by four smaller clusters. The German Bernhard-Nocht Institute (BNI) has the highest centrality degree (d = 11), closely followed by the Institut Pasteur, PHLS Center for Microbiology and Research (Salisbury, UK) and USAMRIID. The CDC USA is a central institution (d = 7) of a smaller cluster, publishing with African partners (Kenyan Ministry of Health) others. Smaller research groups in Kenya (Kemri Wellcome Trust, Institute of Primate Research, Kenya Trypanosomiasis Research Institute), UK and US published together, but had no connections with others.
In the second decade 1986–1995, (Fig 9) two larger, but separate, research communities evolved. One francophone French-Swiss-African community with a homogenous structure in which the Institut Pasteur published mainly with the University of Basel, Institut de recherche pour le développement (IRD), Ecole national veterinaire Lyon and the Hospital Bichat Claude Bernard Paris. The other community consists mostly of American and German institutions, with three main actors (USAMRIID, CDC USA and the University of Marburg), where the USAMRIID and CDC USA connect this community. During this period the WHO had its first appearance as a disconnected actor. All institutions in the network of the second decade are public entities.
With the occurrence of new EVD outbreaks in 1994/1995 the EVD research network grew in the third decade 1996–2005, (Fig 10) into a star-like structure with surrounding chains. During this decade the CDC USA evolved as the most central actor (d = 87). The University of Marburg (d = 54), USAMRIID (d = 52), WHO (d = 46) and NIH (d = 36) remain central but less prominent actors.
The network of the fourth decade 2006–2015, (Fig 11) is skewed by publications in 2014/2015. During this time only few public research institutions and university actors dominate the research collaborations but numerous new actors appeared. Prominent cooperation exist between CDC USA and WHO and CDC, NIH and USAMRIID. While the transnational WHO was well embedded in the network over these last two decades, all main network actors are public institutions, mostly from the US and European countries.
While the global EVD research network remains relatively consistent in the first two decades, the third and in particular the forth decade shows substantial overall increase in the number of institutions and the links between them (Table 1). Simultaneously the average node degree and weighted node degree increased over time, which indicates a growing number of collaborations and research activity per institution.
The decreasing density of the network over all decades indicates a decreasing number of realised edges between nodes relative to the total number of possible edges. The increasing average node degree implies a growing number of research connections per institution. The number of communities increased in line with number of nodes. The high modularity values show that the solutions of the community detection algorithm reflect the substructures of the graph well, i.e. the increase in communities is unlikely to represent a sheer increase in volume, but rather seems to capture an evolution of the field of EVD into several smaller communities.
A degree distribution analysis of the EVD research network in the fourth decade shows a skewed node-degree distribution (Fig 12). While almost 100 nodes appear with a degree of zero (d = 0), indicating no collaboration at all, only few institutions have a very high degree above 160 (mean 12.24; median 5). Most institutions had a degree of less than five (d≤5) as they were named as affiliations by authors of few publications by authors that published with only few co-authors. The few very well connected institutions, such as NIH and CDC USA, are the key actors in this period. In fact the CDC USA has maintained a very central position in the network over all time periods. The private NGO Médecins Sans Frontières (MSF) has only recently emerged within the network and is centrally embedded with a high degree (d = 157).
Since the first reported EVD outbreak in 1976 until today the total number of publications on EVD in WoS has exceeded more than 4500 publications, of which 2528 were original research articles. Like in scientometric analyses we used joint publishing as a proxy indicator of scientific collaboration [17] and thus knowledge exchange for our SNA of the co-authorship network [11,13,30]. Indeed for the EVD overall network we identified research contributions from 1,644 research institutions in 101 countries; most actors are indeed coming from the US [17]. Since 1994 EVD research publications have increased continuously, steadily and independently of the major West African outbreak. This growth in publications is mirrored by a growth in the number of institutions (from 30 to 1,489) and edges (from 60 to 9,176) and therefore on-going network growth accompanied by a decreasing network density. The overall network is an extensive aggregation of 166 different communities with a clearly dominant anglophone and francophone community. This same dominance is seen when analysing the most central actors by degree and betweenness centrality both confirming the dominance of 10 institutions in powerful, control or broker positions in the network [11,33,34].
The pattern of a growing EVD network in size but with a reducing density is characterised by some outliers (106 institutions not connected), frequently less connected contributions from developing countries and the private sector, but with a strong and stable core of dominant or ‘central’ institutions. These characteristics of the network are supported by many of the analyses we performed. For example the relatively and increasingly poorly connected nature of the network (network density), the heavily skewed node degree distribution with the median node degree remaining rather constant, the relatively compact nature of the network (path lengths) and the strong centralisation showing a dominance of a few very strongly connected actors and many poorly connected actors.
Although we acknowledge that our analysis is weakened by the absence of a comparator network (a common challenge in emerging research fields), we also believe that our analysis bring some added value. For example SNA metrics for the overall network shows a density of 0.007 and calculating network density for each decade individually showed progressively decreasing density from 0.138 in the first decade to 0.008 in the last decade. While this is largely influenced by both the size (the more actors a network includes, the more difficult it is for all actors to be connected) and also the correspondingly rapid growth in the network (connections take time to build), we still believe that these figures should raise questions about whether the network–and therefore research outputs–could benefit from greater connectivity and linkages and in doing so greater optimise knowledge transfer and the spread of innovation [15]. The node degree distribution (for the last decade from 2006–2015) further confirms both the observed increase in the average node degree is attributable to only a few central actors whereas the overall network was not well connected in this period. Thus, the network growth during the 2014/2015 epidemic diluted connectivity, at a time when collaboration was arguably most needed.
These observations are built on when we look further at the node degree distribution for 2006–2015. This confirms that while most actors only had few connections during this time, some actors are extremely connected. This distribution form has been described as “power law” or “scale free distribution” and is typically observed amongst poorly connected networks [35,36]. This ‘concentrated core’ is corroborated by the high number of the average weighted node degree (17.89), in contrast to the average node degree (12.05), which is also an indicator that some actors in the EVD network are connected more strongly to each other than others due to repeated publishing [27]. It shows that these actors have on-going collaborations, share research results intensely by jointly publishing—but focus sharing amongst their co-authors. This latter finding is something confirmed by our SNA results, which show strong centralisation amongst six institutions (CDC USA, NIH, USAMRIID, WHO, the University of Marburg and the University of Harvard), suggesting that knowledge is mostly exchanged within the network between and/or through these actors. Centrality is a measure of power in SNA [37], this is especially the case for our central actors whose knowledge broker status is confirmed with regard to EVD research due to their high degree and betweenness centralities.
Additionally, observation of the path lengths reveal further insight into the efficiency of information exchange, with the shorter the average path length of a network diameter, the more efficient is information exchanged within the network structure [26,35]. We found that the average paths lengths (3.02) of the overall network is lower than the average node degree (12.05), indicating both that some institutions have a lot of direct neighbours and that on average nodes can reach other nodes by crossing only two other nodes. The network diameter (8.0) suggests that sub-graphs within the network do not span more than across a chain of eight nodes. Taking both aspects into account this implies that the overall structure of the network is characterized by isolated and weakly connected components, i.e. localized small networks that have only few relations amongst each other.
Although our study cannot, unfortunately, reveal anything about the ‘type’ of research conducted, observations on the type of research institution maybe serve as a proxy for this insight. Two notable observations here were both the relative underrepresentation and disconnectedness in the overall network of both research institutions from affected countries and the private sector. Among the unconnected nodes appear some private industry actors (e.g., Novartis Vaccines, Biohelix Corp, Baxter Bioscience and Oravax Inc.), in addition to African universities such as the University of Benin and the University of Mbarara. While there may be many good reasons that explain the disconnectedness, for example proprietary restrictions to collaboration (in the case of industry), new entrants to the field or for resource-related barriers to International collaboration. This observation remains significant for a number of reasons, presumably both of these actor types posses’ unique and distinct knowledge and capabilities that could diversify and strengthen the expertise within the network if better and more broadly integrated, this is likely even more the case during a public health emergency of international concern. Also, this ability to identify disconnected but valuable nodes, demonstrates a great added value of tools such as SNA. Finally the recent entry into the network of non-traditional research actors such as MSF should be welcomed, especially as endemic country capacity is being developed and integrated into international networks, due to their unique position as being close to patients in the field yet able to advocate–distant funders–on the need for a well-supported, needs-driven research agenda [5,38].
We believe the structure, nature and evolution of the international EVD research network described in this paper presents some learnings for policy. Looking positively, the network itself has maintained a similar structure–a relatively compact network with a few consistent actors at its core–over the four decades studied, implying it is a stable constellation. This institutional memory provides a solid foundation for knowledge maintenance over time, indeed without central actors networks might be disrupted and knowledge exchange hampered [30]. The growth in the network over time through the entry of new actors, particularly since 2014/2015, is positive as it likely indicates the arrival of new ideas and approaches. However although collaboration has increased over time, our analysis found that the network remains relatively poorly connected. Hence there may be an additional role for the ‘central actors’ to expand their role beyond a hub for dissemination and exchange into a facilitator for integrating the newer actors and expertise into the network. Additional opportunities presented by the network analysis include: a reflection on the, perhaps, over-reliance or vulnerability to the network of all of the ‘central actors’ being public government or university institutions. The importance of predictable, sustainable, funding flows to their continued role as network ‘brokers’ feels more exposed in these current financially and politically turbulent times. While the dominance of these institutions is not surprising, we assume that they have the infrastructure, capability and public-financing, it may represent a weakness in two respects: firstly, with respect to its insufficiently diverse expertise mix, particularly with respect to the translation of this research into the development of tangible, context-relevant tools and capacity building in affected countries [8,39]; secondly, with respect to the risk of over-centralising expertise, resulting in the stifling or suppression of innovation and growth and development of new ideas.
Finally, in small research areas for diseases predominantly impacting the lives of those in low-income countries such as EVD, the inherent market failures indicate that this reliance of public-financing will likely continue [Wölfel in: 3,5–7]. Given this, we believe, that a valuable insight from our study is to observe ways in which the network efficiency could be enhanced to extract greater patient-impact from the public financing inputs. For example: focused efforts on integrating new collaborators into the network, provision of tools to enhance the productivity and improved transparency and sharing of research data [9,40] the identification of expertise gaps and targeted filling of these gaps and lastly, but perhaps most importantly, National alignment, focus and financing coordination (strategic research planning) around the globally agreed prioritised research agenda [41]. Although many of these calls have already been made by many actors, particularly since the 2014/2015 EVD outbreak we believe this study represents an important empirical tool to support these calls and inform National and global policy development as the global community works to avert the next EVD outbreak.
The use of bibliometric data has intrinsic limitations and restrictions related to any analysis of secondary data and where data ceases to provide information, in particular in relation to content or results of published research.
Two major limitations to our study were identified and previously highlighted. The first being the absence of other publications with which to contextualise and compare our results. This absence of relativity in our conclusions limits the comparative value of our findings although the absolute data remain valid. Although SNA is increasingly being used as a tool to analyses research areas it remains a relatively new field so we are optimistic that this is a time-limited constraint.
Secondly, we acknowledge that our study would be greatly enriched by an ability to analyse the data by ‘type’ of research not only type of publication i.e. basic, applied, clinical, implementation research, translation, health systems etc. However, at present, this is not a search field within WoS, so we were unable to attain the source data. Should key, public, medical, search engines enable this in the future, SNA such as ours would be an even more powerful tool to provide insight into research focus and productivity. This analysis we believe would have great value–supplementing existing financing and development pipeline analyses [42,43]—in providing a more granular understanding of product development gaps and the persistent absence of tools for the prevention, diagnosis and treatment of EVD [6,44]. Our analysis of decreasing network density over time could have been further triangulated with the use of an additional metric such as the percentage of the giant component or the clustering coefficient. Other limitations include reporting delays and the possibility that some publications were not included in the WoS database, however sample testing of other databases, including PubMed.gov, did not reveal other publications on EVD.
Although the impact of missing publications was likely small future studies could aggregate studies from diverse databases and in particular try to assess contribution of private industries R&D. Despite manual and automated attempts to resolve challenges with institution name cleaning and disambiguation it cannot be excluded that some actors and/or relationships were not captured or were captured incorrectly. Although unlikely, errors of the software used cannot be completely excluded and different algorithms might lead to different presentations of results. Therefore network visualisations should be critically assessed in context to minimise misinterpretations. We further note that GeoLayout visualisation can be misleading since it locates the African continent in the map centre and visualised edges may overlap nodes. For this reason a country distribution was processed additionally with Force Atlas 2. The use of only free available software and easy accessible bibliometric data from WoS both facilitate the easy reproducibility of our study.
We conducted the first systematic landscaping of published EVD global research bibliometrics using SNA tools for analysis and visualisation.
Since 1976 Ebola outbreak EVD research, numbers of authors and affiliated institutions and links between them are constantly increasing, mostly independent from outbreaks and in-particular in the past two decades.
The overall EVD research network is organised around a few co-authoring key actors, mostly publicly financed. Low network density indicates room for increased cooperation between institutions, in-particular links to less connected and more peripheral institutions could foster knowledge exchange and innovation. Key network actors, such as the CDC USA, maintained network coherence over time–and probably kept EVD research on-going. Limited scientific collaboration of research organisations from LMIC and the private industry, and how they utilise their expertise and knowledge, is neglected.
However, the absence of effective treatments for EVD questions the existing EVD research network efficacy and efficiency and suggests the need for both direction and structure to optimize the network to focus on research relevant for treatments. Since most institutions in the global network are publicly funded, guidance to direct and re-orientate research might be facilitated by funders (through calls targeting knowledge and translation gaps) and be offered by supranational policy setting entities such as WHO and its Global Observatory on Health Research and Development.
Further in-depth quantitative and qualitative analysis, e.g. text mining of publications abstracts, analysis of EVD research study methods and separate R&D product pipeline analysis, is recommended to ensure empirically based strategic research guidance and relevant to EVD product development.
In any case, SNA of co-authorship networks is an innovative tool to evaluate research collaborations between individuals, organizations and countries, contributes to the understanding of the evolution of research networks and should be used for strategic research planning and a regular monitoring.
|
10.1371/journal.pgen.1001002 | Identification of a Functional Genetic Variant at 16q12.1 for Breast Cancer Risk: Results from the Asia Breast Cancer Consortium | Genetic factors play an important role in the etiology of breast cancer. We carried out a multi-stage genome-wide association (GWA) study in over 28,000 cases and controls recruited from 12 studies conducted in Asian and European American women to identify genetic susceptibility loci for breast cancer. After analyzing 684,457 SNPs in 2,073 cases and 2,084 controls in Chinese women, we evaluated 53 SNPs for fast-track replication in an independent set of 4,425 cases and 1,915 controls of Chinese origin. Four replicated SNPs were further investigated in an independent set of 6,173 cases and 6,340 controls from seven other studies conducted in Asian women. SNP rs4784227 was consistently associated with breast cancer risk across all studies with adjusted odds ratios (95% confidence intervals) of 1.25 (1.20−1.31) per allele (P = 3.2×10−25) in the pooled analysis of samples from all Asian samples. This SNP was also associated with breast cancer risk among European Americans (per allele OR = 1.19, 95% CI = 1.09−1.31, P = 1.3×10−4, 2,797 cases and 2,662 controls). SNP rs4784227 is located at 16q12.1, a region identified previously for breast cancer risk among Europeans. The association of this SNP with breast cancer risk remained highly statistically significant in Asians after adjusting for previously-reported SNPs in this region. In vitro experiments using both luciferase reporter and electrophoretic mobility shift assays demonstrated functional significance of this SNP. These results provide strong evidence implicating rs4784227 as a functional causal variant for breast cancer in the locus 16q12.1 and demonstrate the utility of conducting genetic association studies in populations with different genetic architectures.
| Breast cancer is one of the most common malignancies among women worldwide. Genetic factors play an important role in the etiology of breast cancer. To identify genetic susceptibility loci for breast cancer, we performed a genome-wide association study in 15,468 breast cancer cases and 13,001 controls. A single nucleotide polymorphism (SNP) rs4784227 located on chromosome 16q12.1, a previously-reported region for breast cancer risk, was found to be associated with breast cancer risk. The association of this SNP with breast cancer risk remained highly significant in Asians after adjusting all previously-reported SNPs in this region. In vitro biochemical experiments using both luciferase reporter and electrophoretic mobility shift assays confirmed the functional importance of this SNP. Our results demonstrate the importance of conducting genetic association studies in populations with different genetic backgrounds to identify functional variants.
| Breast cancer is the most common malignancy among women in the United States and many other parts of the world. Genetic factors play an important role in the etiology of breast cancer. Only a very small fraction of cases in the general population, however, can be explained by high-penetrance breast cancer susceptibility genes, such as BRCA1 and BRCA2. Recent genome-wide association (GWA) studies [1]–[8], including our own study among Chinese women in Shanghai [6], have identified multiple common genetic susceptibility loci for breast cancer. Each of the common genetic factors identified thus far confer only a small to moderate risk for breast cancer. With the exception of our study, all other reported GWA studies have been conducted among women of European ancestry. GWA studies conducted in other populations could identify not only additional novel genetic variants for breast cancer but also help to fine map causal variants for regions reported from previous GWA studies.
In early 2009, we reported a novel genetic susceptibility locus at 6q25.1 for breast cancer risk in a fast-track replication of promising SNPs selected from a GWA scan of 1,505 cases and 1,522 controls recruited in the Shanghai Breast Cancer Study (SBCS) [6]. We have since increased the sample size for the initial GWA scan to 2,073 cases and 2,084 controls to increase the statistical power to identify novel genetic risk variants for breast cancer. We have recently completed the second fast-track replication using data and biological samples collected from 13,395 cases and 10,917 controls recruited in 12 studies of Asian and European ancestry. SNP rs4784227, located at 16q12.1, a region identified from a previous GWA study conducted in Europeans [1], [3], was found to be a risk variant for breast cancer in Asian women independent of SNPs reported from the previous study [1], [3]. In vitro experimental results provide strong support for the functional significance of this SNP and suggest that this SNP may explain the association observed for breast cancer in this locus. Herein, we report findings from this large genetic study of breast cancer.
Approval was granted from relevant review boards in all study sites; all included subjects gave informed consent.
Included in this consortium project were 15,468 cases and 13,001 controls from 12 studies (Table 1). Detailed descriptions of these participating studies and demographic characteristics of study participants are provided in the supplement Text S1 and Table S1. Briefly, the consortium included 19,796 Chinese women from seven studies conducted in Shanghai [6], [9] (three studies, n = 10,497), Tianjin [10] (n = 3,115), Nanjing [11], [12] (n = 2,885), Taiwan [13] (n = 2,131), and Hong Kong [14] (n = 1,168); 3,214 Japanese women from three studies conducted in Hawaii [15] (n = 1,120), Nagoya [16] (n = 1,288), and Nagano (n = 806) [17]; and 5,459 European Americans from the Nashville Breast Health Study (NBHS, n = 3,172) and the Nurses' Health Study (NHS, n = 2,287, included as part of the Cancer Genetic Markers of Susceptibility Project - CGEMS). All cases and controls recruited in the Shanghai studies were included in Stages I and II, and subjects from the remaining Asian studies were included in Stage III. Data from CGEMS were used for help to select SNPs for Stage II. Cases and controls recruited in NBHS and the NHS (CGEMS) were included in the final stage to evaluate the generalizability of the findings.
Genotyping for Stage I has been described previously [6]. Briefly, the initial 300 subjects were genotyped using the Affymetrix GeneChip Mapping 500 K Array Set and the remaining 3,918 subjects were genotyped using the Affymetrix Genome-Wide Human SNP Array 6.0. We included one negative control and three positive quality control (QC) samples from the Coriell Cell Repositories (http://ccr.coriell.org/) in each of the 96-well plates for Affymetrix SNP Array 6.0 genotyping. A total of 127 positive QC samples were successfully genotyped and the average concordance rate was 99.9% with a median value of 100%. The sex of all study samples was confirmed to be female. The identity-by-descent analysis based on identity-by-state was conducted to detect first-degree cryptic relationships using PLINK, version 1.06. All samples with a call rate <95% were excluded. The SNPs were excluded if: (i) minor allele frequency (MAF) <1%, (ii) call rate <95%, or (iii) genotyping concordance rate <95% in quality control samples. The final dataset included 2,073 cases and 2,084 controls for 684,457 markers.
Genotyping for Stage II was completed using the iPLEX Sequenom MassArray platform. Included in each 96-well plate as QC samples were two negative controls (water), two blinded duplicates, and two samples from the HapMap project. To compare the consistency between the Affymetrix and Sequenom platforms, we also included 124 samples from Stage I that were genotyped by Affymetrix SNP 6.0. The mean concordance rate was 99.7% for the blind duplicates, 98.8% for HapMap samples, and 98.6% between Sequenom and Affymetrix 6.0 genotyping.
Genotyping for Stage III and NBHS was performed using TaqMan at five different centers. The genotyping assay protocol was developed and validated at the Vanderbilt Molecular Epidemiology Laboratory, and TaqMan genotyping assay reagents were provided to investigators from the Tianjin study (Tianjin Cancer Institute and Hospital), Nanjing study (Nanjing Medical University), Multiethnic Cohort Study (MEC, University of Southern California), and Nagano Breast Cancer study (Japan National Cancer Center), who conducted the genotyping assays at their own laboratories. Samples from the four other studies (Hong Kong, Taiwan, Nagoya, and Nashville) were genotyped at the Vanderbilt Molecular Epidemiology Laboratory. During the genotyping, two negative controls were included in each 96-well plate, along with 30 unrelated European and 45 Chinese samples from the HapMap project genotyped together with each study for QC purposes. The consistency rate was 100.0% for the HapMap samples comparing genotyping data obtained from the current study with data obtained in the HapMap project. Each of the non-Vanderbilt laboratories was asked to genotype a trial plate containing DNA from 70 Chinese samples before genotyping study samples. The consistency rate across all centers for these trial samples was 100% compared with genotypes previously determined at Vanderbilt. In addition, replicate samples comparing 3–7% of all study samples were dispersed among the genotyping plates at all centers.
A 3.0 kb DNA fragment containing major allele (C) of rs4784227 was PCR amplified by using forward primer 5′-GATCAGCTAGCCATAGTGTGGTAGCTAGTTG-3′ and backward primer 5′-GATCA CTCGAGCTGCTGGGCTTAGCTACAAG-3′. This fragment was subcloned into luciferase reporter vector, pGL3 basic, pGL3 promoter, and pGL3 enhancer (Promega, WI) between Nhe1 and Xho1 restriction sites, respectively. The minor allele (T) sequence was generated by using QuickChange Site-Directed Mutagenesis Kit (Strategene, La Jolla, CA) with the following pair of oligonucleotides, 5′- GAGTATTTACATCACAATAATCAGCAAACACTACAAATTGGGAC-3′ and 5′- GTCCCAATTTGTAGTGTTTGCTGATTATTGTGATGTAAATACTC-3′. All DNA constructs were verified by sequencing analyses. Transfection was performed with the use of FuGene 6 Transfection Reagent (Roche Diagnostics, Indianapolis, IN) in triplicate for each of the constructs. Briefly, 1−2×105 cells of HEK 293, MCF-7, MCF10A, and MDA-231 cells were seeded in 24-well plates and co-transfected with pGL4.73, a Renilla expressing vector which served as a reference for transfection efficiency. Thirty-six to 48 hours later the cells were lysed with Passive Lysis Buffer, and luminescence (relative light units) was measured using the Dual-Luciferase Assay System (Promega, WI). The rs4784227 regulatory activity was measured as a ratio of firefly luciferase activity to renilla luciferase activity, and the mean from four independent experiments are presented.
Biotin-labeled, double stranded oligonucleotide probes 5′-ATTTGTAGTGTTTGCCGATTATTGTGATGT-3′ and 5′-ACATCACAATAATCGGCAAACACTACAAAT-3′, and 5′-ATTTGTAGTGTTTGCTGATTATTGTGATGT-3′and 5′- ACATCACAATAATCAGCAAACACTACAAAT-3′ containing the major and minor allele sequence of rs4784227 were synthesized. The probes were incubated with nuclear protein extracts from MCF10A, MCF7, and MDA-MB-231 cells, in the presence or absence of competitors, i.e. unlabelled probes. Protein-DNA complexes were resolved by polyacrylamide gel electrophoresis and detected using a LightShift Chemiluminescent EMSA kit (Pierce Biotechnology, Rockford, IL).
In Stage I, PLINK version 1.06 was used to analyze genome-wide data. Population structure was investigated by using the principal component analysis implemented in EIGENSTRAT [18] (http://genepath.med.harvard.edu/~reich/Software.htm). A set of 12,533 SNPs with MAF≥5% in Chinese and a distance ≥25 kb between two adjacent SNPs was selected to evaluate the population structure. The first two principal components were included in logistic regression models for adjustment of population structures. Odds ratios (OR) and 95% confidence intervals (CIs) were estimated by logistic regression analysis. ORs were also estimated for the variant allele based on a log-additive model. Age was adjusted for in the analyses of Stages I and II data. In Stage III, individual data were obtained from each study for a pooled analysis. ORs from multiple studies were adjusted for age and study site. Heterogeneity across studies and between ethnicities was assessed with likelihood ratio tests. Stratified analyses by ethnicity, menopausal status, and estrogen receptor (ER) status were carried out. P-values based on 2-tailed tests are presented.
Individual genotyping data from the Cancer Genetic Markers of Susceptibility (CGEMS, http://cgems.cancer.gov/data/) study were obtained through an approved data request application in order to perform meta-analyses of GWA scan data from both the Shanghai studies and the CGEMS project. Program MACH (http://www.sph.umich.edu/csg/abecasis/MACH/) was used for genotype imputation that determines the probability distribution of missing genotypes conditional on a set of known haplotypes while simultaneously estimating the fine-scale recombination map. For the Shanghai studies, imputation was based on 660,118 autosomal SNPs genotyped in Stage I that had a MAF>1% and passed the QC procedure, using the phased Asian data from HapMap Phase II (release 22) as the reference. A total of 2,272,352 SNPs showed an imputation quality score ≥0.90. The CGEMS GWA scan data were genotyped using Illumina HumanHap550 for 1,142 breast cancer cases and 1,145 controls nested within the Nurses' Health Study cohort. For CGEMS, genotypes were imputed based on 513,602 autosomal SNPs with MAF>1%, using phased CEU data from HapMap Phase II (release 22) as the reference. A total of 2,168,847 SNPs showed an imputation quality score ≥0.90. To evaluate associations between imputed SNP data and breast cancer risk, logistic regression (additive model) was used, in which SNPs were represented by the expected allele count, an approach that takes into account the degree of uncertainty in the genotype imputation.
Meta-analyses of GWA scan data for SBCS and CGEMS were conducted for 1,968,549 SNPs with a MAF ≥1% in both populations and imputation quality scores ≥0.90. Meta-analyses were performed using a weighted z-statistics method, where weights were proportional to the square root of the number of individuals in each sample and standardized such that the weights added up to one. The z-statistic summarizes the magnitude and direction of the effect relative to the reference allele. An overall z-statistic and p value were then calculated from the weighted average of the individual statistics. Calculations were implemented in the METAL package (http://www.sph.umich.edu/csg/abecasis/Metal).
Of the 53 successfully genotyped SNPs in Stage II (Table S2), highly significant associations with breast cancer risk were found for rs4784227 (16q12.1) with OR (95% CI) of 1.23 (1.16–1.31) per T allele (P for trend, 1.3×10−8) (Table 2). Three other SNPs also showed a significant or marginally significant association with breast cancer risk. These four SNPs were selected for further validation in Stage III, which included 6,173 cases and 6,340 controls of Asian ancestry from seven studies in the Asia Breast Cancer Consortium (Methods; Table S1). SNP rs4784227 was consistently associated with breast cancer risk in all studies (Figure 1), with an OR of 1.25 (95% CI: 1.20–1.31, P = 3.2×10−25) in the pooled analysis of Asian samples from all three stages. No heterogeneity in the association of this SNP with breast cancer was observed across the studies included in the consortium. The association of rs4784227 with breast cancer risk was observed in both pre-menopausal (OR: 1.24 (1.17–1.32) and P = 6.5×10−12), and post-menopausal women (OR: 1.27 (1.19–1.35) and P = 3.0×10−14) (data not shown in tables). The positive association was stronger in ER(+) breast cancer (per allele OR = 1.29, 95% CI = 1.23–1.36, P = 3.0×10−23) than in ER(−) breast cancer (per allele OR = 1.19, 95% CI = 1.12–1.26, P = 1.3×10−8). In case-only analyses, when compared with cases with ER(−) cancer, ORs associated with ER(+) breast cancer were found to be 1.09 (95% CI: 1.03–1.16; P for trend, 5.8×10−3). None of the other three SNPs that showed a significant association in Stage II, however, were replicated in Stage III (Table S3).
SNP rs4784227 is located in 16q12.1, a region where three genetic risk variants for breast cancer (rs8051542, rs12443621, and rs3803662) were reported previously in a study conducted among women of European ancestry [1]. Of these three previously reported SNPs, the closest (rs3803662) is approximately 12.8 Kb away from rs4784227. The linkage disequilibrium (LD) pattern of this region in Asians is very different from the pattern found in European descendents (Figure 2 and Table 3). In Stage I and II samples, SNP rs4784227 is in low LD with previously-reported SNPs, with r2 being 0.07, 0.14, and 0.37 for rs12443621, rs3803662, and rs8051542, respectively (Table 3). In European Americans included in the HapMap project, however, SNP rs4784227 is in strong LD with SNP rs3803662 (r2 = 0.86) but weakly correlates with the other two SNPs. SNPs rs8051542 and rs3803662 each showed a significant association with breast cancer risk (P = 2.0×10−3 and P = 1.7×10−4) in a combined analysis of data from Stages I and II (Table 4). However, after adjusting for rs4784227 the association with rs8051542 disappeared. Although the positive association with rs3803662 remained, it was of only borderline significance (P = 0.12). SNP rs12443621, however, showed no association with breast cancer risk, which is consistent with results reported by the initial study of Asian women [1]. The association of rs4784227 with breast cancer risk remained highly significant after adjusting for these three previously-reported SNPs, individually or in combination (Table 4). Haplotype analyses of these four SNPs showed that all haplotypes containing the T allele of rs4784227 were associated with an increased risk of breast cancer, although not all point estimates were statistically significant at P<0.05 due to a small sample size for several haplotypes (Table S4).
In studies conducted in European Americans, SNP rs4784227 also showed a significant association with a per allele OR (95% CI) of 1.17 (1.03–1.34) in CGEMS and 1.21 (1.07–1.37) in NBHS (Table 5). A significant (NBHS) or marginally significant (CGEMS) association was observed for rs3803662, a previously reported SNP, but not for two other previously-reported SNPs. After adjusting for rs4784227, no association with rs3803662 was seen. On the other hand, the positive association with rs4784227 remained after adjusting for rs3803662 or the other two SNPs, although the association was no longer statistically significant at P<0.05.
To evaluate whether SNP rs4784227 has any intrinsic regulatory function, we conducted an in vitro luciferase assay in four cell lines including metastatic breast cancer cell MDA231, non-metastatic breast cancer cell MCF-7, breast epithelial cell MCF10A, and HEK293. Luciferase reporter constructs containing a 3 kb DNA fragment with the reference allele C and the risk allele T of rs4784227, respectively, were generated and transiently transfected into these cells. By comparing to the respective empty vectors, no luciferase activity change was observed in pGL3 basic and pGL3 enhancer vectors that harbor rs4784227 fragments, which indicate that rs4784227 fragments do not have intrinsic promoter activity (data not shown). In contrast, in the pGL3 promoter vector, fragments containing rs4784227 reduced luciferase activity, and the reduction was more apparent in fragments containing risk allele T than the reference allele C (Figure 3A). With the exception of the MCF7 cells, the difference between the T and C allele was statistically significant at P≤0.05.
To investigate whether the DNA sequence containing rs4784227 may interact with nuclear proteins and if so, whether a single nucleotide change in the rs4784227 site may alter the protein-DNA interactions, we performed electrophoretic mobility shift assays. In these assays, oligonucleotide probes corresponding to the reference allele C or the risk allele T were incubated with nuclear protein extracts from MCF10A, MCF-7, and MDA-231 cells. Compared with reference allele C, risk allele T in rs4784227 resulted in an altered DNA-protein complex intensity in these cells (B and II) (Figure 3B). In contrast, risk allele T did not alter the intensity of the nonspecific DNA-protein complex band (I). These results were unaffected by the presence of large amounts of unlabeled competitors.
In this multi-stage GWA study of over 15,486 cases and 13,001 controls, we identified SNP rs4784227 as highly significantly associated with breast cancer in both Asians (per allele OR = 1.25, 95% CI = 1.20–1.31, P = 3.2×10−25) and European Americans (per allele OR = 1.19, 95% CI = 1.09–1.31, P = 1.3×10−4). SNP rs4784227 is located at 16q12.1, a region reported previously to harbor breast cancer genetic risk variants among European descendents [1], [3]. In Asians, however, this SNP is either not in LD or only in weak LD with any of the three previously-reported SNPs in these regions, and adjusting for these SNPs did not alter the association of breast cancer with this newly-identified SNP. Although in European Americans rs4784227 is in strong LD with one of the previously-reported SNPs, rs3803602, the positive association of rs4784227 with breast cancer remained after adjusting for previously-reported SNPs. In vitro experiments showed that risk allele T reduced luciferase activity and altered DNA-protein binding patterns. These results implicate rs4784227 as a functional genetic risk variant for breast cancer, and this SNP may explain, at least partially, the association of breast cancer with other SNPs identified in 16q12.1.
SNP rs4784227 is located 18.4 kb upstream of the TOX3 gene and in the evolutionarily-conserved region of intron of the LOC643714 gene. Several transcription factors are predicted to bind to this SNP (http://www.cbrc.jp/research/db/TFSEARCH.html). This SNP, however, has not been shown to be in the coding region for any non-coding RNA or miRNA/snoRNA/scaRNA based on UCSC Genome Browser. Our luciferase reporter assays showed that the intronic region harboring rs4784227 may have intrinsic repressor activities, suggesting that rs4784227 may affect its underlying gene LOC643714 or its neighborhood gene expression and thus affect breast cancer risk. The rs4784227-associated repressing activity could be the result of differential binding affinity of transcription machinery to the rs4784227-containing DNA sequences. We examined this hypothesis by conducting electrophoretic mobility shift assays and confirmed that the risk T allele of rs4784227 significantly alter DNA-nuclear protein(s) interactions. Thus, it is possible that inhibitory nuclear protein(s) selectively bind to the risk allele T to repress transcription. A database search (http://www.cbrc.jp/research/db/TFSEARCH.html) for transcription factor binding sites showed that the sequence at the rs4784227 site has a high degree of similarity with several consensus elements recognized by transcription factors, of which HNF-3b and C/EBP prefer to bind DNA fragments with the risk T allele of this SNP site. However, these putative transcription factors or their associated proteins have not been confirmed to be involved in the regulation of LOC643714 or its nearby genes.
In summary, through a GWA study we have identified and confirmed rs4784227 as a genetic risk variant for breast cancer. In vitro experiments showed a functional significance of this SNP that may explain the association of breast cancer with other SNPs identified at locus 16q12.1. This study demonstrates the importance of conducting genetic association studies in populations with different LD structures to identify causal genetic variants for breast cancer and other complex diseases.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.