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.1000821 | Natural Strain Variation and Antibody Neutralization of Dengue Serotype 3 Viruses | Dengue viruses (DENVs) are emerging, mosquito-borne flaviviruses which cause dengue fever and dengue hemorrhagic fever. The DENV complex consists of 4 serotypes designated DENV1-DENV4. Following natural infection with DENV, individuals develop serotype specific, neutralizing antibody responses. Monoclonal antibodies (MAbs) have been used to map neutralizing epitopes on dengue and other flaviviruses. Most serotype-specific, neutralizing MAbs bind to the lateral ridge of domain III of E protein (EDIII). It has been widely assumed that the EDIII lateral ridge epitope is conserved within each DENV serotype and a good target for vaccines. Using phylogenetic methods, we compared the amino acid sequence of 175 E proteins representing the different genotypes of DENV3 and identified a panel of surface exposed amino acids, including residues in EDIII, that are highly variant across the four DENV3 genotypes. The variable amino acids include six residues at the lateral ridge of EDIII. We used a panel of DENV3 mouse MAbs to assess the functional significance of naturally occurring amino acid variation. From the panel of antibodies, we identified three neutralizing MAbs that bound to EDIII of DENV3. Recombinant proteins and naturally occurring variant viruses were used to map the binding sites of the three MAbs. The three MAbs bound to overlapping but distinct epitopes on EDIII. Our empirical studies clearly demonstrate that the antibody binding and neutralization capacity of two MAbs was strongly influenced by naturally occurring mutations in DENV3. Our data demonstrate that the lateral ridge “type specific” epitope is not conserved between strains of DENV3. This variability should be considered when designing and evaluating DENV vaccines, especially those targeting EDIII.
| Dengue viruses are mosquito-borne flaviviruses and the agents of dengue fever and dengue hemorrhagic fever. It has been widely assumed that antibodies that neutralize dengue bind to regions on the viral envelope (E) protein that are conserved within each serotype. However, few studies have explored how natural variation influences dengue-antibody interactions. Mouse antibodies that strongly neutralize dengue bind to a region on domain III of E protein. This region has been the focus of much recent work because it might be the target of protective human antibodies as well. We compared a large number of E protein sequences and discovered that the region on E protein domain III targeted by neutralizing antibodies was highly variable between strains of dengue serotype 3. Using a panel of antibodies, we experimentally demonstrate that natural strain variation in dengue serotype 3 has a strong influence on antibody binding and neutralization. Our results challenge the dogma that neutralizing antibody binding regions are conserved within each serotype. The results of this study are relevant to the current global effort to develop and evaluate dengue vaccines.
| Dengue viruses (DENVs) are mosquito-borne flaviviruses and the agents of dengue fever and dengue hemorrhagic fever (DHF). According to the World Health Organization, over 2.5 billion people are at risk of contracting dengue, 100 million people develop symptomatic infections and up to 50,000 die from DHF each year. The DENV complex consists of 4 serotypes (DENV1-DENV4). DENVs have antibody epitopes that are unique to each serotype and epitopes that are cross reactive between serotypes. People who have recovered from primary DENV infections develop long term, protective immune responses against the homologous serotype only. In fact, individuals exposed to a second infection with a different serotype face a greater risk of developing DHF indicating that pre-existing immunity can exacerbate disease under some conditions [1].
As previously infected individuals do not appear to be re-infected with the same serotype, it is widely assumed that neutralizing antibody epitopes are conserved among strains belonging to the same serotype [2],[3]. In fact, the current strategy for developing dengue vaccines is based on the assumption that a neutralizing immune response directed to a single strain will protect against most if not all strains of DENV within the serotype. However, there is considerable genetic diversity within each serotype such that each has been subdivided into genotypes [4]. Despite this diversity, surprisingly few studies have explored how naturally occurring strain variation within each serotype influences DENV neutralization. Blaney and colleagues immunized monkeys with candidate live attenuated dengue vaccines and characterized the immune response in monkeys by using a panel of viruses representing the 4 serotypes and genotypes within each serotype. They observed large differences in neutralization titer when comparing different genotypes of DENV3 [5]. In a study of pediatric dengue cases in Thailand, investigators observed significant differences in the ability of sera to neutralize reference and clinical strains of DENV3 [6]. Guzman and colleagues reported that amino acid sequence differences between DENV3 strains can have strong influences on virus neutralization by murine and human immune sera [7]. Studies with other flaviviruses have also demonstrated that neutralization is dependent on the lineages and strains used in the assay [8],[9]. Thus, the current paradigm that neutralizing antibody epitopes are conserved within each serotype may not accurately depict the complexity of the antigenic relationships, especially in DENV3.
Antibodies, in particular, have emerged as key effector molecules responsible for protective and pathogenic immune responses to DENV [1]. The DENV envelope (E) protein is the major target of neutralizing antibody [10]. E protein mediates attachment to host cells and low pH fusion of the viral and host cell membranes. The crystal structures of E from several flaviviruses (tick borne encephalitis, DENV2, DENV3 and West Nile) have been solved [11]–[14]. Individual subunits of E protein consist of three beta-barrel domains designated domains I (EDI), II (EDII) and III (EDIII) and the native protein is a homodimer [11],[12],[14]. Mouse monoclonal antibodies (MAbs) that bind to all three domains of DENV E have been generated and characterized [10], [15]–[17]. The most potent neutralizing MAbs bind to an epitope on the lateral ridge of EDIII of flaviviruses [10],[18],[19]. This epitope, which is not conserved between dengue serotypes, has been the focus of much recent work because it might be the target of the natural human immune response that leads to type specific neutralization. Investigators are also testing EDIII as a vaccine for inducing antibodies that neutralize a specific serotype, without inducing serotype cross reactive antibodies with potential for disease enhancement [20],[21].
In the present study, we have examined the phylogenetic relatedness of the E protein sequences from a large number of viruses representing the different genotypes of DENV3. Many surface exposed amino acids were variable between established genotypes of DENV3. Especially noteworthy was the observation that the EDIII lateral ridge, which is a known site targeted by neutralizing MAbs in related flaviviruses, was variable between DENV3 genotypes. We experimentally demonstrate that naturally occurring amino acid differences in DENV3 EDIII lead to differential binding and neutralization by MAbs.
Aedes albopictus C6/36 cells were maintained at 28C in MEM (Gibco) supplemented with 10% fetal bovine serum (FBS) (Gibco), penicillin (100 U/ml) and streptomycin (100 µg/ml) in the presence of 5% CO2. Human leukemic monocyte lymphoma cell line U937 expressing DC-SIGN (U937 DC-SIGN) were maintained at 37C in RPMI (Gibco) supplemented with 10% FBS, 50 mM beta mercaptoethanol, penicillin (100 U/ml) and streptomycin (100 µg/ml) in the presence of 5% CO2. All media were also supplemented with 0.1 mM non-essential amino acids (Gibco) and 2 mM glutamine (Gibco).
Working virus stocks were obtained by inoculating C6/36 mosquito cells in MEM (Gibco) supplemented with 2% FBS (Gibco), penicillin (100 U/ml) (Gibco) and streptomycin (100 µg/ml) (Gibco) 0.1 mM non-essential amino acids (Gibco) and 2 mM glutamine (Gibco) and growing the virus for eight days at 28C under 5% CO2. Supernatants were harvested, clarified by centrifugation and, supplemented with 15% FBS and stored in aliquots at −80C. Viral titers were determined by plaque assay on Vero-81 cells as previously described [22] and only stocks with a titer above 105 PFU/ml were used in experiments. The reference virus strains used in the study were strains West Pacific 74 (DENV 1), S16803 (DENV2), CH53489 (DENV3) and TVP-360 (DENV4) routinely used in the DENV neutralization test. These viruses were obtained from Robert Putnak (Walter Reed Army Institute of Research, MD) and they have been passaged >10 times in different mammalian (Vero, Diploid fetal rhesus lung and Primary African Green monkey kidney cells) and insect (Aedes albopictus C6/36 cells) cell lines. For studies on different genotypes of DENV3, we also used UNC3043 (strain 059.AP-2 from Philippines, 1984), UNC 3009 (D2863, Sri Lanka 1989), and UNC3066 (strain 1342 from Puerto Rico 1977). These viruses were obtained from Dr. Duane Gubler and Claire Wong at CDC, Fort Collins, CO. These viruses had been passaged 3 times in Aedes albopictus C6/36 cells prior to being used in these studies.
MAbs 8A1 (IgG1) and 14A4 (IgG1) against DENV3 were provided by Robert Putnak (Walter Reed Army Institute of Research, MD). MAb1H9 (IgM) was provided by John Aaskov (Queensland University of Technology, Australia) [23]. MAb 1A1-D2 was provided by John Roehrig, (DVBID, CDC, Fort Collins, CO). MAbs 8A5 (IgG1) and 12C1 (IgG1) were generated for this study by immunizing mice with purified DENV3 strain CH53489.
Vero-81 cells were inoculated with UNC 3043 (DENV3 -genotype I), CH53489 (DENV3 - genotype II), UNC 3009 (DENV3 - genotype III), or UNC3066 (DENV3 -genotype IV) at an MOI of 0.1. The virus-containing media was harvested 5–7 days after infection and centrifuged to pellet cell debris. The clarified media was laid on top of a 20% sucrose (wt/vol) cushion and centrifuged (72,000×g for 5 h) to pellet the virus. The virus pellet was allowed to dissolve overnight in PBS before layering on a 10%–40% iodixanol gradient and being centrifuged at 163,700×g for 120 min. The virus-containing fractions were harvested. PBS was added to the virus to dilute the iodixanol. The diluted solution was centrifuged (72,000×g for 5 h) to pellet the virus and remove the iodixanol. The virus pellet was resuspended in PBS and virus protein content was estimated by spectrophotometry. The virus was stored at −80°C.
Recombinant EDIII constructs were created using cDNA from the following virus strains to represent each serotype of DENV and genotypes of DENV3: West Pacific 74 (DENV 1), S16803 (DENV2), UNC 3043 (DENV3 -genotype I), CH53489 (DENV3 - genotype II), UNC 3009 (DENV3 - genotype III), UNC3066 (DENV3 -genotype IV), and TVP-360 (DENV4). Envelope gene fragments encoding EDIII from DENV1, DENV3, and DENV 4 (AA295–398) and DENV2 (AA297–399) were amplified using Vent polymerase (NEB, Ipswich, MA). Reverse primers used in the study were designed to introduce either Hind III (for DENV2–4) or PstI (for DENV1) restriction site at the 3′ ends of the PCR products. PCR products were digested with HindIII or PstI and cloned into pMAL c2X vector (NEB) to generate recombinant EDIII that is fused to maltose binding protein (MBP-EDIII) at the N terminus according to the manufacturer's instructions. MBP-EDIII were expressed in E.coli DH5α (Invitrogen) and purified using amylose resin affinity chromatography (NEB) according to the manufacturer's instructions.
Selected amino acids residues on rEDIII were mutated by site directed mutagenesis using Quickchange multi kit (Stratagene, La Jolla, CA). When selecting sites to mutate, we gave precedence to positions on loops and (not beta sheets) because we did not want to disrupt the overall folding of EDIII. Thus, of the 20 positions we mutated, 16 (301–303- N-terminal linker loop; 323, 325–330- BC loop; 357, 358, 361- DE loop; 380, 382, 383- FG loop) are located on loops that form the lateral ridge neutralizing epitope recognized by type specific neutralizing MAbs [15],[17]. We also mutated amino acids on the A strand (positions 304, 308, 310) because this strand forms a dengue subcomplex epitope recognized by neutralizing MAbs [15]. We mutated position 386 on the G strand because Serafin and Aaskov reported that mutations at this position lead to escape from 1H9 antibody used in the current study [23]. PCR primers were designed using QuikChange® Primer Design Program (www.stratagene.com) and PCR was conducted according to manufacturer's instruction. Single stranded pMal c2X plasmids (NEB, Ipswich, MA) encoding MBP-EDIII fusion proteins with amino acid substitutions were cloned into DH5α cells for expression and purification of mutant rEDIIIs. Substitution of amino acids in all mutant constructs was confirmed by sequencing. Expression and purification of mutant rEDIII proteins were essentially same as mentioned in the earlier section.
ELISA plates were coated by adding 200 ng/well of purified EDIII-MBP protein or 75 ng/well of purified DENV antigen in Carbonate buffer (pH 9.0) and incubating the plates overnight at 4C. Rabbit anti MBP sera (New England Biolabs) was used to quantify binding of MBP-EDIII to plates. The flavivirus cross reactive MAb 4G2 was used to quantify binding of virus to plates. Two hundred nanograms of EDIII-MBP saturated binding to the ELISA plate and we did not observe appreciable differences in binding between different EDIII proteins created for this study. Similarly, 75 ng saturated virus binding to the plate and we did not observe appreciable differences in binding between different viruses used in the current study.
The plates were washed with Tris buffered saline with 0.2% Tween 20 (TBST wash buffer) and blocked with 3% normal goat serum (NGS) in Tris buffered saline with 0.05% Tween 20 (TBST blocking buffer) for 1 hour at 37C. Serially diluted MAbs in TBST blocking buffer were then added to each well and incubated for one hour at 37C. After washing 3 times with TBST wash buffer, the plates were incubated for one hour at 37C with alkaline phosphatase conjugated goat anti-mouse antibody (Sigma). Plates were washed 3 times with TBST wash buffer and developed by adding p-nitrophenyl phosphate substrate (Sigma). Optical density (OD) was measured at 405 nm using a spectrophotometer.
The flow cytometry based neutralization protocol as described by Kraus, et. al., was used with modifications to determine 50% neutralization values for each antibody [22]. Same amount of virus (2×107 genome equivalent copies) from each DENV3 genotype was used to infect cells in experiments comparing the neutralization activity of MAbs among the DENV genotypes. Different concentrations of MAbs were mixed with each virus strain in a 96 well tissue culture plate and incubated for one hour at 37C in the presence of 5% CO2. U937DC-SIGN cells (5×104 cells in 100 µl) were introduced to each well and incubated for an additional 2 hours at 37C to allow virus to bind to cells. The cells were then washed with media and 200 µl of fresh media was added to each well and incubated for 24–72 hr at 37C with 5% CO2. After washing 2 times with PBS, the cells were fixed and permeabilized using CytoFix/Cytoperm kit (BD bioscience). The cells were then stained with Alexa 488 conjugated anti dengue MAb 2H2 and the percentage of infected cells was measured in a flow cytometer. EC50 values were calculated using GraphPad Prism version 4.00 for Windows (GraphPad Software, San Diego California USA, www.graphpad.com) and non linear regression analysis.
A total of 175 unique DENV3 full-length envelope protein sequences were downloaded from GenBank and these were aligned using ClustalX version 1.83 [24] using the PAM distance matrix and default parameters. A variety of parameters and substitution matrices for the alignment were evaluated using the program TuneClustalv1.0 (http://www.homepage.mac.com/barryghall/Software.html) and the PAM series matrix was determined to be the most appropriate, with default gap opening and extension values. The alignment was used to identify 32 variable/informative sites that were defined as columns of heterogeneity in the alignment where the same amino acid change occurred in at least three independent sequences. To display the 32 informative sites (Figure 1), we selected a representative subset of 28 sequences that contained at least 4 sequences from each genotype of DENV3 depicted in Figure S1.
As individuals infected with DENV appear to develop a long term, protective immune response to the homologous serotype, it has been assumed that neutralizing antibody epitopes are conserved within each serotype [10],[25]. To further evaluate this assumption, we used phylogenetic approaches to compare the full length E protein sequence of 175 DENV3 strains. The sequences, which were obtained from Genbank, had representatives of each of the 4 recognized genotypes of DENV3 [26]. Amino acid positions that were variable in 3 or more independent sequences as identified by alignment were defined as informative sites. Thirty two of the 493 amino acids on DENV3 E protein were identified to be informative sites (Figure 1). Individual subunits of E protein consist of three beta-barrel domains designated domains I (EDI), II (EDII) and III (EDIII) (Figure 2A). Informative sites were present in all domains of the E protein (Figures 1 and 2). Many of the informative sites had mutations that were conserved within but not between DENV3 genotypes (Figure 1). Twenty eight of the thirty two informative sites were located on the ectodomain (AA 1 to 392) of E protein. As the structure of the ectodomain of DENV3 E protein has been solved, we were able to determine if the variant sites were either surface exposed or buried [12]. Eighteen of 28 informative sites in the ectodomain were surface exposed, while others were partially or completely buried in the molecule (Figures 1 and 2).
We compared the locations of known antibody epitopes on the flavivirus E protein and the positions of informative sites on DENV3 E protein. Mouse monoclonal antibodies (MAbs) that bind to E protein have been mapped to six regions (Figure 2A and Table 1) [10],[18]. Most of the informative sites on EDI and EDII were within or adjacent to these antigenic regions (Figure 2A and Table 1). However, the antigenic region at the fusion loop was completely conserved between DENV3 strains (Table 1).
Many MAbs that strongly neutralize flaviviruses bind to EDIII. DENV serotype specific (type specific), neutralizing MAbs bind to epitopes on the lateral ridge of EDIII, which is formed by three loops connecting the D–E, B–C, F–G beta sheets on EDIII and the linker region connecting EDIII and EDI (Figure 2B and Table 1) [10],[18]. Investigators have also defined an epitope on EDIII recognized by MAbs that neutralize more than one DENV serotype [15],[27],[28]. This DENV sub complex epitope overlaps with the lateral ridge epitope but is centered at positions 305–308 (DENV3 numbering) on the A strand of EDIII (Figure 2B and Table 1). We compared the positions of known antibody epitopes and DENV3 informative sites on EDIII. The dengue type specific, lateral ridge epitope overlapped extensively with the informative sites on EDIII (Figure 2B, Table 1). This analysis supports the hypothesis that the EDIII lateral ridge epitope engaged by strongly neutralizing MAbs is not conserved between DENV3 strains.
To directly address if natural amino acid variation in DENV3 EDIII results in altered antibody binding and neutralization, we assembled and mapped a panel of DENV3 EDIII reactive mouse MAbs. All the antibodies selected for this study (8A1, 1H9, 14A4, 8A5, and 12C1) bound to EDIII of DENV3 (Figure 3). MAbs 8A1 and 1H9 were type specific as they only bound to EDIII from DENV3 (Figure 3) and only neutralized DENV3 (data not shown). 14A4 was a sub-complex specific antibody that bound strongly to DENV3 and weakly to DENV1 (Figure 3). 14A4 strongly neutralized DENV3 and weakly neutralized DENV1 (data not shown). 8A5 and 12C1 were non-neutralizing antibodies that cross reacted with EDIII from all 4 serotypes (Figure 3 and data not shown).
To map the binding sites of the MAbs, we expressed and purified 28 EDIII recombinant proteins with defined mutations. The positions to mutate were selected based on antibody mapping studies done with other flaviviruses [10],[15],[17],[27],[29],[30]. The binding of each antibody to wild type and mutant proteins was compared by ELISA (Table 2). MAb 1H9 is a type specific, neutralizing DENV3, EDIII reactive IgM antibody that has previously been shown to select for escape mutation at position 386 [23]. We observed a greater than 80% loss of binding of 1H9 when amino acids at positions 302, 304, 308, 310, 323, 325–330, 357, 358, 361, 380, 382 and 386 were mutated on EDIII (Table 2). Most of these mutations are on the lateral ridge of EDIII (Figure 4).
MAb 8A1 is a strongly neutralizing, type specific DENV3, EDIII reactive IgG antibody. With this antibody, we observed a greater than 80% loss of binding when amino acids at 301, 302, 304, 326–328, 330, 361, 380, 382 and 386 were mutated on EDIII (Table 2). As in the case of 1H9, most of these positions overlap with the EDIII lateral ridge epitope. However, 1H9 and 8A1 did not bind to identical epitopes because some mutations that influenced 1H9 had marginal to no effect on 8A1 (Table 2).
MAb 14A4 is a neutralizing EDIII reactive IgG antibody that cross-reacts with DENV3 and DENV1 (Figure 3). This DENV sub complex antibody bound poorly to recombinant proteins with mutations at position 308 (A strand), and positions 326 and 328 (B–C loop) (Table 2). These mutations are located at a similar position to a DENV EDIII sub complex epitope recently described in the literature [15],[27]. The sub complex epitope overlaps with the lateral ridge but is centered on the A strand of EDIII.
The DENV complex cross reactive MAbs (8A5, 12C1) bound to all the mutant proteins indicating these antibodies likely bind to a cross reactive epitope outside the lateral ridge region (Table 2). In Figure 4 we display the structure of DENV3 EDIII and the location of mutations that reduced the binding (>80%) of each MAb.
One concern with the above mapping studies was that some mutations might disrupt the overall folding of EDIII and non-specifically reduce antibody binding. To address this concern, we performed binding studies with a well characterized DENV subcomplex specific MAb 1A1D2, which binds to and neutralizes DENV1, 2 and 3 but not 4 [15]. The crystal structure of DENV2-EDIII-1A1-D2 Fab complex has been solved [31]. The 1A1-D2 MAb binds to a highly conformational epitope with a footprint that consists of the A strand, B strand, DE loop and G strand of EDIII [31]. When we compared the binding of 1A1-D2 to the panel of DENV3 EDIII mutants created for this study, a greater than 80% loss of binding was observed when DENV3 EDIII positions 304, 308, 310, 326, 328 and 330 were mutated (Table S1). DENV3 positions 304–310 are on the A strand and positions 326, 328 and 330 are on the BC loop which is adjacent to the B strand. The EDIII mutations at a distance from the known footprint of 1A1-D2 did not disrupt the highly conformational 1A1-D2 epitope indicating that the overall folding of EDIII was preserved in our mutants (Table S1).
Several amino acid positions (301, 302, 329, 380 and 386) on EDIII implicated in binding to MAbs 8A1 and 1H9 (Table 2) were also identified as informative sites that were not conserved between DENV3 genotypes (Figure 1). All these positions are located in close proximity to one another on the EDIII lateral ridge. To directly address if naturally occurring variation at these informative sites leads to altered antibody interactions, we compared the binding of MAbs 1H9, 8A1 and 14A4 to representative EDIII from each of the 4 genotypes of DENV3. MAbs 8A1 and 1H9 bound to genotypes I, II and III but not to EDIII from genotype IV (Figure 5). The DENV sub complex specific 14A4 antibody bound to EDIII from all 4 genotypes (Figure 5). Studies were also performed to compare the binding of these MAbs to purified viruses (Figure 6). As predicted from the recombinant EDIII binding experiments, 1H9 bound to DENV3 genotype I, II and III viruses but not to the genotype IV virus (Figure 6). Similarly, MAb 8A1 bound to genotypes I, II, and III but not to the genotype IV (Figure 6). MAb 14A4 bound to all 4 genotypes. These results indicate that naturally occurring amino acid variation on DENV3 EDIII influence the binding of type specific antibodies.
To verify that amino acid differences at the EDIII lateral ridge were responsible for MAb binding differences, further studies were conducted with MAb 8A1 and recombinant EDIII proteins. We systematically changed amino acids in the EDIII genotype IV construct to genotype II and defined the minimum number of changes required to restore the 8A1 epitope. In Figure 7 we depict the EDIII amino acid differences between the different genotypes. Simply making single amino acids changes at positions 301 or 302 did not restore binding. Some binding was regained when both 301 and 302 were changed from SG (genotype IV) to LN (genotype II) (Figure 7). Full binding was restored when positions 301, 302 and 380 were changed (Figure 7) indicating that these were the critical changes that led to the loss of binding of MAb 8A1 to DENV3 genotype IV. Residues 301, 302 and 380 are surface-exposed neighbors on the lateral ridge of EDIII, with residues 301 and 380 separated by approximately 4.7 angstroms. Thus, these three residues are likely to be a part of a single epitope bound by 8A1.
Experiments were conducted to compare the ability of EDIII MAbs to neutralize different genotypes of DENV3. MAbs 8A1 and 1H9 failed to neutralize DENV3 genotype IV (Table 3), which was expected since these antibodies did not bind to this virus (Figure 6). Surprisingly, even though we did not observe differences in the binding of 8A1 and 1H9 to genotype I, II and III viruses, we observed differences in the neutralization titers (Table 3). For example the neutralization titers for 8A1 were 10 fold different between genotype I and III viruses (Table 3). 1H9 displayed a 60 fold difference in the neutralization titer between genotype I and II viruses (Table 3). These results indicate that two mechanisms influence the ability of MAbs to neutralize virus infectivity. In the first, mutations which ablate binding also ablate neutralization. In the second, genetic differences between DENV3 strains that have little effect on in vitro binding can have significant biological effects on neutralization.
A long-held paradigm in flavivirus research has been that DENVs display little if any within intra-serotypic antigenic variation and this has been the basis for the development of current multivalent vaccines and immunotherapeutics [10],[18]. The main goal of the current study was to characterize the extent of envelope protein variation within the DENV3 serotype and to determine if this variation influenced antibody binding and neutralization. Sequence and structural analysis of the E protein indicated that 7% of the amino acids were variable between the four genotypes of DENV3, and most of the non-conserved residues were surface exposed and located at or proximal to known antibody binding sites. Particularly striking was the variation observed on the lateral ridge of domain III, which has previously been identified as the target of antibodies that strongly neutralize flaviviruses [10],[18]. Finally, we demonstrated that natural variation on EDIII influences the ability of MAbs to bind and neutralize DENV3. Our results reported here, together with the other published studies [8],[9] challenge the long held view that neutralizing antibody epitopes are conserved across DENV strains belonging to the same serotype.
Our results show that EDIII lateral ridge antibodies 8A1 and 1H9 bound to DENV3 genotypes I, II and III but not genotype IV indicating that naturally occurring mutations in EDIII can lead to a total loss of MAb binding. Even though MAbs 8A1 and 1H9 bound to DENV3 genotypes I, II and III with similar apparent affinity, we observed striking differences in the ability of the MAbs to neutralize these viruses. The neutralization titers were almost 10 fold different between viruses for 8A1 and 60 fold different for 1H9. Our results indicate that apparent affinity of antibody to virus immobilized on ELISA wells is not always predictive of the neutralization titer. There are amino acid differences on the EDIII lateral ridge of genotype I, II and III viruses (Figure 7) and these changes may lead to subtle changes in virus antibody interactions that are not detected in our ELISA binding assay. The flavivirus envelope proteins undergo low pH induced conformational changes during viral entry [32]. Some antibodies neutralize flaviviruses by binding to the virus in endosomes and blocking late steps in viral entry [29]. It is possible that antibody binding to the low pH conformation of the viral envelope might be a better predictor of neutralization potency than binding to the neutral pH conformation assessed here. Further studies with virions in different conformations, and an infectious clone of DENV3 to introduce targeted mutations are needed to dissect the mechanism underpinning the ability of EDIII lateral ridge antibodies to neutralize different genotypes of DENV3.
One potential problem with our studies is the possibility that some of the recombinant EDIII proteins used in the current study might be grossly misfolded and the binding differences might not be due to direct interactions between antibody and the altered amino acid. NMR and antibody binding assays have established that wild type EDIII expressed as an MBP fusion protein is properly folded [33]–[36]. When selecting sites to mutate, we primarily targeted surface exposed amino acids on loops because we did not want to disrupt the overall structure of EDIII. Moreover, in most cases mutations that led to the loss of binding of 8A1 or 1H9 preserved the binding sites of 14A4, 8A5 and 12C1 indicating that the proteins were not grossly misfolded (Table 2). Finally we probed the conformation of our EDIII mutants using a DENV sub complex antibody 1A1-D2 which binds to a highly conformational epitope on EDIII that has been mapped by X-ray crystallography [31]. As depicted in Table S1, of the 28 mutant proteins we created only 6 mutants failed to bind to this antibody (>80% loss of binding). The 6 mutants that failed to bind had mutations that were on or adjacent to the known footprint of 1A1-D2. Based on these results we are confident that the recombinant proteins used in the current study were not grossly misfolded. Nevertheless, we cannot completely rule out indirect or distance effects of some mutations on antibody binding and some of the mutations that reduced binding might not be in direct contact with antibody.
Several groups have focused on developing DENV vaccines based on recombinant EDIII [21],[37],[38]. Our results indicate that EDIII based vaccines need to be carefully evaluated. If people immunized with these antigens mainly develop neutralizing antibodies that bind to the lateral ridge epitope recognized by MAbs such as 8A1 and 1H9, then natural strain variation is likely to lead to vaccine failure. Alternatively, if EDIII vaccines stimulate antibodies to a conserved, neutralizing epitope such as the A strand epitope (recognized by 14A4), then the vaccine might be broadly protective across DENV3 strains.
Recently we reported that people exposed to natural DENV infections have low levels of EDIII reactive antibody several years after recovery from infection [36]. Given the low levels of EDIII reactive antibody in human immune sera, we were surprised by the extent of amino acid variation between EDIII from different DENV3 genotypes. It is plausible that interactions with cellular receptors and not antibody are behind the observed variability in EDIII. It is also plausible that EDIII reactive, neutralizing antibodies are abundant during early stages after infection and select for mutation in EDIII. Further studies are needed to better characterize human receptors and antibodies that interact with E protein and to assess how these interactions contribute to natural variation in DENV3 E protein.
|
10.1371/journal.pgen.1004750 | TIF-IA-Dependent Regulation of Ribosome Synthesis in Drosophila Muscle Is Required to Maintain Systemic Insulin Signaling and Larval Growth | The conserved TOR kinase signaling network links nutrient availability to cell, tissue and body growth in animals. One important growth-regulatory target of TOR signaling is ribosome biogenesis. Studies in yeast and mammalian cell culture have described how TOR controls rRNA synthesis—a limiting step in ribosome biogenesis—via the RNA Polymerase I transcription factor TIF-IA. However, the contribution of TOR-dependent ribosome synthesis to tissue and body growth in animals is less clear. Here we show in Drosophila larvae that ribosome synthesis in muscle is required non-autonomously to maintain normal body growth and development. We find that amino acid starvation and TOR inhibition lead to reduced levels of TIF-IA, and decreased rRNA synthesis in larval muscle. When we mimic this decrease in muscle ribosome synthesis using RNAi-mediated knockdown of TIF-IA, we observe delayed larval development and reduced body growth. This reduction in growth is caused by lowered systemic insulin signaling via two endocrine responses: reduced expression of Drosophila insulin-like peptides (dILPs) from the brain and increased expression of Imp-L2—a secreted factor that binds and inhibits dILP activity—from muscle. We also observed that maintaining TIF-IA levels in muscle could partially reverse the starvation-mediated suppression of systemic insulin signaling. Finally, we show that activation of TOR specifically in muscle can increase overall body size and this effect requires TIF-IA function. These data suggest that muscle ribosome synthesis functions as a nutrient-dependent checkpoint for overall body growth: in nutrient rich conditions, TOR is required to maintain levels of TIF-IA and ribosome synthesis to promote high levels of systemic insulin, but under conditions of starvation stress, reduced muscle ribosome synthesis triggers an endocrine response that limits systemic insulin signaling to restrict growth and maintain homeostasis.
| All animals need adequate nutrition to grow and develop. Studies in tissue culture and model organisms have identified the TOR kinase signaling pathway as a key nutrient-dependent regulator of growth. Under nutrient rich conditions, TOR kinase is active and stimulates metabolic processes that drive growth. Under nutrient poor conditions, TOR is inhibited and animals alter their metabolism to maintain homeostasis and survival. Here we use Drosophila larvae to identify a role for ribosome synthesis—a key metabolic process—in mediating nutrient and TOR effects on body growth. In particular, we show that ribosome synthesis specifically in larval muscle is necessary to maintain organismal growth. We find that inhibition of muscle ribosome synthesis leads to reduced systemic insulin-like growth factor signaling via two endocrine responses—decreased expression of brain derived Drosophila insulin-like peptides (dILPs) and increased expression of Imp-L2, an inhibitor of insulin signaling. As a result of these effects, body growth is reduced and larval development is delayed. These findings suggest that control of ribosome synthesis, and hence protein synthesis, in specific tissues can exert control on overall body growth.
| Nutrient availability is a critical determinant of cell, tissue and body growth in developing animals. Nearly two decades of research has identified the Target-of-Rapamycin (TOR) kinase signaling pathway as a major nutrient-responsive growth pathway in eukaryotes [1], [2]. TOR functions in two distinct complexes – TORC1 and TORC2 – and it is TOR kinase activity specifically within TORC1 that has been established as a growth driver. A complex intracellular signaling network activates TOR kinase activity within TORC1 in response to availability of extracellular nutrients such as amino acids and glucose. TORC1, in turn, stimulates many cell metabolic processes that drive growth and proliferation [2], [3]. In contrast, when nutrients are limiting, TORC1 activity is inhibited and cells switch their metabolism to promote homeostasis and survival during starvation conditions.
One important metabolic target of nutrient/TOR signaling in the control of growth is ribosome biogenesis [4]–[10]. A limiting step of ribosome synthesis is the RNA Polymerase (Pol I)-dependent transcription of ribosomal RNA (rRNA). Studies predominantly in yeast and mammalian cell culture have described mechanisms by which TOR promotes rRNA synthesis [4], [5], [7]–[13]. One target of TOR signaling emerging from these studies is the Pol I-specific transcription factor Transcription Initiation Factor-IA (TIF-IA). TIF-IA associates with Pol I and recruits it to rDNA genes to initiate transcription [5], [6], [9], [14]–[16]. This function of TIF-IA is stimulated by nutrient-dependent activation of TOR, and a handful of reports have proposed mechanisms involving TOR-dependent changes in TIF-IA phosphorylation, levels or localization to rDNA genes [9], [17]. These effects may also involve TOR functioning directly at nucleolar rDNA genes [13]. While these studies provide a molecular basis for understanding how nutrients and TOR control Pol I and rRNA synthesis in cells, the contribution of rRNA and ribosome synthesis to tissue and body growth in developing animals is not as clear.
Genetic studies in model organisms, most notably Drosophila, have provided most detail into how nutrient/TOR signaling controls tissue and body growth. During the four-day Drosophila larval period, animals increase in mass almost 200-fold [18]. This dramatic growth is nutrition-dependent and mostly occurs in non-dividing polyploid cells that make up the bulk of the larval organs. TOR signaling is central to this size control and functions by coupling dietary protein to growth [19]–[22]. Loss of TOR function in cells or tissues leads to a reduction in cell size or tissue mass, whereas TOR over-activation leads to increased cell and tissue growth [19]–[22].
TOR activity in specific tissues can also influence overall body size through non-autonomous endocrine or systemic effects [23], [24]. An example is the role of TOR in the larval fat body [25], [26]. When dietary proteins are abundant, amino acid uptake into fat body cells stimulates TOR activity. This triggers release of a fat-to-brain secreted signal that promotes the production and release of Drosophila insulin-like peptides (dILPs) from neurosecretory cells in the brain [25], [26]. These dILPs then circulate throughout the animal and promote growth in all larval tissues via a conserved insulin receptor/PI3K/Akt signaling pathway [27]. In contrast, when larvae are starved, TOR signaling in the fat body is suppressed leading to reduced circulating dILP levels, and decreased insulin signaling and growth. In this way, TOR activity in the fat body links nutrition to larval growth and development. TOR activity in larval muscle has also been reported to exert systemic effects to promote overall body growth and development [28]. This ability of TOR activity in specific tissues to control whole body metabolism and growth is an emerging theme in both mouse and fly genetic studies [23], [24], [29], [30], and emphasizes the importance of non-autonomous mechanisms in the control of body growth.
In this paper, we describe our ongoing work exploring the role for rRNA synthesis in controlling tissue and body growth in larvae. We find that the nutrient-dependent TOR pathway is required to maintain TIF-IA mRNA and protein levels in larval tissues, especially the muscle, during development. We also find that TIF-IA-dependent ribosome synthesis is required in muscle to maintain systemic insulin signaling and promote normal body growth and development, and loss of TIF-IA in muscle blocks the body growth-promoting effects of TOR signaling. This work emphasizes the importance of non-autonomous, tissue-specific effects of ribosome synthesis on endocrine signaling and body growth during development.
In previous work, we showed that the nutrient/TOR pathway controls rRNA synthesis in developing larvae and that TOR signaling promotes TIF-IA recruitment to rDNA genes [6]. Here, we examined whether TOR signaling may function by controlling TIF-IA levels. Deprivation of dietary protein leads to reduced TOR signaling and decreased rRNA synthesis in larvae. We found that under protein starvation conditions (induced by transferring larvae to a sucrose-only diet), TIF-IA protein levels were reduced compared to fully fed controls (Figure 1A). We also found that TIF-IA protein levels were also reduced in tor null mutant (torΔP) larvae compared to wild-type controls (Figure 1B). TOR can promote growth in part via its downstream effector kinase, ribosomal protein S6 kinase (S6K) [31]. However, we found that that TIF-IA protein levels were unaltered in s6k mutant (s6kL1) larvae compared to wild-type (Figure 1C). These results prompted us to examine TIF-IA mRNA levels. We found that both starved larvae and torΔP mutant larvae had reduced levels of both TIF-IA mRNA (Figure 1D, F) and pre-rRNA (Figure 1E, G) consistent with a reduction in synthesis of rRNA and hence ribosomes. Thus, during larval development nutrient/TOR signaling is required to maintain appropriate levels of TIF-IA mRNA and protein.
As well as controlling cell-autonomous growth, TOR activity in specific larval tissues is required for overall body growth in Drosophila. For example, reduced TOR signaling in larval muscle [28] and fat [25], [26] leads to reduced body growth. Given the importance of ribosome synthesis as an effector of TOR in the control of cell-autonomous growth, we examined whether TIF-IA-dependent ribosome synthesis could also exert non-autonomous effects on body growth. We first examined larval muscle. As with whole larvae, we found that protein starvation decreased both TIF-IA protein (Figure 2A) and mRNA (Figure 2B), and also pre-rRNA (Figure 2C) in larval muscle. To explore the consequence of this reduction in TIF-IA levels, we examined the effects of RNAi-mediated knockdown of TIF-IA in muscle, using a UAS-TIF-IA inverted repeat (IR) line. Ubiquitous expression of this TIF-IA IR line in larvae using the daughterless (da)-GAL4 driver (da>TIFIA-IR) phenocopied tif-ia mutants, and led to reduced TIF-IA protein levels (Figure S1B) and larval growth arrest (Figure S1A). Both of these effects were fully reversed by co-expression of a UAS-TIF-IA transgene (Figure S1A), confirming the specificity of the UAS-TIF-IA IR line. We then used the UAS-TIF-IA IR line to knock down TIF-IA specifically in muscle (using the dMef2-GAL4 driver – Figure S2). We found that RNAi-mediated knockdown of TIF-IA muscle mimicked the decrease in both TIF-IA mRNA (Figure 2D) and pre-rRNA (Figure 2E) levels following starvation. When we monitored larval growth and development, we observed that dMef2>TIF-IA IR larvae were smaller than age-matched control larvae (Figure 2G). Moreover, dMef2>TIF-IA IR larvae were significantly delayed in pupal development with respect to control (dMef2>+) larvae (Figure 2F), and only approximately 20% of dMef2>TIF-IA IR larvae formed pupae. These dMef2>TIF-IA IR pupae were malformed compared to control (dMef2>+) pupae (Figure 2H). We examined feeding by transferring dMef2>+ (control) and dMef2>TIF-IA IR larvae onto yeast paste colored with blue food dye. After 4 hours, we observed that both the control and TIF-IA IR larvae contained blue food in their guts (Figure S3), suggesting that knockdown of TIF-IA in larval muscle did not impair feeding. Together, these findings indicated that TIF-IA-dependent ribosome synthesis in muscle is required to maintain normal body growth and development.
We also examined the organismal effects of TIF-IA knockdown in other tissues. We used two fat body GAL4 drivers (r4-GAL4 and ppl-GAL4) to express UAS-TIF-IA IR during larval development. We found that r4>TIF-IA IR larvae showed a modest, although statistically significant delay in both developmental timing - time from larval hatching to pupation (Figure 3A) - and growth (Figure 3B), but showed no significant change in body size compared to control (r4>+) animals (Figure 3C). Co-overexpression of UAS-Tsc1 and UAS-Tsc2 - negative regulators of TORC1 – using r4-GAL4 led to marked reduction in body growth, thus confirming the effectiveness of the driver (Figure S4). We also found that ppl>TIF-IA IR larvae showed no significant difference in developmental timing (Figure 3D) or final body size (Figure 3E) compared to controls (ppl>+). We also examined the effects of TIF-IA knockdown in the larval lymph gland and hemocytes using two different drivers, hemolectin (hml)-GAL4 and peroxidasin (pxn)-GAL4. In both cases, we observed no statistically significant decrease in larval development (Figure 3F, G). In fact, larval development was modestly, although significantly, accelerated in hml>TIF-IA IR larvae.
TOR activity in muscle is required for normal larval growth and development [28]. We confirmed this finding by inhibiting TOR in the muscle by two different methods, expression of a dominant negative form of TOR in muscle (dMef2>TORTED) [32] and co-overexpression of UAS-Tsc1 and UAS-Tsc2 - negative regulators of TORC1 - in muscle (dMef2>Tsc1,Tsc2). We measured pupal volume, as an indicator of final body size. Our data showed that in both cases, inhibition of TOR in larval muscle reduced pupal volume (Figure 4A, B). Amino acid availability is an important activator of TOR kinase signaling, and we also found that knockdown of the amino acid transporter slimfast (using a UAS-slifAnti antisense [25]), in the larval muscle led to a significant reduction in pupal volume (Figure 4C). Finally, we also examined whether over-activation of TOR in muscle was sufficient to drive systemic growth. We found that overexpression of Ras homolog enriched in brain (Rheb), an upstream activator specifically of TORC1, in muscle (dMef2>Rheb) was sufficient to increase pupal volume compared to control (dMef2>+) pupae (Figure 4D). Together, these findings suggest that TOR activity in muscle is both necessary and sufficient to control overall systemic growth.
We next examined whether TIF-IA function was required for these muscle effects of TOR. As described above, overexpression of Rheb in muscle led to increased body size, as indicated by larger larvae (Figure 4E) and increased pupal volume (Figure 4F), while RNAi-mediated knockdown of TIF-IA showed the opposite effects. We found that co-expression of UAS-TIF-IA IR (dMef2>Rheb;TIF-IA IR) phenocopied dMef2>TIF-IA IR animals and abrogated the Rheb induced increase in body size. We quantified the pupal volume and found that reducing TIF-IA in muscle reduced the Rheb induced increase in pupal volume (Figure 4G). Overall, these data indicated that TIF-IA activity in muscle is required for TOR signaling to drive systemic growth.
The insulin pathway is the major endocrine regulator of body growth in larvae. Under nutrient-rich conditions, several dILPs are expressed and released into the larval hemolymph [33]. These dILPs then bind to a single insulin receptor in target cells and promote growth [26]. In contrast, starvation leads to reduced systemic insulin signaling and decreased growth. We therefore explored whether the growth inhibitory effects of muscle-specific TIF-IA knockdown were due to reduced systemic insulin signaling. Under nutrient rich conditions, high level of insulin signaling leads to activation of Akt kinase and phosphorylation and nuclear exclusion of the FOXO transcription factor. But when insulin signaling is reduced, FOXO relocalizes to the nucleus and activates target genes such as eIF4E-Binding Protein (4EBP). Therefore, changes in FOXO nuclear localization and transcriptional activity serve as a reliable ‘read-out’ of insulin signaling [34]–[36]. As previously reported, we found that FOXO was excluded from nuclei in fat body cells from fed larvae (Figure 5A), but showed strong nuclear accumulation in fat body cells from starved larvae (Figure 5B). When we knocked-down TIF-IA levels in muscle (dMef2>TIF-IA IR), FOXO showed strong, statistically significant nuclear accumulation in fat body cells (Figure 5C, D). We next measured the levels of 4EBP, a FOXO target gene, and found that dMef2>TIF-IA IR larvae had increased 4EBP mRNA levels with respect to control (dMef2>+) larvae (Figure 5E). Finally we measured the examined levels of phosphorylated Akt – the kinase downstream of insulin signaling that is responsible for phosphorylation and inhibition of FOXO. Using western blotting with an anti-phospho Akt (Ser505) antibody, we found that dMef2>TIF-IA IR had markedly reduced levels of phospho Akt compared to control (dMef2>+) larvae (Figure 5F). Levels of total Akt were also lower, but much less so than the suppression in levels of phosphorylated Akt. Together these data suggest that TIF-IA knockdown in muscle leads to reduction in systemic insulin signaling.
An important source of dILPs is a cluster of neurosecretory cells in the larval brain [33], [37]. These cells secrete three dILPs (2, 3 and 5), and expression and/or release of these dILPs are suppressed upon protein starvation [26]. Moreover, loss of these neurons leads to slow growing and small larvae [33], [37], [38]. We found that dMef2>TIF-IA IR larvae had reduced dILP3 and dILP5 mRNA levels compared to control (dMef2>+) larvae, while dILP2 mRNA levels were unaltered (Figure 5G). Previous studies showed that nutrient-deprivation leads to reduced dILP2 secretion and hence increased retention in the neurosecretory cells [26]. This retention can be easily visualized by staining with anti dILP2 antibodies. Using, this approach we observed an increase in dILP2 staining in the neurons of dMef2>TIF-IA IR larvae compared to control larvae (Figure 5H, I, J). Together, these data suggest that one mechanism by which reduced TIF-IA activity in muscle suppresses peripheral insulin signaling is by reduced expression and release of brain-derived dILPs.
In addition to the dILPs, other secreted factors can influence insulin signaling in Drosophila. One factor is Imaginal morphogenesis protein-L2 (Imp-L2), which is the Drosophila homolog of insulin-like growth factor binding protein 7 (IGFBP7) [39]. Imp-L2 can bind to dILPs and inhibit their ability to signal through the insulin receptor [39]–[41]. Moreover, a recent report showed that mitochondrial perturbation in adult muscle leads to increased Imp-L2 expression and subsequent suppression of systemic insulin signaling [42], [43]. We found that dMef2>TIF-IA IR larvae had upregulated Imp-L2 mRNA levels in their muscle compared to control (dMef2>+) larvae (Figure 5K). These data suggest that upregulation of Imp-L2 may provide another mechanism by which perturbation of TIF-IA in muscle suppresses systemic insulin signaling. Indeed, we found that overexpression of Imp-L2 in the muscle led to delayed larval development and reduced pupal size (Figure S5).
Our data suggest that TIF-IA function in muscle is required to maintain systemic insulin signaling in fed animals. We next examined whether TIF-IA-mediated ribosome synthesis in muscle may provide one mechanism to couple nutrient availability to systemic insulin signaling. We overexpressed a UAS-TIF-IA transgene in muscle (dMef2>TIF-IA) and observed a very slight, but statistically significant acceleration in development compared to (dMef2>+) larvae (Figure S6A), although final body size was not affected (Figure S6B). Similar effects were observed with a second, independent UAS-TIF-IA transgene (Figure S6C, S6D). We then examined effects of muscle TIF-IA overexpression in starved animals. When larvae are deprived of dietary protein, insulin signaling is reduced leading to upregulated levels of FOXO target genes such as 4EBP and InR, an effect we observed here following 24 hr starvation. However, when we overexpressed TIF-IA in muscle (dMef2>TIF-IA), the starvation-induced increase in both 4EBP and InR mRNA was partially reversed compared to control (dMef2>+) larvae (Figure 6A, B). This result suggests that TIF-IA function in muscle can, in part, couple nutrient availability to systemic insulin signaling.
The findings presented here suggest that TIF-IA function in muscle is required for normal nutrient-dependent systemic insulin signaling and growth. Hence, upon knockdown of TIF-IA in muscle, we saw reduced growth and delayed development. To further implicate a role for reduced insulin signaling in these effects, we tested whether restoring insulin signaling to some degree could have any effect on the phenotypes we observed. To achieve this we examined partial loss of negative regulators of insulin signaling. We first tested the effects of reducing foxo gene dosage. We found that the decrease in larval growth seen in dMef2>TIF-IA IR larvae was partially reversed in larvae that were heterozygous for a loss-of-function mutation in foxo (foxo25) (Figure 7A). We next examined the effects of reducing the levels of Imp-L2, whose expression was increased in dMef2>TIF-IA IR larval muscle. We found that co-expression of a UAS-Imp-L2 inverted repeat (IR) line with the UAS-TIF-IA IR in muscle, also partially reversed the growth defects seen with expression of UAS-TIF-IA IR alone (Figure 7B). Loss of one copy of foxo (foxo25/+) alone or expression of UAS-Imp-L2 IR alone in the muscle had no effects on larval size (Figure S7). When we measured developmental timing, we also saw that both the delayed larval development and reduced numbers of pupating larvae seen in dMef2>TIF-IA IR larvae were partially reversed in larvae that either were heterozygous for foxo25, or which co-expressed UAS-Imp-L2 IR in the muscle (Figure 7C). These experiments provide genetic evidence that muscle TIF-IA function is required for normal larval growth and development at least in part by maintaining systemic insulin signaling.
The major finding of our work is that under nutrient-rich conditions TIF-IA-dependent regulation of muscle ribosome synthesis is required to maintain systemic insulin signaling and body growth.
Work in yeast, mammalian cell culture and Drosophila indicates that TIF-IA links nutrient availability and TOR signaling to rRNA synthesis [4]–[10]. Here we show that in growing tissues in vivo one mechanism by which nutrient/TOR signaling functions is through maintaining TIF-IA levels. Recent studies in yeast also showed TIF-IA levels were reduced following pharmacological inhibition of TOR [17]. Moreover, in previous work, we showed that maintaining high levels of TIF-IA expression could reverse the decrease in rRNA synthesis caused by amino acid starvation in Drosophila larvae [6]. Hence, control of TIF-IA levels represents one mechanism by which nutrient availability and TOR signaling can control the synthesis of rRNA. TOR has also been reported to indirectly control site-specific phosphorylation of TIF-IA, and this phosphorylation modulates TIF-IA nucleolar localization [9]. Hence, TOR may impact TIF-IA function in several ways.
When we mimicked the starvation induced decrease in muscle TIF-IA mRNA levels by RNAi-mediated knockdown, we observed that larvae were slower growing and failed to develop. This phenotype was not simply due to a gross motor defect, since the larvae were able to crawl normally and ingest food. Studies from Demontis and Perrimon [28] describe a similar reduced growth phenotype following inhibition of TOR signaling in larval muscle. Here, we extended this work to show that increased TOR in muscle leads to a larger overall body size, and that this effect required intact TIF-IA function.
Our data implicate changes in insulin signaling as underlying the effects of TIF-IA-dependent ribosome synthesis in muscle on overall body growth and development. Our findings also suggest that the ability of dietary nutrients to stimulate and maintain systemic insulin rely, in part, on maintaining TIF-IA levels and function in muscle. Muscle TIF-IA appeared to control insulin signaling by at least two mechanisms. First, we saw that expression of brain-derived dILPs required normal muscle TIF-IA function. The expression and release of dILPs (2,3 and 5) from a cluster of neurosecretory cells [33], [37] in the brain is regulated by signals from other tissues. Hence, the changes in systemic insulin signaling that we saw following inhibition of TIF-IA in muscle could be explained by a role for muscle-derived secreted factors (often termed myokines). In mammals, muscle has been shown to secrete many factors, including a host of cytokines, and secretion of these factors is often controlled by nutrients [44]–[50]. In Drosophila, the full complement of factors secreted from muscle is not clear [49]. Nevertheless, one or more secreted factors could potentially signal to the brain to promote dILP release. Indeed, a recent study showed that suppression of ribosome synthesis by overexpression of Mnt in adult muscle led to release of myoglianin, a myostatin-like myokine, which induced remote effects on fat body function [51]. Also, activin signaling in adult muscle can remotely control dILP release and systemic insulin signaling [52]. Second, we saw that knockdown of TIF-IA in muscle led to an increase in expression of Imp-L2, a secreted protein that functions to suppress insulin signaling [39]. A recent paper showed that perturbation of mitochondrial function in Drosophila muscle can also lead to upregulation of Imp-L2 expression [42]. Together with our data, this finding suggests that upregulation of Imp-L2 may be a common response triggered by metabolic stress in muscle cells. Importantly, we were able to partially rescue both the reduced growth and delayed development seen with muscle knockdown of TIIF-IA by either loss of one copy of foxo or RNAi-mediated knockdown of muscle Imp-L2. In both cases, the rescue was partial probably because neither genetic manipulation would be predicted to completely restore systemic insulin signaling. Nevertheless, the findings provide further support for our model that muscle-specific ribosome synthesis can control systemic insulin signaling and body growth.
A previous report described how inhibition of PI3K/TOR signaling in muscle led to both reduced muscle cell size and a non-autonomous reduction in size of other tissues and overall body size [28]. These non-autonomous effects were proposed to be mediated through altered endocrine signaling from the muscle to other tissues, although it is unclear whether this occurred solely as a result of reduced muscle cell size, or whether it reflects a cell size-independent role for TOR in controlling the endocrine function of muscle. Our findings here suggest that altered insulin signaling is one important endocrine response that links changes in muscle ribosome synthesis to altered physiology and growth in other tissues, although as with the effects of TOR we cannot discern whether this occurs only due to reduced muscle cell size. Interestingly we showed that inhibition of TIF-IA in the fat body had only a weak non-autonomous effect on body growth, although TIF-IA knockdown can limit fat cell size and ploidy [6]. Thus the mechanisms that couple TIF-IA and ribosome synthesis in muscle to the endocrine control of systemic insulin may not operate in the fat body. Ultimately, it is likely that the role for TIF-IA and ribosome synthesis in controlling overall body growth depends on a combination of cell-autonomous and non-autonomous influences. For example, inhibition of ribosome synthesis in the prothoracic gland was shown to extend larval development by altering endocrine ecdysone hormone signaling [53].
Muscle is a metabolically active tissue that probably has a high demand for continued ribosome biogenesis and protein synthesis to maintain autonomous growth. Our studies suggest that muscle ribosome synthesis may also act as a checkpoint for overall body growth. If muscle ribosome synthesis is perturbed (e.g. by nutrient deprivation), this may cause muscle cells to trigger a suppression of systemic insulin signaling to limit body growth. In using ribosome synthesis as a checkpoint for controlling systemic insulin, muscle cells may simply sense and respond to general changes in bulk translation. Alternatively, altered translation of a select subset of mRNAs may influence either Imp-L2 expression or the ability of muscle to remotely control brain dILP expression. In either case, our findings suggest that larval muscle is also an important nutrient-sensing tissue, in addition to the fat body, that can control systemic insulin signaling via endocrine signaling. The endocrine mechanisms by which either fat or muscle control systemic insulin signaling are nor clear and may be different in both cases. However, it seems that both tissues rely on protein synthesis, although perhaps through different mechanisms. Our data suggest that control of rRNA synthesis is an important limiting step in muscle, while previous work suggests that regulation of tRNA synthesis and signaling via Myc is important in fat [36], [54], [55].
The ‘checkpoint’ response to perturbation of muscle ribosome synthesis may be important for controlling not just growth, but also other organismal responses. For example, upon starvation or other environmental stressors, a reduction of muscle TIF-IA and ribosome synthesis may function to suppress systemic insulin signaling to alter whole body metabolism in order to maintain animal survival under adverse conditions. Reducing insulin signaling has been well described as mediator of stress resistance and extended lifespan in many animals including Drosophila, C. elegans and mice [56]–[58]. Indeed, a recent report showed that elevated Imp-L2 in Drosophila muscle increased adult lifespan [42], [43]. Also, overexpression of 4E-BP, a translational repressor, in muscle [59] or in whole organism [60] leads to stress resistance, and extended lifespan. Thus control of muscle protein synthesis, possibly by regulating ribosome biogenesis, may be a common mechanism to control stress responses and lifespan by regulating whole-body insulin signaling.
All stocks and crosses were raised at 25°C and maintained on a media containing 100 g Drosophila Type II agar, 1200 g cornmeal, 770 g Torula yeast, 450 g sugar, 1240 g D-glucose, 160 ml acid mixture of propionic acid and phosphoric acid per 20 L of water. The following fly stocks were used: w1118; yw; UAS-TIF-IA; UAS-TIF-IA IR (v20334, Vienna Drosophila RNAi Center, VDRC); UAS-Tsc1, UAS-Tsc2; torΔP/CyO; s6kL1/TM6B; UAS-Rheb; UAS-slifAnti; UAS-TORTED; UAS-GFP; UAS-Imp-L2 IR (15009-R3, NIG, Japan); foxo25/TM6B, dMef2-GAL4; da-GAL4; r4-GAL4; ppl-GAL4; hml-GAL4; pxn-GAL4. For all GAL4/UAS experiments, homozygous GAL4 lines were crossed to the relevant UAS line(s) and the larval progeny were analyzed. Control animals were obtained by crossing the relevant homozygous GAL4 line to either w1118; +; + or yw; +; +, depending on the genetic background of the particular experimental UAS transgene line.
Adult flies were allowed to lay eggs on grape juice agar plates supplemented with yeast paste for 4 hours (hr) at 25°C. 24 hr after egg laying (AEL) all hatched larvae were transferred to food vials with a thin brush, in groups of 45–50 larvae/vial and allowed to develop.
For all experiments, whole larvae were starved by floating on sterile 20% sucrose in 1× Phosphate Buffered Saline (PBS) at 72 hr AEL for indicated times. Subsequently, larvae were collected and processed as per experimental requirements. Fed larvae were collected at 72 hr AEL.
Whole larval or larval muscle tissue extracts were prepared by lysing 72 hr AEL larvae in 4× protein sample buffer (240 mM Tris-HCl pH 6.8, 8% SDS, 5%β-mercaptoethanol, 40% glycerol, 0.04% bromophenol blue) with a motorized pestle, boiling for 4 minutes at 95°C and immediately loading the samples onto a SDS-PAGE gel. Immunoblotting was performed as previously described [54]. Antibodies used were βtubulin (E7, Drosophila Studies Hybridoma Bank), phospho-Drosophila Akt Ser505 (Cell Signaling Technology, 4054) and Akt (Cell Signaling Technology, 9272). Affinity-purified antibodies were generated against TIF-IA was raised by immunizing rabbits using the synthetic peptide CIVDKRPKNFDLSKSQEFDKQ, corresponding to residues 585–604 (Anaspec Inc.).
Whole larval or larval muscle tissues were isolated at definite time points AEL (as indicated in the figure legends). Total RNA was extracted using TRIzol according to the manufacturer's instructions (Invitrogen; 15596-018). RNA samples were DNase treated as per manufacturer's protocol (Ambion; 2238G). The DNase treated RNA was reverse transcribed by Superscript II to make cDNA. This cDNA was used as a template to perform qRT-PCR reactions (BioRad Laboratories, MyIQ PCR machine using SYBR Green PCR mix) using specific primer pairs (sequences available upon request). Pre-rRNA levels were measured by using primer pairs against the internal transcribed spacer (ITS) region of 45S pre-rRNA transcript. qPCR data were normalized to β tubulin mRNA, whose levels we found were essentially unchanged across all the experimental conditions. The exception was the qPCR analyses of tor mutants, where values were corrected for actin mRNA levels. For each experiment, a minimum of 3 groups of 5–8 larvae was collected. Each experiment was independently repeated a minimum of 3 times.
Larvae were collected at 24 hr AEL and placed in food vials in equal numbers per vial (with a maximum density of 50 larvae per vial). The number of pupae in vials was counted every 24 hr. For each genotype, minimum of 3 replicates were used to calculate the mean percentage of pupae per timepoint.
Pupal volume was calculated as previously described [55].
Larval and pupal images were obtained using a Zeiss Stereo Discovery V8 microscope using Axiovision software. Microscopy and image capture was performed at room temperature and captured images were processed using Photoshop CS5 (Adobe). For each experiment all larval and pupal images were captured using identical magnifications. Final figures were generated from these by cropping individual larvae and then simply rotating images to orient them in the same direction, without altering size or scale. These images were then assembled on a single black canvas in Photoshop. Larval sizes were assessed by using Photoshop to measure larval body areas from these microscope images. Tissue images and Differential Interference Contrast (DIC) images were captured by taking serial Z-stacks using the same magnification and time of exposure.
Larvae were inverted using fine forceps in 1× PBS at particular time points (as indicated in the figure legends). Inverted larvae were fixed in 8% paraformaldehyde for 40 minutes, washed in 1× PBS-0.1% TritonX (PBST), blocked for 2 hr at room temperature in 1× PBST with 5% fetal bovine serum (FBS). Larvae were incubated overnight with primary antibody at 4°C, washed several times with 1× PBST and incubated with secondary antibody (1∶4000) for 2 hours, at room temperature. After few washes, fat bodies were isolated from these larvae using fine forceps and mounted on glass slides with cover slips using Vectashield (Vector Laboratories Inc., CA) mounting media. Primary antibodies used were rabbit anti-FOXO (from Marc Tatar) and rabbit anti-dILP2. Alexa Fluor 488 and 568 (Invitrogen) were used as secondary antibodies. Hoechst 33342 (Invitrogen) was used to stain nuclei. dILP2 immunostaining of larval brains was performed as described [54].
For all experiments, error bars represent standard error of mean (SEM). P values were computed by Student's t-test, using Microsoft Excel or Analysis of Variance (ANOVA) followed by Tukey's post-hoc test, using GraphPad prism (version 6). For developmental timing experiments, mean time to pupation was computed using Mann-Whitney U test using GraphPad prism (version 6). P<0.05 was considered to be statistically significant, as indicated by asterisk (*) or as indicated in the figure legend.
|
10.1371/journal.pcbi.1004223 | Bud-Localization of CLB2 mRNA Can Constitute a Growth Rate Dependent Daughter Sizer | Maintenance of cellular size is a fundamental systems level process that requires balancing of cell growth with proliferation. This is achieved via the cell division cycle, which is driven by the sequential accumulation and destruction of cyclins. The regulatory network around these cyclins, particularly in G1, has been interpreted as a size control network in budding yeast, and cell size as being decisive for the START transition. However, it is not clear why disruptions in the G1 network may lead to altered size rather than loss of size control, or why the S-G2-M duration also depends on nutrients. With a mathematical population model comprised of individually growing cells, we show that cyclin translation would suffice to explain the observed growth rate dependence of cell volume at START. Moreover, we assess the impact of the observed bud-localisation of the G2 cyclin CLB2 mRNA, and find that localised cyclin translation could provide an efficient mechanism for measuring the biosynthetic capacity in specific compartments: The mother in G1, and the growing bud in G2. Hence, iteration of the same principle can ensure that the mother cell is strong enough to grow a bud, and that the bud is strong enough for independent life. Cell sizes emerge in the model, which predicts that a single CDK-cyclin pair per growth phase suffices for size control in budding yeast, despite the necessity of the cell cycle network around the cyclins to integrate other cues. Size control seems to be exerted twice, where the G2/M control affects bud size through bud-localized translation of CLB2 mRNA, explaining the dependence of the S-G2-M duration on nutrients. Taken together, our findings suggest that cell size is an emergent rather than a regulatory property of the network linking growth and proliferation.
| The size between different organisms ranges considerably, yet, the size of the individuals and even the same types of cells within the individuals are remarkably constant. Cell size emerges from the balance between how fast the cell grows and the frequency with which it divides. This system level coordination of growth and division is universal across species and is required to ensure faithful duplication and genetically intact offspring. We have devised a computational model for the interplay of growth and division in the premier model organism, Baker’s yeast, to test the fundamental architecture of this coupling and to assess the role that cell size itself can play in it. In contrast to traditional theories that assume a yet-to-be-determined cell size sensor, our model relies on a single mechanism, effectively measuring the cell’s translational capacity, applied twice at different stages of the cell’s life-cycle to explain this coupling. In our model, a growth condition specific cell size emerges, as has been found in experiments. Our analysis shows how the nature of the two linked properties growth and proliferation can shape eukaryotic cells and explain cell size as an emergent rather than regulatory property of this process.
| Cell size is a fundamental systems level property of life. It emerges as a combination of the cell cycle, controlling the orderly orchestration of duplication and division, and the individual growth rate, reflecting extra- and intracellular physiological conditions. The cell cycle and the growth rate are coupled, such that proliferation and growth are balanced, avoiding abnormally large or small cells. Understanding the coupling is of particular interest for two reasons. First, the cell cycle as well as cellular growth are two fundamental properties that can be found in nearly all forms of life. Second, decoupling of the two can have disastrous consequences for an organism, e.g. deterioration of cell size.
The unicellular eukaryote Saccharomyces cerevisiae can be observed to grow to a ‘critical cell size’ in the G1 phase before committing to passage through the cell cycle [1]. The commitment is called START in S. cerevisiae and constitutes the transcriptional activation of more than 200 genes by the transcription factor complexes SBF and MBF [2]. This triggers the onset of downstream events, such as budding and DNA replication. SBF/MBF activity is controlled by the G1 network, which involves the cyclin dependent kinase (CDK) Cdc28, its activating subunits the G1 cyclins Cln1/2/3 and the transcriptional repressor Whi5 (reviewed in [3]). The most upstream undisputed activator of START is Cln3. Cln3 binds to and activates the CDK to phosphorylate Whi5, which relieves the repression of SBF/MBF. The START transition is triggered when a critical activity of the CDK is reached [4]. Beyond the critical level, CDK activity stabilises through positive feedback involving Cln1/2 [5, 6]. The core network architecture with the competition between the active CDK and the transcriptional repressor is analogous to the Restriction Point, which is the equivalent of START in mammalian cells [7]. The nature of the mechanism within the START network that ties growth and proliferation together remains unknown.
Size control must be as old as the cell cycle itself. It is conserved across species over a huge range of cell sizes and shapes, and it is well established that size control can occur in cell cycle phases other than G1 [8]. Recent evidence strongly suggests that also in budding yeast size control is likely to be exerted outside of G1 [9, 10]. The fission yeast Schizosaccharomyces pombe has a size control checkpoint at the G2/M boundary and many of its components are conserved in budding yeast [11, 12]. The observation that the budded phase duration responds to growth media and the high degree of conservation between the two yeasts prompts the question, whether a size control mechanism guards mitotic entry in budding yeast as well [13–16]. Unfortunately, size control at the G2/M transition is less well understood in budding yeast [8, 9, 17–19].
It is well known, however, that S. cerevisiae can arrest its cell cycle at the G2/M boundary through activation of the so called morphogenesis checkpoint [20]. The kinase responsible for mitotic entry Clb2-Cdc28 is inhibited through phosphorylation at the Tyr19 residue by Swe1 [20, 21]. Re-activation of Clb2-Cdc28 requires removal of the inhibitory phosphate by the mitotic inducer homolog Mih1 phosphatase [22]. The exact property that is monitored in budding yeast remains a matter of debate, but it has been argued that the checkpoint responds to perturbations of the actin cytoskeleton or even bud growth [9, 17, 18]. Indeed, recent work suggests that polarised exocytosis may be required to pass through the checkpoint [9]. This hypothesis directly connects membrane growth at the bud site to cell cycle progression suggesting a growth dependent size control checkpoint for mitotic entry in budding yeast [9].
In the regulation of both the START and the G2/M transitions, a master CDK is balanced against an opposing regulator that must be overcome to initiate crucial cellular events, like DNA replication or cell division [3]. The master CDK is activated in G1 and G2 by the accumulation of cyclins and this activation is antagonised by CDK inhibitors and rapid degradation of the cyclins to form elaborate molecular switches [23, 24]. Accordingly, the transitions occur in a switch like manner, when the time (or size) is right. It is not exactly clear how growth can flip a transition switch, but one theory for size control proposes the level of an unstable cell cycle regulator as a gating device to measure the growth capacity of the cell [25, 26]. The growth or biosynthetic capacity of the cell determines the growth rate and the unstable regulator is presumed to be one of the G1 cyclins, most likely Cln3 [8]. Cln3 levels are heavily influenced by the available nutrients and Cln3 translation is slowed down in conditions when fewer ribosomes are available [27–29]. The G1 cyclins and other components of the START network modulate cell size at START [15, 30]. Additionally, many other components influence cell size, especially those with a functional link to the cellular growth machinery, like Sfp1 or Sch9 [31]. Recently, it was shown that cell size at START is set as a function of the individual growth rate of a cell and that START network components modulate the strength of this correlation, arguing for the existence of a growth rate dependent cell sizer in G1 [15].
Both in vivo and in silico analysis suggested that G1 is not the only size control phase during the cell division cycle [9, 10, 17, 32]. Also, if G1 were the sole size control phase, the question would remain how budding yeast can maintain size homeostasis and control in case the mechanisms in G1 are impaired, as e.g. through CLN3 overexpression. CLN3 overexpression mutants display a substantial reduction in G1 length and a small cell size (whi phenotype) [33, 34]. Since Cln3 is an upstream activator of START, its overexpression can intuitively explain the shortened G1 phase and the concomitant reduction in cell size. Counterintuitively, the generation time remains largely unchanged, arguing for the existence of another control point further downstream in the cell cycle to compensate for the reduced time in G1 [33, 34].
The obvious candidate for this would be the G2/M transition. Here, the mitotic cyclins Clb1/2 are responsible for CDK activity to trigger mitotic entry [4]. Intriguingly, CLB2 mRNA (mCLB2) accumulates in the bud, while the Clb2 protein is distributed throughout the cell [35, 36]. If the cell localises its CLB2 transcripts to the bud it is likely that translation of the transcript is also a local phenomenon. In principle, active transport of the mRNA leading to localised translation of mCLB2 could serve to measure the biosynthetic capacity of the bud, to form a bud (daughter) sizer in G2. Once the bud is strong enough for independent life, measured through production of sufficient Clb2, the cell enters mitosis. It is tempting to speculate that if Cln3 is a cell sizer, then Clb2 might be a bud sizer.
Here, we approach the elusive problem of size control from a theoretical angle, and use a number of mathematical models for rigorous testing of different control concepts. Our approach is to model single cells that are capable of growth and division, and grow them in in silico cell cultures. Through inheritance, we let the culture evolve over time, generating fully traceable cell populations for analysis [32]. In the model, growth and proliferation is integrated through production of a critical unstable cell cycle regulator as function of the biosynthetic capacity of the cell. Using this approach, we formally establish that the accumulation of cyclins to a threshold provides the necessary prerequisites for size control and homeostasis. Since it is unlikely that size control occurs in G1 alone, we use models that include additional size control at the G2/M boundary. We find that we need to take into account the intriguing, and to our knowledge unexplained, fact that mCLB2 is localised to the growing bud of yeast cells to fully explain experimental data [35]. Bud-localised translation of the major cyclin that activates the master CDK for mitotic entry can, in theory, constitute a growth rate dependent sizer for the bud. With the suggested mechanism we (i) offer a functional explanation for the mCLB2 transport into the bud, (ii) elucidate the prolongation of the budded phase in response to poor nutrients or CLN3 overexpression, and (iii) propose a unifying model for integration of growth and division in the G1 and G2 phases of budding yeast.
We present here an extended version of a minimal eukaryotic cell model that is capable of growth and division (Fig 1A) [32]. In the model, two types of biomass define growth: structural and internal biomass. Structural biomass represents all those components that are destined for the cell wall or membrane. Internal biomass are soluble agents within the cell, describing the cell’s capacity to metabolise nutrients and build new macromolecules (biosynthetic capacity). Structural and internal biomass accumulate with replicative age, leading to an increase in size over generations as observed experimentally (Fig 1B) [37, 38]. Budding yeast grows and divides asymmetrically and, therefore, growth of the mother and the bud is considered separately [39]. The volume trajectory for a single cell is biphasic with altered growth rates dependent on cell cycle stage as observed experimentally as well (Fig 1B) [15, 40–42]. A rudimentary version of the cell cycle machinery is included, with one proxy for the G1 cyclins (Cln) and one proxy for the mitotic cyclins (Clb) that determine CDK activity (Fig 1B) [43]. Transcription of cyclins is considered stochastic and restricted to a distinct cell cycle phase (CLN in G1, CLB in G2—Fig 1B). Growth and division are coupled via production of Cln in G1 and Clb in G2, as a function of the biosynthetic capacity of the cell. This means that the translation of cyclin mRNAs is dependent on the internal biomass (see also Materials and Methods). At division, two cells emerge from one and all soluble components are split according to the volume ratio of the mother and the bud. Through cell growth and division, we evolve an entire asynchronously growing population from one progenitor cell, as previously described [32]. The model is a comprehensible, minimal approximation of the complex process controlling duplication and separation in living eukaroytes [44].
To test the effect of mCLB2 localization on size control, we implemented two model versions. In Model-1, mCLB is uniformly distributed in the cell. It is translated in the entire cell. In this model, the biosynthetic capacity of the whole cell is integrated at the G2/M transition (G2 size control of the entire cell). In Model-2, the mRNA is translated exclusively in the newly forming bud mimicking the effect of mRNA localization. This model integrates only the biosynthetic capacity of the bud in G2 (G2 bud size control). Regardless of the specific G2 size control mechanisms, in both models, the productive capacity of the cell is measured in G1 through translation of mCLN (G1 size control), the primary phase for size control in budding yeast [1, 45]. The two models differ in a single equation (Table 1). In silico cultures generated with Model-1 and Model-2 show size homeostasis on the population level (Figs 2A and S1). Moreover, in both models there is a strong dependence of the cell volume at START on the individual growth rate in G1 phase (Figs 2B and S1–S3), as observed experimentally [15]. Thus, in accordance with experiments, both models exert growth rate dependent size control primarily in G1, regardless of the mCLB2 localization in G2 [1, 15, 45].
For the models to become reliable mathematical tools to investigate the coupling of growth and division they were fitted independently to complex experimental growth and proliferation data (Fig 3) [15]. The data is available for cells grown on glucose, galactose, raffinose and ethanol [15]. In the models, simulation of cell growth on different carbon sources can be achieved by modulating the parameter growth (Table 2). The parameter growth scales the biomass formation and represents the availability of nutrients to the system (Table 1). A decrease in growth leads to a reduction of cell size and a prolongation of the cell cycle [32]. We used the data for glucose and ethanol for parameter estimation and the data for galactose and raffinose for validation of the models (Fig 3A). The antagonistic trend within the data (‘fast growth → short cell cycle → large cells’ versus ‘slow growth → long cell cycle → small cells’, shown in Fig 3A as a grey shadow) enables to constrain most of the model parameters (Fig 3B, Materials and Methods). Although parameters show correlations (S4 Fig), within the parameters boundaries the fits converge into a global minimum (Figs 3B, S5 and S6). Both model variants predict cell size at birth, at budding, and duration of G1 and S-G2-M for all conditions with high accuracy. To test which model is more suitable to describe the given data, we ranked them using the Akaike Information Criterion (AIC) [46]. Ranking yields that Model-2 fits the data better than Model-1, although both versions use the same number of parameters (S1 Table). The worst fit of Model-2 is still better than the best fit of Model-1 (Fig 3B). This indicates that the mechanism of compartmentalization (bud-localised mCLB) renders the system more capable to describe (glucose and ethanol) and predict (galactose and raffinose) the data [15].
Experimental data clearly show that S-G2-M duration is not constant between different media [13–15]. Since S and M phase are constant, a G2 regulator is needed to account for the adaptation of G2 duration in response to nutritional status in vivo [13]. Since both models reproduce the experimental data for glucose, galactose, raffinose and ethanol (Fig 3A), we conclude that, regardless of compartmentalization, translation of mCLB is a good candidate. The proposed mechanism also leads to correlation between the volume of the bud at division and the individual growth rate in the budded phase as well as the S-G2-M duration (S7–S9 Figs). Thus, Clb2 constitutes a growth rate dependent sizer candidate in G2.
To further distinguish the models, we analysed the effect of mRNA localization on other systems level properties beyond average size and cell cycle phase duration. It has been reported that a growth rate dependent sizer can prevent large fluctuation of G1 length to reduce the generation time on the population level [15]. We find that compartmentalization of Clb translation (Model-2) reduces fluctuations of G2 length (Fig 4A). Interestingly, this does not lead to a reduction in generation time on the population level (S10 and S11 Figs). The higher noise at mitotic entry, inherent in Model-1, propagates directly to cell division ratios (Fig 4B), where cells that spend too much or little time in S-G2-M produce abnormally large or small buds, respectively. In comparison with experiments, Model-1 predicts a too high variability in division ratios for the first five generations of the population, i.e. for more than 95% of the cells in the culture [41]. In contrast to Model-1, the predictions of division ratios from Model-2 are accurate for young and old cells (Fig 4B). This is even more pronounced for slow growing cells (S12 Fig). Thus, a bud sizer can tune mitotic entry to reduce noise and maintain division ratios over generations.
It is well known that cells grown on rich media grow faster and are larger compared to cells grown on poor media [47]. Previously, we analysed cell size distributions of cells grown on rich and poor media and found that, in contrast to average cell size, the relative variability in cell size does not change [32]. A model that allowed size control exclusively in G1 (constant S-G2-M) could not explain this [32]. Model and data could only be reconciled when we adapted S-G2-M duration to growth conditions. Thus, we hypothesised that, for a robust setting of average cell size and variability within the culture, S-G2-M duration must show some form of adaptation to growth conditions. We have seen that Model-1 and Model-2 are able to reproduce and predict the adaptation of S-G2-M duration in response to different conditions (Fig 3). To test whether both versions stably reproduce an increase in average cell size while keeping the relative variability constant, we analysed their behavior with respect to cell size statistics (Fig 5). It is apparent that both versions show the expected increase in average cell size, but that relative variability is increased in Model-1. Model-1 shows an increase in relative variability of roughly 25% over different conditions (glucose ∼ 0.65, ethanol ∼ 0.81), whereas Model-2 of only 7% (glucose ∼ 0.60, ethanol ∼ 0.64). This indicates that a growth rate dependent G2 bud sizer stabilises the relative variability in cell size observed in yeast populations [32].
Both model versions show adaptation of S-G2-M duration to growth conditions (Fig 3). Hence, according to our hypothesis (adaptation of S-G2-M stabilises the variability—see last paragraph), both models should in theory be able to stabilise the variability [32]. However, only Model-2 is able to limit the size variability (Fig 5). To explain this, we analysed the apparent S-G2-M adaptation in both models. The difference between Model-1 and Model-2 in G1 and S-G2-M duration seems minor under most conditions (Fig 3). However, the difference between the models becomes striking when inspecting the time that daughter and mother cells spend in G1 and S-G2-M separately (Fig 6A). Apparently, in Model-1 mother and daughter lines diverge, whereas in Model-2 they do not. In Model-1, time in G1 is different for mothers and daughters, as expected [48]. This difference is larger for cells grown on ethanol than for cells grown on glucose, also as expected [14]. In Model-1, the time in S-G2-M differs for mothers and daughters as well, which is in clear contrast to experimental evidence [13, 14]. Single cell data has shown in detail that there is little difference for time in S-G2-M between daughters and mothers (Table 3) [14]. Model-2 also displays the expected differential G1 duration between mother and daughter cells, which again decreases with nutrient quality, as expected (Fig 6A) [48]. In contrast to Model-1, the S-G2-M duration of mothers and daughters in Model-2 is more equilibrated, very similar to experimental findings (Table 3). Yet, there is a trend in Model-2 that for slow growing cells budded phase is longer for mother than daughter cells. A consequence of the equilibrated budded phase within the population (Model-2), is that the volume of new born daughter cells increases with the age of the corresponding mother (Fig 6B). As a result, daughters of old mothers are considerably larger than daughters from young mothers in Model-2. This observation has also been made in vivo [49], suggesting that a mechanism exists that controls the size of the bud, rather than absolute cell size, in G2.
To further compare the predictive power of the models, we simulated a scenario similar to over producing CLN3. We enforced a doubling of CLN expression in the models (referred to as OE-CLN). Both models react to the overproduction of CLN with a decrease in G1 duration and a compensatory increase of the budded phase duration (S-G2-M; Fig 7). In agreement with experimental observations, average generation times are only slightly (Model-2) or not at all (Model-1) reduced by the mutation (S10 and S13 Figs) [33, 34]. However, the difference between cell cycle durations of mothers and daughters is reduced in the mutant (S13 and S14 Figs). Yet, only Model-2 shows a the reduction in cell size in response to OE-CLN (Fig 7). Model-1, with the whole-cell G2 sizer, fails to predict the small cell size phenotype that is typical for CLN3 overexpressing cells [33, 34]. This shows that a bud sizer (Model-2) is required to predict the CLN3 overexpressing strain’s whi mutant phenotype.
Here, we present a mechanistic single cell growth model that is able to predict cell growth and division timing in budding yeast populations. There are very few models that are able to (i) show and explain size homeostasis, (ii) offer mechanistic insight into the cellular machinery governing growth and division, and (iii) that are still comprehensible and manageable [32, 50]. Our model is designed to omit all details that are not absolutely required to reproduce the coordination of growth and proliferation. Using the model we show that a G2 bud sizer mechanism is required in addition to a G1 sizer in order to (i) better fit and predict population size and timing data for different nutritional conditions (Figs 3, 4B and 5, Table 3), (ii) offer a functional explanation for the experimentally observed mCLB2 transport into the bud [35], (iii) reduce the noise at mitotic entry (Figs 4 and S12) and (iv) render the model capable of predicting the phenotype of a CLN3 overproducing strain (Fig 7). Thus, our results indicate that a bud sizer mechanism could operate at the G2/M boundary in vivo [10].
We use the model to show that biomass dependent accumulation of cyclins to a threshold results in size homeostasis on the population level and growth rate dependent size adaptation in G1 (Fig 2), as seen in vivo [15]. Thus, our results are in accordance with the view that G1 is the primary phase for size control in budding yeast [1, 45]. It was shown that the START network sets the cell size as a function of the growth rate [15]. We observe a similar behaviour in our model, meaning that we implement a growth rate dependent sizer through a minimal version of the START network. The model proposes that the underlying network developed from a single CDK-cyclin pair that later differentiated between G1 and G2 phase. Consistently, cells can be driven through the cell division cycle by artificial expression of a single CDK-cyclin fusion protein [43]. While we observe a stronger correlation between the cell volume at START and the individual growth rate than seen in experiments (Fig 2B) [15], this can be explained by the simplicity of the model and the lack of feedbacks and other regulatory mechanisms [23, 26, 51, 52]. Taken together, the model predicts (a) that already a single G1 cyclin suffices for size control in G1 and (b) that monitoring biosynthetic capacity through production of a critical unstable cyclin is a growth rate dependent sizer.
In previous work, we found that a growth model with size control operating exclusively during G1 is not able to fully reproduce data of cell sizes and proliferation times measured for different conditions [32]. Also, recent experimental evidence strongly supports the existence of a size-regulating mechanism in the budded phase [10]. Accordingly, we tested models with additional size control at the G2/M transition for the entire cell (Model-1) or only for the bud (Model-2). The G2 size control is implemented analogously to the G1 sizer, through biosynthetic capacity dependent production of a critical cell cycle regulator. The models establish that a G2 size control point, but not necessarily bud size control, is required to reproduce the experimentally observed adaptation of the budded phase in poor growth media (Fig 3) and in the CLN3 overexpression mutant (Fig 7) [13–15, 33, 34]. Hence, both models argue in favor of the existence of a sizer mechanism in G2. Yet, simply regulating G2 length in response to growth conditions does not seem to be the end of the story. Conceptually, the reduction of G1 and the compensatory adaptation of the budded phase, inherent in both models, is not sufficient to explain the small size phenotype of the CLN3 overproducer (Figs 7 and S14).
The prolongation of the budded phase has different reasons in Model-1 and Model-2. In Model-1, cells pass START quickly with a reduced biosynthetic capacity (and size) as a direct consequence of the CLN overexpression. Since Model-1 is implemented to integrate the whole cell’s biosynthetic capacity at mitotic entry, cells compensate by extending S-G2 to build up sufficient productive power to enter mitosis. The requirement to enter mitosis here equals the one in the wild type so that, in Model-1, CLN overexpressing cells are very similar in size compared with wild type cells. A reduction in growth rate, due to early passage through START is compensated late in the cell cycle when enough biosynthetic capacity has built up. Consequently, cells overexpressing CLN in Model-1 tend to have larger buds leading to larger cells at birth (S14 Fig). In contrast, the prolongation of the budded phase seen in Model-2 is due to a general reduction in growth rate, which is again the consequence of the early passage through START. In Model-2, the biosynthetic capacity of the bud is monitored, which accumulates slower because of the generally reduced growth rate. However, in Model-2, a cell can enter mitosis as long as the bud fulfills minimal requirements, even if the cell in total is smaller and less productive. This is in accordance with the fact that entry into mitosis is correlated with the size of the bud, but not with the size of the mother cell [17]. Thus, only the bud sizer and not a whole cell sizer concept at mitotic entry is able to accommodate the small size phenotype of the CLN3 overproducer and can thus reconcile model and experimental observations [33, 34].
We found that the budded phase is slightly longer in mothers than in daughters for slow growing cells in Model-2 (Fig 6A). While such an effect might be too small to be detected in experiments, it is more likely that the model oversimplifies at this point: The assumption that 100% of the CLB2 transcript is transported to and exclusively translated in the bud in vivo could be too strong. Indeed, it is difficult to assess the exact number of bud-localised transcript (100% in the model and ≥ 90% as reported experimentally [35]). Predictions of Model-1 show the effect of non-localization, i.e. shorter budded phase in mothers. This is indicative that allowing some translation of the cyclin in the mother (e.g. 10%) can eradicate the difference.
We predict bud-localised mCLB2 to be an essential part of the growth and division coupling. Obviously, we cannot be sure that we have implemented the true biological mechanism in our model, but we show here that our minimal mechanism displays many characteristic properties of the system. It is tempting to speculate that the G2 bud sizer is of similar design as the one operating in G1, since it seems easier to duplicate a working mechanism than to invent two distinct, yet functionally equivalent ones. Still, we acknowledge that there are other hypotheses on how cyclin synthesis is related to growth rate [30, 31, 40, 53]. Most likely, the in vivo situation is the complex result of different interacting molecular effects. Nonetheless, the mechanism proposed here suffices to explain most of the data, and—unlike many other hypotheses that rely on a specific molecular mechanism in the G1 phase—the mechanism proposed here is generic and not phase specific. In light of this, passive accumulation of CDK regulators (Cln3 in G1 and Clb2 in G2) are promising candidates [8]. The old concept of the unstable regulator is seductively simple and elegant [25]. By making the underlying size control mechanism depending on the critical CDK activity induced by an exchangeable regulator (cyclins can substitute for one another) one could explain why none of the components, neither concerning the START network nor the morphogenesis checkpoint, are absolutely indispensable [43, 54, 55].
The importance of the mRNA localisation, as we highlight it here, can possibly be tested experimentally. The polarised localisation of mCLB2 could be perturbed by disruption of the sequence in mCLB2 required for transport (if identified), or by deletion of the MYO4 gene that encodes the type V myosin motor responsible for bud-localisation of mRNA [56]. Such perturbations would revert the phenotype from Model-2 to Model-1. To discriminate the two experimentally, the MYO4 deletion would need to be combined with a CLN3 overexpression. The model predicts the loss of the whi phenotype in this double mutant, which should be detectable in an experimental setting (Fig 7). However, this experiment involves two major genetic perturbations and the outcome may not be as clean as the in silico experiment.
Seeing that mCLB2 is localised to the bud of the cell and during S-G2-M mainly the bud grows, it is likely that a G2 sizer actually has an effect on bud size rather than absolute cell size in vivo [35]. Also, while measuring the biosynthetic capacity of a cell in G1 works well to ensure that it is strong enough for duplication, measuring the capacity again in G2 (cell and bud—Model-1) seems futile and less accurate. It seems rather more useful to measure the capacity only of the bud (Model-2), ensuring that the new descendant is strong enough for independent life. From a population’s perspective, it is reasonable to control the offspring’s size at least as strictly as individual cell size, maybe even with the same mechanism.
Taken together, we propose a model where cell size in budding yeast is controlled at the cell cycle junctions G1/S and G2/M in a growth rate dependent fashion (Fig 8). Our results suggest that the simple mechanism we employ here at both transitions is sufficient. Moreover, the model works without setting of a ‘critical cell size’. Considering the growth medium dependent nature of the ‘critical cell size’ itself, this strongly advocates an interpretation of cell size as an emergent property of the coupling between growth and division, rather than a regulatory parameter. In accordance, cell size at START would be a function of the growth rate in G1. Strong evidence in this direction recently emerged [15]. We propose here that bud size at division is a function of the growth rate in S-G2-M, meaning that there is a common size control theme in the two growth phases of the cell division cycle. This can explain the prolongation of both phases, G1 and G2, in response to poor nutrients and the small size phenotype seen for CLN3 overexpression [14, 15, 33, 34]. In conclusion, we present a cell growth model, which unifies integration of growth and division in the G1 and G2 phases of the cell division cycle to accurately reproduce and predict cell size at birth and at budding, as well as timing of the cell cycle phases over four different nutritional conditions for budding yeast.
The model is an extension of a minimal eukaryotic cell model [32]. It is implemented with ordinary differential equations, stochastic functions and algebraic equations (Table 1). Two species were added (mCLB & Clb), such that Cln drives the cell cycle in G1 phase and Clb in G2. Hence, transcription of CLN is restricted to G1 and of CLB to G2. By default, the model links metabolism to progression through the cell cycle via biomass dependent accumulation of the two regulatory proteins (G1 cyclin Cln and G2 cyclin Clb), whereas the synthesis phase (S-phase) and the Mitose (M-phase) simply delay cell cycle progression (see Table 2). The cell cycle of the model has four transitions, corresponding to the eukaryotic phase transitions (G1/S, S/G2, G2/M, M/G1). The model equations governing the dynamics are displayed in Table 1.
The model rests on a set of explicit assumptions: namely, that nutrient supply is defined by uptake, which is proportional to cell area; that transcription is stochastic and that nutrient incorporation into biomass relies on the biosynthetic capacity of the cell. Thus, production reactions are dependent on precursors and the internal biomass. The efficiency of nutrient incorporation is inversely scaled with volume to reflect dilution. Furthermore, that the total area of the cell is the sum of the area of the mother and the bud. Correspondingly, we calculate the total volume of the cell as the sum of mother and bud volume. Mothers and buds are approximated as separate spheres, thus V ∝ A3/2. As a cell grows, the ratio of the area to the volume shifts, since the area expands slower than the volume. Given our above stated assumption about the influence of nutrient supply (area) and dilution (volume), it follows that the decrease of the area-to-volume-ratio places an increasing constraint on the cellular biosynthetic capacity slowing down growth [32]. The idea that the surface area-to-volume-ratio plays an important role in connecting the cell growth to the cell division cycle was also explored by others [50, 57]. In our model, cells may allocate their resources according to cell cycle stage, which means that resources can be used to form structural or internal biomass in different proportions in different cell cycle stages. Specifically, we assume that the increase of the biosynthetic capacity is strong in G1 (heavier allocation of resources to internal biomass) and less so during S-G2-M [58, 59]. The structural biomass can furthermore be distributed to either the area of the mother cell or the area of the bud. There is no bud growth during G1 and we assume that there is only bud growth during S-G2-M [15, 40]. Finally, we simplify phase transitions to a threshold for nuclear kinase activity, assuming zero order ultra-sensitivity [60].
The model itself is a single cell model that can grow and divide (S1 File). To model entire asynchronously growing cell cultures, however, we developed an algorithm to simultaneously simulate a growing ensemble of the single cell models [32]. An executable and editable version of the algorithm implemented in python is added in the Supporting Information (S2 File). During the simulation, a cell grows during G1, then it grows a bud during S-G2-M. At division, the bud is detached from the mother cell. The mother cell starts a new cell cycle and, additionally, a new cell instance is created according to the size of the detached bud. All soluble components are split at division between the two cells according to the volume ratio of the mother and the bud. In this fashion, two distinct cells are created from one cell. The two cells differ in starting conditions, e.g. cell size or biosynthetic capacity, which lead to differential growth and proliferation properties for the individual cell, e.g. growth rate or time spent in G1. Simulation of cell cultures in different nutrients is implemented through the parameter growth, which is used to scale the biomass formation to control the nutrient availability of the system (Tables 1 and 2). In both models we use the constraint that growthethanol = growthglucose⋅0.5. Similar relations are used to simulate galactose (growthgalactose = growthglucose⋅0.77) and raffinose (growthraffinose = growthglucose⋅0.6).
We implemented two different versions of the model. Note that the only difference between the two is equation 4. Model-1 uses equation 4.1 and Model-2 is implemented with equation 4.2. Model version 2 is based on the fact that mCLB2 is actively transported into the nascent bud in S. cerevisiae [35]. However, since the model does not include spatial displacement of components, we employed a work-around to implement the consequence of the transport. Assuming that the important function of mCLB2 transport into the bud is to localise translation to this sub compartment, we allowed only the fraction of ribosomes in the bud to translate the CLB mRNA. This is why, in equation 4.2, the metabolic capacity (BR) is scaled with the term Vd(t)/V(t), assuming well-stirred conditions.
Model-1 and Model-2 contain 14 parameters each, five of which we estimated using a maximum likelihood approach (see Table 2). Both model versions were fitted independently to experimental data from Aldea and colleagues who analysed asynchronously growing daughter cells using time-lapse microscopy [15]. We used two different though related types of their data to constrain our parameters: (i) time at START (T1), budding (T2) and division (T3); (ii) volume at birth (V0) and budding (Vbud). Aldea and colleagues provide the mean value (μ) and coefficient of variation (cvar in %) for both types of data for daughter cells grown on four different carbon sources (glucose, galactose, raffinose and ethanol). Here, we used the data for glucose and ethanol for parameter estimation. We recalculated the standard deviation (σ) from the cvar and the mean, such that σ = μ⋅cvar/100. The data and the model fits are shown in Fig 3.
In the model we do not distinguish between START transition (T1) and budding (T2) but consider that as soon as the threshold is crossed, the in silico cells enter S-phase and G1 is finished. As such, for this special case, we relate the model and the data as follows. Time in G1 (TG1) from the model equals the experimental time at START (T1) plus the time from START to budding (T2)
TG 1 = T 1 + T 2 . (1)
Now the mean data point is the sum of T1 and T2 with two stdevs σT1 and σT2, respectively. Neglecting correlations of T1 and T2, the formula for propagation of uncertainties provides an approximation of the combined error
σ f = ( ∂ f ∂ x 1 ) 2 σ x 1 2 + ( ∂ f ∂ x 2 ) 2 σ x 2 2 (2)
where f = x1+x2 (see Eq 1) [61]. This simplifies to yield the combined error for TG1
σ TG 1 = σ ( T 1 + T 2 ) = σ T 1 2 + σ T 2 2 . (3)
As objective function for the fitting of parameters we chose the weighted sum of squared residuals (wRSS) given by the measured values x and the simulated values x ^, such that
wRSS = ∑ ( x − x ^ σ ) 2 . (4)
Minimization of Eq 4 is equivalent to maximizing the log-likelihood given by
l n ( L ( p ) ) = − m 2 l n ( 2 π ) − ∑ l n ( σ ) − 1 2 ∑ ( x − x ^ σ ) 2 . (5)
It is important to stress at this point that the experimental data was generated analysing only daughter cells [15]. Accordingly, only in silico daughter cells we used to calculate the appropriate data x ^.
Model-1 and Model-2 were fitted independently using a custom evolutionary parameter estimation algorithm (S1 Text). The algorithm is suitable for fitting population data and has been used for all parameter estimation tasks. The algorithm allows specification of parameter boundaries for the estimation (parameter boundaries are shown in Table 2). The estimation was performed for 100 uniformly distributed initial values (in the range of the parameter boundaries) for the parameters which enabled us to derive the parameter correlations (S4 Fig).
We estimate five out of 14 parameters because to a certain degree we anticipated parameter correlations, over-fitting or under-determination of parameters since the nature of the data (population averaged) and the parameters (single cell) are distinct. There are correlations in the parameters that the data cannot account for. For example, the parameters for Cln protein production kp1, degradation kd, probability of mCLN synthesis PmCLN and the threshold value (Table 2) together influence the duration of G1 phase. Their combined effect determines the final duration. Since the given data concerns the duration, we cannot hope to estimate the true value of the four influential parameters but only their combined effect. This is why we set three out of four to a fixed value and estimated the fourth, such that the global effect matches the data. We can thus not report unique kinetic parameters for protein production and degradation but values that are useful in combination to describe the global process that determines G1 duration.
Accordingly, our approach enables us to describe the global effect and also distinguish different model versions. To find the model version that would best approximate reality given the data and the number of parameters we employed the Akaike Information Criterion (AIC) to rank the models [46]. The AIC establishes a relationship between the maximum likelihood and the Kullback-Leibler information, which is a measure for the information lost when approximating reality with a model [62]. The AIC was computed as
A I C = − 2 ( l n ( L ( p ) ) ) + 2 K (6)
with K being the number of estimated parameters in the model. Model statistics with respect to the objective function, the log-likelihood and the AIC are summarised in S1 Table.
|
10.1371/journal.ppat.1005116 | Supporting Role for GTPase Rab27a in Hepatitis C Virus RNA Replication through a Novel miR-122-Mediated Effect | The small GTPase Rab27a has been shown to control membrane trafficking and microvesicle transport pathways, in particular the secretion of exosomes. In the liver, high expression of Rab27a correlates with the development of hepatocellular carcinoma. We discovered that low abundance of Rab27a resulted in decreased hepatitis C virus (HCV) RNA and protein abundances in virus-infected cells. Curiously, both cell-associated and extracellular virus yield decreased in Rab27a depleted cells, suggesting that reduced exosome secretion did not cause the observed effect. Instead, Rab27a enhanced viral RNA replication by a mechanism that involves the liver-specific microRNA miR-122. Rab27a surrounded lipid droplets and was enriched in membrane fractions that harbor viral replication proteins, suggesting a supporting role for Rab27a in viral gene expression. Curiously, Rab27a depletion decreased the abundance of miR-122, whereas overexpression of miR-122 in Rab27a-depleted cells rescued HCV RNA abundance. Because intracellular HCV RNA abundance is enhanced by the binding of two miR-122 molecules to the extreme 5’ end of the HCV RNA genome, the diminished amounts of miR-122 in Rab27a-depleted cells could have caused destabilization of HCV RNA. However, the abundance of HCV RNA carrying mutations on both miR-122-binding sites and whose stability was supported by ectopically expressed miR-122 mimetics with compensatory mutations also decreased in Rab27a-depleted cells. This result indicates that the effect of Rab27a depletion on HCV RNA abundance does not depend on the formation of 5’ terminal HCV/miR-122 RNA complexes, but that miR-122 has a Rab27a-dependent function in the HCV lifecycle, likely the downregulation of a cellular inhibitor of HCV gene expression. These findings suggest that the absence of miR-122 results in a vulnerability not only to exoribonucleases that attack the viral genome, but also to upregulation of one more cellular factor that inhibit viral gene expression.
| Eukaryotic cells constantly expel a variety of small vesicles that are loaded with proteins, nucleic acids and other small compounds that were produced inside the cell. One particular kind of vesicle is called exosome. Exosomes are initially located in multivesicular compartments inside cells and are docked at the cell surface membrane by the small GTPase Rab27a. In the liver, high expression of Rab27a correlates with the development of hepatocellular carcinoma, suggesting a high trafficking capacity for exosomes. Also, it has been shown that hepatitis C virus (HCV) can spread from cell to cell via exosomes. We discovered that Rab27a abundance affects HCV virion abundance that independent from its role in exosome secretion. The presence of Rab27a in membrane-enriched replication complexes and nearby lipid droplets points to functions of Rab27a in the viral life cycle. Depletion of Rab27a resulted in a lower abundance of the liver-specific microRNA miR-122. It is known that two molecules of miR-122 form an oligomeric complex with the 5’ end of the viral RNA leading to protection of the viral RNA against cellular nucleases. However, we show that the Rab27a-mediated loss of miR-122 was independent of its role in protecting the viral RNA, very likely by the downregulation of a cellular inhibitor of HCV gene expression. These findings argue for novel, hitherto undetected roles for miR-122 in the viral life cycle.
| Hepatitis C virus (HCV) is a hepatotropic positive-sense, single-stranded RNA virus that belongs to the Flaviviridae family. The HCV genome is about 9.6 kb in length and encodes a polyprotein, which is cleaved into at least ten viral proteins by host and viral proteinases [1, 2]. The open reading frame is flanked by 5’ and 3’ noncoding regions, which regulate translation and replication of the viral RNA. In addition, the 5’ terminal sequences of the HCV RNA genome form an oligomeric complex with two molecules of liver-specific miR-122 [3, 4]. This complex greatly stabilizes the viral RNA from degradation by exonucleases [5, 6].
Exposure to HCV typically leads to persistent infections that cause chronic hepatitis, liver cirrhosis, and hepatocellular carcinoma [7]. An estimated 170 million people are affected by the virus, making it a serious global health burden [8]. Recently, Gilead Sciences’ sofosbuvir/ledipasvir (Harvoni) and AbbVie's paritaprevir/ritonavir/ombitasvir plus dasabuvir (Viekira Pak) were approved as the new line of interferon-free treatment regimen. In addition, Miravirsen (Santaris Pharma, Denmark), an antisense inhibitor of miR-122, showed a decrease of HCV titers in patients chronically infected with HCV in phase II clinical trials [9], demonstrating that miR-122 is a potential therapeutic host target to combat HCV. Here, we report an additional role for miR-122 in promoting HCV infection that is independent of its well-characterized 5’ end stabilization function.
Like many RNA viruses, HCV exploits membranes and the trafficking machinery of the host for viral replication [10, 11]. For example, accumulating evidence suggests that HCV can exit infected cells via the multivesicular transport system [12–15]. While these studies employed fractionation and ultrastructural approaches, evidence for the cellular origin or the mechanism of vesicle generation remains lacking. Recently, it has been reported that Rab27a modulates exosome vesicle secretion by docking multivesicular bodies to the plasma membrane [16]. Curiously, several studies have shown that Rab27a, a small GTPase, is also involved in replication of viral genomes in cells infected with human immunodeficiency virus, herpes simplex virus, hepatitis E virus and HCV [15, 17–19]. However, the mechanism by which Rab27a modulates viral genome replication remains unclear. In this study, we found that Rab27a affects HCV RNA and virion abundance by a pathway that is independent of exosome secretions. Specifically, Rab27a located to membranes that are enriched in viral replication complexes and to lipid droplets, which are sites thought to initiate packaging of the viral RNA genome. Furthermore, intracellular abundance of Rab27a affected miR-122 abundance. Curiously, Rab27a’s modulation of miR-122 was independent of miR-122’s stabilizing role of the viral RNA. Therefore, Rab27a likely downregulates, via miR-122, a cellular inhibitor of HCV gene expression.
To determine whether HCV RNA and protein abundances are regulated by exosomal vesicles, we first inhibited exosomal trafficking in human liver carcinoma Huh7 cells by depletion of Rab27a [16]. Northern analyses revealed that the liver Rab27a gene is transcribed into three RNA transcripts of 1.2 kb, 2.6 kb and 3.5 kb in size (Fig 1A). This result is consistent with Rab27a RNA species that are expressed in human fibrosarcoma cells [20]. All three Rab27a transcripts were decreased by 90% in both uninfected and JFH1 HCV-infected cells that were treated with siRNAs directed against the common Rab27a open reading frame (Fig 1A and 1B). As expected, Western blot analysis showed that the abundance of Rab27a protein was also decreased in Rab27a siRNA-treated cells (Fig 1D). To determine if Rab27a depletion affected extracellular exosome yield, the abundance of CD81, a marker for exosomes derived from the multivesicular body pathway, was examined in cell lysates and in extracellular, partially purified exosome preparations. S1 Fig shows that Rab27a depletion diminished the extracellular amount of CD81-containing exosomes in uninfected (S1A Fig) and in HCV-infected cells by approximately 40% (S1B Fig).
To examine the effects of Rab27a on HCV gene expression, viral RNA and protein abundances were measured in Rab27a-depleted cells. Results showed that Rab27a depletion caused a 60% decrease in HCV RNA abundance (Fig 1A, lane 4, and Fig 1C), but had no effect on actin mRNAs. Rab27a depletion also led to a decrease in HCV core protein abundance (Fig 1D). These data are consistent with a previous report on the effect of Rab27a depletion on HCV RNA abundance [15]. Similar effects of decreased viral RNA (S2A Fig) and protein (S2B Fig) protein abundances during Rab27a depletion were observed when cells were infected at a 1000-fold higher multiplicity of infection with HCV. To control for siRNA off-targeting effects, additional Rab27a siRNAs (siRNA-3 and siRNA-4), which target different regions of all Rab27a mRNA species, were tested. These siRNAs also showed decreased Rab27a and HCV RNA (S3A Fig) and protein abundances (S3B Fig). Importantly, Rab27a depletion in Huh7 cells did not have a significant effect on cell viability (S4A Fig) or caused apoptosis (S4B Fig). Therefore, depletion of Rab27a causes selective inhibition of HCV gene expression without any significant effects on cellular viability.
It has been reported that HCV can be transmitted from cell to cell via exosomes [12, 14, 21–23]. Rab27a plays a role in exosome secretion. Thus, we would expect an increase in cell-associated virus titer in Rab27a-siRNA treated cells compared to control-siRNA treated cells. Depletion of Rab27a decreased extracellular virus titer by about 80% (Fig 2A), but, surprisingly, cell-associated virus titer also decreased by about 60% (Fig 2B). However, the ratio of cell-associated to total infectious virus particles in Rab27a-depleted cells was similar to that of control-siRNA treated cells (Fig 2C). Consistently, extracellular HCV RNA abundance was decreased to nearly 80% in infected cells that were treated with Rab27a siRNAs, compared to Ctrl siRNAs-treated cells (4.9 x 106 copies/ml) (Fig 2D). The decrease of extracellular HCV RNA abundance did not cause an accumulation of intracellular HCV (Fig 1A, lane 4, and Fig 1C). Thus, these data suggest that the diminished yield of cell-associated infectious virus particles during Rab27a depletion is not due to impaired exosome secretion, arguing that Rab27a modulates HCV gene expression by a mechanism that is different from its role in exosome secretion.
To determine whether Rab27a modulates viral RNA abundance at the RNA replication or translation step, and to bypass any effects on virion entry, we monitored the expression of subgenomic JFH1-Rluc (sgJFH1-Rluc) replicons [24, 25] (Fig 3A). These replicons are either competent for both translation and RNA replication, or contained a GND mutation in the catalytic domain of the viral RNA-dependent RNA polymerase (NS5B) that prevents genome replication (Fig 3A). Briefly, Huh7 cells were transfected with Rab27a siRNAs, and subsequently transfected with replication-competent sgJFH1-Rluc RNAs (Fig 3B) or replication-defective sgJFH1-Rluc-GND RNAs (Fig 3C). Luciferase activity was measured at different times after HCV RNA transfection. Two peaks of luciferase activity were noted in the sgJFH1-Rluc RNA-transfected cells treated with control siRNAs (Fig 3B). The first peak at 4 hours post-transfection represents the initial translation of the input RNA, which is absent in cyclocheximide-treated cells (Fig 3B and 3C). The second luciferase peak represents the translation of replicating RNAs, because it is absent in sgJFH1-Rluc-transfected cells that were treated with the NS5B inhibitor MK-0608 (Fig 3B) and in sgJFH1-Rluc-GND-transfected cells (Fig 3C). Depletion of Rab27a did not diminish translation of the input RNA (Fig 3B and 3C). However, translation of replicating RNAs was significantly decreased in Rab27a-depleted cells compared to control siRNA-treated cells. Importantly, the EMCV IRES activity was not affected by Rab27a depletion (S5 Fig), eliminating the possibility that these results were due to altered abundances of viral proteins. These findings argue that Rab27a plays a role in the viral life cycle by modulating HCV RNA replication.
It is known that cells expressing HCV replicons or cells that are infected with HCV display membrane rearrangements and formation of virus-induced membranous webs [11, 26–29]. The HCV-induced membranous webs, which are thought to be the sites of viral replication, are mainly derived from the endoplasmic reticulum (ER)[29]. To examine whether Rab27a is located to membranes during HCV RNA replication, membrane-enriched fractions from uninfected and HCV-infected cells were isolated, using discontinuous sucrose gradients. Western blot analyses showed that the membrane fractions contained the ER membrane marker protein calnexin (Fig 4A and 4B, lanes 3 and 4). In addition, HCV proteins NS5A, NS3 and capsid protein core also located to these fractions (Fig 4A and 4B). Interestingly, Rab27a was also found to localize in the membrane-enriched fraction. Rab27a depletion caused a decrease of HCV NS3, NS5A and core protein abundance in the enriched-membrane fraction (Fig 4B, lane 4), but not calnexin or GAPDH. These results indicate that Rab27a is associated with membrane-enriched fractions in infected cells, and that Rab27a depletion selectively diminished the abundance of several viral non-structural proteins in the replication complex-containing membranes.
The above genetic and biochemical findings argue that Rab27a regulates HCV RNA replication via its association with virus-induced membranes. To further substantiate this hypothesis, the subcellular location of Rab27a was investigated by confocal immunofluorescence microscopy. Astonishingly, Rab27a exhibited a doughnut-like structural localization around lipid droplets (LDs) (Fig 5 and S6 Fig) in uninfected (Fig 5A) and in infected liver cells (Fig 5B). These findings suggest that Rab27a may have a hitherto unknown role in the metabolism of LDs in liver cells. The LD-Rab27a doughnut-like structures colocalized with viral core protein in infected cells (Fig 5B). In addition, a small fraction of NS3 displayed a punctate distribution in the LD-Rab27a structures, indicating that Rab27a localizes to adjacent to sites of viral replication (S6B Fig).
The impaired HCV gene and protein expression may be due to a lack of stabilization of HCV RNA. To examine whether Rab27a affects HCV RNA stability, Huh7 cells were transfected with control- or Rab27a-siRNAs, followed by addition of the NS5B inhibitor MK-0608 to block new synthesis of HCV RNA. The rate of HCV RNA decay was determined by Northern blot analysis at different times after addition of MK-0608 (Fig 6A). Viral RNAs from control- and Rab27a-depleted samples displayed similar decay rates, with approximate half-lives of 4.8 hours (Fig 6B). These results indicate that Rab27a depletion affects the rate of HCV RNA replication without changing HCV RNA stability.
It is known that miR-122 modulates HCV RNA expression [3, 30]. Therefore, it is possible that the observed effects of Rab27a depletion on the rates of HCV RNA replication could be due to altered abundance of miR-122. Thus, intracellular miR-122 abundance was monitored in Rab27a-depleted cells by Northern blot analysis. Results showed that miR-122 abundance was decreased by more than 30% in both uninfected- and HCV-infected Rab27a-depleted cells (Fig 7A and 7B). This was surprising because miR-122 has been reported to be quite stable in liver cells [31]. A luciferase reporter-based assay also showed diminished miR-122 function in Rab27a-depleted cells (S7 Fig). While the abundances of five other endogenous miRNAs (miR-16, miR-21, miR-22, miR-26 and miR-130a) were not changed in uninfected, Rab27a-depleted cells (Fig 7A), the abundances of miR-16, miR-22 and miR-130a showed a modest decrease in Rab27a-depleted cells during HCV infection; but not to the same extent as miR-122 (Fig 7B).
To test whether the modulation of HCV RNA replication by Rab27a was caused by the altered abundance of miR-122 or any other microRNA, we investigated whether miR-122 overexpression prevented the Rab27a-dependent inhibition of HCV RNA replication (Fig 8A). Fig 8B shows that overexpression of miR-122 mimetics could rescue HCV RNA abundance in Rab27a-depleted cells, while the overexpression of miR-22 had no effects. A similar result was observed during overexpression of miR-21 as a control. These findings suggest that the decrease of HCV RNA abundance in Rab27a-depleted cells is due to the reduction in miR-122 abundance and is not due to the reduction of other microRNAs, such as miR-22 (Fig 8B).
We next examined whether Rab27a modulates the transcription of miR-122. Primary miR-122 (pri-miR-122) transcript abundance was examined in Rab27a-depleted uninfected or HCV-infected cells. S8A and S8B Fig shows that the abundance of pri-miR-122 is not affected by the depletion of Rab27a in uninfected and infected cells, suggesting that Rab27a modulates miR-122 abundance at a post-transcriptional step. Because precursor-miR-122 (pre-miR-122) can not be detected in cultured Huh7 cells, we determined the effect of Rab27a on the stability of a pre-miR-122 species that is resistant to the cleavage by Dicer [32]. Thus, the intracellular decay of a dicer-resistant pre-p3 (dNx12) that is functional in regulating mRNAs with miR-122 target sites [32] was examined (S9A Fig). Control- or Rab27a-siRNA treated cells were transfected with 5’-32P-labelled pre-p3 (dNx12) mimetics and the abundance of the labeled pre-miRNAs was determined at one day after transfection. The three independent experiments in S9B Fig show that the abundance of 5’-32P-labelled pre-p3 (dNx12) significantly decreased by the depletion of Rab27a (S9C Fig), arguing that Rab27a likely diminished miR-122 abundance by decreasing pre-miR-122 abundance.
It is known that two miR-122 molecules protect the 5’-terminal sequence of the HCV RNA genome from exonucleolytic degradation [5, 6]. Thus, it was possible that the reduced level of intracellular miR-122, after Rab27a depletion, caused the decrease in HCV RNA abundance by leaving the viral RNA unprotected. To test this possibility, a mutant HCV RNA genome (HCV-G27G42) that contained a mutation at each of the two miR-122 binding sites at the 5’ UTR was generated (Fig 9A, nucleotides highlighted in red). When transfected into cells, HCV-G27G42RNA cannot replicate because it cannot bind endogenous miR-122 (Fig 9A, (I), left upper panel) [3, 4, 30]. However, introduction of p3-loop miR-122 molecules that harbor a compensatory mutation at position 3 (red), and additional mutations at positions 9–13 (orange) and 18 (orange) (Fig 9A, (I), lower panel) can enhance HCV-G27G42 RNA abundance (Fig 9B, lanes 1 and 2). The nucleotide changes 9–13 (orange) and 18 (orange) in p3-loop miR-122 allow us to distinguish p3-loop miR-122 from endogenous wildtype miR-122 in Northern blots. As a negative control, miR-122 molecules with mutations in their entire seed sequences (p2-8; nucleotides highlighted in blue) (Fig 9A, (II), lower panel) did not enhance HCV-G27G42 RNA abundance (Fig 9B, lane 4). This finding shows that the HCV-G27G42 RNA genome abundance was enhanced by p3-loop miR-122, and not by endogenous miR-122 or p2-8 miR-122. Expression of p3-loop miR-122 mimetics allowed a 50% of HCV RNA accumulation in Rab27a-depleted cells (Fig 9B, lane 3) compared to cells that were not depleted of Rab27a (Fig 9B, lanes 1 and 2). Quantitation of the abundances of the endogenous and p3-loop miR-122 molecules revealed that endogenous miR-122 abundance was diminished by 30% in Rab27a-depleted cell (Fig 9C), a finding that is consistent with the result in Fig 7A and 7B. In contrast, the abundance of p3-loop miR-122 was not affected by Rab27a depletion (Fig 9C). Therefore, the 50% decrease in HCV-G27G42 RNA abundance in Rab27a-depleted cells in the presence of p3-loop miR-122 mimetics (Fig 9B, lane 3), is independent of the interaction of p3-loop miR-122 with the 5’ end of HCV RNA. This findings argue that endogenous miR-122, but not p3-loop miR-122, downregulates the expression of an inhibitor of HCV RNA gene expression.
CD81-containing exosomes are multivesicular body-derived microvesicles found in eukaryotic cells and are involved in cell-to-cell communication. It has been shown that both mRNAs and miRNAs can be transferred into neighboring cells by this pathway [33], and that HCV RNA can also be secreted from infected cells by extracellular vesicles [12, 14, 21–23]. However, extracellular vesicles, including exosomes, can be derived from several distinct pathways. To test whether HCV RNA and miR-122 are secreted by bona-fide exosomes, Rab27a that modulates the docking of multivesicular bodies to the plasma membrane [16] was depleted by siRNAs. Indeed, depletion of Rab27a led to a decrease of CD81- and CD63-positive exosome secretion in Huh7 cells (S1 Fig). Previous studies have argued that viral RNA can be transferred by “exosomes” [12, 14, 21–23], which were isolated from supernatants of cultured cells by subsequent centrifugation steps and CD81 affinity chromatography. In contrast, we show here that depletion of exosomes by genetic downregulation of the exosome docking protein Rab27a lowered both the intracellular and extracellular abundance of HCV RNA and virions (Fig 2), arguing that microvesicles other than exosomes are the major vehicles for the transport of viral RNA and virions.
The effect of Rab27a siRNA-3, which targets the 3’ noncoding region of all Rab27a mRNAs, on HCV RNA abundance could not be restored by overexpressing a knockdown-resistant Rab27a variant. We also found that overexpression of Rab27a did not increase HCV RNA and extracellular exosome abundance. Thus, Rab27a may affect HCV RNA abundance and exosome secretion as part of a protein complex. Alternatively, siRNA-3 caused off-target effects that were unrelated to Rab27a. To examine the latter possibility, additional siRNAs targeting different regions of Rab27a mRNAs were tested. All siRNAs showed a decrease in HCV RNA abundance, supporting the specificity of Rab27a’s effect on HCV RNA abundance (S3 Fig). Importantly, all siRNAs directed against Rab27a did not affect cell viability.
Studies with HCV replicons provided genetic evidence that Rab27a modulates the rate of viral replication (Fig 3). To further substantiate this finding with a biochemical approach, we examined the protein composition of membranes, which are sites for viral RNA replication. A substantial amount of Rab27a located to membrane-enriched fractions, both in uninfected and infected cells (Fig 4). In addition, confocal microscopy studies revealed that Rab27a localizes to LDs in uninfected and infected cells (Fig 5). Curiously, Rab27a coats LDs, visualizing the Rab27a-LDs complex as a doughnut-shaped structure. LD-associated Rab27a colocalized with viral core protein and with a small fraction of NS3. It has been proposed that HCV core recruits ER-derived membrane webs that are close to LDs to create a local membrane environment for viral replication and assembly [34, 35]. While the exact mechanism by which Rab27a modulates HCV RNA abundance is not clear at present, our findings strongly argue that Rab27a regulates HCV RNA abundance at LDs.
It is known that the presence of miR-122 is essential to maintain HCV RNA abundance. Profiling of several microRNAs in Rab27a-depleted cells showed that the abundance of miR-122 was decreased both in uninfected and infected cells. The loss of HCV RNA abundance during Rab27a depletion could be rescued by overexpression of miR-122 mimetics, which is consistent with the hypothesis that Rab27a-mediated depletion of miR-122 caused loss of HCV RNA abundance (Fig 8).
No significant decrease in the amount of primary miR-122 was observed in Rab27a-depleted cells, indicating that the effect of Rab27a depletion on miR-122 most likely occurred at a post-transcriptional step in the cytosol. Indeed, Rab27a depletion caused a decrease of ectopically expressed precursor miR122 (S9 Fig). It has been reported that both pre-microRNAs and mature microRNAs can be released from cells via exosomes that contain the GW182 component of the RNA-induced silencing complex (RISC) [36, 37]. This observation raises the possibility that depletion of Rab27a enhances the intracellular abundance of GW182-containing vesicles that affect the stability of pre-miR122 or miRNA-122 molecules. However, depletion of Rab27a effector Slp4 did not affect miR-122 and HCV RNA abundances. Because depletion of Slp4 inhibits exosome trafficking [16], loss of HCV RNA and miR-122 was not due to the accumulation of intracellular exosomes. We hypothesize that pre-miR-122 is being destabilized in the absence of Rab27a by an as-of-yet unknown mechanism. We also noted a selective decrease of several microRNAs in infected cells. One explanation is that HCV infection causes a dispersion of Processing bodies, where microRNAs, microRNA-targeted mRNAs and Argonaute proteins are located [38, 39]. This dispersion may affect turnover of specific microRNAs in infected cells. Alternatively, HCV is known to sequester components of RISC, such as Ago2 and GW182, at the HCV 5’ end for maintaining viral genome stability [40]. As a consequence, RISC-free microRNAs may be more easily degraded [22, 41, 42]. It is important to note that both miR-122 and miR-22 are depleted in HCV-infected cells. However, only the depletion of miR-122 affects HCV RNA abundance (Fig 8), arguing that loss of HCV RNA abundance was not caused by an overall loss of microRNAs in infected cells.
We examined whether loss of miR-122 led to the accumulation of HCV-G27G42 RNA molecules that were vulnerable to exonuclease cleavage. Thus, we examined the abundance of HCV-G27G42 RNA that could be protected by ectopically expressed mutant miR-122 molecules in Rab27a-depleted cells. The abundance of HCV-G27G42 RNA that could interact with mutant miR-122, but not with endogenous miR-122, also decreased in Rab27a-depleted cells (Fig 9). Because mutant miR-122 molecules very likely do not recognize mRNA targets that are modulated by wildtype, endogenous miR-122, effects of endogenous miR-122 on HCV RNA abundance are by a mechanism that is different from its protecting the 5’ end of the viral RNA. We also examined whether a miR-122 antagonist, instead of Rab27a depletion, caused a decrease in miR-122 to affect HCV replication that is independent of endogenous miR-122. We noted to our surprise that exogenously expressed mutant miR-122 mimetics cannot be functionally sequestered by the employed antagomirs. Thus, it is possible that an antagomir-inaccessible pool of mutant miR-122 accumulates within the transfected cell.
Finally, depletion of Rab27a has no effect on exoribonucleases Xrn1 and Xrn2 abundance. Thus, it is very likely that miR-122 downregulates an inhibitor of HCV gene expression. Such an inhibitor is not involved in the biosynthesis of cholesterol, because cholesterol abundance is not affected in Rab27a-depleted uninfected or infected cells.
Human hepatoma Huh7 cells were kindly provided by Francis V. Chisari (The Scripps Research Institute, San Diego). Huh7 cells were cultured in DMEM supplemented with 10% fetal bovine serum, 1x non-essential amino acids and 2 mM L-glutamine (Gibco).
All Small interfering RNA (siRNA) oligonucleotides and other RNA oligonucleotides were synthesized by Stanford PAN facility (Stanford, CA). The siRNA sequences are as follow: siControl, 5’- GAUCAUACGUGCGAUCAGAdTdT-3’; siRab27a-1: 5’- GGAGAGGUUUCGUAGCUUAdTdT-3’; siRab27a-2: 5’- GCCUCUACGGAUCAGUUAAdTdT-3’. The RNA oligonucleotide sequences are as follow: p3-loop miR-122: 5’- UGCAGUGUCUAUUUGGUCUUUGU-3’; p2-8 miR-122: 5’- UAAUCACAGACAAUGGUGUUUGU-3’. For formation of RNA duplexes, 50 μM of sense and antisense strands were mixed in annealing buffer (150 mM HEPES (pH 7.4), 500 mM potassium acetate, and 10 mM magnesium acetate) to a final concentration of 20 μM, denatured for 1 min at 95°C, and annealed for 1 h at 37°C.
Huh7 cells (106) were seeded in 10 cm tissue culture dishes. Cells were infected with wild-type JFH1 at a MOI of 0.01 for 5 h, washed with PBS to remove unbound virus, trypsinized and replated in 15 cm tissue culture dishes. The supernatant was collected at 3 days post-infection and centrifuged at 1,000 rpm, 10 min at 4°C to remove cell debris. The infected cells were scraped and resuspended in medium and subjected to freezed-thraw cycles. Samples were centrifuged at 1,000 rpm, 10 min at 4°C to remove cell debris. For the virus stock, the supernatant was mixed with cell-associated virus. Virus was stored in aliquots at -80°C. Virus titter was determined by using fluorescent focus-forming assay.
Huh7 cells (2.5 x 105) were seeded in 60 mm tissue culture dishes. Cells were transfected the following day with 50 nM of siRNA duplexes (25nM siRab27a-1 plus 25 nM siRab27a-2) using Dharmafect I reagent (Dharmacon) according to the manufacturer’s instruction. After 24 h post-transfection, the cells were infected with HCV JFH-1 virus at a MOI of 0.01 at 37°C. After 5 h incubation, cells were washed with PBS to remove unbound virus, trypsinized and replated in duplicate tissue culture dishes. Virus-infected cells were transfected again with 50 nM of siRNA duplexes at day 1 post-infection, and harvested at day 3 post-infection. The efficiency of siRNA depletion was evaluated by Northern and Western blot analysis.
Huh7 cells were washed once with PBS and total RNA was extracted using TRIzol (Invitrogen) following the manufacturer’s protocol. Ten μg of total RNA in RNA loading buffer (32% formamide, 1x MOPS-EDTA-Sodium acetate (MESA, Sigma) and 4.4% formaldehyde) was denatured at 65°C for 10 min and separated in a 1% agarose gel containing 1x MESA and 3.7% formaldehyde. The RNA was transferred and UV crosslinked to a Zeta-probe membrane (Bio-Rad). The membrane was hybridized using the ExpressHyb hybridization buffer (Clontech) or ULTRAhyb (Ambion) and α-32P dATP-RadPrime DNA labelled probes (Invitrogen) complementary to HCV (nucleotides 84–374), Rab27a (nucleotides 664–1145), or actin (nucleotides 685–1171). Autoradiographs were quantified using ImageQuant (GE Healthcare).
Ten μg of total RNA was separated in 12% acrylamide/ 7 M urea gel. Small RNAs were transferred onto a Hybond-N+ membrane (GE Healthcare), and detected by γ-32P-end labelled DNA probes complementary to miR-122, miR-16, miR-21, miR-22, miR-26, miR-130a, mutant miR-122 or U6 snRNA. Oligonucleotide sequence of probes are: miR-122 probe, 5’-CAAACACCATTGTCACACTCCA-3’; miR-16-5p probe, 5’-CGCCAATATTTACGTGCTGCTA-3’; miR-21 probe, 5’-TCAACATCAGTCTGATAAGCTA-3’; miR-22-3p probe, 5’-ACAGTTCTTCAACTGGCAGCTT-3’; miR-26a-5p probe, 5’- AGCCTATCCTGGATTACTTGAA-3’; miR-130a-3p probe, 5’- ATGCCCTTTTAACATTGCACTG-3’; U6 probe, 5’-CACGAATTTGCGTGTCATCCTTGC-3’. The membrane was hybridized using 7.5 x Denhardt’s solution, 5 x SSPE, 0.1% SDS, 0.05 mg/ml tRNA. Autoradiographs were quantified using ImageQuant (GE Healthcare).
Cells were washed with PBS once and lysed in RIPA buffer (50mM Tris (pH8.0),150 mM NaCl, 0.5% sodium deoxycholate, 0.1% SDS, and 1% Triton X-100) containing Complete EDTA-free protease inhibitors (Roche) for 15 min on ice. The cell lysate was clarified by centrifugation at 14,000rpm for 5 min at 4°C. Forty μg of cell lysate was mixed with 2x SDS sample buffer (126 mM Tris HCl, 20% glycerol, 4% SDS and 10% β-mercaptoethanol, 0.005% bromophenol blue, pH 6.8), denatured at 90°C for 5 min and separated in a 10% SDS-polyacrylamide gel. Protein was transferred to a PVDF membrane (Millipore). The membrane was blocked with 5% non-fat milk in PBS-T and probed using primary antibody, followed by horse-radish peroxidase-conjugated secondary antibodies. The blot was developed using Pierce ECL Western Blot Substrate (Thermo Scientific) according to the manufacturer’s instructions, and exposed to Biomax Light Films. The following primary antibodies were used for western blot analysis: anti-Core (C7-50) (Abcam, ab2740), anti-Rab27a (Abnova), anti-GAPDH (Calbiochem CB1001).
Infectious titers were determined by measuring fluorescent focus forming units (FFU) [43]. Rab27a depleted cells were infected with JFH-1 virus. For extracellular virus, supernatant of the infected cells was collected at day 3 post-infection. To harvest cell-associated virus, infected cells were washed with PBS three times, collected into a new tube, and resuspended in 500 μl DMEM. The cells were frozen and thawed three times. Both extracellular and cell-associated supernatants were sedimented at 14,000 rpm, 4°C for 5 min to remove cell debris. The viral titer was determined by FFU assay. Briefly, 3.2 x 104 cells were seeded in a 48-well plate and incubated overnight. A serial dilution of virus stock was added to cells and incubated for 5 h at 37°C. The diluted virus supernatant was removed from cells. Cells were washed with PBS and replaced with fresh medium. At day 3 post-infection, infected cells were washed once with PBS and fixed with cold methanol/acetone (1:1). The level of HCV infection in the cells was analyzed by using a mouse monoclonal antibody direct against HCV core (Abcam) at 1:1000 dilution in 1% fish gelatin/PBS at 4°C overnight and an AlexFluor488- conjugated goat anti-mouse antibody (Invitrogen) at 1: 200 dilution at room temperature for 2 h. The fluorescent focus forming units were counted using a fluorescence microscope, and the viral titer was expressed as FFU per ml.
Cell culture supernatants were collected from infected Huh7 cells. HCV RNAs from the supernatant were isolated using TRIzol LS reagent (Inviterogen) following the manufacturer’s protocol. HCV transcripts were quantified using SuperScript III Platinum SYBR Green One-Step qRT-PCR kit (Invitrogen). The reactions were performed using the CFX connect Real-Time system (BIO-RAD). HCV transcript levels were determined by comparison to standard curves derived from in vitro transcribed HCV RNA. The primer sequences for JFH1 were, Fwd, 5’-TCTGCGGAACCGGTGAGTA-3’; Rev, 5’-TCAGGCAGTACCACAAGGC-3’.
The plasmid H77ΔE1/p7, containing a deletion of structural proteins E1-E2-p7 [44] was transcribed using the T7 MEGAscript kit (Ambion), according to the manufacturer’s protocol. A mutant HCV RNA (nucleotide 27 and 42 C to G change) from H77ΔE1/p7-S1+2:p3 was transcribed as described [3, 4, 39]. Huh7 cells were transfected with Rab27a siRNAs (50 nM) at day 1 and mutant miR-122 duplex (50 nM) at day 2. Subsequently, cells were electroporated with the mutant HCV RNA at day 3. Briefly, Huh7 cells in 10cm dishes were trypsinized, washed with PBS once, and then washed with the Cytomix buffer, and suspended in the Cytomix buffer (120mM KCl, 0.15 M CaCl2, 10mM K2HPO4, 25 mM HEPES, 2 mM EDTA, 5 mM MgCl2, pH7.6), containing 10 μg HCV RNA. The cells were electroporated in 0.4 cm Biorad cuvette at 900V, 25 μF, and ∞ resistance, then incubated at room temperature for 10 min and seeded in a new 10cm dish. The cells were transfected again with Rab27a siRNAs and mutant miR-122 duplexes at 1 day after electroporation and harvested 3 days after electroporation.
Subgenomic JFH1-Rluc and JFH1-Rluc-GND were kindly provided by Glenn Randall (University of Chicago). The replicon RNA was generated using the T7 MEGAscript kit (Ambion) according to manufacturer’s protocols. Huh7 cells in 6 well plates were transfected with control or Rab27a siRNA using the Dharmafect I reagent (GE Dharmarcon). After 1 day post-transfection, cells were transfected with 2 μg of replicon RNA in TransMessenger reagent (Qiagen) for 1 h, and replaced with complete medium according to manufacturer’s instructions. Cells were harvested at 1, 2, 4, 8, 12, 24, 36 and 48 hours. Luciferase activity from the sample was detected according to manufacturer’s instructions.
Membrane-enriched fractions were isolated using a modified protocol adapted from Schlegel et al. [45]. Briefly, cells were washed with cold PBS twice, scraped in PBS, and pelleted. Cells were suspended in hypotonic buffer (10 mM Tris (pH 8.0), 10 mM NaCl, 1 mM MgCl2, with complete protease inhibitor cocktail tablets (Roche) and 0.5 mM PMSF) for 10 min on ice and then homogenized for 50 strokes using a Dounuce homogenizer. The cell homogenate was centrifuged at 1,000 x g for 10 min to remove nuclei and unbroken cells. The supernatant was collected and salt concentration was adjusted by adding NaCl to a final concentration of 300 mM. The cytoplasmic extract was then layered on a 10% and 60% sucrose in 300 mM NaCl, 15 mM Tris-HCl (pH7.5), 15 mM MgCl2, and centrifuged at 26,000 rpm at 4°C in a SW41 rotor for 16 h. The viscous layer in the middle of the gradient was collected using a syringe. The sample was concentrated with a Nanosep 3K Omega centrifugal device (Pall Life Sciences). The sample was resuspended in 4 x SDS sample buffer and separated in a 10% SDS-polyacrylamide gel.
Uninfected and HCV-infected Huh7 cells were grown on 8-chambered coverglass slides (LabTek II chamber slides, Thermo Scientific) for 3 days. Cells were rinsed with PBS and fixed with 4% paraformaldehyde (Electron Microscopy Sciences) in PBS for 20 min at RT. Cells were then washed with PBS for 5 min twice, and permeabilized with 0.1% Triton X-100 in 1% fish gelatin (Sigma) in PBS (1% PBS-FG) for 5 min. Blocking incubation was performed in 1% PBS-FG for 10 min, 3 times, at RT. Cells were incubated with primary antibodies in 1% PBS-FG at 4°C overnight, washed with 1% PBS-FG for 10 min twice, and incubated with secondary antibodies for 2h at RT. To visualize lipid droplets, cells were by stained with BODIPY 493/503. After washing with 1% PBS-FG for 10 min twice, Hoechst 33258 dye (Sigma) in 1% PBS-FG was added and cells were incubated 5 min at RT. After two washes in 1% PBS-FG for 5 min each, the coverglass slides were embedded in Fluoromount-G (SouthernBiotech). Samples were imaged at RT (22°C) with a 20×/N.A.0.60 or a 63×/N.A.1.30 oil Plan-Apochromat objective on a Leica SPE laser scanning confocal microscope (Leica-microsystems). Images were processed with ImageJ (Ver. 1.48, NIH) using only linear adjustments of contrast and color.
The following antibodies and reagents were used for immunofluorescence staining. Primary antibodies: mouse anti-Rab27a (H00005873-M02, Abnova), goat anti-HCV core (2861, Virostat), goat anti-HCV NS3 (2871, Virostat). Secondary antibodies: Alex Fluor 555 conjugated donkey anti-goat IgG (H+L) (A-21432) and Alex Fluor 647 donkey anti-mouse IgG (H+L) (A-31571, Life technologies). Primary antibodies were used at 1:100 dilution and secondary antibodies were used at 1:200 dilution. Bodipy 493/503 was 1:100 dilution from 1 mg/ml stock (D3922, Invitrogen). Hoechst 33258 dye was 1:10,000 dilution from 2 mg/ml stock (Sigma).
Statistical analyses were performed with Prism 5 (GraphPad). A two-tailed paired Student’s t-test was employed to assess significant differences between two groups. Error bars represent standard error of the mean.
|
10.1371/journal.ppat.1006954 | HSV-1-induced disruption of transcription termination resembles a cellular stress response but selectively increases chromatin accessibility downstream of genes | Lytic herpes simplex virus 1 (HSV-1) infection triggers disruption of transcription termination (DoTT) of most cellular genes, resulting in extensive intergenic transcription. Similarly, cellular stress responses lead to gene-specific transcription downstream of genes (DoG). In this study, we performed a detailed comparison of DoTT/DoG transcription between HSV-1 infection, salt and heat stress in primary human fibroblasts using 4sU-seq and ATAC-seq. Although DoTT at late times of HSV-1 infection was substantially more prominent than DoG transcription in salt and heat stress, poly(A) read-through due to DoTT/DoG transcription and affected genes were significantly correlated between all three conditions, in particular at earlier times of infection. We speculate that HSV-1 either directly usurps a cellular stress response or disrupts the transcription termination machinery in other ways but with similar consequences. In contrast to previous reports, we found that inhibition of Ca2+ signaling by BAPTA-AM did not specifically inhibit DoG transcription but globally impaired transcription. Most importantly, HSV-1-induced DoTT, but not stress-induced DoG transcription, was accompanied by a strong increase in open chromatin downstream of the affected poly(A) sites. In its extent and kinetics, downstream open chromatin essentially matched the poly(A) read-through transcription. We show that this does not cause but rather requires DoTT as well as high levels of transcription into the genomic regions downstream of genes. This raises intriguing new questions regarding the role of histone repositioning in the wake of RNA Polymerase II passage downstream of impaired poly(A) site recognition.
| Recently, we reported that productive herpes simplex virus 1 (HSV-1) infection leads to disruption of transcription termination (DoTT) of most but not all cellular genes. This results in extensive transcription beyond poly(A) sites and into downstream genes. Subsequently, cellular stress responses were found to trigger transcription downstream of genes (DoG) for >10% of protein-coding genes. Here, we directly compared the two phenomena in HSV-1 infection, salt and heat stress and observed significant overlaps between the affected genes. We speculate that HSV-1 either directly usurps a cellular stress response or disrupts the transcription termination machinery in other ways with similar consequences. In addition, we show that inhibition of calcium signaling does not specifically inhibit stress-induced DoG transcription but globally impairs RNA polymerase I, II and III transcription. Finally, HSV-1-induced DoTT, but not stress-induced DoG transcription, was accompanied by a strong increase in chromatin accessibility downstream of affected poly(A) sites. In its kinetics and extent, this essentially matched poly(A) read-through transcription but does not cause but rather requires DoTT. We hypothesize that this results from impaired histone repositioning when RNA Polymerase II enters downstream intergenic regions of genes affected by DoTT.
| Transcription termination is an essential process in gene expression that is coupled to all parts of RNA metabolism including transcription initiation, splicing, nuclear export and translation (reviewed in [1, 2]). It results in the release of RNA polymerase II (Pol II) and the nascent transcript from the chromatin, determines the general fate of individual transcripts and plays a crucial role in limiting the extent of pervasive transcription of the genome. Herpes simplex virus 1 (HSV-1) efficiently modulates cellular RNA metabolism and both cellular and viral gene expression to facilitate lytic infection [3–9]. Using 4-thiouridine-(4sU)-tagging followed by sequencing (4sU-seq), we recently reported that lytic HSV-1 infection results in the disruption of transcription termination (DoTT) of the majority but not all cellular genes [10]. This was dependent on de novo protein synthesis and already became broadly detectable by 2-3h of infection, which is before the release of the first newly generated virus particles at around 4h post infection (p.i.). At 7-8h p.i., about 50% of all 4sU-seq sequencing reads mapping to the human genome originated from intergenic regions (compared to <10% in uninfected cells). Previously, we referred to transcription beyond poly(A) sites due to DoTT as ‘read-out’. As this term has led to confusion, we now use the term ‘read-through’ to refer to transcription that extends beyond poly(A) sites. Transcription into a downstream gene arising from read-through from an upstream gene is referred to as ‘read-in’. For more than half of expressed cellular genes, poly(A) read-through affected >35% of their transcription. Read-in transcription into downstream genes was responsible for the seeming induction of about 1,100 cellular protein-coding and non-coding genes late in infection. In addition, it resulted in chimeric transcripts spanning two or more genes as evidenced by intergenic splicing events that connect exons of neighboring cellular genes.
Subsequently, two other studies reported on the disruption of transcription termination in cellular stress responses and cancer [11, 12]. Transcription downstream of genes (DoG) was observed in the osmotic stress response in human neuroblastoma cells, which was independent of de novo protein synthesis but appeared to at least partially rely on inositol-1,4,5-trisphosphate receptor (IP3R) activation and calcium signaling [11]. In addition, pervasive transcription read-through was identified in renal cell carcinoma [12]. This was dependent on the loss of histone methyltransferase SETD2, consistent with the role of epigenetic factors in RNA processing. Similar to HSV-1 infection, novel RNA chimeras were observed. Invasion of oncogenes by polymerases that initiated at upstream genes indicated a novel link between aberrant expression of oncogenes and chimeric transcripts prevalent in cancer. Taken together, these findings raise important questions regarding the underlying molecular mechanisms and functional roles of DoTT/DoG transcription in HSV-1 infection, cellular stress responses and cancer.
DoG transcription during osmotic stress was identified by Vilborg et al. upon exposure to 80mM KCl for 1h (from now on referred to as ‘salt stress’) in a human neuroblastoma cell line (SK-N-BE(2)C) by RNA-seq on nuclear, RiboMinus-treated RNA [11]. This revealed about 2,000 human genes to be affected. In addition, DoG transcription was also observed following heat stress (44°C) [11]. Recently, Vilborg et al. also reported on DoG transcription upon oxidative stress and found significant similarities but also clear stress-specific differences between the three stressors [13]. In our primary study, we analyzed newly transcribed RNA purified using 4sU-seq in one hour intervals of the first 8h of lytic HSV-1 infection of primary human foreskin fibroblasts (HFF) (Fig 1A). Under these conditions, the HSV-1 infected cells only start to lyse around 16 to 24h of infection. This allowed us to directly assess and quantify the relative frequency of transcripts experiencing DoTT as well as the extent of read-through transcription occurring within one hour intervals during the first eight hours of infection [10]. Throughout this manuscript, we refer to HSV-1-induced disruption of transcription termination as ‘DoTT’ to differentiate it from stress-induced DoG transcription. It is important to note here that transcription in intergenic regions downstream of genes was almost exclusively observed on the sense strand in relation to the upstream gene. This clearly distinguishes read-through from the recently reported activation of antisense transcription of the host genome during lytic HSV-1 infection [14].
Although DoTT was much more prominent at late times (7-8h p.i.) of HSV-1 infection than in salt or heat stress, we wondered whether the two phenomena might reflect the same cellular mechanism. We thus performed a detailed comparison and characterization of HSV-1-induced DoTT and DoG transcription triggered by salt and heat stress using 4sU-seq in the same cell type, namely HFF. This showed clear similarities in read-through between HSV-1 infection and the different stresses but also clear context- and condition-specific differences. Furthermore, we performed ATAC-seq (transposase-accessible chromatin using sequencing [15]) to compare chromatin accessibility before and during HSV-1 infection and stress. Strikingly, HSV-1-induced DoTT was accompanied by a strong increase in chromatin accessibility downstream of the affected poly(A) sites, which essentially matched the region of read-through transcription. This did not cause but rather required DoTT as well as a high level of transcriptional activity into downstream genomic regions. Interestingly, this effect was specific to HSV-1 and not observed in salt or heat stress (up to 2h) indicating that other mechanisms by which HSV-1 perturbs RNA processing contribute to this unexpected gene-specific alteration in the host chromatin landscape.
To directly compare HSV-1-induced DoTT with DoG transcription during cellular stress responses, we performed 4sU-seq analysis (60min 4sU-tagging followed by RNA sequencing) of HFF exposed to either salt (80mM KCl) or heat stress (44°C) for 1 and 2h (see Fig 1B). Two biological replicates of each condition as well as 2 untreated samples for each stressor were analyzed. 4sU-seq data for the first 8h of HSV-1 infection in HFF were obtained from our previous study [10]. A visual inspection of mapped reads for marker genes with either strong (SRSF3, SRSF6) or no (GAPDH, ACTB) DoTT/DoG transcription already indicated a striking similarity between presence or absence of DoTT/DoG transcription in the three conditions (Fig 1C and 1D; Fig A in S3 File, links to UCSC genome browser sessions showing read coverage for all cellular genes and samples separately for both replicates can be found at www.bio.ifi.lmu.de/HSV-1). As previously reported for HSV-1 infection [10] (Fig 1E), the percentage of reads mapping to intergenic regions downstream of gene 3’ ends increased substantially during salt and heat stress in HFF (Fig 1F). Intergenic read counts were highest directly downstream of gene 3’ ends and gradually decreased with increasing distance to gene 3’ ends. Furthermore, downstream intergenic transcription occurred almost exclusively in the same orientation as the upstream gene in all conditions (Fig B in S3 File). The low levels of antisense reads downstream of genes increased with increasing distance from gene 3’ ends as a consequence of read-through transcription for genes expressed from the opposite DNA strand outside of the 100kb downstream window considered. The gradual decrease in read levels downstream of genes was not due to differences in the length of read-through between genes, but was also observed at the level of individual genes (Fig B in S3 File and Fig C in S3 File). It could be approximated reasonably well by a linear fit at least late in HSV-1 infection and at 2h salt and heat stress, but the slope of the linear fit differed between genes (Fig C in S3 File). As a consequence of this gradual decrease and in contrast to regular mRNAs, 3’ ends of poly(A) read-through transcripts are not clearly defined [10, 11]. As the extent of read-through for individual genes gradually increased throughout infection, read-through transcripts extended further and further downstream of the gene.
To compare the extent of DoTT/DoG transcription between the three conditions, we focused on the 9,404 protein-coding and lincRNA (long intergenic non-coding RNA) genes whose expression was well detectable (fragments per kilobase of exons per million mapped reads (FPKM) ≥1) in all uninfected/untreated 4sU-seq samples. We then applied our previous approach [10] of dividing expression in the 5kb downstream of genes by the gene expression (FPKM) value (see methods). This measure (denoted as percentage of downstream transcription) is independent of any normalization to sequencing depth, which is canceled out in the division. As 4sU-tagging provides newly transcribed RNA from defined intervals of infection and stress, the obtained ratios quantify the percentage of transcripts newly transcribed in this interval that experience poly(A) read-through. To avoid confounding effects due to transcription of neighboring genes, we only included genes separated from neighboring genes on the same strand by at least 5kb on either side (5,928 genes). Although the restriction to the first 5kb downstream of a gene is relatively arbitrary, using a larger window of e.g. 10kb resulted in highly correlated values of downstream transcription (Spearman correlation Rs>0.95) but would exclude an additional 737 genes (12.4%) from the analysis. To account for small levels of downstream transcription in uninfected and untreated cells (mean = 4.2% and 0.06%, respectively), we calculated read-through as the difference in the percentage of downstream transcription between infected/stressed and uninfected/untreated samples (see methods). Read-in was quantified in the same way by first quantifying transcription in the 5kb upstream of genes relative to gene expression and then subtracting levels in uninfected/untreated samples. Since our previous study indicated that genes with read-in were more prone to read-through, we only used genes for the comparative analysis with at most 10% read-in in both HSV-1 infection and salt and heat stress (3,682 genes, Table A in S1 File). With the exception of the first three hours of HSV-1 infection where DoTT was hardly detectable, read-through values were highly correlated between replicates (Fig D in S3 File; Rs≥0.85).
The induction of DoG transcription upon salt and heat stress was reflected in median read-through levels of 6 to 15% (Fig 2A; Fig D in S3 File for individual replicates). Consistent with the recent report by Vilborg et al. [13], global read-through levels peaked at 1h of salt stress, but required 2h to reach comparable levels in heat stress. At the highest level, read-through in both salt and heat stress was comparable to read-through at 4-5h post HSV-1 infection, but considerably lower (~3-fold) than at the end of our HSV-1 infection time-course (7-8h p.i.). Median read-through levels in all conditions were highly correlated (Rs = 0.99) to the overall perturbation of gene expression (measured as standard deviation of FPKM log2 fold-changes; Fig 2B). Here, results for salt and heat stress fitted very well to a curve estimated from our HSV-1 time-course. At single gene level, however, read-through showed only a weak positive correlation with fold-changes in gene expression for HSV-1 infection (after the first 3h), salt and heat stress (Fig E in S3 File; Rs≤ 0.37). Vilborg et al. [13] also only found weak correlations between fold-changes in DoG transcription and fold-changes in expression of the respective genes (Rs = 0.12). The even lower correlations observed by Vilborg et al. may be explained by their use of nuclear RNA, which also contains RNA produced before stress. This underestimates gene expression changes for genes with low basal RNA turnover [16]. It should be noted that gene expression fold-changes estimated from RNA-seq data (even after normalization to house-keeping genes as performed here) only indicate changes in the relative, but not absolute, abundance among all expressed genes. As the overall transcription levels decline during lytic HSV-1 infection [17], positive fold-changes do not necessarily indicate actual transcriptional induction but only less down-regulation compared to other genes.
In our previous study, we reported that DoTT-induced read-through was increased for genes without the canonical AAUAAA poly(A)-signal upstream of the gene 3’end. Similarly, Vilborg et al. found several 6-mers to be depleted (including AAUAAA) or enriched downstream of genes with pan-stress DoG transcription. However, their analysis focused on the total frequency of the 6-mers downstream of all pan-stress DoG genes instead of the frequency for individual genes. We now aimed to identify 6-mers whose abundance in the 100nt up- or downstream of individual gene 3’ends was significantly correlated to read-through (FDR adjusted p-value <0.0001 for at least one condition or time-point, see methods). Strikingly, AAUAAA was the only 6-mer whose abundance upstream of gene 3’ends was significantly correlated with read-through in both stresses and HSV-1 infection (Fig 2C) and its absence upstream of gene 3’ ends was associated with significantly higher read-through (Wilcoxon rank sum test, p<0.0001; Fig 2D). Other 6-mers were only significantly correlated to read-through in HSV-1 infection and showed no significant differences in read-through in salt or heat stress (Fig F in S3 File). Upstream of gene 3’ends, negative correlations were found for a 6-mer overlapping the AAUAAA sequence as well as two C-rich motifs. Downstream of gene 3’ends, this included a number of G-rich motifs. Only one motif downstream of genes was positively correlated to read-through (AUUUUU), but only in HSV-1 infection. This sequence resembles binding motifs of a number of RNA binding proteins [18, 19], including HNRNPC (Heterogeneous Nuclear Ribonucleoprotein C), which has been shown to influence poly(A) site usage.
To directly compare HSV-1-induced DoTT to DoG transcription, we calculated Spearman rank correlations of read-through values between each pair of conditions and time-points. This compares the ranking of genes with regard to read-through, i.e. whether top- and lowest-ranked genes tend to be the same between samples. Read-through mostly correlated extremely well (Rs>0.8) between adjacent time-points for the same condition apart from the first three hours of HSV-1 infection where DoTT was hardly noticeable (Fig 3A). Moderate but comparable correlations were observed between salt stress and either heat stress or HSV-1 infection at 4-5h p.i. (Rs = 0.45-0.51). In contrast, read-through in heat stress was slightly better correlated to salt stress than to HSV-1 infection (Rs = 0.4). Since we observed a weak correlation between read-through and gene expression fold-changes in all conditions, we also calculated correlations after excluding genes with highest fold-changes (≥2 in any sample). This aimed to exclude genes for which differences in read-through between conditions might be explained by changes in transcriptional activity. However, correlations for the remaining 2,601 genes did not increase, which is probably explained by the observation that gene expression fold-changes were also well correlated (Fig G in S3 File). Thus, differences between conditions in DoTT/DoG transcription cannot be explained by differential alterations in transcriptional activity.
Next, we performed hierarchical clustering of genes based on read-through (average of replicates) for each condition (Fig 3B). This identified a large cluster of 1,368 genes (37%) with read-through in all conditions (marked in blue) as well as a number of clusters with differences between conditions. It furthermore highlighted the prevalence of DoTT/DoG transcription with only 102 genes (3%) showing no DoTT/DoG transcription (defined as ≤5% read-through) in any infected/stress sample. Overrepresentation analysis for Gene Ontology (GO) terms using DAVID [20] found an enrichment of genes with extracellular regions (25 genes) and heparin binding (6 genes) among these 102 genes. However, no functional categories were overrepresented for the 1,368 genes with read-through in all conditions.
Interestingly, the only gene experiencing ≥75% read-through already after 2-3h p.i. HSV-1 infection and in all stress conditions was interferon regulatory factor 1 (IRF1) (Fig H in S3 File). IRF1 is an important mediator of both type I and II interferon signaling and studies with IRF1-deficient mice have shown an important role for IRF1 in the immune response against viruses [21–23]. Furthermore, even a relatively small reduction in IRF1 expression, e.g. mediated by cellular miR-23a, is sufficient to measurably augment HSV-1 replication in cell culture [24]. Notably, ribosome profiling data from our previous study revealed a >4-fold drop in IRF1 translation during HSV-1 infection despite an >1.8-fold increase in total RNA at 8h p.i. [10]. This presumably reflects the negative effects of DoTT on IRF1 translation and suggests that HSV-1 exploits DoTT to evade the host immune response.
A striking characteristic of HSV-1-induced DoTT was the associated increase in aberrant splicing [10]. In particular, this comprised novel intragenic and intergenic splicing events as well as splicing associated with nonsense-mediated decay (NMD). Intergenic splicing joins known exons of neighboring genes and confirms transcription of large chimeric transcripts spanning two or more cellular genes. It can be observed as early as 3-4h p.i. in HSV-1 infection. One of the most prominent examples connects SRSF2 and JMJD6. We also observed intergenic splicing in the two stress conditions, but the few examples did not cluster with intergenic splicing events in HSV-1 infection (Fig 3C). Analysis of induced splicing events upstream of gene 3’ ends, however, showed similar characteristics in both HSV-1 infection and salt and heat stress. In all three conditions, induced intragenic splice junctions were enriched for novel splice junctions and junctions found only in processed transcripts (containing no ORF but not classified as long or short non-coding RNAs) or in NMD-associated transcripts (Fig 3D; examples in Fig I in S3 File). Genes with induced intragenic splicing events showed increased read-through in all three conditions (Fig I in S3 File), but read-through was also observed in genes without induced splicing events. Thus, aberrant splicing upstream of gene 3’ ends more likely resulted from, rather than is responsible for DoTT/DoG transcription. One possible explanation for the association of aberrant splicing with DoTT/DoG transcription may be that all serine and arginine rich splicing factor (SRSF) genes included in our analysis (SRSF2, SRSF3, SRSF5, SRSF6, SRSF7, SRSF10, SRSF11) showed DoTT/DoG transcription in at least two, but mostly all three conditions. All of these SRSF genes showed a >2-fold greater drop in translation at 8h p.i. HSV-1 infection in the ribosome profiling data than expected from the changes in their total RNA levels.
Vilborg et al. reported that salt stress-induced DoG transcription in SK-N-BE(2)C cells depends on IP3R activation, Ca2+ release from intracellular stores and downstream kinases [11]. HSV-1 entry into cells is dependent on the activation of Ca2+ signaling pathways and triggers Ca2+ release from intracellular stores [25, 26]. In addition, HSV-1 infection results in an increasing loss of stable, resting Ca2+ at late times of infection indicating a bimodal role of Ca2+ signaling in HSV-1 infection [27]. Before assessing the effect of Ca2+ signaling inhibitors on DoTT in HSV-1 infection of HFF, we first aimed to reproduce the results by Vilborg et al. in salt stress. HFF were exposed to 80mM KCl for 1h in presence of (i) an inhibitor of IP3R signaling (2-APB), (ii) the membrane permeable Ca2+ chelator BAPTA-AM, or (iii) inhibitors of the downstream kinases Ca2+/calmodulin-dependent protein kinase II (CaMKII) and protein kinase C/protein kinase D (PKC/D) (KN93 and Gö6976, respectively). DoG transcription was first quantified by qRT-PCR on total RNA for DDX18, which shows strong read-through in HSV-1 infection as well as salt and heat stress. Consistent with the previous report, BAPTA-AM prevented DoG transcription while the other inhibitors resulted only in a moderate (25–65%) reduction (Fig 4A). We thus aimed to assess the effect of BAPTA-AM on DoTT in HSV-1 infection. To avoid the described detrimental effects of BAPTA-AM on virus entry and the onset of productive infection [25, 26], we only added BAPTA-AM to the cell culture media of HFF at 1h p.i. (MOI = 10) when viral gene expression is already well initiated. To first determine its effect on viral gene expression, we quantified immediate-early (ICP0), early (ICP8) and true late (ICP5) gene expression at 8h p.i. by qRT-PCR. Strikingly, BAPTA-AM treatment was highly detrimental to viral gene expression of all three kinetic classes resulting in a >1,000-fold drop in viral mRNA levels (Fig 4B).
Considering this strong reduction in viral gene expression, we hypothesized that depletion of intracellular Ca2+ by BAPTA-AM in HFF might globally impair Pol II activity rather than specifically interfere with DoTT/DoG transcription. We thus analyzed the effect of 1h of BAPTA-AM treatment of uninfected cells on transcriptional activity of three cellular genes (SRSF3, IRF1 and DDX18). For this purpose, we labeled newly transcribed RNA by adding 500μM 4sU to the cell culture medium for 1h. Following isolation and purification of the 4sU-labeled newly transcribed RNA (4sU-RNA) from a fixed amount of biotinylated total RNA per condition (60μg), transcriptional activity of these genes was quantified using qRT-PCR on 4sU-RNA. BAPTA-AM indeed induced a drop in transcriptional activity that was at least as strong as observed upon inhibition of Pol II using actinomycin D (Act-D; Fig 4C). In addition, global 4sU incorporation rates into total cellular RNA were substantially reduced upon BAPTA-AM treatment (Fig 4D). This indicated that BAPTA-AM might not only interfere with Pol II but also with rRNA synthesis (Pol I and III transcription), which contributes about 50–60% of 4sU-RNA in HFF as estimated from our RNA-seq data [10]. We thus quantified transcription rates from 4sU-RNA for a Pol I transcript (18S rRNA), a Pol III transcript (5S rRNA) in addition to four genes transcribed by Pol II (GAPDH, SRSF3, IRF1 and DDX18) upon exposure of HFF to 80mM KCl for 1 and 2h and BAPTA-AM (Fig 4E). In addition, we tested whether the combined exposure of cells to Gö6976 and KN93, which also diminished salt stress-induced DoG transcription in total RNA, also globally affected transcriptional activity. While salt stress alone already resulted in a drop in transcription rates for Pol I (≈1.5-fold), II (3- to 5-fold) and III (≈1.4-fold) transcripts, BAPTA-AM impaired transcriptional activity of all three polymerases. This suggests that global inhibition of cellular RNA polymerases by BAPTA-AM rather than a specific effect on transcription termination is responsible for the loss of salt stress-induced DoG transcripts. As BAPTA and its derivatives share a high selectivity for Ca2+ over Mg2+ (>105 stronger binding), the observed effects did not result from the co-depletion of intracellular Mg2+ [28]. Interestingly, combined Gö6976/KN93 treatment also globally impaired Pol I, II and III transcription, albeit to a lesser degree (2- to 10-fold), thereby explaining the slight reduction in DoG levels in total RNA (Fig 4A). In contrast, 2-ABP treatment, which had shown no effect on DoG transcription when analyzing total cellular RNA, did not impair polymerase activity. Finally, we quantified read-through transcription for the three DoG genes SRSF3, IRF1 and DDX18 in 4sU-RNA (Fig 4F). Neither KN93/Gö6976 nor 2-ABP treatment had any effect on the induction of the respective DoG transcripts. Unfortunately, BAPTA-AM treatment did not allow to reliably measure read-through transcription due to the impaired transcription (very low copy numbers or even negative PCR results). We conclude that the reduced levels of DoG transcripts upon inhibition of Ca2+ signaling do not result from direct effects on DoG transcription but from global inhibitory effects on cellular transcription in general. To our knowledge, this strong inhibitory effect of BAPTA-AM treatment on RNA polymerase activity has not been appreciated so far and should be considered when interpreting results obtained using BAPTA-AM to inhibit calcium signaling.
Vilborg et al. initially reported that DoG transcripts (DoGs) were strongly enriched at the chromatin [11] and that one of the more abundant DoGs, doSERBP1 (downstream of SERBP1), remained at the site of synthesis. However, they subsequently also observed DoGs in the nucleoplasma of cells when searching for them by confocal microscopy with increased sensitivity [13]. To assess the fate of the transcripts arising from DoTT in HSV-1 infection, we separated cell lysates (uninfected cells and 8h p.i.) into cytoplasmic, nucleoplasmic and chromatin-associated fractions [29, 30] and analyzed all three fractions as well as total cellular RNA by RNA-seq (2 replicates). The efficient separation of the cytoplasmic from the nuclear RNA fraction was confirmed by the enrichment of well-described nuclear lincRNAs (MALAT1, NEAT1, MEG3; Fig J in S3 File) in nucleoplasmic and chromatin-associated RNA as well as cytoplasmic enrichment of reported cytoplasmic lincRNAs (LINC00657, VTRNA2-1; Fig J in S3 File). In addition, overrepresentation of intronic reads in chromatin-associated RNA compared to nucleoplasmic RNA (>5-fold higher) demonstrated the efficient separation of these two RNA fractions (Fig J in S3 File)
In uninfected cells, only chromatin-associated RNA showed notable levels of downstream transcription (median 7.2%; Fig 5A), consistent with the standard model of transcription termination in eukaryotic cells [1]. At 8h p.i., substantial read-through was observed in all fractions except for cytoplasmic RNA (Fig 5B, Table B in S1 File), indicating that read-through transcripts are not efficiently exported to the cytoplasm. When we grouped genes according to their extent of read-through in 7-8h p.i. 4sU-RNA, we observed a strong increase during infection in the enrichment of the respective mRNAs (counting only the exonic regions upstream of gene 3’ ends) in both nucleoplasmic (Fig 5C) and chromatin-associated RNA (Fig J in S3 File) depending on the extent of read-through. While no change in nuclear enrichment was observed for genes without read-through, genes with >75% read-through were on average >2.5-fold more enriched at 8h p.i. than in uninfected cells. In particular, IRF1 was >6 and 4-fold more enriched in nucleoplasmic and chromatin-associated RNA, respectively, at 8h p.i. than in uninfected cells. Further evidence for an inefficient export of read-through transcripts is provided by intergenic splicing events, which are mostly absent in cytoplasmic RNA at 8h p.i. despite their considerable abundance in the other subcellular RNA fractions (Fig 5D). This also explains our previous observation based on ribosome profiling that RNA chimeras and consequently genes induced by read-in transcription arising from DoTT are not, or only poorly translated [10]. We conclude that DoTT leads to nuclear retention of the respective read-through transcripts and thereby notably contributes to HSV-1 induced host shut-off.
The similar overall level and high gene-specific correlation (Rs = 0.8) of read-through in nucleoplasmic and chromatin-associated RNA indicates that transcripts resulting from HSV-1-induced DoTT are generally released from the chromatin, i.e. the site of synthesis, into the nucleoplasm (see e.g. Fig 5F). Nevertheless, we identified 18 genes (Table C in S1 File) for which these transcripts appeared to remain at the chromatin (≤5% read-through in nucleoplasmic and cytoplasmic RNA, but ≥25% in chromatin-associated RNA; examples in Fig K in S3 File).
Interestingly, there was a modest correlation (Rs = 0.32-0.53) between the percentage of downstream transcription observed in chromatin-associated RNA of uninfected/unstressed cells and read-through upon stress or HSV-1 infection (Fig 5E). This suggests that genes with a relatively high extent of downstream transcription in uninfected/unstressed cells might be predisposed for DoTT/DoG transcription. To exclude that this was an artifact of read-through being calculated from downstream transcription, we calculated ‘mock’ read-through values from the two biological replicates for the same time-point (see methods). For mock read-through, the correlation was much lower at only ~0.13. This suggests a link between downstream transcription detectable in chromatin-associated RNA in uninfected/untreated cells and read-through in stress/infection. A possible explanation might be that the respective poly(A) sites are weaker and thus more prone to further disruption by HSV-1 or stress-related mechanisms. Fig 5F illustrates this for IRF1, for which downstream transcription in chromatin-associated RNA of uninfected cells was 14% and covered ~5kb. Interestingly, the correlation between downstream transcription in chromatin-associated RNA in uninfected cells and read-through during infection was highest at early time-points, i.e. at 1h for salt/heat stress and 2-3h p.i. for HSV-1 infection (Fig 5E and 5G). At late stages of HSV-1 infection, even cellular genes with very little downstream transcription in chromatin-associated RNA from uninfected cells showed read-through transcription (Fig J in S3 File).
Based on publicly available DNase hypersensitive and ATAC-seq data for unstressed murine fibroblasts, Vilborg et al. recently reported that, even prior to stress, pan-DoG genes are already characterized by a chromatin signature indicative of an open chromatin state. However, due to the lack of respective data following salt or heat stress, they could not assess the consequences of read-through on cellular chromatin. We thus performed ATAC-seq in HFF at 0, 1, 2, 4, 6 and 8h of HSV-1 infection and 1 and 2h of salt and heat stress (n = 2). For all ATAC-seq samples, open chromatin regions (OCRs) were enriched around promoters, thereby confirming the high quality of the data (Fig L in S3 File). Both length and score of OCRs at gene promoters correlated with gene expression in uninfected cells (Rs = 0.42 and 0.4, respectively; Fig L in S3 File). In contrast to the findings by Vilborg et al., we did not observe a positive correlation between DoTT/DoG transcription and the presence of OCRs in the 5kb downstream of genes in unstressed/uninfected cells (Fig M in S3 File). However, we noted a weak positive correlation (Rs≤0.25) between the presence of downstream OCRs (dOCRs) and the expression level of the corresponding genes (Fig M in S3 File). Notably, the highly expressed genes GAPDH and ACTB, which were not affected by DoTT/DoG transcription (Fig 1D; Fig A in S3 File), were characterized by open chromatin downstream of their 3’ends already in uninfected cells (Fig M in S3 File). In summary, our data argues against genes being predisposed for DoTT/DoG transcription by open chromatin downstream of their 3’ ends.
We next analyzed the impact of HSV-1-induced DoTT on chromatin accessibility. To our surprise, we observed a substantial increase in open chromatin downstream of individual genes with HSV-1-induced DoTT (Fig 6A, Fig N in S3 File). Here, downstream regions were often covered by OCRs for tens-of-thousands of nucleotides, similar to the pattern of read-through transcription in these downstream regions. This already became detectable at 4h p.i and resulted in a substantial increase in the number of long OCRs (Fig 6B), which were specifically enriched downstream of genes (Fig O in S3 File). Thus, these do not result from global effects of HSV-1 infection on cell viability (e.g. due to enhanced chromatin accessibility in a subpopulation of dying cells). To quantify the total extent of open chromatin downstream of individual genes, we assigned dOCRs to genes if they were either close to the gene 3’ end or another dOCR that had already been assigned to the respective gene (see methods) and then calculated the total genomic length covered by dOCRs (= dOCR length). This revealed a specific increase of dOCR length throughout infection for genes with high read-through (Fig 6C). For 174 of the 681 genes (26%) with >80% read-through at 7-8h p.i., dOCR length exceeded 5kb at 6h p.i., while only 26 of 326 (8%) genes with ≤5% read-through at 7-8h p.i. had a dOCR length ≥5kb at 6h p.i. (Fisher’s exact test p = 6.71×10-12). For 11 of these 26 genes (42%), this was likely due to a close-by downstream gene with DoTT on the opposite strand (see Fig 6D for FBN2, >60kb dOCR matches the read-through of the SLC12A2 gene on the opposite strand). These 11 genes showed no DoTT despite long dOCRs (originating from DoTT for genes with convergent transcription on the opposite strand) and strong expression at 7-8h p.i. (10 with FPKM >1, 6 with FPKM >3). This indicates that the increase in downstream open chromatin during HSV-1 infection is not responsible for DoTT but rather that the formation of dOCRs requires DoTT. Furthermore, induction of long dOCRs for genes with read-through was dependent on the transcription rates of the respective genes. Genes with >80% read-through and long dOCRs were much higher expressed at 7-8h p.i. than read-through genes without long dOCRs (Fig 6E). Accordingly, when dOCR length was compared to read-through for the 1,273 most highly expressed genes (FPKM ≥2) at 7-8h p.i., the difference in dOCR length between genes with different read-through levels was much more pronounced (Fig 6F). Finally, strong increases in OCRs within gene bodies or promoter regions were only observed for genes with read-in transcription but not upstream of the poly(A) read-through. This explains the smaller, but nevertheless notable global increase in long OCRs in gene bodies (Fig O in S3 File).
Given the striking increase in dOCR length for well-expressed genes affected by HSV-1 induced DoTT, we also expected to see an increase in chromatin accessibility for salt and heat stress. However, there was no general increase in the number of long OCRs during salt or heat stress and no increase in dOCR length for individual genes in contrast to HSV-1 infection (Fig P in S3 File). Accordingly, dOCR length did not increase for genes with high levels of stress-induced DoG transcription (Fig P in S3 File), not even for highly expressed genes (Fig 6A, Fig N in S3 File). Since read-through at 2h salt and heat stress was comparable to 4-5h p.i. HSV-1 infection and extensive dOCRs were clearly detectable at 4h p.i., stress-induced DoG transcription does not appear to lead to open chromatin downstream of genes. Thus, only HSV-1 induced DoTT, but not DoG transcription in salt or heat stress, results in this striking increase in the accessibility of genomic regions downstream of affected genes.
In addition to enhanced chromatin accessibility downstream of pan-stress DoGs, Vilborg et al. also found an enrichment of several histone marks typically found at actively transcribed genes (H3K36me3, H3K79me2) and at enhancers (H3K4me1, H3K27ac) based on ENCODE data from unstressed murine NIH-3T3 fibroblasts [13]. Considering the discrepancy of our findings regarding open chromatin to their findings, we also analyzed ChIP-seq data from ENCODE for histone marks in uninfected/unstressed HFF. Significant positive correlations (FDR adjusted p-value <0.01) between read-through in stress conditions and the presence of histone marks in the 5kb downstream of genes were only observed for the elongation marker H3K36me3 and DoG transcription in heat stress (Fig M in S3 File). However, weak positive but not significant correlations were also observed in salt stress for H3K36me3. The same was true in both stresses for two markers of accessible regulatory chromatin, H3K27ac and H3K4me1. Interestingly, for H3K36me3, positive correlations were also observed to read-through already detectable in the first two hours of HSV-1 infection. However, at later times of infection, this shifted to highly significant negative correlations between read-through and the presence of H3K27ac, H3K27me3, H3K4me1 and H3K4me3 marks in the ENCODE data for uninfected cells. While this highlights important differences between DoTT and DoG transcription, the biological significance of the presence of certain histone marks in cells prior to stress or infection remains unclear. ChIP-seq data from time-course experiments of both HSV-1 infection and stress will be required to resolve these conflicting observations.
In collaboration with Wyler et al., we recently reported on the activation of antisense transcription in the human genome during lytic HSV-1 infection [14]. To assess whether this antisense transcription was also associated with the formation of OCRs, we investigated the 11 antisense transcripts that had been extensively validated by RT-qPCR and Nanostring nCounter assays (Fig Q in S3 File). Interestingly, induction of antisense transcripts was clearly accompanied by an induction of corresponding long OCRs in three of these cases (BBCas, EFNB1as, ING1as). These represented 3 of the 4 (together with C1orf159as) most highly expressed antisense transcripts at 7-8h p.i., consistent with a role of transcription in the formation of long OCRs. For another four cases, an effect on open chromatin was visible but less clear (NFKB2as, IFFO2as, FOXO3as, C1orf159as). Moreover, similar to transcripts of DoTT-affected genes, the length of the 11 antisense transcripts gradually increased quite substantially during HSV-1 infection, indicating that they are also affected by DoTT. To exclude that long OCRs during HSV-1 infection are an artifact of or are directly related to the induced antisense transcription, we determined the fraction of long OCRs (≥5kb; <80 long OCRs per replicate in uninfected cells, >500 per replicate at 6 and 8 p.i.) that overlapped (≥25% of OCR in antisense transcript) any of the 3,098 antisense transcripts identified by Wyler et al. (Fig R in S3 File). In uninfected cells, ≥40% of the few long OCRs overlapped with an antisense transcript. With increasing duration of infection, this fraction decreased and only ~13% of long OCRs overlapped an antisense transcript at 8h p.i., but often also a region of read-through transcription on the opposite strand. This supports a model in which HSV-1-induced dOCRs originate from read-through transcription while antisense transcripts also experience DoTT and consequently show the associated long OCRs if transcribed at a sufficient rate.
HSV-1 infection, cellular stress responses and cancer result in extensive transcriptional activity downstream of a subset of cellular genes [10–12], but the relationships between the underlying molecular mechanisms remained unclear. By directly comparing HSV-1-induced DoTT with DoG transcription in salt and heat stress in the same experimental setting, we show significant overlaps between the genes affected by DoTT/DoG transcription but also clear context- and condition-specific differences. Importantly, differences were not only observed between DoTT and DoG transcription but also for DoG transcription between the two different stresses. Notably, the gene-specific correlation of read-through between salt stress and heat stress essentially equaled the correlation between salt stress and HSV-1 infection at 4-5h p.i. Multiple cis- and trans-regulatory factors are known to determine both splicing and poly(A) site usage [31] and even promoter elements have been shown to shape RNA processing by influencing Pol II processivity [32]. Thus, variability in DoG transcription upon different stressors and HSV-1-induced DoTT may originate from differences in downstream responses, interactions with other signaling pathways activated upon the different stresses or infection or even activation of alternative pathways with similar molecular consequences on the transcription termination machinery. In any case, the striking similarities between DoTT and DoG transcription indicate that related mechanisms are at play during HSV-1 infection. While the extent of DoTT further increased at late times of infection, salt stress-induced DoG transcription already declined by 2h, presumably due to detrimental effects of prolonged exposure to enhanced extracellular K+ concentrations on the exposed cells. In this respect, the expression of viral proteins counteracting the consequences of detrimental stress responses such as translational arrest and apoptosis may enable the much more efficient disruption of transcription termination by HSV-1.
The results presented here and in our previous manuscript [10] could have been a result of transcriptional noise that becomes evident in the context of transcription inhibition or extensive degradation of actively transcribed mRNAs by the virion-associated host shut-off protein (vhs) [33]. Alternatively, it might result from de novo pervasive transcription initiation downstream of the respective genes. However, we disfavor these models. First, it is important to note that we analyzed newly transcribed rather than total RNA. Therefore, transcriptional activity downstream or upstream of genes is always directly compared to the transcriptional activity of the corresponding gene occurring during the same timeframe of infection. The global loss in Pol II activity should equally affect genomic regions within, downstream and upstream of genes. Furthermore, strong transcriptional down-regulation of hundreds of genes has been analyzed in a broad range of different conditions using 4sU-seq [34–36], none of which showed any increase in transcriptional activity downstream of genes. In addition, infection with a vhs-null mutant, which does not trigger a notable decline in transcriptional activity until at least 12h of infection [17], still resulted in a very similar extent of read-through transcription as wild-type HSV-1 infection [10].
The data obtained in this study provide further strong evidence that downstream transcriptional activity arises from DoTT. First, the high correlations between the extent of read-through in HSV-1 infection, salt and heat stress indicate that all three conditions involve a common mechanism, namely poly(A) read-through. Second, RNA-seq analysis of subcellular RNA fractions revealed a striking dependence of nuclear retention of exonic regions during infection on the extent of read-through observed for the respective genes. The most likely scenario is that DoTT and extensive poly(A) read-through transcription result in large aberrant transcripts that cannot be efficiently exported to the cytoplasm. Third, the induction of extensive dOCRs for genes experiencing DoTT, which depends on the transcription level of these genes, provides strong evidence for an increase in absolute transcriptional activity downstream of these genes during infection. Additional evidence against pervasive de novo transcription initiation downstream of genes is provided both by the intergenic splicing events between neighboring genes induced in HSV-1 infection and the strong strand-specificity of the downstream transcriptional activity. De novo transcription initiation would not be limited to the strand of the upstream gene but would be expected to occur on either strand. The strong strand-specificity also excludes that downstream transcriptional activity is an artifact of the reported activation of antisense transcription during infection [14]. Moreover, DoTT and read-through transcription is clearly much more prominent than this antisense transcription. We now even provide evidence that antisense transcripts are also affected by DoTT and show DoTT-associated dOCRs.
Vilborg et al. reported that DoG transcription was associated with an open chromatin state downstream of genes prior to stress [11]. This observation was not confirmed in our ATAC-seq data from primary human fibroblasts. While we currently cannot fully explain the discrepancy between these findings, we hypothesize that the enrichment of accessibility marks observed by Vilborg et al. may result from a restriction to pan-DoG genes detectable in nuclear RNA. As this also includes RNA transcribed before stress, relative levels of DoG transcripts are lower than in newly transcribed RNA. Thus, their analysis may be biased towards more highly transcribed DoG transcripts, which are more readily detectable. When we analyzed histone mark ChIP-seq data from ENCODE for uninfected HFF, we could only reproduce the positive correlation reported by Vilborg et al. [13] between the presence of the transcription elongation mark H3K36me3 (but not H3K4me1 and H3K27ac) and read-through in heat stress and to a lesser degree in salt stress. Interestingly, this was also observed during the first two hours of HSV-1 infection, which nicely fits to our observation that genes with active downstream transcription in chromatin-associated RNA in uninfected cells are more prone to read-through. In addition, it indicates that read-through occurring very early in HSV-1 infection may reflect a cellular stress response to infection and thus essentially DoG transcription. Later in infection, however, the picture completely shifts to negative correlations between read-through and repressive (H3K27me3) and general (H3K4me3) promoter marks as well as accessible regulatory chromatin (H3K27ac and H3K4me1). While the correlation with both repressive promoter marks and activating histone marks late in HSV-1 infection is difficult to interpret and seems contradictory, it hints that at this point other mechanisms than a general stress response may come into play. It is important to note, however, that the respective ChIP-seq data were only obtained from uninfected/unstressed cells and thus do not reflect the changes in histone marks upon infection/stress.
The most striking finding of our study is the extensive increase in genome accessibility downstream of well-expressed genes affected by DoTT during HSV-1 infection, which essentially matched the transcriptional read-through observed at the respective time of infection. Of note, the peak heights of extensive dOCRs were often similar to levels observed in gene promoters of the respective genes where histones are displaced by transcription factors binding to promoter elements. However, in DoTT-associated dOCRs, this was not restricted to a few hundred base pairs but extended for tens-of-thousands of nucleotides. Our data indicate that dOCRs are not the cause but rather the consequence of DoTT and their formation additionally requires high levels of transcriptional activity in the respective downstream genomic regions. Considering the high correlation between DoTT and DoG transcription, we were surprised not to observe any evidence of dOCRs for DoG transcription in salt or heat stress. As DoTT-associated dOCRs were already well detectable by 4h of infection when the overall extent of DoTT and DoG transcription was very similar, the lack of dOCRs in salt and heat stress is not merely due to quantitative differences between the three conditions.
Progression of transcribing Pol II across a gene is accompanied by the displacement of nucleosomes, followed by their rapid co-transcriptional repositioning immediately behind Pol II (reviewed in [37]). We hypothesize that dOCRs result from impaired histone repositioning in the wake of Pol II. The lack of dOCRs in salt and heat stress indicates that dOCRs do not merely arise when Pol II starts transcribing far into previously untranscribed regions of the genome. Furthermore, gene bodies upstream of poly(A) sites affected by DoTT showed no general induction of OCRs, suggesting that there is no general inhibition of histone repositioning during HSV-1 infection. However, induced OCRs were also observed in gene bodies following read-in transcription, which argues against a role of distinct histone modifications in intergenic regions. Interestingly, HSV-1 infection was found to mobilize histones including linker and core histones (H1, H2B, H3.1 and H4) as well as histone variants (H3.3) [38, 39]. This resulted in increases in the pools of “free” histones despite an inhibition of histone synthesis during infection [40, 41]. Therefore, it is unlikely that the induction of dOCRs results from a deprivation of free histones. On the contrary, the reported histone mobilization may at least partly result from impaired histone repositioning downstream of genes and thus release of histones from the respective regions into the nucleoplasm following read-through transcription. A critical role in nucleosome reassembly is played by the histone chaperons Spt6 and the FACT (FAcilitates Chromatin Transcription) complex (reviewed in [42]). Interestingly, recruitment of Spt6 to active cellular genes includes direct interactions with the C-terminal domain (CTD) of Pol II [43, 44]. Here, specific post-translational modifications of the CTD, which depend on the position of the transcribing Pol II within a gene, govern the functional state and properties of Pol II and its interactions with other factors [45]. Recently, the HSV-1 ICP22 protein was found to relocate both Spt6 and FACT to viral replication compartments. This may limit their availability to Pol II when transcribing cellular genes in HSV-1 infection [46], but does not explain the selective failure in nucleosome reassembly only downstream of genes with read-through. Follow-up studies on recruitment and disengagement of Spt6 and FACT from Pol II upon infection with wild-type HSV-1 and mutant viruses as well as the concurrent analysis of post-translational modifications of the Pol II CTD will provide important insights into the functional regulation of transcription by Pol II and its termination downstream of genes.
In summary, our findings provide a much more detailed picture of the molecular processes involved in DoTT/DoG transcription and point the direction for further studies to elucidate the underlying molecular mechanisms.
Human fetal foreskin fibroblasts (HFF) were purchased from ECACC and cultured in DMEM with 10% FBS Mycoplex and 1% penicillin/streptomycin. HFF were utilized from passage 11 to 17 for all high-throughput experiments. This study was performed using wild-type HSV-1 strain 17. Virus stocks were produced in baby hamster kidney (BHK) cells (obtained from ATCC) as described [10]. HFF were infected with HSV-1 24h after the last split for 15 min at 37°C using a multiplicity of infection (MOI) of 10. Subsequently, the inoculum was removed and fresh media was applied to the cells.
Salt stress was initiated by adding 80mM KCl to the tissue culture medium. Heat stress was started by replacing the cell culture medium with pre-warmed 44°C medium and culturing the cells for 1 or 2h at 44°C. Newly transcribed RNA was labeled for 1h using 500μM 4-thiouridine (Carbosynth). Total RNA was isolated using Trizol and newly transcribed RNA was purified as described [10]. The IP3R inhibitor 2-APB (100μM, Sigma-Aldrich), the PKC/PKD inhibitor Gö6976 (10μM; Tocris), the CaMKII inhibitor KN-93 (10μM; Tocris) and the calcium chelator BAPTA-AM (50μM; Cayman Chemical) were applied as described [11]. Actinomycin D (2μg/ml, Sigma-Aldrich) was applied at a final concentration of 2μM to inhibit Pol II. Reverse transcription was performed using All-in-One cDNA Synthesis Supermix (Biotool) including a mix of hexanucleotide random primers and poly-dT primers. qRT-PCR was performed using the SYBR Green 2x Mastermix (Biotool) (qRT-PCR primer sequences in Table D in S1 File). Relative quantitation was performed using the ΔΔCT approach. Dot blot analysis was performed as described previously [47] with a few minor changes regarding the detection of 4sU-incorporation into total cellular RNA. Briefly, metabolic labeling of newly transcribed RNA was initiated by adding 500μM 4sU to the cell culture medium together with either 50μM BAPTA-AM, 2μg/ml Actinomycin D or mock (DMSO). Total RNA was isolated using Trizol and thiol-specifically biotinylated using Biotin-HPDP. Following removal of the unincorporated Biotin-HPDP by Chloroform extraction and recovery of the biotinylated RNA by isopropanol/ethanol precipitation, 200ng down to 22ng of biotinylated RNA or 60ng to 0.6 ng of a biotinylated oligo (50bp) were spotted on a positively charged Zeta membrane (Biorad) in alkaline buffer. The membrane was subsequently probed with a Streptavidin-DyLight-680 conjugate and visualized using a LI-COR imaging system.
Subcellular RNA fractions (cytoplasmic, nucleoplasmic and chromatin-associated RNA) were prepared combining two previously published protocols [29, 30]. For the detailed protocol see S2 File. The efficiency of the fractionations was controlled by qRT-PCRs for intron-exon junctions for ACTG1 (chromatin-associated vs other three fractions) and western blots for histone H3 (nuclear vs cytoplasmic fraction). Fractionation efficiencies were furthermore confirmed on the RNA-seq data by comparing expression values of known nuclear and cytoplasmic RNAs as well as intron contributions (Fig J in S3 File).
Sequencing libraries were prepared using the TruSeq Stranded Total RNA kit (Illumina). While rRNA depletion was performed for total RNA and all subcellular RNA fractions using Ribo-zero, no rRNA depletion was performed for the 4sU-RNA samples. Sequencing of 75bp paired-end reads was performed on a NextSeq 500 (Illumina) at the Cambridge Genomic Services and the Core Unit Systemmedizin (Würzburg).
HFF were infected for 8h with wild-type HSV-1 at an MOI of 10 or exposed to 1h or 2h of 80mM KCl or 44°C as described above. ATAC-seq was performed according to the original protocol starting with 1x105 cells per condition [15]. ATAC-seq libraries were quantified by Agilent Bioanalyser and sequenced by NextSeq 500 at the Cambridge Genomic Services (75bp paired-end reads).
Sequencing adapters were trimmed from sequencing reads using cutadapt [48]. Trimmed sequencing reads were mapped against (i) the human genome (GRCh37/hg19), (ii) human rRNA sequences and (iii) the HSV-1 genome (HSV-1 strain 17, GenBank accession code: JN555585, only for HSV-1 infection data) using ContextMap v2.7.9 [49] (using BWA as short read aligner [50] and allowing a maximum indel size of 3 and at most 5 mismatches). For the two repeat regions in the HSV-1 genome, only one copy each was retained, excluding nucleotides 1–9,213 and 145,590–152,222. As ContextMap produces unique mappings for each read, no further filtering was performed and all reads mapped to the human genome were used for downstream analyses. Number of mapped sequencing reads per genome position (= coverage, sum of 2 replicates) were visualized using the Integrative Genomics Viewer (IGV) [51]. No normalization was performed for this purpose.
Number of read fragments per gene were determined from the mapped 4sU-seq reads in a strand-specific manner using featureCounts [52] and gene annotations from Ensembl (version 87 for GRCh37). All read pairs (= fragments) overlapping exonic regions on the corresponding strand by ≥25bp were counted for the corresponding gene. HSV-1 gene annotations were obtained from GenBank (GenBank accession code: JN555585). Expression of cellular protein-coding and lincRNAs was quantified in terms of fragments per kilobase of exons per million mapped reads (FPKM) and averaged between replicates. Only reads mapped to the human genome were counted for the total number of mapped reads for FPKM calculation. Fold-changes in FPKM values were normalized by dividing by median fold-changes for housekeeping genes (as defined in [53]) to account for different levels of DoTT/DoG transcription and consequently different numbers of intergenic reads in different samples and conditions.
The percentage of transcription downstream or upstream of a gene (on the same strand) were calculated separately for each replicate as:
As both transcription downstream or upstream of the gene (= FPKM in 5kb downstream or upstream of gene) and transcription within the gene (= gene FPKM) are quantified in the same timeframe of infection using 4sU-seq, both should be affected to the same degree by a general decrease in transcription. Thus, calculation of this ratio cancels out the effect of any general decrease in transcription.
%downstream and upstream transcription were averaged between replicates and transcription read-through and read-in were then calculated as:
If this resulted in negative values for a gene, read-through or read-in were set to 0. As calculation of %downstream transcription or %upstream transcription cancels out the effect of any overall decrease in transcription, calculation of read-through or read-in are independent of any such decrease.
Mock read-through values were calculated from the two replicates for each time-point for each condition. Here, mock read-through(x,r) = %downstream transcription in replicate r for sample x - %downstream transcription in replicate r’ for sample x. Here, r, r'∈{1,2} and r'≠r.
Occurrence numbers of all possible 6-mer nucleotide sequences were determined within 100nt up- or downstream of gene 3’ends. Spearman correlations between these counts for each gene and read-through values in each sample as well as significance of correlations were calculated using the cor.test function in R and adjusted for multiple testing for each sample using the method by Benjamini and Hochberg for controlling the false discovery rate (FDR) [54].
Splice junctions and read counts for splice junctions were determined from spliced 4sU-seq/RNA-seq read mappings. All predicted junctions were considered that used at least one annotated exon boundary and ended within the annotated 3’ and 5’ ends of the corresponding gene. Only reads were counted that included at least 5bp on either side of the splice junction. Regulation (up- or downregulation) of splice junctions was evaluated in terms of the odds-ratio: cj*coj*cjcoj. Here, cj* and cj are the junction counts in infected or treated and uninfected or untreated cells, respectively. coj* and coj are the counts for all other junctions of the same gene in infected or treated and uninfected or untreated cells, respectively. Odds-ratios and significance of odds-ratios were calculated from replicate data using the Mantel-Haenszel chi-squared test in R. Multiple testing correction was performed with the method by Benjamini and Hochberg [54]. Splicing events were considered significantly upregulated (downregulated) if the adjusted p-value was ≤ 0.01 and the odds-ratio ≥2 (≤0.5).
Adapter trimming and mapping to human and HSV-1 genomes was performed as described for the 4sU-seq data. BAM files with mapped reads were converted to BED format using BEDTools [55] and OCRs were determined from these BED files using F-Seq with default parameters [56]. No filtering of OCRs was performed. Assignment of OCRs to gene promoters was performed using ChIPseeker [57]. 5kb dOCR length for each gene was calculated as the number of nucleotides in the 5kb directly downstream of the gene 3’ end that overlap an OCR. dOCR length for a gene was calculated as the total genomic length of downstream OCRs (including only the positions downstream of the gene 3’ end) assigned to the gene in the following way. First, all OCRs overlapping with the 10kb downstream of a gene were assigned to this gene. Second, OCRs starting at most 5kb downstream of the so far most downstream OCR of a gene were also assigned to this gene. This was performed iteratively, until no more OCRs could be assigned. 5kb dOCR length and dOCR length were averaged between replicates. Empirical cumulative distribution functions for dOCR length were calculated with the ecdf function in R [58]. For this purpose, genes were grouped according to read-through at 7-8h p.i. HSV-1 infection or salt or heat stress. Thresholds were chosen such that genes without DoTT (read-through ≤5%) were in one group and the remaining genes were divided in equal-sized groups according to read-through.
Narrow peaks for ChIP-seq data of histone modification marks (H3K27ac, H3K27me3, H3K36me3, H3K4me1, H3K4me3, H3K9me3) in HFF were downloaded from ENCODE (epigenome series ENCSR403RCR). Presence of histone modification marks downstream of each gene was evaluated by determining the number of nucleotides in the 5kb directly downstream of the gene 3’ end that overlap peaks for the corresponding histone marks (denoted as downstream histone mark length). Spearman correlations between read-through in all conditions and downstream histone mark length and significance of correlations were calculated using the cor.test function in R and adjusted for multiple testing for each sample using the method by Benjamini and Hochberg [54].
The datasets generated and analyzed in the current study are available in the Gene Expression Omnibus (GEO) database under the following accession numbers:
4sU-seq data of HSV-1 infection: GSE59717.
4sU-seq data of salt and heat stress: GSE100469.
RNA-seq of total, cytoplasmic, nucleoplasmic and chromatin-associated RNA: GSE100576.
ATAC-seq data for HSV-1 infected cells: GSE100611.
ATAC-seq data for salt and heat stress: GSE101731.
|
10.1371/journal.ppat.1004609 | Promiscuous RNA Binding Ensures Effective Encapsidation of APOBEC3 Proteins by HIV-1 | The apolipoprotein B mRNA-editing enzyme catalytic polypeptide-like 3 (APOBEC3) proteins are cell-encoded cytidine deaminases, some of which, such as APOBEC3G (A3G) and APOBEC3F (A3F), act as potent human immunodeficiency virus type-1 (HIV-1) restriction factors. These proteins require packaging into HIV-1 particles to exert their antiviral activities, but the molecular mechanism by which this occurs is incompletely understood. The nucleocapsid (NC) region of HIV-1 Gag is required for efficient incorporation of A3G and A3F, and the interaction between A3G and NC has previously been shown to be RNA-dependent. Here, we address this issue in detail by first determining which RNAs are able to bind to A3G and A3F in HV-1 infected cells, as well as in cell-free virions, using the unbiased individual-nucleotide resolution UV cross-linking and immunoprecipitation (iCLIP) method. We show that A3G and A3F bind many different types of RNA, including HIV-1 RNA, cellular mRNAs and small non-coding RNAs such as the Y or 7SL RNAs. Interestingly, A3G/F incorporation is unaffected when the levels of packaged HIV-1 genomic RNA (gRNA) and 7SL RNA are reduced, implying that these RNAs are not essential for efficient A3G/F packaging. Confirming earlier work, HIV-1 particles formed with Gag lacking the NC domain (Gag ΔNC) fail to encapsidate A3G/F. Here, we exploit this system by demonstrating that the addition of an assortment of heterologous RNA-binding proteins and domains to Gag ΔNC efficiently restored A3G/F packaging, indicating that A3G and A3F have the ability to engage multiple RNAs to ensure viral encapsidation. We propose that the rather indiscriminate RNA binding characteristics of A3G and A3F promote functionality by enabling recruitment into a wide range of retroviral particles whose packaged RNA genomes comprise divergent sequences.
| APOBEC3 proteins are cell-encoded restriction factors that counteract infections, particularly by retroviruses such as HIV-1, and retrotransposons. When packaged into HIV-1 particles, APOBEC3G and APOBEC3F both inhibit reverse transcription and induce destructive hypermutation in viral DNA. The mechanism of APOBEC3 virion packaging awaits elucidation, though a dependency on RNA binding has been established. Here, we employed a cross-linking and next generation sequencing approach to determine which RNAs are bound to A3G and A3F in HIV-1 infected cells. We show that both proteins bind to multiple different RNAs, including viral RNA as well as cellular coding and non-coding RNAs, with relatively little evidence of selectivity. We then developed a complementation assay to address the diversity of RNAs that can act as substrates for A3G/F virion packaging. Consistent with the RNA binding profiles, many RNAs can promote packaging provided that those RNAs are, themselves, packaged. These observations suggest that APOBEC3 packaging lacks selectivity and is driven simply by the non-specific RNA binding capabilities of these proteins. We speculate that this model accounts for the broad range of retro-elements that are susceptible to repression by individual APOBEC3 proteins, and also that such substrates cannot escape APOBEC3-mediated inhibition through sequence variation.
| The apolipoprotein B mRNA-editing enzyme catalytic polypeptide-like 3 (APOBEC3, or A3) proteins have been identified as potent antiviral effector proteins [1,2]. There are seven family members in humans, each of which contains one (A3A, A3C and A3H) or two (A3B, A3D, A3F and A3G) characteristic zinc-coordination domains, one of which is catalytically active [3]. These proteins have been identified as inhibitors of retroviruses such as human immunodeficiency virus type-1 (HIV-1) [4], simian immunodeficiency viruses, murine leukaemia virus [5–7] and mouse mammary tumour virus [8], as well as viruses from other families such as hepatitis B virus [9], adeno-associated virus [10] and also endogenous retroelements [11]. Viruses have developed assorted strategies to evade A3-mediated inhibition, the most prominent of which is the expression of the dedicated regulatory protein, Vif, by most lentiviruses. Specifically, HIV-1 Vif counteracts APOBEC3 proteins by inducing their proteasomal degradation through the direct recruitment of CBF-β and a cellular E3 ubiquitin ligase comprising CUL5, ELOB/C, and RBX2 [12–15]. When Vif is absent or defective, APOBEC3 proteins are packaged into progeny virions and transferred to target cells during new infections, where they inhibit reverse transcription and hypermutate nascent cDNAs through excessive cytidine-to-uridine editing [5,7,16–19]. Thus, the encapsidation of APOBEC3 proteins into viral particles is essential for their antiviral activity, and a complete description of APOBEC3 protein function will require a full understanding of the packaging mechanism.
APOBEC3 proteins are RNA binding proteins [20–22]. A3G associates in an RNA-dependent mechanism with multiple ribonucleoprotein (RNP) complexes and accumulates in cytoplasmic RNA-rich microdomains such as P-bodies, stress granules and Staufen-containing granules [23–26]. Localisation to these regions does not appear to be important for antiviral function [27,28], and it has been suggested that sequestration in RNPs may be important for downregulation of APOBEC3 protein activity within cells. These findings further raise the question of how APOBEC3 proteins are packaged into assembling virions. One elegant study has demonstrated that it is newly synthesised protein that is encapsidated, presumably by averting entrapment into cytoplasmic RNPs [29].
The packaging of A3G into HIV-1 particles requires the nucleocapsid (NC) region of the viral Gag protein [30–33]. It has been established that the A3G interaction with NC is RNA-dependent, leading to the consensus view that its packaging is reliant upon its capacity to bind RNA [30–33]. However, although several groups have sought to define specific RNAs that are responsible for A3G packaging, a clear consensus has not yet emerged. In particular, Kahn et al. suggested that viral genomic RNA (gRNA) is required for A3G packaging [34], Wang et al. have concluded that 7SL RNA is the responsible RNA [35], and Svarovskaia et al. proposed that both viral and cellular RNAs play a role [32].
A3G interacts with diverse RNAs such as the 7SL RNA (the RNA component of the signal recognition particle, SRP), Alu RNAs, human Y RNAs and several mRNAs [24,36], many of which are also packaged into retroviruses (reviewed in [37]). In our study, we used an unbiased strategy to identify the RNAs that interact with A3G or A3F in HIV-1 infected cells and in budded virions, and then undertook specific experiments to investigate the involvement of such RNAs in A3G and A3F packaging. To define the interacting RNAs, we employed a cross-linking and immunoprecipitation technique (iCLIP) followed by next generation sequencing. This method has successfully defined RNAs that interact with proteins such as neuro-oncological ventral antigen (NOVA), hnRNP C and Argonautes [38–40].
Here, we demonstrate that A3G and A3F interact with a broad range of RNA molecules, including HIV-1 gRNA, cellular mRNAs and a number of small non-coding RNAs. A series of cell-based assays revealed that no single/unique RNA mediates the encapsidation of A3G or A3F, suggesting that multiple, diverse RNAs can recruit APOBEC3 proteins into viral particles, provided that they are themselves packaged. We therefore propose that A3G and A3F exploit their relatively non-specific RNA binding capabilities to patrol the cytoplasm for nascent retroviral RNA, thereby ensuring effective capture by assembling viruses and resultant antiviral function.
Packaging of A3G into HIV-1 particles requires the nucleocapsid (NC) region of p55Gag, and it has been established that this interaction is RNA-dependent [30–33]. This led to the consensus view that A3G packaging depends on RNA binding. In order to identify the specific RNAs to which A3G binds for efficient packaging, we first applied an unbiased, deep sequencing method to catalogue the RNAs that are bound to A3G and A3F in living cells productively infected with HIV-1. We also extended this study to determine A3G and A3F target RNAs in cell-free HIV-1 virions.
To generate libraries of RNAs bound to A3G or A3F, we first generated CEM-SS human T-cell lines that stably expressed GFP (negative control), GFP-A3G, GFP-A3F, T7-GFP (negative control), T7-A3G or T7-A3F (S1 Fig.). By using two different immunoprecipitation tags for A3G and A3F, we could identify if either the GFP or T7-epitope tag biased the resulting library. Importantly, our study is distinguished from all others (to the best of our knowledge) by inclusion of GFP-only controls: this is an important addition as it allows the determination of RNA binding enrichment relative to background. These cultures were challenged with vif-deficient HIV-1 such that more than 90% of the cells were infected, as judged by intracellular p24Gag staining (S2 Fig.). 48 h after infection, the supernatants were collected and used to assess A3G and A3F antiviral efficacy. Regardless of the tag, both A3G and A3F were antiviral and inhibited single-cycle virus infectivity by 100- and 30-fold, respectively (S3 Fig.).
Virus-producing cells and viruses produced from these cells were used to generate iCLIP libraries. Fig. 1A shows RNA cross-linked to the proteins of interest after T7 or GFP directed immunoprecipitation, RNAse digestion (used to shorten the length of bound RNAs for later deep sequencing) and ligation of a linker. Of note, A3G and A3F each cross-link vastly more RNA than GFP, consistent with the fact that these proteins are established RNA binding proteins. The RNAs that migrate at higher molecular weights than the protein of interest were extracted. A primer that anneals to the linker was then used to generate cDNA. The cDNAs were circularised, a further primer annealed over the linker to create a region of double stranded DNA and digested with BamHI. This procedure generated DNAs where the unknown sequence was flanked by linkers, allowing PCR amplification of the library (Fig. 1B) and next generation sequencing.
The sequencing provided a total of 20 million reads of 50 nucleotides each. The libraries obtained for each protein contained between 0.1 and 5 million sequences. We aligned the reads to the human genome and to the HIV-1 genome using bowtie [41]. Only reads that aligned uniquely in the genomes with a maximum of two mismatched nucleotides were considered for analysis. Furthermore, the reads containing the same random barcode and truncated at the same nucleotide were considered PCR artefacts, and only one of such reads was considered. These unique reads represented between 80% and 95% of the total reads for each replicate experiment, demonstrating the high level of sequencing library complexity obtained in the experiments. Human mRNAs were split into regions, specifically the 5’- and 3’-untranslated regions (UTRs), introns and open reading frames (ORFs), and each region type was analysed independently. The number of reads aligning to a specific sequence or class of sequences was divided by the total number of aligned reads obtained from the library and then compared with the GFP control to provide measurements of enrichment. We performed two independent experiments for each cell line. By comparing these replicates, we observed that the data sets were highly correlated (r>0.98). The data obtained from the differently tagged (GFP versus T7) proteins also exhibited high correlation (r>0.95). Therefore, we present here the averaged data obtained for the four libraries with standard deviations. As summarised in Fig. 1C, A3G and A3F binding to mRNAs was enriched compared to the GFP control. While the binding of both A3G and A3F to the 3’-UTRs was higher (2-fold), only A3G was enriched in the ORF or 5’-UTR regions (2- and 3-fold, respectively); Fig. 1C and S1 Table).
We next looked in detail at virion-associated RNA: both the gRNA and host RNAs such as tRNAs, U RNAs, 7SL, ribosomal RNAs and Y RNAs [37]. Because these cellular RNAs are transcribed from repeat elements in the human genome, the reads were aligned using the repeat masker software [42] to a library containing consensus sequences of the elements, allowing misalignment of up to 3 nucleotides per read. As before, hits in each repeat consensus were divided by the total number of reads and compared to the GFP negative control. Fig. 1D and S2 Table show that A3G and A3F bound 2-fold more to the viral gRNA compared with GFP. Also, both A3G and A3F bound to Y RNAs (17- and 7-fold, respectively) and to U RNAs (2-fold), as compared to GFP. Interestingly, A3G bound to 7SL (14-fold) while A3F did not.
As depicted in Fig. 1E, we observed that in infected cells, 20% of the RNA that is bound to A3G or A3F is of viral origin. However, in viral particles this RNA constitutes the majority of the library (80% of reads aligned to HIV-1 gRNA). This implies that, in vif-deficient HIV-1, A3G/F may be packaged mainly through interactions with HIV-1 gRNA, or that APOBEC3 proteins bind to encapsidated cellular RNAs and then transfer to gRNA once inside the viral particles.
In light of the diversity of RNA substrates bound by A3G or A3F, we undertook a series of studies designed to determine which RNAs can mediate the encapsidation of A3G or A3F into budding HIV-1 particles.
One obvious RNA species that could potentially mediate the packaging of A3G into particles is the viral gRNA. Indeed, at least one previous report has considered this RNA to be essential for A3G packaging [34]. Since A3G and A3F are clearly able to bind to this RNA in infected cells, we performed packaging assays to test this hypothesis. First, we used lentiviral vectors with or without gRNA (Fig. 2A). Succinctly, 293T cells stably expressing HA-tagged A3G or A3F were co-transfected with the HIV-1-based packaging plasmid (p8.91) [43], the VSV glycoprotein envelope expression vector, and either the pHR’SIN-cPPT-SEW lentiviral vector plasmid (denoted lt vector) [44] that expresses gRNA with an intact packaging signal (Ψ) or a mock plasmid. Immunoblot analysis of particles harvested 48 h post transfection and isolated through a sucrose cushion shows that A3G and A3F were packaged into both viral vectors with almost identical efficiency, indicating that viral gRNA is not required for effective A3G or A3F packaging.
These data were next confirmed using an alternative experimental system where the gRNA also serves as the mRNA for Gag (Fig. 2B). 293T cells stably expressing HA-tagged A3G or A3F were transfected with expression vectors encoding: (i) Gag-Pol, (ii) Gag, (iii) Gag in which NC had been replaced by a leucine zipper domain (Zwt) to allow for Gag multimerisation and VLP production (Gag ΔNC) [45], or (iv) a Gag-Pol vector in which the previously defined gRNA packaging determinant between SL2 and SL3 [46] had been deleted (ΔΨ); (S4 Fig.). The protease region of Pol was rendered inactive in the Gag-Pol constructs to facilitate the detection and comparison of Gag across the different samples. Analysis of viral-like particles (VLPs) confirmed that NC is necessary for efficient A3G and A3F packaging, but showed that the Ψ element is dispensable (Fig. 2B, lanes 3 and 7) [30–33]. Since Gag alone was able to package A3G or A3F efficiently, we conclude that Pol is also not required for the packaging of A3G and A3F (Fig. 2B, compare lanes 4 with 1 and 8 with 5).
We also determined the amount of gRNA packaged into wild type Gag, Gag ΔNC and Gag-Pol ΔΨ VLPs by quantitative real time PCR (qRT-PCR) analysis (Fig. 2C). As anticipated, the Gag ΔNC, and ΔΨ VLPs each contained <5% of the level of gRNA compared to wild type Gag VLPs.
We next examined a second RNA, the non-coding 7SL RNA of the SRP, for its importance in A3G/F packaging. This RNA is incorporated into retroviral particles [47–49]. In particular, our iCLIP analyses (Fig. 1D) confirmed earlier work showing that A3G, unlike A3F, binds to 7SL RNA [50]. Of note, one group has previously reported that 7SL is the RNA required for A3G packaging [35], whereas a second concluded the opposite [51].
Over-expression of SRP19, a protein component of the SRP, reduces the amount of 7SL RNA packaging into HIV-1 particles [35], presumably by binding to free cellular 7SL RNA and precluding its interaction with Gag and resultant incorporation into assembling virions. Accordingly, we transfected 293T cells stably expressing A3G or A3F with a plasmid encoding the vif-deficient NL4-3 provirus together with an expression vector for SRP19 or a control vector. Immunoblot analysis 48 h post transfection showed that both A3G and A3F were still packaged into virions, irrespective of the reduction in virion-associated 7SL RNA (Fig. 3A). Interestingly, reducing the amount of packaged 7SL did not influence the infectivity of the viruses or the antiviral activity of the packaged A3G or A3F (S5 Fig.). The data were then confirmed using Gag VLPs, with Gag ΔNC serving as a negative control (Fig. 3B). Quantitative RT-PCR analysis of RNA extracted from these VLPs confirmed that SRP19 overexpression reduced virion-associated 7SL RNA levels by more than 90% (Fig. 3C). Our data therefore rule out a unique requirement for 7SL RNA for efficient A3G or A3F incorporation into virions, suggesting that their packaging may be mediated by other RNAs.
To investigate a potential redundancy between HIV-1 gRNA and host 7SL RNA, we next transfected 293T cells expressing A3G or A3F with a vector encoding SRP19 to inhibit 7SL packaging, as well as the aforementioned ΔΨ construct to generate VLPs depleted of gRNA. We observed that A3G and A3F were both packaged with normal efficiencies irrespective of SRP19 over-expression and/or the prevention of gRNA packaging (Fig. 4). Our observations imply that APOBEC3 proteins are packaged into these particles through the action of other RNAs.
Our data suggested that a variety of host cell RNAs might be involved in APOBEC3 protein packaging into virions. To investigate this further, we next devised an alternative experimental approach whereby the packaging of RNA was dictated by heterologous RNA binding proteins (or domains thereof) rather than by the product of binding to NC and intracellular abundance. To do so, RNA binding domains were genetically fused to the carboxy-terminus of Gag ΔNC, and these proteins were used to generate VLPs in the presence or absence of A3G or A3F. In case of poor virion production, Gag ΔNC was co-expressed to ensure efficient VLP production [52,53].
To verify the experimental system, we initially expressed SRP19 fused to Gag ΔNC, anticipating that SRP19 would bind 7SL RNA and recruit it into VLPs together with A3G or A3F. Fig. 5A demonstrates that the addition of 6 amino acids to the C-terminus of Gag ΔNC, creating convenient restriction sites to allow fusions to Gag ΔNC (lanes 3 and 7), does not mediate APOBEC3 packaging. Also, A3G/F packaging was not mediated by the fusion of Gag to GFP, a protein that does not specifically bind to RNA (lanes 4 and 8). Although SRP19 was cleaved from Gag (lanes 5 and 10), the produced particles still contained intact Gag ΔNC-SRP19. Quantification of the amount of 7SL RNA in these VLPs demonstrated that the presence of SRP19 enables the efficient packaging of 7SL RNA (Fig. 5B). Importantly, these Gag ΔNC-SRP19 particles were able to package A3G efficiently, but not A3F. This observation is in agreement with the iCLIP data, which show that A3F does not preferentially bind to 7SL. Thus, while 7SL RNA is not required for the packaging of A3G into HIV-1 particles with an intact NC domain (Fig. 3), it is evidently able to promote packaging of A3G when selectively captured by VLPs.
Our iCLIP data suggested that both A3G as well as A3F can bind to Y RNAs, small non-coding RNAs of ∼100 nucleotides that are components of RoRNPs [54,55] (Fig. 1D). These RNAs are also incorporated into HIV-1 particles [35,49]. To investigate their role in A3G/F packaging, we first tried to “knock down” their levels using siRNA mediated silencing. Although we could reduce cellular Y RNA concentrations, the levels that were encapsidated remained unaltered, in line with previous findings where cellular Y RNAs were reduced following RNAi-mediated Ro60 depletion but packaging into MLV particles was unchanged [56].
We then used our established Gag ΔNC fusion system to ask if Y RNAs would be able to recruit A3G or A3F into viral particles. Ro60 was fused to Gag and 293T cells were co-transfected to express HA tagged A3G or A3F with wild type Gag, Gag ΔNC, Gag ΔNC-Ro60, Gag ΔNC-Ro60 together with wild type Gag or with Gag ΔNC. VLPs were produced and immunoblotting showed that, although Gag fused to Ro60 alone did not produce detectable quantities of particles, mixed particles containing the Gag-Ro60 fusion protein and either Wt Gag or Gag ΔNC were formed (Fig. 5C). Quantitative RT-PCR specific for Y3 RNA was performed on RNA extracted from these particles; both mixed Gag ΔNC + Gag ΔNC-Ro60 and wild type Gag particles each contained ∼7-fold more Y3 RNA compared to Gag ΔNC VLPs (Fig. 5D), thus validating our approach. Analysis of the A3G/F contents demonstrated a strong restoration of packaging for the Gag ΔNC + Gag ΔNC-Ro60 mixed particles, though the levels did not match those noted with wild type Gag VLPs (Fig. 5C). These results indicate that controlled packaging of specific RNA ligands of A3G or A3F can determine their encapsidation, presumably by bridging between assembling Gag fusion proteins and APOBEC3 proteins, further demonstrating that specific RNAs can promote the encapsidation of APOBEC3 proteins.
Having shown that specific RNAs can recruit APOBEC3 proteins into assembling HIV-1 particles, we next asked whether this was a general property of packaged RNAs. To address this, a series of RNA binding domains (RBDs) of cell-encoded RNA binding proteins known to possess broad RNA binding capabilities were fused to Gag ΔNC as above. Specifically, we used the RBDs from two heterogeneous ribonucleoprotein (hnRNP) proteins: hnRNP C1, that binds to uridine tracts of RNAs [39], and hnRNP K, which is the prototypic protein for the KH RNA binding motif and has high affinity to poly(C) RNA [57,58]. We also fused the splicing factor SRSF2 that has a degenerate RNA binding sequence motif [59] and the double stranded RNA binding protein Staufen-1 [60,61].
These expression constructs were co-transfected with A3G/F into 293T cells together with vectors for wild type Gag or Gag ΔNC, and VLPs analysed by immunoblot (Fig. 6). Remarkably, all four RBDs readily rescued packaging of A3G and A3F in mixed virions with Gag ΔNC with (generally) similar efficiency as the wild type Gag or NC itself when reconnected to Gag ΔNC (compare lanes 6, 8, 10 and 12 with 1, 2 and 4; and 18, 20, 22 and 24 with 13, 14 and 16). Given that A3G and A3F exhibit very broad RNA binding characteristics (Fig. 1), we conclude that a multitude of such RNA substrates, if packaged, can serve to draw A3G/F into VLPs.
Fig. 6 shows that A3G and A3F can be incorporated into VLPs when diverse RBDs are fused to Gag. This is consistent with our iCLIP data, demonstrating that A3G and A3F are able to bind to multiple diverse RNAs. One obvious question that is raised by these observations is whether A3G/F are preferentially encapsidated by assembling HIV-1 VLPs, or whether they are inevitably packaged as a natural consequence of their promiscuous RNA binding capabilities. To address this directly, we carefully quantified the ratios of A3G and A3F to RNA in the lysates of virus producing cells and the matching particle preparations (Fig. 7). Accordingly, 293T cells were transiently transfected with vectors expressing T7-tagged versions of A3G or A3F, as well as the wild type Gag expression vector. Cells were also transfected with an irrelevant plasmid to serve as a negative control. VLPs and cell lysates were collected at 48 h post transfection. VLPs were isolated through a continuous sucrose gradient and the fractions containing VLPs were identified by immunoblot (S6 Fig.). Protein quantities were then determined against a standard curve of recombinant T7-A3G (Fig. 7A and 7B), and RNA in cell lysates and VLPs were extracted and quantified by Qubit (Fig. 7C). Culture supernatant from cells transfected with an irrelevant plasmid was used to assess background, and RNA was not detected in these samples (threshold of detection, 20 pg/ml). Interestingly, the calculated ratios of A3G/F to RNA were similar in cells and in virus particles (Fig. 7D, mean of 4 independent experiments). In other words, there is no evident enrichment of A3G or A3F in virions relative to virus-producing cells. Taking all our findings together leads us to conclude that the packaging of A3G and A3F into HIV-1 particles is driven by RNA binding, and that multiple/diverse RNAs can fulfil this role provided they are themselves packaged.
A3G and A3F are antiviral proteins that have to be packaged into newly synthesised retrovirus particles to exert their activity. However, the details of the packaging mechanism remain incompletely understood, though a preference for encapsidating newly synthesised A3G has been established [29–35,47,48,51]. Here, we used for the first time a high throughput method to identify the RNAs that these two proteins bind to in living cells productively infected with vif-deficient HIV-1 and in cell-free viral particles. We then systematically addressed which RNAs are either necessary or sufficient for A3G and A3F incorporation into budding virions.
An important innovation with our study was the employment of a reference (“non-RNA binding”) protein during the iCLIP procedure, in this case GFP. This enabled us to identify A3G/F RNA ligands that are enriched over background RNA associations that are a presumed property of any protein (Fig. 1; exemplified here by the generation of iCLIP libraries from GFP-containing cells). While some preferential binding to certain classes of RNAs was apparent, it was nonetheless evident that the patterns of A3G/F binding were mostly non-discriminatory. We speculate that such promiscuity in RNA binding could be explained by the capacity of A3G/F to interact with myriad RNA sequences that are available for binding simply because they are not already occupied by other proteins.
We investigated in detail some specific RNAs that are found in retroviral particles [37]. We found that A3G and A3F do not require HIV-1 gRNA or 7SL for packaging into HIV-1 particles (Fig. 2, 3 and 4). However, using our Gag ΔNC fusion assay, we could show that A3G (but not A3F) can utilise 7SL to be incorporated into viral particles (Fig. 5). These observations correlated well with our iCLIP data, where it was shown that A3F does not bind to 7SL.
In accordance with previous studies [30–33], we observed that A3G and A3F are not packaged into Gag ΔNC VLPs. Importantly, these particles can clearly incorporate APOBEC3 proteins when a variety of unrelated RNAs are recruited into assembling virions via fusions of Gag ΔNC to a series of unrelated RBDs (Fig. 5 and 6). In other words, under experimental conditions, A3G and A3F can be packaged into HIV-1 particles via interactions with diverse and unrelated RNAs. Our data also show that 80% of the RNA sequences that A3G/F bind to inside vif-deficient (but otherwise normal) viral particles are of viral origin. If we assume that this distribution correlates with the binding of A3G/F to RNAs that are being packaged during particle production, our data imply that viral genomic RNA ordinarily mediates A3G/F packaging. An alternative possibility is that non-viral RNAs recruit A3G/F into particles, but that A3G/F then release these RNAs and bind to gRNA following particle formation and release.
Lastly, fastidious quantification of A3G/F and total RNA levels in cells and budded virions revealed that viral particles are not enriched for APOBEC3 protein content (Fig. 7). While RNA binding is clearly required for APOBEC3 protein packaging, these observations indicate that there is no selectivity for engaging RNAs that are destined for incorporation into assembling viruses. Accordingly, we speculate that cytoplasmic APOBEC3 proteins exploit their relatively non-specific RNA binding capabilities to patrol the cytosol and bind to unoccupied sites on RNAs. In the context of cellular RNAs, this may account for the pronounced accumulation of APOBEC3 proteins in RNA-rich microdomains such as P-bodies and stress granules [23–26]. For viruses with RNA genomes, such as retroviruses, this can result in encapsidation into newly formed viral particles. Given the DNA editing function of APOBEC3 proteins, viruses that have RNA genomes and replicate via DNA intermediates—namely retroviruses and hepadnaviruses—will be susceptible to hypermutation and inhibition [4,5,9]. Indeed, we propose that the relatively non-specific RNA binding characteristics of A3G/F render these proteins well suited to the inhibition of a wide variety of viral and transposon targets. Moreover, this feature may further ensure that viral sequence variation, a noted hallmark of HIV-1, will not afford a means of escape from APOBEC3-mediated restriction: perhaps this underlies the evolution of an entirely different (protein-based) evasion mechanism, namely Vif-induced protein degradation?
cDNAs encoding SRP19, GFP, A3G or A3F were cloned between the XbaI and BamHI sites of pCGTHCFFLT7 [62]. DNA fragments encoding T7-tagged derivatives of GFP, A3G or A3F were cloned between the XhoI and EcoRI sites of the retrovirus vector, pCMS28 [26]. A GFP-containing fragment with a GST linker sequence at its 3’-end was cloned between the EcoRI and XhoI restriction sites of pCMS28, and A3G or A3F cDNAs were then inserted using the NotI and XhoI sites. Plasmids expressing HA-tagged A3G and A3F were previously described [50]. vif-deficient HIV-1NL4-3 [27] and HIV-1IIIB [63] strains were used where indicated. The wild type Gag-Pol vector, pCMS446, was generated by inserting the HIV-1 5’ UTR-Gag-Pol, Protease activity inactivated (nucleotides 455–5096 from the HV-1HXB2 isolate [GenBank: K03455.1] [64]) fragment into pcDNA3.1 (Invitrogen) that contains the HIV-1 RRE [65]. pGag was generated by deleting the Pol sequence 3’ to the Gag stop codon. pGag ΔNC was generated by cloning the analogous 5’ UTR-Gag-Pol fragment from the Zwt-p6 proviral construct [45] into the same RRE-containing vector. SL2 and 3 were deleted from pGag by overlapping PCR using the primers 5′-AGGGGCGGCGACTGGTGAGAGATGGGTGCGAGAGCGTCAGTATTAAGC-3′ and 5′-TGACGCTCTCGCACCCATCTCTCACCAGTCGCCGCCCCTCGCCTCTTGC-3′ to generate pGag ΔΨ [46]. pGag ΔNC NB was created by inserting NheI and BamHI restriction sites in frame 5’ to the stop codon of Gag in pGag ΔNC. NC (HIV-1HXB2 strain), GFP, SRP19, Ro60, hnRNP C1 (amino acids 1–104), hnRNP K (amino acids 38 to 464 with amino acids 323–338 deleted), SRSF2 (amino acids 2–93) and Staufen 1 cDNAs were inserted into pGag ΔNC NB using the NheI and BamHI sites.
293T and HeLa cells were obtained from the American Tissue Culture Collection (ATCC). TZM-bl cells were obtained through the NIH AIDS Reagents Repository Program (ARRP). These cell lines were cultured in Dulbecco’s modified Eagle’s medium (Invitrogen, UK) supplemented with 10% foetal bovine serum and 1% penicillin/streptomycin. CEM-SS T cells, from ARRP, were cultured in Roswell Park Memorial Institute 1640 medium (Invitrogen, UK) supplemented with 10% foetal bovine serum and 1% penicillin/streptomycin. Stable CEM-SS T cell lines were generated by standard retroviral transduction using MLV-based vectors expressing GFP, GFP-A3G GFP-A3F, T7-GFP, T7-A3G or T7-A3F and selected with 1 μg/ml puromycin. 293T cells stably expressing HA-tagged A3G or A3F were generated by transduction with MLV based vectors expressing the proteins of interest and selection with 1 μg/ml puromycin. Expression levels of the A3 proteins were assessed by immunoblot using rabbit polyclonal sera specific for A3G [66] or A3F [67] for primary detection.
iCLIP has been described in detail previously [39]. Briefly, CEM-SS T cells stably expressing the proteins of interest were infected with vif-deficient HIVIIIB. 48 h later, a sample was collected to assess infection by intracellular p24Gag staining and flow cytometry, confirming that at least 80% of cells were productively infected. The supernatant was collected, filtered through a 0.45 μm pore filter, and viruses isolated through a 20% sucrose cushion (wt/vol) at 21000 × g for 2 h at 4°C and resuspended in PBS. Cells were collected, washed 6 times with PBS and resuspended in PBS. Cells and viruses were then radiated with 400 mJ/cm2 using a Stratlinker 2400. Cells were pelleted by centrifugation and the supernatant discarded. Cells and viruses were then resuspended in 1 ml of lysis buffer (50 mM Tris-HCL, pH 7.4; 100 mM NaCl; 1% NP-40; 0.1% SDS; 0.5% sodium deoxycholate and protease inhibitor) and sonicated. 0.16 μg of RNase A were added to High RNase samples and 0.04 ng to the other samples. Tubes were incubated at 37°C for 3 min and added to protein G dynabeads previously incubated with anti-T7 antibody (Novagen) or anti-GFP antibody (Roche). The RNAs were dephosphorylated using Shrimp alkaline phosphatase (Promega) and a pre-adenylated linker was ligated to the 3’-end of RNAs on beads. RNAs were radiolabeled with P32-ϒ-ATP and separated using SDS-polyacrylamide gel electrophoresis, electrophoretically transferred to nitrocellulose and visualised on film. Pieces of the membrane containing the RNAs of interest were excised and resuspended in PK buffer (100 mM Tris-HCl pH 7.5, 50 mM NaCl, 10 mM EDTA) containing 2 mg/ml proteinase K. RNAs were isolated with phenol/chloroform (Ambion) and precipitated with 2.5 volumes of 100% ethanol, 0.1 volumes of sodium acetate (3 M, pH 5.5) and 0.5 μl of glycoblue. RNAs were then pelleted and reverse transcribed using barcoded primers. cDNAs were run on a TBE-urea polyacrylamide gel and products ranging from 70–85, 85–110 and >110 base pairs were excised. Nucleic acids were extracted and circularised. A primer that anneals to the linker previously ligated to the RNAs was used to create a double stranded region and this DNA was digested with BamHI. cDNA was then amplified by PCR and sequenced on one lane of an Illumina GA2 flow cell with 50 nucleotides run length. Data is available at ArrayExpress with the accession number E-MTAB-2700.
Before mapping reads, adapter sequences were removed, and the barcodes for each sample within each library were used to identify which sequence was immunoprecipitated from each protein. Mapping of the reads was performed against the human genome (version Hg19/ GRCh37) and HIV-1IIIB (GenBank ID EU541617) genome using bowtie [41]. Reads that aligned to a single position on the human genome or HIV-1 genome with at most two mismatches were considered for analysis. Genomic annotations were then assigned based on gene annotations provided by Ensembl (v59). Reads were also aligned to human repeat sequences using repeat masker [42] and a database of consensus sequences provided by the software. A maximum of 3 mismatches were allowed.
HIV-1 virions were produced by transfecting 293T cells with vif-deficient pNL4-3 or pIIIB using polyethyleneimine (PEI). Virus were then harvested 48 h later and filtered through a 0.45 μm pore filter. Lentiviral vectors were produced in 293T cells by transfecting the p8.91 packaging plasmid, a lentiviral vector and the VSVg envelope plasmid pMDG2.1 [43] at a ratio of 2:2:1 using PEI. Vectors were harvested 48 h after transfection and filtered. Viral particles were quantified by p24Gag enzyme linked immunosorbent assay (Perkin Elmer). VLPs were produced by transfecting the packaging plasmid of interest and the Rev expression vector, at a ratio of 2:1. Wherever stated, HA- or T7-tagged GFP, A3G or A3F vectors were co-transfected at a ratio of 1:5 to the packaging plasmid. The supernatant was collected 48 h later, filtered through a 0.45μm pore filter and isolated through a 20% sucrose cushion (wt/vol) at 21000 × g for 2 h at 4°C. Cells and viral pellets were lysed in radioimmunoprecipitation assay (RIPA) buffer (50 mM Tris-HCl pH 7.4, 100 mM NaCl, 1 mM MgCl2, 1% NP-40, 0.1% SDS, 0.5% sodium deoxycholate). T7-tagged A3G was purified as described before [68]. Expression, VLP production and packaging of APOBEC3 proteins were assessed by standard immunoblot using anti-HA (mouse monoclonal; 12CA5) or anti-T7 (Novagen) and anti-p24Gag (mouse monoclonal; p24-2 [69]) antibodies and detected and quantified by Li-cor Odyssey infrared imaging using IRDye800CW or IRDye680LT-labeled secondary antibodies. VLPs produced for RNA and protein quantification were filtered, layered over a continuous sucrose gradient (60–20%) and centrifugated at 150,000 × g for 1h15min at 4°C. 1mL fractions were collected and centrifugated at 21000 × g for 2 h at 4°C. The supernatant was removed and the pellet resuspended in RIPA buffer. The fractions containing p24Gag were identified by immunoblot. The fraction with highest level of p24Gag was then used to quantify the packaged A3G/F by immunoblot using a standard curve of purified T7-A3G. The remainder of the fraction was used to extract RNA using a microRNA extraction kit (Promega). Total RNA was then quantified using the Qubit RNA HS assay kit (Life Technologies), following the manufacturer’s instructions.
RNA was extracted from viral particles using Tri Reagent LS (Sigma) according to the manufacturer’s instructions. 0.5 μg of total RNA was used to synthesise cDNA with the high-capacity cDNA reverse transcription kit (Applied Biosystems) and random primers using the manufacturer’s protocol. qPCR was then performed using the primers 5′-TAACTAGGGAACCCACTGC-3′, 5′-GCTAGAGATTTTCCACACTG-3′ and the probe 5′-6-carboxyfluorescein [FAM]-ACACAACAGACGGGCACACACTA-6-carboxytetramethylrhodamine [TAMRA]-3′ to detect HIV-1 gRNA; SYBR Green (Applied Biosystems) was used to detect 7SL with the primers 5′-GGGCTGTAGTGCGCTATGC-3′ and 5′-CCCGGGAGGTCACCATATT-3′, and to quantify hY3 RNA with the primers 5′-GGCTGGTCCGAGTGCAGTG-3′ and 5′-AAAGGCTAGTCAAGTGAAGCAGTGG-3′.
|
10.1371/journal.pgen.1007591 | ProxECAT: Proxy External Controls Association Test. A new case-control gene region association test using allele frequencies from public controls | A primary goal of the recent investment in sequencing is to detect novel genetic associations in health and disease improving the development of treatments and playing a critical role in precision medicine. While this investment has resulted in an enormous total number of sequenced genomes, individual studies of complex traits and diseases are often smaller and underpowered to detect rare variant genetic associations. Existing genetic resources such as the Exome Aggregation Consortium (>60,000 exomes) and the Genome Aggregation Database (~140,000 sequenced samples) have the potential to be used as controls in these studies. Fully utilizing these and other existing sequencing resources may increase power and could be especially useful in studies where resources to sequence additional samples are limited. However, to date, these large, publicly available genetic resources remain underutilized, or even misused, in large part due to the lack of statistical methods that can appropriately use this summary level data. Here, we present a new method to incorporate external controls in case-control analysis called ProxECAT (Proxy External Controls Association Test). ProxECAT estimates enrichment of rare variants within a gene region using internally sequenced cases and external controls. We evaluated ProxECAT in simulations and empirical analyses of obesity cases using both low-depth of coverage (7x) whole-genome sequenced controls and ExAC as controls. We find that ProxECAT maintains the expected type I error rate with increased power as the number of external controls increases. With an accompanying R package, ProxECAT enables the use of publicly available allele frequencies as external controls in case-control analysis.
| Recent investments have produced sequence data on millions of people with the number of sequenced individuals continuing to grow. Although large sequencing studies exist, most sequencing data is gathered and processed in much smaller units of hundreds to thousands of samples. These silos of data result in underpowered studies for rare-variant association of complex diseases. Existing genetic resources such as the Exome Aggregation Consortium (>60,000 exomes) and the Genome Aggregation Database (~140,000 sequenced samples) have the potential to be used as controls in rare variant studies of complex diseases and traits. However, to date, these large, publicly available genetic resources remain underutilized, or even misused, in part due to the high potential for bias caused by differences in sequencing technology and processing. Here we present a new method, Proxy External Controls Association Test (ProxECAT), to integrate sequencing data from different, previously incompatible sources. ProxECAT provides a robust approach to using publicly available sequencing data enabling case-control analysis when no or limited internal controls exist. Further, ProxECAT’s motivating insight, that readily available but often discarded information can be used as a proxy to adjust for differences in data generation, may motivate further method development in other big data technologies and platforms.
| Recent investments have produced sequence data on millions of people with the number of sequenced individuals continuing to grow. Although large sequencing studies, such as the Trans-Omics for Precision Medicine (TopMed) through the National Heart, Lung, and Blood Institute, exist, most sequencing data is gathered and processed in much smaller units of hundreds to thousands of samples. This is especially true in the study of diseases that are not very common but still likely to have a complex or oligogenic genetic architecture. These silos of data mean that most rare-variant association studies of uncommon, complex diseases are underpowered. Zuk et al. suggest that sample sizes in the tens, and perhaps hundreds of thousands are required for adequate power[1]. In addition to increasing the sample size of future studies, fully leveraging existing sequencing resources could increase power considerably and could be vital in scenarios where resources to sequence more samples are limited.
Existing genetic resources such as the Exome Aggregation Consortium (ExAC; >60,000 exomes)[2] and more recently, the Genome Aggregation Database (gnomAD; ~140,000 sequenced samples) have the potential to be used as controls in studies of complex diseases. However, to date, these large, publicly available genetic resources remain underutilized, or even misused[3], in large part due to the lack of statistical methods that can appropriately use this summary level data in complex disease studies. In particular, there is a large potential for bias caused by differences in sequencing technology, processing, and read depth[3].
Recently, Lee et al[4] developed iECAT, a method to incorporate publicly available allele frequencies from controls into an existing, unbiased, but underpowered case-control analysis. They found that iECAT controls for bias while increasing power to detect association to a genetic region and can be applied to both single variant analysis and gene region analysis using a SKAT-O framework[5]. iECAT cannot be applied to very rare variants such as singletons or doubletons and requires a set of controls that were sequenced and variant-called in parallel to the cases (i.e. internal controls). Additionally, the type I error rate for iECAT increases as the size of the internal control sample set decreases relative to the internal cases. Thus, there is still the need for methods that can incorporate very rare variants and external controls without the explicit need for large internal control samples.
Here we present Proxy External Controls Association Test (ProxECAT), a method to estimate enrichment of rare variants within a gene region using internal cases and external controls. Our method addresses existing gaps such as using singleton and doubleton variants and requiring only external controls.
Rare-variant tests in a gene are often limited to variants predicted to have a functional effect on the protein, hence discarding non-functional variants. This can result in greater power[6, 7]. The development of ProxECAT was motivated by the observation that these discarded variants can be used as a proxy for how well variants within a genetic region are sequenced and called within a sample. ProxECAT is both simple and fast, requiring only allele frequency information, and is thus well suited to use publicly available resources such as ExAC and gnomAD.
We evaluate ProxECAT in simulations, and empirical analysis of high depth of coverage (80x) whole-exome sequenced childhood obesity cases (N = 927) using both low-depth of coverage (7x) whole-genome sequenced controls (N = 3,621), and ExAC (N = 33,370). Our method controls the type I error rate in simulations and yields the expected distribution of test statistics in real data settings. Given an accompanying R package, ProxECAT provides a robust and previously unavailable method to use publicly available allele frequencies as external controls in case-control analysis. This increases the utility of existing sequenced datasets to generate hypotheses and further research into the genetic basis of disease.
For a gene region-based test, we consider the following. Let Y denote the disease status, with Y = 1 and Y = 0 for internal case and external control status, respectively. We split the variants into those that are predicted to have a functional genetic impact and those that are not predicted to have a functional impact. We use the latter as the proxy variants. Let, x1f and x1p denote the counts of the functional and proxy rare variant alleles respectively for internal cases and x0f and x0p denote the counts of functional and proxy rare variant alleles respectively for external controls (Table 1).
We model the observed variant minor allele counts in Table 1 as a random sample from four independent Poisson distributions, i.e., X1f∼Pois(λ1f),X0f∼Pois(λ0f),X1p∼Pois(λ1p), and X0p∼Pois(λ0p). The derivation of the ProxECAT test statistic follows from the null hypothesis in Eq (1):
H0:λ1fλ1p=λ0fλ0p.
(1)
Using the method of Lagrange Multipliers and the constraint as defined by the null hypothesis, we find the maximum likelihood estimates (MLEs) of our parameters: λ1f,λ1p,λ0f,λ0p. Details are in S1 Appendix.
Our MLEs under the null hypothesis are:
λ^1f=(x1f)2+x1fx0f+x1fx1p+x0fx1px1f+x0f+x1p+x0pλ^0f=(x0f)2+x1fx0f+x0fx0p+x1fx0px1f+x0f+x1p+x0pλ^1p=(x1p)2+x1fx1p+x1px0p+x1fx0px1f+x0f+x1p+x0pλ^0p=(x0p)2+x0fx0p+x1px0p+x0fx1px1f+x0f+x1p+x0p.
(2)
We use the parameter estimates in the likelihood for the constrained null hypothesis. The MLEs for the unconstrained alternative hypothesis parameters are the variant allele counts for each group (i.e. λ˜1f=x1f,λ˜0f=x0f,λ˜1p=x1p,λ˜0p=x0p). We then complete a likelihood ratio test (LRT) as the ratio of the constrained (null hypothesis) and unconstrained (alternative hypothesis) likelihoods, which, by Wilk’s theorem[8] can be transformed to have a chi-squared distribution with 1-df.
It has been shown that functional variants have a lower minor allele frequency (MAF) distribution compared to synonymous variants[9]. Further, high-depth of coverage sequencing will detect a higher amount of variation at lower MAFs compared to low-depth of coverage sequencing[9, 10]. This results in high-depth of coverage sequencing detecting more functional variation relative to synonymous variation compared to low-depth of coverage sequencing. To allow for scenarios where sequencing coverage varies considerably between cases and controls, we weight the observed functional variant minor allele counts. Specifically, we divide the number of minor alleles for functional variants by the median ratio of the number of minor alleles for functional to synonymous variants within cases (M1) and within controls (M0) separately:
x1,weightedf=x1fM1x0,weightedf=x0fM0.
The weighted functional variant minor allele counts, x1,weightedf and x0,weightedf, are used in place of the observed functional variant minor allele counts, x1f and x0f, respectively to estimate the parameters in (2). This new test statistic is called ProxECAT-weighted.
By assuming a Negative Binomial distribution for the number of minor alleles in a region instead of a Poisson distribution, we extend ProxECAT to incorporate possible over-dispersion. We model the Negative Binomial distribution with the mean, λ, and over-dispersion, η, parameters where the distribution approaches Poisson as η becomes large (S1 Fig).
We simulated a variety of confounding scenarios. Case-control confounding represents systematic, genome-wide differences in the number of rare minor alleles observed in cases and controls due to differences in sequencing technologies and pipelines. Gene confounding refers to a gene having a higher or lower number of rare minor alleles than expected based on gene length. Gene confounding can occur in both cases and controls for a variety of reasons including differences in mutation rates, ability to detect variants, and annotation quality. Confounding can also occur when a particular gene region has a different number of rare minor alleles in cases and in controls due to sequencing differences between cases and controls. This confounding is distinct from case-control confounding in that it is isolated to a particular gene region rather than genome-wide. Here, we refer to this confounding as gene confounding only in cases. The simulation scenarios and parameters are presented in Table 2 and Supplemental Table 1.
The case-control LRT (see Software and Statistical Analysis under Subjects and Methods) was robust to gene confounding scenarios maintaining the appropriate type I error rate but had an increased type I error rate in the presence of case-control confounding. The case-only LRT maintained appropriate type I error rate in the presence of case-control confounding but was inflated in the presence of gene-confounding. The inflation in the type I error for the case-control LRT and the case-only LRT increased further when both gene and case-control confounding were present. This was especially true for the case-control LRT (Fig 1).
Despite usually being within the 95% confidence interval for type I error, ProxECAT appeared to have a slight, but consistent inflation (Supplemental Table 2). This minor, but consistent inflation in the type I error rate can be addressed by using a more conservative significance threshold. We found that multiplying the significance level by 0.9 works well such that a 0.045 significance threshold maintains a 0.05 type I error rate, a 0.009 significance threshold maintains a 0.01 type I error rate, etc. Both the case-control LRT used here and ProxECAT assume a Poisson distribution and had inflated Type I Error rate in the presence of overdispersion (S3 Table). ProxECAT-over, which assumes a Negative Binomial distribution instead of a Poisson distribution, corrects for overdispersion in simulations when the overdispersion parameter is known and overdispersion is not too extreme (i.e. over-dispersion, η ≥ 5) (S3 Table).
Case-control LRT had higher power than ProxECAT under scenarios of no case-control confounding and given the same sample size (S4 Table). However, the power of ProxECAT increased as the sample size of the external control set increased eventually reaching higher power than the case-control LRT for the same number of internal sequences (Fig 1). This increase in power for ProxECAT is due, in part, to being able to sequence more cases with ProxECAT (N = 1000) than with a case-control LRT where sequencing resources need to be split between cases and controls (here Ncases = 500 and Ncontrols = 500). ProxECAT’s power increased while the type I error stayed the same under confounding scenarios where the number of functional variants in the cases increases (S4 Table).
To assess the fit of the Poisson distribution and specifically look for over dispersion, we simulated rare minor alleles assuming a Binomial distribution for each variant and compared these results to the theoretical Poisson distribution for the number of rare minor alleles in a genetic region. No over dispersion was apparent as the sampling mean and variance of the simulated scenarios were similar across different sample sizes, MAFs, and number of minor alleles per gene (S2 and S3 Figs). When the expected number of minor alleles per gene was greater than 20, the Poisson approximation for the number of minor alleles started to look more continuous. In other words, as the expected number of variants per gene decreased, the Poisson approximation became more discrete and multimodal (S2 and S3 Figs). The theoretical distribution for the number of minor alleles per gene created from simulating genotypes for individual, independent variants from a Binomial distribution was more robust to discretization maintaining a mostly continuous distribution until the expected number of minor alleles per gene was equal to or less than four.
We evaluated ProxECAT using 926 cases from the Severe Childhood Onset Obesity Project (SCOOP) sample as cases and either 3,621 UK10K Cohort or 33,370 ExAC non-Finnish Europeans as controls. High-depth of coverage WES SCOOP cases vs. low-depth of coverage WGS UK10K Cohort controls had an inflated distribution of test statistics for the case-control LRT both at the center (lambda = 1.971) and in the tail of the distribution. While we did not observe inflation in the tail of the distribution for ProxECAT (Fig 2), there was a large inflation in the overall distribution of test statistics (lambda = 3.151). We observed a much higher ratio of the number of minor alleles in functional to synonymous variants per gene for the high-depth of coverage cases, median = 3.00, versus the low-depth of coverage controls, median = 1.89 (Table 3). ProxECAT-weighted, which adjusts for this systematic difference in sequencing coverage, resulted in a distribution of observed test statistics that more closely matches the expected distribution (lambda = 1.026, Fig 2).
A large strength of this method is the ability to use allele frequency data directly, rather than individual level allele calls. To assess the ability of this method to use publicly available allele frequency data, we used ExAC allele frequencies as controls for the SCOOP cases. The standard case-control LRT was inflated at both the median, lambda = 1.713, and tail (Fig 3) while our method maintained the expected distribution of test statistics. Because the depth of sequencing coverage is comparable and high for both SCOOP cases and ExAC controls, ProxECAT-weighted produced similar results to the standard, un-weighted test.
For both analyses, filtering to very rare variants was essential to avoid inflation in the distribution of observed test-statistics. This can be accomplished using moderate internal frequency filters and an external dataset such as 1000Genomes (MAF < 1%) as in the SCOOP vs UK Cohort analysis or using more stringent internal frequency filters (MAF < 0.1%) and no external dataset as in the SCOOP vs ExAC analysis.
Four genes, passing a 0.01 level of significance in both the SCOOP vs UK10K Cohort analysis and in the SCOOP vs ExAC analysis, are shown in Table 4. These results are putative novel obesity candidates meriting further replication. MIB2 may be of particular interest as it is associated with decreased body weight in mice in the International Mouse Phenotyping Consortium (p-value = 7.49*10−10, http://www.mousephenotype.org/data/genes/MGI:2679684). Additional genes with the smallest p-values are found in S5–S7 Tables.
Within the SCOOP vs. ExAC analysis, we completed a sensitivity analysis using three increasingly broad proxy selection strategies of Sequence Ontology terms: (1) synonymous (SYN); (2) predicted low impact rating from Ensembl [11] (LOW); and (3) not in our functional category (NOT FUNC). These strategies are nested with LOW Sequence Ontology terms included in NOT FUNC, and SYN Sequence Ontology terms included in both LOW and NOT FUNC. We assessed consistency across the number of alternate alleles and in the distribution of test statistics across the three proxy selection strategies.
As expected given the nested nature of the proxy selection strategies, SYN had a smaller number of alternate alleles than either LOW or NOT FUNC and LOW had a smaller number of alternate alleles than NOT FUNC. SYN and LOW proxy selection strategies produced similar numbers of alternate alleles per gene while the correlation was lower for NOT FUNC with either SYN or LOW (S4 Fig). We found similar consistency in the distributions of test statistics between the proxy selection strategies (S5 Fig).
We propose a new method, ProxECAT, to test for enrichment of an accumulation of very rare variant alleles in a gene-region using publicly available external allele frequencies. ProxECAT only requires allele frequencies and uses exclusively external controls enabling the use of large, publicly available datasets such as ExAC and gnomAD. Analyses in simulations and using UK10K Cohort and ExAC as control sets for childhood obesity cases show that ProxECAT keeps the type I error rate and expected distribution of test statistics under control despite differences in sequencing technology and processing. Because ProxECAT uses external controls, additional resources can be devoted to sequencing cases. This results in greater power for ProxECAT compared to the case-control LRT test for the same number of internally sequenced individuals.
There are several limitations to the method proposed here. First, ProxECAT has a minor, but consistent inflation in the type I error rate. This limitation is easily addressed by using a more conservative significance threshold. Second, ProxECAT cannot currently include covariates such as sex, and ancestry. Thus, internal cases and external controls should be closely matched by ancestry and, as with any association study, findings will need independent replication preferably using a study where cases and controls are sequenced and processed in parallel. Third, the current approach does not enable internal controls to be analyzed along with external controls. While two analyses can be done in parallel and compared, it would be ideal to incorporate internal and external controls into the same statistical test. We are actively working on extensions to address these limitations.
It is important to highlight that research utilizing solely external controls is more susceptible to confounding due to known or unknown factors. Thus, any genes identified using ProxECAT or any method that uses only external controls should be carefully followed up in further validation, replication, and functional studies.
ProxECAT provides a robust approach to using allele frequencies from existing, publicly available sequencing data enabling case-control analysis when no or limited internal controls exist. ProxECAT uses the insight that readily available genomic information often discarded from analyses (here synonymous variation) can adjust for sizeable confounding due to differences in data generation. In the era of big data, we hope that both this insight and the ProxECAT method will enable additional genetic discoveries and will also motivate future methodological advancements in analyzing data across technologies and platforms.
All tests were implemented using functions from our accompanying R package ProxECAT (https://github.com/hendriau/ProxECAT). Our primary test, which can model both ProxECAT and ProxECAT-weighted, was implemented with the proxecat function and our secondary test modeling over-dispersion was implemented using the proxecat.over function. We also implemented a case-control LRT to test for enrichment of rare, functional variant alleles in cases vs. controls and a case-only LRT similar to that performed by Zhi and Chen in 2012 [12]. The case-only LRT tests for enrichment of rare alleles for functional variants in each gene of interest compared to the genome-wide average number of minor alleles per gene in cases only adjusting for the length of each gene. Unless otherwise specified, we assumed the data follow a Poisson distribution for all LRTs.
Within each case-control confounding simulation, we simulated 20,000 independent genes under four gene-disease association and gene confounding states. The four distinct gene states are: (1) association with case status and no gene confounding, (2) association with case status and gene confounding, (3) no association with case status and gene confounding, (4) no association with case status and no gene confounding. The number of rare minor alleles per gene was simulated under a Poisson distribution or an over-dispersed Poisson modeled using a Negative Binomial parameterization using the R functions rpois and rnbinom, respectively. The mu and size parameters in rnbinom represent the mean and over-dispersion, respectively.
To assess the fit of the Poisson distribution, we simulated the number of each genotype group for each variant assuming Hardy-Weinberg Equilibrium and a Binomial distribution where p was the MAF. We varied the MAF (0.0001, 0.0005, 0.001, 0.005), the sample size (1000; 10,000), and the maximum number of variable variants within the gene region (5, 10, 20). We then assessed how closely the simulated distributions of the number of minor alleles observed per gene region matched a theoretical Poisson distribution where λ was the mean from each simulation scenario.
Whole-exome sequenced (WES) cases are from the Severe Childhood Onset Obesity Project (SCOOP) cohort[6, 13], which is a self-reported UK European subset of the Genetics of Obesity Study (GOOS). GOOS includes individuals with severe early-onset obesity body mass index (BMI) standard deviation score (SDS) > 3 and age at onset of obesity < 10 years. Leptin deficient individuals (identified by biochemical measurement) and those with mutations in the MC4R gene were excluded.
We used VerifyBamID (v1.0)[14] and a threshold of ≥3% to identify contaminated samples. We computed principal components with the 1000Genomes Phase I integrated call set[9] using EIGENSTRAT v4.2[15] to identify non-Europeans, and pairwise identity by descent estimates from PLINK v1.07[16] with a threshold of ≥0.125 to identify related individuals. Contaminated, non-European, and related samples were removed resulting in 927 SCOOP cases for analysis. Details about sequencing and variant calling for the SCOOP cases, as part of the UK10K exomes can be found elsewhere[17]. All participants gave written informed consent and all methods were performed in accordance with the relevant laboratory/clinical guidelines and regulations.
The whole-genome sequenced (WGS) controls consist of the UK10K Cohort sample, comprised of two population cohorts: the Avon Longitudinal Study of Parents and Children (ALSPAC) and the TwinsUK study from the Department of Twin Research and Genetic Epidemiology at King’s College London (TwinsUK). We used allele frequency data for 3,621 individuals that passed sample QC as described elsewhere[17].
We used allele frequency values for the N = 33,370 non-Finnish European (NFE) group from the ExAC variant site dataset version 1.0 (http://exac.broadinstitute.org/downloads)[2].
To focus on rare or very rare variants, we limited to variants below a pre-specified MAF threshold in both cases and controls. We used MAF ≤ 1% in the SCOOP cases vs. UK10K cohort controls analysis and MAF ≤ 0.1% in the SCOOP vs. ExAC analysis. For the SCOOP cases vs. UK10K controls analysis, we also applied external filtering excluding variants with a MAF > 1% in at least one of the 1000Genomes five primary ancestry groups. Exclusion by 1000Genomes MAF was not possible when using ExAC as 1000Genomes sample are included in the ExAC genotype frequencies. We explored the distribution of test statistics over several thresholds for the minimum number of functional (xf) and proxy (xp) variants within each gene (5, 10, and 20).
Analysis regions were limited to the intersection of respective target regions for SCOOP vs. UK10K Cohort and for SCOOP vs. ExAC. All variant annotation was applied using the GRCh37 human reference. The Ensembl Variant Effect Predictor (VEP, http://www.ensembl.org/info/docs/tools/vep/index.html [11] v79 and v90.1) from Ensembl was used to add variant consequence annotations for SCOOP vs. UK10K Cohort and SCOOP vs. ExAC respectively. We defined functional variation using the following Sequence Ontology terms[18] variant consequences: splice_donor_variant, splice_acceptor_variant, stop_gained, frameshift_variant, stop_lost, initiator_codon_variant, inframe_insertion, inframe_deletion, missense_variant, and protein_altering_variant. Variants were considered synonymous if they had the “synonymous_variant” flag. We defined the LOW proxy group as having a predicted low impact rating from Ensembl, SO terms: splice_region_variant, incomplete_terminal_codon_variant, stop_retained_variant, synonymous_variant.
We used quantile-quantile plots (QQ-plots) to assess the resulting distribution of test statistics from the real data applications. Specifically, we looked at the middle of the distribution of test statistics as assessed by the lambda value (i.e. the median of the observed test statistic divided by the median of the expected test statistic) and the tail of the distribution of test statistics, which we assessed visually.
|
10.1371/journal.pgen.1006895 | Division-induced DNA double strand breaks in the chromosome terminus region of Escherichia coli lacking RecBCD DNA repair enzyme | Marker frequency analysis of the Escherichia coli recB mutant chromosome has revealed a deficit of DNA in a specific zone of the terminus, centred on the dif/TerC region. Using fluorescence microscopy of a marked chromosomal site, we show that the dif region is lost after replication completion, at the time of cell division, in one daughter cell only, and that the phenomenon is transmitted to progeny. Analysis by marker frequency and microscopy shows that the position of DNA loss is not defined by the replication fork merging point since it still occurs in the dif/TerC region when the replication fork trap is displaced in strains harbouring ectopic Ter sites. Terminus DNA loss in the recB mutant is also independent of dimer resolution by XerCD at dif and of Topo IV action close to dif. It occurs in the terminus region, at the point of inversion of the GC skew, which is also the point of convergence of specific sequence motifs like KOPS and Chi sites, regardless of whether the convergence of GC skew is at dif (wild-type) or a newly created sequence. In the absence of FtsK-driven DNA translocation, terminus DNA loss is less precisely targeted to the KOPS convergence sequence, but occurs at a similar frequency and follows the same pattern as in FtsK+ cells. Importantly, using ftsIts, ftsAts division mutants and cephalexin treated cells, we show that DNA loss of the dif region in the recB mutant is decreased by the inactivation of cell division. We propose that it results from septum-induced chromosome breakage, and largely contributes to the low viability of the recB mutant.
| RecBCD protein complex is an important player of DSB repair in bacteria and bacteria that cannot repair DNA double-stranded breaks (DSB) have a low viability. Whole genome sequencing analyses showed a deficit in specific sequences of the chromosome terminus region in recB mutant cells, suggesting terminus DNA degradation during growth. We studied here the phenomenon of terminus DNA loss by whole genome sequencing and microscopy analyses of exponentially growing bacteria. We tested all processes known to take place in the chromosome terminus region for a putative role in DNA loss: replication fork termination, dimer resolution, resolution of catenated chromosomes, and translocation of the chromosome arms in daughter cells during septum formation. None of the mutations that affect these processes prevents the phenomenon. However, we observed that terminus DNA loss is abolished in cells that cannot divide. We propose that in cells defective for RecBCD-mediated DSB repair the terminus region of the chromosome remains in the way of the growing septum during cell division, then septum closure triggers chromosome breakage and, in turn, DNA degradation.
| Most bacteria have a circular chromosome on which replication is initiated at single origin oriC and proceeds bi-directionally on the two replichores until forks meet in the terminus, opposite to oriC. The chromosome terminus is a particularly active region where several important processes take place: replication termination, chromosome dimer resolution and last steps of chromosome segregation (Fig 1A). In E. coli, replication is arrested by the presence of sites called Ter that are bound by a specific protein Tus (reviewed in [1, 2]). Ter-Tus complexes allow replication to proceed in one direction only and thus create a replication fork trap in which replication forks enter but from which they infrequently exit. This system allows right and left replichores to be replicated principally in a clockwise and anti-clockwise direction, respectively, ensuring that replication is mainly co-directional with transcription [3–6]. 2D-gel analyses allowed the visualisation of replication forks arrested at TerC, and to a lesser extent at TerA and TerB [7].
In most bacterial species a site called dif, which is located opposite to oriC on the circular chromosome, is acted upon by the XerC/XerD site-specific recombination complex to resolve chromosome dimers (reviewed in [8, 9]). dif is also the site of inversion of the GC skew on the chromosome [2, 10] and the site of orientation inversion of biologically active motifs such as Chi (crossover hotspot instigator) and KOPS (FtsK-orienting polar sequences [11, 12] reviewed in [13]). KOPS (GGGNAGGG) are used by the septum protein FtsK to orient the translocation of chromosomes to daughter cells, and the convergence of these sequences at dif makes it the last segregated chromosomal sequence in slow growing cells [14]. Although KOPS are present all around the chromosome, FtsK is particularly active in the 400 kb region, centred on dif [15]. Translocation of chromosomes by FtsK is arrested by encounter with the XerCD-dif complex [16, 17]; FtsK then activates this complex to trigger chromosome dimer resolution at dif [18–20]. Finally, FtsK was proposed to displace the terminus-specific DNA-bound protein MatP [14], a protein that organizes and condenses the 780 kb Ter macrodomain by binding specifically to short DNA sequences called matS [21–23]. In this manuscript we call “terminus” the chromosome region opposite to oriC, centred on the point of inversion of the GC skew, regardless of the position of replication forks merging.
The septum forms at mid-cell by the assembly of several proteins in a defined order (reviewed in [24, 25]). Early proteins are FtsZ and its regulators, which include FtsA. Formation of the Z-ring is essential for the recruitment of late proteins, including FtsK and FtsI. FtsK is a bi-functional protein, its N-terminal domain is essential for cell division and is anchored in the septum; it is separated by a linker from the C-terminal domain, a cytoplasmic ATPase non-essential for viability and responsible for (i) DNA translocation in a direction imposed by KOPS, and (ii) activation of chromosome dimer resolution by interaction with XerCD (reviewed in [26, 27]). FtsI is a septum peptidase essential for constriction (reviewed in [28, 29]).
The terminus region was reported to be a preferential region of genetic instability [30–33]. However, hyper-recombination in the terminus region was dependent on replication termination at Ter sites, or perturbation of the dimer resolution system XerCD/dif, or perturbation of FtsK-mediated chromosome segregation, and it occurred in a small subpopulation of cells (at most 1%) [30–33]. More recently, a limited region of the terminus was reported to be amplified in certain mutant contexts (notably in a recG mutant) and a nearly identical region was lost in recB mutants [34, 35].
RecBCD is the enzyme that initiates recombinational repair of DNA double-strand breaks (DSB) in E. coli. This enzyme specifically recognizes DNA double-strand ends, unwinds and cleaves dsDNA via its coupled DNA helicase and exonuclease (exo V) activities, and when it encounters a Chi site it loads RecA on single-stranded DNA (ssDNA) [36–38]. Loss of terminus DNA in recB mutants was proposed to result from the formation of dsDNA ends by erroneous merging of replication forks leading to a transient over-replicated intermediate [35]. According to the proposed model, in the RecBCD+ context over-replicated dsDNA ends would be appropriately degraded by RecBCD, restoring intact chromosomes, while in the recB mutant extensive DNA degradation from these dsDNA ends by various single-stranded DNA exonucleases would cause DNA loss [35].
The amplification of terminal DNA in a recG mutant has been examined in detail [6, 34, 39, 40], while the DNA loss in the recB mutant has been less extensively explored, and the model of merging forks has not been directly tested experimentally [35, 41]. Here we use Marker Frequency Analysis (MFA) and live-cell fluorescence microscopy to further characterize this phenomenon. As previously proposed [34, 35], we consider that loss of terminus DNA in the recB mutant results from the degradation of unrepaired DNA double-stranded ends, but we show that it is independent of the position of replication termination, which argues against the model of merging forks. In search for an alternative source of chromosome breakage, we show that terminus DNA loss in the recB mutant occurs during cell division and requires septum formation. We propose that chromosome breakage in the E. coli terminus region is septum-induced damage. In addition, we observed weak Tus-dependent DNA loss at Ter sites, which was only detected when division was prevented by mutation or when replication terminates at an ectopic Ter.
Marker frequency analysis (MFA) of the chromosome of wild-type and recB mutant cells in exponential growth confirmed a deficit of DNA reads in the terminus region of the chromosome in the absence of RecB (Fig 2A, S1 Fig). To allow a direct comparison of MFA and microscopy results, all experiments were done in minimal medium (M9). This DNA loss is centred on the dif-TerC region when cells are grown exponentially in M9 glucose (Fig 2A), as previously reported for cells grown in LB [34, 35]. We have developed a live-cell microscopy approach to confirm that this phenomenon, which is observed by MFA in a population of growing cells, in fact occurs in a sub-population of individual cells and we have quantified this sub-population. Strains constitutively expressing the yGFP-ParBpMT1 protein from a chromosomal-borne gene and carrying parSpMT1 sites inserted at three different locations were used. Binding of the yGFP-ParB protein to its parS target site allows the visualisation of each chromosome parS sequence as a fluorescent focus. Three strain backgrounds carrying each a different parS sites were used. These parS sites were ydeV::parSpMT1 located between dif and TerC, 10 kb away from each site, yoaC::parSpMT1 located about 300 kb away from dif on the left replichore and ycdN::parSpMT1 located about 500 kb away from dif on the right replichore ([42], S1 Table, Fig 1B). Cells grown in exponential phase in M9 glucose medium were observed by fluorescence microscopy. In a wild-type context, nearly all cells showed foci and the proportion of cells with two foci increased with the distance from dif, as previously reported (S2 Table) [14, 43]. In the recB mutant, the proportion of cells without any focus was much higher than in RecB+ cells: the recB mutant with ydeV::parSpMT1 (the locus between dif and TerC) showed 32% of cells with no focus and the control parS sites located 300 or 500 kb from dif showed 7–8% of cells with no focus (Table 1, S2 Table). These results argue that in a recB context about one third of cells have lost the dif-TerC region specifically, in agreement with the results of MFA experiments [34, 35] (Fig 2A).
To better characterize this chromosomal DNA loss, the dynamic behaviour of foci was tracked by time lapse microscopy of recB cells growing on M9 glucose agarose pads, as described in Materials and Methods. As shown in Fig 2F and S1 and S2 Videos, ydeV::parSpMT1 foci were lost with the following characteristics: 1) the foci disappeared concomitantly with cell division, most often at the site of septum formation, and in one of the two daughter cells only (yellow stars), no loss at any other time point was observed, 2) the loss occurred after duplication of this region, since most of the time two foci were clearly visible at earlier time points (white arrows), 3) after cell division, the daughter cell that had lost a focus stopped growing and did not divide, whereas the cell that retained a focus divided again, and produced a focus-less cell at each generation after the first event (yellow stars). Although the production of focus-less cells was asymmetrical and hereditary, we did not observe any bias toward the old or new pole. We called the first division that produces one focus-containing cell and one focus-free cell “the initial event” and calculated that these represented 17.7% of cell divisions (not counting the “secondary events” that follow, Table 1). Because divisions that produced focus-less cells also produced a focus-containing cell, the proportion of focus-less cells in the population was expected to be one half of the proportion of divisions that produced them in the absence of transmission of the phenomenon to progeny. In contrast to this, the proportion of focus-less cells (32%) was higher than the proportion of divisions that produced them (17.7%), in agreement with the transmission of the phenomenon to following generations (Table 1). Nearly 75% of the initial events were transmitted to the progeny for as many generations as we could see (up to five). In addition, in about 17% of the cases transmission was interrupted for one generation (i.e. the cell that retained a focus produced two focus-containing cells, one propagated normally and the other one resumed the production of one focus-less cell at each generation). Although there is clearly some transmission in spite of the interruption, this second category of events was observed in all recB mutants and is not counted as transmitted in Table 1.
RecBCD has two activities, a recombinase activity that requires RecB and RecC but not RecD (helicase and RecA loading activities), and an exonuclease activity, called exoV, which degrades linear dsDNA and is catalyzed by the entire RecBCD complex [36, 38]. DNA degradation by exoV occurs in the absence of RecA or in the absence of Chi sites. recD mutant cells are recombination proficient but do not degrade dsDNA, and it was previously reported that terminus DNA loss is not detected by MFA in the recD mutant [35]. Accordingly, analysis of ydeV-parSpMT1 foci showed that the recD single mutant behaved like wild-type (Table 1), therefore we confirm that the presence of RecBC prevents DNA loss regardless of exoV activity. In contrast, the recC mutant that lacks both RecBC and RecBCD complexes and thus both recombination and exoV activities, behaved as the recB mutant (Table 1, S1 Fig).
DNA loss was previously shown by MFA to occur in the dif-TerC region of recB mutants in two different E. coli genetic backgrounds: MG1655 [34], in which replication terminates primarily at TerC and 4 to 5 times less often at TerA [7], and in W3110 [35]. W3110 carries a large inversion between rrnD and rrnE around the replication origin, which enlarges the right replichore and shortens the left one by about 220 kb [44, 45] (Fig 1A). As a consequence of this inversion, the closest replication terminator from oriC in this context is TerA and not TerC, and therefore replication is expected to terminate at TerA more often than at TerC. Nonetheless, the position of the peak of DNA loss was the same in W3110 as previously reported in MG1655 [35]. This surprising observation prompted us to directly measure the influence of the position of replication termination on DNA loss in the dif/TerC region. For that purpose, we first compared recB and recB tus mutants (in which Ter sites are non-functional) by MFA (Fig 2B, S1 Fig) and by snap-shot fluorescence microscopy of ydeV::parSpMT1/yGFP-ParBpMT1 foci (Table 1). Inactivation of tus did not prevent DNA loss in the dif-TerC region detected by MFA (Fig 2B) and did not modify the percentage of focus-less cells for the ydeV::parSpMT1 locus close to dif, and for the yoaC::parSpMT1control locus (Table 1). Note that the ratio of reads in tus recB over the tus mutant increased in a large Ter region compared to the rest of the chromosome (Fig 2B). We do not know the reasons for this phenomenon, but the existence of a mixed population partially masks DNA loss at dif/TerC in the MFA experiment. Time lapse experiments showed that the loss of ydeV::parSpMT1 foci followed the same scheme in tus recB as previously observed in Tus+ recB cells, i.e. loss of focus in one daughter cell, at the septum, at the time of division, and transmission of this defect to the progeny (S3 Video), with a similar percentage of initial events (Table 1, 15.8% vs 17.7%), and a high level of events transmitted to progeny (87%). Replication was previously shown by 2D gels to terminate mainly at TerC in wild-type cells [7], but our result in the tus mutant suggests that most of the loss of DNA in the dif-TerC region occurs independently of forks merging at TerC.
Because in the tus mutant forks may still merge in the dif region that is opposite to the origin, we constructed a strain in which the clockwise replication forks are prevented from reaching TerC by the introduction of an additional TerB site that arrests replication prematurely, 29 kb downstream of TerA (pspE::TerB, TerB* in Fig 1A and 1B). In the TerB* strain a new replication fork trap is created between TerA and TerB* and the dif site is mainly replicated by the counter-clockwise fork, instead of the clockwise fork in the wild-type strain. Fluorescence microscopy showed that loss of ydeV::parSpMT1 foci (close to dif) was increased to 48% in the strain containing TerB*, while loss of control yoaC::parSpMT1 and ydvN::parSpMT1 loci was unchanged (Table 1, S2 Table). As expected, the increase from 32% in recB to 48% in TerB* recB cells was Tus-dependent, as we counted 35% of ydeV::parSpMT1 focus-less cells in the TerB* recB tus mutant (Table 1, S2 Table). Time lapse experiments showed that in TerB* recB cells ydeV::parSpMT1 foci were still lost in one daughter cell, at the septum, at the time of division, with a transmission of this defect to the progeny (Fig 2G, Table 1), and the proportion of original divisions that yielded the first focus-less cells in an inheritable manner was increased from 17.7% to 25% (Table 1). Finally, measuring loss of control yoaC::parSpMT1 foci (Table 1) and MFA analysis showed that the position of the peak of DNA loss at dif was not affected by the creation of this new replication fork trap, away from TerC (Fig 2C, S2 Fig). Nevertheless, the MFA experiment also revealed a new Tus-dependent peak of DNA loss at TerB*, weaker than the DNA loss at dif (Fig 2C and 2D, S2 Fig). It was previously proposed that DNA double-strand ends, target for RecBCD, were formed in the terminus region of the recB mutant by erroneous merging of replication forks, and that degradation of these unrepaired DNA double-stranded ends by the combined action of helicases and single-stranded exonucleases was the cause of terminus DNA loss [34, 35]. Since in the presence of the additional TerB* site, replication forks are expected to merge at this site or at TerA, and are therefore unlikely to merge at TerC, the strong DNA loss that we observed in the dif/TerC region of the TerB* recB mutant (Fig 2C, Table 1) cannot result from replication fork merging. We also propose that DNA loss results from DNA degradation by the combined action of helicases and single-stranded exonucleases, but propose that the dsDNA ends on which these enzymes act are produced by a DNA DSB occurring in dif/TerC region of the recB mutant chromosome, regardless of the position of replication termination. In addition, our results also show that displacing replication termination to TerB* creates a new hotspot of DNA loss, weaker than the dif/TerC hotspot, at the new replication termination site.
The observation that the position of DNA loss in the dif-TerC region is independent of the position of replication fork merging raises the possibility that it might be determined by dif rather than TerC. dif is the site of chromosome dimers resolution, a XerCD- and FtsK- dependent reaction [18–20]. We tested a possible role of dimer resolution in DNA loss by inactivating xerC or removing the dif site. In RecB+ cells these mutations increased the proportion of ydeV::parSpMT1 focus-less cells from less than 1% to about 15% (Table 1, S2 Table) and, accordingly, a weak DNA loss of the dif region could be detected by MFA (S3 Fig). Time lapse experiments showed that focus-loss in the dif or xerC single mutants results from breakage of both chromosomes at the time of cell division (Fig 3E, S4 Video); Abnormal pattern of cell division in microcolonies of dif and xer mutants was previously observed and was proposed to result from breakage of chromosome dimers by septum closure, which was called guillotining [46, 47]. 40–42% of cells lacked ydeV::parSpMT1 in xerC recB or dif recB mutants (Table 1). We propose that this higher level of focus-less cells compared to the recB single mutant results from a combination of broken dimers and septum-induced breaks (about 50% of dimers are RecB-independent, [48]). Accordingly, time lapse microscopy confirmed that some focus-less cells result from the concomitant loss of both ydeV::parSpMT1 foci in the two daughter cells at the time of division (presumably dimer breakage), whereas most of them result from the transmitted, asymmetric loss of one focus in one daughter cell at the time of division (Fig 3E). Interestingly, in dif or xer cells that contain a dimer, cell division was delayed for more than an hour before we observed cell separation and focus-loss (S4 Video), while cell division of the recB mutant was not delayed upon focus-loss in one daughter cell compared to cells that do not lose foci. The proportion of initial events in the dif recB mutant was 17.4%, similar to the recB single mutant, confirming that division-dependent loss of ydeV::parSpMT1 foci is independent from dimer resolution, and most of these events were transmitted to progeny (Table 1). Furthermore, the weaker loss of the control yoaC::parSpMT1 and ycdN::parSpMT1 foci compared to loss of the dif proximal ydeV::parSpMT1 site (Table 1, S2 Table), and MFA analyses of xerC and xerC recB mutant (Fig 3A, S3 Fig) confirmed that the loss of DNA remains centred on dif in the absence of dimer resolution.
Two dif deletions were tested: one that lacks only the dif site, and one that also inactivates the adjacent hipA locus. HipA is a toxin that blocks growth by inactivating translation and is counteracted by the short-lived anti-toxin HipB ([49] and ref therein). In the absence of both dif and HipA, the peak of DNA loss observed in MFA experiments in a recB mutant context was deeper and larger than in the recB single mutant (Fig 3B, S3 Fig). Accordingly, in microscopy snapshot experiments the proportion of cells that lack the ydeV::parSpMT1 or the control yoaC::parSpMT1 focus was much increased, to 65% and 40% respectively (Table 1, S2 Table). Time lapse experiments showed that in the absence of HipA, focus-less cells grew and divided for several generations, which increased the proportion of these cells in the population, and presumably allowed the degradation of more and more chromosomal DNA with generations (Fig 3F). This result showed that in recB single mutants the growth arrest of focus-less cells results from the degradation of the hipAB locus after chromosome breakage, which causes the accumulation of active toxin HipA. This phenomenon was previously described after breakage of chromosome dimers at the septum in a dif mutant [46]. It confirms genetically that chromosomes lacking the ydeV::parSpMT1 site originally conserve all genes required for growth and cell division, and that chromosome degradation is a slow process (essential genes are absent from the terminus region, [47, 50]).
If DNA loss results from DSBs occurring at the peak of loss followed by nearly symmetrical DNA degradation by single-stranded nucleases, why are these DSBs introduced in the dif region even in the absence of this site? The dif site is the last chromosome locus segregated in daughter cells because it is the site of convergence of the KOPS sequences, which are used by the FtsK protein to segregate replicated chromosomes to daughter cells ([14, 27], Fig 1A). KOPS sites are present all around the chromosome but FtsK is mainly active in a 400 kb region approximating the one of decreased reads in the recB mutant [15]. We tested a putative role of FtsK in the localisation of terminus breaks with the use of a ftsKATPase mutant, in which ftsK carries a nucleotide substitution that specifically inactivates the ATPase activity and thus prevents DNA translocation without affecting DNA binding. We first analyzed ftsKATPase RecB+ cells by microscopy. Quantification of focus loss showed that the proportion of focus-less cells was increased compared to wild-type cells, particularly for the ydeV::parSpMT1 site located next to the dif locus (from 0.6% to about 20%, Table 1, S2 Table). This DNA loss presumably resulted mainly from a lack of dimer resolution in the absence of FtsK translocation activity. Inactivation of recB in the ftsKATPase mutant led to a large increase in focus-less cells (nearly 55% of cells contain no ydeV-parSpMT1 focus and 14% contain no yoaC-parSpMT1 focus). Similar results were observed when FtsK translocation was inactivated by the deletion the entire protein C-terminal domain (Table 1, S2 Table). Time lapse microscopy showed two kinds of events leading to focus loss in the ftsK recB mutant context. In 15.5% of the divisions, one ydeV-parSpMT1 focus was lost at the septum, at the division time, in one daughter cell only, with a transmission of this defect to the progeny (Table 1, S5 Video). This result shows that the events occurring in the recB single mutant also occur in ftsKATPase recB cells, with a similar frequency (Table 1). In addition, the two daughter cells lost the ydeV-parSpMT1 foci during 12% of the divisions, presumably owing to dimer breakage (S5 Video), and other types of focus loss could be observed, which presumably result from the segregation defect and account for the high percentage of focus-less cells in the ftsKΔCTer recB mutant population (S6 and S7 Videos).
In contrast with the dif and xerC mutants, DNA loss around dif was not detected by MFA in the ftsKATPase single mutant, and rather a weak DNA amplification was visible in the terminus region (S3 Fig). Since microscopy results show a loss of the dif region in 20% of ftsKATPase and ftsKΔCTer mutants, this amplification reflects the existence of a mixed population of cells, some that lose the dif/TerC region as observed by microscopy, and some that amplify it and mask the loss in the MFA experiments. DNA degradation around the dif locus was observed by MFA in the ftsKATPase recB mutant (Fig 3C, S3 Fig), but a larger DNA region was degraded than in the recB single mutant (compare Fig 3C and Fig 2A). Furthermore, the MFA did not show the deep loss expected from the microscopy results. This loss could be masked if DNA amplification occurs in a subset of cells, as detected by MFA in single ftsKATPase mutant. We conclude that the DNA translocation activity of FtsK plays an important role in the sharp targeting of DNA loss to the dif region in the recB mutant, leading to a wider distribution of DNA loss in the absence of the FtsK C-terminal domain or ATPase activity. Nevertheless, DNA loss and therefore DNA breaks still occur specifically in the dif/TerC chromosome region when DNA translocation by FtsK is inactivated.
To know whether in wild-type cells RecBCD acts in the dif/TerC region, we investigated RecA binding by ChIP followed by qPCR of sequences upstream and downstream of the first Chi site on each side of the dif/TerC region (Fig 4, RecBCD loaded at a DSB in the dif/TerC region unwinds DNA toward the origin, until it encounters properly oriented Chi sites at which it loads RecA). As previously reported [51] we detected a weak increase of RecA ChIP downstream of Chi when cells were grown in LB, but we did not detect any increase in cells grown in M9 glucose (Fig 4). Similar results were obtained in cells that over-express RecA owing to a mutation in the recA gene SOS operator (Fig 4). We conclude that in minimal medium, DNA breakage in the dif/TerC region does not occur in RecB+ cells, and thus only occurs in recB mutants.
To address the question of the origin of chromosome breaks in the terminus region of a recB mutant two enzymes that cleave DNA were tested, Topoisomerase IV (Topo IV) and endonuclease I (Endo I). Topo IV, encoded by the parC and parE genes, catalyzes the decatenation of daughter chromosomes after replication. Topo IV interacts with both XerC and FtsK, its decatenation activity is stimulated by its interaction with FtsK in vitro and a hotspot of activity was detected in vivo close to dif, which is dependent on its interaction with XerCD [52–55]. We hypothesized that if catenated chromosomes persist in the path of septum closure, an erroneous action of Topo IV during decatenation could be responsible for chromosome breakage. Because the inactivation of Topo IV by a ts mutation prevents cell division, and cell division is required for the DSBs studied here (see below), the effects of a parE10ts mutation were tested by MFA at 37°C, where Topo IV is impaired but cell division is not affected enough to prevent cell growth [56]. As shown in S4 Fig, the parE10ts mutation did not affect DNA loss in the dif region at this semi-permissive temperature. Furthermore, interaction of Topo IV with XerCD is required to target its action close to dif [55], therefore the lack of effect of xerC or dif inactivation described above argues against a direct action of Topo IV to break DNA next to dif. This was confirmed by using the observation that the Topo IV hotspot next to dif is abolished by over-production of the C-terminal region of ParC from a plasmid, (parC-CTD plasmid, [55]): the presence of this plasmid did not affect focus loss (Table 1, S4 Fig). Therefore, the division-dependent DSBs studied here do not result from an erroneous action of Topo IV at dif.
The most abundant endonuclease in E. coli is the periplasmic enzyme Endo 1. We hypothesized that a leak of Endo 1 from the periplasm to the cytoplasm during division might cause cleavage of one chromosome in the terminus region. However, the inactivation of the endA gene encoding Endo 1 did not affect the proportion of focus-less cells in a recB context (Table 1, S2 Table). The enzyme(s) that introduces the DSBs responsible for DNA loss remain unidentified.
Because time lapse experiments showed that the ydeV::parSpMT1 focus was often lost at the septum and always concomitantly with cell division, we tested whether septum formation plays a role in the loss of the dif region. We used three different conditions that affect cell division: ftsAts and ftsIts thermosensitive mutants, which block an early and a late step of divisome assembly at 42°C, respectively, and cephalexin, a drug that prevents the action of FtsI (reviewed in [24, 25]). Because blocking division produced cells that were highly elongated and difficult to analyse by microscopy, the loss of terminus DNA was examined by MFA. The ftsAts recB mutant was compared at 30°C and after 2 hours of incubation at 42°C, to block division (Fig 5A, S5 Fig). Loss of DNA centred on dif was observed at 30°C where FtsA is active, but was much weaker at 42°C. A similar result was obtained with the ftsIts recB mutant (Fig 5B, S5 Fig). Comparison of a cephalexin-treated with an untreated recB mutant showed that cephalexin also prevented DNA loss at dif (compare S5 with S1 Fig). Therefore, results in all three cases indicated that loss of DNA in the dif region of a recB mutant is decreased when septum assembly is prevented. Control experiments showed that DNA loss was similar in a recB single mutant at 37°C and at 42°C (S5 Fig). Ratios of recB over fts recB mutants grown at 42°C revealed two phenomena (Fig 5C and 5D): (i) they confirmed that the loss of reads centred on dif/TerC is specific to dividing cells, (ii) they revealed 5–10% more reads in the TerA/TerD region in dividing versus non-dividing cells. The latter observation suggests that in a recB context blocking cell division causes a slight loss of TerA/TerD sequences. When results for the fts recB cells grown at 30°C and 42°C were compared this increase was not observed (ftsAts) or weak at TerD (ftsIts), presumably because cell division is partially affected in these mutants at the permissive temperature (compare Fig 5A and 5B with 5C and 5D).
Comparisons of RecB+ and recB mutants in ftsIts (or ftsAts) contexts at 42°C, revealed two regions of DNA loss caused by recB inactivation, weaker than in dividing cells and centred on TerA and on the dif/TerC region (Fig 5E and 5F). The effects of recB inactivation in ftsIts tus and ftsAts tus mutants were analyzed (Fig 5G and 5H). The small peak in the TerA region disappeared in the tus context, showing that it results from replication arrest at TerA (compare Fig 5G and 5H with Fig 5E and 5F, S6 Fig). Accordingly, in the ftsIts context the weak recB-dependent DNA loss in the TerB/TerC region was displaced to the dif region when tus was inactivated. The peak of DNA loss at dif in fts tus recB mutants is weaker than in dividing cells (compare Fig 5G and 5H with Fig 2B and 2D), accounting for the difference between dividing and non-dividing cells shown above (Fig 5A, 5B, 5C and 5D). We conclude from these experiments that (i) DNA loss around dif in the recB mutant is decreased by inactivating cell division (Fig 5A–5D), and (ii) weaker peaks of DNA loss that require the Tus protein can be observed at TerA in non-dividing recB mutants (Fig 5E–5H).
To further test whether the site of convergence of GC skew determines the localisation of DNA loss, we constructed mutants in which a new GC skew convergence zone was created in the terminus region. First, we used a strain where the entire terminus region from 1 379 810 to 1 617 226 was deleted (ΔLC3-R111 strain, Fig 1C). This 237 kb deletion removes half of the DNA region degraded in the recB single mutant including dif, hipA and TerC. It defines a new 102 kb replication fork trap between TerA and TerB and creates a new GC skew converging zone at the junction, next to which we inserted a parS site (pspE::parSpMT1). Because the ΔLC3-R111 mutant lacks the dif site, it showed 18% of focus-less cells. In the ΔLC3-R111 recB mutants 49% of cells were devoid of a pspE::parSpMT1 focus (Table 2). Time lapse analyses showed that the loss of a pspE::parSpMT1 focus in the ΔLC3-R111 recB resulted mostly from loss of one focus in one daughter cell at the time of division as in the original recB mutant (Fig 6A); we counted 19% initial events and 77% of them were transmitted to progeny (Table 2). Focus loss also occurred at a lower frequency in both daughter cells at the time of division, presumably resulting from dimer breakage (Fig 6A). The ΔLC3-R111 deletion removes dif and hipA, but shortens the region devoid of essential genes that can be degraded without preventing cell propagation. Accordingly, we observed one or two divisions of focus-less cells owing to the absence of hipAB.
For unknown reasons the MFA analysis of the ΔLC3-R111 chromosome (S7 Fig), showed a breakpoint in the read copy number around TerA, which was not detected in the recB mutant (Fig 7A). The ratio of reads in recB mutant over RecB+ cells is affected by this breakpoint, and in Fig 7A we present directly the MFA result of the ΔLC3-R111 recB mutant. The peak of DNA loss measured by MFA was located at the new junction of the chromosome arms, about 65 kb from TerB (Fig 7A). Similarly to the lack of effect of tus inactivation on DNA loss in the recB mutant (Table 1), inactivating tus in the ΔLC3-R111 recB mutant did not prevent DNA loss at the GC skew convergence point (Table 2). For unknown reasons, inactivation of the DNA translocation activity of FtsK in the ΔLC3-R111 mutant led to an amplification centred on the midpoint between TerA/D and TerF and the breakpoint between TerA and TerB was not detectable (ΔLC3-R111 ftsKΔCTer S7 Fig). However, DNA loss occurred in ΔLC3-R111 ftsKΔCTer recB as in ΔLC3-R111 recB, and the inactivation of FtsK translocase slightly widened the maximum point of DNA loss, to the entire 105 kb region between TerB and TerA (Fig 7B), These results show that the position of division-induced DSBs is determined by the point of GC skew convergence, in a way that is independent of the sequence of this junction, and is more precisely targeted to the KOPS convergence point in the presence of the FtsK translocation activity.
We then created a new GC skew convergence zone by inverting a region of the terminus. In the InvT3 mutant, a ~175 kb sequence is inverted on the right chromosome arm, which does not contain any Ter site and starts about 34 kb from dif (Fig 1B). In this strain the main GC skew convergence zone is moved 209 kb to the left of dif, and the dif position becomes a minor convergence zone with on its left only 34 kb of DNA in the original orientation. InvT3 and InvT3 recB strains were compared by MFA (Fig 7C, S8 Fig). Inactivation of recB in InvT3 created a new degraded region corresponding to the entire inverted sequence, but no peak of DNA loss at the new convergence zone. Importantly, the main DNA degradation peak in the dif region was still present (Fig 7C, S8 Fig). Microscopy analysis confirmed a specific loss of the dif region by showing that the proportion of ydeV::parSpMT1 focus-less cells was similar in InvT3 recB and recB mutants (38% and 32% respectively, Table 1, S2 Table; the additional focus-less cells observed in InvT3 RecB+ could result from a perturbation of segregation because of KOPS inversion, causing irreparable damage). The inactivation of the ATPase function of ftsK enlarged the degraded region but the maximum of DNA loss was still in the dif region (Fig 7D). Comparison of FtsK+ and ftsK mutant MFA in the recB context suggests that FtsK-mediated translocation slightly protects the inverted region from degradation. It is also interesting to note that in the absence of the FtsK translocation activity, loss of the reads in the 1380–1554 kb region in the recB mutant was higher when this sequence was inverted than when it is in the original orientation (compare Fig 7D with Fig 3C). This observation suggests the existence of a system other than FtsK able to detect the sequence orientation. Nevertheless, these results also indicate that a 175 kb GC skew convergence zone is not sufficient to create a division-induced DSB.
In the InvT2 mutant a 150 kb region encompassing the TerA and TerD sites and located 209 kb from dif is inverted (Fig 1B). Clockwise replication forks are expected to be arrested in the TerA-TerD region in this mutant, and the replication fork trap is moved between the inverted TerA and TerE. Accordingly, the MFA profile of the InvT2 mutant shows that replication forks meet in the TerA-TerD region, ~250–300 kb away from dif (S9 Fig). As for the LC3-R111 mutant, direct results of the InvT2 MFA are shown in Fig 7E because the breakpoint of read copy number around TerA affects the ratio of RecB+ versus recB mutant reads (Fig 7, S9 Fig). Two peaks of DNA degradation were clearly detected by MFA in the InvT2 recB mutant: the one centred at dif observed in all recB mutants and a new one coincident with the inverted TerA site (Fig 7E, S9 Fig). The inactivation of tus suppressed the TerA-associated degradation but not DNA loss at dif, and allowed the detection of some DNA degradation associated with the inversion region, as in InvT3 (Fig 7F, S9 Fig). The observation that the InvT2 inversion did not affect inheritable division-dependent focus loss was confirmed by microscopy, as the proportion of ydeV::parSpMT1 focus-less cells increased in InvT2 from 8% to 42% upon recB inactivation (Table 1, Fig 6B). The observation of Tus-dependent DNA loss at TerA confirms that DNA breakage occurs at an artificially introduced Ter site that creates a new replication fork trap, as observed with TerB* (Fig 2C and 2D). In addition, these results confirm that division-induced DSBs in the dif region are not affected by the creation of a new GC skew convergence zone, as observed with InvT3. Furthermore, in InvT2 as in InvT3 the new DNA convergence zone does not show a peak of DNA loss, in contrast with ΔLC3-R111, but the number of reads in the whole inverted region is lower than when this sequence is not inverted (compare Fig 7C with Fig 2A and Fig 7F with Fig 2B).
In the present study, we reveal that the DNA region of the E. coli chromosome terminus, previously shown to be under-represented in the recB mutant [34, 35], is lost following division-dependent chromosome breakage (Fig 5). We have demonstrated that DNA loss in the dif region of recB mutants occurs with the following characteristics: (i) after duplication of the region, (ii) at the time and often at the site of cell division, (iii) in one of the two daughter chromosomes, and (iv) is transmitted to progeny (Figs 2, 3 and 6 and S1–S6 Videos). This DNA loss is independent of the position of replication termination, as we observed that DNA loss at dif/TerC is unaffected in a tus mutant, or when replication forks are prevented from reaching this region by pspE::TerB* or by the inversion of TerA-TerD in the InvT2 mutant (Figs 2, 6 and 7, Table 1). It occurs in the absence of chromosome dimer resolution (not affected in dif and xerC mutants, Fig 3, Table 1). In the absence of FtsK-driven DNA translocation, terminus DNA loss is less precisely targeted to the KOPS convergence sequence (Fig 3 and Fig 7), but follows the same pattern as in FtsK+ cells (Table 1, S5 Video). DNA loss can occur at least at two different GC skew convergence zones regardless of their sequence (recB and ΔLC3-R111 recB, Figs 6 and 7, Tables 1 and 2), but DNA loss at the natural GC skew convergence point is not affected by a nearby 150–175 kb inversion (InvT2 and InvT3 inversions, Figs 6 and 7, Table 1). The only mutations that strongly decrease terminus DNA loss in a recB context are those that block cell division (ftsAts and ftsIts mutants at 42°C, cephalexin treatment at 37°C; Fig 5). A schematic representation of terminus DNA loss according to our results is shown in Fig 8.
Septum-induced breakage was previously reported in xer and dif mutants, in which chromosome dimers are not resolved to monomers and remain in the path of the closing septum [46]; as expected we observed dimer breakage in our experiments, which occurs specifically in mutants affected for dimer resolution (xer, dif, ftsK) and is characterized by a loss of ydeV-parSpMT1 foci in both daughter cells along with a significant delay in cell division. Dimer breakage during septum formation was called guillotining, a term that does not describe precisely the molecular events leading to DNA DSBs. If we assume that chromosome dimer breakage results from physical tension associated with the pulling of two linked chromosomes during segregation, then the breakage of one chromosome observed in recB cells implies that this chromosome is broken as a consequence of being attached in the terminus region while the origin is gradually pulled towards the daughter cell. This attachment could be a covalent link with the other daughter chromosome after replication completion, or a strong binding to a septum protein. It is unlikely that this link is topological, since DNA loss is unaffected in conditions that perturb Topo IV action. It is noticeable that breakage occurs without any delay in cell division, in contrast with dimer breakage. Alternatively, breakage could be enzymatic, but the nature of the nuclease remains mysterious.
Importantly, we did not detect any focus loss in the recB mutant at any other time point in the cell cycle than cell division, and our measures of replication speed based on the MFA results show that the recB mutation does not affect replication progression (the ratio of recB versus wild-type reads is constant all along the chromosome except at the terminus and equal to 1 in Fig 2A). This is in agreement with the recently published results, where authors using flow-cytometry analysis concluded that absence of RecB does not affect chromosome replication speed [57]. Hence, we propose that the main source of chromosome breakage in the recB mutant grown in M9 is not replication fork impediments but rather division-induced breaks in the terminus region of the chromosome.
Following chromosome breakage, degradation of the DNA double-stranded ends by exonucleases is responsible for DNA loss. This step was postulated but never demonstrated [34, 35], and so far the formal possibility of under-replication of the dif region being responsible for the low number of reads observed in MFA experiments could not be excluded. In time lapse experiments, we most often see two ydeV-parSpMT1 foci before one of them disappears, which ascertains for the first time the presumed assumption that DNA loss does not result from a lack of replication but from DNA degradation of a replicated chromosome. Furthermore, the dif hipA and the ΔLC3-R111 mutants behaved as expected i.e., in the absence of HipA-induced cell death, the broken chromosomes are slowly degraded and cells with a broken chromosome propagate until degradation by exonucleases reaches essential genes.
DNA loss occurs at the site of GC skew convergence, and is observed at the new GC skew convergence zone in the strain that carries a large terminus deletion, confirming that the phenomenon is not DNA sequence specific. In addition, DNA loss is not affected by replication orientation, which progresses across dif in the clockwise direction in the majority of wild-type cells [7], but not when replication is arrested prior to dif by an ectopic TerB* site (pspE::TerB) or by inverting TerA and TerD (InvT2). We have shown that division-induced chromosome breakage is independent of any specific DNA sequence. These observations support a model in which the chromosome terminus region is somehow specifically and precisely positioned in the path of the division machinery. This positioning is more centred on the KOPS convergence zone when FtsK translocation is active, but remains centred on dif in FtsK mutants, and therefore relies on a so far unknown process. Division-induced chromosome breakage occurs in a sub-population of dividing recB cells when the positioning is inappropriately controlled. In addition to causing breakage of one chromosome, the improper processing of the terminus leaves a mark on the intact chromosome, which is responsible for the transmission of the defect to the next generation.
In addition to the septum-induced DNA DSBs described above, we observed a Tus-dependent loss of reads, suggesting DNA breakage in specific recB mutant conditions: (i) at ectopic or inverted Ter sites (pspE::TerB Fig 2C and 2D, TerA in InvT2 Fig 7E and 7F) and (ii) in the TerA-TerD region of cell division mutants (ftsAts and ftsIts at 42°C, cephalexin treated cells, Fig 5). The loss of reads around Ter sites is symmetrical, indicating that DNA breakage does not occur after blockage of the first fork that reaches Ter (only the replicated side of Ter would then be degraded). Previous studies showed that forks blocked at ectopic Ter sites are stable, and that DNA double-strand ends are formed at such Ter sites upon arrival of a second round of replication behind the first blocked one, by rear-ending, but in these previous mutant strains fork-merging was prevented [58, 59]. The Tus-Ter specific DNA breaks observed in the present work could therefore result from abnormal replication forks merging at Ter sites. However, although our MFA experiments are only semi-quantitative, Tus-dependent DNA loss at Ter sites seems to occur in a lower proportion of cells than division-induced breaks, with a weaker peak of DNA loss than the peak of division-induced DNA loss. Furthermore, the absence of DNA loss at TerA in wild-type cells shows that Tus-Ter induced breaks do not occur at a natural Ter site in cells that divide normally. The observation that blocking cell division triggers Tus-dependent DNA loss at TerA suggests an unknown link between replication fork merging and cell division, a link which would also be perturbed in dividing cells by arresting forks at an ectopic Ter site. Our observation suggests that replication termination at TerC, or forks merging at other sequences than Ter sites, is the most favourable condition during normal cell division and any change in this arrangement leads to loss of DNA at the new active Ter site. Further work will be needed to understand how Ter/Tus dependent DNA DSBs are made, and whether a common mechanism is involved upon division blockage at the natural TerA site and in dividing cells at ectopic Ter sites.
In conclusion, we have shown that DNA degradation in the GC skew convergence region occurs in a subpopulation of recB growing cells. The reaction is transmitted to progeny and is strongly decreased when cell division is prevented. It is targeted to the KOPS merging zone by the translocase activity of FtsK, and occurs in a broader chromosome terminus region in mutants that lack this activity. Since our time-lapse experiments did not show any growth defect or loss of focus in the recB mutant at any other time than cell division, we propose that division-induced DNA breakage could be responsible for the decreased viability of recB cells under normal laboratory growth conditions. These findings open new fields of investigation in search for the molecular mechanism responsible for this reaction.
All E. coli strains are derivatives of MG1655. Strains and plasmids are described in S1 Table. MM is M9 [60] supplemented with 0.4% glucose. Standard transformation and P1 transduction procedures were as described [60]. pspE::TerB-CmR, endA::KanR, araC::parBpMT1-CmR, araC::parBpMT1-ApraR, ydeV::parSpMT1-ApraR mutations were constructed by gene replacement (recombineering) as described in [61], using DY330 [62]. All other strains were constructed by P1 transduction. All mutations introduced by P1 transduction were checked by PCR and all new mutations constructed by recombineering were checked by PCR and sequencing. recB mutations were tested by measuring UV sensitivity. Deletion ΔR111-LC3 and inversions InvT2 and InvT3 were made as described [63].
The araC open reading frame was replaced by the yGFP-parBpMT1-CmR sequence, to express yGFP-ParB protein under the control of the constitutively expressed araC promoter. For this construction, the Cm gene was amplified from pKD3 plasmid using primers harboring an HindIII site (S3 Table). Amplified fragments were digested with HindIII and cloned into the HindIII site of pFHC2973 to make pFHC2973- yGFP-parBpMT1-CmR plasmid (Nielsen et al., 2006). Clones were confirmed by PCR and sequencing using flanking primers. The yGFP-parBpMT1-Cm fragment was then amplified from pFHC2973- yGFP-parBpMT1-CmR plasmid using primers 582 and 583 (S3 Table) and inserted downstream of the araC promoter by the gene replacement method (recombineering) as described in [61], using DY330 [62].
Cells were grown in M9 minimal medium supplemented with 0.4% glucose to exponential phase (0.2 OD 650 nm). Chromosomal DNA was extracted using the Sigma GenElute bacterial genomic DNA kit. 5 μg of DNA were used to generate a genomic library according to Illumina's protocol. The libraries and the sequencing were performed by the High-throughput Sequencing facility of the I2BC (http://www.i2bc.paris-saclay.fr/spip.php?article399&lang=en, CNRS, Gif-sur-Yvette, France). Genomic DNA libraries were made with the ‘Nextera DNA library preparation kit’ (Illumina) following the manufacturer’s recommendations. Library quality was assessed on an Agilent Bioanalyzer 2100, using an Agilent High Sensitivity DNA Kit (Agilent technologies). Libraries were pooled in equimolar proportions. 75 bp single reads were generated on an Illumina MiSeq instrument, using a MiSeq Reagent kit V2 (500 cycles) (Illumina), with an expected depth of 217X. An in-lab written MATLAB-based script was used to perform marker frequency analysis. Reads were aligned on the Escherichia coli K12 MG1655 genome using BWA software. Data were normalized by dividing uniquely mapping sequence reads by the total number of reads. Enrichment of uniquely mapping sequence reads in 1 kb non-overlapping windows were calculated and plotted against the chromosomal coordinates.
Cells were grown in M9 minimal medium supplemented with 0.4% glucose to exponential phase (0.2 OD 650 nm) and spread on a 1% (wt/vol) agarose pad for analysis. For snap-shot analyses, cell images were acquired using a DM6000-B (Leica) microscope with MetaMorph software (Version 7.8.8.0, Molecular Devices) and analyzed using ImageJ. Images were taken from 5–10 different fields in each experiment. Two to three independent experiments were carried out to calculate mean and standard deviation for distributions of foci for each strain. For time-lapse analyses, 0.4% glucose agarose pads were used, the slides were incubated at 30°C and images were acquired every 10 minute by an Evolve 512 electron-multiplying charge-coupled device (EMCCD) camera (Roper Scientific) attached to an Axio Observe spinning disk (Zeiss). Image acquisition was done using MetaMorph software (Version 7.8.11.0, Molecular Devices). At each time point, we took a stack of 32 bright-field images covering positions 1.6 μm below and above the focal plane. Image acquisition was performed on five selected different fields corresponding to different cell populations in each experiment. Final images were reconstructed from image stacks using an in-lab written MATLAB-based script. Image analysis was done manually using ImageJ software. For each mutant strain analyzed, two independent time-lapse experiments were realized, each providing five images with 5–10 bacteria per image at the start. The number of divisions that provided two foci-containing cells and the number of first divisions that provided one focus-containing and one focus-free cells were manually counted. Only cells that started with a normal division were taken into account (few cells produced a focus containing-cell and a focus-less cell from the start and were not counted as initial events, as they did not show any normal division preceding the initial event). The percentage of initial events (between parentheses in Table 1) corresponds to the ratio of cell divisions where a focus is lost in one daughter cell for the first time to the total number of cell divisions. For example, in the scheme shown in Fig 2E, we counted 2 initial events (#2 and 7) out of 9 total cell divisions, and 100% heredity.
Cells were grown in either M9 minimal medium supplemented with 0.4% glucose, 5 μM CaCl2 and 1 mM MgSO4 or LB medium supplemented with 0.5% glucose at 37°C as described in Cockram et al, 2015 [51].
|
10.1371/journal.pntd.0002237 | Aedes Mosquito Saliva Modulates Rift Valley Fever Virus Pathogenicity | Rift Valley fever (RVF) is a severe mosquito-borne disease affecting humans and domestic ruminants. Mosquito saliva contains compounds that counteract the hemostatic, inflammatory, and immune responses of the host. Modulation of these defensive responses may facilitate virus infection. Indeed, Aedes mosquito saliva played a crucial role in the vector's capacity to effectively transfer arboviruses such as the Cache Valley and West Nile viruses. The role of mosquito saliva in the transmission of Rift Valley fever virus (RVFV) has not been investigated.
Using a murine model, we explored the potential for mosquitoes to impact the course of RVF disease by determining whether differences in pathogenesis occurred in the presence or absence of mosquito saliva and salivary gland extract.
C57BL/6NRJ male mice were infected with the ZH548 strain of RVFV via intraperitoneal or intradermal route, or via bites from RVFV-exposed mosquitoes. The virus titers in mosquitoes and mouse organs were determined by plaque assays.
After intraperitoneal injection, RVFV infection primarily resulted in liver damage. In contrast, RVFV infection via intradermal injection caused both liver and neurological symptoms and this route best mimicked the natural infection by mosquitoes. Co-injections of RVFV with salivary gland extract or saliva via intradermal route increased the mortality rates of mice, as well as the virus titers measured in several organs and in the blood. Furthermore, the blood cell counts of infected mice were altered compared to those of uninfected mice.
Different routes of infection determine the pattern in which the virus spreads and the organs it targets. Aedes saliva significantly increases the pathogenicity of RVFV.
| Rift Valley fever is an endemic and epidemic zoonosis in Africa and the Arabic Peninsula. In humans, in the most severe cases the viral infection causes fulminant hepatitis associated with haemorrhagic fever, permanent blindness or severe encephalitis. Despite the importance of vector transmission in the spread of arboviruses, few studies on the physiopathology of viral infection have considered the role of the arthropod in the efficiency of viral infection. Moreover, the route of virus inoculation and the presence of the vector's saliva can potentially affect virus pathogenicity. Our results show that saliva from Aedes mosquitoes increases Rift Valley fever pathogenicity. Importantly, our study also revealed that RVFV transmitted via mosquito bites spread differently than virus inoculated by other routes. These observations may have interesting repercussions given the role mosquitoes were shown to play in the transmission of RVFV in humans during the last outbreak of the disease in Saudi Arabia. Identification of salivary proteins able to increase RVFV virulence may pave the way to new approaches to prevent or cure the disease.
| Rift Valley fever virus (RVFV) is a zoonotic mosquito-borne virus which causes epizootics and associated human epidemics throughout Africa [1], [2]. First identified in Kenya in 1931 [3], RVFV is now considered an endemic zoonotic agent in sub-Saharan Africa causing explosive outbreaks in animals and humans. It has been observed in Egypt, Mauritania, and the Arabic Peninsula [4], [5], [6]. The manifestation of severe RVF disease in humans is variable. Humans may develop a wide range of clinical signs including hepatitis, retinitis, and delayed-onset encephalitis and, in the most severe cases, haemorrhagic disease. The overall case fatality ratio is estimated to be between 0.5% and 2% [7], [8], [9]. In Yemen and Saudi Arabia, a RVFV outbreak resulted in approximately 2,000 human infections and 250 deaths (CDC 2000). A study of the RVFV epidemic in Saudi Arabia reported a high incidence of neurological manifestations (17.1%) in infected individuals [7]. Mosquito bites were reported to play an important role in the transmission of the disease during this outbreak.
RVFV belongs to the genus Phlebovirus in the family Bunyaviridae. Its tripartite negative-strand RNA genome is composed of a large segment (L) that encodes the L protein, which is the viral RNA-dependent RNA polymerase; a medium segment (M) that encodes a single open reading frame (ORF) generating the NSm, G1 (Gc) and G2 (Gn) proteins and a small segment (S) that encodes the nucleocapsid protein (N) and a nonstructural protein (NSs) using an ambisense strategy [10]. NSs was shown to suppress interferon induction (Billecocq et al., 2004).
RVFV can be transmitted to vertebrates by several species of mosquitoes such as Aedes spp. and Culex spp. Human infections typically occur through bites from infected mosquitoes, through percutaneous/aerosol exposure during the slaughter of infected animals, or via contact with aborted fetal materials. Transmission efficiency depends on the ability of the virus to cross the various barriers in the vector [11]. Therefore, after a mosquito takes a blood meal from an infected individual, the ingested virus passes into the midgut of the mosquito where it replicates before infecting different organs in the mosquito. At the end of the extrinsic incubation period in the vector, salivary glands are infected and the virus is transmitted by saliva during a blood meal. The reproductive system of the mosquito is also infected and transovarial transmission is important for long term maintenance of the virus [12]. Worldwide RVFV is considered as a potential biological weapon. Both modified live attenuated virus and inactivated virus vaccines have been developed for veterinary use, but there are currently no commercially available vaccines for humans.
During a blood meal, insects are subject to defensive responses from the vertebrate, including hemostasis and the immune response. In this context, the saliva injected by the mosquito plays multiple roles. Indeed, saliva proteins have angiogenic, anti-hemostatic, anti-inflammatory and immunomodulatory properties [13]. The various properties of the saliva proteins towards the host immune response affect the pathogen transmission. In some cases, co-injection of virus and saliva potentiates viral infection of the vertebrate [14], [15], [16], [17]. In other cases, pre-exposure to saliva generates enhances mortality from subsequent viral infection via mosquito bite [18]. A longer viremia was observed in deer and chipmunks infected by mosquito bite containing La Crosse virus, another member of the Bunyaviridae family, compared to syringe injection [19]. Potentiation of infection by mosquito saliva was also demonstrated for Cache Valley virus, an orthobunyavirus that also belongs to the Bunyaviridae family [14]. These observations raise the question of whether RVFV infection is also potentiated by mosquito saliva. Since RVFV is also transmitted by blood and aerosols, the context for its transmission differs from those of other viruses studied previously. In this project, our objective was to evaluate the role of Aedes mosquito saliva in the natural transmission of RVFV. For this purpose, we make use of an animal model that allowed us to study the pathogenesis of RVFV infection. We evaluated two different routes of infection: the intraperitoneal route, which has been utilized in most previous studies of RVFV pathogenesis, and the intradermal route, which mimics the mosquito bite. We also used non-infected and RVFV-infected mosquitoes to evaluate the role of saliva in the progression of the disease. Importantly, we found that Aedes saliva potentiated RVFV infection, once again highlighting its role in arbovirus transmission.
All studies on animals followed the guidelines on the ethical use of animals from the European Communities Council Directive of November 24, 1986 (86/609/EEC). All animal experiments were approved and conducted in accordance with the Institut Pasteur Biosafety Committee. Animals were housed in the Institut Pasteur animal facilities accredited by the French Ministry of Agriculture to perform experiments on live mice, in appliance of the French and European regulations on care and protection of the Laboratory Animals (accreditation number B 75 15-01 and B 75 15-07). The study protocols were approved by the Comité d'Ethique pour l'Expérimentation Animale (CEEA) - Ile de France - Paris - Comité 1.
The ZH548 strain was isolated from a human infection during the 1977 outbreak in Egypt [18]. The case was anonymous and an informed consent was not required at that time. This strain was part of a collection used by the NRC of arboviruses (B. Le Guenno and H. Zeller). This collection was transmitted to us and we possess an AFSSA authorization of detention, transfer and manipulation (since 2001) as a “select agent”.
We used the DBA-1 and C57BL/6-NRJ mice for infections (Janvier, France). Vero E6 cells were grown in DMEM supplemented with 10% fetal bovine serum (FBS), 10 µg/ml of penicillin and 10 U/ml of streptomycin. C6/36 cells were grown at 28°C in plastic cell culture flasks in Leibovitz medium 15 supplemented with 10% FBS, penicillin (50 units/m1), and streptomycin (50 mg/ml).
Stocks of the virulent Egyptian ZH548 RVFV strain were produced under biosafety level 3 (BSL3) conditions. In all experiments, the ZH548 strain was obtained from a cell culture of C6/36 cells. It was produced under BSL3 conditions.
Dehydrated eggs of Aedes aegypti (strain PAEA) and Ae.vexans vexans were placed in water to hatch. Adult mosquitoes were reared in a room held at 25±1°C and 80% relative humidity, and having a light/dark ratio of 12 h/12 h. The larvae were fed on brewer's yeast tablets and adults were fed on sugar water (10%).
Rabbit blood was collected in heparinized tubes (0.02%). Red blood cells were separated from plasma by centrifugation, washed 3 times in 1X PBS, and were resuspended in the same buffer. Five-day old female mosquitoes were placed in boxes sealed with veils and were fed on 37°C thermostated glass feeders covered with chicken skin and filled with a mixture containing 2 mL of red cells, 1 mL of virus solution (108 plaque forming unit (pfu)/mL) and 30 µL of ATP (5.10−3 M).
Mosquito females were blood-fed five days after hatching. Three weeks later (corresponding to the extrinsic incubation period of RVFV in Ae. aegypti and Ae. vexans mosquitoes), 100 salivary glands (SG) were dissected and placed in 100 µL 1X PBS. The inocula used in our experiments were equivalent to a pair of SG (or two salivary glands extracts [SGE]). SG-containing tubes were stored at −80°C. SGEs were prepared by sonicating the SGs (five times at 4 min each with a pulse ratio of 2 sec on/2 sec off) and centrifuging the crude extract at 13,000 rpm for 15 min at 4°C. The supernatant was transferred to clean tubes and stored at −80°C. The protein concentration was determined by spectrophotometry at 280 nm (Nanodrop).
Fifteen days after their blood-meal, RVFV-exposed mosquitoes were anesthetized at 4°C, legs and wings were sectioned and bodies were placed on a double-sided tape fixed on a glass slide. The proboscis was inserted manually into a 10 µL-cone filled with 5 µL of filtered 1X PBS or DMEM+Glutamax containing 2% FBS. The cone content was collected 45 min later and the virus titer in the solution was determined by plaque assay.
Mice were anesthetized intraperitoneally with a mixture ketamine/xylazine consisting of 2 mL of 2% Rompun (Bayer), 4 ml of Imalgene 1000 (Merial), 4 ml of sterile water (Gibco) and 2 mL of 1X PBS (Gibco). “Pathogen-free” male mice C57BL/6NRj (Janvier) aged four weeks and weighing 15–20 g each, were infected in a BSL3 animal facility by intraperitoneal or intradermal route in the absence or presence of either mosquito SGE (one SG pair per inoculums = SGP: 2 µl in 20 µl) or non-infected mosquito bites, or they were infected directly by bites from infected mosquitoes. Selected mice were euthanized five days after infection and the following organs were harvested without any perfusion: brain, liver, spleen, stomach, small and large intestine, pancreas, bladder, heart, lungs, thymus, lymph nodes and salivary glands. Brains were divided into two parts: cerebellum and brain hemispheres (including olfactory bulbs). For virus titration, large organs were cut into pieces of ∼30 mg, whereas small organs like lymph nodes, salivary glands and thymus were kept whole and frozen at −80°C. Samples were then homogenized either in Trizol or in DMEM. Supernatants were collected after centrifugation.
For these samples, mice were sacrificed 5 days after infection and perfused with 4% formalin. The organs removed were kept in a freshly prepared solution of formalin. The fixed tissues were embedded in paraffin, cut into 3-µm sections thick, and stained with hematoxylin and eosin (H & E).
RVFV-containing samples were titrated on E6 cells by the plaque assay method. Cell counts were performed on KOVA slides. E6 cells were grown in DMEM+Glutamax (Dulbecco) containing 10% decomplemented FBS, 10 U/mL penicillin and 10 µg/mL streptomycin in 6-well plates containing 106 cells per mL for plaque assays. Tenfold serial dilutions of each sample to be titrated were prepared in DMEM medium containing 2% FBS, 10 U/mL penicillin and 10 µg/mL streptomycin. 300 µL of inoculum dilution was deposited in each well of a 6-well plate and incubated with for 1 hr at 37°C in a CO2 incubator. Then, 4 mL of agar (culture medium containing 2% FBS and 2% agarose) were deposited in each well and incubated for three days. The plaques were then revealed with a 0.2% solution of crystal violet containing 3.7% formaldehyde and 20% ethanol.
For the detection of anti-RVF antibodies in mouse sera, we used a microsphere immunoassay in which a purified recombinant RVF N antigen was covalently associated to color-coded microbeads (unpublished data). Captured anti-RVF antibodies on coupled microspheres were detected using biotinylated anti-mouse IgG and phycoerythrin-conjugated streptavidin by FACS analysis.
We used the Power SYBR Green RNA-to-Ct One-Step Kit (Applied Biosystems, Carlsbad, California) according to the manufacturer's protocol. It allowed amplifying a 108 bp sequence located between nucleotide 1485 and nucleotide 1593 of the M segment of RVFV. The primers selected were as follows: upper 5′-CATGGATTGGTTGTCCGATCA-3′ and lower 5′-TGAGTGTAATCTCGGTGGAAGGA-3′. Each sample was analyzed in duplicate against a standard curve produced from a specific concentration range of synthetic RNA. We amplified the samples on an Applied Biosystems 7500 instrument using the following PCR program: a reverse transcriptase (RT) step for 30 min at 50°C; inactivation of the RT enzyme and activation of DNA polymerase for 10 min at 95°C; 40 PCR cycles of 15 sec at 95°C and 1 min at 60°C (annealing temperature of primers), during which fluorescence data is collected; and finally, 20 sec at 95°C with ramping 19 min 59 sec for melting curves.
Results were compared using two nonparametric statistical tests: Kruskal-Wallis and Mann-Whitney. The median day of death was calculated for each condition and results were compared using Kruskal-Wallis and Mann-Whitney statistical tests.
We first selected an optimal mouse strain for our experimental infection model. For this purpose, we infected six C57BL/6 and six DBA-1 male mice by the intradermal route with RVFV and found that the survival curves for these two strains differed significantly. Whereas DBA-1 mice started to die at four days after infection (D4), C57BL/6 mice started to die at seven days after infection (D7). Moreover, whereas neurological symptoms (such as hind limb paralysis) occurred in C57BL/6 mice, no such problems were observed in DBA-1 mice (data not shown). Therefore, we chose the C57BL/6 genetic background for our RVFV infection model.
We next compared the mortality rates and RVFV tissue distributions in mice infected by two different routes of injection: the intraperitoneal (IP) and intradermal (ID) routes. The kinetics of infection was slower with the ID route, and a delayed mortality of two days was observed between the two routes of injection (Figure S1). At D3, no significant differences in viremia were found between the two routes of injection. However, at D6, viremia remained at a plateau level of 104 pfu/mL in animals inoculated via IP injection whereas virus titers significantly decreased between D3 and D6 in ID injected mice (Figure 1). Moreover, high virus titers were found in the brain of mice infected by ID injection but not in the liver whereas high titers were found in the liver of mice infected by IP at D6 but not in their brain (Figure 1). In agreement with these findings, ID-infected mice presented neurological symptoms. Since ID infection more closely mimics natural infection by the vector, all subsequent infections were performed by this route.
To determine whether Ae. aegypti mosquito saliva has a role in potentiating RVFV infection, we infected mice ID with doses of virus between 10 and 104 pfu/mouse, with or without SGE from uninfected mosquitoes and calculated the median day of death of the animals for each condition. At the lower dose, not all mice died (Figure 2) and 66% of the mice surviving did not present any anti-RVFV antibodies (data not shown). However, in presence of saliva, all mice but one died (Figure 2) and the surviving mouse presented anti-RVFV antibodies. The effects of SGE on mortality of infected mice were identified at the lower virus doses of 10 to 103 pfu/mouse (Figure 2). Median day of death calculation indicated a significant difference between virus and virus+SGE for injection of 102 and 103 pfu/ml (p = 0.01 and p = 0.002 respectively) (Figure S2). At higher RVFV doses, the effect of SGE on mortality rate was not significant (p>0.05). The weight of the infected mice also decreased as the infections proceeded (data not shown). From these results, we selected 103 pfu/mouse as the reference dose for studying RVFV distribution in mice in the presence and absence of SGE.
In addition, we found that the effect of the SGE was not restricted to Ae. aegypti mosquitoes as Ae. vexans SGE also increased RVFV virulence (Figure S3). Median day of death calculation indicated a significant difference between virus and virus+SGE (p = 0.006) and virus+saliva (p = 0.01). However, we did not observe any difference between virus+SGE and virus+saliva (p = 0.42). Interestingly, we did not observe any effect on mice survival when we injected Culex pipiens pipiens SGE (data not shown).
We infected C57BL/6 mice with ID injections of virus (103 pfu/mouse) in the presence or absence of SGE and followed the distribution of the virus in the blood and in various organs. The organs were not perfused prior collection. We sacrificed the mice at D5 because in Figure 2 infected mice died during the night between D5 and D6. Viremia levels were very high in the infected mice, and high virus titers were also found in the liver, brain and cerebellum (Figure 3), lymphoid organs (spleen, thymus and lymph nodes) (Figure S4), as well as in heart, kidneys, and lungs (Figure S5). Low virus titers were found in the eyes, jejunum and ileum (less than 103 pfu/mL), whereas the intestine, stomach, ceacum, colon and gallbladder contained no measurable titers.
In the presence of SGE, virus titers were significantly increased (almost 104 fold) than those produced in the absence of SGE, in the brain cortex (p = 0.024), liver (p = 0.004) and blood (p = 0.004) (Figure 3). The virus titers in the cerebellum exhibited an opposite trend compared to the titers in the brain cortex. Indeed, in this organ, virus titers of animals infected in the presence of SGE were lower (median value of 104 fold) than those of animals infected without SGE (p = 0.004).
We then analyzed the virus titers in the lymphoid organs of mice infected in the presence and in the absence of SGE. The virus titers in the inguinal lymph nodes (p = 0.03), spleen (p = 0.004) and thymus (p = 0.007) of mice infected in the presence of SGE were significantly higher than those of mice infected in the absence of SGE (Figure S4). Virus titers in the lungs (p = 0.004), kidneys (p = 0.005), bladder (p = 0.004) and heart (p = 0.01) of mice infected in the presence of SGE were significantly higher (102 fold increase) compared to the titers found in these tissues of mice infected without SGE (Figure S5). Virus titers in the pancreas exhibited a pattern similar to that observed in the cerebellum, as these titers were significantly lower in the presence of SGE (p = 0.016) than in absence of SGE. In contrast, the addition of SGE to the viral inoculum did not lead to any significant differences in the virus titers in the mesenteric lymph nodes (ML), aortic lymph nodes (AL), popliteal lymph nodes (PL) or salivary glands (data not shown). These results correlated well with the RNA quantification results we obtained from RT-qPCR analysis of each organ (data not shown).
Following RVFV infection of mice with or without SGE, we found that several blood parameters were altered in infected mice compared to uninfected ones. These changes included significantly lower numbers of white blood cells and platelets (Table 1).
The significant leukopenia observed in infected mice was associated with changes in the white blood cell count, with proportionally higher numbers of granulocytes and monocytes and lower numbers of lymphocytes compared to those in uninfected mice (Table 1). A 50% decrease of platelets and white blood cell counts was observed in presence of SGE in the inoculation (Table 1).
To better understand the physiology of RVFV infection, we conducted histological analysis of the liver of infected mice and found significant differences between the livers of mice infected in the presence or absence of SGE (Figure 4). Indeed, mice infected in the presence of SGE exhibited signs of multifocal hepatitis (Panels A and C). Inflammatory foci were randomly distributed in the liver parenchyma (arrowheads in Panels A and C). These foci were characterized by prominent neutrophil infiltrations (asterisk in Panel C insert) that were associated with fewer numbers of lymphocytes. Necrotic hepatocytes, with acidophilic cytoplasm and a highly condensed basophilic nucleus (pyknosis) or a fragmented nucleus (karyorrhexis) were identified within the inflammatory foci (arrow in Panel C insert). The profile of the liver lesions in mice infected in the absence of SGE (Figure 4; Panels B and D) was very different from that of mice infected in the presence of SGE. Three out of four mice exhibited hepatic necroses with very few inflammatory foci (Panel B). These necrosis foci were randomly located within the parenchyma (arrows in Panel D insert) and were associated with few inflammatory cells (Panel D insert).
We collected saliva from RVFV-exposed mosquitoes to estimate the concentration of virus injected during a bite. These results showed mosquitoes may inject approximately 50±20 pfu in each bite, and that more than 50% of the mosquitoes had been infected after an artificial blood-meal.
To this point, our experiments were performed with SGE, which contains a mixture of salivary and housekeeping proteins. To determine whether the unique components of saliva triggered the potentiating effect on RVFV virulence, we allowed ID-infected mice to be bitten by non-infected mosquitoes. We inoculated C57BL/6 male mice with RVFV (50 pfu/mouse) by ID injection and exposed the mice to non-infected mosquito bites in the area of the ID infection. The number of blood-fed mosquitoes was determined. The weight changes of the mice were followed for 14 days thereafter. The weight curves of the infected mice corroborated with our previous results (Figure 5). In the absence of mosquito bites, mice survived for at least 11 days and died between 13 and 14 days post infection. If infection was accompanied by non-infected bites, time to death was shortened. We however did not find any clear correlation between the number of bites and the time to death.
In a second series of experiments, mice were bitten by RVFV-exposed mosquitoes collected on D16 or on D19 after infected blood meal. Blood-fed mosquitoes were collected and their viral loads were determined. The weight curves for the mice bitten by mosquitoes at D16 after virus exposure were followed. Three out of five mice received 1 or 2 bites from RVFV-exposed mosquitoes. Only one mouse died 13 days after receiving four mosquito bites (for which two out of four mosquitoes were infected), while the other mice survived until day 14 (data not shown). The experiment was repeated with mice bitten by D19 RVFV-exposed mosquitoes (Figure 6). As before, the numbers of blood-fed mosquitoes were counted and their viral loads were determined. Mice received up to 9 bites and 3 to 6 of these bites were from infected mosquitoes. Two mice having received bites from infected mosquitoes did not die during the time of experiment (11 days). Their weight did not decrease. Mosquitoes did not feed on two mice. For the other 6 mice, death was observed from day 5 to day 10 post-feeding. The time to death did not depend on the number of blood-fed mosquitoes collected on each mouse and is more probably related to the amount of virus injected during the probing phase and to the number of uninfected bites. This amount seems highly variable in our experiment. Indeed, we were not able to identify mosquitoes that could have probed, and thereby injected virus, without taking any blood meal.
RVFV is primarily transmitted by mosquito bites and, to a lesser extent, by direct contact with infected animals, mainly sheep and goats, as reported during an RVF epidemic in southwestern Saudi Arabia [7]. However, many studies describing the pathogenesis of this virus have been conducted without considering this natural way of transmission. Indeed, the route of virus inoculation and the presence of components from the vector saliva are likely to have consequences on the immune response that is eventually developed by the host in response to the pathogen. In fact, several studies have shown that the saliva of arthropod vectors transmitting infectious diseases can play a crucial role in the ability of the vector to transmit the pathogen [15].
The mouse strain used in any model of RVFV infection is an important factor that should be carefully considered. Several different genetic strains of mice have been used previously: BALB/cByJ, C57BL/6, 129/Sv/Pas, and MBT/Pas [20]. The BALB/cByJ, and C57BL/6 strains were found to be the least susceptible to RVFV infection. In our study, DBA-1 mice were more sensitive to the virus than C57BL/6 even though they did not exhibit any neurological symptoms. The C57BL/6 strain experienced hepatic infection as well as neurologic symptoms. These mice are therefore good models to study the most severe forms of RVF in humans, and allow the study of neuropathogenesis and progression of the virus from the periphery to the central nervous system following intradermic inoculation.
A number of different studies aimed at defining the pathogenesis of RVFV in animals have employed IP, intranasal and subcutaneous inoculations [4], [21]. Indeed, exposing mice to aerosols containing RVFV can cause infection [22] whereas other routes of exposure induce delayed death [21]. In our study, the IP and ID routes of injection led to different patterns of virus dissemination. The brain and liver were the main targets of the virus after ID and IP infection, respectively. Viremia was maintained longer after IP infection whereas survival was shorter compared to ID infection. This result showed that the route of infection is a key determinant for infection.
First, after ID injection, we found the virus in many organs. High virus titers were found in the liver and in the blood early after infection at D3, whereas at D6, high virus titers were found in the brain, while the viremia has decreased. In agreement, mice presented neurological symptoms at D6, characterized by compulsive or uncoordinated movements, and/or paralysis, and they also had discolored livers presenting hemorrhagic lesions. Our observations correlated well with other studies that showed that this virus causes fulminant hepatitis [7] or meningoencephalitis [23] in humans. We found other organs to be less infected, including the heart, lungs, pancreas and kidneys, which was reported previously [24]. Mice salivary glands were found to be significantly infected, raising the question of the infectivity of saliva. Viral antigens have also been detected in odontogenic and gingival epithelia [24]. In addition, we detected virus in the primary and secondary lymphoid organs and the lymphocyte numbers were lower in infected animals compared to controls. These changes could be explained by lymphocyte apoptosis in lymphoid organs (thymus, spleen and lymph nodes), which was also demonstrated in BALBc mice subcutaneously infected with another RVFV strain (ZH501) [24].
We also observed changes in other blood count parameters like the number of thrombocytes (platelets), granulocytes and monocytes. In general, a decrease in circulating platelet number may be caused by decreased or ineffective bone marrow production, increased intramedullary destruction (hemophagocytic syndrome), increased peripheral destruction (immune-mediated or non–immune-mediated mechanisms), altered distribution of circulating cells (splenic consumption or endothelial sequestration), or decreased cellular life span. Bone marrow was found to be infected in our study (data not shown), and lower numbers of myeloid cells in the spleen and bone marrow in RVFV infection have been reported previously [24]. On the other hand, in patients with dengue hemorrhagic fever, although dengue virus-induced bone marrow suppression was shown to decrease platelet synthesis, an immune mechanism of thrombocytopenia caused by increased platelet destruction appeared to be also active [25], [26], [27], [28]. Granulocytes and monocytes numbers were higher in the blood of infected animal (Table 1). Three types of granulocytes are present in peripheral blood: neutrophils, eosinophils and basophils. The count of eosinophils was found to vary as function of Rift Valley fever disease progression in mice: it first decreased at the beginning and then increased before death [29], which could explain our findings. Granulocytes were also found to be important target cells for RVFV infection [21] and thereby represent a site of viral replication to infect other cells or organs. Monocyte numbers also increase in many other vectorial infectious diseases such as West Nile, dengue and malaria [24], [25], [26]. Similar changes in blood cell counts including lymphopenia and thrombopenia were reported for the Saudi Arabian epidemics of 2000 [7].
We investigated the role of vector salivary components in RVFV infection. Potentiation of virus transmission and/or pathogenicity in the presence of vector saliva has previously been described in vector/pathogen/host interactions [14], [15], [30]. Some of the many salivary proteins co-injected during a vector bite cause immunomodulatory effects on the host. These may include the induction of a Th2 response and the inhibition of Th1 pro-inflammatory cytokines [15], [31]. In addition, it has been shown in vivo that Aedes mosquito bites are likely to significantly reduce T cell recruitment [16]. We tested the saliva of two Aedes species: Ae. vexans, which is an important RVFV vector in Africa and in the Arabic peninsula [32], [33], [34], [35]; and Ae. aegypti, which exhibits good vector competence for the virus as shown in our study as well as in others [11], [36] and whose genome has been sequenced. Early death was observed in the groups of mice co-injected with both Aedes SGE. In addition, the survival curves obtained for RVFV-infected mice exposed to the bites of mosquitoes corroborated those obtained with co-injected SGE and confirmed that both Ae. aegypti and vexans saliva potentiates RVFV pathogenicity. These results are comparable to those reported for mice with West Nile virus mixed with mosquito saliva [31]. Interestingly, although Culex pipiens was found to be competent to transmit RVFV [37], we did not observe any increase of RVFV pathogenenicity in presence of salivary gland extracts from this species.
We determined the effects of SGE and saliva on RVFV virulence and distribution for several organs and included histological analyses of the liver. For most organs, including liver, the brain cortex, kidneys, lungs, heart, bladder, spleen, thymus and lymph nodes, virus titers were significantly higher if SGE was included in the inoculum, in agreement, with previous studies where saliva was shown to increase the invasion of neural tissues by West Nile virus and produced higher virus titers in the brain [31]. An SGE-mediated decrease in antiviral activity at the site of inoculation might promote viral replication and infection of different cell types [20], [21], [38], thereby increasing virus production in several organs and causing specific histological lesions, as observed in the liver. This is consistent with what Schneider and Higgs observed in mice infected with West Nile virus in presence of mosquito bites [31]. Early after virus inoculation, they did not observe any difference in the viral titers measured in various organs in presence or absence of saliva whereas after 7 days of infection, higher titers were observed after mosquito bites. We cannot exclude however a modification of the kinetics of virus replication and dissemination in the various tissues in the presence of saliva. Interestingly, and while the viremia is significantly increased, lower virus titers were found in the pancreas and cerebellum in presence of SGE, showing that saliva may also affect virus dissemination. With respect to the brain, our results suggest that saliva might modify the kinetics and/or the extent of invasion of specific regions. The modalities of infection of the central nervous system by RVFV are still poorly understood. Neurons and glial cells were found positive for RVFV throughout the central nervous system of infected calves [39]. Gray et al. [29] showed that the brains of RVFV ZH501 infected mice were essentially normal throughout the course of the study despite evidence of a high viral titer and significantly increased inflammatory cytokine concentrations in the brain tissue of some studied animals. The outcome of our study may suggest either that the presence of saliva at the site of inoculation may favour different ways of brain invasion or that the kinetics of infection is increased and that the cerebellum was first invaded and already partly cured at the moment of sample harvesting while the virus was spreading towards the brain cortex.
Since a direct effect of saliva on the brain is unlikely, we propose that modulation of the early immune and inflammatory responses at the site of virus injection may, in turn, modulate the permeability of the blood-brain barrier, allowing virus titers in the brain to be significantly higher. Further studies on this matter are currently underway and preliminary experiments are in favor of an increase of the vascular permeability of the blood brain barrier in presence of saliva. We suggest that intermediate elements like TLR3 and IL6 might be involved in this effect. Actually, West Nile virus, by activating TLR3 (toll-like receptor 3) [40], and allowing TNFα secretion, was proposed to increase blood-brain barrier permeability. Moreover, it was shown that IL-6 played an important role in increasing brain permeability in a model of bacterial meningitis [41], [42]. Similar mechanisms might occur in RVFV infections in the presence of saliva.
Our histological analysis of infected liver showed that mice infected in the presence of SGE developed multifocal hepatitis with inflammatory foci that were randomly distributed in the hepatic parenchyma. This was also accompanied by a massive recruitment of neutrophils and lymphocytes in the liver parenchyma. CD4+ and CD8+ lymphocytes and cytokines, including TGF-β, TNF-α and IFN-γ were shown to be involved in the hepatic pathogenesis of yellow fever virus infection in combination with a direct cytopathic effect of the virus [43]. The early modulation of the innate response in the dermis caused by mosquito bites probably induces a dysregulation of the immune system and triggers the different pathologic effects observed in absence and presence of the mosquito saliva.
Exposure of inoculated mice to mosquito bites confirmed that saliva components have a potentiating effect on RVFV infection. Indeed, we observed early death in mice infected by ID and bitten by uninfected mosquitoes although a clear correlation between the number of engorged mosquitoes and the time of death could not be established. This is probably explained by the time of probing that differs between mosquitoes and the length of the probing time conditioned the amount of saliva injected in the dermis.
The next step was to compare infection by an infected mosquito to infection by ID. Death was observed as early as day 5 post-infection, a delay which is comparable to that of mice infected ID with 103 pfu in the presence of SGE. This observation shows that mosquitoes may inject more than 50 pfu in agreement with the detection of a discrepancy between the titers obtained by salivation and those determined in vivo [44]. Our results also showed that the bites of non-infected mosquitoes may potentiate infection caused by the bites of infected mosquitoes. It is important to note that although the number of infected mosquitoes in nature is relatively low, the number of uninfected bites is much higher. Thus, constant local stimulation with saliva may have the potential to modulate the impact of RVFV infection [45].
In conclusion, we have clearly demonstrated an overall potentiating effect of mosquito saliva on RVFV infection. Both Aedes aegypti and Aedes vexans saliva are able to decrease the survival of RVFV-infected mice. The impact of saliva components on the innate immune response at the site of bite certainly explains the facilitation observed, either by increasing the kinetics of distribution of the virus or by altering this distribution through differential targeted organs. The identification of salivary proteins involved in the facilitation of infection and determination of their mode of action could help develop new approaches for preventive or therapeutic purposes in humans.
|
10.1371/journal.pcbi.1005136 | Dynamic Nucleosome Movement Provides Structural Information of Topological Chromatin Domains in Living Human Cells | The mammalian genome is organized into submegabase-sized chromatin domains (CDs) including topologically associating domains, which have been identified using chromosome conformation capture-based methods. Single-nucleosome imaging in living mammalian cells has revealed subdiffusively dynamic nucleosome movement. It is unclear how single nucleosomes within CDs fluctuate and how the CD structure reflects the nucleosome movement. Here, we present a polymer model wherein CDs are characterized by fractal dimensions and the nucleosome fibers fluctuate in a viscoelastic medium with memory. We analytically show that the mean-squared displacement (MSD) of nucleosome fluctuations within CDs is subdiffusive. The diffusion coefficient and the subdiffusive exponent depend on the structural information of CDs. This analytical result enabled us to extract information from the single-nucleosome imaging data for HeLa cells. Our observation that the MSD is lower at the nuclear periphery region than the interior region indicates that CDs in the heterochromatin-rich nuclear periphery region are more compact than those in the euchromatin-rich interior region with respect to the fractal dimensions as well as the size. Finally, we evaluated that the average size of CDs is in the range of 100–500 nm and that the relaxation time of nucleosome movement within CDs is a few seconds. Our results provide physical and dynamic insights into the genome architecture in living cells.
| The mammalian genome is partitioned into topological chromatin domains (CDs) in the living cell nuclei. Gene expression is highly regulated within CDs according to their structure, whereas chromatin itself is highly dynamic. This raises the following question: how is the CD structure in such dynamic chromatin? We developed a conceptual framework that unifies chromatin dynamics and structure. Using a polymer model with a fractal domain structure in a viscoelastic medium, we analytically show that nucleosome movement is subdiffusive and depends on CD structure. Hence, structural information can be extracted based on nucleosome movement in living cells with single-particle tracking experiments. This framework provides physical insights into the relationship between dynamic genome organization and gene expression.
| Genomic DNA is packed and folded three-dimensionally in the cell nuclei. In the nuclei of eukaryotic cells, the nucleosome is a basic unit consisting of an approximately 147-bp DNA wrapped around core histones [1]. Recent experimental evidences suggest that the nucleosome is irregularly folded without the 30-nm chromatin fiber [2–7]. On the other hand, at the scale of the whole nucleus, interphase chromosomes occupy distinct chromosome territories [8]. This highly organized chromosome structure allows for effective regulation of various genome functions.
By virtue of recent developments of chromosome conformation capture (3C) techniques, the genome-wide chromosome organization has been revealed by detecting the physical contact frequencies between pairs of genomic loci [9]. More recently, 3C derivatives, Hi-C and 5C profiles demonstrated that metazoan genomes are partitioned into submegabase-sized chromatin domains (CDs) including topologically associating domains (TADs) [10–12]. TADs are considered to be a regulatory and structural unit of the genome [13]; genome loci located in the same TAD are associated with each other, whereas genomic interactions are sharply depleted between adjacent domains. For even single-cell Hi-C, individual chromosomes maintain domain organization [14]. Furthermore, kilobase-resolution in situ Hi-C maps identified not only small contact domains but also CTCF-mediated loop domains [15, 16].
In contrast, dynamic aspects of chromatin have been shown by live-cell imaging experiments [17–24]. In particular, single-nucleosome imaging in living mammalian cells has revealed local nucleosome fluctuations caused by the thermal random force [25–27]. The mean-squared displacement (MSD) of dynamic nucleosome movement clearly shows subdiffusive motion,
MSD ( t ) = D app · t β ( 0 < β < 1 ) , (1)
where Dapp is the apparent diffusion coefficient with dimension m2/sβ. This means that nucleosome movement must be affected by restrictions from some factors but thermal noise. Therefore, there must be a way that the dynamic aspect is consistent with aspects of the genome organization. A theory is required to relate the dynamic aspects described by Dapp and β to the structural features of CDs. To date, the subdiffusive exponent β has been considered to depend on the folding structure of nucleosome fibers [28] and the viscoelasticity of the thermal environment [29, 30].
The fractal nature of chromatin architecture as well as nucleus environment has been revealed recently [9, 31, 32]. The topological structure of CDs can be described by use of the fractal manner. Here, we propose a polymer model for a CD, whose conformational state is assumed to be expressed by the fractal dimension df in a viscoelastic medium with the exponent 0 < α < 1. Although not only the strings and binders switch model [33] but also the block copolymer model [34] can explain aspects of chromatin folding and chromosome architecture in Hi-C experiment datasets, in our model we abstract information on the conformational states of CDs and interpret their dynamic features by using size scaling according to the fractal dimensions. Accordingly, the analytical form of the MSD of nucleosomes in CDs can be derived in terms of polymer physics. As a result, the structural information of CDs, such as the size and conformational state expressed by the fractal dimension, can be derived from the MSD data of dynamic nucleosomes.
A standard approach for treating Eq 3 is to use the normal coordinates X p ( t ) ≡ 1 N ∫ 0 N cos ( p π n N ) R ( n , t ) d n for p = 0, 1, 2, ⋯; however, the nonlinearity of the long-range interaction makes it difficult to deal with the equation in this manner. Therefore, to simplify the analysis, firstly, we assume that nucleosome fluctuations within the CD reach thermal equilibrium after the relaxation time τdf,α, which is explicitly described below (Eqs 11 and 12). Second, we use an approximation to transform the nonlinear Langevin equation (Eq 3) into a linear equation by averaging under thermal equilibrium with respect to the normal coordinates
∫ 0 t γ ( t - t ′ ) d X p ( t ′ ) d t ′ d t ′ = - k p X p ( t ) + g p ( t ) . (5)
The term in the left hand side and the second term in the right hand side (RHS) are straightforwardly derived according to the normal coordinates, in which g p ( t ) ≡ 1 N ∫ 0 N cos ( p π n N ) g ( n , t ) d n satisfies 〈gp(t)〉 = 0 and the FDR 〈 g p κ ( t ) g q λ ( t ′ ) 〉 = k B T N γ ( t - t ′ ) δ κ λ δ p q ( 1 + δ p 0 ) / 2 (see S1 Text, Section IA). Instead of the linearity of Eq 5, the parameter kp implicitly includes the nonlinear effect such as the long-range interactions, and is determined by the variance of Xp over the thermal relaxation time [30] (see S1 Text, Section IB):
k p = 3 k B T 2 N X p 2 CD for p ≥ 1 and k 0 = 0 . (6)
Finally, to calculate the thermal average 〈 X p 2 〉 CD, the effective size scaling (Eq 4) generated by the long-range interactions is used. The asymptotic form for large p is calculated as follows (see S1 Text, Section IC):
X p 2 CD ≃ 〈 R 2 〉 CD 2 A d f p - 1 - 2 / d f . (7) Adf is a dimensionless constant depending on the fractal dimension: A d f = π 1 + 2 / d f Γ ( 1 + 2 / d f ) sin ( π / d f ). We shall refer to the above approximation as the linearization approximation, which is on the same level of the approximation as the preaveraging approximation in terms of polymer physics [35, 50]. From this point forward, to avoid complicated expressions caused by this asymptotic form, we regard the asymptotic sign ‘≃’ as equality.
Next, let us consider the MSD of nucleosomes in CDs. Since the inverse transform of normal coordinates is R ( n , t ) = X 0 ( t ) + 2 ∑ p = 1 ∞ cos ( p π n N ) X p ( t ) and the correlation between different modes vanishes, the MSD of the n-th nucleosome, ϕ(n, t)≡〈[R(n, t) − R(n, 0)]2〉, is expressed as
ϕ ( n , t ) = X 0 ( t ) - X 0 ( 0 ) 2 + 8 ∑ p = 1 ∞ cos 2 p π n N X p 2 CD - C p ( t ) , (8)
where the correlation function is defined as Cp(t)≡〈Xp(t) ⋅ Xp(0)〉. Multiplying Eq 5 by Xp(0) and averaging with 〈gp(t) ⋅ Xp(0)〉 = 〈gp(t)〉⋅〈Xp(0)〉 = 0, we can derive that the correlation function for p ≥ 1 satisfies
∫ 0 t γ ( t - t ′ ) d C p ( t ′ ) d t ′ d t ′ = - k p C p ( t ) . (9)
The first term for p = 0 in the RHS of Eq 8 corresponds to the MSD of the center of the CD, and the motion obeys ∫ 0 t γ ( t - t ′ ) d X 0 ( t ′ ) d t ′ d t ′ = g 0 ( t ) and the FDR 〈 g 0 κ ( t ) g 0 λ ( t ′ ) 〉 = k B T N γ ( t - t ′ ) δ κ λ. According to the fluctuation-dissipation theorem [49], the motion of the center of mass is subdiffusive with exponent α (see S1 Text, Section IE):
X 0 ( t ) - X 0 ( 0 ) 2 = 2 〈 R 2 〉 CD A d f Γ ( 1 + α ) t τ d f , α α , (10)
where
τ d f , α ≡ N γ α 〈 R 2 〉 CD A d f · 3 k B T 1 / α (11)
represents the relaxation time of nucleosome fluctuations in the CD.
On the other hand, the second term in the RHS of Eq 8 describes the fluctuations of many modes inside the CD. Using the Laplace transformation and the thermal equilibrium initial state, the solution of Eq 9 can be derived as follows (see S1 Text, Section ID):
C p ( t ) = X p 2 CD E α - p 1 + 2 / d f t / τ d f , α α , (12)
where Eα(x) is the Mittag-Leffler function. According to the polymer physics [35] for t ≪ τdf,α, ϕ(n, t) is dominated by terms with large p. Moreover, since the MSD in our experiment (Fig 2E) is calculated by averaging the nucleosome trajectories at various positions in CDs, the term cos 2 ( p π n N ) can be replaced by the average 1/2. Therefore, according to the asymptotic form of the Mittag-Leffler function, Eα(−x) ≃ exp[−x/Γ(1 + α)] for x ≪ 1, and the conversion of the sum into the integral, we obtain for t ≪ τdf,α MSD ( t ) ≃ 2 B d f , α 〈 R 2 〉 CD A d f Γ ( 1 + α ) t τ d f , α α · 2 / ( 2 + d f ) , (13)
where B d f , α = d f 2 [ Γ ( 1 + α ) ] d f / ( 2 + d f ) Γ [ d f / ( 2 + d f ) ] is a dimensionless constant (see S1 Text, Section IF). Thus, in our model, subdiffusive motion of single nucleosomes is a typical feature, assuming both fractal CDs and viscoelastic medium.
In order to apply our model to living human cells, single-particle imaging of nucleosomes was performed by observation of PA-mCherry labels [51] attached to histone H2B in human HeLa cells (Fig 2A). The clear single-step photobleaching profile of the H2B-PA-mCherry dots shows a single H2B-PA-mCherry molecule in a single nucleosome (Fig 2B). We tracked approximately 40,000 dots representing single nucleosomes (S1 Table). Fig 2D shows representative trajectories of the dynamic nucleosome movement in single cells.
Here, to evaluate the state of CDs according to their position in the nucleus, we focused on the nuclear interior and periphery (or surface) (Fig 2C and S1 Fig), and calculated the MSD. The nuclear periphery is a heterochromatin-rich region, which presumably shows much less active transcription than the interior. The plots of the MSD at each region, in time interval t up to 0.5 s, are shown in Fig 2E (normal scale) and S2 Fig (log-log scale) (also see S1 Table). The MSD at the interior is higher than that at the periphery. This result implies that nucleosome movement within CDs in the euchromatin-rich interior region is higher than that in the heterochromatin-rich periphery region.
As we analytically derived the subdiffusive MSD (Eq 13), the experimental result clearly shows subdiffusion of single-nucleosomes: using Eq 1, the plots fit well with the MSD curves 0.018 t0.44 μm2 and 0.013 t0.39 μm2 for the interior and the periphery, respectively.
Comparing Eqs 1 and 13, β and Dapp are calculated as
β = α · 2 2 + d f , (14) D app = C d f , α · 3 k B T N γ α 2 / ( 2 + d f ) · 〈 R 2 〉 CD d f / ( 2 + d f ) , (15)
where C d f , α = 2 B d f , α ( A d f ) d f / ( 2 + d f ) Γ ( 1 + α ). It turns out that these values contain statistical information of the CD structures, 〈R〉CD and df. Since β and Dapp can be determined by the fitting in our experiments, we can therefore estimate 〈R〉CD and df, inversely.
The lower MSD at the periphery than at the interior, Dapp,periphery < Dapp,interior and βperiphery < βinterior, reflects the fact that the CDs near the periphery are in a more compact conformational state and are smaller in size than those at the interior: df,periphery > df,interior and 〈R〉CD,periphery < 〈R〉CD,interior. This property is consistent with the conventional distribution of heterochromatin: the CDs in the heterochromatin-rich nuclear periphery are more compact than those in the euchromatin-rich interior [52].
To estimate the structural information of CDs through solving Eqs 14 and 15 inversely, the values of N, α, and γα in mammalian living cell nuclei are required. The average size of TADs was determined to be 880 kb from mouse embryonic stem cells (mESCs), with a range of 100 kb to 5 Mb [10]. Here, we assume a CD size of 1 Mb, which corresponds to 〈N〉CD = 5000 nucleosomes. To the best of our knowledge, few studies have estimated the friction effect in viscoelastic cell nuclei. Therefore, we use the value of the diffusion coefficient of enhanced green fluorescent protein (EGFP)-monomer around interphase chromatin, DEGFP = 20.6 μm2/s [25], measured by fluorescence correlation spectroscopy, in which α is assumed to be 1. In general, as a result of the FDR in a viscoelastic medium with α, the diffusion coefficient of a diffusive particle for one degree of freedom is kBT/[Γ(1 + α) ⋅ γα,particle] (see Eq. S34 in S1 Text). Since the contribution of Γ(1 + α) is within the range 1 ≤ 1/Γ(1 + α)<1.13 for 0 < α ≤ 1, the friction coefficient of EGFP in the nucleus can be approximately regarded as the diffusion coefficient as γα → 1,EGFP = kBT/DEGFP. The hydrodynamic radius of a nucleosome bead with an H2B-PA-mCherry is assumed to be approximately quadruple for the EGFP. This means that the friction effect is also 4 times larger [48]. Accordingly, we use γα → 1 = 4kBT/DEGFP. Finally, the structural information of CDs is estimated by calculating
d f = 2 α β - 2 , (16) 〈 R 〉 CD = D app C d f , α 2 + d f 2 d f 4 〈 N 〉 CD 3 D EGFP 1 / d f . (17)
β could be measured in our experiment, although the value of α could not be determined simultaneously. Hence, Eq 16 represents the relationship between α and df, as shown in Fig 3A. Under this constrained condition, according to Eqs 16 and 17, the values of the structural information within the nuclear interior and periphery regions are calculated and mapped as a function of α (Fig 3B). Since fluorescence correlation spectroscopy measurements of GFP have shown that the value of α is close to 0.79 in not HeLa but NRK nuclei [31], as an example, we summarize the estimated values for α = 0.8 and α = 0.9 in Table 1. The exponent β = 0.4 for the fractal globule model [28] corresponds to the value for df = 3 and α = 1 in Eq 14. Furthermore, our previous results have shown smaller exponents β = 0.37 and 0.31 for interphase chromatin and mitotic chromosome, respectively [25]. Unless considering the case of 0 < α < 1, this smaller exponent cannot be explained. Note that α has only minor effects on Cdf,α (see S3 Fig).
The relaxation time of nucleosomes in CDs is calculated as
τ d f , α = 4 〈 N 〉 CD 〈 R 〉 CD 2 A d f · 3 D EGFP 1 / α , (18)
and is mapped as a function of α and df (Fig 3C). The short relaxation time (∼ s) means that the thermal equilibrium, which is the precondition for the linearization approximation, were fulfilled in our experiments. In measurements of long-term single-nucleosome movements, the MSD is expected to show a transition toward movement of the center of CDs with the exponent α (Eq 10). This would enable estimating α, df, 〈R〉CD, and τdf,α without requiring the use of the assumptive values described above, such as 〈N〉CD and DEGFP. The long-term (≫τdf,α) imaging of chromatin dynamics in mammalian nuclei might reveal this transition motion [19, 21, 24].
As mentioned at the beginning of this section, the measured TAD size of mESCs is in the range of 100 kb to 5 Mb. Fig 3D shows the relationship between 〈R〉CD and τdf,α for α = 0.9 as a function of 〈N〉CD, corresponding to the range of 200 kb to 4 Mb, according to Eqs 17 and 18. The relaxation time within several tens of seconds is consistent with the assumption of the linearization approximation as mentioned above. Moreover, the estimated CD size within 100–500 nm is also consistent with observed radius for chromatin domains as detected by super-resolution imaging [53].
The critical assumption of the linearization approximation is that nonlinear and complicated long-range interactions can be replaced by the mean-field fluctuation near thermal equilibrium within a fractal CD. Our result, that the estimated CD size is about 100–500 nm and the relaxation time is at most a few seconds, implies that the condition of the approximation is fulfilled. This kind of approximations has been discussed well in polymer physics [35], where the results of the approximation are not much different from those of more sophisticated calculations including the renormalization group theory. Furthermore, we have already reported that thermal fluctuation plays a dominant role in chromatin dynamics within CDs during a few seconds of observation for each fluorescent nucleosome [25–27]. On the other hand, successive ATP-dependent active processes on chromatin might play an important role in chromosome folding during mitosis [54]. In such a case, we cannot apply our theory due to the non-equilibrium nature.
In addition, non-equilibrium fluctuations driven by ATP-dependent cell activities affect chromatin dynamics [17, 20]. In order to directly take the effect into account, we have to add non-equilibrium fluctuation noise term to Eq 3. For a polymer with specific non-equilibrium fluctuations, where the correlation of the added athermal noise exponentially decays, a theoretical result has been obtained [55]. However, properties of non-equilibrium fluctuations on chromatin are unclear. It was suggested that a decrease of the chromatin persistence length of a CD may occur due to ATP-driven nucleosome remodeling [56]. We might consider an effect of ATP-driven remodeling on changes of the domain size and the fractal dimension.
Recent high-resolution Hi-C and ChIA-PET data have shown that architecture proteins such as CTCF and cohesin play important roles in CD organization at the boundaries [15, 16, 57]. However, because of experimental limitations including specific nucleosome labeling and microscopy resolution, we could not distinguish the nucleosome movements between the domain center and boundaries. Furthermore, different epigenetic states [53] including posttranslational modifications [58] affect the spatial organization of chromatin domains. These effects seem to regulate nucleosome-nucleosome interactions. If we could directly observe the nucleosome movements depending on the interactions in vivo using a novel labeling technology, we would be able to extend our framework; which will be a challenging issue in the future.
Here, we considered a locally clustered polymer with effective size scaling (Eq 4) in the absence of hydrodynamic interactions (HIs) as a model of CDs. The inverse proportion of kp to N, except for the contribution from 〈 X p 2 〉 CD, in Eq 6 reflects the lack of HIs in our model; that is, the hydrodynamic field goes through nucleosome beads without interactions. The hydrodynamic effect of surface monomers in a polymer blob on the exponent β has been argued in [28]. Applying their discussion to our results, Eq 16 changes into df = c(2α/β − 2), where the coefficient c is within the range 1 ≤ c < 1.09. The effect is expected to be small. One can also consider a polymer model including HIs, which would affect the mobility matrix and work cooperatively within a polymer blob [35, 50]. In such a situation, the HI cancels out the effect of the size scaling described by the fractal dimension df: β = α ⋅ 2/3, and β does not depend on df (see S1 Text, Section II).
Our results indicate that our proposed model serves as a strong method for extracting the structural information of CDs from observations of dynamic nucleosome movement. Super-resolution microscopy techniques can be used to elucidate the spatial size of CDs according to different epigenetic states [53]. On the other hand, development of an effective imaging technique to reveal the fractal dimensions remains a challenge for the future. The conformational state of CDs characterized by the fractal dimension must be associated with the accessibility of transcription factors, depending on the physical size of those factors [59]. Beyond the pioneer computational work of analyzing interphase chromosomes based on the chromatin fibers [60], further development of not only a large-scale chromosome model based on the results of a genome-wide association study [61] but also restraint-based three-dimensional modeling of genomes [62] is expected to provide novel insight and open the door toward further discovery on the relationship between dynamic genome organization and stochastic gene expression.
To observe single nucleosomes and analyze their local dynamics in living human cells, histone H2B was fused with photoactivatable (PA)-red fluorescent protein (mCherry) [51] and expressed in HeLa cells as described previously [25]. The cell lines expressing H2B-PA-mCherry at a very low level were isolated. The cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS) (vol/vol) at 37°C in 5% CO2 (vol/vol). The cells were plated 24–48 h before the experiment onto Iwaki glass bottom dishes treated with poly-lysine. Before the experiment, the medium was replaced by DMEM F-12 (non phenol red) with 15% FBS. The cells were then set on the microscope stage kept in a custom-built 37°C microscope incubator enclosure with 5% CO2 (vol/vol) delivery throughout the experiment.
For single-nucleosome imaging, an oblique illumination microscope was used to illuminate a limited thin area within the cell (Nikon laser TIRF microscope system Ti with sapphire 564-nm laser). In general, PA-mCherry exhibits red fluorescence only after activation by a 405-nm laser [51]. However, we unexpectedly found that a relatively small number (∼100/time frame/nucleus) of H2B-PA-mCherry molecules were continuously and stochastically activated even without UV laser stimulation. Fig 2A shows a typical single-nucleosome image of a living HeLa cell. Due to the clear single-step photobleaching profile of the H2B-PA-mCherry dots, each dot in the nucleus represents a single H2B-PA-mCherry in a single nucleosome (Fig 2B). Nucleosome signals were recorded in the interphase chromatin of the nuclear interior and periphery in living HeLa cells at a frame rate of ca. 50 ms/frame. Note that the two different focal planes for the nuclear interior and periphery (Fig 2C) were precisely ensured by nuclear surface labeling with Nup107 (a nuclear pore component)-Venus (a bright yellow fluorescent protein) [63] (see S1 Fig).
Local nucleosome fluctuation was observed (ca. 60 nm movement/50 ms), presumably caused by Brownian motion. The free MATLAB software u-track [64] was used for single-nucleosome tracking. The dots were fitted to an assumed Gaussian point spread function to determine the precise center of the signals with higher resolution. Finally, we obtained data set of two-dimensional Mi trajectories { ( x 0 j , y 0 j ) , ( x 1 j , y 1 j ) , …, ( x i j , y i j ) }, where the suffix j ∈ {1, ⋯, Mi} represents the sample number for the tracked time-interval [0, ti]; ti ≡ i × 50 ms. Several representative trajectories of fluorescently tagged single nucleosomes are shown in Fig 2D (bar = 100 nm).
According to observed regions, we calculated the ensemble-averaged MSD of single nucleosomes: MSD ( t i ) = 3 2 1 M i ∑ j = 1 M i [ ( x i j - x 0 j ) 2 + ( y i j - y 0 j ) 2 ]. Here, in order to obtain the three-dimensional value, we multiplied the two-dimensional value by 3/2 on the assumption of isotropy. Plots of the MSDs of single nucleosomes in interphase chromatin at the nuclear interior (10 cells) and the nuclear periphery (10 cells) from 0 to 0.5 s are shown in Fig 2E. The plots for single nucleosomes were fitted with the subdiffusion model (Eq 1) using R-software. The standard error of the mean (SEM), which is the standard deviation of the sampling distribution of the mean, for MSD(ti) was sufficiently small. The number of trajectories Mi and the SEM of MSD(ti) are summarized in S1 Table.
|
10.1371/journal.ppat.1002881 | Quantifying the Diversification of Hepatitis C Virus (HCV) during Primary Infection: Estimates of the In Vivo Mutation Rate | Hepatitis C virus (HCV) is present in the host with multiple variants generated by its error prone RNA-dependent RNA polymerase. Little is known about the initial viral diversification and the viral life cycle processes that influence diversity. We studied the diversification of HCV during acute infection in 17 plasma donors, with frequent sampling early in infection. To analyze these data, we developed a new stochastic model of the HCV life cycle. We found that the accumulation of mutations is surprisingly slow: at 30 days, the viral population on average is still 46% identical to its transmitted viral genome. Fitting the model to the sequence data, we estimate the median in vivo viral mutation rate is 2.5×10−5 mutations per nucleotide per genome replication (range 1.6–6.2×10−5), about 5-fold lower than previous estimates. To confirm these results we analyzed the frequency of stop codons (N = 10) among all possible non-sense mutation targets (M = 898,335), and found a mutation rate of 2.8–3.2×10−5, consistent with the estimate from the dynamical model. The slow accumulation of mutations is consistent with slow turnover of infected cells and replication complexes within infected cells. This slow turnover is also inferred from the viral load kinetics. Our estimated mutation rate, which is similar to that of other RNA viruses (e.g., HIV and influenza), is also compatible with the accumulation of substitutions seen in HCV at the population level. Our model identifies the relevant processes (long-lived cells and slow turnover of replication complexes) and parameters involved in determining the rate of HCV diversification.
| Hepatitis C virus (HCV) is a RNA virus that infects over 170 million people across the world. It leads to a chronic infection in the majority of people who are infected (>70%). Most people only discover that they are infected long after initial infection. Thus, it is difficult to study the very early events in infection. Here we study 17 individuals during the earliest possible stages of infection, from before the virus is detectable in the plasma to around 35 days post-infection. We focus on understanding the viral kinetics and the diversification of HCV during this acute phase of infection. During chronic infection HCV is present in the host as a swarm of multiple variants generated by its error prone copying. We studied the early diversification of HCV during acute infection using a new mathematical model of HCV replication. We found that after a phase of fast increase in viral load, accompanied by viral diversification, there is a stabilization of viral load and diversity levels. Using our model, we were able to estimate for the first time the HCV mutation rate during acute infection. We estimated the median in vivo viral mutation rate is 2.5×10−5 mutations per nucleotide per genome replication (range 1.6–6.2×10−5), about 5-fold lower than previous estimates. We also used a different approach, based on results of classical genetics, to calculate HCV's mutation rate and obtained consistent results (2.8–3.2×10−5).
| Hepatitis C virus (HCV) is a member of the hepacivirus genus within the flaviviridae family of virus, and it has a single positive stranded RNA molecule (∼9500 nucleotides) as its genome [1]–[3]. After entering a cell this RNA is translated into a single large polyprotein, which is cleaved to produce the viral structural and non-structural (NS) proteins [1]–[3]. The NS5B protein is a viral-specific polymerase, which is involved in replicating the HCV RNA genome [1], [4]. During genome replication the virion's positive strand RNA is copied into a complementary negative strand, which then must be copied back to produce a new positive strand. In the simplest replication model, this negative strand or a complex of the original positive strand and the newly created negative strand form an intermediate that acts as the template for producing new positive strands. This template plus various non-structural proteins form a structure called a replication complex [5]. If all new positive strands, and hence virions, are created from the same replication complex, we say that replication occurs by a “stamping machine" mechanism [6]–[9]. However, HCV infected cells often have more than one replication complex; indeed in vitro and in situ studies suggest there are about 40 such complexes in one infected cell [4], [10].
The HCV polymerase is an RNA-dependent RNA polymerase (RdRp) and hence does not possess error correcting mechanisms. Thus HCV replication, like that of other RNA viruses, is highly error prone [1]–[3]. Measuring the actual mutation rate, which derives both from the (+)RNA to (−)RNA and the (−)RNA to (+)RNA steps of replication, has been difficult [6], [11], [12]. A recent study determined the intrinsic error rate of the HCV polymerase in vitro using enzyme kinetic measurements [12]. They found high error rates, of ∼10−3 per site, for transitions and about 100-fold lower rates for transversions. Still, the in vivo mutation rate is likely different. Mutation is difficult to estimate in vivo due to selection, multiple rounds of replication and incomplete sampling [6], [11]. One proposed way to determine the in vivo mutation rate is to estimate it based on the frequency of lethal mutants in the viral population at any given time [13]. In fact, classical genetics shows that the frequency of a lethal mutation in a haploid population in mutation-selection balance is μ, the mutation rate. A recent study used this method to estimate an upper limit for the in vivo mutation rate of HCV as (1.15±0.29)×10−4 per nucleotide per replication round [13], which is within the range of other RNA viruses [6].
This high mutation rate is consistent with the high degree of HCV diversity found across the population of infected individuals [14], [15]. Indeed, HCV is highly variable, with multiple subtypes, and a global diversity that is higher than that of HIV-1 [15]. Clearly, this population level diversity, which reflects the HCV evolution rate, is in part prescribed by the mutation rate of the virus in vivo [16]. Moreover, in chronically infected individuals the HCV viral population is also diverse [17]. This diversity allows fast evolution and escape from immune [18] or antiviral drug pressure [19], and may contribute to HCV pathogenesis [18], [20].
An important question is how HCV diversity is generated. While it clearly depends on the mutation rate, we shall show using a model of HCV replication that it also depends on other parameters of the HCV life cycle [7]–[9], such as the long-lived nature of infected cells, as compared to HIV infected cells [21], [22], the existence of multiple replication complexes within an infected cell [4], [10], and the turnover rate of these replication complexes. In order to validate this model and obtain quantitative estimates of the in vivo HCV mutation rate, we shall exploit our observations in an accompanying report [23] and those of others [24], [25] that during the initial stages of primary infection the viral population is comprised of discrete low diversity lineages of viral sequences emanating from the transmitted/founder viral genomes [23]. Further, early on, diversity increases with time since infection. We shall show that the rate of diversification is not constant but rather slows as infection is established. Our model provides a quantitative explanation for this phenomenon. Analyses of HIV evolution in acute infection have been used to estimate the time since infection [26], [27]. Here, we know with reasonable accuracy the time of infection, but use the same ideas to estimate the in vivo mutation rate of HCV.
The early dynamics of viral increase in HCV infection is different from that seen in other chronic infections, such as HIV [28] and HBV [29]. The HCV viral load in the subjects in this study increases roughly exponentially until it reaches a plateau (Figure 1A). This has also been observed in a prior study of acute HCV infection [30] and observed in chimpanzees experimentally infected with HCV [31]. Quantitative characteristics of this early increase are given in Table 1. The median time between the last negative sample and the first HCV positive sample in our dataset was 5 days, which is consistent with a viral dynamics analysis of larger numbers of plasma donors [30]. Because of this short interval, we assumed that the virus started expanding at the last negative sample. If the virus started expanding after this, our estimated expansion rate would be an underestimate. The median HCV RNA exponential growth rate was 2.2/day, corresponding to a doubling time of 0.31 days (or 7.4 hours). The median peak viral load observed was 3×106 HCV RNA IU/ml and it took a median of 21 days to reach this level. The virus then stayed at approximately this high viral load level for a median of at least 26 days. In two subjects, we did not have enough follow-up to conclusively affirm whether a plateau exists or not. These estimates are in agreement with a previous study of 77 plasma donors with longer follow-ups, which reported an estimate of ∼6 days of viral expansion before the first positive measurement (compared to a median of 5 days in our dataset) and a mean plateau duration of ∼56 days [30].
The observation of the viral load plateau suggests that the number of infected cells reaches a steady state level a couple of weeks post infection. It is possible that this is a dynamic steady state, with removal of infected cells in equilibrium with generation of new infected cells. However, HCV is likely non-cytolytic [32], consistent with the normal levels of alanine aminotransferase (ALT<40 IU/L is upper limit of normal [33], [34]) in these individuals early in infection (Figure 1B). In addition, prior work has suggested that the cytolytic immune response takes weeks to months to emerge [31], [35], [36] (consistent with an increase in ALT to 10× to 20× the normal level late in acute infection [37]). Thus, it is likely that the rate of infected cell death during this early period is comparable to that of uninfected cells. The lifespan of uninfected hepatocytes has been estimated as being on the scale of months to years [38], [39], and thus infected cell death is probably negligible at these early times. In this case, the plateau in viral load suggests an equilibrium where all cells that can be infected are infected and producing virus. Assuming that there are 1011 hepatocytes in the liver [40], we estimate that a median of 6% (with range 1.7%–22%) of these are infected across our subjects (Table 1), consistent with experimental measurements in chronic infection [41], including recent estimates by two-photon microscopy of frozen sections of liver tissue [42]. Thus, primary HCV infection is characterized by fast growth of viral load to a plateau where only a minority of hepatocytes is infected.
To evaluate how HCV diversity changes during primary infection, we performed single genome amplification (SGA) followed by direct amplicon sequencing [23], [26], otherwise known as single genome sequencing [43], at multiple time points in the subjects shown in Figure 1. SGA is achieved through serial dilution of the cDNA obtained by reverse transcription of HCV RNA from plasma (see Methods and [23] for details). We amplified 5′ half-genome sequences, on average 4879 nucleotides, covering core, E1, E2, p7, NS2 and most of the NS3 proteins of HCV. For early samples, with low viral loads, we amplified the same region, but in two separate assays of one quarter genome each to enhance sensitivity of amplification. In this way, we obtained 84 sets of sequences for the 9 subjects at multiple (between 3 and 5) time points. On average, we had 44 sequences per time point. All of the sequences were deposited in Genbank; see Li et al. [23] for further details and accession numbers.
We then aligned separately the set of sequences for each time point and for each sequence region and used a sequence visualization tool (Highlighter – www.HIV.lanl.gov), to analyze the sequence diversity based on individual nucleotides. This tool allowed us to identify low diversity monophyletic lineages corresponding to the putative transmitted/founder (T/F) viruses – the consensus at the earliest time point from SGA data [23]. We next confirmed that these lineages were maintained across the times sampled, to guarantee that we were analyzing the diversification of the same lineage over time. In cases where there were two or more putative T/F viruses, we analyze only the dominant lineage, as SGA sequence data was too limited to study the minor lineages.
From these 84 sequence alignments, we were able to study the evolution of the virus and the emergence of new mutations from very early in infection (mean: 7 days, range 2 to 15 days since the last negative sample across the 9 patients) until late in the plateau phase of viral load (mean: 33 days, range 21 to 42 days). We found that HCV sequence diversity increases quickly early on, but then stabilizes in 7 patients, starting at about day 14; in subject 10051 there was not enough follow up to assess this issue, and in subject 10029 a clear stabilization of diversity was not observed. The plateau of diversity occurred when an average of 46% of the sequences were still identical to the inferred T/F viral genomes. In three subjects (10029, 10062, 9055) there was an increase in diversity at late times, ∼35 days. Note that for 10062, this is coincident with an increase in ALT levels suggesting turnover of infected hepatocytes (Figure 1B).
We also found that in the vast majority of cases, HCV diversity at each time point was consistent with a star-like phylogeny, i.e. the viruses' sequences coalesce at a single genome founder [27], [44]. The only exception was the 5′-half of 9055 at the last sampling time point, day 38, when there was evidence for the onset of immune selection [23]. The mutations detected in the sequence sets also conformed to a Poisson distribution in the inter-sequence pairwise Hamming distances [27]. The exceptions were the 5′-half of 10029 at day 13, the second 5′ quarter of 10029 at day 34, the second 5′ quarter of 10051 at day 7, and the first quarter and 5′-half of 10051 at day 21. Due to the specifics of the HCV replication life-cycle, one predicts occasional violations in star-like diversification and in the fit to the Poisson distribution, because there is a non-negligible probability of shared stochastic mutations between HCV sequences. That is, shared mutations may occur even in the absence of selective forces. See the accompanying report [23] for a more detailed discussion of these issues.
We next developed a model of HCV replication to study the time course of accumulation of mutations and to estimate the in vivo mutation rate of HCV needed to describe the observations above. This stochastic model of HCV replication allowed us to study the time course of viral load changes and the accumulation of mutations in the study subjects (see Methods). In the model, we assume cells are infected by a single virion, i.e., that superinfection does not occur [45], [46]. We further assume that in every infected cell, on average, only a fraction k of newly synthesized viral (+) strand RNA (vRNA) is exported in new virions, and the rest, 1-k, forms new replication complexes (RC). We assume that vRNA degradation can be neglected, i.e., that the newly synthesized vRNA is either rapidly complexed with proteins and converted into stable RC, or rapidly encapsidated and exported. (Note that this is very different from analyses of HCV treatment, when production of vRNA and/or virion assembly/release may be blocked, and vRNA degradation becomes an important parameter in the clearance of infection [47]). These processes are assumed to continue until the cell generates a maximum number of replication complexes (RCM). Note that if we set k = 1, so that all synthesized vRNAs are exported, we recover the “stamping machine" mode of replication [7]–[9], where all virions result from the same replication complex, i.e., the same negative strand of RNA. The existence of multiple replication complexes within one cell corresponds to “geometric growth". In our model, after a virus is exported, a fraction 1-θ of the released virions is assumed to be cleared from circulation [5], and the remaining fraction, θ, is assumed to infect new cells. We also assume infected cells are long lived, and thus, we initially neglect death of infected cells during the first few weeks of infection. This assumption is consistent with the viral load profiles seen in the infected subjects, where viral load increases rapidly to a maximum level and plateaus at this level for weeks.
We used our model to reproduce the viral load data (Figure 1A). For each subject, the only free parameter available to determine the trajectory of virus over time is the fraction of vRNA exported, k, since all other parameters are fixed a priori or are calculated as a function of k (see Methods). We found that the model could describe the viral load data well with just this single adjustable parameter. The values estimated for k indicate that most of the synthesized vRNA is exported as virions (median k = 0.77, range 0.42–0.89). Moreover, the estimated values of k are quite similar among the different individuals, with the exception of 10025, who has a lower estimated k ( = 0.42). However, this subject has only one viral load measurement during the up-slope of the virus, which strongly influences the value estimated for k. Indeed, for this individual, choosing higher values for k lead to only slightly lower quality fits (not shown).
Next, we used our model to analyze the diversification profiles of HCV in these patients. As the viral RNA is copied, errors in the incorporation of nucleotides are possible, i.e., mutations occur. If we let μ denote the probability that a base in the newly produced virion differs from that in the infecting virion, then for the stamping machine model the mutation rate, μ, is simply twice the rate at which bases are miscopied by the HCV RdRp, to account for the cycle of (+)RNA strand→(−)RNA strand→(+)RNA strand copying. With multiple replication complexes in a cell, opportunities exist for additional copying errors to be made since a newly synthesized (+)RNA strand needs to be copied again to make a replication complex. Every time a RNA strand incorporating a mutation is made, there is a probability that this mutation is lethal, and the virus or replication complex made from such RNA is non-functional. Prior experimental studies indicate that the fraction of random mutations that are lethal is about 40% in RNA viruses [48].
We incorporated mutation in our model to analyze the viral diversification data and estimate the mutation rate needed to match the observed accumulation of mutations. We assume that at time zero the putative T/F virus starts replicating and mutating. We then compute the decrease over time in the fraction of sequences identical to the T/F virus (i.e., “the fraction of unmutated viruses"). We compare this model prediction to the identical measurement in our subjects and varied the mutation rate to obtain the best agreement between model and sequence data obtained from plasma HCV RNA, which corresponds to (+)RNA strands.
The best description of the data was obtained for a median mutation rate (for the half-genomes) of μ = 2.5×10−5 per nucleotide per replication (Figure 2A–C). Moreover, this estimate was consistent across subjects and across regions of the genome (range: 1.6×10−5–6.2×10−5 per nucleotide per replication, Table 1).
Our model exhibits a fast decrease in sequence identity early in infection, as the viral load increases exponentially and more and more cells are infected, followed by a stable viral diversity level as the virus reaches and stays at its plateau. This stasis in viral diversification is compatible with the assumption that the plateau in viral load corresponds to a stable pool of infected cells. This indeed seems to be the case for 5 of the patients (Figure 2A–C); for 1 case there is not enough data. If the plateau in viral load corresponded to a dynamic steady state in which infected cells were dying and being rapidly replaced, our model would predict a continuous increase in diversification resulting from the continuous replacement of replication complexes. In a few cases, we did see an increase in diversity at times later than 30 days, and in three patients (10029, 10062, and 9055) the observed long term behavior (later than about day 35) deviates from that predicted by our simulations. This difference between model and data could be due to sampling error, for example the 95% CI for theory and data at day 42 overlap for patient 10062. Alternatively, some processes not accounted for in the model may be operational at these later time points, leading to increased diversity. For example, for subject 9055 anti-HCV antibodies are detectable at this late time point and there is strong evidence of CTL selection (escape or reversion) [23]; and for 10062 there is a late increase in ALT (Figure 1B), which suggests the initiation of a CTL response consistent with renewed cycles of infection.
Our model also makes predictions about the distribution of mutations across the population. Interestingly, our model not only matches the fraction of unmutated viruses, but also the fraction of viruses with 1, 2, 3, … mutations, even though this detailed data was not used to parameterize the model (Figure 3A–C). We obtained excellent agreement with the data, except when we observed a late increase in diversity in the three patients discussed above (10029, 10062, 9055). We tested this agreement for the 5 h genomes by a Monte Carlo test [49], since the number of expected mutations is low (<5) in several cases. The null hypothesis is that the data follows the theoretical expected values, and with the exception of those three patients, there was good agreement between observed and predicted mutation counts (p>0.05). Moreover, if we consider the distribution of mutations at the previous time for which we have SGA data, this agreement was also seen in 10029 and 10062 (p>0.05, and we cannot reject the null hypothesis).
We next tested whether our results were dependent on the particular values assigned to the parameters that we fixed in the simulation (see Methods). We found that both the viral load time course and the viral diversification were not sensitive to particular values of these parameters (Figure S1 in Text S1). For example, we assumed a maximum of RCM = 40 replication complexes per infected cell, as seen in vitro [4] and in situ [10]. Clearly this number could be different in vivo. However, our results were essentially the same, when we varied RCM from 10 to 80 (Figure S1 in Text S1).
To further confirm the robustness of our results, we next used the method suggested by Cuevas et al. [13] for estimating the mutation rate of HCV by analyzing the frequency of lethal mutations. Classical genetics shows that the frequency of lethal mutations is equal to the mutation rate, since all such mutations should be produced directly by mutation in the last replication round. As in Cuevas et al. [13], we used non-sense (stop codon) mutations as a proxy for lethal mutations. The concept is to count all stop codons in the data set and to divide this by the number of mutation targets (non-sense mutation targets – NSMT), i.e. codons that by a single mutation could generate a stop codon (see Text S1 for details). For these analyses, we were able to use all 17 patients in our cohort, thus expanding our data set.
In total we had 898,335 NSMTs and 13 stop codons in the over 1×107 bases sequenced [23] (Tables S1 and S2 in Text S1). Surprisingly, 4 of the stop codons were identical and at the same position in 10051 at two different time points (see Table S1 in Text S1). This strongly indicates that this stop codon appeared only once in this patient, and that stop codons may not be lethal in HCV but instead complemented by intact genomes within the same cell. Thus, we counted this stop codon only once, for a total of 10 mutations leading to stop codons. A calculation identical to that proposed in [13] then shows that μ = 3.2×10−5 per nucleotide per replication, which is fully consistent with our estimate above. We also propose an improved way to calculate this rate from the same data (see Text S1), and with this method obtain μ = 2.8×10−5 (binomial 95% CI: 1.4–5.2×10−5).
Altogether, these data and analyses indicate that HCV sequences diversify early in infection, during the exponential increase of viral load, which is then followed by a plateau in diversity for up to a few weeks. The mutation rate needed to explain these observations (μ≈2.5–3.2×10−5 per nucleotide per replication, Figure 3D) is 5 and 100 times smaller than previously reported for HCV [13] and its purified RdRp [12], respectively.
We next investigated in detail why HCV diversification appears to stop after a few weeks of infection, and what processes could break this plateau in diversity, since in chronic HCV infection the virus is much more diverse [23]. In particular, we analyzed the effect of turnover of replication complexes and the emergence of the cytolytic immune response.
In the baseline simulations of the model, we neglected the turnover of replication complexes (RC). However, RC may degrade. In this case, to sustain viral replication, the RC would need to be continuously produced to balance their degradation. Thus, we next analyzed the impact on our model predictions of including RC degradation.
For fast RC turnover (e.g., half-life 1.5 d), most (median of 59%) of the simulated infections die out, and those that lead to sustained infection show a slow growth of the virus that is not compatible with the data (Figure 4A, left panel). It is possible to recover fast viral growth rates, if one postulates that a larger fraction of newly synthesized RNA is used to form new replication complexes (i.e., if k is smaller). When the turnover of replication complexes is not negligible (t½<5 days), on the time scale of our simulations, the accumulation of mutations is faster at later times as replacement of replication complexes occurs (Figure 4A, right panel). In this case, to describe the data a smaller mutation rate would be needed, at least in some patients. Importantly, turnover of replication complexes also implies a continued increase in diversity throughout the observation period, since more (−) strand RNA needs to be made and hence there is more opportunity for mutations to occur. However, such a continued increase in diversity is not seen for subjects 10012, 10017, 10021 and 10025. On the other hand, this process could help explain the marked increase in diversity seen at late time points in subjects 10029, 10062 and 9055. Note however that even these subjects seem to have a stabilization of diversity prior to this marked increase, which is not compatible with fast turnover of replication complexes. If the turnover of replication complexes is much slower (eg., ∼15-day half-life) then the profiles do not differ from our baseline case where there is no turnover over the 50 day period studied.
Here we studied RC turnover inside the cell, but it is also possible that cells die due to the immune response against HCV, thus forcing re-generation of RC. Thus, we next considered the effects of cell turnover on the results of our model.
An effect of immune processes is removal of infected cells. Because there may be some limit to the number of infected cells in the liver [42], the death of infected cells may allow new cells to be infected, which in turn generates new RC and the opportunity for mutation accumulation. For all subjects for whom there is enough data, we find a stabilization of diversity, which in a few cases is then followed by a “sudden" marked increase at a later time point (10029, 10062, 9055). It could be that this latter pattern is an artifact of sampling. For example for 10062, the observed fraction of unmutated sequences at the three time points sampled have confidence intervals that overlap, and those fractions are not significantly different, p = 0.07 (Figure 2C, overlap of vertical bars). In our model this stabilization in diversity accumulation occurs because a steady-state is attained for the numbers of replication complexes and infected cells, without continued turnover. Rather than a sampling issue, it is possible that the observed increase in diversity is due to an immune response emerging at late time points, which leads to an increase of the infected cell death rate (δ). Indeed, this is indicated in studies of experimental infection of chimpanzees, where the immune response is delayed several weeks [31], [36]. In this context, an alternative explanation for the increase in diversity in 10062 is the appearance of such an immune response as suggested by the increase in ALT in this subject (Figure 1B). To study the effect of a late immune response that kills infected cells, we allowed for this process starting at 30 days post infection (Figure 4B). As expected, the emergence of an immune response lowers the viral load, possibly leading to a new lower viral load steady state, as is observed in some experimentally infected chimpanzees [31]. With the loss of infected cells, new cycles of infection occur along with creation of new replication complexes, and the model predicts a renewed increase in the accumulation of diversity, which mimics the data in some subjects (eg., 10062, 9055). However, we do not have enough data to precisely estimate the timing and magnitude of this immune response.
We analyzed the viral dynamics and viral diversification of HCV very early in acute infection. The early diversity of HCV is very low, and the inter-sequence Hamming distances follow a Poisson distribution, as would be expected when the mutations occur approximately at the same rate at all positions and the sequences are not selected for diversity [27], [44]. Given this observation, the number of mutations at early times should depend on the time since infection, the mutation rate and the biology of viral replication. This idea has been used before in the context of primary HIV infection to estimate the time of infection, assuming a given mutation rate [26], [27]. In the present study, the time of infection is known to within a short time window, with the first HCV positive sample within 5 days of the last negative sample. With this information, we could use our data to estimate the in vivo HCV mutation rate. By developing a model of HCV replication that takes into account the details of the viral lifecycle, we found the estimated mutation rate varied among subjects between 1.6×10−5–6.2×10−5 mutations per nucleotide per replication cycle, with a median of 2.5×10−5 (Table 1, 5 h genome). This estimate was very robust to different assumptions about model parameter values (see Text S1). Moreover, we systematically made conservative assumptions for the less well known parameter values leading to higher estimates for the mutation rate. To further confirm our results, we estimated the mutation rate by a completely different approach based on the frequency of stop codons (non-sense mutations), corrected by the number of non-sense mutation targets, as proposed by Cuevas et al. [13]. With this calculation we obtained a mutation rate of 2.8×10−5 or 3.2×10−5 mutations per nucleotide per replication cycle depending on the calculation method (see Text S1), which is consistent with the estimate from our more complex dynamical model and substantially less than the rate (∼10−4) estimated by Cuevas et al. [13]. A likely explanation for the difference between the findings of our nonsense mutation analysis and that of Cuevas et al. is that in our study Taq polymerase errors are eliminated from the finished sequences by the SGA-direct amplicon sequencing method and thus do not enter in the error rate calculations; this was not the case for the previous analyses [6], [13]. We further note that estimates of the HCV mutation rate based on nonsense mutations are likely to be overestimates since we found that stop codons were not always lethal (see Text S1). One explanation for this observation is that there are multiple HCV RNAs in an infected cell and another RNA may complement nonsense mutations. Indeed, we also found a case of a chronically infected patient who has a strain with a large deletion replicating in plasma at multiple time points [23]. Moreover, for dengue virus (in the same Flaviviridae family of HCV) there is a report of a viral strain with a stop codon that spread and attained a high frequency in the population, implying replication in both humans and mosquitoes [50].
In addition, our analysis does not account for mutational errors resulting from the cDNA synthesis step of the sequencing process, which again may lead to an overestimation of the mutation rate. However, we used Superscript IIITM Reverse Transcriptase (Cat. No. 18080-093, 2000 units, Invitrogen Life Technologies, Carlsbad, CA) that has been reported to have an error rate of ∼2×10−6 mutations/nucleotide/replication [23], [51], which is at least 10-fold lower than our HCV mutation rate estimates, and hence should not significantly influence our estimates.
Our estimates of the mutation rate for the HCV RdRp of ∼2.5×10−5 are notable because previous reports have suggested that the in vivo mutation rate of HCV is of the order of 10−4 mutations per nucleotide per replication [13]; and that the in vitro rate of the isolated RdRp could be as high as 10−3 [12]. One possible explanation for the latter discrepancy is that the mutation rates observed with purified RdRp enzymes are generally larger than those seen in vivo, because in vitro analyses cannot recapitulate the intracellular milieu of the replication or polymerase complex. For example, in the case of HIV reverse transcriptase, the errors measured with purified enzyme were found to be up to 20-fold higher than those measured in infected cells [52]. Another possibility is that we may have missed some low prevalence strains. However, a detailed power calculation shows that with the number of sequences obtained per patient, we would only miss strains that are present at very low levels, below 2% [23], which is much better than was possible before [25], [53] (see Li et al. [23] for a detailed discussion). Moreover, for the dynamical model we follow time courses and analyzed the fraction of virus identical to the T/F virus; and for the stop codon analyses, we corrected for the mutational targets. Both of these lower the impact of missing strains.
Given the low level of diversity observed in early infection and the relatively low mutation rate, the enormous diversity of HCV [14], [15], [18] and its high substitution rate (i.e., substitutions/site/year) have to be understood in light of HCV's replication mechanism [16]. Relatively long-lived infected cells, with multiple replication complexes allow for the accumulation of diversity in the virions produced. At the same time, the turnover of both replication complexes and infected cells, which must surely ensue as the immune response develops, allows for renewed generation of diversity throughout the course of infection (compare 10062 in Figure 1B and Figure 2C). Indeed, it could be that these details of the life cycle are responsible for the large diversity of HCV. We note that HIV and influenza, which are thought to have similar mutation rates to the one estimated here [6], [52], also have high substitution rates [54]. In this context, we see that accumulation of diversity is not only dependent on mutation rate, but also to a great extent on the particular processes of the viral life cycle [7], [8], [16]. Clearly, the pressure of the immune response, once established, will be important in determining relative fitness of many of the mutations and in determining the spectrum of mutations observed. That we see only scarce evidence of positive selection in our dataset indicates that there is a window of several weeks before the effects of the immune response can be detected.
Another important parameter that we estimated was the fraction of infected cells during the early plateau in viral load, which ranged between 1.7% and 22% of hepatocytes. This fraction is in reasonable agreement with other studies of HCV [41], [42]. In our model, this fraction depends on the value assumed for the maximum number of replication complexes (RCM). The larger the number of replication complexes in an infected cell, the more viruses this cell can produce per unit of time, and thus the fewer the number of infected cells needed to maintain a given steady state viral load. However, increasing RCM has little effect on our estimate of the mutation rate (see Text S1).
In this study, we constructed a simple model of HCV replication that tried to capture the most salient features of the viral life cycle. Moreover, we were careful to choose parameters consistent with the literature a priori, so that only 2 parameters had to be adjusted to fit the data on viral growth and diversity increase. We tested variation in the model assumptions and found that the results were quite robust. Still, it is clear that many complexities could be added to the model. For example, instead of having a fixed RCM, we could allow it to vary from cell to cell and possibly even from time to time; or we could allow for a distribution of generation times for RNA synthesis. These and other processes are easy to include in the model, however we opted to keep to the essential aspects of the replication process, so that we did not have to make further assumptions, which would complicate the interpretation of the results. In essence, this is akin to choosing a simple experimental system that is amenable to easy manipulation and interpretation of results, even if it does not represent fully all the details of in vivo system.
Altogether, the unique dataset presented here, including HCV viral kinetics and genomic diversification very early in infection, revealed that the initial exponential expansion of HCV RNA is followed by a plateau in viral load that lasts up to a few weeks [30]. The initial viral expansion is accompanied by a fast early increase in sequence diversity, whereas during the viral plateau viral diversity remains approximately constant. During the plateau viral production continues but is simply balanced by the rate of viral clearance. In order to understand why viral diversity did not continue to increase during this period, we develop a novel stochastic model of HCV infection. The basic idea behind the model is that during the early exponential expansion of the virus, new cells are being infected and generating multiple replication complexes in each infected cell. This involves multiple copying events of (+)RNA to (−)RNA to (+)RNA, etc, with errors potentially being generated at each stage. We postulate that once the viral plateau is reached a stable population of long-lived infected cells has been generated which then produce the plateau virus without any need for new RC generation. If no new replication templates are made then there is little opportunity for mutations to accumulate, though each virus can still mutate in relation to its parent RC due to the (−)RNA to (+)RNA copying event. We found that our model, based on this idea, agreed with both the viral load kinetic data and the sequence diversity data if we assumed that the in vivo mutation rate of HCV is ∼2.5×10−5 per nucleotide per replication cycle. This is about 5-fold lower than previously reported, but still high enough that coupled with the long-lasting nature of HCV infection and the very high turnover of virus in chronic infection leads to substantial HCV diversity in an individual and in the population.
Plasma samples were obtained from seventeen regular source plasma donors, who became HCV infected during periods of twice-weekly plasma donations. The donors were untreated and asymptomatic throughout the collection period. All subjects gave written, informed consent and the study protocols were approved by institutional review boards at the University of Pennsylvania, the University of Alabama at Birmingham and Duke University. HCV RNA and antibodies were analyzed as described elsewhere [23].
Single genome amplification (SGA) followed by direct amplicon sequencing was performed on sequential plasma vRNA samples (i.e., (+) RNA strands), as described in detail elsewhere [23].
For our dynamical analyses, we selected subjects who had at least two time points sampled with single genome amplification assays [23]. Thus, three subjects were not included – 6213, 6222, 10004. Six subjects (10002, 10003, 10016, 10020, 10029, 106889) had more than 7 putative T/F viruses, which makes a diversification analysis impractical, both due to the complexity of the viral species in the subjects and the small number of sequences representing each lineage [23]. The exception was 10029, who had a dominant lineage with more than 38 sequences for each time point, and we included this subject in our analyses. Thus, there were 9 subjects who were sampled at multiple time points and who had a clearly dominant putative T/F virus lineage [23]. Here we only analyzed these dominant lineages, for which we have the most data (SGA sequences).
Sequence alignments were initially made with ClustalW and then checked individually using JalView 2.6.1 (www.jalview.org). We used ConsensusMaker (www.HIV.lanl.gov) to calculate the consensus of the first set of sequences sampled by SGA, which is the putative T/F virus [23]. The set of sequences from each SGA sample with the corresponding consensus was analyzed by PoissonFitter (www.HIV.lanl.gov) to calculate for each sequence the number of mutations away (i.e., Hamming distance) from the T/F, and to test whether sequence diversification conforms to a star-phylogeny and if the set of inter-sequence Hamming distances follow a Poisson distribution [44].
Altogether we analyzed time courses of thousands of sequences with over 11.9 million base pairs and 1887 mutations [23].
To analyze the process of replication of HCV and how it affects the generation of diversity in primary infection, we developed an agent based model of HCV infection and replication. We assumed cells are infected by a single virion, and that in every infected cell, on average a fraction k of newly synthesized viral RNA (vRNA) is exported in new virions, and the rest, 1-k, form new replication complexes in the cell. These processes continue until the cell contains a maximum number of replication complexes (RCM). We assume this maximum value is set by the availability of host factors. After a virus is exported, a fraction 1-θ of released virions are cleared from circulation, and the rest, θ, infect new cells.
As the vRNA is copied, errors in the incorporation of nucleotides are possible. Every time a mutation occurs, there is a probability that this mutation is lethal, implying a virus or replication complex made using such mutant vRNA is non-viable. Sanjuan [48] estimates that the fraction of random mutations that are lethal is about 40% for RNA viruses.
We assumed HCV is noncytolytic [32]. Thus, infected cells can produce virus for long periods of time – until the infected cell dies, either from natural death or immune attack. Early in acute infection there is little evidence of cytotoxic T cell activity and CD8+ T cells do not appear to enter the liver until many weeks after infection [36]. In addition, normal hepatocytes live for months [39] to a year or more [38], thus, we either totally neglect death of infected cells or allow death after the first few weeks of infection. The assumption of no early death is consistent with the normal levels of alanine aminotransferase (ALT) measured in these individuals (Figure 1B) and the viral load profiles, where viral load increases rapidly to a maximum level and then stays at that level for some time. (This is in stark contrast for example with HIV, a cytolytic virus, where a clear peak in viral load is seen during primary infection followed by a decrease in viral load [28].)
Replication of the RNA and formation of a new virion or replication complexes is not instantaneous, as it takes a certain amount of time for synthesis of the different molecular components and their assembly. Although this time is most likely variable from replication cycle to replication cycle and from cell to cell, we assume that it is similar for all replication events in our model, fixing it at an average time to complete all the replication steps. This time we call the “generation time". Most likely it will take longer to produce the first copied RNA upon cell infection than later ones, as various molecular events need to occur before virus production begins (eg., uncoating, polyprotein synthesis and cleavage, assembly of the replication complex, etc) [5].
In the simulation, based on the assumptions described above, we follow the number, age (in the sense of generations) and mutational burden of each virion and each replication complex inside infected cells. The simulation was implemented in the R language (www.r-project.org). Because these are stochastic simulations, there is variability from one run to the next, even when all parameters remain the same. Thus, for each patient and each set of parameters (in Figures 1–4) we present results from 100 runs. Including more runs (we tested some cases with 200 runs) does not significantly alter the results presented.
The parameters of the stochastic model are as follows:
|
10.1371/journal.ppat.1005223 | Rescue of a Plant Negative-Strand RNA Virus from Cloned cDNA: Insights into Enveloped Plant Virus Movement and Morphogenesis | Reverse genetics systems have been established for all major groups of plant DNA and positive-strand RNA viruses, and our understanding of their infection cycles and pathogenesis has benefitted enormously from use of these approaches. However, technical difficulties have heretofore hampered applications of reverse genetics to plant negative-strand RNA (NSR) viruses. Here, we report recovery of infectious virus from cloned cDNAs of a model plant NSR, Sonchus yellow net rhabdovirus (SYNV). The procedure involves Agrobacterium-mediated transcription of full-length SYNV antigenomic RNA and co-expression of the nucleoprotein (N), phosphoprotein (P), large polymerase core proteins and viral suppressors of RNA silencing in Nicotiana benthamiana plants. Optimization of core protein expression resulted in up to 26% recombinant SYNV (rSYNV) infections of agroinfiltrated plants. A reporter virus, rSYNV-GFP, engineered by inserting a green fluorescence protein (GFP) gene between the N and P genes was able to express GFP during systemic infections and after repeated plant-to-plant mechanical passages. Deletion analyses with rSYNV-GFP demonstrated that SYNV cell-to-cell movement requires the sc4 protein and suggested that uncoiled nucleocapsids are infectious movement entities. Deletion analyses also showed that the glycoprotein is not required for systemic infection, although the glycoprotein mutant was defective in virion morphogenesis. Taken together, we have developed a robust reverse genetics system for SYNV that provides key insights into morphogenesis and movement of an enveloped plant virus. Our study also provides a template for developing analogous systems for reverse genetic analysis of other plant NSR viruses.
| Reverse genetics is a powerful tool for fundamental studies of virus biology, pathology and biotechnology applications. Although plant negative-strand RNA (NSR) viruses consist of members in the Rhabdoviridae, Bunyaviridae, Ophioviridae families and several unassigned genera that collectively account for many economically important crop diseases, unfortunately, several technical difficulties have hindered application of genetic engineering to these groups of viruses. This study describes the first reverse genetics system developed for plant NSR viruses. We report an efficient procedure for production of infectious virus from cloned cDNAs of sonchus yellow net virus (SYNV) RNAs, a model plant rhabdovirus. We have also engineered a recombinant SYNV vector for stable expression of a fluorescent reporter gene. Using this system, we have generated targeted SYNV mutants whose analyses provide key insights into enveloped plant virus movement and morphogenesis processes. Moreover, our findings provide a template for reverse genetics studies with other plant rhabdoviruses, and a strategy to circumvent technical difficulties that have hampered these applications to plant NSR viruses.
| Negative-strand RNA (NSR) viruses have major impacts on public health, agriculture and ecology, and they collectively are responsible for some of our most serious human, veterinary, wildlife and plant diseases [1]. Plant NSR viruses comprise members of the Rhabdoviridae, Bunyaviridae, Ophioviridae families, and of the unassigned Emaravirus, Tenuivirus, Varicosavirus and Dichorhavirus genera and account for many economically important crop diseases [1–3]. Most members of the plant NSR viruses are transmitted by specific arthropods (aphids, leafhoppers, thrips or mites) in which they also replicate, and many of these viruses share similarities in particle morphology, genome organization and fundemental replication strategies to their animal/human-infecting counterparts within the same families [3–7].
Generation of an infectious virus from a cDNA copy of the viral genome, an approach referred to as reverse genetics, is the most powerful genetic tool in modern virology. Unlike positive-strand RNA viruses, whose genomic RNAs (gRNAs) are infectious upon introduction into permissive host cells, neither the naked gRNAs nor the antigenomic RNAs (agRNAs) of NSR viruses are able to initiate infection process when present alone. Instead, replication initiation of these groups of viruses requires de novo viral mRNA synthesis from the viral nucleocapsid (NC) which consists of the viral gRNA and the NC core proteins [8,9]. Therefore, the minimal infectious units of NSR viruses are the viral NCs and generating infectious NCs for reverse genetic studies initially was a major challenge due to difficulties in in vivo reconstitution of functional NCs containing recombinant RNAs. Hence, nearly a decade passed after development of positive-strand virus reverse genetics systems before the first NSR reverse genetics applications were achieved with animal rhabdoviruses [10–12]. These successes involved an entirely different approach from that used to engineer positive-strand RNA viruses, and consisted of transforming cell lines expressing bacteriophage T7 polymerase with transcription plasmids encoding the core nucleocapsid proteins and exact copies of the agRNAs. Under these conditions, viable nucleocapsids were assembled in vivo, leading to replication of recombinant viruses in single cells, followed by invasion of surrounding cells to produce plaques that could be identified visually [10–12]. Notably, a key strategy leading to success was to express viral agRNAs rather than gRNAs, and it was thought that this circumvented hybridization of gRNAs and core protein mRNA transcripts to form double-stranded RNAs that could interfere with the template activities of the RNAs and trigger potent antiviral responses [13,14].
Since the initial rhabdovirus reverse genetics breakthroughs, related strategies have been developed for all families of animal NSR viruses, using either T7 polymerase or endogenous RNA polymerase I to direct intracellular transcription of exact copies of viral RNAs [15–20]. These accomplishments have permitted refined analyses of virus biology and pathology, construction of vectors capable of stable expression of foreign proteins, and attenuated recombinant virus vaccines [8,20–23]. Unfortunately, the inherent low efficiency of NSR virus rescue, coupled with several technical obstacles associated with plants, has hampered adaption of reverse genetics systems developed for animal/human NSR viruses to their plant counterparts during the past two decades. These problems include unavailability of plant or insect vector cell cultures suitable for virus replication and plaque formation, lack of T7 polymerase expression systems and poorly defined RNA polymerase I promoters in plants, as well as interference of the rigid plant cell wall with delivery of the multiple plasmids needed for NC reconstitution. Thus, the lack of reverse genetic systems for plant NSR viruses represents a critical technological gap that has severely hindered our understanding of plant NSR virus infection cycles and pathogenesis.
Plant rhabdoviruses are separated into the Cytorhabdovirus or Nucleorhabdovirus genera based on their cytoplasmic or nuclear sites of replication and morphogenesis, and all members have nonsegmented NSR genomes with a similar structural protein gene organization to those of animal rhabdoviruses [7]. Common elements of all rhabdovirus agRNA genomes consist of 5′ leader (le) and 3′ trailer (tr) sequences flanking five viral structural protein genes that are separated by gene junction sequences. Generally, the gene junction sequences are highly conserved within each virus and are moderately conserved amongst different rhabdoviruses. Three essential cis-elements are embedded in the gene junction sequences, i.e. the Gene-End elements that signal transcription termination and polyadenylation of upstream mRNAs, the Gene-Start elements for initiation of downstream mRNAs transcription and a non-transcribed intergenic region located between the Gene-End and Gene-Start elements. The five common rhabdovirus structural proteins consist of the nucleoprotein (N), phosphoprotein (P), matrix protein (M), glycoprotein (G), and the large RNA polymerase (L), organized in the order 5′-N-P-M-G-L-3′ on the agRNA. However, many rhabdovirus genera encode various additional accessory genes interspersed between the N and L genes [9]. The plant rhabdoviruses differ from their animal rhabdovirus counterparts by encoding one or more accessory movement proteins (MP), at least one of which is thought to be required for cell-to-cell movement [7]. Sonchus yellow net virus (SYNV), the most extensively studied Nucleorhabdovirus, encodes five structural proteins, plus sc4, a putative MP, in the order ‘N-P-sc4-M-G-L’. The sc4 protein is present in infected tissue, but does not form a major component of purified virus preparations [24]. During replication, the N, P and L core proteins assemble with viral gRNA or agRNA to form NCs that function in viral replication and transcription in the nuclei of SYNV-infected cells. As replication proceeds, the nuclei of SYNV-infected tissues become greatly enlarged and develop nuclear viroplasms [7,25]. During morphogenesis, the NCs presumably are coiled by the M protein to form bullet-shaped cores that bud through the inner nuclear envelopes to acquire host membrane lipids and viral glycoprotein spikes and accumulate as bullet-shape or bacilliform particles in perinuclear spaces [7,25]. Unfortunately, due to the lack of a reverse genetics system, none of the processes involved in replication, morphogenesis and cell-to-cell movement are well understood.
In this investigation, we describe for the first time the production of a recombinant plant NSR virus directly from cloned cDNAs. This system relies on co-infiltration of Nicotiana benthamiana leaves with Agrobacterium tumefaciens strains containing plasmids encoding the SYNV agRNA, the N, P and L core proteins, and viral suppressors of RNA silencing (VSRs). We have also engineered a reporter virus that can express green fluorescent protein (GFP) stably during several plant-to-plant passages. Deletion analyses with recombinant SYNV (rSYNV) have provided key insights into SYNV movement and morphogenesis. The establishment of SYNV reverse genetics provides a template for development of similar systems for other plant NSR viruses and will permit fundamental questions in plant NSR virus biology to be studied.
To engineer rSYNV cDNA clones, the full-length SYNV gRNA (13.7-kilobase) was amplified by reverse transcription-PCR (RT-PCR), and the cDNAs were inserted into an Agrobacterium binary expression vector to produce pSYNV for transcription of agRNAs in agroinfiltrated leaves (Fig 1A). The SYNV cDNA was positioned between a truncated cauliflower mosaic virus (CaMV) double 35S promoter (2X35S) and a self-cleaving hepatitis delta virus (HDV) ribozyme sequence to ensure synthesis of SYNV agRNA transcripts with exact 5′- and 3′-ends (Fig 1A). To reconstitute infectious NCs in vivo, a mixture of Agrobacterium cultures harboring the pSYNV, and the pGD-N, pGD-P and pGD-L supporting plasmids that encode the N, P and L core proteins needed for NC formation with the SYNV agRNA, were co-infiltrated into N. benthamiana leaves. The mixture also contained Agrobacteria harboring the tomato bushy stunt virus (TBSV) p19, barley stripe mosaic virus (BSMV) γb and tobacco etch virus (TEV) P1/HC-Pro VSRs to minimize host RNA silencing responses [26,27], as this strategy has proven to be successful for in vivo reconstitution of an SYNV-derived minireplicon (MR) [26]. Approximately 20 days post infiltration (dpi), a small percentage of the infiltrated plants (∼5%) developed typical systemic SYNV symptoms such as stunting, leaf cupping and vein clearing (Fig 1B and Table 1). Immunoblotting with antibodies raised against SYNV virions revealed comparable amounts of the G, N, M and P proteins in the rSYNV- and wild-type SYNV (wtSYNV)-infected tissues (Fig 1C). As predicted from our previous SYNV MR experiments showing that fluorescent reporter expression requires the core proteins and is greatly enhanced by co-expression of VSR proteins [26], plants agroinfiltrated with mixtures lacking any of the core protein plasmids failed to develop symptoms, as was also the case with 145 plants infiltrated with mixtures lacking the VSR plasmids (Table 1).
Mechanical transmission assays showed that rSYNV is highly sap-transmissible (up to 100% of inoculated plants), and that the rSYNV and wtSYNV strains elicited indistinguishable disease symptoms on upper emerging leaves starting from ~13 dpi (S1A Fig). Moreover, as observed in previous studies [7,25], transmission electron microscopy revealed similar cytopathological effects in wtSYNV- and rSYNV-infected cells, which contained large numbers of bacilliform particles in perinuclear spaces around the periphery of the nuclei (S1C Fig).
During plasmid construction, we observed that the pSYNV cDNA contained a mutation at nucleotide (nt) 13,592 in the L gene sequence that changed a Lys codon (AAA) in the wtSYNV strain to an Arg (AGA) codon and created a BsmBI restriction site that could be used as a genetic marker for rSYNV. Therefore, to verify that the agroinoculated plants contained rSYNV rather than wtSYNV contaminants, RNA was extracted and an ~1,500 nt cDNA encompassing the mutant sequence was amplified by RT-PCR and digested with BsmBI. As expected, the cDNAs from rSYNV-infected tissue produced ~500 and 1,000 nt bands, but the wtSYNV cDNA was not digested, whereas a control digestion at an adjacent ApaI site provided identical digestion patterns with both cDNAs (Fig 1D). Moreover, the BsmBI restriction site mutation was stably maintained in the progeny genomes of rSYNV after mechanical transmission with leaf sap extracted from agroinfected plants (S1B Fig). These results demonstrate conclusively that rSYNV was derived from the cloned plasmids.
Rescue of recombinant NSR viruses from cDNA is generally inefficient because multiple plasmids must be simultaneously introduced into single cells to reconstitute infectious virus [13,14,18–20]. Therefore, we sought to improve the recovery of rSYNV by reducing the numbers of plasmids delivered by agroinfiltration. We cloned the N, P and L gene expression cassettes into the pGD vector to develop a single multi-expression plasmid designated pGD-NPL (See S1 Protocols for cloning details). This strategy reduces the Agrobacterium strains required for expression of the N, P and L proteins from three to one, while also ensuring simultaneous expression of the three core proteins in a given cell. When the pGD-NPL bacterial culture was substituted for the mixture containing the pGD-N, pGD-P and pGD-L plasmids (N+P+L mixture), along with Agrobacterium strains harboring pSYNV and the VSRs, the proportion of infected plants increased more than two-fold compared with the N+P+L mixture, and resulted in ~12% of the agroinfiltrated plants developing systemic infections (Table 1).
The relative ratios of the supporting N, P and L proteins are also important for recovery of recombinant NSR viruses [18]. Therefore, we used our previously developed SYNV MR fluorescent reporter expression assay [26] to determine recovery conditions that might lead to increased efficiency of rSYNV generation. In this assay, we constructed a SYNV MR derivative, which contains GFP and Red fluorescent protein (RFP) reporter genes substituted for the SYNV N and P ORFs, respectively, and the flanking 5′ le and 3′ tr sequences (Fig 2A) [26]. The MR plasmid, pSYNV-MR-GFP-RFP, when co-delivered with pGD-NPL and the VSR plasmids via agroinfiltration, exhibited intense GFP and RFP fluorescent foci throughout infiltrated regions (Fig 2B and S2 Fig). To optimize the ratio of core protein expression, different concentrations of Agrobacterium cultures harboring the pGD-N, pGD-P or pGD-L plasmids were added to the pGD-NPL culture. These mixtures were co-infiltrated into N. benthamiana leaves, and the appearance of GFP and RFP fluorescent foci was observed by fluorescence microscopy (Fig 2B and S2 Fig). To our surprise, addition of extra amounts of Agrobacterium containing the N or P plasmids greatly reduced the expression of the GFP and RFP reporter genes, whereas supplying additional L plasmid (NPL+L mixture) increased the MR reporter foci in the infiltrated leaves (Fig 2B and S2 Fig). Note that the increasing amounts of N protein expression appeared to result in a strong reduction in the strength of reporter expressions, whereas additional P protein expression drastically reduced the numbers of fluorescent foci (Fig 2B and S2 Fig, N and P panels). These MR experiments suggested that higher rSYNV recoveries might be obtained if L protein expression was increased, and this proved to be the case when we added an extra volume of bacteria harboring the pGD-L plasmid to the NPL mixture. This mixture (NPL+L) led to systemic symptoms in ∼ 26% of the infiltrated plants, compared with ~ 12% with the NPL mixture, and ~ 5% with the N+P+L mixture (Table 1). Thus, reducing the numbers of supporting plasmids while increasing the abundance of the L plasmid dramatically improved rSYNV recovery.
To develop an rSYNV vector for foreign gene expression in plants, a duplicated N/P gene junction sequence along with the GFP coding sequence was inserted into the pSYNV plasmid between the N and P genes to generate pSYNV-GFP (Fig 3A). In this configuration, GFP mRNA synthesis is initiated immediately after termination of the upstream N protein mRNA synthesis by the duplicated N/P gene junction, and is followed by P mRNA synthesis that is directed by the native N/P gene junction. The pSYNV-GFP plasmid was agroinfiltrated into N. benthamiana leaves along with the bacterial mixture harboring the NPL+L and VSR plasmids. At about 6 dpi, GFP foci began to appear in single cells randomly distributed throughout the infiltrated tissue, and by 9 dpi fluorescence of these cells became more intense and faint fluorescence began to appear in surrounding adjacent cells (Fig 3B). By 12 dpi, fluorescence of the neighboring cells was clearly evident, and more extensive tissue fluorescence was obvious by 15 dpi, and in some cases, fluorescence was evident in isolated leaf veins (Fig 3B, Note white arrow). By ∼20 dpi, the newly emerging leaves of some infiltrated plants exhibited tight curling and yellow net symptoms typical of SYNV infections, and the rSYNV-GFP and rSYNV viruses appeared to be equally infectious based on the appearance of systemic symptoms in inoculated plants (Table 1). When monitored under long wavelength ultraviolent (UV) light, the symptoms in recombinant SYNV-GFP (rSYNV-GFP)-infected leaves were accompanied by strong GFP fluorescence (Fig 3C). Western blot analyses also revealed similar levels of the G, N, M and P proteins in uninoculated upper leaves systemically infected by rSYNV and rSYNV-GFP, and showed that GFP expression was abundant in rSYNV-GFP-infected leaves, but was absent in rSYNV-infected leaves (Fig 3D).
With rare exceptions, plant positive-strand RNA vectors are unable to maintain foreign genes stably during plant-to-plant transfers [28,29]. To investigate the stability of rSYNV-GFP, healthy plants were mechanically inoculated with rSYNV-GFP sap preparations and GFP expression was monitored under UV light and by Western blot analysis. Intense GFP fluorescence could be detected for at least five rSYNV-GFP serial passages (Fig 3E) and GFP protein was expressed at similar levels in all plant passages (Fig 3F). RT-PCR analysis with GFP specific primers also indicated that the GFP insert was stably maintained in the progeny virus genomes (Fig 3F). These results demonstrate that rSYNV can be engineered for stable expression of foreign genes.
Having tagged rSYNV with the GFP reporter, we carried out experiments to investigate the requirements of the sc4, M and G genes for SYNV cell-to-cell movement because these genes appear not to be required for virus replication [7,26]. To knockout the sc4, M or G genes in rSYNV-GFP, the entire transcription unit of a given gene was deleted (Fig 4A), beginning with the upstream transcription start site through the entire ORF and the downstream transcription termination sequence [7]. Bacteria harboring the sc4 (rSYNV-GFP-Δsc4), M (rSYNV-GFP-ΔM) and G (rSYNV-GFP-ΔG) deletion mutants were each agroinfiltrated into N. benthamiana leaves along with the NPL+L and the VSR bacterial mixture, and their movement patterns were compared with those of rSYNV-GFP. In rSYNV-GFP infiltrated leaves, discrete GFP fluorescent foci were first seen in single cells at about 6 to 8 dpi and spread into adjacent cells extensively by 14 dpi (Fig 4B). However, the rSYNV-GFP-Δsc4 foci were restricted to single cells at both 8 and 14 dpi (Fig 4B). These data demonstrate that the sc4 protein has an essential role in viral cell-to-cell movement and possesses the characteristics of a virus MP. In contrast, both the rSYNV-GFP-ΔM and rSYNV-GFP-ΔG mutants were capable of local movement, albeit less efficiently than rSYNV-GFP (Fig 4B). In addition, an M and G double mutant, rSYNV-GFP-ΔMG, was still able to move from cell-to-cell, although the rates of movement appeared to be lower than those of the rSYNV-GFP-ΔM and rSYNV-GFP-ΔG mutants (Fig 4B).
Because mRNA transcription of nonsegmented NSR viruses progressively attenuates at each gene junction site [30], deletion of a given transcription unit may lead to alteration of viral mRNA ratios that could result in virus attenuation [31]. Therefore, we generated a second set of mutants, in which the RFP gene was substituted for the sc4, M or G genes respectively (S3 Fig), and assessed these mutants for their localized movement abilities. Again, only the sc4 substitution mutant (rSYNV-GFP-Δsc4:RFP) was unable to initiate cell-to-cell movement (S3B Fig). Thus, these two sets of data collectively show that both the M and G proteins are dispensable for cell-to-cell movement and suggest that uncoiled NCs are able to function in localized movement.
To further confirm that the sc4 protein, but not the RNA sequence is required for local movement, we tested whether or not the sc4 protein expressed in trans can complement rSYNV-GFP-Δsc4 cell-to-cell movement. Because SYNV cell-to-cell movement becomes evident only after about 9 dpi (Figs 3B and 4), to synchronize sc4 expression with critical steps in SYNV movement, we took advantage of SYNV MR-directed expression of proteins, which has been shown to persist for up to 20 days [26]. To this end, we constructed SYNV MR-sc4-RFP (Fig 4A), which substitutes the sc4 ORF for the N ORF and the RFP ORF for the P ORF (See S1 Protocols for cloning details), and evaluated whether persistent expression of sc4 could facilitate rSYNV-GFP-Δsc4 movement in trans. As shown in Fig 4C, the MR-sc4-RFP infiltrated regions exhibited RFP fluorescence in some cells at 8 dpi, which provided a marker for MR-mediated gene expression. In some instances, GFP fluorescence produced by rSYNV-GFP-Δsc4 was observed in single cells that also showed RFP fluorescence (Fig 4C, left panels), indicating that those cells had received all plasmids necessary for MR-sc4-RFP and rSYNV- GFP-Δsc4 reconstitution and expression. In these cases, GFP fluorescence continued to increase in cells and spread beyond the RFP fluorescence by 14 dpi, which mostly remained confined to single cells or to a very limited number of cells (Fig 4C, right panels). These results suggest that MR-sc4-RFP did not invade adjacent cells extensively, but that the sc4 protein produced in trans was able to expedite limited cell-to-cell transit of the rSYNV-GFP-Δsc4 reporter virus.
To determine the viral proteins involved in systemic infection, the agroinfiltrated plants shown in Fig 4 were monitored for appearance of symptoms and GFP fluorescence in upper uninoculated leaves. The rSYNV-GFP-ΔG mutant was able to develop systemic infections and induced symptoms similar to rSYNV-GFP (Fig 5A), although the proportion of systemically infected plants was drastically reduced compared with the rSYNV-GFP inoculations (∼3% systemically infected plants for rSYNV-GFP-ΔG as compared to ∼22% for rSYNV-GFP infiltrations, Table 1). In rSYNV-GFP-ΔG infections, fluorescence first appeared in the upper leaves at ∼25 dpi and began to spread from the leaf veins, but was mostly confined to the upper leaf veins even at 35 dpi (Fig 5A, lower panels). In contrast, the rSYNV-GFP-infected upper leaves exhibited GFP fluorescence throughout the mesophyll tissues by 30 dpi (Fig 5A, upper panels). Nevertheless, Western blot analyses revealed only moderately reduced abundances of the SYNV N, M, and P proteins and GFP in the systemically infected leaves when compared with those of the rSYNV-GFP infections, and confirmed the absence of G protein in the systemic leaves infected by rSYNV-GFP-ΔG (Fig 5B). These results show that the SYNV G protein is not required for symptom development and systemic movement, but may facilitate infection by unknown mechanism(s). As expected from its inability in cell-to-cell movement, the rSYNV- GFP-Δsc4 mutant was defective in systemic infection (Table 1). Interestingly, although rSYNV-GFP-ΔM was capable of localized movement in the infiltrated leaves (Fig 4B), the mutant was unable to invade upper leaves as judged by visual inspection and RT-PCR analysis (Table 1). Hence, the M protein appears to be required for SYNV movement from primary infection foci into vasculature.
To investigate the roles of the G protein in SYNV morphogenesis and cytopathology, upper leaf tissues infected with rSYNV-GFP and rSYNV-GFP-ΔG were compared by transmission electron microscopy. As with previous studies of SYNV-infected plants and protoplasts [7,25], enlarged nuclei of rSYNV-GFP-infected cells contained large numbers of intact bacilliform particles (71.0 ± 3.4 nm diameter; n = 26) surrounded by invaginated inner nuclear envelopes (Fig 5C, as indicated by red arrows). Smaller nonenveloped bullet-shaped aggregates (53.0 ± 2.2 nm diameter; n = 26) were also observed in electron dense regions characteristic of the subnuclear viroplasms (Fig 5C, left panel). The larger enveloped particles have the appearance of mature enveloped virions, whereas the smaller nonenveloped particles appear to represent naked cores that have not yet completed morphogenesis [25]. In marked contrast, rSYNV-GFP-ΔG-infected cells contained only nonenveloped particles (54.3 ± 2.2 nm in diameter, n = 12) that were randomly distributed or occurred as orderly aligned arrays within the viroplasms (Fig 5D). Our data thus demonstrate that the G protein is required for morphogenesis of enveloped SYNV particles, and that in the absence of the G protein large numbers of NC cores accumulate in or near the viroplasms.
Recovery of infectious NSR viruses from cloned cDNAs for reverse genetic analyses is now routine for all animal NSR virus families. Although the procedures are quite inefficient, with 104 to 107 transfected cells per primary infected cell, recombinant virus particles released from primary infected cells can be passaged to permissive cell lines to obtain progeny viruses suitable for a variety of purposes [13,14,18–20]. Unfortunately, only a few insect vector cell cultures suitable for rescue of recombinant plant NSR viruses have been established [32,33]. Even these lines are difficult to maintain and to our knowledge, plasmids suitable for transient expression of multiple genes in these lines are unavailable. Moreover, introduction of multiple components into single plant cells after removal of the cell wall is inefficient and protoplast recoveries after transformation or viral transfection is low. In addition, transformation and high level expression of multiple viral proteins and RNAs in plant leaves is difficult due to the presence of the cell wall and the existence of potent plant antiviral gene silencing mechanisms [34,35]. Hence, to circumvent these problems, we turned to infiltration of N. benthamiana leaves with Agrobacterium strains harboring plasmids encoding the SYNV agRNAs and the N, P and L core proteins needed for de novo NC assembly, coupled with the use of VSRs proteins to suppress host RNA silencing. This approach has enabled in planta rescue of rSYNV from cDNAs with an infection phenotype identical to wtSYNV (Fig 1 and S1 Fig).
The recovery of rSYNV from agroinfiltrated plants was initially inefficient, as only ~5% of the agroinfiltrated plants developed systemic symptoms, but we were able to improve rSYNV recovery to ~26% of the infiltrated plants by optimizing the infection mixture components (Table 1). It is worth noting that the SYNV MR reporter system that we developed earlier [26] was invaluable in devising steps to rescue full-length rSYNV and to improve recovery efficiency. The MR derivatives provided a rapid assay to determine the functionality of agRNA derivatives and the optimum conditions for expression of the SYNV core components that could be applied to improve the efficiency of rSYNV recoveries in agroinfiltrated leaves (Fig 2). Similar MR systems have also proven to be very helpful for recovery of animal NSR viruses [8, 20], and we believe that time invested to develop MR derivatives will be very worthwhile in future studies to develop and optimize engineering of recombinant plant NSR viruses.
Although the general principles used for recovery of rSYNV are similar to those used for recoveries of NSR animal viruses, our study reveals several distinct aspects that merit consideration when developing strategies for generation of other recombinant plant NSR viruses. First, co-expression of VSRs to suppress potent RNA silencing response in plants proved to be extremely important for generation of rSYNV (Table 1). Similar strategies have previously been shown to be important for cDNA recoveries of complex plant positive-strand RNA viruses [36,37]. It is known that host gene silencing mechanisms generally reduce Agrobacterium transient gene expression [34], and that co-expression of VSRs can alleviate this limitation [35]. These VSR proteins most likely facilitate rSYNV recovery by reducing degradation of SYNV mRNAs and agRNA transcripts to permit high levels of N, P and L protein expression needed for efficient NC generation. In addition, VSR proteins may also prevent host antiviral RNA silencing machineries from degrading rescued virus and promote efficient virus spread during the initial stages of infection [38,39]. Second, our studies revealed a requirement of higher amounts of the L plasmid relative to the N and P plasmids for efficient SYNV recovery. The requirement for higher levels of L protein is counter-intuitive, because the abundance of the N and P proteins is much higher than the L proteins in nonsegmented NSR nucleocapsids [8,9]. Moreover, studies with NSR animal viruses have usually shown that high levels of N protein expression are correlated with more efficient rescues [18]. However, the relative molar ratios of N, P and L protein expressed in the agroinfiltrated leaves were not determined in our study due to low titers of L protein antibody, so it is possible that transient expression of the L protein (~242 KDa) is less efficient than those of the smaller N and P proteins (54 and 34 KDa, respectively). Third, unlike most NSR animal virus rescue systems, in which phage T7 RNA polymerase or host RNA polymerase I were used to direct intracellular transcription of viral agRNAs [18–20], we used a CaMV 35S promoter, which relies on the endogenous RNA polymerase II machinery, to drive the SYNV agRNA and core protein expression. The 35S promoter was truncated to permit exact 5′ initiation of SYNV agRNA transcription, and this strategy circumvents the use of hammerhead ribozyme cleavage to generate authentic 5′ ends of agRNA transcripts [26]. Since the 35S promoter is known to function in various dicots and monocots species, it is likely that the plasmid-based system developed in the present study will be applicable to other families of plant NSR viruses.
We have engineered an SYNV reporter virus (rSYNV-GFP) that expresses GFP stably even after repeated mechanical passages (Fig 3). We have used this reporter virus to provide a simple visual assay to follow local and systemic spread of rSYNV, and to investigate the genetic requirements for SYNV movement. As with other plant viruses, NSR viruses must move from initially infected cells to neighboring cells through MP-gated plasmodesmata [3]. Studies based mainly on positive-strand RNA viruses have proposed two major mechanisms whereby plant viruses move from cell-to-cell. These mechanisms involve either direct interactions of MPs with either viral genomes or with intact virions, which then move through MP modified plasmodesmata [40,41]. However, direct evidence has not previously been available as to the nature of the plant NSR virus infectious entities that navigate intercellular connections. In the case of plant rhabdoviruses, previous indirect studies have suggested that sc4, and similar plant rhabdovirus homologs exhibit MP properties [7,42–44]. The failure of the rSYNV-GFP-Δsc4 mutant to move from cell-to-cell (Fig 4 and S3 Fig) now provides the first direct evidence that sc4 is the SYNV MP and supports a movement function for sc4 homologs of other plant rhabdoviruses. However, in contrast to the sc4 deletion mutant, the M and G single or double deletion mutants are capable of localized movement (Fig 4 and S3 Fig). These findings argue against previous speculations that the M protein functions as an essential component of cell-to-cell movement complexes or that mature virions move through ER tubules and desmotubules into adjacent cells [45,46]. Rather, our data suggest a model whereby a portion of the NCs in or adjacent to the viroplasms interact with the sc4 protein and are exported from the nucleus. Consistent with this model, the putative MPs of several plant NSR viruses, i.e. the P3 protein of rice yellow stunt nucleorhabdovirus and the NSm proteins of several tospoviruses in the Bunyaviridae family, have been shown to bind directly to the their cognate N proteins [42,47–49], although similar interactions have not been reported for the SYNV sc4 and N proteins.
Mature virions of enveloped NSR viruses are formed by a budding process, during which the M proteins function to condense and coil the NCs into cores that then bud through host membranes to acquire phospholipid envelopes and the glycoprotein spikes [50,51]. The budding process is essential for release of animal viruses from infected cells, while the surface glycoproteins have critical roles during cellular entry [50,52]. Although such processes presumably occur during insect vector infections of plant enveloped NSR viruses, the functions of the glycoprotein during plant host infections have remained obscure. Our deletion analyses now show that the rSYNV-GFP-ΔG mutant is able to cause systemic infection in plants and induce typical symptoms (Fig 5A). However, the ΔG mutant failed to undergo morphogenesis, resulting in large numbers of naked cores accumulating in the viroplasms (Fig 5D). Such cytopathic structures are reminiscent of the striking arrays of cores present in the nuclei of SYNV-infected protoplasts treated with tunicamycin [25], an inhibitor that blocks G protein N-glycosylation [53]. These data are also consistent with previous findings by Sin et al [54], who have used a reassortment-based forward genetics approach to map the insect transmissibility determinant of tomato spotted wilt virus (TSWV), a tripartite NSR virus in the Bunyaviridae family. Sin et al. have shown that several TSWV single-lesion isolates with mutations in the glycoprotein precursor are defective in virion assembly and insect transmissibility, but are able to infect plants. Hence, the glycoproteins and the mature virions of two distinct plant NSR virus families are dispensable for systemic infection of plants. Interestingly, the SYNV M deletion mutant was unable to invade upper leaves (Table 1), although its local movement appeared to be as efficient as the rSYNV-GFP-ΔG (Fig 4B). These results suggest that the M protein may play a role in long-distance movement, perhaps by promoting efficient NC coiling [7,51] and/or NC entry into the vasculature.
In conclusion, we have developed a plasmid-based reverse genetics system for recovery of rSYNV directly in agroinfiltrated plants. This achievement permits investigation of fundamental aspects of plant rhabdovirus biology and pathology that were technically unapproachable previously. We anticipate that similar approaches can be applied to other plant NSR viruses for refined studies of plant infections and insect interactions. Given the exceptional stability of the GFP protein during plant-to-plant transmission of rSYNV-GFP (Fig 3E and 3F), and the widespread use of recombinant animal rhabdoviruses as vectors for expression of antigens and delivery of therapeutic genes [22,23,55], rSYNV and other plant NSR viruses also hold great promise for biotechnological applications.
Total RNA extracted from SYNV-infected N. benthamiana plants was used for RT-PCR amplification of the full-length cDNA of agRNA with a high-fidelity KOD-Plus-Neo DNA polymerase (Toyobo, Osaka, Japan) and the forward 5′-tttcatttggagaggAGAGACAGAAACTCAGAAAATACAAT-3′ and reverse 5′-atgccatgccgacccAGAGACAAAAGCTCAGAACAATCCCTAT-3′ primers. The forward primer contains 15-nt overhangs (lowercase letters) complementary to the 3′ end of the 35S promoter, whereas the reverse primer contains 15-nt overhangs (lowercase letters) complementary to the 5′ end of the HDV ribozyme, respectively. To generate a binary plasmid for intracellular transcription of SYNV agRNA after agroinfiltration, the full-length cDNA was cloned into a modified pCB301-2X35S-Nos plasmid [56]. The pCB301-2X35S-Nos plasmid was linearized by StuI and SmaI double digestion, which removes the sequence between the transcription initiation site of 35S promoter and the multiple cloning sites immediately before the HDV ribozyme sequence. The linearized plasmid was recovered and assembled precisely with the SYNV cDNA by using an In-Fusion HD PCR Cloning Kit (Clontech, Japan). The resulting plasmid (pSYNV) was sequenced to confirm its authenticity and correct assembly, and was expected to transcribe the full-length SYNV antigenome with the exact 5′ end by 35S promoter and the exact 3′ end processed by the HDV ribozyme cleavages. Supporting pGD plasmids [57] directing expression of the SYNV N, P, and L proteins, as well as the BSMV γb, TBSV p19, and TEV P1/HC-Pro have been described previously [26]. The constructions of other plasmid derivatives used in this study are detailed in the S1 Protocols.
Recombinant binary plasmids were electroporated into Agrobacterium tumefaciens strain GV3101 and N. benthamiana agroinfiltrations were performed essentially as described [26]. Bacterial cell suspensions were activated by acetosyringone, adjusted to an optical density (OD) A600 of 0.8 and incubated for 2 to 4 hr at room temperature. Immediately before infiltration, equal volumes of Agrobacterium cultures harboring the pGD-N, pGD-P, pGD-L (N+P+L mixture), pSYNV (or the rSYNV deletion derivatives), were mixed with one volume of bacterial mixture containing the BSMV γb, TBSV p19, and TEV P1/HC-Pro plasmids, unless otherwise stated. In the case of the NPL mixture, one volume of pGD-NPL culture at 0.8 OD was used to substitute for the N+P+L mixture. For the NPL+L mixture, equal volumes of the NPL and the L Agrobacterium cultures were mixed at 0.8 OD.
Mechanical transmissions were carried out as previously described [58]. Briefly, 2 g of young leaves from plants infected for approximately 20 days were ground in a chilled mortar containing 5 mL of freshly prepared cold (~4°C) inoculation buffer (5% sodium sulfite and 2% Celite). Two leaves of N. benthamiana plants at the 4~5 leaf-stage were gently rubbed by hand with the leaf extracts. The inoculated plants were placed in an insect-free growth chamber at ~25°C and 60% relative humidity under a 16 h light/8 h dark photoperiod.
Total RNAs were extracted from upper leaves of wtSYNV- and rSYNV-infected plants with Trizol reagent (Invitrogen). RT-PCR was carried out using AMV Reverse Transcriptase XL (Takara, Japan) and Phusion High-Fidelity DNA Polymerase (NEB, Beverly, MA) with SYNV-specific primers SYNV/11008/F and SYNV/12503/R (S1 Table). The PCR products were digested with BsmBI or ApaI in a 20-μl reaction mixture and separated in 1.5% agarose gels. PCR products amplified from sap-inoculated plants were also sequenced to confirm stable maintenance of the genetic tag.
Total proteins were extracted from healthy or infected N. benthamiana leaves and evaluated by Western blotting. Proteins separated by SDS-PAGE were either stained with Coomassie blue or transferred to nitrocellulose membranes and probed with polyclonal antiserum specific to the disrupted SYNV virions [59], the SYNV G [60], or monoclonal antibodies against GFP and RFP (Abcam, Cambridge, UK).
N. benthamiana leaves were examined with a Zeiss SteREO Lumar. V12 epifluorescence microscope with filter set Lumar 38 (excitation 470/40; emission 525/50) and Lumar 31 (excitation 565/30; emission 620/60) for GFP and RFP, respectively. The data were processed with LSM software Zen 2009 (Carl Zeiss).
Systemically infected leaf tissues were fixed in 2.5% glutaraldehyde and 1% osmium tetroxide in 100 mM phosphate buffer (pH 7.0) essentially as described by Kong et al [61] and then embedded in Epon 812 resin as described by the manufacturer (SPI-EM, Division of Structure Probe, West Chester, USA). Ultrathin sections (70 nm) were mounted on formvar-coated grids, and then stained with uranyl acetate for 10 min followed by lead citrate for 10 min. The stained sections were examined under a transmission electron microscope (TEM; H-7650, Hitachi, Japan) at 80 kV accelerating voltage.
|
10.1371/journal.pntd.0005499 | Comparison of O-polysaccharide and hemolysin co-regulated protein as target antigens for serodiagnosis of melioidosis | Melioidosis is a severe disease caused by Burkholderia pseudomallei. Clinical manifestations are diverse and acute infections require immediate treatment with effective antibiotics. While culture is the current diagnostic standard, it is time-consuming and has low sensitivity. In endemic areas, inaccessibility to biosafety level 3 facilities and a lack of good serodiagnostic tools can impede diagnosis and disease surveillance. Recent studies have suggested that O-polysaccharide (OPS) and hemolysin co-regulated protein 1 (Hcp1) are promising target antigens for serodiagnosis of melioidosis.
We evaluated rapid ELISAs using crude antigens, purified OPS and Hcp1 to measure antibody levels in three sets of sera: (i) 419 serum samples from melioidosis patients, Thai and U.S. healthy donors, (ii) 120 serum samples from patients with other bacterial infections, and (iii) 423 serum samples from 200 melioidosis patients obtained upon admission and at 12 and 52 weeks post-recovery. We observed significantly higher antibody levels using the crude antigen prepared from wild type B. pseudomallei K96243 compared to that of an OPS-mutant. The areas under receiver operator characteristics (AUROCCs) for diagnosis were compared for individual Hcp1-ELISA or OPS-ELISA or combined Hcp1/OPS-ELISA. For Thai donors, AUROCCs were highest and comparable between the Hcp1-ELISA and the combined Hcp1/OPS-ELISA (0.95 versus 0.94). For U.S. donors, the AUROCC was highest for the combined Hcp1/OPS-ELISA (0.96). Significantly higher seropositivity was observed in diabetic patients compared to those without diabetes for both the Hcp1-ELISA (87.3% versus 69.7%) and OPS-ELISA (88.1% versus 60.6%). Although antibody levels for Hcp1 were highest upon admission, the titers declined by week 52 post-recovery.
Hcp1 and OPS are promising candidates for serodiagnosis of melioidosis in different groups of patients. The Hcp1-ELISA performed better than the OPS-ELISA in endemic areas, thus, Hcp1 represents a promising target antigen for the development of POC tests for acute melioidosis.
| Melioidosis, caused by Burkholderia pseudomallei, is a life-threatening infection endemic in tropical countries. Definitive diagnosis of melioidosis relies upon bacterial culture which requires suitable laboratory facilities and reliable antibody testing. To obtain an effective target antigen for use in a simple point-of-care (POC) test, rapid ELISAs using crude B. pseudomallei antigen preparations or purified O-polysaccharide (OPS) and hemolysin co-regulated protein (Hcp1) were compared using serum samples from three large collections obtained from melioidosis patients and patients with other bacterial infections. We detected high levels of antibodies to Hcp1 and OPS in serum from melioidosis patients upon admission and showed that anti-Hcp1 levels declined post-recovery. When serum samples from endemic areas were tested, the performance of the Hcp1-ELISA and combined Hcp1/OPS-ELISA were higher than the OPS-ELISA. When serum from non-endemic areas was tested, the combined Hcp1/OPS-ELISA gave the highest performance. Both the OPS- and Hcp1-based ELISAs were useful for detection of antibodies in various groups of patients including diabetics. Since anti-Hcp1 titers in melioidosis patient serum were higher than anti-OPS titers, Hcp1 is an attractive candidate for further development of a rapid POC test for use in endemic areas.
| Melioidosis is a severe infectious disease caused by the Gram-negative environmental bacterium, Burkholderia pseudomallei. It is an under-recognized tropical disease that is a common cause of community-acquired infections in Southeast Asia and northern Australia. It is recognized that melioidosis is a more significant global public health concern than previously thought, with increasing numbers of cases reported in many countries [1]. A recent report estimated the incidence of melioidosis to be 165,000 cases per year worldwide, with a predicted annual mortality of 89,000 [2]. In Thailand, the estimated incidence rate is 12.7 cases of melioidosis per 100,000 people per year and the mortality rate is 43% [3]. Melioidosis is the third most common cause of death from infectious diseases in northeast region after HIV infection and tuberculosis [3]. Up to 80% of patients with melioidosis have one or more risk factors which include diabetes, alcohol use, renal disease, thalassemia, cancer and glucocorticoid therapy [1, 4]. Among these, diabetes is the most common underlying disease with 60% of melioidosis patients being diabetic [1].
B. pseudomallei is a facultative intracellular pathogen [5] that can invade host cells, escape from phagosomes, survive within the cytosol and spread from cell-to-cell in many organs [6, 7]. These processes are dependent upon virulence-associated type III and type VI secretion systems (T3SS and T6SS) expressed by this pathogen [8, 9]. Lipopolysaccharide (LPS) and capsular polysaccharide (CPS) are additional virulence factors that contribute to the pathogenesis of B. pseudomallei [10]. The clinical manifestations of melioidosis are diverse and can mimic other infections, ranging from skin and soft tissue infections to acute pneumonia and septicemia frequently resulting in misdiagnosis. Treatment of melioidosis requires immediate administration of ceftazidime or carbapenems, which are generally not used as empirical treatment for other bacterial sepsis [1].
Making an early and accurate diagnosis of melioidosis to guide treatment is critical for reducing patient mortality. The diagnosis of melioidosis and subsequent appropriate treatment depends on culture of B. pseudomallei from clinical specimens, or evidence of sepsis in people with a high risk of exposure and predisposing factors (e.g. diabetes) for melioidosis. However, identification of B. pseudomallei by culture is time-consuming (typically 72 hours), has low sensitivity (60%) [11, 12] and requires both experience and stringent laboratory health and safety for this Hazard Group 3 pathogen. Using culture methods, laboratories unfamiliar with B. pseudomallei frequently misidentify the bacterium as an inconsequential environmental Pseudomonas species [13].
An alternative approach to the gold standard of bacterial culture for diagnosis of melioidosis is antigen detection using a monoclonal antibody to B. pseudomallei capsule as a point-of-care (POC) diagnostic lateral flow assay (LFA). Although rapid and low cost, the LFA only achieves 40% sensitivity in blood of culture-positive patients, limiting its diagnostic utility in acute melioidosis [14]. Quantitative real-time polymerase-chain reaction (qPCR) assay of clinical samples may provide a more rapid result than culture, but has a disappointing sensitivity at 61% in northeast Thailand, especially when performed on blood (sensitivity at 25%) [15]. Additional tests such as latex agglutination assays, immunofluorescence assays or matrix-assisted laser desorption ionization-time of flight mass spectrometry (MALDI-TOF MS) are required to accelerate the identification of positive cultures [16–19].
To improve the time for diagnosis of melioidosis, an indirect hemagglutination assay (IHA) is used to determine antibody titers that are indicative of exposure to B. pseudomallei. While rapid compared with bacterial culture, the sensitivity and specificity of the IHA in Thailand are low (69.5% and 67.6% respectively) [20]. We recently developed a simpler O-polysaccharide (OPS)-based latex agglutination assay which shows potential for detecting exposure to B. pseudomallei in individuals from non-endemic areas but lacks specificity in long term residents from endemic regions [20].
A rapid POC serological test with high sensitivity and specificity would be ideal for use in resource-poor areas where melioidosis is endemic. To develop such assays, identification of good serologic markers is critical. It is also important to evaluate whether an assay can differentiate between acute melioidosis and previous infection or exposure. Our recently developed rapid indirect enzyme-linked immunosorbent assay (ELISA) provides a platform for evaluation of different antigen candidates [21]. Among several antigens tested, our studies and others have highlighted the potential of B. pseudomallei OPS and hemolysin co-regulated protein 1 (Hcp1) as targets for further development of serodiagnostic tests for melioidosis [20–24]. Hcp proteins are both structural components and substrates of T6SSs [24], and in B. pseudomallei are known to be expressed in vivo [9, 22, 23].
To identify the best candidate for further development of POC, we used rapid ELISAs to measure antibodies to OPS and Hcp1 using our large collections of serum samples from both endemic and non-endemic areas. The aims of this study were 1) to compare the antibody responses measured by ELISA to OPS and non-OPS antigens in sera from melioidosis patients and healthy donors, 2) to develop a rapid ELISA using Hcp1 as the target antigen for antibody detection and then compare the results of Hcp1-ELISA with the OPS-ELISA, 3) to evaluate the diagnostic potential of Hcp1 and OPS for determination of antibody titers in different groups of melioidosis patients, and 4) to evaluate the dynamics of the antibody responses to OPS and Hcp1 over 12 months in individual melioidosis patients by comparing titers during acute infection with the titers observed at 3 and 12 months post-recovery.
Initially, two sets of anonymous human serum samples were used to evaluate the ELISAs as described previously [21]. The first set included 141 on-admission sera from culture-confirmed B. pseudomallei infected patients who were admitted to Sappasithiprasong hospital, Ubon Ratchathani, northeast Thailand, 188 serum samples obtained from healthy donors from the same area in northeast Thailand and 90 serum samples obtained from healthy U.S. donors (Innovative Research, Novi, MI, USA). The second set was three groups of on-admission anonymous human sera that were used to further evaluate the specificity of the ELISAs. These included the following groups: 1) 20 acid-fast stain positive tuberculosis patients from Chiangrai, north Thailand, 2) 50 culture-proven scrub typhus patients from Udon Thani, northeast Thailand, and 3) 50 culture-proven leptospirosis patients from Udon Thani, northeast Thailand.
To evaluate the diagnostic potential of Hcp1 and OPS antigens for determination of antibody titers in different groups of melioidosis patients, a third set of independent serum samples was used. This set included serum samples obtained from patients with culture-confirmed melioidosis collected a median of 5 days (Interquartile range, IQR 3–6 days, range 2–13 days after admission (N = 200), and at 12 weeks (N = 113) and 52 weeks (N = 110) post-recovery. The patients were recruited in a longitudinal clinical and immunological study at Sappasithiprasong hospital during September 2012-October 2015 [25]. All participants were ≥ 18 years old. All serum samples were stored at -80°C.
The study was approved by Ethics Committee of Faculty of Tropical Medicine, Mahidol University (approval number MUTM 2014–079 and MUTM 2012–018), Sappasitthiprasong hospital (approval number 018/2555), and the Oxford Tropical Research Ethics Committee (reference 64–11). Written informed consent was obtained from the participants enrolled in the study.
Whole-cell (WC) antigen was extracted from the wild type strain B. pseudomallei K96243 (from a Thai patient in northeast Thailand; expresses type A OPS) and an OPS mutant (ΔwbiD K96243) by heating at 80°C for 1 h. The supernatant was used as the antigen described previously [21, 26]. The OPS mutant defective in wbiD (BPSL2677) was constructed as described in our previous study [27].
B. pseudomallei LPS type A was extracted from the select agent excluded strain RR2808 (capsule mutant) using a modified hot phenol method [28, 29]. Purified OPS antigen was then obtained using acid hydrolysis and gel permeation chromatography as previously described [30]. For expression of recombinant Hcp1 (rHcp1) with a N-terminal 6xHis-Tag, the hcp1 ORF (BMAA0742) was PCR amplified from B. mallei ATCC 23344 genomic DNA using the Bmhcp1-6HisF (5’-CCCAACGGTCTCACATGGCGGCGCATCATCATCATCATCATCTGGCCGGAATATATCTCAAGG-3’) and Bmhcp1-R1 (5’-CCCAACGGTCTCAAGCTTCAGCCATTCGTCCAGTTTGCGGC-3’) primer pair; BsaI linkers are underlined. The resulting DNA fragment was digested with BsaI and cloned into pBAD/HisA digested with NcoI/HindIII producing plasmid pBADBmhcp1-6HisF. Notably, B. pseudomallei and B. mallei Hcp1 proteins are 99.4% identical. Recombinant DNA techniques were conducted as previously described [31]. Oligonucleotide primers were obtained from Integrated DNA Technologies. DNA sequencing was performed by ACGT Inc. For purification of rHcp1, E. coli TOP10 (pBADBmhcp1-6HisF) was grown to mid log phase in LB broth and protein expression was induced using 0.02% L-arabinose (Sigma). Bacterial pellets were resuspended in B-PER (Pierce) plus Benzonase (Novagen) and Lysozyme (100 μg/ml) and incubated for 20 min at room temperature with gentle agitation. Insoluble material was removed by centrifugation and the resulting supernatant was loaded onto a gravity-fed Ni-NTA agarose (Invitrogen) column. The column was washed with Wash Buffer (50 mM Tris pH 8.0, 300 mM NaCl and 40 mM Imidazole), protein was eluted with Elution Buffer (50 mM Tris pH 8.0, 50 mM NaCl and 300 mM Imidazole) then dialyzed against PBS and loaded onto a gravity-fed His-Pur Cobalt Resin (Thermo Scientific) column. The column was washed with PBS and rHcp1 was eluted with Wash Buffer, dialyzed against PBS, concentrated and stored at 4°C. Protein concentrations were determined using a BCA protein assay kit (Pierce). Endotoxin removal was performed using High Capacity Endotoxin Removal Resin (Pierce). The amount of endotoxin in the rHcp1 preparations was quantitated using a LAL Chromogenic Endotoxin Quantitation Kit (Pierce).
The ELISAs were performed using these following antigens: 1) WC antigen prepared from wild type B. pseudomallei K96243, 2) WC antigen prepared from an OPS mutant defective in OPS (ΔwbiD K96243) antigen, 3) rHcp1 protein, 4) the purified OPS antigen, and 5) OPS antigen in combination with rHcp1 (Hcp1/OPS). The optimal concentration of coating antigen was determined using pooled melioidosis and pooled healthy sera as previously described [21]. Following evaluation for antigen concentration and serum dilution, the plates were prepared for ELISA using the optimized antigen concentration as follows: WC 0.25 μg/ml, OPS 1 μg/ml, Hcp 2.5 μg/ml and OPS/Hcp1 (0.5 μg/ml OPS/1.25 μg/ml Hcp1). The serum samples used for evaluation of the various ELISAs included culture-confirmed melioidosis patients (N = 141), U.S. healthy donors (N = 90), Thai healthy donors (N = 188), tuberculosis patients (N = 20), scrub typhus patients (N = 50) and leptospirosis patients (N = 50) at a dilution of 1:2,000. All ELISAs were performed using a 1:2,000 dilution of horseradish peroxidase-conjugated rabbit antihuman IgG as previously described [21].
To examine the antibody titers specific to Hcp1 and OPS, ELISAs were performed with a third set of serum samples obtained from acute phase and recovery phase melioidosis patients using undiluted sera and two-fold serial dilution sera at range of 1:125 to 1:2,048,000. The endpoint antibody titer was read at the serum dilution which showed positive OD values of each ELISA. Positive results for individual serum samples were determined using OD cut-off values at specificity of 95%. The antibody titers of individual patients were compared between week 0, week 12 and week 52. Only samples from different time points with two-fold changes in titer were considered as increased or decreased antibody titers.
Statistical analyses were performed using Stata version 12 (StataCorp LP, College Station, TX) and Prism 5 Statistics (GraphPad Software Inc, La Jolla, CA). All data in box plots are presented as 25th and 75th percentile boundaries in the box with the median line within the box; the whiskers indicate the 10th and 90th percentiles. The Mann-Whitney test was used to test the difference of median between different serum groups. Spearman’s rank correlation was used to determine the pairwise correlation coefficient for the pairs of tests [32]. The McNemar test was used to compare the sensitivity between tests. Fisher’s exact test was performed to compare the ELISA results and clinical presentation and outcomes. Differences were considered statistically significant at a p-value < 0.05.
A receiver operator characteristic (ROC) curve was created to monitor the shifting the positive cut off value on true-positive (sensitivity) and false positive (1-specificity) rates. Areas under the ROC curves (AUROCC) were compared using a nonparametric method as previously described by DeLong et al. [33]. The ELISA data of the melioidosis group and Thai donors were evaluated separately from the data of the melioidosis group and U.S. donors using OD cut-off values at specificities of 95%.
Results from our previous study using a crude antigen ELISA (WC-ELISA) to determine the levels of B. pseudomallei-specific antibodies in five individual melioidosis patients, to either wild-type (K96243) or an OPS mutant (K96243ΔwbiD), indicate that OPS appears to be the predominant antigen recognized by human antibodies [21]. In the present study, we expand these experiments to determine the antibody levels in 419 individual sera obtained from melioidosis patients (N = 141), Thai healthy donors (N = 188), U.S. healthy donors (N = 90) using the same WC-ELISA with coating antigens prepared from either the wild type or the OPS mutant (Fig 1). Our results revealed that the median OD value for the melioidosis group was statistically higher compared to Thai donors (P < 0.001 for both ELISAs) and U.S. donors (P < 0.001 for both ELISAs). The median OD value for the melioidosis group was 5.9 times lower in the OPS mutant-WC-ELISA compared to the wild type-WC-ELISA. Similarly, the median OD value for Thai healthy donors was 3 times lower in the OPS mutant-WC-ELISA in comparison to the wild type-WC-ELISA [median OD 0.04 (IQR 0.02–0.08) versus 0.12 (IQR 0.06–0.22); P < 0.001]. In contrast, the median OD value for U.S. healthy donor serum was not significantly different between the two ELISAs [median OD 0.12 (IQR 0.05–0.27) for OPS mutant versus 0.11 (IQR 0.05–0.38) for wild type; P < 0.954]. These findings suggest that OPS is a predominant antigen recognized by antibodies in Thai melioidosis patients and Thai healthy donors sera. This was not the case, however, for sera from U.S. donors. Interestingly, results from the OPS mutant-WC-ELISA also revealed that several melioidosis patients appeared to have high antibody levels to antigens other than OPS.
Recently, our group and others have shown that Hcp1 is a promising candidate serodiagnostic marker for melioidosis [22–24]. B. pseudomallei Hcp1 is a T6SS component that is expressed in vivo or under iron-limiting conditions when the organism is grown in vitro [22, 24]. To assess the serodiagnostic potential of Hcp1, we developed a rapid ELISA using rHcp1 as the target antigen and compared it with our established OPS-ELISA [21]. The optimal conditions for our Hcp1-ELISA were initially determined using pooled serum from either melioidosis patients or healthy donors. The optimized concentration of rHcp1 for coating wells was 2.5 μg/ml. For the primary antibody incubation step, we used a serum dilution of 1:2000 at room temperature (25°C) for 30 minutes. The assay was standardized throughout the study using these conditions for all serum samples as previously described [21].
For comparison with our previous study using an OPS-ELISA, a total of 539 serum samples were tested in our Hcp1-ELISA [21]. These included on-admission sera from culture-proven melioidosis patients (N = 141), Thai healthy donors (N = 188), U.S. healthy donors (N = 90), tuberculosis patients (N = 20), scrub typhus patients (N = 50) and leptospirosis patients (N = 50) [21]. Quantitative results of OD values in both ELISAs are summarized in Table 1. The median OD of melioidosis patients for the Hcp1-ELISA was higher than that of OPS-ELISA [median OD 3.16 (IQR 2.22–3.40) versus 1.78 (IQR 0.67–3.11); P < 0.001]. The median OD value of the melioidosis group was statistically different from Thai healthy donors, U.S. healthy donors, tuberculosis patients, scrub typhus patients and leptospirosis patients for both ELISAs (P < 0.001 for both ELISAs for all comparisons between melioidosis patients versus each of other groups).
We determined the correlation between individual results of the Hcp1-ELISA and OPS-ELISA using serum samples from melioidosis patients, Thai healthy donors, U.S. healthy donors, tuberculosis patients, scrub typhus patients and leptospirosis patients. The pairwise correlation coefficient (rho) of all serum samples was 0.80, however, the relatedness between antibody response to the Hcp1 and OPS antigens was different between groups of serum samples. The results indicate a strong relatedness only in U.S. healthy donor group (rho = 0.93) but the results of the Hcp1-ELISA and OPS-ELISA were less correlated with the groups of Thai melioidosis patients (rho = 0.38) and Thai healthy donors (rho = 0.60). The correlation coefficients were 0.72, 0.50 and 0.74 for tuberculosis, scrub typhus and leptospirosis patients, respectively.
We next compared the diagnostic potential of each antigen individually (Hcp1-ELISA or OPS-ELISA) or combined (Hcp1/OPS-ELISA). ROCs were plotted by calculating the sensitivity and specificity of increasing numbers of the true-positive rate and false-positive rate. The results for comparisons of these ELISAs using the melioidosis group and Thai donors are shown in Fig 2A, and those using the melioidosis group and U.S. donors are shown in Fig 2B. When the results of Thai donors were analyzed, the areas under the receiver operator characteristic curves (AUROCCs) for diagnosis of melioidosis were highest and comparable between the Hcp1-ELISA and the combined Hcp1/OPS-ELISA (0.95 versus 0.94, P = 0.153) (Fig 2A). The AUROCC of the OPS-ELISA was significantly lower than that of the Hcp1-ELISA (0.91 versus 0.95, P = 0.001); and lower than that of Hcp1/OPS-ELISA (0.91 versus 0.94, P = 0.003). When the results from the U.S. donors were analyzed (Fig 2B), the AUROCC for diagnosis of melioidosis was highest for the combined Hcp1/OPS ELISA (0.96) and significantly higher when compared to the AUROCC of Hcp1-ELISA (0.93, P = 0.009) and the OPS-ELISA (0.92, P < 0.001). The AUROCC of the Hcp1-ELISA was not significantly different from that of the OPS-ELISA (0.93 versus 0.92, P = 0.353)
We further analyzed the sensitivity and specificity of Hcp1-ELISA in comparison to OPS-ELISA and Hcp1/OPS-ELISA using the 539 serum samples described above. To compare the performance of assays, we used an OD cut-off corresponding to a specificity of 95% using Thai healthy donors as controls (OD 1.165) (Table 2). The results demonstrated that the diagnostic sensitivity of the Hcp1-ELISA was significantly higher than that of the OPS-ELISA (83.0% versus 71.6%, P = 0.003). The sensitivity of the Hcp1-ELISA was not significantly different from the sensitivity of combined Hcp1/OPS-ELISA (83.0% versus 81.6%, P = 0.527). The specificity of the Hcp1-ELISA using U.S. healthy donors, tuberculosis patients, scrub typhus patients and leptospirosis patients as non-melioidosis controls were 95.6%, 100%, 98.0% and 100%, respectively. The specificity of the OPS-ELISA using U.S. healthy donors, tuberculosis patients, scrub typhus patients and leptospirosis patients as non-melioidosis controls were 96.7%, 100%, 94.0% and 98.0%, respectively.
We next investigated whether the use of antibody levels to Hcp1 and OPS as serodiagnostic markers of acute infection in melioidosis patients was influenced by the presence or absence of diabetes (Table 3). The sensitivity and antibody titers to Hcp1 and OPS were determined at week 0 for 200 follow-up patients. At serum dilution 1:2000, we found that the sensitivity of the Hcp1-ELISA for diabetic patients (N = 134) was significantly higher than for non-diabetic patients (N = 66) (87.3% versus 69.7%, P = 0.004). Similarly, the sensitivity of the OPS-ELISA for diabetic patients was significantly higher than for non-diabetic patients (88.1% versus 60.6%, P < 0.001). The median antibody titers for Hcp1 and OPS were significantly higher in diabetic patients compared to non-diabetic patients (P < 0.001 for both Hcp1 and OPS) (Fig 3). The median antibody titers for Hcp1 in diabetic patients was 26,322, (IQR 8,898–59,420) and non-diabetic patients was 10,327 (IQR 1,250–35,842). The median antibody titers to OPS in diabetic patients was 10,657 (IQR 4,487–29,157) and in non-diabetic patients was 3,499 (483.8–10,863).
We next compared the antibody titers for Hcp1 and OPS in melioidosis patients with or without bacteremia. The levels of Hcp1- or OPS-specific antibodies were determined at week 0 in 200 follow-up patients (Table 3). Using a serum dilution of 1:2000, the sensitivity of the Hcp1-ELISA for patients with bacteremia (N = 105) was not significantly different from that of the patients without bacteremia (N = 95) (82.9% versus 80.0%, P = 0.654). The sensitivity of the OPS-ELISA for patients with bacteremia was not statistically different from patients without bacteremia (83.8% versus 73.7%, P = 0.085). Although the median antibody titer for Hcp1 was not significantly different between bacteremic and non-bacteremic patients (median 22,108, IQR 9,045–66,612 versus median 14,769, IQR 4,154–39,255, P < 0.057), the median antibody titer for OPS was higher in bacteremia patients compared to non-bacteremia patients (median 9,636, IQR 4,150–28,886 versus median 6389, IQR 1817–16106, P = 0.019) (Fig 4).
The sensitivities of the Hcp1-ELISA and OPS-ELISA were determined using on-admission (week 0) serum collected from 198 melioidosis patients whose survival status was available (Table 3). The sensitivity of the Hcp1-ELISA for non-survivors (N = 64) was not significantly different from survivors (N = 134) (78.1% versus 82.8%, P = 0.629). The sensitivity of the OPS-ELISA for non-survivors was not different from the patients who survived (78.1% versus 79.1%, P = 0.91). The median antibody titers for Hcp1 in 134 patients who were survived was 18,035 (IQR 6,022–48,678) which was not significantly different from the median antibody titer for 64 non-survivors 25,873 (IQR 4,086–45,678); P = 0.822) (Fig 5). The median antibody titer for OPS in survivors was 9,588 (IQR 2,793–24,096) which was higher than the median of non-survivors 5,330 (IQR 2,114–17,670) but was not significantly different (P = 0.074).
We next investigated whether the antibodies to Hcp1 and OPS might be useful serodiagnostic markers during acute infection and/or following recovery from melioidosis. The specific antibodies were determined in 423 archived serum samples obtained from 200 melioidosis patients recruited in our recent longitudinal study [25]. Using a serum dilution of 1:2000, the percentage of positive serum samples in the Hcp1-ELISA was 81.5% (163/200), 82.3% (93/113) and 81.8% (90/110) at week 0, week 12 and week 52, respectively. The percentage of positive serum samples in the OPS-ELISA was 79.0% (158/200), 85.8% (97/113) and 88.0% (88/110) at week 0, week 12 and week 52, respectively.
To compare the antibody levels, we determined the endpoint antibody titers to Hcp1 and OPS in 103 individual patients who survived one year after acute infection (Fig 6). The median titer of melioidosis patients for the Hcp1-ELISA at week 0 was not different from that at week 12 [median 19,792 (IQR 6,654–57,596) versus 17,677 (IQR 4,327–44,079); P = 0.255], but the titer was significantly lower at week 52 [median 10,427 (IQR 3,731–19,046); titers of week 12 versus week 52, P < 0.019; week 0 versus week 52, P < 0.001] (Fig 6A). The results of individual patients for Hcp1 are shown in S1 Fig. Of the 103 patient samples tested, 49 (47.6%) showed decreased antibody titers for Hcp1 at week 52 compared to week 0 and 12 while only 31 (30.1%) showed decreased antibody titers for Hcp1 at week 12 compared to week 0 (Fig 6B). However, 14 (13.6%), 20 (19.4%) and 18 (17.5%) of patients had increased titers at week 52 compared to week 0, week 12 compared to week 0, and week 52 compared to week 12, respectively. We found 40 (38.8%), 52 (50.5%) and 36 (35.0%) of patients had no change in antibody titers at week 52 compared to week 0, week 12 compared to week 0, and week 52 compared to week 12, respectively.
In contrast, the median titer of melioidosis patients for OPS-ELISA increased at week 12 compared to week 0 [median 10,192 (IQR 2,783–26,502) versus 17,464 (IQR 7,346–38,834); P = 0.006] but decreased at week 52 [median 8,848 (IQR 3,293–15,359); P <0.001 for comparison between week 12 and week 52] (Fig 6A). The results of individual patients are shown in S2 Fig. The median titer was not different between week 0 and week 52 (P = 0.329). Of a total 103 patients, the number of patients that had decreased antibody titer for OPS at week 52 compared to week 0 was 30 (29.1%), at 12 compared to week 0 was 7 (6.8%), decreased titer at week 52 compared to week 12 was 53 (51.5%) (Fig 6B). However, 22 (21.4%), 41 (39.8%) and 3 (2.9%) of serum samples showed increase titer at week 52 compared to week 0, week 12 compared to week 0 and week 52 compared to week 12 respectively. We found 51 (49.5%), 55 (53.4%) and 47 (45.6%) of the patients had no change in antibody titer at week 52 compared to week 0, week 12 compared to week 0 and week 52 compared to week 12, respectively.
We next determined the correlation between the individual results from the Hcp1-ELISA and OPS-ELISA conducted using serum samples from 103 follow-up melioidosis patients who had survived at one year after admission (Fig 7). The pairwise correlation coefficient (rho) of all serum samples was 0.58 (P < 0.001). The relatedness between the antibody responses against Hcp1 and OPS was different for the 0, 12 and 52 week sample sets. The rho values at week 0 and week 12 were only 0.46 (P < 0.001) and 0.55 (P < 0.001), respectively. A stronger correlation (rho = 0.80, P < 0.001) was observed with the serum samples collected at week 52.
Melioidosis is a potentially fatal disease that is more widely distributed globally than previously recognized [2, 34]. A rapid and reliable POC serological test would be particularly useful as a tool for serodiagnosis and for seroprevalence studies in highly endemic regions as well as in countries where melioidosis is underreported. In the present study, we used a rapid ELISA platform to assess the serodiagnostic potential of various candidate target antigens. Our results showed that consistent with previous studies, antibodies specific for B. pseudomallei OPS were predominant in melioidosis patient sera. Interestingly, our WC-ELISA using both a wild type B. pseudomallei strain and an OPS mutant demonstrated that the median OD value for Thai healthy donors group was significantly lower in the OPS mutant-WC-ELISA compared to the wild type-WC-ELISA. The data suggested that the OPS-specific antibodies might also contribute to the high rate of seropositivity in healthy individuals from endemic areas and may influence the specificity of the assay. In contrast, the median OD value for U.S. healthy donor serum was not significantly different between the two ELISAs. A possible explanation for this finding is that individuals in Thailand and other endemic regions may be previously exposed to environmental Burkholderia species that express Type A OPS (e.g. B. thailandensis) whereas U.S. healthy donors would not. [35, 36].
Hcp1 is a major virulence factor that plays a critical role in the intracellular lifestyle of B. pseudomallei. Results of this study are consistent with our previous study [37] demonstrating that Hcp1 is immunogenic and is recognized by serum from melioidosis patients. We and others have demonstrated that Hcp1 is expressed at a low level when B. pseudomallei is cultured in vitro, but is produced at a high level in vivo within an intracellular environment [24, 37, 38]. Our findings suggest that the detection of antibodies to Hcp1 in a high percentage of melioidosis patients upon admission is likely to reflect infection with B. pseudomallei rather than non-infective exposure. In addition, since Hcp1 expressed by B. pseudomallei (and B. mallei) is structurally different than B. thailandensis Hcp1, it is likely that seropositivity to this protein antigen will be less prevalent than seropositivity to OPS in healthy donors in endemic areas.
Results of ROC analyses revealed that the Hcp1-ELISA had a significantly higher AUROCC than the OPS-ELISA when serum from Thai healthy donors was used as a control. While the OPS-ELISA alone may not be ideal for detection of acute infections in endemic areas, our data suggest that it may be more useful when combined with Hcp1 for use in non-endemic regions such as the USA. In support of these results, we demonstrated that the median OD value of Hcp1-ELISA in the melioidosis group was significantly higher than the median OD value of the OPS-ELISA. Our results also indicated that antibodies to Hcp1 are significantly elevated in patients with B. pseudomallei infections. The pairwise correlation coefficient for the results of the two ELISAs for all of the serum samples was high (0.80), however, the relatedness between antibody response to the Hcp1 and OPS antigens was lower in the patient group compared with the two healthy control groups. It is possible that the immune pathways activated by polysaccharide and protein antigens are different among individuals. Hcp1 is a T-cell dependent protein antigen while OPS is a carbohydrate that can induce humoral immune responses via a T-cell independent pathway. A recent report demonstrated that Hcp1 can bind to the surface of host antigen-presenting cells, which may contribute to its immunogenicity by inducing high antibody titers in melioidosis patients [24].
The diagnostic performance of Hcp1-ELISA for antibody detection using the first set of serum samples including 141 Thai melioidosis, 188 Thai healthy donors and 90 U.S. healthy controls showed a significant improvement over the conventional IHA. Using the same serum set, the Hcp1-ELISA had 83% sensitivity and 96% specificity compared with the IHA (sensitivity 69.5% and specificity 67.6%) in our previous study [20]. The results obtained from a second set of 200 melioidosis patient serum samples confirmed the high sensitivity (82%) at the time of admission. Our findings are consistent with a previous report by Lim and colleges [24] which demonstrated that anti-Hcp1 IgG titers in 20 melioidosis patient serum samples were significantly higher compared to serum from 20 healthy controls. In addition, an Hcp1-ELISA developed by Cheng et al using serum from 32 melioidosis patients and 20 healthy donors from Malaysia showed a sensitivity of 93.7% with a specificity of 100% [23].
Antibodies to OPS are highly elevated in melioidosis patients. Analysis of a follow-up set of 200 serum samples from diabetic or non-diabetic melioidosis patients in this study showed sensitivity to be 79%, which was consistent with the high sensitivity result of our previous study (72%) [21]. We observed higher seropositivity for both Hcp1 and OPS in diabetic melioidosis patients compared to those without diabetes. This has been previously seen for the IHA in Northern Australia [39] and for IHA in our laboratory in Thailand (manuscript in preparation). One possible explanation for the high seropositivity observed in diabetic patient is the alteration in the balance between cell-mediated immunity and humoral immunity in response to B. pseudomallei infection, for example due to enhanced polyclonal B-cell stimulation in Type 2 diabetes secondary to chronic hyperactivation of the innate immune response [40]. Ongoing studies are exploring the causality and exact mechanisms of higher antibody titers to B. pseudomallei in patients with diabetes. We reported no significant difference in diagnostic sensitivity between bacteremic versus non-bacteremic patients and survivors versus non-survivors. Thus, Hcp1 and OPS appear to be two promising antigens for further development of POC serological tests for all melioidosis patients including diabetics.
The induction of antibody responses to Hcp1 in melioidosis patients is relatively rapid with our results showing high median anti-Hcp1 titers at week 0. Interestingly, these titers were decreased at week 12 and week 52 for patients who recovered from the disease. Based on this, it appears that antibody titers to Hcp1 may be a useful serodiagnostic marker for acute infections as well as for monitoring the disease progression or recovery. The induction of antibody to OPS appeared to be slower with OPS-specific antibodies detectable at week 12 but declining by week 52. Further studies will be necessary to determine when OPS-specific antibody titers reach peak levels.
Our experiments focused on follow-up patients highlights the variation of individuals in the direction and rate of antibody titer changes over time. These results provide evidence of inter-individual variation in responses to the same B. pseudomallei antigens which may involve specific immune statuses, variable past exposures, infecting bacterial strains, and clinical disease factors. We recognize that our serological tests using a single serum dilution (1:2,000) may be of relatively limited value for following disease progression and that determining endpoint antibody titers will be more useful for assessing the variability of antibody levels in melioidosis patients.
In conclusion, this study establishes that Hcp1 and OPS are useful targets for serodiagnosis of melioidosis in various groups of patients. Overall, the Hcp1-ELISA provided better diagnostic assay values than the OPS-ELISA. When used in non-endemic areas, a combined Hcp1/OPS-ELISA showed increased sensitivity compared to the ELISAs using Hcp1 or OPS alone. Our results support accelerated development of Hcp1-based assays for a much needed POC test for the diagnosis of acute melioidosis.
|
10.1371/journal.pcbi.1002272 | A Whole-Body Model for Glycogen Regulation Reveals a Critical Role for Substrate Cycling in Maintaining Blood Glucose Homeostasis | Timely, and sometimes rapid, metabolic adaptation to changes in food supply is critical for survival as an organism moves from the fasted to the fed state, and vice versa. These transitions necessitate major metabolic changes to maintain energy homeostasis as the source of blood glucose moves away from ingested carbohydrates, through hepatic glycogen stores, towards gluconeogenesis. The integration of hepatic glycogen regulation with extra-hepatic energetics is a key aspect of these adaptive mechanisms. Here we use computational modeling to explore hepatic glycogen regulation under fed and fasting conditions in the context of a whole-body model. The model was validated against previous experimental results concerning glycogen phosphorylase a (active) and glycogen synthase a dynamics. The model qualitatively reproduced physiological changes that occur during transition from the fed to the fasted state. Analysis of the model reveals a critical role for the inhibition of glycogen synthase phosphatase by glycogen phosphorylase a. This negative regulation leads to high levels of glycogen synthase activity during fasting conditions, which in turn increases substrate (futile) cycling, priming the system for a rapid response once an external source of glucose is restored. This work demonstrates that a mechanistic understanding of the design principles used by metabolic control circuits to maintain homeostasis can benefit from the incorporation of mathematical descriptions of these networks into “whole-body” contextual models that mimic in vivo conditions.
| Homeostasis of blood glucose concentrations during circadian shifts in survival-related activities, sleep and food availability is crucial for the survival of mammals. This process depends upon glucose intake, short-term storage as glycogen, and gluconeogenesis. The integration of hepatic glycogen anabolic and catabolic dynamics with whole body energetics is critical for survival. In this paper we use computational modeling to investigate the potential survival advantage of substrate (futile) cycling of glycogen and glycogen precursors. Our simulations, combined with published experimental results of other researchers, indicate that as the body enters a state of fasting, the activity of enzymes involved in the synthesis of glycogen increases leading to increased substrate cycling. This increase in substrate cycling allows the system to respond more rapidly once new external sources of glucose become available. The whole-body computational model developed for this work allows the metabolic control circuitry to be studied under simulated in vivo conditions, providing functional insights that are not evident when individual modules of glycogen regulatory circuitry are examined in isolation.
| Glucose is the major metabolic fuel of mammals, with its maintenance at appropriate levels within the body being crucial for normal function, while dysregulation is associated with diseases such as diabetes mellitus, galactosemia and glycogen storage diseases [1]. Maintaining glucose levels requires a highly responsive control system capable of balancing a wide range of environmental conditions, perhaps the most basic of which is managing the uptake of nutrients from food at irregular time intervals. Specifically, transitions between fed and fasted states require rapid shifting between the storage of excess glucose, in the form of glycogen, within the liver and muscle and the breakdown of these stores for delivery of glucose to other organs. In healthy individuals, proper functioning of this system ensures that available nutrients are efficiently captured and stored during times of excess, while effectively managed and distributed during times of fasting.
The rate with which the organism responds to these changes can play a critical role in survival. Optimization of energy storage is essential during competition for sparse food supplies, while rapid delivery of these energy supplies during hasty retreat from predators can mean the difference between life and death [2]. A key player in energetics, especially for erythrocyte and brain function, is blood glucose concentration.
The liver is the central organ for regulation of glucose and glycogen and acts as the primary distributor of nutrients through the blood to other tissues. When in a fasted state, the liver breaks down glycogen stores, producing glucose for other tissues. After a meal, the liver switches to a glucose consuming state, capturing nearly 26% of the glucose presented to it by the portal system during the first passage [3]. Nearly 10–15% [4], [5] of liver weight is comprised of glycogen stores when filled.
Glucose regulation within the liver is performed by the glycogen circuit that controls both the storage of glucose as glycogen (glycogenesis) as well as its breakdown into glucose-6-phosphate from hepatic stores (glycogenolysis). Of significance is the fact that glycogenolysis and glycogenesis are not the result of a single reversible reaction, but rather are two separate, highly-regulated pathways. Two key molecular players within these pathways are glycogen synthase (GS) and glycogen phosphorylase (GP). GS drives the synthesis of glycogen, with its activity regulated through multiple mechanisms including allosteric activation, covalent modification, as well as enzymatic translocation [6]–[8]. GP catalyzes the rate-limiting step in glycogenolysis and it too, is actively regulated through phosphorylation at a single residue on the NH2 terminus as well as through allosteric regulation [6]–[8]. Both these enzymes exist in activated (GSa and GPa) as well as inactivated (GSb and GPb) states.
As the synthesis of glycogen and its breakdown into glucose occur through separate pathways, there is the potential for substrate cycling to occur, wherein glucose and glycogen are continuously interconverted. In fact, the glycogen circuit exhibits different behaviors depending on the state of liver glycogen levels (Figure 1). In the fed state, glucose is plentiful in the blood and glycogen levels within the liver are relatively high, resulting in the activation of GS and the synthesis of glycogen. When a fasting state is entered, glycogen levels in the liver are high and blood glucose levels are maintained by the breakdown of this glycogen into glucose-6-phosphate by GPa. Finally, when in the fully fasted state, glycogen stores within the liver are essentially depleted. It is here, in the context of glycogen depletion, that cycling is observed between glycogen and glycose-6-phosphate [9], [10].
It has long been suggested that substrate cycling is a generic mechanism that can potentially improve such properties as sensitivity and system response time, allowing net synthesis when there is a small offset in the substrate concentrations [7], [10]–[13]. However, demonstrations of cycling and its functional relevance in a physiological context are still relatively rare. In this work, we were particularly interested in investigating the potential role of the cycling - no cycling architecture of the glycogen circuit manifested during the transition from a fed to a fasted state. While the benefit of preventing substrate cycling is apparent since energy is dissipated in the form of heat during this process, it is not clear why it is beneficial for glycogen to cycle under the fasted state, as shown in Figure 1.
Mathematical models, which provide one way to explore such questions, have been applied successfully to many biological fields, but their application has been limited in the case of the nutritional sciences [14]. The number of mathematical models of hepatic energy metabolism, as it relates to hepatic glycogen storage, has been slowly increasing in response to interest in the impact of exercise on energetics in the case of diabetes [14], diet [15] and athletic training [16]. In addition, large-scale reconstructions of metabolism, typically based on flux or constraint-based models, have recently been developed for multiple organ systems including the liver [17]–[21]. These stochiometry-based approaches can be used to analyze the relevant biological network solely based upon systemic mass-balance and reaction capacity constraints when kinetic information is missing [22], [23]. However, as these approaches are based on steady state assumptions and do not consider specific kinetic properties, they provide a fundamentally different view of metabolism and metabolic dynamics than detailed mechanistic models.
In the absence of a suitable model for the present work, we developed a physiological model based on a central control glycogen circuitry by Hers et al. [7] and Mutalik et al. [24], with the whole body bioenergetics described in [25]–[29] as well as the feedback and feed forward control loops described in [30]–[33] for maintaining glucose homeostasis under different fed-fasting conditions. We placed specific emphasis on investigating the role of the cycling - no cycling architecture in metabolic functions. Building on previous biochemical and quantitative modeling descriptions, this model embedded the glycogen circuit of the liver within a physiological system composed of muscle, adipose tissue and blood compartments. By controlling the glucose injection rate into the blood stream, we were able to simulate system response across a broad range of fed/fasting conditions. Our simulation results reproduced previously published experimental observations and further indicated that the cycling design in Figure 1 provides a mechanism for decreasing the amount of time it takes to convert glucose to glycogen in the fasted state.
We now give a brief overview of our model, with full details and the complete set of model equations provided in Protocol S1. Note that the complete MATLAB package together with the description file are provided in Protocol S2 and S3, respectively. The SBML code is also provided in Protocol S4 for a broader usage and implementation. As noted earlier, glycogen is created from glucose during feeding and is subsequently degraded to release glucose-6-phosphate during fasting. The hepatic glycogen circuit controlling this process is embedded within the hepatocyte at the center of our physiological model (Figure 2). Blood is depicted as a closed loop, being carried around the body to connect multiple tissue compartments, including the liver, muscle, and fat. Thus blood functions as a transport system within our model, providing the resources needed to manufacture and store hepatic glycogen during the fed state while carrying its major degradation product, glucose, away during the fasted state for use by other tissues. The liver is currently the most detailed compartment in this model, including selected aspects of glycogenolysis, glycogenesis, glycolysis, gluconeogenesis, the TCA cycle, lipogenesis, lipolysis and ketogenesis (See Protocol S1 for model equations).
As an animal moves through the fed, fasting and fasted states, its body switches to different types of metabolic fuels to stabilize blood glucose concentration. This transition is controlled in large part by the blood levels of insulin and glucagon, both of which are generated in a reciprocal manner by the pancreas in response to changing blood glucose levels. Insulin and glucagon are mutually antagonistic with respect to many aspects of intermediary metabolism and their effects on bioenergetics [25], [34]. Insulin is a key regulator for carbohydrate and fat metabolism in the body. It enhances blood glucose uptake to form triglycerides and glycogen and suppresses pathways such as gluconeogensis and glycogenolysis [35]. Glucagon, on the other hand, is secreted from the pancreas when blood glucose concentration is low. It inhibits glycolysis and stimulates hepatic glycogenolysis and gluconeogenesis in liver by increasing the concentration of cAMP [36]. The elevated level of cAMP in turn activates a cascade of enzymes in the glycogen control circuitry that enhance the degradation of glycogen molecules [7]. Insulin and glucagon, working in a reciprocal fashion, in conjunction with other hormonal regulators, such as leptin and epinephrine, maintain glucose homeostasis in biological systems. Our physiological model also incorporates aspects of the Cori cycle, where lactate from muscle and erythrocytes is carried to the liver and converted to glucose for reuse by these tissues.
Blood glucose is provided from absorbed carbohydrates during feeding up until digestion is complete, at which point hepatic glycogen stores take over this role. Depletion of hepatic glycogen occurs over a period of 12–24 hours, though this varies greatly with activity levels [7], [37]. Once hepatic glycogen stores are consumed, blood glucose levels are maintained by gluconeogenesis. This process uses energy derived from storage fat in the form of acetyl CoA and the carbon skeletons of glycogenic amino acids. In the present physiological model, glycogenic amino acids are represented by alanine derived from muscle. The major sites of gluconeogenesis are the kidney and the liver, with only the latter being represented here. As blood glucose levels fall due to hepatic glycogen depletion, blood insulin levels fall while glucagon levels rise, leading to biochemical changes resulting in the use of alternative fuels in the form of free fatty acids and ketones, and gluconeogenesis which requires the use of such energy as mentioned above.
In tissues such as the heart and muscle, a number of factors regulate the use of alternative energy sources in order to spare blood glucose for use by erythrocytes (which depends solely on blood glucose [38]) and the brain (which mainly depends on blood glucose but can use ketone bodies as an alternative during fasting [39]). Our general bioenergetic model includes a number of tissue and biochemical components that were selected on the basis of their relationship to glycogen metabolism. The timing of these events and the dynamics of blood insulin, glucagon, glucose, free fatty acids, ketones, and levels of hepatic glycogen stores in response to fasting and feeding cycles from our simulation are shown in Figure 3 and are consistent with those which were previously described in [25]–[29].
Mutalik and Venkatesh [24] computed the dose response curves of the enzymes in the glycogen circuitry based on their empirically derived input functions for glucose-6-phosphate (G6p) and cAMP. Again, we note that the dissociation constant () of GS phosphatase and GPa is the key factor in determining the amount of substrate cycling at steady state. In fact, Mutalik et al. [24] defined different physiological states based on the value of , where a smaller value () corresponded to a fed state and a larger one () corresponded to a fasted state. We followed a similar approach to construct the dose response curves for these enzymes. Unlike [24], the glycogen circuitry was incorporated into a 4-compartment physiological model. As a result, the dynamics of cAMP and glucose-6-phosphate were simulated directly within our model and the entire system can be more realistically simulated by simply controlling the plasma glucose concentration.
The dose-response curves for GSa and GPa at two specified values of are shown in Figure 5. Here, the system was run to steady state with a fixed blood glucose concentration between 5 mM and 10 mM, the typical range for fed-fasting experiments in rodents [27], [44]–[46]. By increasing the constant from to , the crossover point of GPa and GSa shifted from a higher glucose concentration to a lower one (from to ) with a correspondingly higher activated fraction (from 5% to 60%). This fraction represents the maximum percentage of both enzymes being active simultaneously, thus it is an indicator of the degree of substrate cycling in the system. We note that the inhibitory effect by GPa on the activation of GS through direct binding to GS phosphatase is partially released with a larger .
From the above discussion, it is apparent that the inhibition of GS activation by GPa through direct regulation of GS phosphatase varies with the state of fasting. With a larger , the maximum amount of substrate cycling (co-activated fraction of GS and GP) is higher at the steady state. Here, we further investigated the dynamics of GS and GP but in the context of a glycogen depleted liver.
To simulate the response of the system to glucose in a glycogen depleted state, we provided a constant input of glucose with and ran the simulation to steady state. We then gradually throttled back the glucose input, and waited until liver glycogen was completely depleted. Glucose supply then re-entered the blood stream as a step function at t = 0, after which the dynamics of hormone, enzyme and substrate responses were observed. The results for two values of , and are shown in Figure 6A. Note that the observation period began at . A detailed description of the plasma glucose feeding function is provided in Figure S5 in Protocol S1.
Figures 6A–B show selected enzyme activities and glycogen concentration as a function of time. Note that “glycogen concentration” here and in the later context refers to the amount of glucose converted to glycogen as one molecule of glycogen comprises an indeterminate number (hundreds or thousands) of glucose subunits. At , liver glycogen stores are completely depleted and GP is mostly active (over 70% in the a-form). The sudden increase in the blood glucose concentration drove the transition from a GP-dominant to GS-dominant scenario. There was a major difference in when and where the intersection of GPa and GSa activity curves occurred for the two selected values of . Under , the point of intersection occurred at 60% and (Figure 6A-left panel). In contrast, this point shifted to 5% and with Figure 6B-right panel.
Figure 6B showed the liver glycogen concentration as a function of time. Again, the observations started at where the blood glucose supply resumed. Readily apparent was the slow but nearly immediate increase in the glycogen concentration at under (black line with dots). Recall that at this value, the level of inhibition of GS phosphatase by GPa was much reduced, allowing the coexistence of 60% of GSa and GPa. In contrast, the liver glycogen concentration remained at a negligible level until with (solid line with squares) where substrate cycling was reduced to 5%. Therefore, the system was able to respond quickly to the glucose stimulus and drive an immediate synthesis of glycogen with a higher level of substrate cycling. In both cases, a dramatic change in the synthesis rate of glycogen occurred where GSa and GPa intersect (15 mins and 30.6 mins correspondingly).
We next further investigated the relationship between the system response time and the level of substrate cycling in a glycogen depleted liver. Instead of two values of (marked by a red square () and triangle () in Figure 7), we considered a range of values from to . There are two different ways to define the system response time to glucose stimulus: (1) the time when the GSa and GPa curves intersect or (2) the time when glycogen concentration exceeds a threshold value. We selected a threshold value of 0.5 mM, the glycogen concentration reached at the end of the simulation () with the smallest . The time response curves under both definitions were shown in Figure 7A as a blue and black line respectively. The differences in the response time shown on both curves were on the order of 30 mins between the largest and smallest . In Figure 7B, we provided the co-active percentage of GS and GP at the intersection point.
The results from this analysis provide a possible explanation as to why the biological system has different metabolic mechanisms (different ) under different fasting states. In a glycogen-depleted state, it is essential to have a highly responsive system, ready for replenishing energy reserves as soon as nutrients become available. Our simulation results clearly showed that the high degree of substrate cycling occurring in the fasted state accelerated the system response in this respect by about 30 minutes, which would be physiologically significant for survival. Conversely, avoiding substrate cycling in a fed state is also desirable from an energy expenditure standpoint, as the combination of reactions involving GS, GP, glucose 1-phosphate uridylytransferase and nucleoside diphosphate kinase result in an ATP consuming reaction ().
We have shown that the level of inhibition of GS phosphatase by GPa through the dissociation constant , or equivalently the level of substrate cycling, determined the system response time in a glycogen depleted liver. Previous studies have shown that this inhibition is glycogen dependent [41], [42]. Watts et al. [47] reported that the GS phosphatase activity decreased in the livers of fasted, fed and gsd/gsd (liver glycogen storage disorder) mice and the addition of glycogen to homogenates of liver from starved rats reduced the glycogen synthase phosphatase activity. More recently, Armstrong et al. [48] pointed out that there are unique binding sites for GPa, PP-1 and glycogen in the hepatic glycogen-targeting subunit of protein phosphatase 1 (), a GS phosphatase specific to liver. Therefore, it is reasonable to assume that this inhibitory regulation changes according to the liver glycogen level. We modeled this effect by using the following expression for the dissociation constant :(8)where [glyc] is the liver glycogen concentration, , , and the Hill constant . Note that the parameters were chosen to match the experimental results of [49] as shown in Figure 8.
We compared our model predictions to experimental studies that investigated GS and GP levels within fed and fasted livers in a rodent model system [49]. In this work, Hue et al. measured GP and GS activities over time in isolated hepatocytes under sequential changes to the glucose concentration (from 5.5 mM to 55 mM) in the incubation medium. Results from this study were redrawn in Figure 8.
It is important to note that we are comparing a “whole-body” simulation with results obtained from cultured cells which are not interacting with events driven by other tissues, such as fat and muscle. However, this comparison demonstrates clear similarities between these cell culture data and our simulations with respect to responses of the glycogen regulatory circuitry to blood glucose concentrations. We started our simulation at the fed steady state and fasted the model system to two different times, 250 mins and 1200 mins, to represent fed and fasted livers respectively. In the simulation for fed livers, 250 mins fasting time was chosen to recreate a fasting environment as seen at the beginning of the experiments (Figure 8A), where GP is mostly in the active form and over 90% GS is in the inactive form [49]. Note that after 250 mins, the liver glycogen level was at about 75% of the fed steady state. In the simulation for fasted livers, 1200 mins was chosen after which only less than 1% glycogen remained. We then compared the response from both livers under 4 different glucose feeding rates (), as shown in Figure 9. Since we made our observation only after glucose supply re-entered the blood stream, we shifted the simulation time forward to 250 mins and 1200 mins in the fed and fasted livers and denoted them as .
Multiple aspects of our simulation results matched reported experimental observations of [49]. For instance, simulations and experiments showed the activation of GS to be highly suppressed by GPa in the fed state. For the lowest glucose injection rate, , GS is not activated at all, which was also observed in experiments by Hue et al. (Figure 8A). Both the experimental and simulation results showed that the active percentage of GS was higher in the fasted than in the fed state at the end of the experiment/simulation (). Furthermore, in both experimental studies and simulations, GS always responded more rapidly (on the order of 10–15 minutes as defined by the cross-over point of GSa and GPa) in the glycogen-depleted compared to the fed state. As the injection rate of glucose increased, the response time of GSa was shortened. Note that the glycogen concentrations from our simulation are provided in Figure S6 in Protocol S1, which also indicated a quicker response from the fasted livers. Although we can accurately capture the changes in response time under different glucose concentrations, it is clear that we have only addressed limited aspects of the relevant metabolic pathways and associated regulatory components. For instance, it is known that bioenergetics is regulated by a number of mechanisms including push-pull [50] and negative feedback, the latter being an integral component of our whole-body and glycogen-specific models. Furthermore, transcript level regulation is required to capture variations in enzyme concentration that occur under different fasting conditions. Such investigations lie outside the scope of the current model.
The cells, tissues, organs, bodies and populations of all living organisms are in a constant state of sensing and response to numerous external and self-generated stimuli [51]. Feedback loops, both positive and negative, play intrinsic roles in homeostatic regulation of biological systems. Negative feedback loops underpin the majority of the balances of nature, from predator-prey relationships to biochemical networks, and are clearly subjected to evolutionary pressures [52]. Negative feedback is a common mode of control for signaling networks [53], reducing time required to reach steady states [30], providing a mechanism for reducing fluctuations in protein expression levels and pathway activity. In contrast to stabilizing activity, in the presence of sufficient time delays, negative feedback can have destabilizing effect and generate overshooting and random oscillations, rendering noise a challenging issue in the modeling of biochemical networks [54]. No attempt was made to incorporate stochasticity into the present investigations. Biological systems employ negative feedback combined with controlled time delays as a means of inducing functional oscillations. Such internally generated oscillations are responsible in large part for circadian rhythms and the cell cycle [55], which are intimately linked to the subject of feed-fasting cycles in the present work.
In this work, we have developed a physiological model that simulates selected major components of bioenergetics as outlined in [25]–[27]. The outer general bioenergetics model (outer ring of Figure 2) was created as a “test bed” [56], [57] for the glycogen circuit, which permits simulation of the glycogen regulatory circuitry in response to physiological changes that mimic the effects of fasting and feeding on whole-body energetics. As we could find no such testbed for the hepatic glycogen regulatory circuit that we were investigating, and as such circuits interact in potentially unpredictable ways with other body systems via the vascular, nervous, and other communication systems, we endeavored to build such a software platform for our investigations. Analysis of this model suggests that the glycogen circuit's context-dependent (fed or fasted) architecture allows for a significant increase in response time when the organism is in a fasted state. Suppression of substrate cycling in the fed state could provide a strategy for energy conservation leading to optimal energy storage.
The current work also provides a platform for further investigation into bioenergetic diseases such as diabetes and glycogen storage disease (GSD). Type VI and type IX GSD, representing 25–30% of the total cases, are either due to a deficiency in glycogen phosphorylase or an abnormality in the enzyme that activates it [58]. Therefore, it is crucial to understand glucose-glycogen metabolism in a whole body environment, especially the regulatory mechanisms for some of the key enzymes in these pathways such as glycogen synthase and glycogen phosphorylase. Interestingly, this work could also be of value for research into optimization of nutrition protocols for athletes or soldiers who are required to perform under stress. Glycogen supercompensation, where glycogen storage ability is increased following glycogen depletion when consuming a high carbohydrate diet, is an important issue for performance in athletes. Numerous studies have been carried out to investigate the relationship between the amount and type of carbohydrate ingestion and the maximum glycogen resynthesis rate [59]–[61]. Of related interest, a study by Roberts et al. [62] demonstrated that metabolism of simple sugars leads to a higher glycogen resynthesis rate than that generated through the metabolism of complex carbohydrates. Under the current computational model platform, these observations could be further investigated in a continuous parameter space, and an optimal nutrition plan for these individuals might be predicted by taking into account energy flows. Computational models, such as the one developed here, could assist in the design of nutrition plans for athletes and individuals suffering from bioenergetic challenges, including diabetes.
One of the goals of our metabolic model was to capture key features of the dynamics of internal energy sources, from fed through fasted states, to include blood glucose, liver glycogen, FFAs, and ketone bodies, regulated by plasma glucagon and insulin. The dynamics of these substrates and enzymes are described in [25]–[27], while the whole body energetics have been reviewed in [26], [28], [29]. A summary of these time events is given in [63]. Such a simulation would then provide a dynamic framework within which to test the behavior of the underlying control circuitry, as for glycogen in the present study. When the physiological system enters the fasting state, blood glucose concentration drops, flipping a reciprocal switch with respect to plasma insulin and glucagon concentrations [28], [36]. cAMP then responds and transmits a signal to the glycogen circuitry to regulate the activities of GP and GS [7]. As a result, hepatic glycogen is being depleted as it is catabolized to maintain blood glucose levels within the physiological range needed for survival. The level of plasma free fatty acids and ketone bodies also rise to provide alternative metabolic fuels. A diagram of the concentrations of selected metabolites with respect to time after fasting commences is available in Figure 3. Except for the similar characteristic behaviors described previously in [25]–[29], our model is also able to capture the damped oscillations at the beginning of a new local stable state.
Here we give a brief overview of the four major compartments in our liver-centered physiological model, as shown in Figure 2. For a detailed description of these pathways, model equations and parameters, please refer to Protocol S1. A detailed parameter-based sensitivity analysis has also been conducted and results revealed that blood glucose is not sensitive to 10-fold changes in the parameters that describe the activity of each enzyme. The results are provided in Table S8–S10 in Protocol S1.
|
10.1371/journal.pgen.1004955 | Pooled Sequencing of 531 Genes in Inflammatory Bowel Disease Identifies an Associated Rare Variant in BTNL2 and Implicates Other Immune Related Genes | The contribution of rare coding sequence variants to genetic susceptibility in complex disorders is an important but unresolved question. Most studies thus far have investigated a limited number of genes from regions which contain common disease associated variants. Here we investigate this in inflammatory bowel disease by sequencing the exons and proximal promoters of 531 genes selected from both genome-wide association studies and pathway analysis in pooled DNA panels from 474 cases of Crohn’s disease and 480 controls. 80 variants with evidence of association in the sequencing experiment or with potential functional significance were selected for follow up genotyping in 6,507 IBD cases and 3,064 population controls. The top 5 disease associated variants were genotyped in an extension panel of 3,662 IBD cases and 3,639 controls, and tested for association in a combined analysis of 10,147 IBD cases and 7,008 controls. A rare coding variant p.G454C in the BTNL2 gene within the major histocompatibility complex was significantly associated with increased risk for IBD (p = 9.65x10−10, OR = 2.3[95% CI = 1.75–3.04]), but was independent of the known common associated CD and UC variants at this locus. Rare (<1%) and low frequency (1–5%) variants in 3 additional genes showed suggestive association (p<0.005) with either an increased risk (ARIH2 c.338-6C>T) or decreased risk (IL12B p.V298F, and NICN p.H191R) of IBD. These results provide additional insights into the involvement of the inhibition of T cell activation in the development of both sub-phenotypes of inflammatory bowel disease. We suggest that although rare coding variants may make a modest overall contribution to complex disease susceptibility, they can inform our understanding of the molecular pathways that contribute to pathogenesis.
| Crohn’s disease and ulcerative colitis are two forms of inflammatory bowel disease which cause chronic inflammation of the gastrointestinal tract. Common genetic variants in more than 160 regions of the human genome have been associated with an altered risk of these disorders, but leave much of the estimated genetic contribution to disease risk unexplained. We sought to establish whether rare genetic variants which alter the structure or function of the proteins encoded by genes also contribute to disease susceptibility. We used high throughput DNA sequencing to screen over 500 genes for such variants in nearly 500 patients and controls, and validated interesting variants in about 10,000 patients and 7,000 controls. We detected association of a limited number of rare variants from coding regions with disease, suggesting that they do not account for a large proportion of genetic susceptibility. However, they highlight the involvement of genes of potential importance in the development of inflammatory bowel disease, including those involved in the activation of immune cells, the regulation of immune response genes, and the degradation of proteins in cells.
| The inflammatory bowel diseases (IBD), Crohn’s disease (CD) and ulcerative colitis (UC) are chronic inflammatory disorders of the gastrointestinal tract that can cause diarrhoea, abdominal pain, bleeding and weight loss. Collectively they affect approximately 827 per 100,000 individuals in European populations and their incidence is rising [1]. CD may affect any part of the gut with discontinuous penetrating lesions, whereas in UC the disease is limited to the colon and rectum and the lesions are continuous but superficial [2]. Both diseases are multi-factorial, with a complex aetiology that involves a combination of an underlying genetic predisposition and environmental triggers. A variety of factors have been proposed to contribute to the pathogenesis including changes within the intestinal microbiota, a defective mucosal barrier, and / or dysregulation of the immune response [3].
A meta-analysis of genome-wide association studies (GWAS) in CD and UC by the International IBD Genetics Consortium (IIBDGC), followed by extensive confirmation of association signals in more than 75,000 individuals has increased the number of IBD-associated loci to 163 [4]. The majority of these loci are associated with both CD and UC, which suggests that there is extensive overlap in the biological mechanisms involved in their pathogenesis. However, although our understanding of the aetiology of IBD has been substantially advanced by GWAS-based approaches, only a modest proportion of total disease variance can be explained by current genetic findings (<15%) [4]. It has been proposed that rare coding sequence variants may make a substantial contribution to disease variance, and confer disease risks large enough to warrant use in preventative screening [5]. Such variants would not be detectable by a conventional GWAS approach because they are not well tagged by the common SNPs on which GWAS panels are based [6].
New high throughput DNA sequencing technologies have made it feasible to investigate the contribution of rare variants to complex disease. In CD, it has long been known that low frequency coding variants in NOD2 make a substantial contribution to disease risk [7–9], and more recent high-throughput sequencing strategies have discovered several independent IBD associated rare variants in NOD2 and other genes from GWAS loci including IL23R, CARD9, IL18RAP, CUL2, C1orf106, PTPN22, RNF186 and MUC19 [10–12]. However, a recent large-scale sequencing study of the coding regions of 25 autoimmune candidate genes in more than 40,000 individuals yielded little evidence that rare variants drive the associations observed at susceptibility loci for common immune disorders, including CD [13]. Thus the exact contribution of rare coding variants to IBD and other immune disorders remains unknown.
Here we describe a targeted high throughput sequencing approach in pooled DNA samples from 474 CD patients and 480 population controls to screen all exons, splice sites, and proximal promoter regions in 531 positional and functional candidate genes. We sequenced CD patients with early-onset disease and/or strong family history to enrich for functional causal variants with stronger effects, and we looked beyond common loci using functionally-derived bioinformatics data such as pathway and protein network analysis to identify additional candidate genes involved in key processes such as the immune-response and autophagy. Potential functional variants and those with evidence of association with CD underwent validation genotyping in a follow up study including 6507 IBD cases and 3064 controls with replication of the top hits in an additional 3662 IBD cases and 3639 controls giving a total of over 10,000 IBD cases and over 7,000 controls for the final combined analysis. We discovered significant novel association of a rare coding variant in BTNL2 and suggestive associations of additional variants in potentially novel IBD genes.
An overview of our strategy for the discovery of rare variants associated with CD is shown in Fig. 1. We selected 531 candidate genes for sequencing in phase I based on 5 selection criteria (Table 1 and described in Materials and Methods). A total of 6,249 exons, together with associated splice sites and proximal promoter regions, were sequenced in 474 CD cases and 480 population controls. Samples were sequenced in case-only or control-only pools of 12, 18 or 24 individuals using the Illumina Genome Analyzer II platform. An average of 98 million sequence reads were generated per pool, of which 87% could be aligned to the reference genome and 64% passed subsequent quality control steps (Materials and Methods). Of these, an average of 25.7 million reads mapped to the targeted genomic regions, which corresponded to a capture efficiency of 40.5%. We observed a mean read depth of >1000x per pool across the 1.57 million bases captured. Taking into consideration the number of individuals per pool, on average 90% of all bases had coverage greater than 4x per haploid genome (S1 Fig.).
In order to reduce false positives calls due to sequencing errors, we applied a stringent filtering procedure (Materials and Methods), after which the number of variants was approximately constant across all pools for all types of variants (S2 Fig.).
Next, variant allele frequencies in each pool were estimated from base-call counts. We assessed the accuracy of this approach by comparing these estimates to minor allele frequencies (MAFs) derived from genotyping data generated by the Wellcome Trust Case Control Consortium (WTCCC); genotypes were available for 153 SNPs located in the captured genomic regions in 66.5% (388 controls, 246 cases) of the individuals sequenced in this study [14]. We observed a very strong correlation (Spearman Rank Correlation r = 0.977) between MAFs for the WTCCC genotypes and the pooled sequencing data (Fig. 2).
After filtering, 3,749 single nucleotide variants (SNVs, here used to refer to any single nucleotide variation regardless of minor allele frequency) were retained, of which over half were low frequency (<5%, S1 Table). Just over half of the SNVs were located in exons (51.1%; 1914 SNVs), with the remainder located in introns, untranslated regions (UTRs), putative splice sites and intergenic regions. We considered 106 of the SNVs (3%) to be novel because they were not present in dbSNP138 (http://www.ncbi.nlm.nih.gov/SNP/). Analysis of all SNVs yielded a transition/transversion ratio (Ti/Tv) of 2.41, which is expected given the bias toward coding sequences in our target regions and is in agreement with previous studies [11]. In addition to SNVs we identified 183 deletions and 117 insertions. Only 14 of these insertion/deletions (indels) were located in an exon (S1 Table). A high rate of true positives in our sequencing data was corroborated by the presence of 97% of our variants in dbSNP138, and the strong correlation between MAFs for the pooled sequencing data and the WTCCC genotype data. Regarding sensitivity of variant detection, the regions captured in our sequencing contain 1,599 variants with a MAF >5% in the phase I release of the 1000 Genomes project, 1,291 of which (80.7%) were detected in our pooled sequencing data.
Our strategy relied on the necessity of sequencing individuals in case-only or control-only DNA pools which could potentially inflate any biases that would arise due to sequencing batch effects. We therefore used principal component analysis to control for this and identify any outlier pools. Examination of PC axes 1 and 2 revealed pools 7 and 8 to be outliers. Both were case pools, although each represented a single lane of flow-cell data from two different runs of the GAII sequencer. Once these pools were removed the data showed reasonable separation of points, but there was a clear tendency for case and control pools to be separated along PC axis 1 (S3 Fig.), which led to an overall genomic inflation of 1.3 (Fig. 3). The extent of the systematic bias in the data meant that PC axes could not be used as covariates in a logistic regression to correct for it, as previously noted [15], nor could we apply methods designed to correct for overdispersion but not bias [16]. We therefore applied a genomic control method for downstream association analysis plus additional QC measure for removal of SNVs with strong over dispersion among pools (Materials and Methods). We note that it is possible that the high systematic bias reflects genuine causal influences given the candidature of all the genes sequenced, but equally we cannot exclude the possibility of experimental sources of bias.
Variant level association with case-control status of pools was performed using logistic regression on 3,442 SNVs after exclusion of 307 SNVs that were too rare (had zero count in case or controls), or only had allele counts in excluded pools (Materials and Methods) (Fig. 3). Encouragingly, several known common and low frequency CD susceptibility variants were detected including variants in ATG16L1, IRGM, IL23R, CARD9 and NOD2, and rare variants in IL23R and NOD2 [7, 8, 10, 11, 17], all of which showed the expected CD odds ratios and allele frequencies in both cases and controls (S2 Table).
We noticed that 1,099 of 3,442 SNVs tested for association in our sequencing data were either included in the IBD Immunochip project directly (803) or by a suitable tagging SNP (r2≥0.8, n = 296) [4]. The IBD Immunochip dataset was therefore considered as an independent replication study for these 1099 variants. We found that 43 of the 141 variants (30.5%) that were at least nominally associated in our sequencing data (p<0.05) were also associated in the CD Immunochip data (p<5x10−8), resulting in a significant correlation between the two datasets (r = 0.446, p = 4.46x10−32).
The majority of variants identified by our study were rare, resulting in modest statistical power for the SNV-wise tests of association. We therefore applied gene-level association tests to investigate whether the burden of predicted functional variants non-synonymous and stop-gain variants) was different in cases compared to controls (Materials and Methods). In our discovery sequencing we identified 341 genes containing one or more functional variants. Thus the gene-burden test provided >90% power to detect a gene-level association where the cumulative MAF is 5% and the cumulative risk (OR) is 2.5 at an alpha level of 0.00015 (allowing for Bonferroni correction based on 341 genes/tests). We identified significant gene-level associations for BTNL2 (no. of variants = 18, p = 8.15x10−5) and NOD2 (no. of variants = 10, p = 9.03x10−6) (S3 Table). Since both genes contained substantially more functional variants than other genes that were tested we controlled for LD by permutation analysis (Materials and Methods), which resulted in loss of significance for BTNL2 (p = 0.022), whilst NOD2 remained significant (p<0.001). Repeating the analysis to include all intragenic variants (functional and non-functional) gave a similar outcome, although neither gene survived permutation testing (p>0.001).
In Phase II we selected 85 variants for validation of disease association by Sequenom (84 SNVs) or Taqman (1 SNV) genotyping in 6,335 IBD cases from the UK IBD Genetics Consortium (3,715 CD and 2,619 UC) and 2,974 controls (Materials and Methods). UC cases were included in the validation because of the extensive overlap in known associated loci for these two related phenotypes [4]. SNVs were selected based on at least nominal evidence of association in the pooled sequencing experiment (p < 0.05), and we prioritised those predicted to be functionally relevant (S1 Text). SNVs already genotyped as part of the IBD Immunochip experiment [4] were excluded. Post-genotyping quality control revealed that two SNVs failed to genotype, two were non-polymorphic and one was not in Hardy Weinberg equilibrium (p < 1x10−6 in controls) leaving a total of 80 SNVs (S4 Table). The genotyping call rate for all remaining SNVs was >90%. To allow validation of our variant calling analysis pipeline we genotyped an additional subset of 368 individuals previously included in our sequencing experiment and were able to show strong correlation between predicted and actual allele frequencies for all 80 SNVs (r = 0.94, p = 2.42x10−38) and low frequency SNVs (MAF<5%, r = 0.86, p = 1.69x10−24). In addition, allele frequencies derived from the pooled sequencing experiment were compared to those derived from all individuals in the phase II genotype data and revealed a highly significant correlation (r = 0.971, p < 6.58x10−48), further supporting the validity of the pooled sequencing approach. We followed up 3 insertion deletion polymorphisms by Taqman genotyping in 2,532 IBD cases and 3,545 controls (rs58682836/COBL frameshift delTTC, rs71297581/TYK2 upstream insC, and rs3833864/PIK3C upstream insC). The indel rs71297581 failed genotyping quality control, producing poor genotype clusters, and neither rs58682836 nor rs3833864 were associated with IBD (p > 0.5).
There was some evidence of association (p < 0.05) for 16 SNVs across 12 genes, CHTOP, ARIH2, NICN1, PLSCR1, IL12B, BTNL2, QRSL1, CALML5, GLT1D1, RTEL1, ATG4B and TBX21 (S5 Table). These were associated with either CD (11 variants), UC (5 variants) or IBD (12 variants), with 4 of these variants located in BTNL2. BTNL2 and IL12B map to established UC and IBD risk loci and have previously been implicated in UC and IBD respectively (6p21/HLA class II/UC and 5q31/IBD respectively), whilst ARIH2 and NICN1are within the same previously described IBD locus (3p21.3/IBD) but the genes themselves have not been implicated. Association of the other 10 genes and their respective variants with IBD has not been reported previously.
Since BTNL2 is within the MHC region and close to the common IBD associated locus in the HLA class 2 region we investigated the extent of LD across the 4 variants and their independence from the known risk locus using haplotype and conditional analysis within a set of cases and controls previously genotyped in both the Immunochip study and our follow up genotyping study (Materials and Methods). The analysis showed that the rare BTNL2 variants p.G454C and p.D336N (rs28362675 and rs41441651) were in almost complete LD with each other (r2 = 0.99) and remained associated with IBD even when the effect at the common SNPs was accounted for (p < 0.049), as did BTNL2 c.-118G>T (rs28362684, p = 0.039) but not the missense variant (p.S334L). Regarding association of the 80 variants with IBD, only the two highly correlated variants in BTNL2 (p.454C and p.D336N) surpassed the Bonferroni threshold for multiple testing (p < 0.0006 for 79 independent SNVs tested). However there was significant enrichment for association signals among the 79 variants, with nearly 3 times the number of significant results than would be expected by chance, with 14% of p-values for association with IBD (i.e. 11/79) being less than 0.05 (p = 0.00189).
Recognising the relatively low power of the validation panel to detect significant association of rare variants with disease, we next carried out extended genotyping (Phase III) of the 5 top SNVs that had a p < 0.01 (and in the case of BTNL2 were independent of each other and the known common risk variants) in an additional panel of 3,662 IBD cases and 3,639 controls (Materials and Methods), and then performed a combined case-control analysis of all 10,147 IBD cases and 7,008 controls that were either sequenced or genotyped (Table 2). We confirmed a genome-wide significant association with BTNL2 p.G454C and increased risk of IBD at (p = 9.65x10−10, OR = 2.3 [95%CI = 1.75–3.04]). We detected association for 3 other variants of the 5 tested in phase III (p < 0.005). Notably, in the combined analysis the direction of the effect for each of the 5 SNPs is consistent with the effect in the validation panel (p < 0.031). However, the 3 additional associations do not meet correction for 79 independent tests (P<0.00063) and are therefore suggestive. They include two low frequency missense variants IL12B p.V298F and NICN1 p.H191R associated with a reduced risk for IBD and one noncoding variant ARIH2 c.338-6C>T which was associated with an increased risk (Table 2). Two of the 3 missense variants associated with IBD (IL12B p.V298F and BTNL2 p.G454C) were predicted to be damaging or non-tolerated by Polyphen2 [18] and/or SIFT (sorts intolerant from tolerant) or Provean [19]. IL1B encodes the p40 subunit common to both the interleukin-12 and interleukin-23 heterodimeric cytokines. The p.V298F variant is not in LD with the common risk variant at this locus (r2 = 0.001, D’ = 0.079), and is predicted to disrupt the structure of the p40 protein by the mCSM structure prediction tool [20], with a predicted stability change ΔΔG of −0.917. We also used the available structure of the IL12B (p40) and IL23A (p19) proteins to model the effect of the V298F mutation in IL12B (S4 Fig.). This indicated an altered conformational state of a region of p40 which is important for binding to its partner proteins IL23A (p19) and IL12A (p35) [21].
BTNL2 is located on chromosome 6p21.3, which contains two common and independent risk loci for IBD. The closest (approximately 200Kb proximal to BTNL2) is within the HLA class II region and is associated with UC (rs477515, p = 5x10−133). The other locus is much further away (approximately 1.1Mb distal of BTNL2) within the HLA-class I region, and associated with CD (rs9264942, p = 5x10−28) [4]. We observed that BTNL2 p.G454C was associated very strongly with UC (p = 3.5x10−12, Table 2) and also associated with CD but to a lesser extent (p = 3.6x10−5, Table 2). In view of the extended LD in this region, it is possible that these associations could be due to LD with the known common risk variants in the HLA class I or class II regions. We investigated this by further conditional logistic regression analysis using 1,638 IBD cases and 1,243 controls genotyped in both the Immunochip study and both our genotyping studies. We confirmed that BTNL2 p.G454C was not in LD with either of the two common IBD risk variants (r2 < 0.001, D’ < 0.7). Conditional analysis showed that BTNL2 p.G454C remained significantly associated with IBD when the effect at the common UC associated SNP (rs477515) was accounted for (p = 0.0045, S6 Table), or the common CD associated SNP (rs9264942) was accounted for (p = 4.83x10−5, S6 Table). Haplotype analysis showed that the risk “A” allele for the rare variant occurred on haplotypes containing either the non-risk or the risk allele for both of the common variants, further suggesting their independence. Given the strength of the effect of p.G454C in UC individuals in particular (Table 2) we carried out specific haplotype analysis using this and the common UC GWAS SNP in the class II HLA region and showed that haplotype A-A containing the risk allele at the rare variant (p.G454C) and the non-risk “A” allele at the common UC GWAS SNP (rs477515) respectively, although very rare, was increased in frequency in cases, (0.2%) compared to controls (0.07%) (S7 Table), and the haplotype G-A containing the risk allele at both the common and the rare variant had a much higher risk for disease (OR = 6.51 [95%CI = 1.87–22.72]) than the haplotype G-C that only had the risk allele at the common SNP and lacked the rare risk allele (OR = 1.38 [95%CI = 1.20–1.57]).
In this study we investigated the contribution of rare variants to susceptibility to inflammatory bowel disease in a large set of candidate genes. Use of targeted next generation sequencing in combination with a DNA pooling strategy allowed us to screen over 500 genes for variants in more than 900 individuals, which is ten-fold more than were investigated in previous studies of IBD [10–12]. The results demonstrate that this is a cost-effective strategy for identifying low frequency variants that may be associated with disease. We were able to validate our approach by accurate estimation of the minor allele frequencies of 153 SNPs previously genotyped in individual case and control samples by the Affymetrix 500K SNP array, and by successfully reproducing the effect sizes (odds ratios) and allele frequencies of multiple common and low-frequency variants previously associated with IBD. We also demonstrated highly significant overlap of association for 1,099 SNPs that were common to our study and the recent GWAS/Immunochip meta-analysis for IBD [4], and showed a strong correlation between the allele frequencies and odds ratios of 80 SNVs that were genotyped by both pooled DNA sequencing and genotyping in our follow up study. Strong correlations between allele frequency estimates from pooled sequencing and genotyping have also been reported in previous studies of Crohn’s disease [10, 11], although read counts tended to underestimate actual frequencies for rare variants in one study [10]. However this approach could prove useful when supported by stringent quality control and validation measures.
Sequencing of coding and potential regulatory regions of 531 genes in a discovery set of 954 individuals, followed by genotyping in 17,131 individuals has allowed us to identify a novel disease associated genetic variant within a gene that maps to a region previously associated with IBD, and suggestive associations of other variants in a known IBD susceptibility gene and in other genes not previously implicated in IBD. The association of the rare variant p.G454C in BTNL2 reached genome-wide significance, and was independent of the known common risk variants for IBD in the HLA region in both a conditional and haplotype analysis. However, this is a complex region of the genome with extensive allelic variation and linkage disequilibrium, and additional as yet unknown IBD risk variants at this locus may exist that are independent of the two main HLA signals previously described but correlated with our rare variant. The glycine residue is highly conserved across all mammals and the cysteine substitution is predicted to be damaging by SIFT (score = 0.01) and probably damaging by PolyPhen2 (score = 0.997). This variant was in almost complete LD with another missense variant D336N which is not predicted to be damaging. BTNL2 codes for the butyrophilin like protein 2, which is a member of butyrophilin family that shares sequence homology with the B7 co-stimulatory molecules. The butyrophilins are implicated in T cell inhibition and the modulation of epithelial cell-T cell interactions [22]. BTNL2 negatively regulates T-cell activation independently of CD28 and CTLA-4, is predominantly expressed in gastrointestinal tissues including human terminal ileum (www.gtexportal.org), and is overexpressed in mouse models of colitis [23]. Recently it has been shown that BTNL2 promotes the expression of Foxp3, which is a transcription factor required for regulatory T cell development and function [24]. In view of its important role in immune modulation and homeostasis and an expression pattern restricted to intestinal epithelial and immune cells, mutations in BTNL2 may affect its ability to regulate T cell activation in response to mucosal inflammation. Common variants at the BTNL2 locus, have been previously shown to be associated with ulcerative colitis whilst being independent of the nearby known HLA susceptibility alleles [25]. Additional coding and loss-of-function variants in BTNL2, have been associated with susceptibility to other immune related disorders including adult-onset sarcoidosis [26, 27] and rheumatoid arthritis [28].
Although no variants other than the two rare and highly correlated missense mutations in BTNL2 surpassed the Bonferroni threshold for testing the 79 independent variants for association with IBD, there was significant enrichment for association signals among these 79 variants, and our extension study and combined analysis showed that the direction of the effect for all 5 SNVs tested was consistent with the initial finding. This suggests that there are likely to be additional true positives within phase II and III of our study that have not met the stringent Bonferroni threshold. This emphasises the difficulty in obtaining statistically robust evidence for association of rare variants even with a combined sample of 17,000 tested here and a relatively large effect size such as, for example, ARIH2 c.338-6C>T, OR = 2.39.
The association of common variants at the IL12B locus with both CD and UC is well established [4], although no obvious causal variant has yet been found. The association of the low frequency IL12B variant V298F with IBD which was detected in our sequencing experiment was retained in the combined analysis of 10,146 IBD and 7,008 controls, (p = 0.00183, OR = 0.82 [95%CI = 0.72–0.93]). IL12B encodes the IL12p40 subunit common to both IL12 and IL23, both of which are produced by activated dendritic cells and macrophages and lead to activation of distinct subsets of T-cells. We found that the minor allele of V298F is associated with a reduced risk of both CD and UC and is independent of the common risk variants at this locus. The variant is predicted to have a damaging or destabilizing effect on protein function or structure, and modeling of the effect of the mutation on the structure of the p40 subunit predicted an altered conformational state which could affect binding to its partner proteins. Thus the rare (Phe) allele may reduce the risk of IBD by attenuating the activation of T cell populations by IL12 and IL23.
We found two additional suggestive associations in ARIH2 and NICN1. Ariadne homolog 2 (ARIH2) is a member of an unusual family of E3 ubiquitin-protein ligases. Loss of ARIH2 has been shown to cause degradation of IκBβ in dendritic cells leading to dysregulated activation of NFκB. The SNP rs200140527 is associated with IBD, and is 6bp upstream of the splice acceptor site for exon 9 of ARIH2, although the C>T change is not predicted to affect the strength of the splice site [29]. Nicolin 1 (NICN1) is a nuclear protein and part of the neuronal tubulin polyglutamate complex [30] although very little else is known about its function. It is expressed in multiple tissues including the human terminal ileum and transverse colon (www.gtexportal.org). The nonsynonymous SNP p.H191R is associated with a protective effect for CD and UC in this study. NICN1 is on chromosome 3 at 49.46Mb, i.e. approximately 460kb proximal to ARIH2 on 3p21 and within a 2Mb locus previously associated with IBD that contains multiple independent genome-wide significant SNPs [4].
Previous sequencing studies have reported that rare coding variants make a limited contribution to the genetics of immune disorders and hypertriglyceridaemia, explaining 1–2% of their genetic variance [10–13, 31, 32]. However, these studies have generally sequenced a limited number of genes located in regions derived from the association of common variants with the disease. Our study highlights the challenges in identifying rare variant association for a polygenic complex trait like IBD. In sequencing more than 500 genes from both GWAS and pathway or network analysis combined with follow up genotyping in over 17,000 individuals we found genome-wide significant association of a rare variant in one gene and suggestive association of 3 SNVs in 3 other genes. However, our follow up studies were powered to detect associations of rare variants with relatively strong effects. For example, our phase II validation panel had 57% power to detect association of a low frequency variant with an allele frequency of 2.5% and OR = 1.3 at alpha level of 0.01 (to flag candidate associations), and 75% power to detect a rare variant with an allele frequency of 1% and OR of 1.6. In the combined analysis of 10,147 cases and 7,008 controls, we had 69% power to confirm association of a variant with a MAF of 0.025 and OR of 1.3 at alpha level of 0.0006 (correction for 79 SNV tests), but 89% power to confirm association for a variant with MAF 0.01 and an OR of 1.6. It is therefore likely that some rare variants with effect sizes of less than 1.6 remain undiscovered in these genes. It is also possible that a proportion of variants that are recognised as being suggestive of association in this study may turn out to be false positives, so further replication and subsequent functional studies will be required to prove causality.
If our 4 newly discovered associations were added to the 26 low frequency SNVs identified in 13 other genes from previously published studies of IBD [7, 10–12, 17, 33] this would total 30 IBD associations with low frequency SNVs in 17 of 548 sequenced genes. However, these screens have predominantly interrogated the coding regions of less than 3% of all known genes. Our study has targeted <25% (198) of all the known genes that map to the 163 IBD associated regions identified by the most recent mapping efforts of the International IBD Consortium [4]. A comprehensive evaluation of the true extent of the contribution of rare coding variants to IBD will have to await whole exome sequencing of very large numbers of case and controls [34], and whole genome sequencing to capture rare regulatory variants in non-coding regions.
The value of studies of rare variants in IBD lies not only in the discovery of additional risk variants which may aid future genetic profiling in at risk populations, but also in their potential to discover further genes and pathways involved in IBD. Our study provides additional evidence of the importance of the regulation of T cell activation and mucosal T cell responses involving BTNL2, and the potential role of proteosomal degradation in the pathogenesis of IBD.
A total of 531 candidate genes were selected based on: (a) Crohn’s disease GWAS hits; (b) GWAS hits from other immune disorders; (c) Pathway analysis based on Gene-set enrichment analysis; (d) IBD related literature; and (e) Network Analysis (Table 1). Details of these selection criteria are provided in S1 Text. Exon coordinates from RefSeq [35] and Ensembl [36] were combined to include all potentially coding regions. Proximal promoters were included by selection of genomic regions from 200 bp upstream to 50 bp downstream of the transcription start site. Putative splice sites were included by addition of five bp each side of coding exons. In total 6,290 genomic intervals were successfully synthesized for the Agilent SureSelect DNA Capture Array. Capture probes (120 bp; 60bp tiling) corresponding to 1,569,003 bp of target sequence.
Crohn’s disease patients for the sequencing experiment (n = 474) were recruited from specialist IBD clinics in London and Newcastle [37] after informed consent and ethical review (REC 05/Q0502/127). Population controls for sequencing (n = 480) were obtained from the 1958 British Birth Cohort [38]. All individuals were of European ancestry. The chances of detecting rare variants with large effects in the sequencing stage was increased by selection of Crohn’s disease (CD) patients with an early age of onset <20 years (n = 204), or with a family history of IBD (n = 174) or both early onset and family history (n = 96). Additionally, 178 (86%) of those individuals with a family history also had at least one affected first degree relative. DNA samples were quantified in triplicate (Qubit, Life technologies) prior to pooling in equimolar amounts to a total of 3 μg of DNA. Pools of 24 CD case DNA samples or 24 control DNA samples were made with a total of 44 pools, 474 cases and 480 controls (including 9 pilot/test pools of 12 and one test pool of 18 CD cases; S1 Text) and libraries were prepared following standard protocols. The validation panel for phase II, consisted of 3,799 unrelated CD and 2,708 unrelated UC, patients recruited by the UK IBD Genetics Consortium [4] and the replication panel consisted of an additional 1644 CD cases and 2018 UC cases recruited from London and Newcastle (as described above). Additional population controls (n[validation] = 3,064; n[replication] = 3622) were from the 1958 British Birth cohort and the National Blood Donor Service [14]. All cases and controls analysed in the replication phase III were independent and unrelated to those sequenced in the phase I and phase II discovery cohort.
Sequencing reads were aligned to the hg18 (NCBI 36) reference genome using Novoalign (version 2.07.09, Novocraft Technologies). We performed quality control using SAM tools [39] and removed PCR duplicates using Picard tools [40]. SNVs and indels were called using SAM-tools and filtered based on the following criteria: i) Phred base quality score ≥ 20, ii) any allele to have at least two base calls on each strand, iii) minimum base call count for any allele to be the equivalent to at least one expected chromosome count (N allele-specific base calls / N total base calls * 2 * N individuals in pool), with at least 0.3 expected chromosome counts attributable to each strand, iiii) criteria to be met in at least three different pools from at least two different batches. These parameters were optimized to reduce biases across all 44 pools (S2 Fig.). After filtering, base call counts were normalised to allele frequencies for each pool based on the total number of base calls that passed the filtering criteria. Variants were annotated using ANNOVAR [41]. Further details of read alignment, quality control and variant calling are provided in S1 Text.
After excluding variants previously implicated with IBD and variants analysed in the IBD Immunochip project [4], we selected 96 SNVs for follow up in phase II using the Sequenom iplex genotyping platform. We chose variants that a) surpassed multiple testing in the pooled sequencing based case-control comparison (p < 10−5), b) were modestly significant in the pooled sequencing based case-control comparison (p < 0.05) and had a low allele frequency (MAF < 5%), c) had functional consequence (within 20bp of a splice acceptor or donor site or non-synonymous variant), and were novel or low frequency (< 1%), d) were absent from one group (either controls or cases) and had a functional consequence (within 20bp of a splice acceptor or donor site or non-synonymous variant). In total 84 SNVs passed design and were genotyped via Sequenom iplex in 2,974 controls, 3,715 Crohn’s disease and 2,620 ulcerative colitis cases. Individuals for which more than 20% of SNVs could not be called were excluded from further analysis. One additional SNV (rs138274580/ATG4B) and 3 indels (rs58682836/COBL frameshift delTTC, rs71297581/TYK2 upstream insC, rs3833864/PIK3C upstream insC), that failed iplex design, were genotyped using the TaqMan chemistry (Life Technologies); SNP since they were ranked as high priority in all categories of our variant selection criteria (S1 Text). Finally we selected 5 SNVs with p<0.01 in any one phenotype (CD, UC or IBD) and, in the case of multiple SNVs in BTNL2, were indicated by LD and conditional regression analysis to be independent of each other and the known common risk variants, for replication genotyping via KASPTM chemistry at LGC Genomics (Hoddesdon, Herts, UK) in 3666 additional IBD cases and 3622 additional controls. In order to validate previous phases we also included a further 858 individuals who had been sequenced and/or undergone sequenom iplex genotyping. To investigate LD and independence of the BTNL2 variants from the known IBD GWAS hits within the MHC we used Immunochip data supplied by the UKIBD Genetics Consortium that was available for 1,638 of our genotyped IBD cases and 1,243 of genotyped controls.
Allele frequencies for each SNV in each pool were standardized and subjected to principal components analysis (PCA) to identify outlier pools and investigate systematic bias between cases and controls. PCA revealed considerable bias, such that cases and control pools could be largely separated by PC axis 1 alone. Various statistical methods for dealing with pooled SNV data have been proposed [16]. In light of the PCA results, we adopted a genomic control approach because it can correct for both overdispersion (additional variance that is distributed equally among pools) and bias (a consistent tendency for allele frequencies in case pools to be different from controls pools). For each SNV, a logistic regression across pools was performed using expected chromosome counts for the two most common alleles to form the dependent variable, and case-control status as the independent variable. The reversal of the conventional functional from allows for different pool sizes to be readily accounted for, and also appropriately reflects the study design (pool status is fixed by the experimenter, not pool allele frequencies). Genomic control was performed by dividing the chi-square statistic for association by the median chi-square statistic across all SNVs. We used evidence for especially strong SNV-specific overdispersion among pools (via a test of residual deviance from the logistic regression for association, p < 1.5x10−5) as an additional QC measure for removal of suspect SNVs.
Burden tests for significant association of a group of SNVs (e.g. all SNVs in a gene) were also performed taking in account both the pooled design and the presence of case-control bias. For a given set of n SNVs, genomic-control-corrected z2 values were summed and tested against the chi-squared distribution with n degrees of freedom. Significant sum-statistics were further tested via permutation of case-control status among pools, to correct for false positives that could be caused from linkage disequilibrium distributing the same signal among multiple SNVs. Note that our burden test allows SNV groups containing a mixture of both risk and protective variants to be tested appropriately.
Statistical analyses of pooled sequencing data was performed using R project for statistical computing (http://www.r-project.org/). Cases-control analysis of validation and replication genotyping data was performed with PLINK version 1.07 [42] using Armitage Trend Test. Additional conditional regression, linkage disequilibrium and haplotype analysis at known common IBD loci was performed using UNPHASED v3.0.12 [43].
The effect of the mutation Val298Phe on IL12B (p40) protein stability was examined using the tool mCSM, which predicts the effect of mutations in proteins using graph-based signatures [20]. The structure of the complex of human IL12B (p40) and IL23A (p19) is available in the RCSB Protein Data Bank [44] (PDB entry 4GRW), and was used as the template to model the structure of mutant IL12BV298F. The modelling procedure first generated the sequence alignment between the target (IL12BV298F) and the template structure (4GRW chain B) by running the tool T-Coffee [45]. The aligned sequences were then used as an input to the structure modelling package Modeller 9v8 [46]to generate 200 structures of IL12BV298F. Among these, only the one with the best Discrete Optimized Protein Energy score was selected for inspection of the mutation Val298Phe. The structure representation tool PyMol (Version 1.5.0.4, Schrödinger, LLC)was used for visual inspection and structural analysis. The interaction between IL12BV298F and IL23A was modelled by superimposing the IL12BV298F structure onto the human wild-type IL12B.
|
10.1371/journal.pcbi.1006033 | A probabilistic, distributed, recursive mechanism for decision-making in the brain | Decision formation recruits many brain regions, but the procedure they jointly execute is unknown. Here we characterize its essential composition, using as a framework a novel recursive Bayesian algorithm that makes decisions based on spike-trains with the statistics of those in sensory cortex (MT). Using it to simulate the random-dot-motion task, we demonstrate it quantitatively replicates the choice behaviour of monkeys, whilst predicting losses of otherwise usable information from MT. Its architecture maps to the recurrent cortico-basal-ganglia-thalamo-cortical loops, whose components are all implicated in decision-making. We show that the dynamics of its mapped computations match those of neural activity in the sensorimotor cortex and striatum during decisions, and forecast those of basal ganglia output and thalamus. This also predicts which aspects of neural dynamics are and are not part of inference. Our single-equation algorithm is probabilistic, distributed, recursive, and parallel. Its success at capturing anatomy, behaviour, and electrophysiology suggests that the mechanism implemented by the brain has these same characteristics.
| Decision-making is central to cognition. Abnormally-formed decisions characterize disorders like over-eating, Parkinson’s and Huntington’s diseases, OCD, addiction, and compulsive gambling. Yet, a unified account of decision-making has, hitherto, remained elusive. Here we show the essential composition of the brain’s decision mechanism by matching experimental data from monkeys making decisions, to the knowable function of a novel statistical inference algorithm. Our algorithm maps onto the large-scale architecture of decision circuits in the primate brain, replicating the monkeys’ choice behaviour and the dynamics of the neural activity that accompany it. Validated in this way, our algorithm establishes a basic framework for understanding the mechanistic ingredients of decision-making in the brain, and thereby, a basic platform for understanding how pathologies arise from abnormal function.
| Decisions rely on evidence that is collected for, accumulated about, and contrasted between available options. Neural activity consistent with evidence accumulation over time has been reported in parietal and frontal sensorimotor cortex [1–5], and in the subcortical striatum [6, 7]. What overall computation underlies these local snapshots, and how it is distributed across cortical and subcortical circuits, is unknown.
Multiple models of decision making match aspects of recorded choice behaviour, associated neural activity or both [8–16]. While successful, they lack insight into the underlying decision mechanism. In contrast, other studies have shown how exact inference algorithms may be plausibly implemented by a range of neural circuits [17–21]; however, none of these has reproduced experimental decision data.
Here we test the hypothesis that the brain implements an approximation to an exact inference algorithm for decision making. We show that the algorithm reproduces behaviour quantitatively while the dynamics of its inner variables match those of corresponding neural signals on the random dot motion task—a highly developed paradigm to probe decision formation. By doing so, we predict how experimentally-acquired snapshots of neural activity map onto inference operations. We show this mapping accounts for the involvement of full recurrent cortico-subcortical loops in decision making. Evidence accumulation is thus predicted to occur over the entire loops, not just within cortex. Introducing this algorithm enables us to predict which aspects of neural activity are necessary for inference—hence decision-making—and which are not. For instance, recent data questioned whether non-increasing cortical firing rates encode evidence accumulation during decisions [22, 23]. We demonstrate that, counter-intuitively, non-increasing as well as increasing cortical rates can encode likelihood functions, and hence evidence accumulation.
Our algorithm explains the decision-correlated experimental data more comprehensively than any prior model, thus introducing a new, cohesive formal framework to interpret it. Collectively, our analyses and simulations indicate that mammalian decision-making is implemented as a probabilistic, recursive, parallel procedure distributed across the cortico-basal-ganglia-thalamo-cortical loops.
We tested our algorithm against behavioural and electrophysiological data recorded in sensorimotor cortex [3] and striatum [6], from monkeys performing 2- and 4-alternative reaction-time versions of the random dot motion task (Fig 1b and 1c). The decision evidence for the algorithm also simulates spike-trains from sensory cortex (the area that provides evidence to sensorimotor cortex), whose statistics we extracted from a third random-dot-task data set by [24]. In all forms of the task, the monkey observes the motion of dots and indicates the dominant direction of motion with a saccadic eye movement to a target in that direction. Task difficulty is controlled by the coherence of the motion: the percentage of dots moving in the target’s direction.
During the dot motion task, neurons in the middle-temporal visual area (MT) respond more vigorously to visual stimuli moving in their “preferred” direction than in the opposite “null” direction [24]. Both the mean (Fig 1d) and variance (Fig 1e) of their response are proportional to the coherence of the motion (see also S4 Fig). MT responses are thence assumed to be the uncertain evidence upon which a choice is made in this task [1, 9].
Normative algorithms are useful benchmarks to test how well the brain approximates an optimal probabilistic computation. The family of the multi-hypothesis sequential probability ratio test (MSPRT) [25] is an attractive normative framework for understanding decision-making. However, the MSPRT is a feedforward algorithm. It cannot account for the ubiquitous presence of feedback in neural circuits and, as we show ahead, for slow dynamics in neural activity that result from this recurrence during decisions. To solve this, we introduce a novel recursive generalization, the rMSPRT, which uses a generalized, feedback form of the Bayes’ rule we deduced here from first principles (Eq 5).
We now conceptually review the MSPRT and introduce the rMSPRT (Fig 2), giving full mathematical definitions and deductions in the Materials and methods. The (r)MSPRT decides which of N parallel, competing alternatives (or hypotheses) is the best choice, based on C sequentially sampled streams of evidence (or data). For modelling the dot-motion task, we have N = 2 or N = 4 hypotheses—the possible saccades to available targets (Fig 1b and 1c)—and the C uncertain evidence streams are assumed to be simultaneous spike-trains produced by visual-motion-sensitive MT neurons [1, 9] (see Methods). Every time new evidence arrives, the (r)MSPRT refreshes ‘on-line’ the likelihood of each hypothesis: the plausibility of the combined evidence streams assuming that hypothesis is true. The likelihood is then multiplied by the probability of that hypothesis based on past experience (the prior). This product for every hypothesis is then normalized by the sum of the products from all N hypotheses; normalisation is crucial for decision, as it provides the competition between hypotheses. The result is the probability of each hypothesis given current evidence (the posterior)—a decision variable per hypothesis. Finally, posteriors are compared to a threshold, whose position controls the speed-accuracy trade-off. A decision is then made to either choose the most probable hypothesis, if its posterior surpassed the threshold, or to continue sampling the evidence streams otherwise. Crucially, the (r)MSPRT allows us to use the same algorithm irrespective of the number of alternatives, and thus aim at a unified explanation of the N = 2 and N = 4 dot-motion task variants.
The MSPRT is a special case of the rMSPRT (in its general form in Eqs 5 and 10) when priors do not change or, equivalently, for an infinite recursion delay; that is, Δ → ∞. Also, the previous recurrent extension of MSPRT [18, 26] is a special case of the rMSPRT when Δ = 1. Hence, our rMSPRT generalizes both in allowing the re-use of posteriors from any given time in the past as priors for present inference. This uniquely allows us to map the rMSPRT onto neural circuits containing arbitrary feedback delays, in particular solving the problem of decomposing the decision-making algorithm into distributed components across multiple brain regions. We show below how this allows us to map the rMSPRT onto the cortico-basal-ganglia-thalamo-cortical loops.
Inference using recursive and non-recursive forms of Bayes’ rule gives the same results (e.g. see [27]), and so MSPRT and rMSPRT perform identically. Thus, like MSPRT [17, 25], for N = 2 rMSPRT also collapses to the sequential probability ratio test of [28]; the rMSPRT is thereby optimal, not only in the oft-used sense of using all available information to do statistical inference (e.g. using the Bayes’ rule), but also in the strict sense that it requires the smallest expected number of observations, thus the shortest time to decide, at any given error rate (which follows from [29]). This is to say that the (r)MSPRT is quasi-Bayesian in general: the physical limit of performance or ideal Bayesian observer for two-alternative decisions (N = 2), and an asymptotic approximation to it for decisions between more than two (N > 2) (which follows from [17, 25]).
The hypothesis that the brain approximates an exact inference algorithm during decision formation is so far untested. This requires showing how uncertain sensory spike-trains can be transformed into the experimentally recorded choices. We do so here for the first time by comparing the predicted choice reaction times of the (r)MSPRT to those of monkeys performing the random dot motion task. We sought to account for the reaction time dependence on three factors: the coherence of the dot motion, the number of decision alternatives, and the trial’s outcome (error, correct). We use a particular instance of rMSPRT (Eqs 9 and 10) to determine predicted normative bounds on the decision time in the dot motion task. We can then ask how well monkeys approximate such bounds. The bounds result from using a minimal amount of sensory information, by assuming as many evidence streams (spike-trains from MT neurons) as alternatives; that is, C = N. Thus, this rMSPRT instance gives the upper bound on optimal expected decision times (exact for N = 2 alternatives, approximate for N = 4) per condition (given combination of coherence and N). Assuming C > N would predict even shorter optimal expected decision times (see [20]).
We assume that during the random dot motion task (Fig 1a–1c), the evidence streams for every possible saccade come as simultaneous sequences of inter-spike intervals (ISI) produced in MT. On each time step, fresh evidence is drawn from the appropriate (null or preferred direction; see Methods) ISI distributions extracted from MT data (Fig 1f). By repeating the simulations for thousands of trials per condition, we can compare algorithm and monkey performance.
Using these data-determined MT statistics, the (r)MSPRT predicts that the mean decision time on the dot motion task is a decreasing function of coherence (Fig 3a). For comparison with monkey reaction times, the algorithm’s reaction times are the sum of its decision times and estimated non-decision time, encompassing sensory delays and motor execution. For macaques 200–300 ms of non-decision time is a plausible range [30, 31]. Within this range, monkeys tend not to reach the predicted upper bound of reaction time (Fig 3a).
The (r)MSPRT framework suggests that decision times directly depend on the discrimination information in the evidence. Discrimination information here is measured as the divergence between pairs of distributions of ISIs (those in Fig 1f) produced simultaneously by MT neurons responding to the same stimulus: one where they were tuned to the dominant motion direction of the dots (it was their preferred; solid lines in Fig 1f) and another where they were not (it was a null direction; dashed lines). Intuitively, the larger this divergence or difference, the easier and hence faster the decision. We can estimate how much discrimination information monkeys used by asking how much the exact inference performed by (r)MSPRT would require to obtain the same reaction times on correct trials as the monkeys, per condition. We thus find, first, that the discrimination information available for decision is very similar across N (Fig 3b), implying that monkeys use MT sensory information consistently. Second, and most important, we find that monkeys tended to use less discrimination information than that in ISI distributions in their MT when making the decision. In contrast, the (r)MSPRT uses the full discrimination information available. This implies that the decision-making mechanism in the monkey brain lost large proportions of MT discrimination information (Fig 3c). Since these (r)MSPRT decision times are upper bounds, this in turn means that this loss of discrimination information in monkeys (Fig 3c) is the minimum.
To verify if this information loss alone could account for the monkeys’ deviation from the (r)MSPRT upper bounds, we depleted the discrimination information of its input distributions to exactly match the estimated monkey loss in Fig 3c per condition. We did so only by modifying the mean and standard deviation of the null direction ISI distribution, to make it more similar to the preferred distribution (exemplified in Fig 3d).
Using these information-depleted statistics, the mean reaction times predicted by the (r)MSPRT in correct trials closely match those of monkeys (Fig 4a). Importantly, this involved no parameter fitting. Instead, we used the fact that for (r)MSPRT the mean total information for a decision is constant given error rate and N; this implies that longer decision times could only result from reducing the discrimination information in the evidence. Strikingly, although this information-depletion procedure is based only on data from correct trials, the (r)MSPRT now also matches closely the mean reaction times of monkeys from error trials (Fig 4b), which are consistently longer than those of correct trials (S1 Fig). Moreover, for both correct and error trials the (r)MSPRT accurately captures the relative scaling of mean reaction time by the number of alternatives (Fig 4a and 4b).
The reaction time distributions of the algorithm closely resemble those of monkeys in that they are positively skewed and exhibit shorter right tails for higher coherence levels (Fig 4c–4f). These qualitative features are captured across both correct and error trials, and 2 and 4-alternative tasks. Together, these results support the hypothesis that the primate brain approximates an algorithm similar to the rMSPRT, ‘starved’ of sensory discrimination information.
The above shows that the (r)MSPRT family of exact inference algorithms can account for the dependence of choice reaction times on task difficulty, trial outcome, and the number of alternatives. But replicating behaviour alone does not tell us if the brain implements a similar computation during decisions. We thus asked whether the inner variables of the rMSPRT could account for the known dynamics of neural activity in cortex and striatum during the dot-motion task. To answer this, we must first map its components to a neural circuit. The rMSPRT is the first probabilistic model of decision able to handle recursion and arbitrary signal delays, which means that in principle it could map to a range of feedback neural circuits. Because cortex [1–5], basal ganglia [6, 32] and thalamus [33] have been implicated in decision-making, we sought a mapping that could account for their collective involvement.
In the visuo-motor system, MT projects to the lateral intra-parietal area (LIP) and frontal eye fields (FEF)—two ‘sensorimotor cortex’ areas. The basal ganglia receives topographically organized afferent projections [34] from virtually the whole cortex, including LIP and FEF [35–37]. In turn, the basal ganglia provide indirect feedback to the cortex through thalamus [38, 39]. This arrangement motivated the feedback embodied in rMSPRT.
Multiple parallel recurrent loops connecting cortex, basal ganglia and thalamus can be traced anatomically [38, 39]. Each loop in turn can be sub-divided into topographically organised parallel loops [39, 40]. Based on this, we conjecture the transient organization of these circuits into N functional loops, for decision formation, to simultaneously evaluate the possible hypotheses.
Our mapping of computations within the rMSPRT to the cortico-basal-ganglia- thalamo-cortical loop is shown in Fig 5, capturing the most prominent functional features of such circuits. For instance, it has been demonstrated that the striato-nigral and the subthalamo-nigral pathways of the basal ganglia compete during decision formation [41]. The computations predicted by rMSPRT to map on the striatum, subthalamic nucleus, and substantia nigra pars reticulata (SNr; see S3 Fig), provide a qualitative formalization of this phenomenon.
Also, negative log-posteriors will tend to decrease for the best supported hypothesis and increase otherwise. This is consistent with the idea of basal ganglia output nuclei (e.g. SNr) selectively removing inhibition from a chosen motor program while increasing inhibition of competing ones [17, 32, 42, 43].
Lastly, our mapping of rMSPRT provides an account for the spatially diffuse cortico-thalamic projection [44], previously unaccounted for by probabilistic models of decision. It predicts that the projection conveys a constantly-increasing, hypothesis-independent baseline that does not affect the inference carried out by the cortico-basal-ganglia-thalamo-cortical loop, but may produce the offset required to facilitate the cortical re-use of inhibitory, fed-back decision information from the basal ganglia (see S2 Fig). This increasing baseline may form part of the hypothesis-independent drive dubbed the “urgency signal” by [31], revealed after averaging LIP population responses across choices. All this is consistent with current views on the active modulation of information transmitted to the cortex by thalamus [45].
The mapping of rMSPRT to cortico-subcortical circuits produces key, testable predictions. First, that sensorimotor areas like LIP or FEF in the cortex evaluate the plausibility of all available alternatives in parallel, based on the evidence produced by MT, and join this to any initial bias. Second, that as these signals traverse the basal ganglia, they compete, resulting in a decision variable per alternative. Third, that the basal ganglia output nuclei use these to assess whether to make a final choice and what alternative to pick. Fourth, that decision variables are returned to sensorimotor cortex via thalamus, to become a fresh bias carrying all conclusions on the decision so far. The rMSPRT thus predicts that evidence accumulation happens uninterruptedly in the overall, large-scale loop, rather than in a single site.
With the mapping above, we can compare the dynamics of rMSPRT computations to those of recorded activity during decision-making in area LIP and striatum. We first consider the dynamics around decision initiation. During the dot motion task, the mean firing rate of LIP neurons deviates from baseline into a stereotypical dip soon after stimulus onset, possibly indicating the reset of a neural integrator [1, 14]. LIP responses become choice- and coherence-modulated after the dip [1]. This also occurs when firing rates deviate from the initial baseline in striatum, where no dip is exhibited [6]. We therefore reasoned that LIP and striatal neurons engage in decision formation from the bottom of the dip or deviation from baseline (respectively) and model their mean firing rate from then on. After this, mean firing rates “ramp-up” for ∼ 40 ms in LIP, then “fork”: they continue ramping-up if dots moved towards the response (or movement) field of the neuron (inRF trials; Fig 6a, solid lines) or drop their slope if the dots were moving away from its response field (outRF trials; dashed lines) [1, 3]. Striatal neurons exhibit an analogous ramp-and-fork pattern of response (Fig 7c and 7d). The magnitude of LIP firing rate is inversely proportional to the number of available alternatives (Fig 6a and 6b) [3, 46]; a phenomenon also recorded in other visuo-motor sites, notably in the superior colliculus [47] and FEF [48–50].
The model LIP (sensorimotor cortex) in rMSPRT captures each of these properties: activity ramps from the start of the accumulation, forks between putative in- and out-RF responses, and scales with the number of alternatives (Fig 6c). Under this model, inRF responses in LIP occur when the likelihood function represented by neurons was best matched by the uncertain MT evidence; correspondingly, outRF responses occur when the likelihood function was not well matched by the evidence.
The rMSPRT embodies a mechanistic explanation for the ramp-and-fork pattern in the two cases of Eq 9. Initial accumulation (steps 0–2 in our simulations; feedforward inference) occurs before the feedback has arrived at the model sensorimotor cortex, resulting in a ramp. The forking (step 3; start of feedback inference) is the point at which the posteriors from the output of the model basal ganglia first arrive at sensorimotor cortex to be re-used as priors. By contrast, non-recursive MSPRT (without delayed feedback of posteriors) predicts well-separated neural signals throughout (Fig 6e). With recursion as the key difference, our framework suggests, first, that the ramp-and-fork pattern gives away the existence of an underpinning delayed inhibitory drive within a looped architecture—here from the model basal ganglia. Second, that the fork represents the time at which updated signals representing the competition between alternatives (posterior probabilities in the rMSPRT) are first made available to the sensorimotor cortex.
The rMSPRT further predicts that the scaling of activity in sensorimotor sites by the number of alternatives is due to cortico-subcortical loops becoming transiently organized as N parallel functional circuits, one per hypothesis. This would determine the baseline output of the basal ganglia. Until task related signals reach the model basal ganglia output, it codes the initial priors for the set of N hypotheses. Their output is then an increasing function of the number of alternatives (Fig 6f). This increased inhibition of thalamus in turn reduces baseline cortical activity as a function of N. The inverse proportionality of cortical activity to N in macaques during decisions (Fig 6a and 6b; [3, 46, 48, 49]) and the direct proportionality of the firing rate to N in their SNr [42] lend support to this hypothesis.
The rMSPRT also captures key features of dynamics at decision termination. For inRF trials, the mean firing rate of LIP neurons peaks at or very close to the time of saccade onset (Fig 6b). By contrast, for outRF trials mean rates appear to fall just before saccade onset. The rMSPRT can capture both these features (Fig 6d) when we allow the algorithm to continue updating after the decision rule (Eq 10) is met. The decision rule is implemented at the output of the basal ganglia and the model sensorimotor cortex peaks just before the final posteriors have reached it. The rMSPRT thus predicts that the activity in LIP lags the actual decision.
This prediction may explain an apparent paradox of LIP activity. The peri-saccadic population firing rate peak in LIP during inRF trials (Fig 6b) is commonly assumed to indicate the crossing of a threshold and thus decision termination. Visuo-motor decisions must be terminated well before saccade to allow for the delay in the execution of the motor command, conventionally assumed in the range of 80–100 ms in macaques [9, 30]. It follows that LIP peaks too close to saccade onset (∼ 15 ms before) for this peak to be causal. The rMSPRT suggests that the inRF LIP peak is not indicating decision termination, but is instead a delayed read-out of termination in an upstream location.
In the rMSPRT, the striatum relays the input from sensorimotor cortex as an inhibitory drive for downstream basal ganglia nuclei. The rMSPRT has three free parameters that shape the ramp-and-fork of its inner variables, but do not alter inference. We have set their value to show that mapped variables can match the pattern in sensorimotor cortical neural dynamics (see Methods); below we show how these predictions depend on the parameter values. Nonetheless, the rMSPRT with these parameters also captures the ramp-and-fork pattern of activity in the monkey striatum (compare panels c, d to e, f in Fig 7).
LIP and striatal firing rates are also modulated by dot-motion coherence (Fig 7a–7d, 7k, 7l). Following stimulus onset, the response of these neurons tends to fork more widely for higher coherence levels (Fig 7a, 7c and 7k) [1, 3, 6]. The increase in activity before a saccade during inRF trials is steeper for higher coherence levels, reflecting the shorter average reaction times (Fig 7b, 7d and 7l) [1, 3, 6]. The sensorimotor cortex or striatum in the rMSPRT shows coherence modulation of both the forking pattern (Fig 7e and 7m) and slope of activity increase (Fig 7f and 7n). rMSPRT also predicts that the apparent convergence of peri-saccadic LIP activity to a common level during inRF trials (Fig 7b and 7l) is not required for inference and so may arise due to additional neural constraints. We take up this point in the Discussion.
Our proposed mapping of the rMSPRT’s components (Fig 5) makes testable qualitative predictions for the mean responses in basal ganglia and thalamus during the dot motion task. For the basal ganglia output, likely from the oculomotor regions of the SNr, rMSPRT (like MSPRT) predicts a drop in the activity of output neurons during inRF trials and an increase in outRF ones. It also predicts that these changes are more pronounced for higher coherence levels (Fig 7g, 7h, 7o and 7p). These predictions are consistent with recordings from macaque SNr neurons showing that they suppress their inhibitory activity during visually- or memory-guided saccade tasks, in putative support of saccades towards a preferred region of the visual field [42, 51, 52], and enhance it otherwise [52].
In detection tasks like visually- or memory-guided ones, the decision cues are extremely obvious. Hence, the accompanying recorded neural-activity transients may be argued to encode very short evidence-accumulations. After all, the accumulation of a single observation (e.g. an ISI) is the simplest, albeit degenerate case of evidence accumulation.
For visuo-motor thalamus, rMSPRT predicts that the time course of the mean firing rate will exhibit a ramp-and-fork pattern similar to that in LIP (Fig 7i, 7j, 7q and 7r). The separation of in- and out-RF activity is consistent with the results of [33] who found that, during a memory-guided saccade task, neurons in the macaque medio-dorsal nucleus of the thalamus (interconnected with LIP and FEF), responded more vigorously when the saccade target was flashed within their response field than when it was flashed in the opposite location.
Understanding how a neural system implements an algorithm is complicated by the need to identify which features are core to executing the algorithm, and which are imposed by the constraints of implementing computations using neural elements—for example, that neurons cannot have negative firing rates, so cannot straightforwardly represent negative numbers. The three free parameters in the rMSPRT allow us to propose which functional and anatomical properties of the cortico-basal-ganglia-thalamo-cortical loop are workarounds within these constraints, but do not affect inference.
One free parameter enforces the baseline activity that LIP neurons maintain before and during the initial stimulus presentation (Fig 7a and 7k). Varying this parameter, l, scales the overall activity of LIP, but does not change the inference performed (Fig 8a). Consequently, this suggests that the baseline activity of LIP depends on N but does not otherwise affect the inference algorithm implemented by the brain.
The second free parameter, wyt, sets the strength of the spatially diffuse projection from cortex to thalamus. Varying this weight changes the forking between inRF and outRF computations but does not affect inference (Fig 8b). The third free parameter, n, sets the overall, hypothesis-independent temporal scale at which sampled input ISIs are processed; changing n varies the slope of sensorimotor computations, even allowing all-decreasing mean firing rates (Fig 8c). By definition, the log-likelihood of a sequence tends to be negative and decreases monotonically as the sequence lengthens. Introducing n is required to get positive simplified log-likelihoods, capable of matching the neural activity dynamics, without affecting inference. Hence, n may capture a workaround of the decision-making circuitry to represent these whilst avoiding signal ‘underflow’, by means of scaling the input data.
Traditionally, evidence accumulation is exclusively associated with increasing firing rates during decision, and previous studies have questioned whether the often-observed decision-correlated yet non-increasing firing rates (e.g. in outRF conditions in Fig 7a, 7c and 7k and [1–3, 5, 53, 54]) are consistent with accumulation [22, 23]. The diversity of patterns predicted by rMSPRT in sensorimotor cortex (Fig 8) solves this by demonstrating that both increasing and non-increasing activity patterns can house evidence accumulation.
We tested the hypothesis that the brain approximates exact inference for decision making. We did so by showing that a novel recursive form of the MSPRT, the rMSPRT, uniquely accounts for both monkey choice behaviour and the corresponding neural dynamics in cortex and striatum, while its architecture matches that of the cortico-subcortical decision circuits.
The recursive computation implied by the looped cortico-basal-ganglia-thalamo-cortical architecture has several advantages over local or feedforward computations. First, recursion makes trial-to-trial adaptation of decisions possible. Priors determined by previous stimulation (fed-back posteriors), can bias upcoming similar decisions towards the expected best choice, even before any new evidence is collected. This can shorten reaction times in future familiar settings without compromising accuracy. Second, recursion provides a robust memory. A posterior fed-back as a prior is a sufficient statistic of all past evidence observations. That is, it has taken ‘on-board’ all sensory information since the decision onset. In rMSPRT, the sensorimotor cortex only need keep track of observations in a moving time window of maximum width Δ —the delay around the cortico-subcortical loop— rather than keeping track of the entire sequence of observations. For a physical substrate subject to dynamics and leakage, like a neuron in LIP or FEF, this has obvious advantages: it would reduce the demand for keeping a perfect record (e.g. likelihood) of all evidence, from the usual hundreds of milliseconds in decision times to the ∼ 30 ms of latency around the cortico-basal-ganglia-thalamo-cortical loop (adding up estimates from [55–57]).
The rMSPRT decides faster than monkeys in the same conditions because monkeys do not make full use of the discrimination information available in their MT (Fig 3b). However, this performance gap arises partially because rMSPRT is a generative model of the task. Thus, this assumes that knowledge of coherence is available by decision initiation, which in turn determines appropriate likelihoods for the task at hand. Any deviation from this generative model will tend to degrade performance, whether it comes from one or more of: the coherence to likelihood mapping [58], the inherent leakiness of neurons, or correlations between spikes or between neurons (see [20]). In this respect, we must consider, first, that the activity dip ∼ 170 ms after stimulus onset is assumed to indicate decision engagement at the LIP level. By then, MT neurons have been reliably modulated by motion coherence for about 120 ms (starting ∼ 50 ms after stimulus onset; see S4 Fig for details), giving a sizeable window to adjust LIP ‘likelihood functions’ to match the decision at hand. Whether this window is large enough or if trial-by-trial ‘likelihood adjustment’ occurs at all remain as interesting questions for future experimental explorations. Second, that LIP neurons change their coding during learning of the dot motion task and MT neurons do not [59], implying that learning the task requires mapping of MT to LIP populations by synaptic plasticity [60]. Consequently, even if the MT representation is perfect, the learnt mapping only need satisfice the task requirements, not optimally perform.
Excellent matches to monkeys’ performance in both correct and error trials, and hence their speed-accuracy trade-offs, were obtained solely by accounting for lost information in the evidence streams. No noise was added within the rMSPRT itself. Prior experimental work reported perfect, noiseless evidence integration by both rat and human subjects performing an auditory task, attributing all effects of noise on task performance to the variability in the sensory input [61]. Our results extend this observation to primate performance on the dot motion task, and further support the idea that the neural decision-making mechanism can perform perfect integration of uncertain evidence.
Neurons in LIP, FEF [4], and striatum exhibit a ramp-and-fork pattern during the dot motion task. Analogous choice-modulated patterns have been recorded in the medial premotor cortex of the macaque during a vibro-tactile discrimination task [53] and in the posterior parietal cortex and frontal orienting fields of the rat during an auditory discrimination task [5]. The rMSPRT indicates that such slow dynamics emerge from decision circuits with a delayed, inhibitory drive within a looped architecture. This suggests that decision formation in mammals may use a common recursive computation.
A random dot stimulus pulse delivered earlier in a trial has a bigger impact on LIP firing rate than a later one [2]. This highlights the importance of capturing the initial, early-evidence ramping-up before the forking. However, multiple models omit it, focusing only on the forking (e.g. [9, 10, 13]). Other, heuristic models account for LIP activity from the onset of the choice targets, through dots stimulation and up until saccade onset (e.g. [12, 14–16]). Nevertheless, their predicted firing rates rely on two fitted heuristic signals that shape both the post-stimulus dip and the ramp-and-fork pattern. In contrast, the ramp-and-fork dynamics emerge naturally from the delayed inhibitory feedback in rMSPRT during decision formation.
rMSPRT qualitatively replicates the ramp-and-fork pattern for individual coherence levels and given number of alternatives, N (Fig 6). However, the peak of the accumulated evidence in the model sensorimotor cortex of rMSPRT does not converge to a common value around decision termination during inRF trials. Consequently, it predicts that the apparent convergence of LIP activity to a common value (Figs 6b and 7b and 7l) is not part of the inference procedure, but reflects other constraints on neural activity.
One such constraint is that these brain regions engage in multiple other computations, some of which are likely orthogonal to solving the random dot motion task. The neural activity recorded during decision tasks may then be a transformation of inference computations, by mixing them with all other simultaneous computations. Consistent with this, the successful fitting of previous computational models to neural data [12, 14–16] has been critically dependent on the addition of heuristic signals for unknown constraints. While beyond the scope of this study, which examined whether a normative mechanism could explain behaviour and electrophysiology during decisions, adding similar heuristic signals to the rMSPRT would likely allow a quantitative reproduction of the peri-saccadic convergence of LIP activity.
Inputs to the rMSPRT were determined solely from MT responses during the dot-motion task, and it has only three free parameters, none of which affect inference. It is thus surprising that it renders emergent predictions that are consistent with experimental data. First, our information-depletion procedure used exclusively statistics from correct trials. Yet, after depletion, rMSPRT matches monkey behaviour in correct and error trials (Fig 4), suggesting a mechanistic connection between them in the monkey that is naturally captured by rMSPRT. Second, the values of the three free parameters were chosen solely so that the model LIP activity resembled the ramp-and-fork pattern observed in our LIP data-set (Fig 6a and 6c). As demonstrated in Fig 8, the ramp-and-fork pattern is a particular case of two-stage patterns that are an intrinsic property of the rMSPRT, guaranteed by the feedback of the posterior after the delay Δ has elapsed (Eq 5). Nonetheless, the algorithm also qualitatively matches LIP dynamics when aligned at decision termination (Fig 6b and 6d). Third, the predictions of the time course of the firing rate in SNr and thalamic nuclei naturally emerge from the functional mapping of the algorithm onto the cortico-basal-ganglia-thalamo-cortical circuitry. These are already congruent with existing electrophysiological data; however, their full verification awaits recordings from these sites during the dot motion task. These and other emergent predictions are an encouraging indicator of the explanatory power of a systematic framework for understanding decision formation, embodied by the rMSPRT.
The rMSPRT contains all previous instances of the MSPRT [17, 18, 25, 26, 62] as special cases. It generalizes them by allowing the re-use of posteriors at any given time in the past as priors for present inference, via recursion. The (r)MSPRT also contains the sequential probability ratio test when N = 2, and its continuous-time equivalent, the popular drift-diffusion model (e.g. [4, 6, 9, 61, 63–66]). While a valuable basic model of decision-making, the drift-diffusion model is restricted to N = 2 alternatives and does not address neural mechanisms. First, it assumes that evidence for decisions comes as a continuous Gaussian process whose presence in the brain is unproven. Since the decision times predicted by the model critically hinge on this process and its statistics (typically disconnected from the statistics of sensory neural activity), this limitation also obscures the interpretation of the drift-diffusion model’s behavioural predictions. Second, its single decision variable must restrict itself to the half-plane closest to the choice threshold associated to one of its two hypotheses if such hypothesis is to be chosen; hence, the drift-diffusion model can account for forking dynamics, but not for the preceding ramping observed in experimental data. In contrast, the rMSPRT natively captures decisions among any number of alternatives (N ≥ 2), can explain ramp-and-fork dynamics, and does so using neural evidence for decisions in its natural format: spike-trains with statistics extracted from MT recordings.
Biophysical models that directly address neural implementations of decision making are predominantly based on winner-take-all competition between neurons representing different hypotheses [8, 11–14, 16, 67, 68]. These provide valuable insights into potential mechanisms by which neural circuits can represent and compute decisions, but do not typically make contact with formal inference procedures (see [69]). The studies of [13, 68] are possible exceptions, since they make the analogy between the predictions of their neural-network model and those of exact, Bayes-based inference. Conversely, the rMSPRT shows how a normative decision-making algorithm can account for cortical and subcortical activity. As such, the rMSPRT provides target, exact-inference computations for future biophysical models.
Mapping any formal algorithm to a neural substrate implies proposing assumed computational contributions for the components of the substrate. In mapping the rMSPRT we made two broad classes of assumptions. First, as explained above, that individual substrates implement multiple functions either simultaneously or under different stimulation scenarios (e.g. experimental paradigms). In particular, we assume that during decision-formation the striatum is only required to perform a light-touch, relay-like transformation of its excitatory cortical inputs into inhibitory outputs. This assumption is shared by multiple models of the basal ganglia (e.g. [70, 71]). The similarity between ramp-and-fork patterns of response across neurons in the LIP [3], FEF [4], and striatum [6] during the dot-motion task, is consistent with this (Fig 7a–7d). That said, computational models have shown how the striatum’s intricate microcircuit [72] can give rise to several types of complex responses to simple cortical input, often taking the form of spontaneously appearing neural ensembles [73–75]. Thus a promising avenue for future research is determining if, and how, the dynamics of the striatal micro-circuit can act as a relay-like function during decision formation.
Our second class of assumptions is that the omitted connections into and within the basal ganglia may not contribute to the computations essential to inference with cortical inputs. Of note, we have omitted in our mapping the projections from thalamus to striatum [76] or to subthalamic nucleus [77], as well as the intrinsic connections from subthalamic nucleus or from globus pallidus pars externa (globus pallidus in non-primates) to striatum (e.g. see [77, 78]). Such omitted connections might offer a more robust implementation of inference computations, or may contribute to overcoming the limitations of implementing an algorithm with neurons.
Demonstrating the compatibility of anatomical pathways with the mapping of the (r)MSPRT is the subject of ongoing research. Success has been achieved in the expansion of the basal-ganglia mapping of the MSPRT to include the pathway from striatum to globus pallidus pars externa and that from the latter to SNr, where the same inference could be done without those pathways [17]. It has also been recently shown that the pallido-striatal connection is compatible with the MSPRT mapping onto the basal ganglia [21], possibly giving a more robust neural implementation. Both results carry to the rMSPRT. In the same bracket is our inclusion of the cortico-thalamic projection here (Fig 5). Since this projection is assumed to be hypothesis-independent (Eq 12), it does not affect the inference done by the rMSPRT. Similar exercises may be able to account for projections from thalamus to striatum or to subthalamic nucleus, and from the latter to striatum, though these are beyond the scope of this study. The rMSPRT provides a starting point to explore all such extended mapping alternatives.
We sought to characterize the neural mechanism that underlies decisions using a normative algorithm—the rMSPRT—as a framework. We find it remarkable that, starting from data-constrained spike-trains, our monolithic statistical test can simultaneously account for much of the anatomy, behaviour, and electrophysiology of decision-making. While it is not plausible that the brain implements exactly a specific algorithm, our results suggest that the essential composition of its underlying decision mechanism includes the following. First, that the mechanism is probabilistic in nature—the brain utilizes the uncertainty in neural signals, rather than suffering from it. Second, that the mechanism works entirely ‘on-line’, continuously updating representations of hypotheses that can be queried at any time to make a decision. Third, that this processing is distributed, recursive, and parallel, producing a decision variable for each available hypothesis. And fourth, that this recursion allows the mechanism to adapt to the observed statistics of the environment in an unsupervised manner, as it can re-use updated probabilities about hypotheses as priors for upcoming decisions. With the currently available range of experimental studies giving us local snapshots of cortical and subcortical activity during decision-making tasks, the rMSPRT shows us how, where, and when these snapshots fit into a complete inference procedure.
Behavioural and neural data was collected in three previous studies [3, 6, 24], during two versions of the random dot motion task (Fig 1a–1c). Detailed experimental protocols can be found in each report. Below we briefly summarize them.
For comparability across databases, we only analysed data from trials with coherence levels of 3.2, 6.4, 12.8, 25.6, and 51.2%, unless otherwise stated. We used data from all neurons recorded in such trials. Our datasets contained between 189 and 213 visual-motion-sensitive MT neurons (see Table 1; single-cell recordings from [24, 79]), as well as 19 LIP neurons (data from [3]) and 48 striatal ones (from [6]) whose activity was previously determined to be choice- and coherence-modulated. The behavioural data analysed was that associated to LIP recordings. For MT, we analysed the neural activity between the onset and the vanishing of the stimulus. For LIP and striatum we focused on the period between 100 ms before stimulus onset and 100 ms after saccade onset.
To estimate moving statistics of neural activity we first computed the spike count over a 20 ms window sliding every 1 ms, per trial. The moving mean firing rate per neuron per condition was then the mean spike count over the valid bins of all trials divided by the width of this window; the standard deviation was estimated analogously. LIP and striatal recordings were either aligned at the onset of the stimulus or of the saccade; after or before these (respectively), data was only valid for a period equal to the reaction time per trial. The population moving mean firing rate is the mean of single-neuron moving means over valid bins; analogously, the population moving variance of the firing rate is the mean of single neuron moving variances. For clarity, population statistics were then smoothed by convolving them with a Gaussian kernel with a 10 ms standard deviation. The resulting smoothed population moving statistics for MT are in Fig 1d and 1e. LIP and striatal mean firing rates are plotted only up to the median reaction time plus 80 ms, per condition.
Analogous procedures were used to compute the moving mean of the computations within simulated algorithms, per time step, rather than over a moving window. These are shown up to the median of termination observations plus 3 time steps.
Let x(t) = (x1(t), …, xC(t)) be a vector random variable composed of scalar observations, xj(t), made in C channels at time t ∈ {1, 2, …} (right-hand side of Fig 9). Let also x(r: t) = (x(r)/n, …, x(t)/n) be the sequence of vectors x(t)/n, i.i.d. across time, from r to t (r < t). Here n ∈ { R > 0 } is a constant data scaling factor. If n > 1, it scales down incoming data, xj(t); this will prove useful ahead when tuning the algorithm to reveal that the dynamics in rMSPRT computations match those of sensorimotor cortex. Note that scaling is only effective from the likelihood on and does not affect the time interpretation of the data. Crucially, since n is hypothesis-independent, it does not affect inference.
There are N ∈ {2, 3, …} alternatives or hypotheses about the uncertain evidence, x(1: t)—say possible courses of action or perceptual interpretations of sensory data. The task of a decision maker is to determine which hypothesis Hi (i ∈ {1, …, N}) is best supported by this evidence as soon as possible, for a given level of accuracy. To do this, it requires to estimate the posterior probability of each hypothesis given the data, P(Hi|x(1: t)), as formalized by Bayes’ rule. The mechanism we seek must be recursive to match the nature of the brain circuitry. Formally, P(Hi|x(1: t)) will be initially computed upon starting priors P(Hi) and likelihoods P(x(1: t)|Hi); however, after some time Δ ∈ {1, 2, …}, it will re-use past posteriors, P(Hi|x(1: t − Δ)), Δ time steps ago, as priors, along with the likelihood function P(x(t − Δ + 1: t)|Hi) of the segment of x(1: t) not yet accounted by P(Hi|x(1: t − Δ)). A mathematical induction proof of this form of Bayes’ rule follows.
If say Δ = 2, in the first time step, t = 1:
P ( H i | x ( 1 ) / n ) = P ( x ( 1 ) / n | H i ) P ( H i ) P ( x ( 1 ) / n ) (1)
By t = 2:
P ( H i | x ( 2 ) / n , x ( 1 ) / n ) = P ( x ( 2 ) / n , x ( 1 ) / n | H i ) P ( H i ) P ( x ( 2 ) / n , x ( 1 ) / n )
Note that we are still using the initial fixed priors P(Hi). Now, for t = 3:
P ( H i | x ( 3 ) / n , x ( 2 ) / n , x ( 1 ) / n ) = P ( x ( 3 ) / n , x ( 2 ) / n , x ( 1 ) / n | H i ) P ( H i ) P ( x ( 3 ) / n , x ( 2 ) / n , x ( 1 ) / n ) (2)
According to the product rule, we can segment the probability of the sequence x(1: t) as:
P ( x ( 1 : t ) ) = P ( x ( t − Δ + 1 : t ) , x ( 1 : t − Δ ) ) = P ( x ( t − Δ + 1 : t ) | x ( 1 : t − Δ ) ) P ( x ( 1 : t − Δ ) ) (3)
And, since x(t) are i.i.d., the likelihood of the two segments is:
P ( x ( 1 : t ) | H i ) = P ( x ( t - Δ + 1 : t ) | H i ) P ( x ( 1 : t - Δ ) | H i ) (4)
If we substitute the likelihood in Eq 2 by Eq 4, its normalization constant by Eq 3 and re-group, we get:
P ( H i | x ( 3 ) / n , x ( 2 ) / n , x ( 1 ) / n ) = ( P ( x ( 3 ) / n , x ( 2 ) / n | H i ) P ( x ( 3 ) / n , x ( 2 ) / n | x ( 1 ) / n ) ) ( P ( x ( 1 ) / n | H i ) P ( H i ) P ( x ( 1 ) / n ) )
It is evident that the rightmost factor is P(Hi|x(1)/n) as in Eq 1. Hence, in this example, by t = 3 we start using past posteriors as priors for present inference as:
P ( H i | x ( 3 ) / n , x ( 2 ) / n , x ( 1 ) / n ) = P ( x ( 3 ) / n , x ( 2 ) / n | H i ) P ( H i | x ( 1 ) / n ) P ( x ( 3 ) / n , x ( 2 ) / n | x ( 1 ) / n )
So, in general:
P ( H i | x ( 1 : t ) ) = { P ( x ( 1 : t ) | H i ) P ( H i ) P ( x ( 1 : t ) ) for t ≤ Δ P ( x ( t - Δ + 1 : t ) | H i ) P ( H i | x ( 1 : t - Δ ) ) P ( x ( t - Δ + 1 : t ) | x ( 1 : t - Δ ) ) for t > Δ (5)
where the normalization constants are
P ( x ( 1 : t ) ) = ∑ j = 1 N P ( x ( 1 : t ) | H j ) P ( H j ) P ( x ( t - Δ + 1 : t ) | x ( 1 : t - Δ ) ) = ∑ j = 1 N P ( x ( t - Δ + 1 : t ) | H j ) P ( H j | x ( 1 : t - Δ ) )
Eq 5 is a general recursive form of the Bayes’ rule, designed to accumulate evidence for inference in a recurrent, uninterrupted fashion. By t > Δ, it uses posteriors Δ ≥ 1 time steps in the past as current priors, thereby generalizing a previous common recursive form of the Bayes’ rule that is limited to Δ = 1 (that in e.g. [18, 26, 68, 80, 81]). Priors updated in this manner are a sufficient statistic of all the evidence observed up to t − Δ. By this ability, and in the general machine-learning sense, any decision algorithm harnessing Eq 5 adapts or learns. Since no labelled examples or teaching signals are required for such learning, the rMSPRT is thence said to be engaged in ongoing unsupervised learning.
Ahead we use three key results from [20] as part of our methods, with no overlap between their results and the results of the present study. First, a lognormal-based form of the likelihood function whose component operations they showed are neurally plausible and most consistent with the statistics of MT responses during the random dots task. Second, a crucial link between the statistics of ISIs in the spike-trains used as evidence for decision (e.g. those of MT during the dots task), and continuously-distributed MSPRT decision times. As discussed below, this link enabled us to use simpler, discrete-time algorithms and still interpret their behavioural predictions in continuous time. And third, the fundamental dependence of MSPRT decision times on: (a) the discrimination information available in the evidence and (b) a constant, fixed for given error rate and N. Since rMSPRT performs identically to MSPRT, all this carries to it.
It is apparent that the critical computations in Eq 5 are the likelihood functions. The forms that we consider ahead build upon the simplest shown by [20], where the number of evidence streams equals the number of hypotheses (C = N); for instance, a minimum of C = 2 differently-tuned neurons are assumed to provide evidence for a N = 2 choice decision. As discussed by them, more complex (C > N), biologically-plausible likelihood functions can be formulated if necessary; the C < N case would make no sense as it would imply the testing of redundant hypotheses. Although not essential, to simplify the notation when C = N, from now on data in the channel conveying the most salient evidence for hypothesis Hi will bear its same index i, as xi(j). When t ≤ Δ we have:
P ( x ( 1 : t ) | H i ) = a ( t ) ∏ j = 1 t f * ( x i ( j ) / n ) f 0 ( x i ( j ) / n ) (6)
this is, the likelihood that xi(j)/n was drawn from a distribution, f*, rather than from f0, that is assumed to have originated xk(j)/n (k ≠ i) for the rest of the channels. In Eq 6, a ( t ) = ∏ m = 1 t ∏ k = 1 N f 0 ( x k ( m ) / n ) is a hypothesis-independent factor that does not affect Eq 5 and thus needs not to be considered further.
When t > Δ only the observations in the time window [t − Δ + 1, t] are used for the likelihood function because data before this window is already considered within the fed-back posterior, P(Hi|x(1: t − Δ)). Then, the likelihood function is:
P ( x ( t - Δ + 1 : t ) | H i ) = d ( t ) ∏ j = t - Δ + 1 t f * ( x i ( j ) / n ) f 0 ( x i ( j ) / n ) (7)
where again d ( t ) = ∏ m = t - Δ + 1 t ∏ k = 1 N f 0 ( x k ( m ) / n ) needs not to be considered further.
Now, for our likelihood functions to work upon a statistical structure like that produced by neurons in MT we need to be more specific. Inter-spike intervals (ISI) in MT during the random dot motion task are best described as lognormally distributed [20] and we assume that decisions are made upon the information conveyed by them. Thus, from now on we assume that f* and f0 are lognormal and that they are specified by means μ* and μ0, and standard deviations σ* and σ0, respectively. We can then put together the logarithm of Eqs 6 and 7 as the log-likelihood function, yi(t), substituting the lognormal-based form of it reported by [20]:
yi(t)={g0Δ+g1∑j=1t[ log(xi(j)/n) ]2+g2∑j=1tlog(xi(j)/n)fort≤Δg0Δ+g1∑j=t−Δ+1t[ log(xi(j)/n) ]2+g2∑j=t−Δ+1tlog(xi(j)/n)fort>Δ (8)
with
g 0 = κ 0 2 2 Θ 0 2 - κ * 2 2 Θ * 2 + log ( Θ 0 Θ * ) g 1 = 1 2 Θ 0 2 - 1 2 Θ * 2 g 2 = κ * Θ * 2 - κ 0 Θ 0 2
where κ = log ( μ 2 / σ 2 + μ 2 ) and Θ2 = log(σ2/μ2 + 1) with appropriate subindices *, 0.
The terms g0Δ in Eq 8 are hypothesis-independent, can be absorbed into a(t) and d(t), correspondingly, and thus will not be considered further. As a result of this, the yi(t) used from now on is a “simplified” version of the log-likelihood.
We now take the logarithm of Eq 5, define −log Pi(t) ≡ −log P(Hi|x(1: t)) and substitute the simplified log-likelihood from Eq 8 in the result, giving:
- log P i ( t ) = { - z i ( t ) - log P ( H i ) + log ∑ j = 1 N exp ( z j ( t ) + log P ( H j ) ) for t ≤ Δ - z i ( t ) - log P i ( t - Δ ) + log ∑ j = 1 N exp ( z j ( t ) + log P j ( t - Δ ) ) for t > Δ (9)
Where zi(t) = yi(t) + c(t) and the term c(t) models a hypothesis-independent baseline. Because of its uniformity across all hypotheses, c(t) has no effect on inference. It is defined in detail below.
The rMSPRT itself takes the form:
D ( t ) = { Choose hypothesis i : if - log P i ( t ) = min j ∈ { 1 , … , N } - log P j ( t ) ≤ θ , at t = T Continue sampling : if min j ∈ { 1 , … , N } - log P j ( t ) > θ , (10)
where D(t) is the decision at the discretely distributed time t, θ ∈ (0, −log (1/N)] is a constant threshold, and T is the decision termination time. Alternatively, an individual threshold per hypothesis can be set as {θ1, …, θN}, giving a more general formulation.
According to our mapping of rMSPRT to the cortico-subcortical loops (Fig 5), the sensorimotor cortex baseline, c(t) (Eq 9), delayed with respect of the output of the model basal ganglia, is:
c ( t + δ y b ) = h ( t - δ b u - δ u y ) + l (11)
It houses a constant baseline l and the thalamo-cortical contribution h(t − δbu − δuy), which in turn is the delayed cortical input to the thalamus
h ( t - δ b u ) = w y u ∑ i = 1 N ( z i ( t + δ y b - δ y u ) + log P i ( t - δ b u - δ u y - δ y u ) ) N (12)
Here we have chosen h(t − δbu) to be a scaled average of cortical contributions; nevertheless, any other hypothesis-independent function of them can be picked instead if necessary. It would thus not affect inference and render similar results.
The definitions above introduce two free parameters l ∈ R + and wyu ∈ [0, 1) that have the purpose of shaping the dynamics of the computations within rMSPRT during decision formation. The range of wyu ensures that the value of computations in the cortico-thalamo-cortical, positive-feedback loop is not amplified to the point of disrupting inference in the overall loop. Crucially, since both parameters are hypothesis-independent, none affects inference.
For rMSPRT decisions to be comparable to those of monkeys, they must exhibit the same error rate, ϵ ∈ [0, 1]. Error rates are taken to be an exponential function of coherence (%), s, fitted by non-linear least squares (R2 > 0.99) to the behavioural psychometric curves from the analysed LIP database, including 0, 9, and 72.4% coherence for this purpose. This resulted in:
ϵ = { 0 . 50 exp ( - 0 . 11 s ) , for N = 2 0 . 75 exp ( - 0 . 08 s ) , for N = 4 (13)
Since monkeys are trained to be unbiased regarding choosing either target, initial priors for rMSPRT are set flat (P(Hi) = 1/N for all i) in every simulation. During each Monte Carlo experiment, rMSPRT made decisions with error rates from Eq 13. The value of the threshold, θ, was iteratively found to satisfy ϵ per condition. Decisions were made over data, xj (t)/n, randomly sampled from lognormal distributions specified for all channels by means and standard deviations μ0 and σ0, respectively; the exception was a single channel where the sampled distribution was specified by μ* and σ*. This models the fact that MT neurons respond more vigorously to visual motion in their preferred direction compared to motion in a null direction, e.g. against the preferred. As explained in Fig 9, this effectively simulates macaque MT neural activity during the random dot motion task. The same parameters were used to specify likelihood functions per experiment.
We have defined rMSPRT to operate over a discrete time line; however, the brain operates over continuous time. [20] introduced a continuous-time generalization of MSPRT that uses spike-trains as inputs for decision. Thence, the length of ISIs is random and their sum up until decision is, by definition, a continuously distributed time. With all other assumptions equal, they demonstrated that, as an average, the traditional discrete-time MSPRT requires about the same number of observations to decision (discretely distributed), as the maximum number of ISIs among input channels, required by the more general spike-based MSPRT (also discretely distributed yet occurring over continuous time; Fig 9). This has two key implications. First, that continuous-time spike-trains can be substituted as decision evidence for (r)MSPRT by discrete-time stochastic processes—like x(r: t) here—as long as their distributions and the statistics that specify them remain equal; with this we gain efficiency on the implementation of discrete- versus continuous-time algorithms in digital computers, as well as simplicity on their analysis and interpretation. Second, and most important to compare the rMSPRT’s performance to experimentally-measured behaviour, that the (discretely distributed) number of observations to decision, T, in (r)MSPRT has an interpretation as continuously-distributed time. In brief, simulating decision evidence in discrete-time for (r)MSPRT as defined here is a simpler, equivalent way to simulate decisions made on the basis of continuous-time spike-trains. In light of this, the expected decision sample size for correct choices, 〈T〉c, required by the (r)MSPRT, can be interpreted as the mean decision time
τc=(〈T〉c+0.5)μ*n (14)
predicted by the more general continuous-time, spike-based MSPRT, where μ*n is the mean ISI produced by a MT neuron whose preferred motion direction was matched by the stimulus and was thus firing the fastest on average (Fig 9). When the mean firing rate to a preferred characteristic of the stimulus is larger than that to a non-preferred one (μ* < μ0)—as in MT [24], middle-lateral, and anterolateral auditory cortex [66]—the hypothesis selected in error trials is that misinformed by channels with mean μ0n which intuitively happened to fire faster than those whose mean was actually μ*n. Hence, the mean decision time predicted by rMSPRT in error trials would be:
τe=(〈T〉e+0.5)μ0n, (15)
where 〈T〉e is the mean decision sample size for error trials. An instance of rMSPRT capable of making choices upon sequences of spike-trains is straightforward from the formal framework above and that introduced by [20]; nevertheless, as said, for simplicity here we choose to work with the discrete-time rMSPRT. After all, thanks to Eqs 14 and 15 we can still interpret its behaviour-relevant predictions in terms of continuous time. These are used to compute decision times throughout.
We outline here how we use the monkeys’ reaction times on correct trials and the properties of the rMSPRT, to estimate the amount of discrimination information lost by the animals. That is, the gap between all the information available in the responses of MT neurons, as fully used by the rMSPRT (parameter set Ω), and the fraction of such information actually used by monkeys.
The expected number of observations to reach a correct decision for (r)MSPRT, 〈T〉c, depends on two quantities. First, the mean total discrimination information required for the decision, I(ϵ, N), that depends only on the error rate, ϵ, and N. Second, the ‘distance’ between distributions of ISIs from MT neurons that are simultaneously contributing evidence for decision, while visual motion matches the tuning of some and not others (e.g. red versus black in Fig 9). This distance is the Kullback-Leibler divergence from f* to f0 K = ∫ x f * ( x ) log 2 ( f * ( x ) f 0 ( x ) ) d x
which measures the discrimination information available between the distributions. Using these two quantities, the decision time in the (r)MSPRT is [20]:
〈 T 〉 c ≥ I ( ϵ , N ) K , (16)
The product of our Monte Carlo estimate of 〈T〉c in the rMSPRT (Fig 3a in the Results) and K from the MT ISI distributions (Fig 1f), gives an estimate of the limit I(ϵ, N) in expression 16, denoted by I ^ ( ϵ , N ).
The ‘mean decision sample size’ of monkeys—hence the superscript m—within this framework corresponds to 〈 T ^ 〉 c m = ( τ ^ c m / μ * n ) - 0 . 5 (from Eq 14). Here, τ ^ c m is the estimate of the mean decision time of monkeys for correct choices, per condition; that is, the reaction time from Fig 3a minus some constant non-decision time. With 〈 T ^ 〉 c m and I ^ ( ϵ , N ), we can estimate the corresponding discrimination information available to the monkeys in this framework as K ^ m = I ^ ( ϵ , N ) / 〈 T ^ 〉 c m (from expression 16).
Fig 3b compares K (red line) to K ^ m (blue/green lines and shadings) for monkeys, using non-decision times in a plausible range of 200–300 ms. Fig 3c shows the discrimination information lost by monkeys as the percentage of K, [ 1 - ( K ^ m / K ) ] × 100 %.
Expression 16 implies that the reaction times predicted by rMSPRT should match those of monkeys if we make the algorithm lose as much information as the monkeys did. We did this by producing a new parameter set that brings f0 closer to f* per condition, assuming 250 ms of non-decision time; critically, simulations like those in Fig 4 will give about the same rMSPRT reaction times regardless of the non-decision time chosen, as long as it is the same assumed in the estimation of lost information and this information-depletion procedure.
An example of the results of information depletion in one condition is in Fig 3d. We start with the original parameter set extracted from MT recordings, Ω (‘preferred’ and ‘null’ densities in Fig 3d), and keep μ* and σ* fixed. Then, we iteratively reduce or increase the differences |μ0 − μ*| and |σ0 − σ*| by the same proportion, until we get new parameters μ0 and σ0 that, together with μ* and σ*, specify preferred (‘preferred’ in Fig 3d) and null (‘new null’) density functions that bear the same discrimination information estimated for monkeys, K ^ m; hence, they exactly match the information loss in the solid lines in Fig 3c. Intuitively, since the ‘new null’ distribution in Fig 3d is more similar to the ‘preferred’ one than the ‘null’, the Kullback-Leibler divergence between the first two is smaller than that between the latter two. The resulting parameter set is dubbed Ωd and reported in Table 1. Note that this is not a fitting procedure, which would be merely descriptive. Instead, information depletion takes advantage of the properties of the (r)MSPRT to describe the data, but also to predict that the longer decision times of monkeys are explained by a reduction in the discrimination information in the streams of decision evidence.
The slight deviation of the mean reaction times of (r)MSPRT vs those of monkeys in Fig 4a stems from the expression 16 being an inequality. Due to this, I ^ ( ϵ , N ) is a likely over-estimate of I(ϵ, N). Dividing I ^ ( ϵ , N ) by 〈 T ^ 〉 c m hence gives an over-estimate of the monkey discrimination information, K ^ m. If then rMSPRT uses statistics consistent with this over-estimated K ^ m, it renders under-estimated reaction times. This residual discrepancy can be corrected by further multiplying K ^ m, per condition, by the corresponding ratio of the decision time of the model over that of the monkey, from Fig 4a. Repeating the simulations with the implied parameter set would trivially render rMSPRT reaction times that more exactly match those of monkeys. This will likely carry with it a better match in error trials, which is unconstrained in the procedure. Nonetheless, this exercise gives us the information loss associated to such enhanced match, shown in Fig 3c as dashed lines for a 250 ms non-decision time (compare to solid lines); this constitutes a further refined measure of the minimum information lost by the animals according to our framework.
|
10.1371/journal.pgen.1004697 | Approximation to the Distribution of Fitness Effects across Functional Categories in Human Segregating Polymorphisms | Quantifying the proportion of polymorphic mutations that are deleterious or neutral is of fundamental importance to our understanding of evolution, disease genetics and the maintenance of variation genome-wide. Here, we develop an approximation to the distribution of fitness effects (DFE) of segregating single-nucleotide mutations in humans. Unlike previous methods, we do not assume that synonymous mutations are neutral or not strongly selected, and we do not rely on fitting the DFE of all new nonsynonymous mutations to a single probability distribution, which is poorly motivated on a biological level. We rely on a previously developed method that utilizes a variety of published annotations (including conservation scores, protein deleteriousness estimates and regulatory data) to score all mutations in the human genome based on how likely they are to be affected by negative selection, controlling for mutation rate. We map this and other conservation scores to a scale of fitness coefficients via maximum likelihood using diffusion theory and a Poisson random field model on SNP data. Our method serves to approximate the deleterious DFE of mutations that are segregating, regardless of their genomic consequence. We can then compare the proportion of mutations that are negatively selected or neutral across various categories, including different types of regulatory sites. We observe that the distribution of intergenic polymorphisms is highly peaked at neutrality, while the distribution of nonsynonymous polymorphisms has a second peak at . Other types of polymorphisms have shapes that fall roughly in between these two. We find that transcriptional start sites, strong CTCF-enriched elements and enhancers are the regulatory categories with the largest proportion of deleterious polymorphisms.
| The relative frequencies of polymorphic mutations that are deleterious, nearly neutral and neutral is traditionally called the distribution of fitness effects (DFE). Obtaining an accurate approximation to this distribution in humans can help us understand the nature of disease and the mechanisms by which variation is maintained in the genome. Previous methods to approximate this distribution have relied on fitting the DFE of new mutations to a single probability distribution, like a normal or an exponential distribution. Generally, these methods also assume that a particular category of mutations, like synonymous changes, can be assumed to be neutral or nearly neutral. Here, we provide a novel method designed to reflect the strength of negative selection operating on any segregating site in the human genome. We use a maximum likelihood mapping approach to fit these scores to a scale of neutral and negative fitness coefficients. Finally, we compare the shape of the DFEs we obtain from this mapping for different types of functional categories. We observe the distribution of polymorphisms has a strong peak at neutrality, as well as a second peak of deleterious effects when restricting to nonsynonymous polymorphisms.
| Genetic variation within species is shaped by a variety of evolutionary processes, including mutation, demography, and natural selection. With the advent of whole-genome sequencing, we can make unprecedented inferences about these and other processes by analyzing population genomic data. An important goal is to understand the extent to which segregating genetic variants are impacted by natural selection, and to quantify the intensity of natural selection acting genome-wide. Understanding the prevalence of different modes of selection on a genomic scale has wide-ranging implications across evolutionary and medical genetics. For instance, genome-wide association studies (GWAS) are searching for mutations associated with disease in large samples of humans [1]. Because mutations associated with disease are a priori likely to be deleterious, quantifying the portion of mutations that are deleterious along with their average effects could have significant implications for the design and interpretation of GWAS. Moreover, the ENCODE project has recently claimed that much of the genome is involved in some form of functional activity [2], [3]. However, the extent to which molecular signals identified by this project are actually produced by biological processes affecting fitness has been disputed [4], [5]. Indeed, comparative genomics studies suggest that only a small proportion of the human genome (5–10%) is under purifying selection, based on signals detectable on phylogenetic timescales [6]–[8]. Quantifying the DFE in noncoding regions is a first step toward understanding the fitness implications of functional activity at the genomic level.
Traditionally, studies have sought to estimate the distribution of fitness effects (DFE) for nonsynonymous mutations by using summary statistics based on the number of polymorphisms and substitutions [9]–[11] and/or the full frequency spectrum [12]–[14]. These studies typically assumed that synonymous variation is neutral or under weak selection. Many of these analyses suggest that while a large proportion of nonsynonymous mutations are nearly neutral, there is also a significant probability that an amino acid changing mutation will be strongly deleterious. While these studies were limited to analysis of protein-coding genes, recent work has focused on quantifying the DFE in regulatory regions, including short interspersed genomic elements such as enhancers [15], [16] and cis-regulatory regions [17]. Reviews of these and other approaches can be found in ref. [18], [19].
There are several obstacles to quantifying the DFE of new or segregating mutations genome-wide. First, inferences about the DFE are confounded by demography [20]. For example, a high proportion of low frequency derived alleles is a signature of negative selection, but can also be caused by recent population growth [21]. Hence, a well-supported demographic model must be used to appropriately control for population history when inferring the DFE. Second, most current methods rely on dividing up polymorphisms into either putatively selected sites or putatively neutral (or less selected) sites (for example, nonsynonymous and synonymous sites, respectively). These studies have relied on fitting a demographic model to the neutral class of sites and then fitting the DFE of new mutations to a probability distribution, typically an exponential or gamma distribution [9], [13] to the class of sites that are putatively under selection (e.g. nonsynonymous sites). While flexible, these distributions may miss some important features of the DFE [22]. For example, mutation accumulation experiments suggest that the DFE may be bimodal for at least some species, with most mutations either having nearly neutral or strongly deleterious effects, and very few mutations falling in between [23], [24]. Thus, assuming two classes of sites may not serve to capture all the relevant information about the DFE (but see [25] for an example of fitting a multimodal DFE to population genetic data and [22], [26] for nonparametric approaches to estimating the DFE of new amino-acid changing mutations). Finally, previous studies have been restricted to analyzing specific subclasses of mutations (e.g. nonsynonymous, enhancers, etc.) because until recently, no single metric existed that could serve to compare the disruptive potential of any type of variant, regardless of its genomic consequence.
Recently, Kircher et al. [27] developed a method to synthesize a large number of annotations into a single score to predict the pathogenicity or disruptive potential of any mutation in the genome. It is based on an analysis comparing real and simulated changes that occurred in the human lineage since the human-chimpanzee ancestor, and that are now fixed in present-day humans. The method relies on the realistic assumption that the set of real changes is depleted of deleterious variation due to the action of negative selection, which has pruned away disruptive variants, while the simulated set is not depleted of such variation. A support vector machine (SVM) was trained to distinguish the real from the simulated changes using a kernel of 63 annotations (including conservation scores, regulatory data and protein deleteriousness scores), and then used to assign a score (C-score) to all possible single-nucleotide changes in the human genome, controlling for local variation in mutation rates. These C-scores are meant to be predictors of how disruptive a given change may be, and are comparable across all types of sites (nonsynonymous, synonymous, regulatory, intronic or intergenic). Thus, they allow for a strict ranking of predicted functional disruption for mutations that may not be otherwise comparable. C-scores are PHRED scaled, with larger values corresponding to more disruptive effects.
Importantly, human-specific genetic variation patterns are not used as input to train the C-score SVM. In this work, we make use of the C-scores to provide a fine-grained stratification of the deleteriousness of variants segregating in modern human populations. We take advantage of the Poisson random field model [28], [29] with a realistic model of human demographic history to fit a maximum likelihood selection coefficient for each C-score, creating a mapping from C-scores to selection coefficients.
To map C-scores to selective coefficients, we obtained allele frequency information from 9 Yoruba (YRI) individuals (18 haploid sequences) sequenced to high-coverage using whole-genome shotgun sequencing as part of a dataset produced by Complete Genomics (CG) [30]. We removed sites that had a Duke Uniqueness 20 bp-mapability score <1 (downloaded from the UCSC Genome Browser, [31]), to avoid potential errors due to mismapping or miscalling in regions of the genome that are not uniquely mapable.
When inferring the DFE, we focused only on models of neutral evolution and negative selection, because C-scores are uninformative about adaptive vs. deleterious disruption (i.e. a high C-score could either reflect a highly deleterious change or a highly adaptive change). Additionally, because we are using polymorphism data only, positive selection should contribute little to the site-frequency spectrum [32].
We first binned polymorphisms into C-scores rounded to the nearest integer and computed the site frequency spectrum for each bin (Figure S1). We then fit the lowest possible C-score (C = 0), presumed to be neutral, to different models of demographic history. We computed the likelihood of the SFS in this bin for a constant population size model, a range of exponential growth models, the model inferred by Tennessen et al. [33] and the model inferred by Harris and Nielsen [34] from the distribution of tracts of identity by state (IBS) (Figure S2), and used an EM algorithm to correct for ancestral state misidentification (Figure S3, see Materials and Methods). We find that a model of exponential growth at population-scaled rate = 1 for 13,000 generations is the best fit to the corrected SFS, although the Tennessen model is almost as good a fit (Figure S2).
Using the best-fitting demography, we next fit a range of models with different selection coefficients to the EM-corrected site frequency spectrum for each C-score bin, using maximum likelihood (Figure 1.A) (see Methods). We restricted to C≤40, because very few sites have larger C-scores, and hence estimates of the selection coefficients for those C-scores are unreliable. We tested the robustness of the mappings to different levels of background selection, by partitioning the data into deciles of B-scores [35] and re-computing the C-to-s mapping for each decile. We observe that the mapping is generally robust to background selection, with substantial differences only observed at the lowest two B-score deciles, which correspond to high background selection (Figure S4). For this reason, and so as to obtain reliable DFEs at exonic sites (where background selection is generally higher than in the rest of the genome), we also performed a neutral demographic fitting and a C-to-s mapping while restricting only to sites in the exome (Figure 1.C). This mapping has a steeper decline than the genomic mapping, reflecting patterns of background selection which are not fully controlled by C-scores but that affect the SFS. We therefore show estimated DFEs using both the genome-wide and the exome-wide fittings below. After removing the C-score bins that best fit the neutral model, we fit a smoothing spline with 20 degrees of freedom to the remaining C-scores, so as to produce a continuous mapping of C-scores to selection coefficients (Figure 1.A).
We were concerned that our binning-based mapping would miss important features about the distribution of coefficients within each bin. To address this, we also fitted individual gamma distributions of selection coefficients to each of the bins. We show the mean, standard deviation (SD) and ancestral misidentification rate of each gamma fitting in Figure S3. The shape of the fitted gammas vary from an L-shape (Mean/SD <1) at low C bins, to a shape resembling a skewed normal distribution at intermediate C bins (Mean/SD>1) to a shape resembling an exponential distribution at high C bins (Mean/SD≈1) (Figure 1.D). We performed a likelihood ratio test comparing the gamma model to the single-coefficient model, and find that only 4 out of the 40 bins (containing only 0.5% of all polymorphisms and 4.7% of nonsynonymous polymorphisms) are significantly supportive of the gamma model (Figure 1.E). A chi-squared test of the fit to the data shows both models perform similarly well, though both result in significant chi-squared scores at low C-score bins when using the genome-wide data (Figure 1.F). This also occurs if we use the human demography model from [33] (Figure S6). We attribute this to the large amount of data present in those bins as well as complex details of demographic history that affect neutral sites but that we did not model in our neutral fitting. In contrast, when mapping using only the exome, almost all bins have non-significant statistics, suggesting that selection dominates over demography in these regions. Based on these results, we decided to use the smoothed single-coefficient fitting for estimating the DFE in most downstream analyses (Figure 2), although we may be missing a small proportion of within-bin variability. Additionally, we show the inferred DFE of each functional class obtained from the gamma-fitted mapping in Figure S5.
We aimed to test the robustness of the selection coefficient estimates within each bin. We were specifically concerned about highly deleterious bins, which are composed of a smaller number of segregating sites than neutral or nearly neutral bins, and could produce unstable or biased estimates. We obtained bootstrapped confidence intervals for each bin and observe that the mappings are relatively stable up to C = 35 (Figure 1.A). As expected, the standard deviation of the bootstrap estimates is strongly negatively correlated with the sample-size per bin (Figure S7, Pearson correlation coefficient = −0.89). Thus, most of the increase in the width of the confidence intervals observed at higher C-score bins can be explained by the small number of polymorphisms available in those bins, and is likely not the result of other unaccounted processes, such as positive selection, operating exclusively on highly scored polymorphisms. To verify that our mapping was not an artifact of the different number of C-scores within each bin, we also performed 100 randomizations of the C-score assignments to each SNP in the genome. For each randomization, we re-computed the C-to-s mapping. When doing so, the bootstrap confidence intervals increase in size with increasing C scores, but the mapping remains flat, as expected (Figure 1.B).
Further, we verified that the mapping did not change considerably when filtering for sites in regions with low CpG density (<0.05), defined as the proportion of CpG dinucleotides in a window of +/− 75 bp around the site [27] (Figure S8.A). This is expected, as the C-score model accounts for differential mutation rates at CpG sites and the resulting scores are generally robust to them [27]. As before, the gamma model is a significantly better fit than the single-coefficient model at only 4 out of the 40 bins (Figure S8.B).
Additionally, we re-mapped the scores using a constant-size model, and verified that the mapping does not change considerably if we assume a different demographic history than the best fit (Figure S9). The mappings are highly similar in shape, with the exception that, because the constant-size model is depleted of singletons relative to the best-fit model, it takes more bins to reach an SFS that is deleterious enough to map to , and so more C-scores map to s = 0.
To test the dependence of our mapping on the choice of score used to estimate selection coefficients, we performed the same fitting procedure on a variety of other conservation and deleteriousness scores (see Methods). We note, however, that all of these scores are included as input in the C-score SVM. Figure S10 shows that the shape of the mapping is fairly consistent across different choices of scores, except for highly deleterious bins, which contain very few sites. When comparing different categories of sites in the Results, we show their distribution of selection coefficients obtained from the C-score mapping, as this score has been shown to be a better correlate to functional disruption than all the other scores mentioned above, and also controls for mutation rate variation across the genome, while other scores do not [27]. Additionally, we plotted the mapped density of selection coefficients for each functional category, using each of the other scores (with smoothing bandwidth = 0.000005 in Figure S11, 0.0000025 in Figure 3 and 0.000001 in Figure S12). We observe that, while all scores easily distinguish genic sites, PhastCons scores have difficulty distinguishing between synonymous and nonsynonymous sites, which we believe is due to PhastCons scores being regional, rather than position-specific scores. Additionally, while bimodality at nonsynonymous sites is most prominent when using C-scores, it also is apparent in other position-specific scores when plotting the density with a fine smoothing bandwidth. Below, we focus on the DFE obtained from C-scores, but draw comparisons with other DFEs to verify the robustness of particular patterns across annotations.
Using the C-score-to-selection coefficient mapping, we obtained the DFE of segregating polymorphisms in Yoruba individuals. This distribution is highly peaked when all polymorphisms are considered (Figure 2, black dashed line), with a large proportion of neutral changes and a long tail of deleterious mutations, as has been observed before when estimating the DFE of coding sequences [9], [11]–[13], [20]. Interestingly, we observe a pronounced drop in frequency for values of . We note that this is not due to our capping our mapping at as the selection coefficients we are able to map are of a greater magnitude than this drop (Figure 1, S13).
We partitioned the data by the genomic consequence of the polymorphisms, using the Ensembl Variant Effect Predictor (v.2.5) [36]. Some classes exhibit a peak of highly deleterious changes for . This peak results in a bimodal distribution that is especially pronounced for nonsynonymous sites (Figure 2, top row, red line), and is almost non-existent for intergenic sites (Figure 2, top row, pink line). Other types of polymorphisms—like splice site, synonymous, 3′ UTR, 5′ UTR and regulatory mutations—have less deleterious peaks than the one observed at nonsynonymous polymorphisms (Figure 2, top row). The C-to-s mapping computed from all genome-wide sites differs from the C-to-s mapping computed from exonic sites only (Figure 1.C), which is likely due to C-scores not being able to fully account for differences in conservation and background selection in the exome (Figure S4). To correct for this, we also computed DFEs obtained from the exome mapping (Figure 2, middle row). Here, bimodality is weaker (though still present) at putatively functional sites. Finally, we plotted a “hybrid” set of DFEs where we show DFEs for exonic sites (nonsynonymous, synonymous, splice sites) computed from the exome-wide mapping and DFEs for non-exonic sites (UTR, regulatory, intergenic) computed from the genome-wide mapping (Figure 2, bottom row).
We can compare the selection coefficient distributions to the distributions of unmapped C-scores (Figure S13) which are much less tightly peaked at intermediate C-score values and do not show a sharp decrease in density for high values, as does the s distribution in Figure 2. We show various statistics calculated on each of the selection coefficient distributions in Table 1 with the genome-wide mapping and in Table S1 with the exome-wide mapping.
Next, we partitioned the data by whether the polymorphisms were found in the GWAS database [37] or not (Figure S14, Tables 1, S1). While we observe a second deleterious peak among the GWAS SNPs as well, these SNPs seem to be highly enriched for neutral polymorphisms.
Finally, we classified polymorphisms by different regulatory categories. We used two regulatory tracks. First, we partitioned the genome by regulatory regions identified by RegulomeDB [38], which predicts regulatory activity in noncoding regions based on different types of experimental evidence (Figure S15, Tables 1, S1). Second, we used the Segway regulatory segment tracks [39], which are the product of an unsupervised pattern discovery algorithm that serves to identify and label regulatory regions along the genome, including genic regions (Figure 4, Tables 1, S1).
The distribution of fitness effects (DFE) describes the proportion of mutations with given selection coefficients. Knowledge of the DFE has profound implications for our understanding of evolution and health. We infer a highly peaked distribution for all polymorphisms, with a drop in density at , which may be the cutoff between weakly deleterious mutations that segregate in human populations and highly deleterious mutations that are easily pruned away by negative selection.
Our inferred non-synonymous distribution is bimodal and looks very similar to the one obtained for nonsynonymous mutations in Drosophila in ref. [11], with a peak at neutrality and another peak at . Several experimental studies have also shown that non-synonymous non-lethal mutations tend to have a multimodal DFE in model organisms [40], [41] (see ref. [18] for a comprehensive review). We note that it is impossible to obtain such kinds of distributions using a gamma or lognormal probability distribution unless one approximates bimodality by assuming a second, separate class of nonsynonymous mutations that are completely neutral and do not follow the best-fitting probability distribution [11], [13], [20], [25].
We also tested the precision of the C-scores by fitting gamma distributed DFEs to each C-score bin. While only very few bins were fit by a highly peaked gamma distribution (Figure 1.D), the increased variation offered by the gamma distribution rarely improved the likelihood significantly (Figure 1.E). This suggests that the C-scores are precise quantifications of negative selection, and that mutations with similar C-scores are likely to have similar selection coefficients.
Interestingly, we found that for many low C-score bins, a chi-squared test would reject the fit of the demographic model to the data. This is possibly because these low C-score bins have a significant number of sites, and hence subtle features of the demography and quality control are relevant. Nonetheless, we note that for moderate-to-high C-score bins and for exonic data, we were not able to reject the fit of the predicted site frequency spectrum from the data. While these bins have fewer sites, it is also likely that stronger selection is obscuring some of the signal of subtle demographic events.
Our novel expectation-maximization approach to quantifying ancestral state misidentification allows us to assess differential misidentification rates across C-score categories. Ancestral state misidentification occurs because a site is hit by more than one mutation, hence obscuring the identity of the ancestral allele. Here, we found that low C-score bins are enriched with ancestral state misidentification, with rates in excess of 5% for some bins (Figure S3). This may reflect a bias induced by the C-scores themselves, because C-scores are trained to distinguish classes of sites that have relatively few substitutions between humans and chimpanzees. Because the signal of ancestral state misidentification and positive selection are very similar [42], it is possible that low C-score bins are enriched for positive selection, although we did not pursue that direction any further. For larger C-score bins, we infer misidentification rates along the lines of those obtained in simulation studies by ref. [42].
Importantly, unlike previous studies, we also obtain DFEs for other types of mutations, including synonymous, splice site, 3′ UTR, 5′ UTR and regulatory polymorphisms, which exhibit bimodality to a lesser degree than the nonsynonymous DFE. In particular, 5′ UTR changes constitute the category with the smallest proportion of neutral or nearly neutral () polymorphisms after nonsynonymous changes, likely reflecting selection on gene regulation upstream of coding sequences. Futhermore, distributions corresponding to mutations in UTR and regulatory regions have a less pronounced trough between the two peaks than the ones observed among coding mutations, suggesting that the magnitude of deleterious effects is more uniformly distributed in non-coding regions. In contrast, missense mutations appear to have more of an “all-or-nothing” effect, as would perhaps be expected when replacing an amino acid inside a protein.
Our method does not use synonymous sites as a neutral benchmark, as do other studies [9], [11], [20]. In fact, our inferred DFE suggests that a considerable number of synonymous polymorphisms may not be neutral, as we observe a second deleterious peak in them too, albeit less deleterious than the one observed at nonsynonymous polymorphisms. We caution, however, that this second peak is less promient when using an exome-specific mapping (Figure 2) or when using other position-specific scores (Figures S11, 3, S12), which suggests that at least part of this peak may be caused by regional patterns of conservation or background selection in the exomes. Instead, intergenic polymorphisms are the class of sites most likely to evolve neutrally. Because this class is so abundant, most of the signal observed when all polymorphisms are pooled together closely reflects the distribution observed for intergenic polymorphisms (Figure 2).
Our results have implications for GWAS, as we find a high proportion of GWAS SNPs to be neutral or nearly neutral, which could suggest a high rate of false positives in this type of association studies. On the other hand, GWAS studies only aim to find polymorphisms linked to causative variants, and so GWAS SNPs need not have strongly deleterious effects. Alternatively, if the effect size of many GWAS SNPs are sufficiently small, it is possible that many of them are not subject to strong selection.
Additionally, by stratifying our results based on different ENCODE categories, we can elucidate the fitness consequences of molecular activity signals detected by ENCODE [2], [3], [38]. We find the category with the lowest proportion of neutral polymorphisms to be the one corresponding to sites that have eQTL evidence as well as evidence for transcription factor (TF) binding, a matched TF motif, a matched DNase footprint and that are located in a DNase peak. In general, categories that combine many regulatory signals tend to show lower proportions of neutral mutations than those that do not, suggesting that data integration across distinct approaches to detecting selection and functionality is likely to do better than any individual approach [43]. Moreover, this suggests that much of the molecular activity detected by ENCODE may not have significant fitness consequences.
Stratification by Segway regions allows us to look at a different aspect of regulatory activity in the genome. Interestingly, we observe that polymorphisms in Transcription Start Sites (TSS) are the ones containing the largest proportion of deleteriousness. This echoes results from analyses of variation at transcription factor binding sites, which have been found to be under stronger constraint when found near TSS than when found far from them [44]. Other regions with high proportions of deleterious polymorphisms include Gene Body (Start), strong CTCF and Enhancer/Gene Middle. This suggests that regions with strong repressor, insulator or enhancer activity, as well as near the start of genes, are particularly important for biological function, perhaps unsurprisingly given our knowledge of the molecular role that these regions play in the regulation of transcription.
DFEs produced from different conservation scores reveal interesting properties about each score (Figures S11, 3). For example, because PhastCons scores are regional and not position-specific, they do not perform well at distinguishing between different classes of genic polymorphisms. Bimodality at nonsynonymous sites is observed to a lesser or greater extent in almost all scores, and it is especially prominent when using C-scores, but bimodality at synonymous sites is only observed in PhastCons scores and C-scores, which suggests it may be caused by regional patterns of background selection. Finally, we note that high PhyloP scores computed from deeper phylogenetic (e.g. Vertebrate) alignments tend to be more deleterious than high PhyloP scores computed from shallower phylogenetic (e.g. Primate) alignments. This likely reflects the higher resolution one can obtain by using deeper alignments to find extremely deleterious sites.
There are several limitations to our method. First, we have restricted ourselves to estimating the DFE of segregating mutations that have reached appreciable frequencies in the population. An extension of this approach would be to infer the DFE of new mutations from the DFE of segregating mutations genome-wide. Second, we assumed no dominance, epistasis or positive selection, which future studies could attempt to incorporate into our approach. In addition, we have assumed sites are independent and have therefore ignored the covariance between linked sites, which likely leads to an underestimatation of confidence intervals obtained from the bootstrapping. The free-recombination assumption may also affect inference due to Hill-Robertson interference between mutations subject to selection [45] as well as linked background selection affecting the SFS of neutral sites in the human genome [35]. This may be a more important issue in our case than other genic-only approaches because we are also including intergenic mutations in our analysis, so the space between analyzed polymorphisms is on average smaller than if we were only looking at coding polymorphisms [20]. We also assume no positive selection. This, however, should not be a major problem, because we are only basing our inferences on polymorphic sites and advantageous mutations contribute little to polymorphism, assuming [32].
One final limitation is that the type of inference performed here is only possible in species from which accurate deleteriousness scores can be obtained, and that it relies on these scores being able to correctly rank sites throughout the genome. As the amount of genomic data increases, new and better scores will likely emerge in the near future for both humans and other species, and so we expect our method could be re-implemented once better proxies for deleteriousness become available.
We used the theory developed by Evans et al. [46] to obtain the expected population site frequency spectrum with non-equilibrium demography. We assume a Wright-Fisher population in the limit of large population size and use diffusion theory to model this process. Writing for the frequency spectrum at frequency x and time where is in units of generations and , we can approximate the dynamics of with genic selection and mutation by solving the following partial differential equation:(1)subject to the boundary condition:(2)where S is the population-scaled selection coefficient (), is the population-scaled mutation rate () and is the effective population size at time relative to the population size at time 0.
We assume N(0) to be 10,000 for all exponential and constant models. For the constant population size model, . For the exponential growth model where is the population-scaled growth rate and is the per-generation growth rate. For models taken from the literature, we use the N(0) assumed by the corresponding paper. For the model of Harris and Nielsen, is piece-wise defined according to their Figure 7. The Tennessen model is similarly defined in a piece-wise fashion according to their Figure 2, although it also includes periods of exponential growth, as opposed to simply being piece-wise constant as in the Harris and Nielsen model.
We solve for numerically in Mathematica, and can then compute the expected number of segregating sites with copies of the derived allele out of a sample of genes. Denoting by the theoretical SFS with selection coefficient s, this quantity is(3)where is the parameterized distribution of selection coefficients. For gamma distributed fits,where and are the shape and rate parameters of the gamma distribution and is the gamma function. For a point mass at ,where is the usual Dirac delta function.
We focused on fitting the shape of the SFS, and hence worked with the probability that a given site in a sample of has copies of the derived allele,(4)
The Mathematica code used for obtaining the theoretical SFS can be found online at: http://malecot.popgen.dk/schraiber/.
We observed that the SFS showed signs of ancestral state misidentification, in particular an excess of high frequency derived alleles (Figure S2). To account for ancestral state misidentification errors, we developed an expectation maximization (EM) algorithm. In the E step, we estimate the posterior probability that a site at frequency out of is misidentified given the current estimated site frequencies, , and the current estimate of the misidentification rate, , as(5)Then, during the M step, we re-estimate the misidentification rate as(6)where is the number of sites at frequency i. We next re-estimate either the demographic parameters or the parameters of the selected model using the log-likelihood(7)to obtain the theoretical SFS for the next iteration, .
The exponential growth model has two free parameters, r, the population-scaled growth rate and t, the total time of exponential growth. We first obtained the site frequency spectrum for all sites with C = 0. Next we solved for the exponential growth model across a grid of and r, as well as the constant, Harris and Tennessen models, and applied our EM algorithm to estimate the best fitting demographic model, using a grid search over demographic models during the M step.
To find the maximum likelihood estimate of for each C-score bin, we first obtained the site frequency spectrum corresponding to each C-score bin. Next, we solved under the fitted demography for in steps of 0.005, along with s = 0. To obtain an estimated SFS under the assumption of gamma distributed selection coefficients, we used the trapezoid rule to numerically integrate against a gamma distribution as in formula 3.
We used our EM algorithm to estimate the best fitting selection coefficient for each bin. When fitting a single coefficient, we used a grid search during the M-step, and when fitting gamma distributed selection coefficients, we used the L-BFGS-B algorithm. To plot the DFE, we used kernel density estimation with smoothing bandwith = 0.000005, unless otherwise stated.
Consequences for different types of sites were determined using the Ensembl Variant Effect Predictor (v.2.5) [36]. If more than one consequence existed for a given SNP, that SNP was assigned to the most severe of the predicted categories, following the VEP's hierarchy of consequences. Codon and degeneracy information was obtained from snpEff [47]. Segway segmentation information [39], [48] was obtained from ref. [27] and RegulomeDB categories [38] were obtained from http://www.regulomedb.org/(last accessed: 24th February 2014).
We were concerned that reference/alternative bias in PolyPhen and SIFT – which use humans in their alignments – would lead to strong biases in C-scores, as the C-score method uses these scores in its training set. To mitigate this issue, the C-scores we are using were polarized with respect to the ancestral allele at sites where the reference differs from the ancestral allele, unlike the standard C-scores, which are always polarized with respect to the human reference (Martin Kircher, pers. comm.).
Nevertheless, we aimed to quantify how much bias remained after this correction. To do so, we obtained PhyloP [49] and PhastCons [50] scores derived from vertebrate, mammal and primate alignments, as well as GERP++ rejected substitution (GERP S) scores [51], for all YRI SNPs. All of these scores were calculated using human-free alignments [27]. We compared the bias observed at the C-scores we are using to the bias observed at the human-free conservation scores. We computed the absolute difference between the mean of each score at sites where reference = ancestral and at sites where reference = derived, divided by the total standard deviation at both types of sites. We plotted this standardized absolute difference as a function of the number of derived alleles in YRI (Figure S16). Though we observe some bias in all the scores, C-scores fall within the range of bias of human-free conservation scores and are not more biased than them. We hypothesize this occurs because the fraction of sites in the training set of the C-score SVM for which SIFT and PolyPhen scores are available (i.e. their “relevance” score as defined in Supplementary Table S3 of [27]) is very small (0.0063), as SIFT and PolyPhen are nonsynoymous-specific scores, and not genome-wide scores. In contrast, PhastCons, PhyloP and GERP Scores were all explicitly obtained from human-free alignments [27] and these are the training annotations with the highest area under the ROC curve (AUC) that have Relevance = 1 (i.e. they are genome-wide scores). The sites we used to obtain the C-to-s mapping are genome-wide polymorphisms, so the bulk of the signal comes from these scores. Interestingly, GERP scores show the least amount of bias. C-scores tend to show some bias, but unlike other scores like PhyloP, the bias is low when the number of derived alleles is high, and therefore when the reference is more likely to be derived.
To test how robust the mapping of C-scores to selection coefficients is to different types of conservation scores, we produced DFEs by using selection coefficient mappings from each of the aforementioned conservation scores. We attempted to equalize the range of all scores by PHRED-scaling them, i.e. converting each score to –log10(p) where is the probability of observing a change as or more disruptive/conserved (based on that particular score scale) among all polymorphic YRI sites. We note that this is different from the natural PHRED scale of C-scores (where is the the probability of observing a score as or more disruptive among all possible, but not necessarily realized, mutations in the human genome), and so we also re-scaled the C-scores to produce a fair comparison. Then, we repeated the maximum likelihood mapping for each PHRED-scaled score in bins of 0.25 units (e.g. 0–0.125, 0.125–0.375, 0.375–0.625, etc). It is important to note that PhastCons are regional scores, while PhyloP and GERP S are position-specific scores. Another difference is that PhastCons scores only measure the probability of negative selection, while PhyloP and GERP S scores also measure positive selection (i.e. low/negative scores represent faster evolution than expected purely under drift), which may be why we observe an uptick at the lower end of the mapping for those scores in Figure S10.
|
10.1371/journal.pmed.1002526 | Physical activity levels in adults and older adults 3–4 years after pedometer-based walking interventions: Long-term follow-up of participants from two randomised controlled trials in UK primary care | Physical inactivity is an important cause of noncommunicable diseases. Interventions can increase short-term physical activity (PA), but health benefits require maintenance. Few interventions have evaluated PA objectively beyond 12 months. We followed up two pedometer interventions with positive 12-month effects to examine objective PA levels at 3–4 years.
Long-term follow-up of two completed trials: Pedometer And Consultation Evaluation-UP (PACE-UP) 3-arm (postal, nurse support, control) at 3 years and Pedometer Accelerometer Consultation Evaluation-Lift (PACE-Lift) 2-arm (nurse support, control) at 4 years post-baseline. Randomly selected patients from 10 United Kingdom primary care practices were recruited (PACE-UP: 45–75 years, PACE-Lift: 60–75 years). Intervention arms received 12-week walking programmes (pedometer, handbooks, PA diaries) postally (PACE-UP) or with nurse support (PACE-UP, PACE-Lift). Main outcomes were changes in 7-day accelerometer average daily step counts and weekly time in moderate-to-vigorous PA (MVPA) in ≥10-minute bouts in intervention versus control groups, between baseline and 3 years (PACE-UP) and 4 years (PACE-Lift). PACE-UP 3-year follow-up was 67% (681/1,023) (mean age: 59, 64% female), and PACE-Lift 4-year follow-up was 76% (225/298) (mean age: 67, 53% female). PACE-UP 3-year intervention versus control comparisons were as follows: additional steps/day postal +627 (95% CI: 198–1,056), p = 0.004, nurse +670 (95% CI: 237–1,102), p = 0.002; total weekly MVPA in bouts (minutes/week) postal +28 (95% CI: 7–49), p = 0.009, nurse +24 (95% CI: 3–45), p = 0.03. PACE-Lift 4-year intervention versus control comparisons were: +407 (95% CI: −177–992), p = 0.17 steps/day, and +32 (95% CI: 5–60), p = 0.02 minutes/week MVPA in bouts. Neither trial showed sedentary or wear-time differences. Main study limitation was incomplete follow-up; however, results were robust to missing data sensitivity analyses.
Intervention participants followed up from both trials demonstrated higher levels of objectively measured PA at 3–4 years than controls, similar to previously reported 12-month trial effects. Pedometer interventions, delivered by post or with nurse support, can help address the public health physical inactivity challenge.
PACE-UP isrctn.com ISRCTN98538934; PACE-Lift isrctn.com ISRCTN42122561.
| Brisk walking for 30 minutes or more daily on most days of the week can help adults and older adults to achieve moderate-to-vigorous physical activity (MVPA) guidelines for health benefits, yet many do not achieve these levels.
Previous pedometer-based walking studies have shown positive effects on increased step counts and time in MVPA for up to 12 months.
For sustained health benefits, increased physical activity levels need to be maintained, yet there is a lack of data from interventions assessed using objectively measured physical activity levels beyond 12 months.
We followed up participants from two primary care 12-week pedometer-based walking trials, including both nurse-supported and postal pedometer arms, to establish whether objectively measured physical activity increases seen at 12 months were sustained at 3–4 years.
PACE-UP followed up 45–75-year-olds 3 years post-baseline and showed that both nurse-supported and postal pedometer interventions continued to have higher physical activity levels compared to the control group (approximately an extra 28 and 24 minutes/week, respectively, of MVPA in bouts and an extra 670 and 630 steps/day, respectively).
PACE-Lift results were very similar. In 60–75-year-olds followed up at 4 years post-baseline, the nurse-supported pedometer intervention group spent about 33 minutes/week more time in MVPA in bouts compared to the control group.
These findings suggest that adult and older adult participants receiving 12-week pedometer-based walking interventions, provided either by post or with nurse support, are still doing more physical activity 3–4 years later.
Pedometer interventions can help address the public health physical inactivity challenge.
| Strong evidence exists for the health benefits of physical activity (PA) for a wide range of conditions [1,2]. Physical inactivity leads to high health service costs [1,3] and is the fourth leading risk factor for global mortality [2]. Adult and older adult guidelines advise ≥150 minutes of moderate-to-vigorous PA (MVPA) weekly, or 75 minutes of vigorous PA, or a combination, in ≥10-minute bouts [1,4], but any increase in PA for inactive people is valuable [5]. Many PA interventions, including pedometer-based interventions, increase PA levels in the short term [6–8]. However, long-term health effects require sustained PA changes [1], and evidence for maintenance is lacking. A meta-analysis of PA interventions (including pedometers) in 55–70-year-olds [8] only identified 2 trials with objective PA data beyond 12 months [9,10]. One showed a significant step-count effect 18 months post-baseline, but only 6 months post-intervention [10]; the other showed a significant increase in step count in the lifestyle group 23 months post-baseline, but only 12 months post-intervention [9]. The meta-analysis authors [8] repeated requests made by previous systematic reviews [11,12] and guidelines [13] for trials to be conducted with longer follow-up periods and objective PA measures.
We previously conducted two pedometer-based walking interventions with adults and older adults, which increased step count and MVPA in bouts at 12 months and provided longer-term follow-up opportunities [14,15]. Both trials recruited postally from primary care and delivered 12-week pedometer-based walking interventions incorporating behaviour change techniques (BCTs) through dedicated practice nurse PA consultations (3 in PACE-UP, 4 in PACE-Lift) or by post (PACE-UP only). PACE-Lift nurse consultations additionally provided feedback on accelerometry findings to participants. PACE-UP recruited 1,023 predominantly inactive 45–75-year-olds. Average baseline daily step count was 7,479 (standard deviation [SD]: 2,671) and average time in MVPA in bouts was 94 (SD: 102) minutes/week. PACE-Lift recruited 298 patients aged 60–75 years. Average baseline daily step count was 7,347 (SD: 2,839) and average time in MVPA in bouts was 92 (SD: 108) minutes/week. Despite age-group and intervention differences, both trials and all intervention groups showed increases in step counts of approximately one-tenth and time in MVPA of over one-third between baseline and 12 months [14,15].
The study aim was to follow up both trial cohorts to examine objectively measured PA levels at 3 years in PACE-UP and 4 years in PACE-Lift. Given the different but overlapping age ranges, interventions that were similar but differed in intensity, and different lengths of follow-up, we analysed the two trials separately, using identical methods, and present the results in parallel.
Participants who had not withdrawn from either trial by 12 months were eligible. Practices excluded participants who had died, moved away, or developed a terminal illness or dementia. Eligible participants were sent a trial follow-up letter, participant information sheet, consent form, and freepost return envelope. Researchers telephoned participants to discuss any queries. Those interested returned signed consent forms. Participants and researchers were unmasked to intervention allocation.
Instruments, questionnaire measures, and protocols were the same as during the trial. Participants were not asked to increase their PA levels, just to continue usual activity, and thus health limitations did not preclude participation. Participants were instructed to wear the accelerometer (Actigraph GT3X+) on a belt over one hip for 7 consecutive days, from getting up until going to bed. A diary (to record activities) questionnaire and freepost envelope were provided. If accelerometry recording did not result in ≥5 days with ≥540 minutes/day, participants were asked to re-wear monitors (re-wears were required for 20 PACE-UP and 1 PACE-Lift participants). Participants were posted a £10 gift voucher.
The Actigraph GT3X+ measures vertical accelerations in magnitudes from 0.05 to 2.0 g, sampled at 30 Hz, and then summed over a 5-second epoch time period. It can record PA continuously for up to 21 days. Actigraph data were reduced using Actilife software (V6.6.0), ignoring runs of ≥60 minutes of 0 counts [14,15]. Summary variables were as used in the trials [14,15]: step counts, accelerometer wear time, time spent in total MVPA (≥1,952 counts per minute, equivalent to ≥3 metabolic equivalents), time spent in ≥10-minute bouts of MVPA, and time spent sedentary (≤100 counts per minute, equivalent to ≤1.5 metabolic equivalents). Only days with ≥540 minutes of registered time were used. To lessen attrition bias, main analyses of effect included all subjects with ≥1 satisfactory day of recording at 3 (or 4) years.
Outcomes focussed on changes between baseline measures and follow-up measures at 3 years (PACE-UP) or 4 years (PACE-Lift). For accelerometry, we analysed: (i) change in average daily step count, (ii) change in time spent weekly in MVPA in ≥10-minute bouts, and (iii) change in weekly sedentary time.
Questionnaire PROMs were as for 3- and 12-month outcomes [16,19]: quality of life [20], exercise self-efficacy [21], pain [22], depression [23, 24], and anxiety [23, 25].
Analysis and reporting followed CONSORT guidelines (S1 Protocol, S2 Protocol). Primary analyses were conducted using STATA version 14.0 (StataCorp), with a two-step process to estimate change. In step 1, average daily step counts at 3 years (PACE-UP) or 4 years (PACE-Lift) were computed from a random-effects model, allowing for day of the week and day of wearing the accelerometer as fixed effects and participant as a random effect. In step 2, average daily step count at 3 years (PACE-UP) or 4 years (PACE-Lift) was regressed on estimated baseline average daily step count, with treatment group, age, gender, practice, and month of baseline accelerometry as fixed effects and household as a random effect in a multilevel model. Identical analyses were carried out for MVPA in ≥10-minute bouts, sedentary time, and wear time. Changes in PROMs were estimated using step 2 only.
Primary analyses used 681 (PACE-UP) or 225 (PACE-Lift) participants who provided accelerometry data at 3 or 4 years, respectively. Sensitivity analyses assessed the effect of missingness: (1) multiple imputation methods were used to impute outcome data for those missing at 3 or 4 years, assuming outcomes were missing at random (MAR), conditional on model variables, and using the STATA procedure mi impute, and (2) missing not at random (MNAR) analyses. The purpose of the MNAR analyses was to assess how extreme the missing data needed to be in order to explain away our positive effect estimates. To do this, we used the Stata module rctmiss (Statistical Software Components [SSC] https://ideas.repec.org/s/boc/bocode.html) [26]. Essentially, the rctmiss programme takes as its starting point MAR estimates for all subjects with missing data. It then adds or subtracts steps to the estimates before re-estimating the treatment effects. Thus, we left the control group missing values at their MAR estimates and first subtracted 500 steps/day from the MAR estimates in the treatment groups; we then took a more extreme scenario, in which we subtracted 1,000 steps/day for those in the treatment groups, again leaving the control group missing values at their MAR values.
Of 1,023 PACE-UP participants, 32 withdrew by 12 months, 2 died before the 3-year follow-up, 1 was excluded, and 681 provided ≥1 day of adequate accelerometry data. The 3-year follow-up rate was 69% (681/988), or 67% (681/1,023) of initial trial participants, the mean age was 59 (SD = 7.9), and 64% (438/681) were female. Of 298 PACE-Lift participants, 15 withdrew by 12 months, 2 died before the 4-year follow-up, and 225 provided ≥1 day of adequate accelerometry data. The 4-year follow-up rate was 80% (225/281), or 76% (225/298) of original trial participants, the mean age was 67 (SD = 4.2), and 53% (120/225) were female. The CONSORT diagram (Fig 1) shows 3- and 4-year follow-up data by randomised groups. Ninety-two percent (625/681) in PACE-UP and 93% (209/225) in PACE-Lift provided ≥5 days of accelerometry data at 3 and 4 years, respectively (S2 Table and S3 Table).
Accelerometry summary measures are shown for the three PACE-UP groups (S2 Table) and two PACE-Lift groups (S3 Table) at each time point. Fig 2 displays effect estimates for different groups from both trials at all time periods for step counts and time in MVPA in bouts, respectively. Table 2 shows these estimates plus sedentary time and wear time in tabular form. At 3 years in PACE-UP, both intervention groups are doing more steps/day than controls, with no significant intervention group differences: postal +627 (95% CI: 198–1,056); nurse +670 (95% CI: 237–1,102). For PACE-Lift, at 4 years, the intervention group is doing more steps/day than the control group, although the difference is not statistically significant: +407 (95% CI: −177–992). For total weekly MVPA in ≥10-minute bouts (minutes/week), PACE-UP 3-year findings compared with control are as follows: postal +28 (95% CI: 7–49); nurse +24 (95% CI: 3–45). For PACE-Lift at 4 years, the intervention group is still doing significantly more MVPA in bouts (minutes/week) than the control group: +32 (95% CI: 5–60). Effect estimates for both steps per day and MVPA were stable when we limited analyses to subjects with at least 4 days of measurement at follow-up (S4 Table).
In PACE-UP, the 3-year treatment effects for steps/day were 98% (postal) (627/642) and 99% (nurse) (670/677), respectively, of the 1-year estimates; in PACE-Lift, the 4-year estimate was 67% (407/610) of the 1-year estimate. For minutes of MVPA in 10-minute bouts, PACE-UP estimates were 85% (postal) (28/33) and 69% (nurse) (24/35), respectively, of the 1-year estimates, while the PACE-Lift estimate was 82% (32/39). Neither PACE-UP nor PACE-Lift showed differences between intervention and control groups at 3 and 4 years for sedentary time or daily wear time (Table 2). A PACE-UP subgroup analysis demonstrated similar effects for steps/day in 45–59- and 60–75-year-olds (S1 Fig).
None of the interventions had significant effects on pain, depression, anxiety, or health-related quality of life at 3 or 4 years, consistent with 3- and 12-month findings (S4 Table). In PACE-UP, there was a persistent exercise self-efficacy effect for the nurse group at 3 years (also seen at 3 and 12 months) but not in PACE-Lift at 4 years (S5 Table).
Table 3 presents sensitivity analyses assuming that missing outcome data were MAR, conditional on a variety of predictors; analyses had little impact on the primary outcome step-count effect estimates and do not change interpretation. For the MNAR analyses, we combined both intervention groups in PACE-UP to increase power and simplify presentation; separate analyses give a similar picture. The MNAR analyses (S2 Fig) make a bigger impact for both trials but only when we assume there is a strong differential departure between the non-random effects in control and treatment groups (see solid lines in S2 Fig). For example, when we assume that the missing data in the treatment groups are 1,000 steps below their MAR values while the control group values are at their MAR values, the treatment effects for PACE-UP are no longer statistically significant; but even then, the confidence interval is still largely positive.
To our knowledge, these are the first population-based pedometer studies showing effects on objectively measured PA levels more than 12 months post-intervention. Compared to controls, intervention participants followed up from both PACE-UP and PACE-Lift trials showed significant increases in MVPA in bouts at 3 and 4 years of approximately an extra 30 minutes weekly, with no difference between intervention groups in PACE-UP (as was also found at 12 months). PACE-UP showed a significant step-count increase of approximately 650 steps/day; PACE-Lift showed a similar but nonsignificant step-count increase. The increases seen in PA levels were similar to those seen at 12 months. No differences were seen in sedentary or wear time.
This work’s main strength is its documentation of longer-term follow-up results beyond 12 months from trials with objective PA data relevant to guidelines. Both trials were based on population-based primary care samples and achieved good follow-up. Sensitivity analyses demonstrated that effect estimates were robust; only extreme assumptions changed interpretation. We presented findings for two trials with overlapping but different age groups and slightly different intervention intensities and follow-up periods. However, the many similarities (recruited postally from primary care; 12-week pedometer-based interventions, including nurse-support arms; accelerometer-assessed main PA outcome measures beyond 12 months) meant there was considerable value in presenting the results together. Age was not an effect modifier in PACE-UP. Despite their differences, both trials show similar consistent long-term increased time in MVPA in bouts for intervention group participants.
The study also had a number of potential limitations. Long-term follow-up data were provided by 76% of PACE-Lift and 67% of PACE-UP original trial participants. Whilst only a small proportion of participants actively withdrew from each trial, reasons for withdrawal were not systematically collected. Whilst losing between a quarter and a third of subjects at follow-up could reduce the generalizability of the findings, we have directly addressed the risk of attrition bias through sensitivity analyses using appropriate imputation methods, and this gave robust results. Participants and researchers were unmasked to group; however, PA outcomes were assessed objectively by accelerometry, and participants were blind to measurements. Participants might have tried harder with PA when monitored, but this would also have affected controls and would have been reduced by using a 7-day data collection protocol [6]. Also, the intervention groups increased their MVPA in ≥10-minute bouts, implying that participants made changes as advised. Whilst the Actigraph accelerometer provides valid estimates of time spent in different intensity levels, including MVPA [27], any waist-mounted activity monitor may underestimate upper body movement, such as weight training and carrying heavy loads [28]; it also underestimates cycling and did not measure swimming. However, crucially, accelerometers are most sensitive to ambulatory activities such as walking, which was the main intervention component of both trials. A further potential limitation is that minimal interventions were offered to both trial control groups after 12-month follow-up. However, this contamination would tend to weaken intervention effects, so the existence of differences in PA levels at 3 and 4 years is an important positive finding and helps us to understand the additional support required for a successful postal intervention.
This paper provides novel, important evidence on sustained effects of pedometer-based walking interventions on objectively measured PA levels. A recent systematic review of the effectiveness of behavioural interventions in increasing PA at 12–36 months [8] identified two studies that provided objectively measured outcomes beyond 12 months [9,10]. We identified two more recent studies using a similar search strategy [29,30]. In reviewing these studies, several issues emerge. First, interventions differed dramatically in duration, intensity, and resources needed—particularly important when considering cost-effectiveness. Second, studies reported follow-up length post-baseline, not post-intervention; maintenance of effects is defined by the latter. None of the four studies provided outcomes more than 12 months post-intervention: one was 6 months post-intervention [10], two were ongoing at the point of assessment [29, 30], and the final one was 12 months post-intervention [9]. Our two studies thus provide the first clear evidence of efficacy for pedometer-based interventions at 33 months and 45 months post-intervention, providing the type of evidence from PA interventions recently called for [8,11,13]. The simplicity of our postal intervention makes it likely to be more cost-effective than more intensive interventions, and the PACE-UP trial cost-effectiveness analyses at 12-month follow-up demonstrated this [14,31].
Our findings support guidance to promote pedometers alongside support for goal setting, self-monitoring, and feedback [32]. However, it is important to consider which factors in pedometer-based interventions are important for success. Both PACE-UP and PACE-Lift included a pedometer, step-count diary, and patient handbook, including BCTs and practice nurse PA consultations [16,19]. Despite PACE-Lift providing a more intensive nurse intervention than PACE-UP, both trials delivered similar effects on PA outcomes at 12 months [14,15] and at 3 or 4 years. Additionally, nurse and postal interventions in PACE-UP achieved similar outcomes at 12 months [14] and 3 years. These findings confirm that shorter, simpler interventions can be equally effective [8,33]. Systematic reviews suggest that individual tailoring, personalised activity goals, and using a step-count diary are important [6,8]; all interventions from both trials provided these elements. That the minimal postal interventions given to both trial control groups at 12 months were not effective at increasing PA levels suggests that the additional support given to the original PACE-UP trial postal arm (follow-up telephone call after a week and encouragement to return completed PA diary after 3 months) was important for this group’s success. The original postal group also had step-count targets set based on baseline blinded pedometer use and received the intervention when they had just been recruited to the trial, so they may have been more motivated. These factors may also have been important to the trial postal intervention’s success. PA guidelines stress the importance of increasing time in MVPA [1,4] rather than just steps. Both of our interventions addressed this: PACE-Lift by nurse feedback on PA intensity from accelerometers [19] and PACE-UP by the “3,000-steps-in 30-minutes” [34] advice given to nurse-support and postal arms [16]. Both trials were effective at increasing MVPA in bouts for all intervention groups at all outcome assessments: 3 and 12 months [14,15] and now at 3 years (PACE-UP) and 4 years (PACE-Lift).
We took the effect estimates from the simplest intervention (PACE-UP postal) to estimate long-term health benefits. Based on a systematic review that quantified the strength of association between walking and coronary heart disease [35], a 28 minutes per week increase in MVPA in bouts seen in the postal group at 3 years should reduce coronary heart disease risk by approximately 4% (95% CI: 3–5%) (see S1 Text). A cohort study that related pedometer steps to mortality [36] allowed us to estimate that a sustained increase of 627 steps/day in the postal group at 3 years should lead to a decrease in all-cause mortality of approximately 4% (95% CI: 1–5%) (see S1 Text).
Whilst environmental and policy interventions are urgently required to address the global inactivity challenge [37], individual PA behaviour change interventions are also important. The sustained effects seen on objective PA outcomes at 3 years for the lower cost postal intervention suggest that this would be an effective and cost-effective [31] intervention to roll out. Minimal support is also required to check that materials have arrived and to encourage return of completed PA diaries but need not be face to face or delivered by a healthcare professional. We are currently conducting implementation work (PACE-UP Next Steps) exploring reach, retention, and ease of adoption in primary care recruiting via postal and face-to-face routes.
The use of wearables to monitor personal PA levels has dramatically increased, through smartphones, wrist- or body-worn devices, and mobile apps, offering opportunities for increasing PA. The “3,000-steps-in-30-minutes” message captures intensity and could become an important new public health goal [34], with new, easy ways to measure steps. Small short-term studies in adults and older adults demonstrate that mobile PA apps can increase PA self-monitoring [38,39] and engagement in regular PA [38,39] and that body-worn fitness trackers can increase time spent in MVPA [39]. PACE-UP Next Steps is currently testing online resources and a mobile app to support the PACE-UP postal intervention. However, despite new PA monitoring opportunities, it is important not to ignore robust, trial-based evidence on effective and cost-effective pedometer- plus paper-based interventions.
We previously reported increased PA at 12 months following 12-week pedometer-based walking interventions for adults and older adults recruited through primary care, delivered either by post with minimal support or through nurse-supported PA consultations. The current paper demonstrates that these findings are still present in participants followed up at 3–4 years. The long-term success of these interventions suggests that they could help to address the public health physical inactivity challenge.
|
10.1371/journal.pgen.0030134 | A Screen for Suppressors of Gross Chromosomal Rearrangements Identifies a Conserved Role for PLP in Preventing DNA Lesions | Genome instability is a hallmark of cancer cells. One class of genome aberrations prevalent in tumor cells is termed gross chromosomal rearrangements (GCRs). GCRs comprise chromosome translocations, amplifications, inversions, deletion of whole chromosome arms, and interstitial deletions. Here, we report the results of a genome-wide screen in Saccharomyces cerevisiae aimed at identifying novel suppressors of GCR formation. The most potent novel GCR suppressor identified is BUD16, the gene coding for yeast pyridoxal kinase (Pdxk), a key enzyme in the metabolism of pyridoxal 5′ phosphate (PLP), the biologically active form of vitamin B6. We show that Pdxk potently suppresses GCR events by curtailing the appearance of DNA lesions during the cell cycle. We also show that pharmacological inhibition of Pdxk in human cells leads to the production of DSBs and activation of the DNA damage checkpoint. Finally, our evidence suggests that PLP deficiency threatens genome integrity, most likely via its role in dTMP biosynthesis, as Pdxk-deficient cells accumulate uracil in their nuclear DNA and are sensitive to inhibition of ribonucleotide reductase. Since Pdxk links diet to genome stability, our work supports the hypothesis that dietary micronutrients reduce cancer risk by curtailing the accumulation of DNA damage and suggests that micronutrient depletion could be part of a defense mechanism against hyperproliferation.
| Cells must ensure the integrity of genetic information before cellular division. Loss of genome integrity is particularly germane to tumorigenesis, where it is thought to contribute to the rapid evolution of the malignant cell towards the fully cancerous phenotype. It is therefore imperative that we understand fully how cells maintain the integrity of the genome and how it is lost during tumorigenesis. In this study, we developed an assay that allowed us to systematically interrogate each gene of the budding yeast S. cerevisiae for its respective contribution to genome integrity. We report the identification of nine novel genes that increase the rate of genome instability in yeast when deleted. To our surprise, one of the genes we identified encodes the enzyme pyridoxal kinase, which acts in the metabolism of vitamin B6. We show that pyridoxal kinase influences genome stability by promoting the conversion of dietary vitamin B6 into its biologically active form, pyridoxal 5′ phosphate. Our work indicates that vitamin B6 metabolites are critical to maintain genome stability and supports a long-standing model, which hypothesizes that vitamin B6 reduces cancer risk by curtailing genome rearrangements.
| The faithful replication of the genome is necessary for maintenance of genome integrity. Disrupting processes that ensure faithful DNA replication results in chromosome breakage, hyper-recombination, or gross chromosomal rearrangements (GCRs) [1–3]. This relationship has been particularly highlighted in the budding yeast S. cerevisiae, where GCRs arise at high rates in cells with defects in the S-phase checkpoint [4], DNA replication licensing [5,6], DNA replication elongation [7–9], chromatin assembly [10], and homologous recombination (HR) repair [8].
Altogether, these studies not only suggest a common origin (i.e., DNA replication), but also a common mechanism by which genome rearrangements are formed [2]. Defects that occur during DNA replication lead to elevated levels of DNA damage, including DNA double-strand breaks (DSBs). In turn, these lesions may serve as substrates for the illegitimate repair processes resulting in GCRs. Therefore, identification of genes that prevent GCRs can potentially uncover novel genome caretakers that guard cells against the accumulation of mutations. In addition, unbiased identification of GCR suppressors could be a useful route for discovering novel genes and pathways that participate in DNA replication.
Most of the current knowledge regarding GCR formation originates from candidate gene studies examining rearrangements at a single locus in budding yeast, the left arm of Chromosome V (ChrV-L). Although this locus has been instrumental in the deciphering of many basic mechanisms governing genome stability in eukaryotes, examination of GCR formation at other loci provides a complementary view. For example, the use of yeast artificial chromosomes to study GCRs led to the discovery that defective chromosome condensation (in a ycs4 mutant) results in GCR events [7]. In addition, studies employing a Chromosome VII disome found that defects in DNA replication and checkpoint control elevate rates of chromosome loss and rearrangements following replication fork stalling [11]. In another study, Hackett et al. employed the telomeric region of ChrXV-L to study GCR events triggered by telomerase dysfunction [12]. This latter locus is particularly useful since GCRs at ChrXV-L involve break-induced replication (BIR), a type of homologous recombination repair predicted to be a major source of genome rearrangements [2,13–15]. In contrast, GCRs formed at ChrV-L are primarily the consequence of de novo telomere addition [8]. This difference can be explained by the architecture of the telomere-proximal region on ChrXV-L, which contains two regions of homology (HRI centered on the PAU20 gene, and HRII centered on the HXT11 gene; Figure 1A) located 12 kb and 25 kb from the telomere [12]. These regions share a high degree of sequence identity with other regions in the genome [12]. As a consequence, DNA lesions formed at loci telomeric to HRI or HRII are predominantly repaired by BIR, producing nonreciprocal translocations in haploid cells. Notably, increased repair by BIR can also lead to loss of heterozygosity in diploid genomes, which may accelerate the process of tumorigenesis by inactivation of tumor suppressor genes.
In this study, we screened the yeast genome for mutants that increase the level of chromosome rearrangements; specifically, those that increase the frequency of BIR-mediated nonreciprocal translocations. We report the construction of a strain containing a GCR reporter on ChrXV-L that is amenable to genome-wide screening and compatible with synthetic genetic array technology [16]. We employed this strain to systematically screen the gene deletion collection [17] leading to the identification of nine new GCR suppressors. Here, we focus on the characterization of one of the most potent GCR suppressors identified, BUD16, which encodes yeast pyridoxal kinase (Pdxk), a critical enzyme in vitamin B6 metabolism. We show that Pdxk is critical for the maintenance of genome integrity via its role in maintaining adequate levels of pyridoxal 5′ phosphate (PLP), the biologically active form of vitamin B6. Our results are consistent with a model whereby dTMP biosynthesis is the pathway affected by a decrease in PLP, thus providing an important link between dietary micronutrients, DNA replication and genome stability. Furthermore, since many epidemiological studies have linked defective vitamin B6 levels to an increased cancer incidence [18–23], our study supports the hypothesis that micronutrients such as vitamin B6 curtails carcinogenesis by preventing genomic instability.
To generate a GCR reporter strain that is amenable to genome-wide screening, we adapted a system previously described by Hackett et al. [12]. We inserted the CAN1 and URA3 genes, two counter-selectable markers, ∼10 kb from the telomere of ChrXV-L (Figure 1A). The simultaneous loss of CAN1 and URA3 (detected on media containing canavinine [can] and 5-fluoro-orotic acid [5-FOA]) at this locus occurs at a rate of 8.9 × 10−8 (Table 1), approximately 250-fold higher than the rate observed at ChrV-L (3.5 × 10−10; Table 2). This elevated GCR rate may be due to the higher efficiency of BIR over de novo telomere addition in repairing DSBs. Moreover, the HRI and HRII regions on ChrXV-L display between 85%–97% homology with a total of 21 chromosome arms [12]. This large number of potential seeds for BIR may also explain the relatively high GCR rate at ChrXV-L. To ensure that the GCR events recovered from the simultaneous loss of CAN1 and URA3 are due to BIR, we analyzed GCR events in wild-type cells by pulsed-field gel electrophoresis (PFGE) using a scheme described by Hackett et al. [12]. Briefly, we isolated genomic DNA from parental canS 5-FOAS cells and cells that have undergone GCR events at ChrXV-L (canR 5-FOAR cells). This DNA is then digested with PmeI to liberate a terminal restriction fragment, separated by PFGE, and finally transferred onto nitrocellulose by Southern blotting to be probed with a ChrXV-L-specific fragment (NOP8; Figure 1B). If BIR occurs by employing any of the 21 homologous chromosome arms as a template, the resulting terminal restriction fragments liberated by PmeI are all predicted to be of lower molecular weight than the parent fragment (∼97 kb). As predicted, the analysis of nine canR 5-FOAR mutants derived from the wild-type strain indicate that nine out of nine have undergone a GCR event at ChrXV-L that is consistent with BIR, since their terminal restriction fragments all migrate faster than that of the parental strain (Figure 1B). Furthermore, the analysis of four canR 5-FOAR strains by comparative genome hybridization using tiling microarrays identified breakpoints either in HRI, in the vicinity of the PAU20 gene (in two of four strains analyzed), or in HRII (i.e., in the vicinity of the HXT11 genes, two of four strains; see Figure 1C for two representative examples). Lastly, we determined the GCR rate at ChrXV-L of a rad52Δ strain, since RAD52 is required for BIR. The rad52Δ strain does not produce any detectable GCR events under the standard conditions of our assay (i.e., the rate must be ≪8.4 × 10−9; Table 2). This result suggests that most of GCR events observed at ChrXV-L are indeed dependent on RAD52, a gene required for BIR. Collectively, the above results indicate that the ChrXV-L GCR reporter monitors BIR-type events.
We crossed the resulting ChrXV-L GCR assay strain with the 4,812 viable open reading frame deletion strains [16,17] and employed a semi-quantitative papillation assay to monitor GCR formation (Figure S1). An initial set of 48 strains that scored positive were reconstructed in the ChrXV-L assay strain to determine their GCR rate by fluctuation analysis [24] (Table 1). This group included deletions in several known GCR suppressors such as the genes encoding Mre11, a component of the MRX complex, the RecQ helicase Sgs1, and the budding yeast FEN-1 homolog, Rad27. Using this scheme, we identified nine gene deletions that display at least a 10-fold increase in GCR rate compared to wild type (Table 1 and Figure 2A). Of these nine novel GCR suppressors, mutations in RMI1, RAD5, SLX8, and HEX3 were independently reported during the course of this study to promote GCRs at ChrV [25–27].
The remaining five novel GCR suppressors include BUD16, WSS1, ESC2, RML2, and ZIP1. Intriguingly, ZIP1 encodes a component of the synaptonemal complex that is active during meiosis [28] and is also expressed in mitotic cells [29], suggesting a potential role for Zip1 in mitotic genome stability. RML2 encodes the mitochondrial L2 ribosomal protein [30]. Surprisingly, a Rml2-GFP fusion protein localizes to the nucleus [31], suggesting a putative nuclear function for Rml2. WSS1 encodes a weak suppressor of an smt3 mutation [32], and ESC2 encodes a protein harboring a SUMO domain that has been linked to heterochromatic silencing [33,34] and the function of the Smc5/6 complex [35]. BLAST searches and alignments reveal that BUD16 encodes a putative Pdxk. With a GCR rate of 1.1 × 10−5 (124-fold above wild type), bud16Δ is within the range of very potent GCR mutator deletions that include rad27Δ (1.3 × 10−5; 148-fold over wild-type rate), mre11Δ (1.3 × 10−5; 140-fold), and sgs1Δ (1.2 × 10−5; 129-fold) (Figure 2A and Table 1). Reintroduction of a plasmid encoding wild-type BUD16 complemented the genome instability of bud16Δ cells, eliminating the possibility that a second site mutation contributes to its elevated GCR rate (Figure 2B). We also examined the type of GCR events promoted by the bud16Δ mutation by PFGE, which indicated that BUD16 prevents BIR-type rearrangements at ChrXV-L (Figure 2C). Given that bud16Δ had the most profound effect on genome stability among the uncharacterized suppressors, we focused on deciphering its role in preventing chromosome rearrangements.
In all organisms, Pdxk is an essential component of a vitamin B6 salvage pathway that ultimately produces PLP [36]. To ascertain whether BUD16 functions as the budding yeast Pdxk, we measured PLP levels in wild-type and bud16Δ strains. We found that the PLP levels of bud16Δ cells are only 1.8% of wild-type levels (Table 3). This result is somewhat surprising, since bacteria, yeast, and plants also possess a de novo vitamin B6 pathway that produces PLP in a Pdxk-independent manner. In yeast, this pathway is under the control of the SNO1 and SNZ1 genes [37]. However, these genes are not normally expressed during logarithmic growth but rather are expressed during stationary phase or under poor nutrient conditions. We found that the simultaneous deletion of the SNO1 and SNZ1 locus did not significantly reduce PLP levels (95.2% of wild-type levels; Table 3) or increase GCR rates (Table 2). Although SNO1 and SNZ1 deletion did not significantly impact genome stability (or PLP levels) in BUD16 cells, the de novo vitamin B6 synthetic pathway is nevertheless essential for viability in the absence of BUD16. Indeed, we are unable to recover viable triple mutants from a cross between bud16Δ and sno1snz1Δ or double mutants from a cross between bud16Δ and snz1Δ (Figure 2D). Overall, the decrease in intracellular PLP levels in bud16Δ along with its synthetic lethality with sno1snz1Δ are consistent with the idea that BUD16 functions in parallel with the de novo B6 pathway as a yeast pyridoxal kinase. Additional characterization of the bud16Δ strain in terms of growth and cell cycle kinetics is described in Text S1 and in Figure S2 and Tables S1 and S2.
To determine whether the bud16Δ mutation increases genome instability across the genome, we calculated the GCR rate of a bud16Δ strain at the ChrV-L locus [8]. We found that the bud16Δ mutation elevates the GCR rate at this locus 19-fold over the wild-type rate (Figure 3A; Table 2). Analysis of the GCR events involving ChrV by whole-chromosome PFGE reveals a mixture of events consistent with de novo telomere additions (six out of eight events analyzed) and nonreciprocal translocations (two out of eight events) (Figure 3B). This ratio of telomere additions to translocations (4:1) is similar to the ratio of GCRs typically recovered from a wild-type strain [4]. Together, these results indicate that BUD16 suppresses different types of genome rearrangements at a minimum of two different loci in the genome, suggesting that BUD16 may act to prevent the occurrence of DNA lesions rather than by promoting a specific type of illegitimate repair.
To further characterize the mechanism that underlies the high GCR rate of bud16Δ cells in the ChrXV-L assay, we crossed the bud16Δ GCR reporter strain to a strain containing a deletion of RAD52, a gene necessary for all types of homologous recombination, including BIR [15]. The GCR rate at ChrXV-L of the bud16Δ rad52Δ double mutant was reduced to wild-type levels (2.4 × 10−8; Table 2). However, this rate is far greater than the GCR rate of a rad52Δ mutation alone (<<8.4 × 10−9; Table 2). Furthermore, analysis of the terminal restriction fragment of the rearranged chromosomes in the bud16Δ rad52Δ double mutant shows terminal deletions in seven out of eight cases that are strikingly larger than those observed in either wild-type or bud16Δ strains (23–37 kb shorter in rad52Δ strains versus ∼7–17 kb in the RAD52+ strains; Figure 4A and 4B). This difference in the size of the ChrXV-L terminal restriction fragment suggests that bud16Δ rad52Δ do not undergo BIR-mediated GCR events that employ the HRI/II regions as seeds. Instead, in the absence of functional HR, these mutants are likely repaired by de novo telomere additions, leading to large terminal deletions.
Together, the observations that BUD16 suppresses GCRs at multiple loci in a BIR-dependent and independent manner suggest that bud16Δ cells experience higher-than-normal levels of DNA lesions during vegetative growth. We find support for this possibility when tetrads from a cross between bud16Δ and rad52Δ are examined (Figure 5A). We observe that the bud16Δ rad52Δ double mutant displays synthetic sickness and poor viability when compared to their congenic single mutants. This result suggests that bud16Δ cells may experience high levels of genotoxic stress that require the HR pathway for optimum viability. Consistent with this explanation, the bud16Δ mutation also displays synthetic sickness when crossed with an MRE11 gene deletion and to a lesser extent with deletion of RAD51, two additional genes acting in the homologous repair of DSBs (Figure 5A). We also observe a strong genetic interaction between the bud16Δ and rad6Δ mutations (Figure 5A), suggesting that DNA lesions caused by the reduction of PLP levels may require post-replicative repair or lesion bypass. Based on this spectrum of genetic interactions, bud16Δ cells likely accumulate DNA lesions during DNA replication, possibly leading to replication fork stalling or collapse.
To gain more direct evidence for the presence of active RAD52-dependent recombination in bud16Δ cells, we monitored the formation of Rad52 DNA repair centers [38]. Upon formation of lesions that engage HR, Rad52 relocalizes from a diffuse nuclear pattern into discernable punctate foci that coincide with DNA lesions. We thus expressed a Rad52 protein fused to the yellow fluorescent protein (YFP) in wild-type and bud16Δ cells and examined the presence of Rad52 repair centers by fluorescent microscopy (Figure 5B). In bud16Δ cultures, 37%–75% of the cells display Rad52-YFP foci compared to 5%–21% of wild-type cells (Figure 5B). In bud16Δ cultures, Rad52-YFP foci are surprisingly found in G1 (unbudded) cells but are most prevalent in S/G2/M (budded) cells (57%–75% in budded versus 37%–59% in unbudded cells). Intriguingly, the presence of Rad52 foci in G1 nuclei suggests the presence of persistent or unrepairable DNA lesions in cells that have undergone checkpoint adaptation [39]. Furthermore, we observe that budded bud16Δ cells display greater than one repair centre in 12–18% of the cells whereas this situation occurs in less than 2% of wild-type cells (Figure 5C). Since up to ten DSBs may localize to one repair centre [38], these results suggest that bud16Δ cells experience high levels of DNA lesions during DNA replication. Alternatively, we cannot exclude the possibility that bud16Δ cells have a dramatically reduced rate of DNA repair. However, bud16Δ cells are not sensitive to the radiomimetic alkylating agent methyl methanesulfonate (MMS; Figure 5D), indicating that the Rad52 pathway is functional in these cells. Therefore, the increased presence of Rad52 foci in bud16Δ cells is most likely explained by an increased number of DNA lesions.
Next, we examined whether the spontaneous DNA damage present in bud16Δ cells is sufficient to activate the DNA damage checkpoint pathway by assaying Rad53, the yeast homolog of the tumor suppressor Chk2. Rad53 kinase activation is observed by a detectable auto-kinase activity [40] concomitant with a reduced mobility on SDS-PAGE due to autophosphorylation. As shown in Figure 5E, Rad53 is hyperactivated in bud16Δ cells when compared to wild type, indicating that sufficient DNA damage is present in the bud16Δ mutant to activate the DNA damage checkpoint. Phosphorylation of Rad53 in cycling populations is often seen in strains that experience high levels of spontaneous DNA lesions, such as dia2Δ, rrm3Δ, and rmi1Δ, among others [25,41,42]. Altogether, our results are consistent with a model whereby bud16Δ cells experience high levels of DNA lesions, including DSBs. These DNA lesions most likely serve as substrates for illegitimate repair, resulting in elevated levels of genome rearrangements.
The genome instability observed in the bud16Δ mutant correlates with low levels of PLP. However, this observation does not eliminate the possibility that its GCR mutator phenotype could be due to a PLP-independent function of Bud16/Pdxk. To address this possibility, we aimed to reduce PLP levels by alternative means to probe the relationship between PLP and genome integrity. As a first means, we inactivated other components of the vitamin B6 salvage pathway. In particular, when yeast are grown to log phase under laboratory conditions, the PLP precursor, pyridoxine, is actively transported into cells mainly, but not solely, by the Tpn1 transporter [43]. Therefore, we asked whether TPN1 deletion impacts total PLP levels and genome stability. Cells harboring a tpn1Δ mutation have low levels of intracellular PLP (8% of wild-type; Table 3) that are nevertheless higher than those of bud16Δ. At a genetic level, we find that the tpn1Δ mutation is not synthetic lethal with the sno1snz1Δ double mutant (Figure 6A). The continued viability of the tpn1Δ sno1snz1Δ mutant supports the observation that although Tpn1 is a component of the B6 salvage pathway, it is not absolutely essential for pyridoxine transport [43]. Accordingly, the GCR rate of the tpn1Δ strain at ChrXV-L is increased 47-fold over the wild-type rate, which is less that of the bud16Δ GCR rate (124-fold over wild type; Figure 6B). However, both genes act in the same pathway to suppress genome rearrangements, as the double bud16Δ tpn1Δ mutant display the same GCR rate as bud16Δ (Figure 6B). Lastly, we find multiple Rad52 recombination centers present in 4%–7% of tpn1Δ budded cells, suggesting the presence of catastrophic DNA damage similar to that seen in bud16Δ, albeit at a lower level (Figure 6C). Altogether, these results further suggest that PLP levels correlate with genome integrity.
In addition to manipulating PLP levels via genetic means, we also manipulated pyridoxine intake to further explore the link between PLP levels and genome integrity. To carry out these experiments, we disabled de novo vitamin B6 synthesis (via SNO1 SNZ1 inactivation) to exclude the contribution of this pyridoxine-independent pathway. We also impaired pyridoxine transport by deleting TPN1. When grown in rich media, the resulting tpn1Δ sno1snz1Δ triple mutant has an elevated GCR rate (191-fold over wild type), which is greater than either the tpn1Δ or sno1snz1Δ mutants (Table 2). Accordingly, when we measure the PLP levels of this strain, we find that they are 5.8% of wild-type levels (Table 3). Importantly, we then supplemented the growth media of tpn1Δ sno1snz1Δ with 2 μg/ml pyridoxine as a means to stimulate its transport across the membrane. As shown in Figure 6D and Table 2, addition of pyridoxine to the media of tpn1Δ sno1snz1Δ potently suppresses its GCR rate to wild-type levels. Critically, under the same conditions, the PLP levels of the tpn1Δ sno1snz1Δ strain are dramatically increased to 81.5% of wild type (Figure 6E; Table 3). Together, these data conclusively demonstrate a relationship between PLP levels and maintenance of genome integrity.
We next examined the role of Pdxk on the genome integrity of human cells by employing the well-characterized vitamin B6 analog 4-deoxypyridoxine (4-DP) [44]. First, we determined whether inhibition of human Pdxk leads to DNA damage, particularly DSBs. To detect DSBs in human cells, we examined the localization of 53BP1, a DNA repair and signaling protein, by immunofluorescence microscopy. 53BP1 forms nuclear foci that colocalize with DSBs in mammalian cells and is thus a useful surrogate marker of this type of DNA damage [45]. As shown in Figure 7A and 7B, addition of 4-DP to the media of HeLa cells causes an accumulation of 53BP1 foci. Second, we found that 4-DP treatment also triggers activation of the checkpoint kinase Chk2, as evidenced by phosphorylation of its Thr68 residue (Figure 7C). Third, we analyzed the phosphorylation status of H2AX on its C-terminal Ser139 residue (known as γ-H2AX), one of the earliest events in the response to DSBs. The presence of γ-H2AX was assessed via immunoblotting (Figure 7C) and flow cytometry (Figure 7D) [46]. In cells treated with the Pdxk inhibitor 4-DP, γ-H2AX clearly accumulates during S-phase between the 2N and 4N DNA content, similar to what is observed in yeast cells (with Rad52-YFP). Importantly, to ensure that the described effects were not due to apoptotic effects caused by Pdxk inhibition at the concentrations of 4-DP employed above, we measured levels of apoptosis in HeLa cells by annexin V staining (Figure S3). From this data, we can rule out the possibility that 4-DP triggers DSB formation via the activation of an apoptotic program. Instead, our results indicate that, as in yeast cells, Pdxk inhibition induces DNA lesions and activation of the DNA damage response.
We finally sought to narrow down the biological pathway in which PLP acts to promote genome stability. This is a difficult task, since PLP is a critical cofactor for numerous essential enzymes acting in amino acid and dTMP biosynthesis. However, our observations in yeast and human cells indicate a role for PLP in preventing DNA lesions during DNA replication, pointing to dTMP synthesis as the likeliest candidate pathway linking PLP to genome stability (Figure 8A). This possible association is strengthened by the multitude of observations that link dTMP biosynthesis to genome integrity in both prokaryotes and eukaryotes (reviewed in [47]). In this context, PLP deficiency may either cause aberrant uracil incorporation into DNA, generate a nucleotide imbalance that impairs DNA replication fork stability, or both. We therefore sought to assess the involvement of PLP in dTMP biosynthesis by testing whether bud16Δ cells accumulate uracil nucleotides in their DNA. To do so, we employed a recently described modified aldehydic slot blot assay that detects abasic sites produced when isolated DNA is treated with a uracil glycosylase enzyme [48]. As shown in Figure 8B, strains lacking Pdxk (bud16Δ) accumulate uracil in their genome significantly more than their wild-type counterparts. This accumulation is likely to be biologically important, as it is in the same range as the uracil accumulation observed in cells deficient in uracil glycosylase (ung1Δ cells), the main enzyme dedicated to the removal of uracil in DNA (Figure 8C). Furthermore, the double ung1Δ bud16Δ mutant accumulates more uracil in its genome than either of the single mutants, suggesting that UNG1 and BUD16 function in separate pathways to prevent uracil incorporation into DNA. These results are therefore consistent with a model in which the bud16Δ mutation increases dUMP pools, thereby increasing the frequency of dUTP incorporation into DNA.
Accumulation of uracil in genomic DNA may lead to DSB accumulation and attendant genome instability via excision of uracil and production of excessive abasic sites. However, deletion of UNG1 does not suppress the bud16Δ genome instability rate and in fact results in a GCR rate increase (Table 2). This result indicates that either uracil excision is not a major cause of DNA damage in cells with low PLP levels, that an alternative excision pathway is involved, or that it is the accumulation of uracil that poses a threat to replication fork progression. Alternatively, it is also possible that a nucleotide pool imbalance caused by dUTP accumulation is a source of replication stress in bud16Δ cells. If bud16Δ cells have a defect in maintaining nucleotide pools, they may display some form of sensitivity to hydroxyurea (HU) a ribonucleotide reductase inhibitor. As shown in Figure 8D, HU dramatically affects the growth of bud16Δ cells at all concentrations tested and also leads to inviability of bud16Δ at 0.2 M HU, as measured by a colony-forming assay (Table S3). In contrast and as discussed above, bud16Δ cells are resistant to MMS, a DNA alkylating agent that causes DNA replication stress by impeding replication fork progression [49] (Figure 5D). Therefore, bud16Δ cells are sensitive to the depletion of deoxyribonucleotides rather than to replication stress. From these results, we suggest that PLP deficiency triggers DNA lesions due to a nucleotide imbalance resulting from defects in dTMP biosynthesis.
In this report, we describe a screen for suppressors of rearrangements at ChrXV-L, a locus producing chromosome aberrations primarily via BIR. Since this locus is different from the commonly used ChrV-L locus, we expected overlapping and distinct sets of genes with those already known to prevent rearrangements of ChrV. Indeed, comparison of the results of our screen with a similar screen undertaken using the ChrV-L reporter chromosome [50] identifies only two overlapping GCR suppressor genes, RAD5 and MRE11. This lack of overlap between both screens indicates that neither screen was saturating or that both loci can identify distinct classes of genome stability regulators. Indeed, we observed that disruption of some genes (such as SGS1) potently affects the GCR rate at the ChrXV-L locus while having a much more modest effect at ChrV [51], indicating that some genes may specifically suppress BIR.
In addition to Pdxk (Bud16), most of the other genes identified in our screen are likely to prevent DNA replication stress, suggesting that an unbiased screen for GCR suppressors is a potentially fruitful means of discovering novel activities influencing DNA replication. In particular, deletion of WSS1, which encodes a potential protease acting in the SUMO pathway [52], is synthetic lethal with deletion of SGS1, the yeast RecQ homolog [53]. This result, coupled with the high GCR rate of the wss1Δ strain, suggests that Wss1, perhaps via its proteolytic activity, acts in the management of DNA replication forks to prevent their demise. Likewise, Esc2 likely participates in maintaining replication fork integrity and tolerance to replication stress, given the ascribed role of its fission yeast homolog Rad60 in these processes [35,54]. Perhaps more puzzling is our identification of ZIP1, a meiotic gene, and RML2, encoding a mitochrondrial ribosome component, as GCR suppressors. Therefore, this study revealed several novel GCR regulators, which may be part of several less well-understood GCR suppression pathways. It will be important to decipher whether the products of these genes do indeed participate in the maintenance of mitotic genome integrity.
The link between decreased intracellular PLP levels and genome stability is important, since vitamin B6 deficiency correlates with heightened cancer risk [18–23]. This work therefore provides support for a model whereby subnormal levels of vitamin B6 may promote cancer development by engendering DNA lesions and attendant genome rearrangements. Given the poorly understood link between diet and cancer incidence, the ChrXV-L GCR assay provides a simple genetic system to probe the consequences of micronutrient deficiency on genome stability. Although the potential link between vitamin B6 and chromosome breakage had been suggested previously [20], the lack of a genetically tractable system to study the role of micronutrients in genome integrity has prevented a definitive mechanistic explanation of the vitamin B6–cancer epidemiological link. This situation has led to a multitude of alternative explanations. For example, other groups have contended that PLP decreases cellular proliferation or protects cells from oxidative stress [55,56]. We directly tested the possibility that reactive oxygen species affect the GCR rate of bud16Δ cells by growing them in the presence of the reactive oxygen species scavenger N-acetylcysteine (NAC). To our surprise, treatment with this compound increased rather than decreased the bud16Δ GCR rate (Table 2), indicating that reactive oxygen species may not play a major role in the formation of genome rearrangements when PLP levels are low.
Replication stress is thought to be a major deleterious event, as it is a source of DNA lesions and genome rearrangements. Paradoxically, recent observations point to a beneficial role for replication stress as an innate defense mechanism against tumorigenesis. Indeed, replication stress appears to be a hallmark of precancerous and hyperproliferating cells [57,58]. In this context, replication stress leads to activation of a DNA damage response that initiates senescence, thereby stopping the growth of a potential tumor [57–61]. These observations suggest that cells are wired to produce DNA lesions when their proliferation is aberrantly stimulated. One key and unresolved question that emerges from this body of work pertains to the nature of the cellular processes that trigger replication stress in response to uncontrolled cell growth. We speculate that our work, which links depletion of PLP to replication-associated DNA lesions, provides a simple mechanism that could link hyperproliferation to the activation of the DNA damage response. Indeed, we hypothesize that the exhaustion of metabolites through unscheduled anabolic processes may be primarily sensed as DNA replication stress. It will therefore be interesting to see whether intracellular PLP levels are decreased in precancerous lesions or whether Pdxk inhibition can sensitize cells to oncogene-induced cellular senescence.
The CAN1 gene from strain BY4741 was amplified from genomic DNA and cloned next to the URA3MX gene marker in the BglII site of pAG60 [62] to yield DDp418. To construct pBUD16 (DDp626), the BUD16 locus encompassing the BUD16 open reading frame was amplified by PCR from yeast genomic DNA and cloned in pCR2.1-TOPO (Invitrogen, http://www.invitrogen.com) and sequenced. The BUD16 locus was then excised with SpeI and NotI and cloned into the SpeI and NotI sites of pRS415.
To construct the GCR assay strain (DDY643) the CAN1 gene of BY4741 strain was replaced with the MFA1pr-HIS3 marker [16]. This strain was then transformed with a PCR fragment containing the cycloheximide-resistance cyh2 marker to yield strain DDY642. The CAN1-URA3 cassette was amplified from DDp418 with primers CAN1-URA3 F1: 5′-GAA TCT GCC GTT TCG ATT TAC TTC GAT AAA GTT TGC GTT GTG AGT CAT ACG GCT TTT TTG-3′ and CAN1-URA3 JM R1: 5′-GGA AAA TTC TGG TCT ATT CAC AAT GAC AAG CGG TGA GCG TGT ATA GCG ACC AGC ATT CAC-3′ (underlined regions anneal to DDp418 and flanking regions are homologous to ChrXV-L). A second round of PCR with the following primers extended homology to the ChrXV-L region: CAN1-URA3 F2: 5′-TAT TGT GAA TTG AAA TTT AAA GTT ATC TCA AAT TCA AAT GAA TCT GCC GTT TCG ATT TAC-3′ and CAN1-URA3 R2: 5′-AGA TGG CTT TTC CAT CAG AGC CAT TGT GAA GAA ATC GGA GGA AAA TTG TGG TCT ATT CAC-3′ (underlined regions anneal to the PCR product from the first round). This amplified fragment was introduced in DDY642 to yield DDY643 and was verified by PCR analysis. The MATα strain (DDY644) used in the screen was derived from DDY643 by mating type switching. All other strains were generated using genetic crosses, via one-step disruptions or via transformation of the indicated plasmids (see Table 4 for genotypes).
To generate gene deletions in the ChrXV-L GCR assay strain, we employed synthetic genetic array technology, essentially as described by Tong et al. [16]. Briefly, DDY644 was mated to the MATa deletion strains from EUROSCARF (http://web.uni-frankfurt.de/fb15/mikro/euroscarf/) on YPD and incubated at 30 °C overnight. Diploids were selected on SD-URA + 200 mg/L G418 and incubated at 30 °C for 2 d. Sporulation proceeded on YE + 0.05% glucose for 7 d at 25 °C. Once sporulated, haploids were selected by a four-step pinning procedure: two selections on SD-URA-HIS + cycloheximide (10 mg/L) followed by two selections on SD-URA-HIS + cycloheximide (10 mg/L) + G418 (200 mg/L). Following the fourth selection step, each deletion mutant was hand patched onto nonselective rich XY media (2% peptone, 1% yeast extract, 0.01% adenine, and 0.02% tryptophan). Fourteen mutants in duplicate were patched onto a single 10-cm plate. They were allowed to grow for 2 d at 30 °C and were then replica plated onto agar plates to remove excess cells prior to replica plating onto FC (5-FOA and can) media. FC plates were incubated for 3 d at 30 °C following replica plating and analyzed for colony formation. Wild-type strains produced between 0–3 colonies on average. Therefore, we scored a patch as a positive hit if the threshold number of colonies per patch were equal to or greater than ten. Deletion mutants (1,160) were then placed in a “1 hit” category (for those with one patch that displayed greater than ten colonies) or were in a “2 hit” category (for those with both patches that showed greater than ten colonies). The “2 hit” category list, which consisted of 273 mutants, was narrowed down by focusing on genes that are expressed in the nucleus [31]. However, we did not discard any gene deletion that had unknown localization data. These filters reduced the number of positive hits to 125. Of these 125, we reconstructed 48 deletions in the DDY643 background. Once constructed and confirmed, these mutants were frozen at −80 °C immediately in SC-URA to ensure retention of the CAN1-URA3 markers.
Strains were grown in SC-URA media to select for the URA3-CAN1 ChrXV-L arm prior to streaking cells onto nonselective rich media (XY). Single colonies were isolated and grown in 5 ml of XY medium until saturation. For wild-type strains, 1 ml of culture was spun down and plated onto a 10-cm FC plate and the number of cells/ml was calculated. A fluctutation test and the method of the median [24] was used to assess GCR rate. Similarily, for the ChrV-L GCR assay, a single colony was inoculated into 15 ml of XY medium until saturation. These cells were spun down and plated onto a 15-cm FC plate and the number of cells/ml was calculated. Again, a fluctuation test and the method of the median were used to measure GCR rates. To assess the effects of pyridoxine supplementation on genome stability, cells were grown in the absence or presence of 2 μg/ml pyridoxine hydrochloride (Supelco; http://www.sigmaaldrich.com/).
Wild-type, bud16Δ, tpn1Δ, sno1snz1Δ, and tpn1Δ sno1snz1Δ strains were grown in 100-ml cultures of XY + 2% glucose (with the exception of tpn1Δ sno1snz1Δ) to early log phase (OD600 1.0). The cells were spun down and washed thrice with 50 ml of cold double-distilled water to remove any external PLP. Cells were then pelleted and lysed by glass beads in 5% TCA. Lysates were clarified by centrifugation at high speed to remove cell debris. PLP measurements were done blindly at the diagnostic division of Anticancer (http://www.anticancer.com/).
We assessed Rad52-YFP focus formation assay essentially as described by Lisby et al. [38] with the following modifications. Three independent isolates of each strain containing the pRAD52-YFP plasmid (a gift of Grant Brown) were grown in SC-LEU. Cells were imaged on an Nikon Eclipse E600 FN microscope (http://www.nikonusa.com) equipped with an ORCA ER2 camera (http://www.hamamatsu.com) and Chroma filters (http://www.chroma.com/). Micrographs were taken in 21 z-stacks with 0.007-μm increments. For each independent isolate, a minimum of 180 cells were examined.
HeLa cells were grown in DMEM supplemented with 10% fetal calf serum. SiHa cervical carcinoma cells were obtained from the American Type Culture Collectin (http://www.atcc.org) and maintained in exponential growth by twice-weekly subcultivation in minimal essential medium containing 10% fetal bovine serum (Gibco, http://www.invitrogen.com/). A stock solution of 4-DP (200 mM) was prepared in 0.9% saline, diluted in growth medium, and adjusted to pH 7.2.
Human whole-cell extracts (25 μg) were prepared by boiling the cellular pellet in 1× SDS sample buffer for 5 min. Extracts were loaded onto an SDS-PAGE gel and after electrophoresis, proteins were transferred to a PVDF membrane (Millipore, http://www.millipore.com/) and immunoblotted with either the phospho-Chk2(Thr68) or Chk2 primary antibodies (Cell Signaling Technology, http://www.cellsignal.com/) followed by horseradish peroxidase-coupled secondary antibody (Jackson ImmunoResearch Laboratories, http://www.jacksonimmuno.com/). Rad53 immunoblotting and autokinase assays were carried out on denatured cell extracts exactly as described previously [9,63].
HeLa cells grown on coverslips were washed twice in PBS and fixed with 2% PFA for 1 h at room temperature, washed, and permeabilized with 1% Triton-X in PBS (1h, room temperature) and subsequently blocked for an additional 1 h in 10% antibody dilution buffer (10% normal goat serum, 3% BSA, and 0.05% Triton X in PBS). Monoclonal 53BP1 antibody (Transduction Laboratories) was diluted in blocking buffer and incubated with cells overnight at room temperature followed by two 10 min washes in 0.05% Triton X-100 in PBS. The appropriate Alexa-555 conjugated secondary antibody (Molecular Probes, http://probes.invitrogen.com/) was diluted 1:500 in 10% antibody dilution buffer and incubated with the coverslips for 2 h at room temperature. After several washes with PBS, the cells were stained with DAPI (10 μg/ml) for 20 min and mounted with ProLong Gold anti-fade agent (Molecular Probes).
Briefly, ∼1 × 108 cells grown from a saturated culture were spun down, washed, and then resuspended in TE (pH 7.5) and Zymolyase (Zymo Research, http://www.zymoresearch.com. Plugs were formed by mixing liquefied low melt agarose (SeaKem; http://www.lonza.com) with the resuspended cells and solidified in plug molds. Plugs were transferred into LET solution (0.5M EDTA pH 8.0, 0.01M Tris-HCl (pH 7.5), 40mM DTT, and 0.4mg/ml Zymolyase) overnight at 37 °C. Plugs were then transferred into fresh tubes containing NDS solution (0.5 M EDTA pH 9.5, 0.01 M Tris-HCl pH 7.5, 1% N-lauroyl sarcosine sodium salt, and 2 mg/ml proteinase K) and incubated at 50 °C overnight. Plugs were washed several times in TE (pH 7.5) and incubated for 1 h. Plugs containing whole chromosomes were immediately run on a CHEF-DR III system (Bio-Rad, http://www.bio-rad.com/) using a 1% agarose gel and 0.5× TBE at 14 °C, switch time 6–120 s, angle 120° for 24 h with a voltage gradient of 6V/cm. To examine the size of the terminal restriction fragment on ChrXV-L, whole chromosomes were prepared as described above in agarose plugs and were then digested with the restriction enzyme PmeI. Plugs of digested chromosomes were run on a 1% agarose gel in 0.5× TBE at 14 °C, switch time 1–15 s, angle 120°, 19 h with a voltage gradient of 6V/cm.
SiHa cells (5 × 105) were fixed in 1.4 ml of 70% ethanol and kept at −20° C for up to two weeks before analysis. All fixed samples were prepared for antibody staining and analyzed on the same day. One milliliter of cold Tris-buffered saline, pH 7.4 (TBS) was added to each tube, then the cells were spun down and resuspended in 1 ml of cold 4% FBS and 0.1% Triton X-100 (TST) and placed on ice. Cells were allowed to rehydrate for 10 min, then spun down and resuspended in 200 μl of mouse monoclonal anti-phospho-histone H2A.X antibody (Upstate Biotechnology http://www.millipore.com/), which was diluted 1:500 in TST. Tubes were incubated on a shaker for 2 h at room temperature, rinsed with cold TST, and resuspended in 200 ul of secondary antibody (Alexa 488 goat antimouse IgG [H + L]F[ab′]2 fragment conjugate [Molecular Probes] diluted 1:200 in TST) for 1 h at room temperature. Cells were rinsed in 2% FBS in TBS and resuspended in 400 μl of cold TBS containing 1 μg/ml DAPI. Samples were analyzed using a Coulter Elite dual laser flow cytometer (http://www.beckmancoulter.com). List mode files were analyzed using WinList software (Verity Software House, http://www.vsh.com/).
DNA was isolated from yeast strains using Qiagen (http://www.qiagen.com) gravity tip columns as per the manufacturer's protocol and assayed for uracil incorporation as described in Cabelof et al. [64]. Briefly, 4 μg of DNA was blocked for 2 h at 37 °C in a 2× tris/methoxyamine buffer (final concentration: 100 mM methoxyamine [Sigma-Aldrich, http://www.sigmaaldrich.com/] and 50 mM Tris-HCl, pH 7.4). DNA was precipitated with 7.5% volumes of 4 M NaCl and 4 volumes of ice-cold 100% ethanol and resuspended in TE buffer, pH 7.6. DNA was then treated with 0.4 units of Uracil DNA Glycosylase (New England Biolabs, http://www.neb.com) for 15 min at 37 °C, immediately precipitated, and resuspended in TE buffer, pH 7.6. DNA was then probed with 2 mM aldehydic reactive probe (Dojindo Molecular Technology, http://dojindo.com/) for 15 min at 37 °C followed by ethanol precipitation and resuspension in TE buffer, pH 7.6. DNA was then heat denatured, immobilized onto a nitrocellulose membrane (Schleicher and Schuell, http://www.whatman.com/), and baked under vacuum as originally described by Nakamura et al. [48]. The dried membrane was washed in 5× SSC for 15 min at 37 °C, then incubated in a prehybridization buffer (20 mM Tris, pH 7.5; 0.1 M NaCl; 1 mM EDTA; 0.5% casein w/v; 0.25%BSA w/v; and 0.1% Tween-20 v/v) for 30 min at room temperature. Streptavidin-POD conjugate (Roche, http://www.roche.com/) was added at a 1:2,000 dilution for 45 min at room temperature. Membrane was washed in TBS/Tween-20 three times at 37 °C, incubated in ECL solution (Pierce, http://www.piercenet.com/) for 5 min at room temperature, then visualized and quantified using a ChemiImagerTM system (Alpha Innotech, http://www.alphainnotech.com/). Data are expressed as the integrated density value of the band per microgram of DNA loaded on the membrane.
Yeast genomic DNA was prepared from saturated 10-ml cultures essentially by the method of [65]. Genomic DNA (2 μg) was digested for 2 h with 10 U of HaeIII and purified by phenol-chloroform extraction and ethanol precipitation. HaeIII-digested genomic DNA (50 μg) was labeled and hybridized by the method of [66]. Briefly, after blunting the DNA ends with T4 polymerase, the fragments were ligated to unidirectional linkers and amplified by ligation-mediated PCR in the presence of aminoallyl-modified dUTP. Indirect labeling was performed using monoreactive Cy5 (for the parental strain) or Cy3 (strains that had undergone GCR) NHS esters that react specifically with the aminoallyl-dUTP. Control and experimental samples were combined, and the labeled DNA was hybridized to a yeast whole-genome ChIP-on-chip microarray (4 × 44K; Agilent Technologies, http://www.home.agilent.com/) and scanned at the University Health Network Microarray Centre (http://www.microarrays.ca/). Microarray images were processed with GenePix Pro 6.0 (Molecular Devices, http://www.moleculardevices.com). Data were analysed as described previously [67]. Hybridization data were preprocessed with ArrayPipe 1.7 [68], the background was subtracted using the “foreground-background” correction method, the data were normalized using the “linear model for microarray analysis (limma) loess (subgrid) method,” and the results were mapped using the University of California Santa Cruz genome browser (http://genome.ucsc.edu/cgi-bin/hgGateway).
Text S2 contains supplementary materials and methods for apoptosis analysis, yeast DNA content, and cell size distributions. |
10.1371/journal.pgen.1000080 | BRCA1 and BRCA2 Missense Variants of High and Low Clinical Significance Influence Lymphoblastoid Cell Line Post-Irradiation Gene Expression | The functional consequences of missense variants in disease genes are difficult to predict. We assessed if gene expression profiles could distinguish between BRCA1 or BRCA2 pathogenic truncating and missense mutation carriers and familial breast cancer cases whose disease was not attributable to BRCA1 or BRCA2 mutations (BRCAX cases). 72 cell lines from affected women in high-risk breast ovarian families were assayed after exposure to ionising irradiation, including 23 BRCA1 carriers, 22 BRCA2 carriers, and 27 BRCAX individuals. A subset of 10 BRCAX individuals carried rare BRCA1/2 sequence variants considered to be of low clinical significance (LCS). BRCA1 and BRCA2 mutation carriers had similar expression profiles, with some subclustering of missense mutation carriers. The majority of BRCAX individuals formed a distinct cluster, but BRCAX individuals with LCS variants had expression profiles similar to BRCA1/2 mutation carriers. Gaussian Process Classifier predicted BRCA1, BRCA2 and BRCAX status, with a maximum of 62% accuracy, and prediction accuracy decreased with inclusion of BRCAX samples carrying an LCS variant, and inclusion of pathogenic missense carriers. Similarly, prediction of mutation status with gene lists derived using Support Vector Machines was good for BRCAX samples without an LCS variant (82–94%), poor for BRCAX with an LCS (40–50%), and improved for pathogenic BRCA1/2 mutation carriers when the gene list used for prediction was appropriate to mutation effect being tested (71–100%). This study indicates that mutation effect, and presence of rare variants possibly associated with a low risk of cancer, must be considered in the development of array-based assays of variant pathogenicity.
| Inherited mutations in the genes BRCA1 and BRCA2 increase risk of breast cancer and contribute to a proportion of breast cancer families. However, more than half of the reported sequence alterations in BRCA1 and BRCA2 are currently of unknown clinical significance. We analysed gene expression in lymphoblastoid cell lines derived from blood of patients with sequence alterations in BRCA1 and BRCA2 and compared these to lymphoblastoid cells from familial breast cancer patients without such alterations. We then classified these lymphoblastoid cells based on their gene profiles. We found that BRCA1 and BRCA2 samples were more similar to each other than to familial breast cancer patients without BRCA1/2 mutations, and that the type of sequence change in BRCA1 and BRCA2 (missense or truncating) influenced gene expression. We included in the study ten familial breast cancer samples, which carried sequence changes in BRCA1 or BRCA2, that are believed to be of little clinical significance. Interestingly these samples were distinct from other familial breast cancer cases without any sequence alteration in BRCA1 or BRCA2, indicating that further work needs to be performed to determine the possible association of these “low clinical significance” sequence changes with a low to moderate risk of cancer.
| Approximately 7% of breast cancer cases occur in women with a strong family history of the disease [1]. Mutations in BRCA1 and BRCA2 account for a considerable proportion of these familial breast cancer cases, with the average cumulative risk in BRCA1 and BRCA2 mutation carriers by age 70 years estimated at 65% and 45%, respectively [2]. The Breast Cancer Information Core (BIC) database (http://research.nhgri.nih.gov/bic/) currently has more than 1400 and 1800 unique sequence variants listed in the BRCA1 and BRCA2 genes, respectively. These include frameshift, nonsense, missense, splice site alterations and polymorphisms. Greater than a third of the BRCA1 and greater than half of the BRCA2 unique variants are “unclassified variants” without compelling evidence of pathogenicity or functional significance. The majority of unclassified variants recorded in the BIC database are predicted missense changes (more than 400 BRCA1 and 800 BRCA2). However other variants which may be categorised as unclassified variants are in-frame deletions or duplications, variants that may disrupt splicing, or variants in the 3′UTR that may affect RNA stability (www.kconfab.org). BRCA1/2 unclassified variants represent a problem in the clinical setting as it is not known which variants are associated with the high risk of disease reported for classical truncating mutations.
Several functional assays may be used to determine the significance of unclassified variants, including transcription activation and complementation assays [3]–[9], but a disadvantage of biochemical assays is that they rely on the functions of specific domains of the protein, require specialized laboratory skills, and are time–consuming to perform. Other methods for classifying variants include the analysis of clinical and histopathological data [10], loss of heterozygosity analysis [11] and bioinformatic analysis to predict the effect of the amino acid change on structure and multiple sequence alignment strategies [12]; [13]–[15]. Integrated evaluation of unclassified variants which combines several approaches, such as the analysis of co-segregation of the mutation with disease, co-occurrence of the variant with a deleterious mutation, sequence conservation of the amino acid change, severity of amino acid change, tumor loss of heterozygosity, and tumor histopathology classification, provides a quantitative tool for the classification of variants [16]–[22]. This multifactorial method was developed to classify such rare unclassified variants into two categories, variants with features of classical high-risk mutations (termed pathogenic), and variants that do not have the features of a high-risk mutation (termed neutral or low clinical significance (LCS)). While the availability of appropriate biospecimens (e.g. number of families and tumors) for inclusion in likelihood prediction is a major factor determining the classification of any single variant, another major caveat of the multifactorial approach is that it is not appropriate for the evaluation of possible moderate or low risk variants, since it uses high-risk mutations as reference for the underlying assumptions [16],[19],[20]. Therefore, the current multifactorial method cannot exclude the possibility that rare variants classified to be of low clinical significance may be associated with a moderate or low risk of cancer.
Gene expression profiling has increased our understanding of the molecular events in breast tumor development, has been used to predict prognosis, and has characterised breast tumors into subtypes [23]–[27]. The value of expression profiling for identifying underlying high-risk gene mutation status is indicated by a number of studies. A distinct gene expression profile has been reported for breast tumors of BRCA1 mutation carriers [23],[28],[29], expected to be homozygous for loss of BRCA1 function at the somatic level. In addition, the existence of distinct gene expression profiles for heterozygous loss of BRCA1 and BRCA2 function is supported by accurate separation of short-term cultures of fibroblasts carrying a germline mutation in the BRCA1 or BRCA2 genes, compared to healthy women undergoing reduction mammoplastic surgery with no family or personal history of any cancer or sporadic breast-cancer-affected controls [30],[31]. Lymphoblastoid cell lines (LCLs) have also been shown to have distinct mRNA expression phenotypes for heterozygous carriers of ATM mutations, some of which are known to be associated with an increased risk in breast cancer [32],[33]. These findings suggest that germline gene expression signatures, including those from fibroblasts or LCLs, may be used to define BRCA1 or BRCA2 mutation status and to assist in assessing the clinical significance of BRCA1 and BRCA2 unclassified variants.
In this study we compared LCL gene expression signatures of breast cancer cases carrying pathogenic mutations in BRCA1 or BRCA2, to familial breast cancer cases with no known BRCA1/2 mutations (BRCAX). We also considered the possibility that BRCAX individuals with a BRCA1 or BRCA2 sequence variant classified to be neutral/low clinical significance (LCS) using multifactorial likelihood analysis may differ in gene expression profile from BRCAX individuals without such sequence variants. In addition, since truncating alterations comprise the majority of known pathogenic mutations but most BRCA1 and BRCA2 unclassified variants are predicted missense alterations, we compared profiles from individuals with known missense or truncating mutations to determine if mutation effect will affect the mutation-associated expression profile for each gene. We derived gene lists to predict mutation status defined by gene and mutation effect, and then tested the efficacy of these gene lists to predict the gene mutation status of LCLs. We provide evidence that gene lists differ according to gene and mutation effect, and according to the presence of sequence variants of low clinical significance. We also demonstrate that the use of appropriately-derived gene lists improves the prediction of pathogenicity of known mutations.
The ultimate aim of this experiment was to establish if gene expression profiles could distinguish between BRCA1 or BRCA2 pathogenic mutation carriers and familial breast cancer cases whose disease was not attributable to BRCA1 or BRCA2 mutations (BRCAX cases). BRCAX breast cancer families are likely to result from mutations in several other genes, and thus represent a heterogeneous group. Moreover, included in the BRCAX group were a subset of 10 BRCAX individuals who carried a BRCA1/2 variant previously classified to be of low clinical significance using multifactorial likelihood approaches [8],[19],[21],[22]. Unsupervised hierarchical clustering showed that BRCAX LCLs containing a BRCA1 or BRCA2 variant of low clinical significance clustered away from the majority of remaining BRCAX samples (Figure 1). A t-test with Benjamini and Hochberg multiple testing correction [34] was performed to determine if there were gene expression differences between the BRCAX individuals with an LCS variant and those without an LCS variant. Expression of 631 genes differed between the two BRCAX subgroups (5% of the 631 genes identified would be expected to pass this restriction by chance). For this reason, BRCAX samples were stratified according to the presence of an LCS variant for further analyses.
Gene expression is similar for carriers of BRCA1 and BRCA2 truncating mutations and rare sequence variants of low clinical significance, but differs from BRCA1 and BRCA2 missense mutations and BRCAX non-BRCA1/2 familial cases.
Unsupervised hierarchical clustering (Figure 2) of LCL expression data from all samples revealed that BRCA1 and BRCA2 samples were more similar to each other than BRCAX samples without an LCS variant. This result suggests that germline effects of heterozygous mutations in BRCA1 and BRCA2 cannot easily be separated using the experimental conditions used in this study. Although BRCAX samples tended to cluster distinctly from BRCA1/2 samples, nine of ten BRCAX individuals who carried a BRCA1/2 variant previously classified to be of low clinical significance fell within the major BRCA1 or BRCA2 mutation cluster. In contrast, six of the nine pathogenic missense mutations of BRCA1 or BRCA2 fell into a BRCA1/2 outlier group, which clustered closer to the BRCAX samples.
To determine the accuracy of using gene expression data from LCLs to predict BRCA1/2 pathogenic carriers and BRCAX individuals, we used a Gaussian Process Classifier (GPC). GPC analysis was used previously in an analysis of microarray profiles from irradiated short-term fibroblasts of BRCA1/2 mutation carriers [31], and allows for multiway comparison of groups. For GPC analysis 2031 genes which were significantly over/under-expressed at the 5% significance level were selected. The GPC was used in a three way comparison to compare BRCA1 truncating mutation carriers to BRCA2 truncating mutation carriers, and to BRCAX samples without an LCS variant. Samples with BRCA1 or BRCA2 pathogenic missense mutations or classified as BRCAX with an LCS variant were then included to determine their affect on the prediction accuracy. A summary of the prediction accuracy is shown in Table 1. The highest prediction accuracy (62.26%) was achieved when the analysis excluded samples classified as BRCAX with an LCS, and samples with BRCA1 or BRCA2 missense mutations. This prediction accuracy is above the expected performance, as a random prediction with three classes comprised of a similar sample number would be 33% accuracy. When BRCA1 and BRCA2 samples were compared to only BRCAX samples with an LCS variant, the prediction dropped to 43.46%, and the addition of the BRCAX samples without an LCS variant improved the accuracy. In all comparisons the inclusion of the pathogenic non-truncating mutations of BRCA1 and BRCA2 lowered the prediction accuracy.
In the clinical setting, unclassified sequence variants of BRCA1 or BRCA2 are generally identified after full sequencing of both genes. Therefore the most common clinical question is whether a variant in BRCA1 or BRCA2 is pathogenic or not. We thus performed pair wise analyses to determine if BRCAX samples could be distinguished from those with pathogenic mutations in BRCA1 or BRCA2. Based on observations from hierarchical clustering analyses and the GPC analysis, we also considered the possibility that the effect of pathogenic BRCA1/2 mutations (truncating or missense) affected LCL gene expression. T-tests were performed using the 20,874 detected probes to elucidate gene differences between i) BRCA1 or BRCA2 truncating mutations vs BRCAX without an LCS variant; ii) BRCA1 or BRCA2 missense mutations vs BRCAX without an LCS variant. The number of genes which passed these restrictions and the overlap between them is outlined in Figure 3A and 3C. The comparisons were then repeated with BRCAX with an LCS variant (Figure 3B and 3D). As expected when BRCA1 and BRCA2 were compared to BRCAX samples without an LCS variant, a greater number of genes were deemed significant compared to BRCA1 or BRCA2 vs BRCAX samples with an LCS variant.
SVM is a widely accepted classification approach for assessing differences in mRNA expression, and was used to compare BRCA1 or BRCA2 individually to BRCAX samples. Since our detailed analysis of gene lists showed that mutation effect (truncating or missense substitution) appears to affect the genes that are differentially expressed in the carriers after IR (Figure 3), we assessed if these gene differences will affect the predictions. We used SVM with the top 200 genes from the comparison of BRCA1 or BRCA2 truncating mutations to BRCAX, and the top 200 genes from the comparison of BRCA1 and BRCA2 missense mutations to BRCAX (Figure 3A and 3C). The genes which differed between BRCA1 or BRCA2 and BRCAX with an LCS variant were not used in this comparison as too few genes passed the restriction (Figure 3B and 3D). The top 200 genes are listed in Tables S2, S3, S4, and S5 and the overlap of the top 200 genes used for prediction from [BRCA1 (missense) vs BRCAX (noLCS)] and [BRCA1 (truncating) vs BRCAX (noLCS)] was 16 transcripts, with no overlap between the top 200 genes from [BRCA2 (missense) vs BRCAX (noLCS)] and [BRCA2 (truncating) vs BRCAX (noLCS)]. A total of 715 different genes were represented in the four lists of top 200 gene-lists from comparison of BRCAX (no LCS) to the different BRCA1/2 groups above. The results are summarised in Tables 2 and 3. The BRCA2 truncating pathogenic carriers were consistently predicted with higher accuracy compared to BRCA1 truncating pathogenic carriers. The accuracy of prediction was improved when the gene list used for prediction was appropriate to the mutation effect (truncating or missense) being tested. When the missense-associated gene list was used, pathogenic truncating mutations were predicted with 35% and 68% accuracy for BRCA1 and BRCA2, respectively. Predictions increased to 71% and 84% for BRCA1 and BRCA2, respectively, using the truncating-associated genes. Similarly, the pathogenic missense mutation carriers were predicted with 83% and 100% accuracy when the missense-associated gene list is used, but this accuracy was lower or remained the same when the truncating-specific gene list was used (83% and 0%). Prediction of BRCAX samples that did not carry an LCS variant was high in all comparisons (82–94%). In contrast, prediction of BRCAX samples that did carry an LCS variant was poor (40–50%).
When using SVM, the significance of the predictions can also be represented by the distance the prediction is from the plane, where predictions called with greater confidence are further from the plane that separates the BRCA1 (or BRCA2) and BRCAX samples. The significance of the predictions for the BRCA1 pathogenic missense mutations is summarised in Figure 4. Although both missense and truncating gene lists correctly predicted 5 of 6 missense mutations, the results show that there is much greater confidence in the 5 correctly predicted missense mutations when using the missense-derived list.
Ingenuity Pathway Analysis of genes which differed between the LCLs carrying pathogenic truncating or missense mutations of BRCA1 or BRCA2 compared to BRCAX samples without an LCS variant was performed to determine the potential functional relevance of the differentially expressed genes. All BRCA1 and BRCA2 pathogenic mutations resulted in gene expression changes relating to cell cycle, cancer and cellular growth and development, while BRCA1 and BRCA2 missense mutations shared some additional similarities (cell death and cell development pathways). There were also alterations in several pathways that were unique to BRCA1 truncating mutations, BRCA2 truncating mutations, BRCA1 missense mutations, or BRCA2 missense mutations (Figure S1).
It is difficult to counsel patients with a strong family history of breast cancer who are found to carry an unclassified variant in BRCA1 or BRCA2. While management at the level of the family should remain unchanged from that of a BRCAX family with no knowledge of a BRCA1/2 mutation, some individuals from high-risk families may nevertheless interpret information about an unclassified variant to alter their choices regarding prophylactic surgery for example, and so require careful counselling. Gene expression profiling can be used to classify samples based on phenotype, and its frequent use in laboratories world-wide holds great promise for clinical application, to the extent that profiling tools are being developed for diagnostic use e.g. Agendia Inc. (http://www.agendia.com/).
Expression profiles of short-term fibroblasts have previously been reported to separate carriers of a heterozygous mutation in the BRCA1 or BRCA2 genes from sporadic breast-cancer-affected controls [30],[31]. We wished to determine if expression profiling of LCLs could similarly be used to predict BRCA1 or BRCA2 mutation status, with the ultimate aim of predicting the significance of unclassified variants of BRCA1 or BRCA2. We chose LCLs as a minimally invasive source of germline material that can be maintained as long term cultures, and because previous studies have shown that LCL array profiling is robust to sourcing of LCLs established in different laboratories [33]. We compared expression profiles of irradiated LCLs from BRCA1 and BRCA2 carriers to those of non-BRCA1/2 BRCAX familial breast cancer patients, an appropriate reference group for the proposed evaluation of unclassified variants identified in familial breast cancer patients. A relatively early time-point of 30 minutes post-irradiation was chosen to capture gene expression initiation, and minimize possible downstream compensation effects. It has previously been shown that 10Gy IR treatment of normal LCLs has an effect on the transcriptional response, with greatest change in mRNA levels for most genes within one hour post-treatment [35], and studies of mouse brain gene expression after whole-body low-dose irradiation have shown that a large number of early IR response genes can be measured at the 30 minute time point [36].
A number of BRCAX cases carried BRCA1 or BRCA2 sequence variants that had been previously classified using multifactorial likelihood modelling methods to be neutral or of low clinical significance-that is, these rare variants are extremely unlikely to be a high-risk mutation in either of these genes, but the modelling methods used cannot assess whether they are truly neutral or associated with a much lower risk of disease. We found that the BRCAX samples with such LCS variants were separated from the majority of BRCAX samples without such LCS variants using unsupervised hierarchical clustering. This result indicates that LCS samples differ in expression profile as a result of their BRCA1 or BRCA2 sequence variant, and was substantiated by the class prediction methods: GPC prediction of the BRCAX samples decreased in accuracy when BRCAX samples with an LCS were included. In addition, SVM to detect BRCA1 or BRCA2 mutation-related gene lists yielded differences in the significant genes for comparisons to BRCAX samples without an LCS variant, compared to BRCAX samples with an LCS variant. Accordingly, prediction of BRCAX subgroup status based on the more robust gene list derived from comparisons to BRCAX individuals without an LCS variant was generally poorer for BRCAX samples with an LCS (40–50%) compared to those without an LCS (82%–94%). These rather provocative results indicate that the possible effect of all rare variants should be considered in development of assays to assess which variants have features of high-risk mutations. Moreover, the similarity in expression profile of these variants to other BRCA1/2 pathogenic mutations suggests that at least some of these LCS variants may confer small-moderate risks of breast cancer, presumably acting in concert with alterations in other genes in the BRCA1/2 pathway to lead to breast cancer. Given the rarity of these variants, alternative statistical approaches will be required to assess the risk of cancer associated with them.
The assay conditions used in this study could not distinguish between samples with pathogenic BRCA1 mutations and those with pathogenic BRCA2 mutations. Ionising radiation has previously been show to separate fibroblast cells which carry BRCA1 or BRCA2 mutations from sporadic cases with 100% accuracy [31], but our experiment differs in several respects. We compared BRCA1 and BRCA2 cases to familial BRCAX cases as an appropriate reference group for familial breast cases likely to be identified as carriers of BRCA1/2 mutations or unclassified variants, we used LCLs instead of fibroblasts, we selected a lower IR exposure (10Gy vs 15Gy), and we chose a relatively early time point of 30 mins after exposure to IR in order to gain a better understanding of the functional differences in response to IR between the BRCA1, BRCA2 and BRCAX cell lines. Some or all of these factors may explain the difference in the ability of this study to distinguish BRCA1 from BRCA2, both of which are involved in DNA damage repair. However, differences in post-irradiation response between BRCAX individuals and BRCA1/2 mutation carriers are supported by alternative analysis we have conducted of the subset of genes reported to be involved in post-irradiation response, comparing mutation-negative normal female controls to BRCAX individuals without an LCS variant, or to BRCA1 or BRCA2 truncating mutation carriers. Our results indicate substantial differences in radiation response between normal controls and the patient groups, and also considerable differences between the BRCAX group and BRCA1 and BRCA2 carriers [37]. Alternative IR exposures and/or post-IR timepoints, and possibly different DNA damaging agents, should be considered for future experiments.
The ultimate aim of this experiment was to identify array profiles that would be useful for the classification of unclassified sequence variants of BRCA1 or BRCA2. In the clinical setting, individuals generally present with full sequencing of both genes, and presence of a variant in one gene or the other. We thus assessed the ability to distinguish BRCA1 or BRCA2, separately, from BRCAX individuals. Importantly, since most unclassified variants are predicted to cause amino acid substitutions, we also assessed the relevance of mutation effect for expression profiles. We found that the genes which significantly differed between BRCA1 or BRCA2 and BRCAX LCLs were dependent on mutation effect. Accordingly, the SVM prediction for each mutation effect was best if the appropriate gene list was used, in terms of both accuracy of prediction (BRCA1 or BRCA2 vs BRCAX) and confidence in the classification as determined by the distance of the prediction from the SVM plane. Thus we strongly urge that mutation effect is taken into account if this type of assay is to be developed for use in predicting the clinical significance of BRCA1/2 variants. The current challenge is that few missense variants have been classified with respect to their clinical significance, with the only 23 individual missense variants termed clinically important by BIC, 17 in BRCA1 and six in BRCA2. Moreover, these are restricted in terms of the domains/regions in which they occur, residing in the BRCA1 start site (n = 2), ring finger (n = 4) or transactivation domains (n = 11), and the BRCA2 CDK2 phosphorylation site (n = 3) or at one codon (2336, n = 3) in a region of unknown function. It will thus be difficult to accrue a panel of known pathogenic missense variants for use in such predictive assays, and will require a concerted collaborative effort. Assuming sufficient pathogenic variants are identified, the successful execution of such a study may eventually distinguish missense-associated gene expression patterns that are generic to missense mutations, and/or those that are specific to the domain location of missense mutations. In addition, a possibly greater challenge will be identifying assay conditions (cell type, perturbation, time-point etc) that can also identify gene expression differences between patients with rare variants of low clinical significance in BRCA1 and/or BRCA2 and those with truly high-risk pathogenic mutations (truncating or missense) in these genes. Our study, using conditions that were not optimal for separating BRCA1 and BRCA2 mutations nevertheless identified gene expression differences between BRCA1/2 pathogenic mutations and LCS variants, suggesting that larger sample sizes and further experimentation may identify a more robust gene list to separate pathogenic mutations, variants of low clinical significance, and individuals with no sequence alterations in BRCA1/2.
Pathway analysis confirming altered expression of cancer, cell proliferation and cell cycle pathways in BRCA1 and BRCA2 mutation carrier groups is consistent with the known functions of BRCA1 and BRCA2 [38],[39]. The pathway differences by mutation type such as cell death and development may reflect that the majority of truncating mutations result in activation of the nonsense mediated decay pathway [40] and complete loss of protein, whereas most missense mutations are likely to result in more subtle effects through ablation of individual functional domains. Some pathways identified were unexpected and are only present in a single mutation type, and it is thus likely that at least some of these pathways were generated by chance alone.
In conclusion, we have provided evidence that carriers of BRCA1 and BRCA2 variants considered to be of low clinical significance have array profiles distinct from other non-BRCA1/2 familial cases, but resembling profiles of pathogenic BRCA1/2 cases, indicating that further work will be required to evaluate their possible association with a low-moderate risk of cancer. We have also shown that it will be important to consider mutation effect when developing array-based assays for predicting the clinical significance of BRCA1 or BRCA2 unclassified variants. Lastly, our findings demonstrate the ability of array profiling of immortalized lines derived from lymphoblastoid cells to detect germline mutations in genes that result in breast and ovarian cancer, and thus have relevance to the investigation of other genetic diseases irrespective of the organs or tissues they affect.
LCLs were derived from breast cancer-affected women recruited into the Kathleen Cuningham Foundation for Research into Breast Cancer (kConFab), a consortium which ascertains multiple-case breast cancer families [41]. These include families in which one or more carriers of a BRCA1 or BRCA2 mutation have been identified, and families in which no predisposing mutation has been identified (BRCAX). The recruitment criteria for BRCAX families are: 1) at least one member of the family at high-risk according to the National Breast Cancer Centre Category III guidelines (http://www.nbcc.org.au), and four or more cases of breast or ovarian cancer (on one side of the family), and two or more living affecteds with breast or ovarian cancer, and four or more living first or second degree unaffected female relatives of affected cases, over the age of 18 ; 2) two or three cases of breast or ovarian cancer (on one side of the family) in same or adjacent generations, if at least one of these cases is ‘high risk’ (i.e. male breast cancer, bilateral breast cancer, breast plus ovarian cancer in the same individual, or breast cancer with onset less than 40 years), and two or more living affected cases with breast or ovarian cancer, and four or more living first or second degree unaffected female relatives of affected cases, over the age of 18.
Classifications for BRCA1 and BRCA2 pathogenic mutations and variants of low clinical significance (LCS) are described on http://www.kconfab.org/Progress/Classification.shtml. Briefly, LCS variants include BRCA1 or BRCA2 variants described in trans with a deleterious mutation in the same gene in an individual and occur at a frequency of less than 1% in unaffected controls, or considered neutral/low clinical significance as measured using multifactorial likelihood approaches [16],[19],[21],[22].
A cohort of 72 LCLs were used in this study. The full listing of mutation details for LCLs is shown in Table S1. In brief, the study included:
LCLs were grown in RPMI 1640 media with 15% fetal bovine serum, 1% penicillin-streptomycin and 1% L-glutamine. The cell number was normalised and fresh medium was added to cells 24hr prior to irradiation with 10Gy, using a calibrated Cs137 c-source delivering 1 Gy/1.5 min. Total RNA was harvested 30min later using an RNeasy kit (Qiagen, Doncaster, VIC). The Illumina Totalprep RNA amplification kit (Ambion, Austin, TX) was used to amplify and biotinylate 450ng of total RNA. Biotinylated RNA was hybridised overnight at 55°C to Illumina Human-6 version 1 BeadChips containing >46,000 probes (Illumina Inc., San Diego, CA). The microarrays were washed, stained with streptavidin-Cy3, and then scanned with an Illumina BeadArray Scanner. Duplicate arrays were performed for eight cell lines across the different groups for quality control purposes, with duplicates performed on different days. All duplicate arrays showed highest correlation with each other (correlation >0.98). Duplicate samples were not included in analysis. Comparative real-time PCR was performed for ten genes on 6–8 samples, using GAPDH to normalise all data, and the comparative cycle threshold method for analysis. Paired student t tests were performed to determine the significance of gene expression changes. Expression differences were validated for 8/10 genes tested.
Raw data was imported into Illumina Beadstudio and then exported into Genespring v7.3 (Agilent Technologies, Forest Hill, VIC) for further analysis. Data was normalised (per chip normalized to 50th percentile and per gene normalized to median) and filtered using an Illumina detection score of >0.99 in at least one sample, which yielded 20,874 probes that were used in all further analyses. The majority of these probes used in the analysis were designed by Illumina to assay the curated portion of the NIH Ref sequence database-16,923 were present in the Ref sequence database, comprising 65% of all Ref sequence-listed probes on the array. Transcripts which had a >2-fold change versus the mean were visualised using unsupervised Hierarchical Clustering (Figures 1 and 2). The clustering method used was a Pearson correlation similarity measure with an average linkage clustering algorithm. Two different methods were used to classify LCLs based on mutation status: (1) A multi comparison Gaussian Process Classifier (GPC) [42] with Leave-One-Out cross-validation to determine the prediction errors, as previously used to predict BRCA1/BRCA2 mutation status of irradiated fibroblasts [31]; (2) A linear classification method commonly used for classification of microarray data, Support Vector Machines (SVM) [43] with Leave-One-Out cross validation. The GPC analysis used 2031 genes which were derived from a t-test to select the genes that were significantly over/under-expressed at the 5% significance, while the SVM used genes from the 20,874 detected probes which differed between groups of LCLs using a t-test p of 0.05. All resulting gene lists are available as supplementary data and all data is available via GEO: Accession number GSE10905.
Ingenuity Pathway Analysis (Ingenuity Systems, www.ingenuity.com) was used for biological interpretation of gene lists. Analysis of the transcripts found to be up- and down-regulated in irradiated LCLs as identified for the different mutation categories identified those biochemical networks most likely to be affected by a BRCA1 and BRCA2 truncating and missense mutation, relative to BRCAX. Those pathways with multiple hits or a significance score ≥20 were then compared. |
10.1371/journal.pcbi.1000694 | Diffusion, Crowding & Protein Stability in a Dynamic Molecular Model of the Bacterial Cytoplasm | A longstanding question in molecular biology is the extent to which the behavior of macromolecules observed in vitro accurately reflects their behavior in vivo. A number of sophisticated experimental techniques now allow the behavior of individual types of macromolecule to be studied directly in vivo; none, however, allow a wide range of molecule types to be observed simultaneously. In order to tackle this issue we have adopted a computational perspective, and, having selected the model prokaryote Escherichia coli as a test system, have assembled an atomically detailed model of its cytoplasmic environment that includes 50 of the most abundant types of macromolecules at experimentally measured concentrations. Brownian dynamics (BD) simulations of the cytoplasm model have been calibrated to reproduce the translational diffusion coefficients of Green Fluorescent Protein (GFP) observed in vivo, and “snapshots” of the simulation trajectories have been used to compute the cytoplasm's effects on the thermodynamics of protein folding, association and aggregation events. The simulation model successfully describes the relative thermodynamic stabilities of proteins measured in E. coli, and shows that effects additional to the commonly cited “crowding” effect must be included in attempts to understand macromolecular behavior in vivo.
| The interior of a typical bacterial cell is a highly crowded place in which molecules must jostle and compete with each other in order to carry out their biological functions. The conditions under which such molecules are typically studied in vitro, however, are usually quite different: one or a few different types of molecules are studied as they freely diffuse in a dilute, aqueous solution. There is therefore a significant disconnect between the conditions under which molecules can be most usefully studied and the conditions under which such molecules usually “live”, and developing ways to bridge this gap is likely to be important for properly understanding molecular behavior in vivo. Toward this end, we show in this work that computer simulations can be used to model the interior of bacterial cells at a near atomic level of detail: the rates of diffusion of proteins are matched to known experimental values, and their thermodynamic stabilities are found to be in good agreement with the few measurements that have so far been performed in vivo. While the simulation approach is certainly not free of assumptions, it offers a potentially important complement to experimental techniques and provides a vivid illustration of molecular behavior inside a biological cell that is likely to be of significant educational value.
| While reductionist biophysical studies continue to contribute important insights into the properties and functions of biological macromolecules, research attention is increasingly being directed at uncovering the extent to which behavior observed in vitro is likely to reflect that occurring in vivo [1],[2]. In a physiological setting, all biomolecules must inevitably experience non-specific, unintended interactions with the intracellular milieu and there are good theoretical reasons to expect that, even if such interactions are only steric in nature, significant alterations in macromolecular folding and association equilibria may result [2],[3]. In order to allow macromolecules to be directly interrogated in vivo therefore, a number of important developments have been made in the experimental fields of hydrogen exchange [4], nuclear magnetic resonance [5],[6], and fluorescence spectroscopies [7]–[9].
An alternative to the use of experimental techniques is to assemble a molecular model of an intracellular environment in silico and to use molecular simulation techniques to explore its behavior; if such a model could be shown to be realistic – and that is a big ‘if’ – it would have the important advantage of allowing the simultaneous, direct observation of all molecules in the system. In fact, at least two simulation studies that attempt to model the bacterial cytoplasm have already been reported [10],[11], producing a number of intriguing results. Both of these previous studies, however, modeled all cytoplasmic molecules as spheres and it is perhaps to be anticipated therefore that simulations that include structurally detailed macromolecular models might lead to additional insights. In pursuit of this strategy, we have chosen the gram-negative prokaryote Escherichia coli as a test system, combining quantitative proteomic [12] and high-resolution structural data [13] to build a first structurally detailed molecular model of the bacterial cytoplasm.
Full details of the construction of the model are provided in Methods. Briefly, however, it is to be noted that the model contains 50 different types of the most abundant macromolecules of the E. coli cytoplasm (accounting for ∼85% of the cytoplasm's characterized protein content by weight; [12]) and a total of 1008 individual molecules. Eight of these molecules are copies of the heterologous (non-E. coli) protein GFP (Green Fluorescent Protein), which has been included so that the diffusional characteristics of the model can be compared with in vivo experimental results (see below). A snapshot of the modeled system, together with a full listing of its constituents, is shown in Figure 1; the total combined macromolecular concentration in all of the simulations reported here is 275g/l.
Starting from three different randomized initial configurations of the cytoplasm model (all shown in Figure S1), we performed independent Brownian dynamics (BD) simulations [14] to explore diffusive behavior. A variety of energetic descriptions of intermolecular interactions were explored, ranging from a simple steric-only model – which provides an opportunity to directly test the predictions of excluded-volume ‘crowding’ theories [2],[3] – to models that include both long-range electrostatic interactions and short-range potential functions that mimic hydrophobic interactions between exposed non-polar groups. In order to determine the most realistic energy model, the long-time translational diffusion coefficients, DLtrans, of the ‘tracer’ GFP molecules were computed from the BD simulations and compared with previously reported experimental estimates obtained by fluorescence-recovery-after-photobleaching (FRAP) analysis of GFP in the E. coli cytoplasm [15]–[18].
A comparison of the computed GFP DLtrans values obtained with the different energy models is shown in Figure 2A. For simulations in which only steric interactions operate between macromolecules the computed GFP DLtrans value is 3–6 times higher than the experimental estimates, and although this value decreases somewhat when electrostatic interactions between macromolecules are added, it remains 2–5 times too high relative to experiment. A more realistic model of macromolecular interactions would allow favorable short-range attractions to occur between exposed hydrophobic atoms and one simple way of approximating such interactions is to use a Lennard-Jones potential, with the well-depth of the potential, ε, being treated as an adjustable parameter (see Methods). As shown in Figure 2A, the inclusion of such a term results in computed GFP DLtrans values that decrease monotonically as the well-depth, ε, increases in magnitude. The best agreement with experiment is obtained with ε = 0.285 kcal/mol: at this value of ε the computed value of DLtrans – which is ∼10% of its value at infinite dilution – is within the experimental error of all in vivo estimates [15]–[18] including a very recent report for diffusion in cells growing in minimal media [18]. As noted in the Discussion, this optimal value of ε is very similar to the values obtained in our previous efforts to model the interaction thermodynamics of single-component protein solutions [19].
Having determined that good agreement with experiment could be obtained using a so-called ‘full’ energy model that included steric, electrostatic and short-range attractive hydrophobic interactions, we extended each of three independent simulations performed with this energy model to 20µs (see Figure S2 for plots of the system's energy versus time). In order to provide a useful baseline for comparative purposes we also performed extended simulations with the purely ‘steric’ energy model (i.e. one that neglects the electrostatic and hydrophobic interactions); the latter simulations were performed for simulation times of 17.5µs. Each BD simulation using the ‘full’ energy model required more than a year (clock-time) to complete. For both energy models, snapshots taken from the last 15µs of each simulation were used for detailed analysis.
An informative, albeit non-quantitative, impression of the simulation behavior can be obtained by viewing movies of the simulations (Supporting Information). In some respects, these movies can be considered a key result of this work: they represent, in effect, dynamic analogs of the highly influential pictorial representations pioneered by Goodsell [20]. Examination of a typical movie obtained from a simulation performed with the ‘steric’ energy model shows the simulated motion to be rapid, chaotic and obviously Brownian. For the more realistic ‘full’ model, on the other hand, motion is somewhat slower-paced, and molecules can be seen to fluctuate between engagement in short-lived associations and periods of relatively free diffusion.
We can place these observations on a more quantitative footing, and obtain an indication of the extent of sampling achieved in 15µs of simulation, from the remaining panels of Figure 2. Figure 2B shows the maximum distances moved, on average, by each molecule type during simulations performed with the ‘full’ and ‘steric’ energy models; all distances are expressed relative to the diameter of the diffusing molecule. In the case of GFP with the ‘full’ energy model, for example, each molecule travels, on average, approximately 6 molecular diameters (i.e. 320Å) from its position at the beginning of the simulation. Since the data in Figure 2B are plotted versus molecular weight it is apparent that 15µs of simulation is sufficient for the smaller macromolecules to move very significant distances, while for the largest macromolecules (the 30S and 50S ribosomal subunits), little motion away from the initial position is achieved. On this basis alone, therefore, we expect the estimates of diffusional behavior for the smaller macromolecules to be somewhat more precise than those of the larger macromolecules. A second measure of the extent of sampling achieved during each simulation period is provided by plotting the number of unique interaction partners encountered by each type of macromolecule as a function of the simulation time (Figure 2C). Encouragingly, most molecule types encounter many unique neighbors over the course of 15µs: during a typical simulation with the ‘full’ model, for example, each GFP molecule encounters ∼80 different neighbors. Just as importantly, the total numbers of unique neighbors continues to increase even toward the end of the simulation period: this indicates that the cytoplasm model remains highly dynamic and does not tend to ‘freeze’ as the simulation progresses.
As might be expected, the average numbers of neighbors that a macromolecule possesses at any instant scales essentially monotonically with its molecular weight: the average number of macromolecules in the immediate neighborhood of a GFP molecule, for example, is only ∼5 while for the 50S ribosomal subunit it is more than 25 (Figure 2D). The time constants for the dissociation of these neighboring interactions – which in all cases are in the microsecond range – also scale straightforwardly with the molecular weight (Figure 2E), indicating that molecules remain in the neighborhood of larger macromolecules for somewhat longer periods of time than they do with smaller macromolecules. The data shown in Figures 2C and 2D can be combined to provide an estimate of the number of times that each molecule's entire complement of neighbors is replaced during the simulations (Figure 2F). Interestingly, while the overall trend is such that smaller macromolecules encounter a more dynamic constellation of neighbors even the largest macromolecules experience a significant number of environmental changes during the 15µs simulation period. While each GFP molecule, for example, effectively ‘shed its skin’ of neighbors a total of ∼14 times, even the 50S ribosomal subunit undergoes ∼5 such transformations (Fig. 2F). This observation suggests that the limited diffusional exploration carried out by the largest macromolecules evident in Figure 2B may, in at least one important respect, give a misleadingly low indication of the extent of configurational sampling achieved in the simulations: it is in fact, possible for a completely static macromolecule to rapidly encounter widely different microenvironments simply by virtue of the dynamic exchange of its smaller, more mobile neighbors.
While it was noted above that the long-time DLtrans value of GFP obtained with the ‘full’ energy model is in good agreement with in vivo measurements (Figure 2A), there are other aspects of diffusional behavior in the simulations that warrant examination. One question that is of interest is how the observed Dtrans values of macromolecules depend on the observation interval, δt, over which their diffusion is monitored (see Methods). The answer to this question is plotted in Figures 3A and 3B for the three most abundant members of the cytoplasm model (MetE, TufA and CspC); these proteins have been chosen for closer examination because their high abundance yields the most statistically robust numbers, but very similar results are obtained for the other constituents of the model. Figure 3A plots the computed Dtrans values of the three proteins versus δt for both the ‘full’ and ‘steric’ energy models. The clear variation of Dtrans with δt seen for all three proteins is indicative of ‘anomalous’ diffusion [21]–; the magnitude of the anomaly is conventionally expressed by the anomality exponent, α, (Methods) which is plotted for the same proteins, again versus δt, in Figure 3B. Examination of this figure shows that with the ‘steric’ energy model, the diffusion of all three proteins progresses from being normal (α∼1), to transiently subdiffusive (α<1), to normal again as the observation interval increases from δt∼100ps to δt∼10ns to δt∼1µs. With the ‘full’ model, in contrast, macromolecules exhibit transiently anomalous subdiffusion even at the shortest observation intervals; again however, a slow, but unequivocal return toward normal diffusion occurs on a high microsecond timescale. The same qualitative features are seen for all other molecule types although, for the largest macromolecules or those with the very lowest copy numbers, it is not always clear that sampling is sufficient to be absolutely certain of a return to normal diffusion at the longest δt values. At very short values of δt however we can obtain quite precise values of α for all molecule types; when these are plotted versus molecular weight (Figure 3C) it is apparent that while there is a clear difference between the values obtained with the two energy models, and a clear size-dependence of α with the ‘steric’ model, there is no such obvious trend with the ‘full’ model.
For both energy models, the plots of α versus δt fit well to an analytical function (solid lines in Figure 3B) that, when integrated, enables an asymptotic long-time translational diffusion coefficient, DLtrans, to be estimated (see Methods). The observed DLtrans values of all molecule types are expressed relative to their translational diffusion coefficients at infinite dilution (D0trans) and plotted versus molecular weight in Figure 3D. For both energy models, the ratio DLtrans/D0trans decreases with increasing molecular weight, which is qualitatively consistent with experimental studies of tracer protein diffusion in simple single-component protein solutions [24] and of Ficoll diffusion in the cytoplasm of mouse 3T3 cells [25]. The poorer correlation obtained for the ‘full’ model (which does not appear to be solely due to incomplete sampling) suggests that translational diffusion in vivo should not be predictable with arbitrary precision solely from knowledge of molecular weight; again, this is in line with the often significant variations observed in the in vivo diffusion coefficients of similarly-sized GFP-constructs [15],[26]. It is perhaps worth noting, however, that the computed diffusive behavior of the heterologous GFP – marked by an asterisk in the ‘full’ model data points – is consistent with the general trend established by the endogenous E. coli macromolecules.
The rotational motion of macromolecules is also significantly affected by immersion in the cytoplasm model. In the case of the ‘full’ energy model, the rotational behavior can be fit equally well by either a double-exponential function or a model that describes transiently anomalous rotational diffusion [27]. Since it is the rotational behavior on a nanosecond timescale that is more relevant to experimental measurements (see Methods), we plot the short-time rotational diffusion coefficient, DSrot of all molecule types, relative to their rotational diffusion coefficients at infinite dilution, D0rot, in Figure 3E. As would be anticipated given the translational behavior shown above, rotational diffusion is significantly slower with the ‘full’ model than it is with the ‘steric’ model.
Notably, a comparison of Figures 3D and 3E shows that with both energy models rotational diffusion is slowed less by immersion in the cytoplasm than is translational diffusion. This can be viewed as indicating that the two kinds of motion experience different relative viscosities (ηrelT and ηrelR for translational and rotational diffusion respectively). Figure 3F plots the ratio of these relative viscosities, ηrelT/ηrelR, versus molecular weight for all molecule types. For the abundant proteins MetE, TufA, and CspC, and the less abundant GFP, we find the ratio of these relative viscosities, ηrelT/ηrelR, to be 3.6, 3.0, 3.2 and 2.5, respectively using the ‘full’ model; perhaps surprisingly, similar numbers are also obtained with the ‘steric’ model (Figure 3F). These computed ratios are in quite good agreement with the value of ηrelT/ηrelR of 2.6±0.2 obtained from in vitro data for apomyoglobin diffusion in human serum albumin [28] (see Methods) and the value of ηrelT/ηrelR of 2.1±0.3 reported for GFP in Chinese hamster ovary cells [29]; the lower value obtained in the latter case is consistent with the lower macromolecular concentration of the mammalian cytoplasm relative to that of E. coli.
In addition to the simulations providing direct views of diffusive motions in the cytoplasm, snapshots extracted from the simulations offer an important opportunity to explore the thermodynamic consequences of the cytoplasm on macromolecular stability. Using a variant of Widom's ‘particle-insertion’ method [30], the free energy change that accompanies the insertion of a molecule into the cytoplasm can be rigorously computed by subjecting the molecule to millions of randomized placements (see Methods). We used this approach to compute the cytoplasm's effects on the folding equilibria of selected proteins by performing separate insertion calculations on their native state structures and on ensembles of 1000 unfolded structures generated by a sophisticated conformational sampling method [31]. We focused initially on the only two proteins for which experimental estimates of thermodynamic stability in the E. coli cytoplasm are available: (1) a construct of the λ-repressor N-terminal domain, λ6-85 [4], which has been found to have essentially identical stability in vivo and in vitro, and (2) the cellular retinoic acid binding protein [7],[32] (CRABP), which has been found to be thermodynamically destabilized in vivo relative to in vitro. Both of these findings – the latter in particular – are non-trivial results to capture since they are inexplicable in terms of conventional macromolecular crowding theory [2],[3],[7],[33],[34] (see below).
We performed thermodynamic calculations under a total of four different scenarios. The first scenario that we examined involved taking cytoplasm snapshots sampled during the ‘steric’ BD simulations, and computing the cytoplasm-interaction energies of the folded and unfolded conformations with the same ‘steric’ energy model: this scenario corresponds to that considered in conventional models of macromolecular crowding effects [2]. In this case, the differences between the folding free energies in vivo and in vitro are computed to be +1.3±0.0 and +2.2±0.1 kcal/mol for λ6-85 and CRABP respectively (blue bars in Figure 4A), with the positive signs indicating that the folding free energies of both proteins are calculated to be more favorable in vivo than in vitro. When compared to the experimental values (red bars in Figure 4A), these results are in poor quantitative agreement for λ6-85 and are qualitatively wrong for CRABP. In a second scenario, we took cytoplasm snapshots sampled during the ‘full’ model BD simulations, but computed the cytoplasm-interaction energies of folded and unfolded conformations using the simpler ‘steric’ energy model. In this case, the differences between the folding free energies in vivo and in vitro are computed to be +1.0±0.0 and +1.6±0.0 kcal/mol for λ6-85 and CRABP respectively (cyan bars in Figure 4A). The smaller crowding effects obtained in this situation reflect the fact that during the ‘full’ BD simulations transient clustering of molecules creates bigger voids in the system; again however, these computed results are in poor quantitative agreement with experiment for λ6-85 and are in qualitative disagreement with experiment for CRABP.
A third scenario that we examined involved taking cytoplasm snapshots sampled during the ‘steric’ BD simulations and computing the cytoplasm-interaction energies with the ‘full’ energy model. In this case, the differences between the folding free energies in vivo and in vitro are computed to be +0.1±0.5 and −1.8±1.4 kcal/mol for λ6-85 and CRABP respectively (green bars in Figure 4A), both of which, notwithstanding the larger error bars, are in rather good agreement with the experimental results. Finally, we took cytoplasm snapshots sampled during the ‘full’ model BD simulations and computed the cytoplasm-interaction energies with the same ‘full’ energy model. In this fourth scenario – which on the basis of the diffusional properties described above would be hoped to provide the most realistic description (Figure 2A) – the computed changes in stability amount to +0.3±0.1 and −0.9±0.4 kcal/mol for λ6-85 and CRABP respectively (yellow bars in Figure 4A); again, these results are in close quantitative agreement with the experimental results. The overall picture that emerges, therefore, is that the experimental results cannot be reproduced, even qualitatively, when the ‘steric’ energy model is used to score the interactions between the folding protein and the cytoplasm environment, but they can be reproduced – and with a perhaps surprisingly high degree of quantitative accuracy – when the ‘full’ energy model is used in the particle-insertion calculations. Furthermore, the fact that similarly good results are obtained regardless of which energy model was used in the BD simulations suggests that, for such calculations, the method of sampling the cytoplasm's configurations is perhaps less important than the nature of the energy function used to describe the protein of interest's interaction with it.
Histograms of the computed interaction energies of the folded and unfolded state with the cytoplasm explain why the predictions of the ‘full’ model successfully reproduce experiment, and deviate so significantly from the predictions of the purely steric model: for both proteins, but especially so in the case of CRABP, the unfolded state conformations are computed to have somewhat more favorable energetic interactions with the cytoplasm than the folded state conformations (Figure 4B). The consequence is that while the excluded-volume (crowding) effect experienced by both proteins undoubtedly significantly stabilizes their folded states relative to their unfolded states (e.g. see the blue and cyan bars in Figure 4A), the effect is counterbalanced by the more favorable energetic interactions engaged in by the unfolded state conformations.
To explore the potential generality of this latter result, we performed identical calculations for a number of other monomeric proteins using snapshots taken from the ‘full’ model BD simulations; histograms illustrating the size distributions of the unfolded states of the tested proteins are shown in Figure 4C. The computed changes in their folding free energies are plotted in order of increasing molecular weight in Figure 4D. As before, when the ‘steric’ energy model is used to compute the cytoplasm-interaction energies the proteins' stabilities are computed to increase (white bars in Figure 4D); the computed stability changes scale broadly with the molecular weight of the protein, reflecting the greater relative difference between folded and unfolded state dimensions of larger proteins. In contrast, when the ‘full’ energy model is used to compute the cytoplasm-interaction energies, the molecular weight dependence is lost (dark grey bars in Figure 4D): some proteins are computed to be stabilized and others destabilized in vivo relative to in vitro (in no case however is the extent of destabilization sufficient to predict that the proteins will be predominantly unfolded in vivo). These results suggest that differences between the in vitro and in vivo thermodynamic stabilities will vary significantly with the identity of the protein.
We performed similar calculations to explore the potential thermodynamic effects of immersion in the cytoplasm on a variety of protein-protein associations. For the formation of homo-dimeric complexes (Figure 4E), we again find that the excluded-volume crowding effect, which alone stabilizes dimers relative to separated monomers by on average 1.1±0.3 kcal/mol, is largely cancelled by the more favorable energetic interactions that the monomers form with the cytoplasm constituents: when the ‘full’ energy model is used the stabilization of the dimeric forms by the cytoplasm is computed to be, on average, only 0.1±0.3 kcal/mol. For the assembly of the trimeric nucleus [35] of the bacterial cytoskeletal protein ParM from three separated monomers, we find that the stabilization predicted with the ‘full’ energetic model is also significantly lower than that predicted from the crowding effect alone (Figure 4F); again, the smaller value appears more consistent with the close similarities between the polymerization behavior of ParM observed in vitro and in vivo [36]. Finally, we performed calculations on the assembly of two published (but putative) structural models of amyloid-like aggregates [37],[38], each formed by association of 8 monomer units (Figure 4F). For one of these two cases, the aggregation of an SH3 domain [37], we find that the use of the ‘full’ model predicts a slightly greater stabilization than that predicted solely on the basis of the crowding effect; the additional stabilization observed in this case results from the protein's interactions with the cytoplasm being dominated by repulsive electrostatic interactions, which, on average, are diminished in the aggregated state (see Figure S3).
Developing working computational models of intracellular environments is one potential route to understanding differences between biomolecular behavior observed in vitro and in vivo. The simulations and calculations described here represent the first attempt to build such a model for the bacterial cytoplasm using atomically detailed structures of the constituent molecules, and represent the first attempt to directly model the consequences of immersion in the cytoplasm on the thermodynamics of protein stability and protein-protein interactions. It is worth noting that these innovations have been made possible in large part due to the immense progress made by the structural biology community in recent years: in constructing our model it was a major surprise to us to find that, of the 50 most abundant cytoplasmic E. coli proteins identified in the study of Link et al. [12], it was possible to produce complete or near-complete structural models for more than 45 (see Supporting Information). Since large-scale structural genomics initiatives continue to map out the structural proteomes of organisms with ever increasing detail [39] it will be possible to make future generations of cytoplasm models even more compositionally complete.
Before considering the strengths and weaknesses of the present model, and the implications of the results reported here, it is important to reiterate that at least two other cytoplasm models have already been reported in the literature. The first such model was described by Bicout and Field [10] some thirteen years ago. Owing to the comparative paucity of both structural information and computer power then available, the model was restricted to only three types of macromolecule, each of which was modeled as a sphere: their modeled system contained 12 ribosomes, 188 copies of a generic protein of molecular weight 160kDa, and 136 tRNAs. Langevin dynamic simulations were used to model behavior over a timescale of 7.5µs, and four different electrostatic approximations were investigated in an attempt to cover a range of possible simplified descriptions of the ribosome's electrostatic properties. With all four models, the long-time translational diffusion coefficient of the modeled protein was slowed by only ∼40% relative to its infinite-dilution value. Since their work pre-dated the first reports of Dtrans values measured in vivo, Bicout and Field could not know at the time that this simulated diffusion was too fast relative to experiment; they were therefore not in a position to more fully calibrate their model. Despite this issue, it should be clear to readers that the work of Bicout and Field was far ahead of its time. It should also be apparent that, like the influential work of Goodsell [20], it was a direct inspiration for the work reported here.
A second and much more recent model for the bacterial cytoplasm has been developed by Ellison and co-workers [11]. Relative to Bicout and Field's work, the model of Ridgway, Broderick et al. provides an enormous step forward in terms of compositional complexity: >100 different types of proteins are represented, and thanks to the availability of the authors' own proteomic data [40], are present in copy numbers that are likely to much more closely reflect their relative abundances in vivo. On the other hand, all macromolecules are treated as spheres, and intermolecular interactions are assumed to be purely steric in nature. In addition, the actual modeling of motion is somewhat simplified: particles take steps of uniform length in randomly chosen directions, with the steps being accepted only if no collision – or reaction – with a neighboring molecule occurs. While somewhat approximate, this approach has the significant advantage of allowing reactive events to be rapidly modeled, making the simulation model applicable to a more general set of problems than that considered here. The resulting model of the cytoplasm was used to investigate the effects of crowding on the translational diffusion of macromolecules and on the rate of the diffusion-limited association of the barnase-barstar protein-protein complex. As noted by the authors, the diffusional simulations produced only a two-fold decrease in the translational diffusion coefficients of GFP-like molecules, suggesting, in common with the results reported here, that (steric) crowding effects alone are insufficient to explain the ∼10-fold slowed diffusion of GFP observed in vivo.
Relative to these two previous cytoplasm models, therefore, the present approach offers a significant increase in both structural and energetic complexity: all macromolecules are modeled in atomic detail and interact with one another via an energetic model that accounts for the two major types of interaction that drive protein-protein associations (i.e. electrostatic and hydrophobic interactions). It does so, of course, at very significant computational expense: each of the simulations performed with our ‘full’ energy model required more than a year of clock-time to complete. But even with its associated expense it should not be thought that the present model represents the pinnacle of sophistication in terms of its description of reality. Leaving aside the fact that the model is incomplete in terms of the types of macromolecules (and small molecules) that it includes, there are several key assumptions of the modeling that are both important to stress and which represent obvious candidates to address further in future work.
A first simplification of the present approach, and one shared by the previous models described above, is that all macromolecules have here been treated as rigid bodies. This simplification has two consequences. First, it immediately precludes us from making any meaningful attempt to simulate the (presumably very interesting) diffusive behavior of highly flexible macromolecules such as mRNAs and intrinsically unstructured proteins. While this is undoubtedly a limitation, it is to be noted that in terms of their contributions to the overall mass content of the cytoplasm, such molecules play a comparatively minor role relative to that played by the folded, globular macromolecules examined here [10]. It is also to be noted that there are currently very serious technical obstacles to be overcome if the diffusive behavior of flexible macromolecules is to be simulated with any degree of realism: we have shown recently, for example, that the inclusion of hydrodynamic interactions (HI), which are computationally very expensive to compute, is essential if flexible protein models are to adequately reproduce translational and rotational diffusion [41]. A second consequence of the rigidity of the present model is that it is not immediately suited to describing conformational changes that might potentially occur in highly crowded conditions, and for which interesting experimental and simulation results have recently been reported [42],[43]. As shown in the second part of this paper however, this limitation can be overcome, at least for the purposes of calculating thermodynamic effects, by the use of particle-insertion calculations. In fact, the use of such an approach has enabled us to explicitly evaluate the cytoplasm's thermodynamic consequences on both folding and association equilibria, something that would currently be prohibitively expensive to achieve through the direct dynamic simulation of flexible protein models.
A second, but not unrelated simplification adopted in the present approach concerns the energy model used to describe intermolecular interactions. On the one hand, the model is comparatively sophisticated in that it includes descriptions of electrostatic and hydrophobic interactions, and models both at an atomic, or near-atomic level of resolution: in this respect it is a clear improvement over previous models used to simulate the cytoplasm. On the other hand, the model assumes that electrostatic desolvation effects can be neglected (which may lead to an overestimation of the strength of electrostatic interactions; [44]) and treats hydrophobic interactions as pairwise additive [45],[46] and of equal strength for aliphatic and aromatic groups. We assume that the effects of these missing features are at least partly subsumed, in an implicit fashion, within our single hydrophobic parameter, ε. For this reason, we should be careful not to attach too much importance to the absolute value of ε found here (0.285 kcal/mol): it is, nevertheless, encouraging that it is very similar to the range of values that we previously obtained [19] when modeling the thermodynamics of simple dilute protein solutions (0.22–0.28 kcal/mol). This is perhaps especially notable given the enormous difference between the protein concentration studied here (275mg/ml) and that studied in the previous work (10mg/ml).
In future, it should be possible to increase the sophistication of the energy model without incurring an exorbitant additional computational cost: if one stays with a rigid-body approach, for example, a number of grid-based methods might be used that allow electrostatic desolvation [44] and/or hydrophobic interactions [47]–[50] to be rapidly calculated. It should be remembered, however, that a more complicated functional form will not necessarily lead to better results, and that, at least for now, it is highly likely that some degree of empirical adjustment of energy terms will be required in order to reproduce experimental behavior. This will be especially true if the intention is to use a similar model to explore, for example, macromolecular crowding effects on specific protein-protein interactions: despite significant advances, no current computational method is capable of accurately predicting the strength or geometry of specific protein-protein interactions with any generality [51]. To model such situations, therefore, it may be necessary to supplement the energy model with additional short-range forces to drive the formation of known intermolecular contacts, in the same way that such terms (commonly known as Gō-potentials; [52]–[54]) are often used in the modeling of protein folding processes; an alternative might simply be to use different ε values for different protein-protein interactions.
A third limitation of the present model concerns its very simplified description of macromolecular hydrodynamics. In particular, while the basic hydrodynamic properties of all macromolecules (i.e. their translational and rotational diffusion coefficients at infinite dilution) are properly accounted for, the BD simulations reported here do not allow for the presence of hydrodynamic interactions (HI) between macromolecules; again this is true also of the two previously reported cytoplasm models [10],[11]. The immense expense associated with HI calculations remains a major stumbling block to their inclusion in large-scale simulations [55] and a number of attempts have therefore been made to accelerate their computation (see, e.g. [56],[57] for very recent and potentially important examples). This expense would be further increased in the present case if, as would in principle be necessary, an Ewald summation technique was used to properly account for HI in periodic boundary conditions [58].
While simply stating that HI are expensive to calculate does not constitute a compelling reason for leaving them out of the simulations, it is pertinent to note that the omission of HI seems unlikely to be the cause of the gross overestimation of the diffusion coefficient of GFP obtained with the ‘steric’ energy model (Figure 2A). It is certainly true, as noted elsewhere [18], that for hard-sphere-like colloidal particles – where the interactions between particles are extremely short-range – theoretical work has established that the inclusion of HI should cause decreases in Dtrans values over both short [59] and long timescales [60],[61]. Such decreases are, however, unlikely to bridge the ∼5-fold gap necessary to bring the ‘steric’ energy model behavior into quantitative agreement with experiment: in an interesting recent simulation study, for example, it was found that an approximate description of HI in crowded hard-sphere solutions resulted in only a ∼40% additional decrease in the diffusion coefficient relative to simulations without any description of HI [62]. In addition, it is also to be noted that for colloidal particles with long-range repulsive electrostatic interactions, theory indicates that the inclusion of HI causes modest increases in Dtrans values at both short [63],[64] and long timescales [64],[65]. Since the current model has macromolecules interacting with each other not only by short-range steric forces and long-range repulsive electrostatic forces, but also by short-range attractive interactions between exposed hydrophobic residues it is difficult to predict the effects that the inclusion of HI might ultimately cause, other than to say that we think they may be comparatively modest. In keeping with the caveat given above about our energy model, however, we clearly must leave open the possibility that the hydrophobic parameter, ε, is also, in part, serving as an implicit correction for the omission of HI from the simulations.
Having produced in the preceding paragraphs a litany of shortcomings of the model one might be tempted to view it as so fundamentally limited that its practical utility is in doubt. Perhaps the strongest argument against such a view comes from the results of the particle-insertion calculations aimed at computing the thermodynamics of protein folding in vivo (Figure 4A). It is important to note that these thermodynamic calculations should be considered bona fide predictions of the simulation model since it was calibrated to reproduce a quite different experimental observable, i.e. the translational diffusion coefficient of GFP. Because of this, we can rule out the possibility that the calibration of the model predisposes it to trivially reproduce experimental protein stability effects. To our knowledge, the calculated results reported here with our ‘full’ energy model are the first to provide a quantitative rationalization of the experimental observation that CRABP is destabilized in vivo (relative to in vitro) and that λ6-85's relative stability is essentially unchanged. As noted earlier, the experimental CRABP result is inexplicable with conventional macromolecular crowding theory (as exemplified by the results obtained here when the ‘steric’ energy model is used in the particle-insertion calculations) since the dimensions of its unfolded state are greater than those of its native state. Use of the ‘full’ energy model, on the other hand, produces results in close agreement with experiment because it explicitly allows for the two states of the protein to engage in differential, favorable energetic interactions with the rest of the constituents of the cytoplasm. Interestingly, good results are obtained when the ‘full’ energy model is used in the particle-insertion calculations regardless of whether the cytoplasm snapshots were sampled from the ‘steric’ BD simulations or sampled from the ‘full’ BD simulations. Although the most internally consistent approach is obviously to use the same energy model in both the BD simulations and the particle-insertion calculations, the fact that good results can apparently also be obtained using snapshots from the ‘steric’ BD simulations is intriguing since such simulations are much faster to conduct than those using the ‘full’ energy model. Our model's predicted effects on the folding free energies of the six other proteins investigated (Figure 4D) await experimental testing of course, but regardless of how quantitatively accurate such predictions might eventually turn out to be we feel reasonably confident in suggesting that future attempts to understand a protein's folding thermodynamics in vivo will need to describe its interactions with the cytoplasm with more realism than is provided by simple steric interactions.
Other findings from the simulations, while probably more difficult to directly test experimentally, provide examples of the kinds of new information that can be obtained from simulation approaches that attempt to model intracellular environments. Examples include the observation that the immediate neighbors of individual proteins exchange rapidly on a microsecond timescale – even for the very largest macromolecules – and that diffusion is transiently anomalous even on a sub-nanosecond timescale. The latter observation is especially interesting given the current interest in anomalous subdiffusion as an efficient mechanism of search and association in physiological situations [8],[66]. Finally, one might also point to the fact that the simulation model correctly reproduces the cytoplasm's relative translational and rotational viscosities as an important favorable result since differential effects on translational and rotational motion appear to have interesting effects on protein-protein association rates in crowded solutions [67]–[69]. It should be remembered, however, that a similarly good reproduction of the relative translational and rotational viscosities is also obtained with the otherwise poorly performing ‘steric’ energy model.
An examination of all of the dynamic and thermodynamic results described above shows, we think, that it is possible to leverage the existing structural biology and quantitative proteomic data to produce a meaningful, working molecular model of the bacterial cytoplasm. The actual simulation model used here has a number of limitations, of course, but continuing increases in computer power and/or the development of faster simulation methodologies, will likely allow many of these drawbacks to be eliminated in the not too distant future. Given the continuing progress in the fields of structural biology and quantitative proteomics it is likely that the same basic approach might be used to model other intracellular environments.
When this work was initiated, the only large-scale quantitative study of the E. coli proteome was that reported by Link et al. [12] who experimentally measured levels of >200 of the most abundant proteins present in E. coli. A number of other quantitative proteomic studies of E. coli have since been reported [40],[70],[71], and, since this work was completed, quantitative estimates of metabolite concentrations have also become available [72]. Restrictions on computer memory (4GB of RAM for all servers used) meant that the total number of different types of macromolecules that could be realistically modeled was limited to 51: these would be 50 types of E. coli macromolecule plus the Green Fluorescent Protein (GFP). Although including only 50 different types of macromolecules means that the model underestimates the structural diversity of the cytoplasm, it is important to note that the macromolecules selected for inclusion account for 85% (by number of protein chains) and 86% (by mass) of all the cytoplasmic proteins quantified and identified in Table 4 of Link et al. [12].
Of the 50 types of E. coli macromolecules to be included in the model, 45 would be proteins. These were selected by working down the list identified by Link et al. in order of decreasing abundance, selecting only those proteins (a) for which high-resolution structures were then available in the Protein Data Bank [13] (PDB) or for which reasonable homology models could be constructed (see below), and (b) for which the cytoplasm was unambiguously identified as the major cellular location in the EcoCyc [73] and/or CCDB [74] databases. A full list of all potentially cytoplasmic proteins identified and quantified in Table 4 of Link et al. (under minimal media conditions), arranged in decreasing order of chain-abundance, is shown in Table S1; asterisks in the columns headed ‘Mod.’ denote those proteins included in our cytoplasm model. It is an indication of the tremendous coverage of the structural proteome that has been achieved by the structural biology community that we were able to obtain, or build, reasonable structural models for all of the 30 most abundant cytoplasmic proteins identified by Link et al. [12]. In addition to the 45 different types of proteins, 5 types of macromolecule were RNAs or RNA-protein complexes: these were the two ribosomal subunits (50S and 30S), and three typical tRNAs for which structures were available: (tRNA-Gln, tRNA-Phe and tRNA-Cys). It is to be noted that we did not model complete (translating) 70S ribosomes owing (a) to the inherent difficulties in modeling the flexible mRNA, and (b) to the absence – at the time this work was begun – of a three-dimensional structure showing the arrangement of multiple 70S ribosomes in a polyribosome [75].
The total number of molecules in the simulations was set to 1008 (eight copies of GFP and 1000 E. coli macromolecules). This number was chosen so that the eventual assembled cytoplasm model would be large enough to provide a reasonable representation of the environment while still allowing simulations of up to 20µs to be performed (albeit over the course of more than a year clock-time). The linear dimensions of the final modeled system (808.4Å in each of the x, y and z directions) correspond to approximately one-twelfth of the diameter of a typical E. coli cell [76]. A summary of the macromolecules selected, their subunit compositions, the PDB codes of their originating structures, and the degree of sequence coverage achieved by the structural models, is presented in Table S2. Using composition estimates provided by Neidhardt et al. [76] as a guide, we set the total concentration of macromolecules in the model (excluding the ‘tracer’ GFP) to 275 g/l; this is slightly on the low side of the rough values of 300–340 g/l estimated independently by Zimmerman and Trach [77]. Of this, 55g/l (i.e. 20% of the total) is contributed by RNA, with 15% of the RNA dry weight contribution being made by tRNA and the remainder being made by ribosomal RNA [76]. mRNA, which accounts for only ∼4% of the total dry weight of RNA in the cell, is omitted from the present model. The remaining 219g/l (i.e. 80%) of the model is contributed by proteins; this percentage is deliberately set somewhat higher than the 55% contribution to the dry weight of the entire cell estimated by Neidhardt et al. [76] in order to compensate for the missing volume of components that are not explicitly represented in the model (DNA, mRNA, lipid, lipopolysaccharide, murein, and glycogen). If one takes the specific volumes of proteins and RNA to be 0.73ml/g and 0.58ml/g respectively [77], the total volume fraction occupied by macromolecules in the model is 0.19; if instead, an ‘effective’ specific volume of macromolecules suggested by Zimmerman and Trach is used [77] (1.0ml/g), the total volume fraction occupied by the macromolecules in the model amounts to 0.27.
Structures for all selected proteins were identified by performing a BLAST search [78] of the protein's FASTA sequence (as reported in the EcoCyc database) against the entire PDB and selecting the structure with the closest identity to the query sequence using the program BioEdit [79]. The quaternary structure of each selected structure was determined using the PQS web server [80] and was verified, where possible, with the EcoCyc database; it should be noted that correct identification of a protein's quaternary structure is a non-trivial undertaking, and the PQS predictions are unlikely to be 100% reliable [80],[81]. Homology modeling was used for all proteins for which either no E. coli structure was directly available in the PDB, or for which a significantly greater coverage of the sequence could be obtained through the use of a non-E. coli structure. All homology modeling was performed using the SWISS-MODEL web server [82] via the so-called “First Approach mode”; for oligomeric proteins each individual chain was homology-modeled independently.
Any sidechains missing from a structure were built in using the molecular modeling program WHATIF [83]. Hydrogens were then added, and partial charges and radii were assigned to atoms using the program PDB2PQR [84]. For proteins, partial charges and atomic radii were taken directly from the PARSE parameter set [85]. For nucleic acids, which are not represented in the PARSE parameter set, partial charges were instead assigned from the CHARMM23 parameter set [86]; partial charges for the modified bases of tRNAs, such as pseudouridine, were assigned based on similarity to functional groups already represented in the parameter sets. The protonation states of all protein ionizable residues were assigned using the fast empirical algorithm PropKa [87]; for these calculations, the pH was set to 7.6, the estimated pH of the E. coli cytoplasm [76]. With each structure complete, infinite-dilution translational and rotational diffusion coefficients – which are necessary input parameters for BD simulations [14] – were calculated with the program HYDROPRO [88] using default parameters. For the latter calculations we assumed a solvent viscosity, η, of 0.89cP, which corresponds to the viscosity of pure water at 25°C; given that the most recent estimate of the total metabolite concentration in the E. coli cytoplasm is ∼300mM [72] we do not anticipate, based on what we currently know, that the viscosity of the solvent environment will be hugely altered from the pure water value.
The final stage of preparation for each molecule involved the calculation of electrostatic potential grids; these were computed in all cases by using the APBS software [89] to solve the non-linear Poisson-Boltzmann (PB) equation [90]. As in our previous BD study of single-component protein solutions [19], two cubic electrostatic potential grids were computed for each type of macromolecule: (a) a comparatively fine grid, of spacing 2Å, with dimensions sufficient to encompass a 20Å shell around the macromolecular surface, and (b) a coarse, long-range grid, of spacing 4Å, that extends at least 50% further in each direction than the smaller grid. The use of a 2Å grid spacing for the higher resolution grids, rather than the 1Å grid spacing used in our previous simulations [19], was necessary in order to fit all potential grids into the available 4GB of RAM. This spacing is, however, sufficiently detailed that at least two grid points always intervene between interacting atoms even when they are at the closest possible separation distance (4.5Å); significant numerical instabilities in the calculated electrostatic forces do not, therefore, arise. In all PB calculations the solvent dielectric was set to 78.0 and the internal dielectric of the macromolecule was set to 12.0, with the boundary between the two being determined by the cubic-spline surface [91] implemented in APBS [89]. Use of an internal dielectric of 12.0 is intended to provide a simple, averaged description of the different dielectric responses of macromolecular interiors and exteriors [19],[92],[93]. The ionic strength in all PB calculations was set to 150mM. With the electrostatic potentials in hand, ‘effective charges’ were computed for each molecule type using the procedure established by Gabdoulline & Wade [94],[95]. Finally, as in our previous work [19], simulations were accelerated by retaining, in addition to the effective charges, only those non-hydrogen atoms that were solvent-exposed: these atoms were identified using the ACC tool within APBS [89], with a 4Å solvent probe.
The BD software used for the simulations is an extension of the methodology developed and tested in our previous work on pure protein solutions [19]. Modifications were made to the software to improve memory usage so that 102 electrostatic potential grids could be simultaneously held in memory; in addition, toward the end of this study, loop-level parallelization of a number of key loops was implemented with OpenMP (http://www.openmp.org) to accelerate computations by a factor of ∼4.
All simulations were performed under periodic boundary conditions [96] in a cubic cell with edges of 808.4Å. The initial configuration of each system had eight GFP molecules evenly positioned at the center of the eight octants of the cell; all other macromolecules were initially positioned by performing random translations and rotations within the cell subject to the requirement that there was at least a 10Å separation between the surfaces of all neighbors. Three independent configurations were generated in this way by use of different random seeds; views of each system before and after 15µs of simulation are shown in Fig. S1. As in our previous work, BD simulations were conducted using the Ermak-McCammon algorithm [97] with a time step of 2.5ps, with additional algorithmic measures being taken to ensure that no atom-atom distances at the completion of each timestep were less than 4.5Å. For subsequent analysis of the simulations, the 3D translational vector and the 3×3 rotational matrix necessary to specify the position of each macromolecule were recorded every 100ps.
The form of the energy model used to describe intermolecular interactions was identical to that used in our previous work [19]: the effective charge method [94] was used to calculate electrostatic interactions, and a Lennard-Jones potential (comprising 1/r12 and 1/r6 terms) was used to provide a simple combined description of steric, van der Waals and hydrophobic interactions. To accelerate the simulations, the combined non-electrostatic interactions were computed only between atom pairs separated by less than 12Å; a list of all such pairs was continually updated every 40 timesteps (i.e. every 100ps). As in our previous work, we treated the strength of these non-electrostatic interactions, which are determined by the well-depth, εLJ, of the Lennard-Jones potential, as the only adjustable parameter of the model. In order to determine the best setting, three independent BD simulations of at least 6µs duration were performed with each of the following εLJ values: 0.190, 0.285, 0.3325 and 0.380 kcal/mol. Finally, for comparison purposes, two additional sets of three BD simulations were also performed: these were (a) simulations in which the only the repulsive (1/r12-dependent) steric interactions operated (these are the ‘steric’ simulations discussed in the main text) and (b) simulations in which only steric plus electrostatic interactions acted.
The effective translational diffusion coefficients, Dtrans, of molecules were calculated from the simulations using the Einstein equation:(1)where < δr2 > is the mean-squared distance traveled by the molecular center of mass in the observation interval, δt; all Dtrans values reported in Results are mean values for each molecule type averaged over the number of copies of each type. In cases of ‘normal’ diffusion, the computed Dtrans values are independent of δt; in certain cases of diffusion in vivo and in vitro however, anomalous sub-diffusion is observed [8], [21]–[23],[66]; in such cases, the apparent Dtrans value is dependent on δt, decreasing with increasing δt. A common way of describing anomalous diffusion involves writing it in the form:(2)where the apparent translational diffusion coefficient Dtrans is now written to indicate that it depends on the observation interval and α is the so-called anomalous diffusion (anomality) exponent; α = 1 corresponds to normal diffusion since it leads to Dtrans being independent of δt, and α<1 indicates anomalous (sub)diffusion. Taking logarithms and differentiating with respect to log (δt) allows us to write:(3)This enables us to obtain α by numerically differentiating Dtrans values computed over a range of δt values; in practice we computed Dtrans at δt values of 100, 200, 300, 600, 1000, … ps, and obtained α at the logarithmic mid-point, δtmid, of these time-intervals, δtmid = 141, 245, 424, … ps.
Plots of α versus log (δtmid) for macromolecules simulated with both the ‘steric’ and ‘full’ energy models all indicated that α itself was dependent on δtmid, thus signifying that diffusion was transiently anomalous. To our knowledge, there is no explicitly derived functional form that describes the expected dependence of α on δt for transient anomalous diffusion. We found however that the data fit well to the following empirical functional form (see Fig. 3B):(4)where α0 is a constant, a and b are parameters that describe the amplitude of the δt-dependent changes to α, and τshort and τlong are, respectively, the timescales over which α first decreases, and then returns to one, with increasing δt. Plots of α versus δt for all molecule types were fit to the above functional form with SigmaPlot [98]: fits were performed using all datapoints from the shortest δtmid value up to the first datapoint that had a percent error exceeding ∼25% (obtained by comparing the α values computed from the three independent BD simulations), or that deviated qualitatively from the trend. To ensure that the latter criterion did not drastically affect the results, the fits were repeated retaining even those datapoints that qualitatively deviated; essentially the same behavior was obtained but with slightly greater values of τlong. Regressed values of τshort and τlong are plotted versus molecular weight for all molecule types in Figs. S4 and S5 respectively.
Having fit a function to the observed dependence of α on δt, it was numerically integrated to obtain an extrapolated, asymptotic long-time Dtrans value using the Dtrans value at δt = 100ps as the starting point for the integration. The quality of fits of the integrated Dtrans values (for the most abundant proteins) is indicated by the solid lines in Fig. 3A.
Effective rotational diffusion coefficients were computed from the time-dependent behavior of the 3×3 rotational matrix recorded every 100ps for every molecule during the simulations. For each of the three rotational axes, an autocorrelation function, θ (δt), was calculated as:(5)where e (0) and e (δt) are unit vectors pointing along one of the rotational axes at time t = 0 and t = δt respectively, and the brackets indicate an average over all possible initial timepoints; the three computed autocorrelation functions were averaged to give a single decay function consistent with the isotropic rotation that we assumed for all molecule types at infinite dilution. Since the resulting averaged autocorrelation function for the ‘full’ energy model did not fit well to a single-exponential decay, and given that translational diffusion was clearly transiently anomalous, we decided to use the following functional form proposed recently for transiently anomalous rotational diffusion [27]:(6)where θ0 is the value of the autocorrelation function at δt = 0 (always 1), a is a parameter, τrot is a long-time rotational correlation time (which dominates as δt→∞), and τrel is the timescale over which a faster, short-time rotational relaxation gives way to the slower rotation characterized by τrot. The above functional form was fit to computed values of θ for each molecule type over a range of δt values up to 1µs; the r2 values for these fits were all in excess of 0.999. An example of such fits for the most abundant proteins is shown in Fig. S6. The long-time rotational diffusion coefficient, DLrot, is then obtained using the relationship:(7)and the short-time rotational diffusion coefficient, DSrot, is obtained from [27]:(8)The computed ratios DLrot/D0rot and DSrot/D0rot obtained with the ‘full’ energy model are plotted for all molecule types versus their molecular weights in Fig. S7; a plot of the parameter a versus molecular weight shows no obvious relationship (not shown).
Comparison of the simulated translational and rotational diffusion coefficients with the infinite-dilution values that are input parameters for the simulations provides an indication of the relative viscosities experienced during the two types of motion. From studies of GFP diffusion in Chinese hamster ovary cells, the Verkman group reports [29] a relative viscosity experienced by translational motion, ηrelT = 3.2±0.2, and a relative viscosity experienced by rotational motion, ηrelR = 1.5±0.1. Combining these numbers gives a ratio, ηrelT/ηrelR of 2.1±0.3, indicating that the effective relative viscosity experienced by translational motion is roughly twice that experienced by rotational motion in mammalian cells.
A second estimate of the ηrelT/ηrelR ratio can be obtained from the work of Zorrilla et al. [28],[99]: these authors have reported measurements of the translational diffusion coefficients of apomyoglobin (17kDa) using fluorescence correlation spectroscopy (FCS) measurements [28] and have compared them with rotational diffusion coefficients that they had previously measured [99] for the same system using time-resolved fluorescence depolarization experiments. They report measurements for two different background proteins, RNaseA and human serum albumin (HSA); we focus on the data reported for the latter since its molecular weight (67kDa) is much closer to the number-averaged molecular weight of the macromolecules in our cytoplasm model (87kDa), than is the molecular weight of RNaseA (14kDa).
The data reported by Zorrilla et al. are expressed relative to the macroscopic viscosity, ηm, of the protein solution (measured with an Ostwald viscometer). They report that ηm fits to the following functional form, ηm = η0 exp (Ac/(1−Bc)), where η0 is the viscosity of pure water, c is the background protein's concentration in mg/ml, and A and B are background-dependent constants: A = 2.7×10−3 ml/mg and B = 1.3×10−3 ml/mg for HSA [99]. Using these values we obtain a macroscopic viscosity for a 275 mg/ml HSA solution of 3.155 η0. Using the data given in Table 2 of ref. 49, the effective viscosity experienced by the translational motion of apomyoglobin in HSA is expressed as ηrelT = (ηm/η0)1.28, which from above means that we can write ηrelT = 3.1551.28 = 4.35; following similar calculations the effective viscosity experienced by the rotational motion is ηrelR = (ηm/η0)0.44 = 3.1550.44 = 1.66. Together, these numbers translate into a value of ηrelT/ηrelR of 2.6±0.2.
As noted in the main text, we find that both the translational and rotational diffusion coefficients of molecules vary with the time interval, δt, over which diffusion is observed. While the observation of this transient anomalous diffusion is significant in its own right it takes on added significance when comparing the relative viscosities experienced by translational and rotational motion. This is because the timescales over which the two types of experiments are conducted are quite different: translational diffusion coefficients are obtained from FCS experiments by fitting to an autocorrelation function over a timescale extending from microseconds to seconds [21],[22],[66] while rotational diffusion coefficients are obtained from fits to data obtained over a nanosecond timescale [28],[29]. We therefore compare the experimentally derived relative viscosities quoted above with diffusion coefficients computed from the BD simulations on the same timescales, i.e. we compare with the ratio of the long-time translational diffusion coefficient DLtrans and the short-time rotational diffusion coefficient, DSrot (see Fig. 3F).
The intermolecular contacts engaged in by each molecule were recorded every 100ps during the BD simulations and subsequently analyzed to determine: (a) the average number of neighbors of each molecule type at any given time, (b) the number of unique neighbors encountered by each molecule type during the course of the entire simulations, and (c) the rate of dissociation of intermolecular interactions. The definition of ‘neighbor’ was kept somewhat loose in order to detect all molecules in the immediate environment of the molecule being probed: molecules were assigned as neighbors if any of their atoms were within ∼12Å of each other. The rates at which the neighbors of a particular molecule dissociated were obtained from plots of the fraction of its neighbors, initially present at t = 0, that remained after some time t = δt, averaged over all possible initial timepoints. In order to obtain the characteristic neighbor-decay rate for each particular type of molecule, such plots were averaged over all molecules of that type. The resulting plots are found to follow biexponential kinetics: (a) a very fast decay process (τfast) that typically has an amplitude of ∼0.7 and is due to loss of neighbors that interact only peripherally with the molecule of interest, and (b) a slower decay process (τslow) that has an average amplitude of ∼0.3 and is due to loss of those neighbors that form bona fide intermolecular contacts. Typical fits for these data are shown in Fig. S8.
The effects of immersion in the cytoplasm on the thermodynamics of protein folding and protein-protein association were computed using the particle insertion technique first outlined by Widom [30]. For small perturbations, the free energy change, ΔG, for transferring a molecule from an environment free of any interacting macromolecules to the cytoplasm environment can be rigorously expressed as:(9)where Eint is the interaction energy of the molecule with the constituents of the cytoplasm, R is the Gas constant, T is the temperature, and the brackets indicate an average over randomly selected insertion positions and configurations of the cytoplasm environment. In order to assess the likely effects of the cytoplasm on a thermodynamic process (such as protein folding) therefore, separate particle-insertion calculations are required for both the initial state (e.g. unfolded protein) and the final state (e.g. folded protein). Such calculations give the free energy changes for the vertical processes in the thermodynamic cycle shown below:(10)Since free energy is a state function, the difference between the free energy changes of the horizontal processes is equal to the difference between the free energy changes of the vertical processes. We can therefore write the difference between the free energy change for the process in vivo and in vitro, ΔΔG, as:(11)The effect of the cytoplasm on the free energy change for a process can therefore be calculated without needing to know the actual value of the free energy change for the process in vitro. A conceptually similar but different approach to computing thermodynamics in crowded solutions has recently been outlined by Zhou and co-workers [100]. Code for performing particle-insertion calculations was generated by modifying the existing BD simulation program; prior to performing large-scale explorations of protein folding and association thermodynamics, the code's correctness was first checked by comparing its predictions for the free energy cost of placing a sphere into a solution of spheres with the corresponding predictions of scaled particle theory [101],[102].
Calculations of the cytoplasm's thermodynamic effects initially focused on protein folding equilibria. In addition to calculating the folding thermodynamics of six proteins already present in the cytoplasm model (Adk, Bcp, CspC, Efp, GFP and PpiB), we examined two other proteins that have been subject to direct experimental study in vivo: these were the 80-residue λ6-85 construct studied experimentally by Ghaemmaghami and Oas [4] and the 136-residue cellular retinoic acid binding protein (CRABP) investigated by Ignatova, Gierasch and co-workers [7],[32]. The structure of the folded state of λ6-85 was taken from its crystal structure in complex with operator DNA (pdbcode: 1LMB [103]); the G46A & G48A mutations present in the experimental construct were made using the rotamer-sampling method SCWRL3 [104]. The structure of the folded state of CRABP (pdbcode: 1CBI [105]) was altered to include the R131Q mutation used in the experimental construct [7], but in the absence of direct structural information no attempt was made to model the experimentally-incorporated fluorophore.
The unfolded states of all eight proteins were modeled as ensembles of 1000 unfolded conformations generated using the conformational sampling method developed by the Sosnick group [31]; the code was kindly made available by Dr. Abhishek Jha. This method has been shown to produce models with dimensions in good agreement with experimental estimates [31]. Prior to calculations, the structures of all conformations were completed by adding sidechains with SCWRL3 [104] and by adding hydrogens with the PDBTOPQR utility [84] of APBS [89]. In order to ensure consistency between the BD simulations and the Widom particle-insertion calculations, effective charges and electrostatic potential grids were calculated for all conformations (both folded and unfolded) using the exact same protocol employed with the rigid protein models of the cytoplasm model (see above).
For each protein, a large number of random trial positions were attempted with both the single, folded state structure and the 1000 unfolded state conformations; each trial consisted of a different randomly selected translation and rotation. For the folded state structure, a total of 25 million trials were attempted; for the unfolded state, 250,000 trials were attempted for each of the 1000 conformations (to give a total of 250 million trials for each cytoplasm ‘snapshot’ studied). For each trial position, the interaction energy of the protein with the surrounding cytoplasm was calculated with (a) the ‘full’ energetic model, which includes electrostatic, steric and hydrophobic contributions, and (b) the ‘steric’ energetic model. To simplify the latter calculations, only two possible energies were allowed: the interaction energy, Eint, was set to +∞ if any of the protein's atoms came within 4.5Å of any of the cytoplasm atoms, and was set to zero if not; this binary scoring method is effectively identical to that used in most examinations of excluded-volume (crowding) effects. Due to the very significant computational expense associated with the particle-insertion calculations, they were applied only to the final ‘snapshot’ of the three independent BD simulations performed with the ‘full’ and ‘steric’ models. Error bars for all reported free energy changes were therefore calculated as the standard deviation of the computed values obtained from the three different system ‘snapshots’. The total number of unfolded and folded-state trial positions that were accepted and rejected for each protein, for each of the three ‘full’ model cytoplasm ‘snapshots’ are listed in Table S3.
A very similar protocol was used to calculate the effects of the cytoplasm on a variety of protein association reactions. Calculations on each assembled protein complex were performed exactly as described above. Calculations on each disassembled complex – e.g. two separated protein monomers in the case of a dimerization reaction – were carried out by performing insertions of all components simultaneously; importantly, each randomized placement was first screened to ensure that there were no steric clashes between any of the inserted components before their interactions with the cytoplasm were evaluated. As might be expected, the requirement of simultaneously placing multiple molecules into the cytoplasm meant that in some cases very large numbers of trial positions were required in order to obtain reasonably converged results. Owing to the significant computational expense, therefore, calculations were only performed on snapshots taken from BD simulations performed with the ‘full’ energy model. In addition, since the Boltzmann-weighting of the sampled interaction energies can contribute significant noise in cases where the number of accepted placements are comparatively low, the cytoplasm-interaction energy distributions were first smoothed by fitting to sums of three Gaussians using SigmaPlot [98] (see Fig. S9 for a typical fit). The total numbers of accepted and attempted insertions for the various association reactions studied are listed in Table S4.
Dimerization equilibria were investigated by performing separate particle-insertion calculations on the dimeric forms and the monomeric forms; for such calculations it was assumed that no structural change (e.g. unfolding) occurs when the two monomers are separated. The trimerization equilibrium of ParM was investigated in analogous fashion, by performing calculations on a trimer extracted from the ParM filament model (pdbcode: 2QU4 [106]). The aggregation of a poly-Q-inserted RNaseA to form an amyloid fiber was studied using the theoretical model developed by Eisenberg and co-workers (pdbcode: 2APU; [38]). The model deposited in the PDB contains 56 aggregated monomeric units; the largest aggregate for which we could obtain reasonably precise free energy estimates however contained eight monomeric units (Fig. 4F). Since formation of the amyloid structure involves a significant change in conformation, the use of monomeric structures extracted without modification from the aggregate model would be inappropriate. Instead, the structure of the monomeric poly-Q-inserted RNaseA was taken from the crystal structure reported by the Eisenberg group (pdbcode: 2APQ [38]). In order to ensure sequence-consistency with the amyloid model, a A131H mutation was made with SCWRL3 [104]. In addition, since the monomeric structure has no resolved coordinates for the inserted GQQQQQQQQQQGNP stretch this region was model-built using the loop-building program Loopy [107]. The second aggregate structure studied was a theoretical model of SH3 domain aggregation proposed by the Shakhnovich group [39] and kindly made available to the authors by Dr. Feng Ding (UNC; personal communication). This structure contains only Cα atoms so complete backbone coordinates were first constructed using the SABBAC webserver [108] (http://bioserv.rpbs.jussieu.fr/cgi-bin/SABBAC) before sidechain positions were constructed using SCWRL3. Owing to the structure's origins being a Cα-only model we were unable to add sidechains in such a way that the assembled aggregate model was free of internal steric clashes; this, however, does not significantly affect our ability to estimate the model's interaction with the cytoplasm environment. As with the RNaseA amyloid model, it would be inappropriate to assume that the conformations of unaggregated monomeric units are identical to those found in the amyloid model; instead therefore the conformation of the monomeric SH3 domain was taken from the crystal structure (pdbcode: 1NLO [109]).
Two movies, each showing 1.8µs of simulation, are provided as separate Quicktime .mov files. Video S1 shows a BD simulation performed with the ‘full’ energy model; Video S2 shows a BD simulation performed with the ‘steric’ energy model. File size restrictions at the PLoS website have limited the size and resolution of the uploaded movies to be used for review. Higher resolution movies are available to readers at the authors' website: http://dadiddly.biochem.uiowa.edu/Elcock_Lab/Movies.html.
|
10.1371/journal.pgen.1000484 | Multipotent Genetic Suppression of Retrotransposon-Induced Mutations by Nxf1 through Fine-Tuning of Alternative Splicing | Cellular gene expression machinery has coevolved with molecular parasites, such as viruses and transposons, which rely on host cells for their expression and reproduction. We previously reported that a wild-derived allele of mouse Nxf1 (Tap), a key component of the host mRNA nuclear export machinery, suppresses two endogenous retrovirus-induced mutations and shows suggestive evidence of positive selection. Here we show that Nxf1CAST suppresses a specific and frequent class of intracisternal A particle (IAP)-induced mutations, including Ap3d1mh2J, a model for Hermansky-Pudlak syndrome, and Atcayhes, an orthologous gene model for Cayman ataxia, among others. The molecular phenotype of suppression includes ∼two-fold increase in the level of correctly-spliced mRNA and a decrease in mutant-specific, alternatively-processed RNA accumulating from the inserted allele. Insertional mutations involving ETn and LINE elements are not suppressed, demonstrating a high degree of specificity to this suppression mechanism. These results implicate Nxf1 in some instances of pre-mRNA processing, demonstrate the useful range of Nxf1CAST alleles for manipulating existing mouse models of disease, and specifically imply a low functional threshold for therapeutic benefit in Cayman ataxia.
| Retroviruses and transposable elements are molecular parasites that integrate into the host genome and require host cell machinery for gene expression, replication and dissemination. Integrating elements can alter the expression of nearby host genes through both transcriptional and post-transcriptional mechanisms. Components of the host cell machinery that can adapt to favor genetic programs of the host cell over those of the parasite may afford one level of innate immunity. In laboratory mice, endogenous retroviruses are virus-derived mobile elements that account for many spontaneous mutations. A frequent class involves retrotransposition into introns of genes in the transcriptional sense orientation, which alters host gene pre-mRNA splicing. Here we show that for the intracisternal A particle (IAP) family of endogenous retroviruses, an allele of the canonical mRNA export factor Nxf1 found in wild Asiatic mice (Mus musculus castaneus) suppresses most insertions of this class (six of seven tested). To our knowledge, these results make Nxf1 the most broadly interacting modifier gene yet documented in this well-studied species. These results have significant implications for manipulating gene expression in mouse models of disease, the role of Nxf1 in pre-mRNA processing and in the dynamic range for therapeutic intervention in Cayman ataxia.
| Retroviruses and transposable elements both utilize host cell factors for their own expression and influence the expression of adjacent host genes through a variety of mechanisms. Components of host cell gene regulatory machinery that interact with molecular parasites may be regarded as components of innate immunity if they can discriminate between host and parasite expression [1]. The generality and exploitability of any given mechanism is an important practical question. Nuclear-cytoplasmic export of RNA is an important point of contact between molecular parasites and host genomes that may fit this criterion for several molecular parasites in mice and humans [2],[3]. We have previously reported that a wild-derived allele of Nxf1, which encodes the major mRNA nuclear export factor, can significantly suppress two mutations caused by insertions of endogenous retroviruses into introns of cellular genes by modulating their mature transcript levels ∼2 fold [4]. A 16 kb transgene containing the full Nxf1 haplotype, but no other recognized gene, was able to confer the modifier phenotype. Whether this interaction could be generalized to a broader class of insertional events, and if so for what range of insertions, was limited by the relatively small number of events examined.
Nxf1 (also called Tap) was first described as a cellular factor that interacts with the Tip protein of herpesvirus saimiri [5] and subsequently shown to be an essential host factor for nuclear export of unspliced viral genomes of simple retroviruses [6]. Although recruitment of Nxf1 to cellular mRNPs may generally be mediated by protein contacts [7],[8], both Nxf1 and its yeast homolog Mex67p also bind RNA directly [9]–[11]. In mammals, known direct targets of Nxf1 include both exogenous and endogenous viral RNAs as well as host sequences [6], [12]–[14]. In addition, we previously reported that one Nxf1 haplotype shows hallmarks of recent positive selection in wild Mus musculus castaneus accessions [4], which may suggest a host-pathogen interaction mediated by Nxf1 in wild populations.
Endogenous retroviruses (ERVs) are non-infectious molecular parasites that are frequent mutagens in mice. Several families of ERV are highly polymorphic among classical inbred strains and among wild accessions [15]. In laboratory mice, ERV insertions account for 10–15% of spontaneous mutations [16],[17], depending on the strains from which estimates are drawn. The intracisternal A particle (IAP) and MusD/early transposon (ETn) families of ERV, which account for most of these, have different apparent rates of transposition in different inbred strains: IAPs appear to be particularly active in C3H strains and ETn elements in A strains [16]. Characteristics of autonomously active copies have been described [18]. Interestingly, the size distribution for newly integrated ETn elements is both broader and, on average, a lower percentage of full length than for IAP elements [19]. As both families are thought to have derived originally from infectious viruses, mechanisms that regulate ERVs or mitigate their impact on host genomes may have broader implications for both gene expression and host-parasite interactions.
To test the range of insertion events for which the modifier activity of Nxf1CAST is effective, we examined gene expression, visible phenotypes, or both for five additional IAP, one LINE, and seven ETn insertion alleles. The host genes cover a wide range of phenotypes, expression patterns, and biochemical pathways:
Here we show that Nxf1CAST suppresses six of six IAP insertions of the IΔ1 class [38], the most frequent class of new insertions, but does not suppress a full-length IAP, a L1-LINE, nor any of six ETn insertion mutations. We quantify RNA and protein levels to show a consistent ∼2-fold increase in normal gene expression from the mutant allele in each case of suppression. Concomitant decrease in the expression of mutant-specific RNAs implicates Nxf1 in pre-mRNA processing in addition to its known role in mRNA export. For disease models and other mouse mutations induced by IAP-IΔ1 retrotransposition, Nxf1CAST provides a genetic rheostat for gene activity in situ.
To test whether Nxf1CAST can suppress the RNA processing defects in AtrnmgL and Mgrnmd, we examined whole brain RNA of progeny from genetic crosses to Nxf1CAST, comparing homozygous mutant littermates that differ in Nxf1 genotype. Because each of these crosses also segregated other loci contributing to coat color, we did not assess pigmentation phenotypes for these two mutants.
For Atrn (Figure 1), abnormally processed message from mgL alleles are detected on Northern blots by probes containing exons 5′ to the insertion site, but not by the 3′ untranslated region ([25] and Figure 1A, B). Because the large but low-abundance normally spliced message was difficult to quantify reliably from Northern blots, we used TaqMan quantitative RT-PCR to assay RNA abundance in mgL mutant brains. Comparing mgL to control animals shows non-significant reduction in abundance of 5′ sequences (Figure 1C), but ∼6-fold loss of full-length transcript, represented by an assay 3′ to the mgL insertion (Figure 1D). However, this assay shows no effect of Nxf1 genotype on Atrn expression.
In contrast, for Mgrn, Nxf1-dependent differences in the level of correctly and alternatively spliced RNA isoforms from md alleles were readily quantified (Figure 2). A probe 5′ to the md insertion (Figure 2A) detects both normal and mutant-specific Mgrn RNAs (Figure 2B). Correctly processed normal RNA is elevated in the presence of Nxf1CAST, while levels of several mutant-specific transcripts is decreased (Figure 2B–D), consistent with the mode of suppression previously reported for Pitpnvb and Eya1BOR. A probe 3′ to the insertion detects only the correctly spliced form, at levels comparable to the 5′ probe (not shown). Quantitative RT-PCR across the inserted intron confirms a ∼2-fold increase in correctly-spliced transcript levels by Nxf1CAST (Figure 2E).
To test Nxf1CAST activity on a mutation for which protein level and phenotype were accessible, we analyzed RNA and protein levels, coat color (eumelanin) dilution and tremor severity of Ap3d1mh2J mutant animals (Figure 3). Locations of the mh2J insertion and probes are indicated in Figure 3A. Although Northern blots show high variance between experiments, comparisons between paired subjects examined on each blot shows a statistically significant increase in normal-sized Ap3d1 transcript and a modest decrease in mutant-specific transcript in the presence of Nxf1CAST (Figure 3B–D). Quantitative RT-PCR confirms the increase in correctly spliced RNA (Figure 3E). Western blots show a corresponding increase in full-length Ap3d1 protein levels detected by an antibody to N-terminal residues (Figure 3F,G). Correspondingly, a smaller protein species detected only in mutant animals is decreased in Nxf1CAST animals. As predicted from this molecular analysis, Ap3d1mh2J mutant animals also had improved pigmentation and neurological assessment scores in the presence of Nxf1CAST as rated by observers blinded to genotype (Figure 3H–J).
We similarly tested Nxf1CAST activity on molecular and visible phenotypes of Usp14axJ (Figure 4). The insertion and probes used are indicated in Figure 4A. Quantification of Northern blots and RT-PCR experiments from brain RNA shows significantly increased levels of correctly processed RNA in the presence of Nxf1CAST (Figure 4B–D). Quantification of Western blots shows that this is translated into an increased level of Usp14 protein (Figure 4E,F). Behaviorally, Usp14axJ mutant animals also showed improved neurological assessment scores, with visibly reduced tremor amplitude in the presence of Nxf1CAST (Figure 4G and Videos S1 and S2). In contrast to other mutations suppressed by Nxf1CAST, normalized levels of mutant-specific isoforms of Usp14 RNA did not differ significantly by Nxf1 genotype. Comparing Northern blots hybridized with either 5′ or 3′ probes (as indicated in Figures 2–5), we find Usp14axJ and Eya1BOR differ from other suppressed mutations in producing RNA isoforms that contain 5′ exons, IAP sequences and 3′ exons [4],[28] where most others produce primarily 5′ exons and terminal IAP sequences.
To test Nxf1CAST activity in the context of a human disease model, we analyzed several levels of molecular and behavioral phenotypes for the Atcayhes mutation (Figure 5). The locations of the hes insertion and probes are indicated in Figure 5A. Atcayhes alleles express prominent mutant-specific Atcay RNAs and very low levels of correctly processed full-length RNA [29]. Northern blots to quantify size-specific RNA levels show reduced level of each mutant-specific RNA detected by a probe 5′ to the insertion (Figure 5B,C). A probe 3′ to the insertion detects only the full length “normal” RNA and is quantifiable only in non-mutant samples (not shown). To measure levels of normal RNA in mutant samples, we used a quantitative RT-PCR (TaqMan) assay (Figure 5D). The presence of Nxf1CAST significantly increases the level of correctly processed Atcay RNA accumulating from hes alleles. This difference is also translated into higher levels of the encoded Caytaxin/BNIP-H protein (Figure 5E,F). Atcayhes mutant animals have profound ataxia and an unusual jumping behavior (see Video S3). Nxf1 genotype had a highly significant impact on Atcayhes neurological phenotypes as rated by multiple observers blinded to genotype, including both reduced ataxia and complete elimination of jumps from open field behavior (Figure 5G and Videos S3 and S4).
To test a non-viral class of retrotransposon, we examined whether Nxf1CAST would suppress the black-eyed white L1-LINE insertion allele of Mitf. This mutation results in loss of pigmented melanocytes and extreme white spotting, leaving only occasional patches of pigment on the head or ears. Despite this low threshold for phenotype modulation, and known effects of other strain backgrounds, we saw no evidence for modification by Nxf1CAST in an F2 cross. Among 14 Mitfmi-bw, Nxf1B6 and 9 Mitfmi-bw, Nxf1CAST doubly homozygous progeny, we observed a single animal of each genotype with dark patches on the head or ears.
We tested Nxf1CAST activity on both sense and antisense-oriented ETn insertions of recent origin in both BALB/cJ and A/J mice. Expression levels of Zhx2 and its repression target Afp were assayed by quantitative RT-PCR from adult liver at P40 from 24 BALB/cJ x B6-Nxf1CAST F2 animals selected by genotype (Figure 6A,B). The BALB/cJ-derived insertion allele expressed ∼1.5% non-mutant levels of Zhx2, with no difference between Nxf1 alleles. Similarly, the effect on Afp persistence, potentially a more sensitive indicator of Zhx2 function, showed no significant difference between Nxf1 alleles, although inter-individual variation was high (Figure 6B, right panel), likely due to other factors segregating in this cross [39].
We tested the ability of Nxf1 to elevate transcript levels for another 5 sense and 3 antisense intronic ETn insertions in a second cross, A/J x B6-Nxf1CAST (Figure 7). Genomic organization and the location and orientation of the insertions are indicated (Figure 7A). Quantitative RT-PCR measurements from brain or muscle (depending on known pattern of expression for each gene) showed no significant differences between Nxf1 genotypes for either sense or antisense insertions (Figure 7B,C). A fifth sense-oriented insertion, in Prkca, showed no difference between inserted and uninserted alleles for either RNA or protein levels in this cross.
Among sense-oriented IAP elements, only AtrnmgL was not suppressed by Nxf1CAST; as the inserted intron does not appear to be differentiated in position, length, or sequence composition from mutations that were suppressed (Figures 1–5 and data not shown) we determined the DNA sequence of each of these inserted elements, as well as the original Pitpnvb insertion [4],[40]. We amplified each insertion using high-fidelity PCR optimized for long sequences, using unique primers flanking each insertion site (Supplemental material online). Ap3d1mh2J, Atcayhes, Mgrn1md, Pitpnvb and Usp14axJ insertions all amplified fragments of 5.5 to 6.0 kb, while the AtrnmgL insertion required modified conditions to support adequate amplification of a unique ∼8 kb product. DNA sequence analysis showed that the AtrnmgL element is a full length (type I) IAP, while each of Nxf1CAST-sensitive elements includes the 1.9 kb deletion of gag-pol sequence typical of type IΔ1 elements [38] (Figure 8A). All 6 elements belong to the IAPEz subfamily (www.repeatmasker.org), and contain an RTE-D transport element [41],[42] near the 3′ LTR. Calculated trees for each segment of aligned sequence shows that the full length AtrnmgL element is not otherwise an outlier in overall sequence composition, except for the undeleted region of the gag gene (Figure 8B). Inclusion in the tree of two recently identified IAP-IΔ1 insertions, Atp2b2jog and Gria4spkw1 [43],[44], suggests that they too should be sensitive to Nxf1CAST-mediated suppression as they fall within sequence clades of suppressed elements for each segment.
Our results show that Nxf1CAST suppresses a broad and frequent class of IAP-induced mutations. The magnitude of increased normal transcript is ∼2-fold and the impact on gene expression and behavioral phenotypes are significant in each case of this class examined. Nxf1CAST increases the steady-state level of correctly spliced host gene transcript and almost always decreases the level of mutant-specific alternatively spliced transcript for six of seven sense-oriented IAP insertions examined to date (Table 1). The one exception, AtrnmgL, differs from all of the suppressed elements we sequenced in having an intact gag-prt-pol coding sequence. Sequences within the deleted region may therefore mediate an additional level of Atrn repression that is not relieved by Nxf1CAST. Each insertion, including AtrnmgL also had a number of more subtle sequence variations, including smaller indels and further studies will be required to clarify which sequence differences contribute to the lack of suppression. However, the current data do provide a clear guide for the class of insertional mutation most likely to be quantitatively modulated by Nxf1CAST, type IΔ1 IAPs, which are by far the most frequent class recovered from spontaneous mouse mutations. While it is possible that other genes within the congenic interval contribute to any one effect, transgenic mouse and lentiviral gene transfer studies with Pitpnvb indicate that the main effect is due to Nxf1, as do the consistency of findings across all six suppressed mutations. Negative data from six ETn-inserted loci indicate that Nxf1CAST is highly selective, and therefore unlikely to result in collateral changes in gene expression when used to manipulate IAP-induced mutations. Indeed, preliminary microarray data failed to identify any significant expression changes in whole brain RNA (B.A.H., unpublished data).
The simplest explanation for the molecular data from the six mutations suppressed by Nxf1CAST would be for Nxf1 to participate in pre-mRNA processing prior to the completion of splicing. This could occur by recruitment of Nxf1 to the nascent transcript by sequences in the IAP (or proteins bound to them co-transcriptionally) and subsequent interactions between Nxf1 and other components of the mRNP. Under such a model, amino acid differences (S48P and E610G) between the allelic Nxf1 proteins would alter the balance of alternative splicing either directly through interactions with splicing machinery or indirectly through an effect on transcriptional elongation rate or preference for termination site in the insertion. An alternative explanation might be for the export activity of Nxf1 to drive the nascent RNP into a territory with different relative activities for splicing and degradation, but this seems more difficult to reconcile with simultaneously increased levels of the correctly spliced message and decreased levels of the mutant splice form in five of the six suppression events.
Nxf1 protein interacts with several factors that could influence alternative splicing, including U2AF35 [45], several SR proteins [7],[8],[46],[47], and components of the TREX complex [48],[49]. Nxf1 is also recruited to the class of retroviral RNA transport elements (RTE-D), found in the IAPs we sequenced from suppressed mutations, through its interaction with RBM15 (OTT1) [42], which has also been linked to both splicing and export of Epstein-Barr virus mRNA [50]. Although these interactions are generally interpreted as recruiting export factors to mature RNPs [51], recruitment of Nxf1 to the nascent transcript through retroviral or cellular RNA transport elements could, in principle, alter the recruitment or activity of splicing factors. Both the RNA binding activity and much of the known protein interaction network around Nxf1 are conserved with respect to the Saccharomyces homolog, Mex67p [11],[48],[52]. It is interesting in this context that in splicing-specific RNA profiling of yeast mutations with defects in mRNA production the expression profile of MEX67-deficient strains cluster with transcriptional elongation factors [53]. Altered elongation rate is thought to be one mechanism that can regulate alternative splicing [54] and recruitment of Nxf1 to elongating nascent transcript could in principle alter the assembly or kinetics of other factors on the elongating pre-mRNA.
The extension of suppressor activity to a wider class of insertional mutations has several practical implications. First, these results predict that Nxf1CAST should be able to modify other mutations that involve similar IAP insertions, for which new examples continue to be reported [43],[44],[55]. Indeed, the recent description of an IAP allele of Pofut1 notes variable reduction of phenotype among F2 progeny in a cross to CAST/Ei, the strain from which the suppressing allele of Nxf1 was derived [55]. The congenic Nxf1CAST stock we have developed should be a useful tool to allow in situ titration of gene expression from either spontaneous or engineered alleles involving such insertions. Second, the range of titration in each of the six cases we have examined is ∼1.5 to 2-fold and semi-dominant. This holds over a fairly broad range of mutational effects on gene expression, ranging from ∼2% and 4% of wild-type levels (unsuppressed and suppressed, respectively) for Atcayhes to 50% and 75% for Eya1BOR. Finally, our in vivo gene titration results across six different mutations suggests that for a wide range of loci and allele strengths, even modest recovery of function may have dramatic phenotypic benefits. This is strikingly true in the case of Atcayhes, where even a 2% increment of expression has a dramatic impact on behavioral phenotype (Videos S3 and S4). This implies that for Cayman ataxia, even a small amount of recovery in biochemical or cellular function would have substantial therapeutic benefit.
We have now demonstrated suppressor activity of the Nxf1CAST allele toward six different mutations with distinct biochemical and physiological properties in the mouse. To the best of our knowledge this is now the most broadly validated suppressor or modifier gene activity in this well-studied species.
Congenic C57BL/6J (B6)–Nxf1CAST mice were derived in our laboratory [4] and maintained by backcrossing to B6. Crosses described here were initiated with a stock at N19 or later backcross generation. C3H/HeJ–AtrnmgL and B6–Mgrnmd were obtained from Dr. Teresa Gunn, Cornell University; mixed stock–Ap3d1mh2J and C3H–Atcayhes from Dr. Margit Burmeister, University of Michigan; B6–Usp14axJ from Dr. Scott Wilson, University of Alabama, Birmingham; and B6–Mitfmi-bw from Dr. Lynn Lamoreux, Texas A&M University. A/J and BALB/cJ were purchased from the Jackson Laboratory. Mice were maintained in specific pathogen-free conditions in accordance with protocols approved by the University of California at San Diego IACUC. Phenotypic comparisons were carried out using littermate pairs. Scores for behavioral phenotypes were assessed by at least 3 trained observers blinded to genotype. Videos of representative behaviors are available online as supporting information.
Genotypes for Nxf1 and each insertional mutation were determined by custom PCR assays for each locus. Conditions for PCR of full-length insertions were optimized using a commercial kit (MasterAmp Extra-Long PCR Kit, Epicentre) and primers in unique flanking sequences. DNA sequence analysis from the resulting PCR products used standard methods, as previously implemented in our laboratory [56] and assembled in Sequencher 4.8. Primers and PCR conditions are provided in the supporting information. Sequence alignments and neighbor-joining trees were performed in MUSCLE [57],[58] on the European Bioinformatics Institute web site (http://www.ebi.ac.uk/).
Freshly dissected tissues were homogenized in Trizol reagent (Invitrogen) and processed for RNA according to the manufacturers instructions. Poly(A)+ RNA was purified by oligo(dT) cellulose chromatography. Northern blots were prepared from formaldehyde-agarose gels by capillary transfer to Hybond-N membranes and crosslinked by exposure to 2400 J UV light. Probes were prepared from cDNA fragments by random primer labeling. Hybridizations to each filter were quantified by phosphorimage analysis (Storm, Molecular Dynamics) and normalized to subsequent hybridization of Gapd to the same membrane as an internal control. Quantitative PCR assays were performed on total RNA. TaqMan assays for Atrn (Applied Biosystems, assays Mm00437738_m1 and Mm01270975_m1) and Atcay (Mm01172843_m1) were performed by the UCSD Center for AIDS Research Genomics Core Laboratory and normalized to a Gapd TaqMan assay. All other quantitative RT-PCR experiments were performed using intron-spanning primers that flank the inserted intron, detected by SYBR green fluorescence in a Bio-Rad CFX96 instrument, and quantified by the ΔΔCt method. Measurements were performed in triplicate for each sample. Samples to be compared were measured on the same plate during a single run. Custom primer sequences and conditions are provided as Tables S1, S2, S3, and S4 online.
Freshly dissected tissues were homogenized in CelLytic M Cell Lysis reagent (Sigma #C2978) plus protease inhibitors and quantified using a bichromate assay (BCA, Pierce). Samples were subjected to SDS-PAGE and Western blotting onto Hybond-ECL membranes. Antibodies and dilutions used were goat anti-Ap3d1 (Rockland, 1∶1000), rabbit anti-Caytaxin/BNIP (Gift of Dr. Low Boon Chuan [59], 1∶5000), rabbit anti-Usp14 (Bethyl Laboratories, 1∶5000). Relative levels of immunoreactivity were quantified using infrared dye-coupled secondary antibodies (Rockland, 1∶10,000) on a LI-COR imager and normalized to rabbit anti-PITPβ [40] as an internal control that correlated with BCA-measured total protein.
Summary data are plotted in figures as mean values, with error bars indicating standard deviations. For variables with expected normal distributions, including quantitative PCR experiments and behavioral observations in which several observers rated performance against a calibrated scale, hypotheses were tested using paired or unpaired t-tests depending upon whether the underlying materials were from explicitly paired samples (e.g., matched littermates) or aggregates (e.g., sibs and cousins). For variables expected to have non-normal distributions across trials (including blotting procedures, in which normalization and scaling across experiments complicate the analysis, and paired samples for which some replicate pairs represent different ages or breeding designs) hypotheses were tested using a nonparametric Wilcoxon signed-ranks test applied to replicates of paired experimental measures. Statistical calculations were carried out in Microsoft Excel or SISA online, http://www.quantitativeskills.com/sisa/ [60] (t-tests) or using the VassarStats public web interface, http://faculty.vassar.edu/lowry/VassarStats.html (Wilcoxon tests).
|
10.1371/journal.ppat.1003702 | Quaternary Structure of Pathological Prion Protein as a Determining Factor of Strain-Specific Prion Replication Dynamics | Prions are proteinaceous infectious agents responsible for fatal neurodegenerative diseases in animals and humans. They are essentially composed of PrPSc, an aggregated, misfolded conformer of the ubiquitously expressed host-encoded prion protein (PrPC). Stable variations in PrPSc conformation are assumed to encode the phenotypically tangible prion strains diversity. However the direct contribution of PrPSc quaternary structure to the strain biological information remains mostly unknown. Applying a sedimentation velocity fractionation technique to a panel of ovine prion strains, classified as fast and slow according to their incubation time in ovine PrP transgenic mice, has previously led to the observation that the relationship between prion infectivity and PrPSc quaternary structure was not univocal. For the fast strains specifically, infectivity sedimented slowly and segregated from the bulk of proteinase-K resistant PrPSc. To carefully separate the respective contributions of size and density to this hydrodynamic behavior, we performed sedimentation at the equilibrium and varied the solubilization conditions. The density profile of prion infectivity and proteinase-K resistant PrPSc tended to overlap whatever the strain, fast or slow, leaving only size as the main responsible factor for the specific velocity properties of the fast strain most infectious component. We further show that this velocity-isolable population of discrete assemblies perfectly resists limited proteolysis and that its templating activity, as assessed by protein misfolding cyclic amplification outcompetes by several orders of magnitude that of the bulk of larger size PrPSc aggregates. Together, the tight correlation between small size, conversion efficiency and duration of disease establishes PrPSc quaternary structure as a determining factor of prion replication dynamics. For certain strains, a subset of PrP assemblies appears to be the best template for prion replication. This has important implications for fundamental studies on prions.
| Prions are infectious agents causing irremediably fatal neurodegenerative diseases in human and in farmed or wild animals. They are thought to be formed from abnormally folded assemblies (PrPSc) of the host-encoded prion protein (PrPC). Different PrPSc conformational variants associated with distinct biological phenotypes, or ‘strains,’ can propagate in the same host. To gain some structural information on the physical relationship between packing order (i.e. quaternary structure) and the strain-specific biological information, we previously subjected PrPSc assemblies from prion strains classified as fast or slow (according to their survival time in susceptible laboratory animals) to sedimentation velocity ultracentrifugation experiments. For the fast strains specifically, the most infectious assemblies sedimented slowly and partitioned from the bulk of PrPSc macromolecular complexes. By changing the solubilization and sedimentation conditions, we established here that a small PrPSc aggregation size and not a low density accounts for these hydrodynamic properties. We further showed that these small assemblies resist proteolytic digestion and outcompete by several orders of magnitude the larger-size assemblies in cell-free prion conversion assays. Thus PrPSc quaternary structure appears to be a determining factor of prion replication dynamics. For certain strains, a discrete subset of PrPSc assemblies appears to be the best template for prion replication.
| Prion disease pathogenesis stems from the post-translational conversion of the monomeric, alpha helix-rich host-encoded prion protein (PrPC) into misfolded, β sheet-enriched PrPSc aggregates [1]. The process is believed to be initiated by PrPSc seeds [2], [3] acquired through infection or arising from spontaneous conversion of wild-type or mutant PrPC into PrPSc [4]. The PrPSc seeds would template the remodeling of host PrPC to the PrPSc form [5]. This self-sustained polymerization process, -in which polymer fragmentation is thought to play a key role [2], [6], [7]-, leads to deposition of injurious plaques into the brain. PrPSc-templated conversion of PrPC or bacterially-derived PrP has been established in cell-free conditions using protein misfolding cyclic amplification (PMCA) assays (for reviews [8], [9]), further strengthening the conformational changes of the prion protein as the main molecular determinant of prion replication and infectivity.
Prion diseases can occur in many mammalian species. Among them are human with Creutzfeldt-Jakob disease, sheep and goat with scrapie, cattle with bovine spongiform encephalopathy (BSE) and cervids with chronic wasting disease [10]. A variety of prion variants or strains exist within a given host species. They cause diseases with specific phenotypic traits, including time course to disease and neuropathological features. Differences in PrPSc biochemical (e.g. resistance to proteases) and biophysical properties [11], [12], [13], [14], [15] indicate that strain-specific biological properties reflect differences in the PrPSc “conformation” associated to each strain [16], [17], [18]. PrPSc has not been amenable to high-resolution structural studies [3], due notably to its insolubility in non-denaturing detergents. Thus the conformational underpinnings of the prion strain phenomenon and notably the contribution of PrPSc quaternary structure remain largely elusive. Conceivably these differences must be sufficiently local to allow faithful prion transmission at least within and between individuals of the same species. Non-PrP components might be part of prion infectious particle or act as a scaffold during the conversion and/or aggregation process and thus might also contribute to prion strain biological phenotype (reviews: [3], [19]).
To gain some structural information on the physical relationship between prion infectivity and PrPSc aggregation state, and how it varies among strains, we previously applied a sedimentation velocity (SV)-based fractionation technique to solubilized brain homogenates from ovine PrP tg338 transgenic mice infected with distinct scrapie and BSE cloned prion strains [20]. Based on the incubation time to disease in tg338 animals, these strains were classified as fast and slow. These experiments led to the observation that the relationship between prion infectivity and PrPSc aggregation state was not univocal. Regardless of the strain, the bulk of proteinase-K (PK) resistant PrPSc was found to sediment in the middle part of the gradient. While for the slow strains, the distribution of infectivity tended to correlate with that of PK-resistant PrPSc, for the fast strains specifically, infectivity peaked markedly in the upper top gradient fractions, which were much less populous in PK-resistant PrPSc aggregates. Although SV is known to separate protein aggregates according essentially to their size, density can often influence their sedimentation properties, thus questioning which parameters would account for the hydrodynamic properties of the fast strain most infectious component.
Here, fractioning the same ovine strains by sedimentation equilibrium (SE) demonstrates that the density properties of prion infectivity and PK-resistant PrPSc tend to overlap regardless of the strain, fast or slow, and the solubilization conditions. This indicates that a reduced PrPSc aggregation size and not a low density essentially account for the SV properties of the most infectious assemblies from the fast strains. We further show that these SV-isolable, small sized infectious assemblies perfectly resist limited protease-induced proteolysis and that their templating activity by PMCA outcompetes that of the bulk of larger size aggregates by several orders of magnitude.
Animal experiments were carried out in strict accordance with EU directive 2010/63 and were approved by the authors' institution local ethics committee (Comethea; permit number 12/034).
The cloned ovine prion strains used in this study have been previously described [20]. They have been obtained through serial transmission and subsequent biological cloning by limiting dilutions of classical and atypical field scrapie and experimental sheep BSE sources to tg338 transgenic mice expressing the VRQ allele of ovine PrP. Pooled or individual tg338 mouse brain homogenates (20% wt/vol. in 5% glucose) were used in centrifugation analyses, as indicated.
The entire, standard procedure was performed at 4°C unless specified otherwise. Mouse brain homogenates were solubilized by adding an equal volume of solubilization buffer (50 mM HEPES pH 7.4, 300 mM NaCl, 10 mM EDTA, 2 mM DTT, 4% (wt/vol.) dodecyl-β-D-maltoside (Sigma)) and incubated for 45 min on ice. Sarkosyl (N-lauryl sarcosine; Fluka) was added to a final concentration of 2% (wt/vol.) and the incubation continued for a further 30 min on ice. For SV, a volume of 150 µl was loaded on a 4.8 ml continuous 10–25% iodixanol gradient (Optiprep, Axys-shield), with a final concentration of 25 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA, 1 mM DTT, 0.5% Sarkosyl. For SE, a volume of 220 µl was mixed to reach 40% iodixanol, 25 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA, 1 mM DTT, 0.5% Sarkosyl final concentration and loaded within a 4.8 ml of 10–60% discontinuous iodixanol gradient with a final concentration of 25 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA, 1 mM DTT, 0.5% Sarkosyl.
The gradients were centrifuged at 285 000 g for 45 min (SV) or at 115 000 g for 17 hours (SE) in a swinging-bucket SW-55 rotor using an Optima LE-80K ultracentrifuge (Beckman Coulter). We found that 5 hours was the minimum time to run proteins at the equilibrium in the optiprep medium. Gradients were then manually segregated into 30 equal fractions of 165 µl from the bottom using a peristaltic pump. Fractions were aliquoted for immunoblot, bioassay or scrapie cell assay analyses. Gradient linearity was verified by refractometry. To avoid any cross-contamination, each piece of equipment was thoroughly decontaminated with 5 M NaOH followed by several rinses in deionised water after each gradient collection. To ascertain the efficiency of the decontamination procedure, solubilized, uninfected brain homogenates were fractionated at the equilibrium. Some resulting fractions were inoculated to tg338 mice (see below). Those were euthanized healthy at 500 days post-inoculation. Their brain and spleen were negative for PrPSc content.
Digitonin (0.1% final concentration; Sigma) or saponin (0.5% final; Sigma) or methyl-β cyclodextrin (10 mM final; Sigma) were added before or after solubilization with dodecyl-β-D-maltoside and Sarkosyl. The incubation was performed for further 30 min on ice.
Brain homogenates from tg338 mice infected with LA21K fast prions (20% wt/vol. in 5% glucose) were adjusted to a final concentration of 25, 50 and 100 µg/ml proteinase K and incubated under constant agitation at 37°C for 1 hour. The digestion was blocked with phenylmethylsulfonyl fluoride (10 mM final concentration; Roche). Undigested samples treated in the same conditions were used as controls. The samples were solubilized and fractionated by SV as described above. The fractions were then inoculated to tg338 reporter mice to estimate their infectivity (see below).
Aliquots of the collected fractions were treated or not with a final concentration of 50 µg/ml PK (1 hour, 37°C). Samples were then mixed in Laemmli buffer and denatured at 100°C for 5 min. The samples (15 µl) were run on 12% Bis-Tris Criterion gels (Bio-Rad, Marne la Vallée, France) and electrotransferred onto nitrocellulose membranes. In some instances, denatured samples (100 µl) were spotted onto nitrocellulose membranes using a dot-blot apparatus (Schleicher & Schuell BioScience (Whatman)). Nitrocellulose membranes were probed for PrP with 0.1 µg/ml biotinylated anti-PrP monoclonal antibody Sha31 as previously described [20]. Thy.1, flotillin and caveolin proteins were probed with anti-CD90.1 (Southern Biotec), anti-flotillin-1 (Abcam) and anti-caveolin-1 (Abcam) antibodies, respectively. Immunoreactivity was visualized by chemiluminescence (GE Healthcare). The amount of PrP present in each fraction was determined by the GeneTools software after acquisition of chemiluminescent signals with a GeneGnome digital imager (Syngene, Frederick, Maryland, United States). The PrP sedimentation profiles obtained by immunoblot were normalized to units and decomposed using multiple Gaussians fits procedures with a maximum entropy minimization approach.
Fractions (unless specified otherwise) were diluted extemporarily in 5% glucose (1∶5) in a class II microbiological cabinet according to a strict protocol to avoid any cross-contamination. Individually identified 6- to 10-week old tg338 recipient mice (n≥5 mice per fraction) were inoculated intracerebrally with 20 µl of the solution, using a 26-gauge disposable syringe needle inserted into the right parietal lobe. Mice showing prion-specific neurological signs were monitored daily and euthanized at terminal stage of disease. To confirm prion disease, brains were removed and analyzed for PK-resistant PrPSc content using the Bio-Rad TsSeE detection kit [21] before immunoblotting, as above. The survival time was defined as the number of days from inoculation to euthanasia. To estimate what the difference in mean survival times means in terms of infectivity, strain-specific curves correlating the relative infectious dose to survival times were used, as previously described [20].
The Rov-cell assay technique will be published elsewhere. Gradient fractions aliquots (20–30 µl) were methanol precipitated as done previously [20], before resuspension in Rov cells culture medium. We verified that methanol precipitation did not affect the overall infectious titer of the samples to titrate. Rov cell [22] monolayers established in a 96 well plate were exposed to the fractions for one week. After several washes with sterile PBS, the cells were further cultivated for two weeks before fixation and PrPSc detection by immunofluorescence using the ICSM33 anti-PrP antibody (D-Gen Ltd, [23]). Immunofluorescent PrPSc signals were acquired with an inverted fluorescence microscope (Zeiss Axiovision). The signal was quantified per cell per well, as previously described [20]. Serial tenfold dilutions of infected brain homogenates were prepared in the same conditions and run in parallel experiments to establish a tissue culture infectious dose curve that directly relates to the PrPSc content.
The modified PMCA procedure will be published elsewhere. It has been adapted from previously described protocols [24], [25]. The PMCA substrate was composed of 10% (wt/vol.) tg338 brain homogenate in PMCA buffer (Tris-HC 50 mM pH 7.4, 1% Triton X-100, 150 mM NaCl). Serial ten-fold dilutions of fractions either as pool or individuals were mixed with substrate lysate in 0.2 ml thin-wall PCR tubes containing beads. Tubes were placed in the Misonix S3000 or Q700 sonicator horns (Misonix, Farmingdale USA; Delta Labo, France) for a round of 96 cycles. Each cycle consisted of a 30 s sonication step at ∼200–250 W followed by a 29.5 min incubation at 37°C. Negative controls were run in parallel. They were composed of unseeded substrate or seeded with uninfected fractions. Aliquots of the amplified samples were digested with PK (100 µg/ml final concentration) for 1 h at 37°C before denaturation in Laemmli sample buffer and dot- or western-blot analysis as described above.
PrPSc and infectivity from fast prion strains exhibited dissimilar hydrodynamic properties by SV, the most infectious assemblies sedimenting slowly [20]. While the detergent used to solubilize brain homogenates disrupted the membrane integrity and released PrPC in the soluble phase [20], -suggesting efficient solubilization conditions-, a tight and specific association of fast prion strains infectivity with lipids, which would also float in the gradient upon ultracentrifugation, could not be totally excluded. To address this possibility, we examined the distribution of LA21K fast infectivity in more stringent solubilization conditions, with the detergents dodecyl maltoside and sarkosyl used sequentially at 37°C instead of 4°C [26], before standard SV fractionation in an iodixanol (Optiprep) gradient [20]. For each fraction, PK-resistant PrPSc was detected by immunoblot and infectivity was measured with a Rov cell-based assay, as previously described [20]. As a result, solubilization at 37°C did not significantly modify the distribution of infectivity in the gradient: the most infectious fractions were found in the top of the gradient, fractions 1 and 2 being 100–1000 fold more infectious than the middle fractions 12–16 containing the bulk of PK-resistant PrPSc (Figures 1 A–B).
To gain resolution in the SV profile, the ultracentrifugation time was doubled. As shown in Figure 1 C, the infectivity peak shifted from fraction 1–2 to fractions 2–4 while PK-resistant PrPSc was found to sediment toward the heaviest fractions of the gradient [12]–[26]. However the shift of infectivity downward was considered as too slight to firmly exclude an intrinsically low density. We therefore decided to study the density of PrPSc and infectivity of the fast strains by sedimentation at the equilibrium. This was compared to that of the slow strains, for which infectivity and PK-resistant PrPSc SV profiles overlapped [20]. Sedimentation equilibrium (SE) allows macromolecules reaching a position in the centrifuge tube at which their own density equals that of the gradient density, independent of time. To achieve this, the sample is mixed with the gradient material (encompassing a wider range of densities than for SV) and the sample is run for a long period of time (reviewed in [27]).
To separate PrP assemblies by density, solubilized brain homogenates were centrifuged isopynically in 10–60% discontinuous iodixanol gradient for 17 hours at 115 000 g. The gradient was then fractionated in 30 fractions of equivalent volume and PrP distribution was assessed by immunoblotting. Three or more independent fractionation experiments with different pooled or individual brains were performed for each strain to assess the reproducibility of the partition and to enable quantitative analysis of the data. In uninfected (Figure 2A, D) as in infected brain (Figure 2B, E) homogenates, PrPC was found in fractions 14–26 and peaked in fraction 18–20, i.e. at a density of ∼1.23–1.28 g/ml (Figure 2A). Other GPI and/or lipid rafts-associated proteins such as Thy1 and flotillin were found in the PrPC-enriched fractions or in the vicinity (Figure 2 A,C), further supporting the view that the conditions employed here led to efficient solubilization of proteins present in detergent resistant microdomains.
The combined curves resulting from the replicate analysis of PrP content indicated that PK-resistant PrPSc aggregates from five ovine strains, - two fast strains, 127S (Figure 2B) and LA21K fast (Figure 3A) and 3 slow strains, LA19K, sheep BSE and Nor98 (Figure 3 B–D) -distributed in two major populations peaking in fractions 8–10 and 12–14, i.e. at respective density of ∼1.115 and ∼1.145 g/ml, nearby that of caveolin, another lipid rafts resident, but oligomeric protein (Figure 2C). Only the proportion of PK-resistant PrPSc per peak varied to a significant degree among the strains.
The distribution of infectivity was assessed by a tg338 mouse incubation time bioassay, using one fractionation performed with pooled brains. It was repeated partially with one strain (Nor98) to confirm the reproducibility of the method. In striking contrast with SV [20], the distribution of infectivity at the equilibrium broadly overlapped that of PK-resistant PrPSc, whether the strain was fast or slow. Thus, for all the strains, fractions 8 to 14 were the most infectious, based on the mean survival times of the mice that succumbed to disease (Table 1). The mean survival times of mice inoculated with the fractions at the two PrPSc density peaks rarely differed to a significant level (Figure S1). Standard infectious dose/survival time curves established individually for each strain tested here [20] indicated that the fractions of higher density were at least 100–1000 less-fold infectious than the most infectious fractions (Figure 3). There was some strain-dependent variation in the distribution of infectivity in the top fractions of very low density (Figure 3). While for LA21K fast, LA19K and sheep BSE the differences in survival times between the upper top fractions 1–4 and the most infectious fractions 8–14 were statistically significant, those did not always reach significant values for Nor98 (Figure S1). For the LA21K fast strain, this provided a 100 to 1000-fold difference in infectious titer between the top and most infectious fractions (Figure 3A). For this strain, the cumulated infectivity of the most infectious fractions by SE approached that previously found in the top fractions by SV [20]. This further supported the view that the most infectious population isolated by SV was indeed present in the middle of the SE gradients.
The SE distribution profile of LA21K fast infectivity was similar when the mouse incubation time bioassay was substituted with the Rov cell assay (n = 3 independent experiments, compare Figure 3A and Figure 4A). Thus differences in survival times were correlated with differences in infectivity content and not different pathogenic effects. The infectivity distribution profile associated with the other fast strain, 127S was closely related to that of LA21K fast (Figure 4B), as measured by the scrapie cell assay (n = 3 independent fractionation studies; Figure 4B) or partly by the incubation time bioassay (Table 1). For both LA 21K fast and 127S, the relative infectious levels at the two PrPSc density peaks rarely differed one from the other significantly, as estimated by the Rov cell assay (Figure S2).
Collectively, these data showed a good correlation between the density profile of infectivity and that of PK-resistant PrPSc aggregates, regardless the “speediness” of the prion strain.
To further ascertain that the relative overlap, at the equilibrium, in the distribution of PrPSc and infectivity of the fast strains truly reflects a physical association with respect to density, we studied the impact on their sedimentation profile of alterations in the solubilization procedure. We added saponin or digitonin (two closely related detergents) or the drug methyl-β cyclodextrin before or after the solubilization with dodecyl maltoside and sarkosyl. These agents are known to specifically deplete or sequester membrane lipids such as cholesterol [28], [29], [30]. The solubilization was performed at either 4°C or 37°C to increase the treatment stringency. This was tested on the 127S fast strain. None of the molecules tested modified PrPC sedimentation profile (data not shown). Only digitonin modified the distribution profile of PK-resistant PrPSc at the equilibrium. The peak of lower density in fraction 8–10 was blurred leading to a Gaussian-like distribution of the protein centered in fraction 13 (Figure 4C). This digitonin effect was observed at 4°C and 37°C, independently of the order in which the detergent was used (data not shown). Adding digitonin to the solubilization procedure led to the evolution of 127S infectivity density profile towards a single peak consistently associated with PK-resistant PrPSc (n = 3 experiments, Figure 4C). Such effect was not observed with saponin and methyl-β cyclodextrin (data not shown). Together these data further reinforces the view that the density of PrPSc and infectivity of the fast prions strains are physically associated.
To conclude with SE experiments, all the data gained using this technique concur to the view that small size and not low density is mostly responsible for the distinctive hydrodynamic properties of the fast strain most infectious component by SV and its partitioning from the bulk of PrPSc.
Having undoubtedly identified that PrPSc aggregates from the fast strains segregated in two populations of differing size and infectivity level by SV, we next examined their respective resistance to PK treatment. This was motivated by the low content of PK-resistant PrPSc of the most infectious population (<10%; Figure 1 and [20]) and the reported existence of small sized PK-sensitive aggregates [31], [32]. LA21K fast brain homogenates were treated with concentrations of PK (0–100 µg/ml) for 1 hour at 37°C prior to SV fractionation. These concentrations were chosen to completely digest PrPC while preserving PK-resistant PrPSc ([21] and unpublished observations). The most infectious fractions (1+2) and the fractions in which PK-resistant PrPSc levels were peaking (12+13) were then pooled, respectively, and inoculated to reporter tg338 mice to assess their relative infectivity levels by incubation time bioassay. This was done in two independent experiments summarized in Table 2. In both experiments, the mean survival time of mice inoculated with the top fractions was marginally prolonged upon the different PK treatments. It would correspond to a reduction <0.5 Log10 of the infectious titer. In contrast, the mean survival time of mice inoculated with the middle fractions was increased by 7 to 18 days upon PK treatment, i.e. a potential reduction of infectivity of >1 Log10. Together these data did not reveal an unusual susceptibility to PK of the LA21K fast, small size most infectious assemblies. The effect of PK treatment appeared even more significant on the larger size PrPSc assemblies.
SV fractionation and the PMCA technique were used to compare the templating efficiency of LA21K fast PrPSc assemblies with differing size and infectivity levels. Serial ten-fold dilutions of the upper most infectious fractions [1]–[3], intermediate PK-resistant PrPSc enriched fractions [12]–[14] and heavy [20]–[22], [28]–[30] fractions were mixed with uninfected tg338 brain lysate and run for one PMCA round of 48 hours. Four independent experiments were performed using four independent fractionations. In each experiment, fractions were amplified in triplicates. The PMCA products were then treated with PK and analyzed by dot-blot based immunoblotting (Figure 5). A positive PrPres signal was observed after PMCA amplification of the upper fractions 1–3 diluted up to 106–107-fold. In sharp contrast, no PrPres signal was detected when the other pools of fractions were diluted more than 104-fold before the PMCA reaction. Assuming a straight correlation between PMCA activity of the fractions and PrP assemblies' content, the specific templating activity per unit PrPres would be 1000 to 10 000-fold higher for the discrete population of ‘small’ PrPSc oligomers than for the bulk of higher size PrPSc assemblies.
Our initial SV studies revealed striking divergence in the hydrodynamic properties of the most infectious assemblies between distinct ovine prion strains from the same host species. For fast strains specifically, the most infectious assemblies sedimented slightly and were associated with low levels of PK-resistant material ([20] and this study). To carefully separate the respective contributions of size and density to these hydrodynamic characteristics, we varied the solubilization conditions and performed sedimentation at the equilibrium. Incidentally this is the first study that compared the density of prion particles associated with phenotypically distinct strains propagated on the same genetic background. All these experiments concurred with the view that a reduced aggregation size but not a low density accounts for the low SV properties of the fast strain most infectious component. We also provided evidence that these SV-isolated, small sized infectious species resist limited PK-proteolysis and have high templating efficiency as suggested by PMCA assay. Together, the straight relationship between small sized PrP assemblies, conversion efficacy and short incubation time observed for the fast strains establishes PrPSc quaternary structure as a determining factor of prion (strain specific) replication dynamics.
Running the ovine prion strains at the equilibrium revealed that PrPSc sedimented in two major density peaks, their respective proportions varying among fast and slow strains. The density values of the 2 PrPSc peaks were markedly reduced compared to that of PrPC, suggesting volumetric differences between these two isoforms. Biophysical, structural and molecular dynamics studies have revealed that the transition from the α-helical to the β-sheet enriched conformation had profound effects on recombinant PrP hydration and packing [33], [34], [35], these two properties directly affecting the volume of a protein. Caveolin-1, a major, -supposedly oligomeric [36], [37], [38]- component of ubiquitous plasma membrane invaginations termed caveolae [39] segregated, at the equilibrium, from monomeric lipid raft resident proteins such as Thy1 and flotillin, further supporting the overlooked notion that oligomerization could markedly alter protein density.
The existence of two PrPSc density peaks is intriguing and will obviously deserve further investigations. First, this may reflect PrPSc molecular mass variations within the brain, which can affect density [40]. Endogenously, PrPSc is differentially trimmed by certain nerve cell subpopulations [41], [42], [43]. The resulting amino-terminal deletion may additionally affect PrP hydration and cavity distribution [44]. Besides, PrPSc aggregation state polymorphism may contribute to differential hydration, as observed with β2-microglubulin fibrils [45]. There is no real consensus over the volumetric properties of amyloid fibrils. They can be associated to compaction or less packed structures [46], [47]. PrPSc binding to ligands, some being known to target the N-terminal part of PrP (for review [48]) could also affect PrP density [33]. The strain-dependent proportions of PrPSc at the peaks of density would be consistent with these hypotheses: prion strains target specific brain area and can exhibit differential PrPSc processing [42], [43], different aggregation states [20] and binding to specific ligands might be strain-dependent [49].
Given all the possible reasons for heterogeneous PrPSc density, the alteration in the PrPSc density profile of fast 127S (Figure 4C) and slow LA19K strains (Figure S3) upon addition of digitonin to the solubilization procedure remains difficult to explain. Its specificity of action as compared to other cholesterol-depleting agents, its absence of effect on the SV properties of PrPC and PrPSc (Figure S4) together with a yield of protein solubilization equal or inferior to that achieved with dodecyl maltosite [20], [50], [51] are strong arguments against an increase in the solubilization yield. Thus differences of densities are more likely to reflect differences in the properties of the bound-detergent species.
At the equilibrium, PrPSc and infectivity sedimented relatively congruently, whatever the prion strain studied, yet infectivity was not distributed in two clearly distinct peaks of densities like PrPSc. There are differences in the infectivity density values previously published [52], [53], [54] and ours, which are likely explained by the use of different starting material, distinct gradient medium and the degree of solubilization achieved. Our density values found for caveolin, -a protein recovered in fractions nearby PrPSc and infectivity-, are consistent with those published [55]. Importantly, the density distribution of PrPSc and infectivity from the 127S fast strain were jointly altered by digitonin. This result strengthened the truly physical association between PrPSc and infectivity with respect to the density of the fast prion strain assemblies. Collectively, these data indicate that a small size and not a low density accounts for the hydrodynamic behavior of the fast strains most infectious component by SV. Keeping in mind all the uncertainties in determining the molecular mass by SV, we estimated previously that these assemblies might correspond to a pentamer of PrP, if constituted of PrP only [20]. However this value might be underestimated as we showed here that PrP density/volumetry has been dramatically altered by its refolding into PrPSc.
There is clear evidence that a variable, strain-dependent proportion of PrPSc can be fairly sensitive to PK treatment [56], [57], [58], [59]. Such PrPSc species have been proposed to be formed of low molecular weight aggregates [31], [32]. PK-sensitive PrPSc has been shown to support a substantial fraction of infectivity [59], [60], -although this might be strain dependent [57], [61]-, and to have a substantial in vitro converting activity [31], [62]. The PrPSc content associated with fast strains such as 127S or LA21K fast resists fairly harsh PK treatment conditions, notably compared to Nor98/atypical scrapie ([21] and unpublished data). Subjecting LA21K fast crude brain homogenate to a PK treatment destroying 99% of PK-sensitive PrPSc infectivity [59] prior to SV fractionation negligibly affected the infectivity associated to the small sized assemblies, as measured reproducibly by the incubation time bioassay. These results are consistent with the inability to detect thermolysin-resistant PrPSc [20], that might be indicative of the presence of PK-sensitive molecules [57], [63]. Counter-intuitively, the infectivity of LA21K fast higher size PrPSc assemblies appeared more sensitive to the PK treatment than that of the smaller ones, suggesting possible differences in the tertiary structure between the 2 populations of assemblies. These data reinforce the view [20] that PK sensitivity does not inversely mirror the size of PrPSc assemblies, at least for certain prion strains.
Here we observed a strict quantitative correlation between the fast prion strains aggregates templating activity, as measured by the conversion of ovine PrPC by PMCA, and their infectivity as measured by mouse incubation time bioassay or replicating activity in cell culture. The templating activity of the smallest size PrPSc aggregates particles was 2–3 logs over that of the bulk of higher sized PrPSc aggregates. Whether this is due to their size, -the smaller, the swifter to polymerize [64]-, or to their specific infectivity remains clearly an open, overlooked question [62] we are currently addressing. Given the superior templating activity of the smallest size PrPSc aggregates, further studies are ongoing to examine whether the SV profile of PMCA-generated PrPSc would be enriched in such assemblies and thus would differ from that of the original brain material. This would be consistent with recent observations suggesting a preferential selection of certain PrPSc conformers during PMCA reactions [65].
The longest PrPSc polymers (assuming they are linear) could conceivably [66], [67], [68] generate numerous converting pieces as active as the small size oligomers, provided they can be fragmented by the sonication and the beads used in PMCA [69]. They also exhibit low conformational stability values (Table S1), as assayed by denaturation assay [70], a characteristic believed to increase the rate of polymer fragmentation [71], [72]. As the main aggregate type in the fast strains, they were expected to exhibit the best converting activity. Having actually found the opposite situation raises the intriguing possibility that the most infectious and the most aggregated PrPSc populations identified by SV might not derive from the same polymerization pathway, as observed with recombinant PrP oligomers [73] and other protein oligomers [74], [75] or, alternatively, that an increase in the polymer size led to an irreversible loss of converting activity. It also suggests that the proposed pivotal role of fibril breakage [6], [7], [72] in hastening fibril growth is a specific property of certain macromolecular assemblies, at least for prion.
The low PMCA activity of the largest PrPSc assemblies further add to the discrepant impact of the overall stability and/or length of PrPSc aggregates on its conversion potency [25], [42], [62], [76]. A clear and confounding limitation in such studies is that the properties of the biochemically dominant PrPSc component are taken as the properties of the whole PrPSc species while it is obvious here that the specific infectivity and templating activity of PrPSc assemblies can be heterogeneous. Another layer of intricacy would be provided by the strain to strain variations.
Cumulatively (this study and [20]), the specific infectivity and converting activity (the levels of infectivity and of PMCA activity divided by the PrP content) of the fast prions PrPSc aggregates appears essentially supported by a minor fraction (<10%) of PK-resistant oligomers of ≤5 PrP molecules, - a size consistent with that deduced from prion radiation inactivation studies [77], [78] -, whereas the bulk of PrPres (>90%), constituted essentially of 12–30 molecules of PrP [20], showed over 1000-fold lowered activities. A considerable proportion of PrPres generated during the course of the disease might thus have a negligible contribution to prion replication dynamics. The reported converting activities of small-sized, PK-sensitive particles [31], [62] or small size PrPres aggregates fractionated by other methods [79], [80] appeared comparatively low. Although the latter studies were based on fast hamster strains, we found that their most infectious particles were also associated with small sized particles, as in the fast ovine strains [20]. It is worth mentioning their infectious starting material was composed of artificially aggregated PrPres particles that were sedimented before subsequent disaggregation and fractionation [79], [80]. Such a procedure may have destroyed or permanently altered discrete subpopulations of infectious particles [60], [81].
Together, our findings suggest that prion infectious particle size is strain-encoded and participates in the strain biological phenotype, in particular the incubation period of disease. For the fast strains, our findings support discrete oligomers as the most effective template in the proteopathic cascade leading to animal death. Their strong converting properties could provide a quick regeneration of templates to sustain prion replication. Their small size could also favor dissemination and initiation of conversion at distance. Whether the oligomeric forms identified in our study demonstrate a more acute neurotoxicity than the larger size aggregates remains to be determined and is currently assessed using prion permissive primary cultures of neurons [82]. As the most potent inducers of the pathogenesis, these oligomers could be in fine the most neurotoxic, incidentally concurring with the view that the oligomers generated during neurodegenerative diseases linked to protein misfolding and aggregation are generally more potent than larger multimers in impairing neuronal metabolism and viability (for reviews [2], [83], [84]).
|
10.1371/journal.pbio.3000047 | Gasdermin D mediates the pathogenesis of neonatal-onset multisystem inflammatory disease in mice | Mutated NLRP3 assembles a hyperactive inflammasome, which causes excessive secretion of interleukin (IL)-1β and IL-18 and, ultimately, a spectrum of autoinflammatory disorders known as cryopyrinopathies of which neonatal-onset multisystem inflammatory disease (NOMID) is the most severe phenotype. NOMID mice phenocopy several features of the human disease as they develop severe systemic inflammation driven by IL-1β and IL-18 overproduction associated with damage to multiple organs, including spleen, skin, liver, and skeleton. Secretion of IL-1β and IL-18 requires gasdermin D (GSDMD), which—upon activation by the inflammasomes—translocates to the plasma membrane where it forms pores through which these cytokines are released. However, excessive pore formation resulting from sustained activation of GSDMD compromises membrane integrity and ultimately causes a pro-inflammatory form of cell death, termed pyroptosis. In this study, we first established a strong correlation between NLRP3 inflammasome activation and GSDMD processing and pyroptosis in vitro. Next, we used NOMID mice to determine the extent to which GSDMD-driven pyroptosis influences the pathogenesis of this disorder. Remarkably, all NOMID-associated inflammatory symptoms are prevented upon ablation of GSDMD. Thus, GSDMD-dependent actions are required for the pathogenesis of NOMID in mice.
| The NLRP3 inflammasome plays an important role in the maturation of interleukin (IL)-1β and IL-18. Accordingly, NLRP3 gain-of-function mutations, which cause a spectrum of autoinflammatory disorders known as cryopyrin-associated periodic syndromes (CAPS), are associated with excessive IL-1β and IL-18 production. Although CAPS-associated inflammatory symptoms are treated with IL-1-blocking agents, emerging evidence indicates that some CAPS patients only partially respond to these drugs. Persistent inflammatory responses have also been reported in CAPS mice deficient in IL-1β and IL-18 signaling and may be the consequences of the pro-inflammatory cell death, pyroptosis, which is induced by gasdermin D (GSDMD), the other effector of the inflammasomes. Consistent with this view, we found that damage to multiple organs that manifested in a mouse model of CAPS was prevented by ablation of GSDMD.
| NLRP3, also called cryopyrin, assembles an inflammasome complex upon sensing danger signals triggered by structurally different exogenous and endogenous molecular entities [1–3]. Failure to clear the insults or restore homeostasis leads to chronic activation of this inflammasome, a response that underlies various inflammatory and metabolic diseases, including gout, diabetes, and atherosclerosis [4]. Activating mutations in the NLRP3 gene also cause constitutive activation of the NLRP3 inflammasome in patients with a spectrum of autoinflammatory disorders known as cryopyrinopathies or cryopyrin-associated periodic syndromes (CAPS), which include neonatal-onset multisystem inflammatory disease (NOMID), Muckle-Wells syndrome (MWS), and familial cold autoinflammatory syndrome (FCAS) [5, 6]. CAPS are monogenic disorders with some degree of genotype-phenotype correlation, with NOMID exhibiting the most severe manifestations [5, 6]. Each of the CAPS phenotypes displays multiple symptoms, including systemic inflammation, recurrent or chronic fever, and urticaria-like rash [5, 6].
Consistent with the NLRP3 inflammasome role in interleukin (IL)-1β and IL-18 maturation, cryopyrinopathies are associated with excessive production of these cytokines. Accordingly, IL-1-blocking drugs are widely used in the management of these disorders. However, it appears that some CAPS patients only partially respond to IL-1 biologics [7–9]. In addition, skeletal lesions, the hallmark of NOMID, are refractory to IL-1 blockade [10–13]. These clinical observations underscore the complexity of cryopyrinopathies by suggesting that other actions of the inflammasomes beyond maturation of cytokines also contribute to the pathogenesis of these disorders. Indeed, the NLRP3 inflammasome also processes gasdermin D (GSDMD) into GSDMD-N (N-terminal domain) and GSDMD-C (C-terminal domain) [14–16]. GSDMD-N translocates to the plasma membrane, where it binds phospholipids and forms pores at the plasma membrane through which IL-1β and IL-18 are secreted by living cells [17–19]. Sustained activity of the inflammasomes causes excessive maturation of GSDMD and pore formation; this leads to membrane perforation and, ultimately, pyroptosis [17, 20–23]. This form of cell death provokes the uncontrolled release of not only IL-1β and IL-18 but also cytoplasmic contents, resulting in the recruitment of immune cells and propagation of inflammation [17, 24]. Thus, pyroptosis is not a silent endpoint, but the extent to which this pathologic process influences the pathogenesis of cryopyrinopathies is unknown.
Knockin mice harboring specific mutations found in CAPS patients were engineered in an attempt to generate preclinical disease-relevant models for genotype-phenotype relationship studies [25–28]. These models recapitulate some clinical features though disease manifestations are, in general, more severe in mice than in humans. Nonetheless, these seminal studies revealed that pyroptosis may be responsible for the persistent inflammatory responses in mice with impaired IL-1β and IL-18 signaling [8, 29]. Here, we used NOMID mice to determine the role that GSDMD and pyroptosis play in this disease model. NOMID mice exhibited systemic inflammation, stunt growth, and damage to multiple organs. These anomalies were absent in NOMID mice lacking GSDMD, which were indistinguishable from wild-type (WT) littermates. These results reveal a nonredundant function of GSDMD in the onset and progression of NOMID in mice.
The NLRP3 inflammasome complex—which comprises NLRP3 itself, the adapter protein, apoptosis-associated speck-like protein containing a CARD (ASC), and caspase-1—processes pro-IL-1β and pro-IL-18 into IL-1β and IL-18, respectively [1]. This inflammasome also cleaves GSDMD into GSDMD-N and GSDMD-C [14–16]. GSDMD-N forms pores at the plasma membranes through which IL-1β and IL-18 are secreted by living cells; excessive pore formation causes pyroptosis, a response that can be assessed in vitro by quantifying the release of lactate dehydrogenase (LDH) [18, 19]. Consistent with the literature, GSDMD was cleaved upon stimulation of WT mouse bone marrow–derived macrophages (BMMs) with lipopolysaccharide (LPS) and nigericin (Fig 1A). Two cleaved GSDMD fragments were detected; whether the larger fragment was further processed to generate the smaller fragment is not known. GSDMD maturation correlated with the release of not only IL-1β (Fig 1B and S1 Data) but also LDH (Fig 1C and S1 Data), indicating that BMMs undergo NLRP3 inflammasome-dependent pyroptosis under these experimental conditions. To reinforce this conclusion, GSDMD processing, cytokine production, and pyroptosis were determined using cells isolated from mice lacking GSDMD or components of either the NLRP3 canonical inflammasome (e.g., NLRP3 or caspase-1) or noncanonical inflammasome (e.g., caspase-11). Maturation of IL-1β and GSDMD was impaired in BMMs lacking any component of the classical NLRP3 inflammasome but was unaffected in Casp11 null cells (Fig 1A–1C), as expected. These results strengthen the view that GSDMD is a key effector of the NLRP3 inflammasome pathway.
The identification of more than 100 NLRP3 sequence variants underscores the challenges of genotype–phenotype relationship studies for CAPS [30]. In efforts to fill this gap, several preclinical CAPS-relevant models were developed [25–28]. They included knockin mice, which harbored a D301N NLRP3 mutation, the mouse ortholog of the human D303N mutation found in NOMID patients [25]. Mating of Nlrp3fl(D301N)/+ mice with lysozyme M-Cre−/+ (LysM-Cre−/+) mice yielded control and Nlrp3fl(D301N)/+;LysM-Cre−/+ mice, in which the autosomal dominant mutation in Nlrp3 was induced in myeloid cells; these mice are referred to as NOMID mice. We previously reported that the phenotype of NOMID mice with myeloid-restricted activation of NLRP3, which included systemic inflammation and skeletal anomalies, resembled that of mice broadly expressing the mutated protein [25, 31, 32]. This mouse model provided the opportunity to determine the impact of GSDMD deficiency in the pathogenesis of NOMID. Consistently, GSDMD cleavage in WT BMMs required priming and secondary signals triggered by LPS and nigericin, respectively (Fig 1D). By contrast, GSDMD proteolysis in NOMID BMMs was induced by LPS alone though the response was maximal in the presence of the ionophore. Likewise, secretion of IL-1β and LDH by WT cells necessitated the combined actions of LPS and nigericin, whereas these responses were significantly induced by the endotoxin alone in NOMID cells (Fig 1E and 1F; and S1 Data). Notably, secretion of IL-1β and LDH was abolished in cells lacking GSDMD. Thus, mature IL-1β is constitutively produced in NOMID cells, but its release requires GSDMD.
NOMID mice are runted, and they usually die by 2 to 3 weeks of age [25, 31], whereas Gsdmd null mice are apparently normal [16]. Consistent with these reports, NOMID pups were indistinguishable from WT and Gsdmd null littermates at birth but exhibited growth retardation and significantly lower body weight by 12 days of age (Fig 2A and 2B; S1A Fig and S1 Data). Additional macroscopic aberrations in NOMID mice included the presence of skin lesions (Fig 2A) and splenomegaly (Fig 2C and 2D; S1 Data). Skin and spleen abnormalities and the small body size phenotype of NOMID were all normalized in mutant mice lacking GSDMD (Fig 2A–2D). Growth delay, systemic inflammation, perinatal lethality, and spleen and skin abnormalities have been reported for other models of CAPS [26, 27, 29]. Deletion of Il-1 receptor completely abolished these outcomes in NOMID mice but not in FCAS and MWS mice [8, 29], findings that are consistent with the view that, in contrast to humans, FCAS and MWS are unexpectedly more severe than NOMID in mice. The release of not only IL-1β and IL-18 but also other pro-inflammatory factors during pyroptosis may be responsible for the persistent residual inflammatory responses in FCAS and MWS models. Thus, it will be informative to determine the effects of GSDMD deficiency on disease progression in other preclinical models of CAPS.
While we were wrapping up this work, a report indicated that lack of GSDMD in mice prevented the onset and progression of Familial Mediterranean Fever, a disease in which aberrant pyrin inflammasome activities caused IL-1β oversecretion and pyroptosis [33]. Deficiency in GSDMD also protected mice against endotoxic shock, consistent with activation of this protein by intracellular LPS [14, 16]. A recent paper suggested an interplay between caspase-8 and caspase-11-GSDMD axis in the execution of endotoxic shock [34]. Collectively, these findings indicate that inactivation of GSDMD arrests pathogenic signals induced by various inflammasomes.
IL-1β propagates inflammation through various mechanisms, including perturbation of chemokine and cytokine signaling networks, responses that lead to the expansion and recruitment of neutrophils to several organs. This cytokine also promotes anemia owing to its negative effects on erythroid progenitors and erythropoietin signaling as well as alteration of the expression of ferritin and ferroportin [35–38]. Accordingly, NOMID mice produced higher levels of IL-1β in bone marrow compared to WT counterparts (Fig 3A and S1 Data), a response that correlated with excessive GSDMD processing in vivo in bone marrow, though the cleaved fragment was barely detected in this compartment (S1B Fig). This observation was not unexpected considering that excessive generation of GSDMD-N caused cytolysis; as a result, the cleaved fragment may have been lost during the sampling process. NOMID mice also exhibited peripheral leukocytosis (Fig 3B and S1 Data) driven by neutrophilia (Fig 3C and S1 Data) and anemia (Fig 3D; S1C Fig and S1 Data), as we previously reported [25, 32]. The identity of myeloid cell subpopulations, which are prone to pyroptosis and propagate inflammation, in this model is unknown, a knowledge gap that future studies should address. In any case, while Gsdmd ablation had no effect on the number of blood cells compared to WT mice, it abrogated or attenuated the onset of leukocytosis and anemia in NOMID mice (Fig 3A–3D). Accordingly, the bone marrow compartment of NOMID mice contained abnormally high levels of Gr1+/CD11b+ cells and low levels of Ter119+ cells (Fig 3E and 3F; S1 Data), responses that were normalized upon Gsdmd deletion.
Histological analyses showed massive neutrophilic infiltration in the liver, dermal and hypodermal layers of the skin, and the spleen of NOMID mice compared to WT or Gsdmd−/− counterparts (Fig 4). Inflammation in the spleen was characterized by disorganized structures of white and red pulps. Because skeletal complications—including low bone mass—are hallmarks of NOMID, we investigated these outcomes in NOMID mice. Histological examinations of skeletal elements showed disorganized columns of chondrocytes with profoundly altered morphology. The epiphysis was hypocellular (Fig 4), a phenotype that was previously reported to be caused by massive chondrocyte death [25, 32] and reminiscent of the human disease [39]. The number of osteoclasts, cells responsible for bone resorption, was markedly increased in NOMID mice relative to control mice. Remarkably, all organs that were analyzed in NOMID;Gsdmd−/− mice were all spared from inflammation-induced damage (Fig 4). Thus, deletion of GSDMD abolishes inflammatory responses and organ demise in NOMID mice.
Blockade of IL-1 activity has been the main strategy for neutralizing pathogenic signals induced by this cytokine in CAPS and other autoinflammatory disorders. However, these drugs have shortcomings, including high cost and the requirement for parenteral administration. Thus, there is still a medical need for the development of safe and affordable drugs for the treatment of autoinflammatory diseases. Breakthrough research demonstrating that GSDMD-mediated pyroptosis releases cytoplasmic contents, including IL-1β and IL-18, offers a novel node for therapeutic intervention. Stemming from its mechanisms of action, blockade of GSDMD and subsequently pyroptosis should, in theory, provide superior efficacy compared with targeted blockade of IL-1β. The compelling evidence indicating that inactivation of GSDMD blocks inflammatory responses induced by the NLRP3 inflammasome lends support to discovery efforts aimed at identifying selective inhibitors of GSDMD actions.
All procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of Washington University School of Medicine in St. Louis, Missouri. All experiments were performed in accordance with the relevant guidelines and regulations described in the IACUC-approved protocol number 20160245.
Gsdmd−/− mice [16] were kindly provided by Dr. V. M. Dixit (Genentech, South San Francisco, CA). Nlrp3−/−, Casp1−/−, and Nlrp3fl(D301N)/+ mice and lysozyme M (LysM)-Cre mice have previously been described [25, 31, 40, 41]. Casp11−/− mice were purchased from The Jackson Laboratory. All mice were on the C57BL6J background, and mouse genotyping was performed by PCR.
Complete blood counts were performed by the Washington University School of Medicine DCM Diagnostic Laboratory as previously described [31]. Bone marrow cells were flushed out as previously described, and photographed [31].
BMMs were obtained by culturing mouse bone marrow cells in culture media containing a 1:25 dilution of supernatant from the fibroblastic cell line CMG 14–12 as a source of M-CSF, a mitogenic factor for BMMs, for approximately 5 days in a 10-cm dish following the procedures that we published [42]. Nonadherent cells were removed by vigorous washes with PBS, and adherent BMMs were detached with trypsin-EDTA and were cultured in culture media containing a 1:50 dilution of CMG at 4 × 104/well in a 96-wells plate (for the analysis of IL-1β and LDH) or 1.2 × 106/well in a 6-wells plate (for Western blot analysis).
BMMs were treated with 100 ng/mL LPS for 3 hours, then with 15 μM nigericin for 30 minutes. Extracts from BMMs or bone marrow cells were prepared by lysing cells or cell pellets, respectively, with RIPA buffer (50 mM Tris, 150 mM NaCl, 1 mM EDTA, 0.5% NaDOAc, 0.1% SDS, and 1.0% NP-40) plus phosphatase inhibitors and Complete Protease Inhibitor Cocktail (Roche, Brighton, MA). Protein concentrations were determined by the Bio-Rad method, and equal amounts of proteins were subjected to SDS-PAGE gels (12%). Proteins were transferred onto nitrocellulose membranes and incubated with GSDMD antibody (1:1,000, ab209845, Abcam, Cambridge, MA) or β-actin (1:5,000, sc-47778, Santa Cruz Biotechnology, Dallas, Texas) overnight at 4°C, followed by a 1-hour incubation with secondary goat anti-mouse IgG (1:5,000, A21058, Thermo Fisher Scientific, Grand Island, NY) or goat anti-rabbit IgG (1:5,000, A21109, Thermo Fisher Scientific), respectively. The results were visualized using Li-Cor Odyssey Infrared Imaging System (LI-COR Biosciences, Lincoln, Nebraska).
Mouse bone marrow cells were flushed from femur and tibia. For flow cytometry analysis of the leukocytes, red blood cells (RBCs) were depleted with RBC lysis buffer (Roche, Brighton, MA). Cells (0.5–1 × 106) were incubated with Fc block (anti-mouse CD16/32, BioLegend, San Diego, CA) to block nonspecific Fc binding, stained with isotype control or APC-anti-mouse Ter119 (BioLegend, San Diego, CA), FITC-anti-mouse CD11b (eBioscience, Grand Island, NY), and PE-anti-mouse Ly-6G/Ly-6C (Gr1) antibody (BioLegend, San Diego, CA) according to the supplier’s instructions. Flow cytometry was performed using BD LSRFortessa or BD FACSCanto II Flow Cytometer system, followed by analysis with FlowJo software (Tree Star, Ashland, Oregon).
All tissues were harvested and fixed in 10% formalin. Long bones were decalcified in 14% (w/v) EDTA for 5 days at room temperature. All tissues were embedded in paraffin, sectioned at 5 μm thickness, and mounted on glass slides. Sections were stained with hematoxylin–eosin (HE) or TRAP as previously described [42].
BMMs were treated with 100 ng/mL LPS for 3 hours, then with 15 μM nigericin for 30 minutes.
Cell death was assessed by the release of LDH using LDH Cytotoxicity Detection Kit (TaKaRa, Mountain View, CA). IL-1β levels in conditioned media were measured by ELISA (eBiosciences, Grand Island, NY). For IL-1β measurements in bone marrow, flushed bone marrow was centrifuged, and the supernatants were collected as described previously [42]. IL-1β levels were quantified using the eBioscience ELISA kit.
Statistical analysis was performed using one-way ANOVA with Tukey's multiple comparisons test or two-way ANOVA with Tukey's multiple comparisons test in GraphPad Prism 7.
|
10.1371/journal.ppat.1006282 | HPV16 and 18 genome amplification show different E4-dependence, with 16E4 enhancing E1 nuclear accumulation and replicative efficiency via its cell cycle arrest and kinase activation functions | To clarify E1^E4’s role during high-risk HPV infection, the E4 proteins of HPV16 and 18 were compared side by side using an isogenic keratinocyte differentiation model. While no effect on cell proliferation or viral genome copy number was observed during the early phase of either virus life cycle, time-course experiments showed that viral genome amplification and L1 expression were differently affected upon differentiation, with HPV16 showing a much clearer E4 dependency. Although E4 loss never completely abolished genome amplification, its more obvious contribution in HPV16 focused our efforts on 16E4. As previously suggested, in the context of the virus life cycle, 16E4s G2-arrest capability was found to contribute to both genome amplification success and L1 accumulation. Loss of 16E4 also lead to a reduced maintenance of ERK, JNK and p38MAPK activity throughout the genome amplifying cell layers, with 16E4 (but not 18E4) co-localizing precisely with activated cytoplasmic JNK in both wild type raft tissue, and HPV16-induced patient biopsy tissue. When 16E1 was co-expressed with E4, as occurs during genome amplification in vivo, the E1 replication helicase accumulated preferentially in the nucleus, and in transient replication assays, E4 stimulated viral genome amplification. Interestingly, a 16E1 mutant deficient in its regulatory phosphorylation sites no longer accumulated in the nucleus following E4 co-expression. E4-mediated stabilisation of 16E2 was also apparent, with E2 levels declining in organotypic raft culture when 16E4 was absent. These results suggest that 16E4-mediated enhancement of genome amplification involves its cell cycle inhibition and cellular kinase activation functions, with E4 modifying the activity and function of viral replication proteins including E1. These activities of 16E4, and the different kinase patterns seen here with HPV18, 31 and 45, may reflect natural differences in the biology and tropisms of these viruses, as well as differences in E4 function.
| In HPV induced lesions, the most abundant protein expressed in the productive stage of viral life cycle is E1^E4 (E4), with its expression being coincident with viral genome amplification. To clarify the role of E4 in the high-risk HPV life cycle, we carried out a comparative analysis of E4 function in HPV16 and 18 using an isogenic keratinocyte cell-line background. Our results show that E1^E4 contributes to virus genome replication efficiency and life cycle completion rather than being essential. These effects were seen more dramatically with HPV16. The difference between HPV16 and HPV18 in our system suggests important tropism differences between these viruses. HPV16 E4’s contribution to the virus life cycle is mediated by several activities, including its G2 arrest function, as well as its role in activating members of the MAPK pathway, including ERK, p38, and most notably pJNK. These 16 E4 functions facilitated the nuclear localization of the E1 virus helicase and enhanced E1/E2 dependent viral genome amplification as well as stabilising E2. We suspect that the massive accumulation of E4 in the upper epithelial layers may however underlie a more critical role for E4 post-genome amplification.
| Human papillomaviruses (HPVs) are small, non-enveloped DNA viruses that infect cutaneous and mucosal stratified epithelium to induce a wide variety of epithelial lesions ranging from benign papillomas to invasive carcinomas [1]. So far, more than one hundred and fifty HPV types have been completely sequenced [2], with anogenital types being divided into two groups according to their contribution to cancer development. The low-risk types include HPV 6 and 11 that are associated primarily with benign genital lesions. The high-risk types, such as HPV16, 18, 31, 33, and 45 are found in around 99.7% of cervical cancers, with HPV types 16 and 18 being responsible for more than 75% of cervical cancers worldwide [3]. Although both HPV16 and 18 are contained within the Alpha Papillomavirus Genus, they are members of different species (Alpha7 and 9), with different biologies and disease-associations. For all papillomaviruses, the virus life cycle is linked to the differentiation of the infected epithelial cell as it migrates from the basal layer to the epithelial surface. Viral genomes are maintained as nuclear episomes that replicate along with cellular DNA at low copy number in the basal layer. Following cell division, HPV genomes partition into the two daughter cells, with one of these entering the suprabasal layers and committing to terminal differentiation. Although this facilitates expression of L1 and L2 and the assembly of infectious virions, completion of the virus life cycle depends on epithelial site of infection, with life-cycle deregulation leading to neoplasia, and in some instances, the development of cancer [1, 4, 5].
During the HPV life cycle, levels of the viral E1^E4 protein rise as genome amplification begins, allowing accumulation of E4 protein in the upper layer of the epithelium in cells supporting virus synthesis. The E1^E4 protein is translated from a spliced mRNA, with the first 5 amino acids being encoded by the E1 open reading frame [6–8]. The E1^E4 protein associates with and reorganizes the cellular cytokeratin network both in vivo and in vitro via its N-terminal leucine-rich motif (LLXLL), which may eventually contribute to efficient virus release [9–12]. Recent findings suggest that this interaction may also affect viral genome amplification, with the disruption of the HPV16 E1^E4 keratin binding motif leading to a defect in amplification success during the virus life cycle [13]. Interestingly, over-expression of the E4 protein of HPV1, 16 and 18 in monolayer cell culture induces a robust G2/M cell cycle arrest [14–17], which may create an environment that facilitates efficient E6/E7 driven viral genome amplification. In HPV16, an association of E1^E4 with Cdk1/cyclinB complexes via its PTTP motif is responsible for this effect [15]. Our recent work suggests a further complexity of E1^E4 function, with the protein being modified first by MAPK and then by Cdk1/cyclinB [14, 18] as the infected cell moves from S-phase into G2 prior to growth arrest. These changes modify E4 structure and keratin association, and eventually facilitate E1^E4 cleavage by the protease calpain. This final post-translational modification removes key sequences at the proteins N-terminus, allowing it to multimerise and to be deposited as highly abundant amyloid fibrils [19]. This ordered pattern of expression and protein modification suggests a primary role for E4 in the productive stages of the virus life cycle, including genome amplification, virus assembly, virus transmission and release. Recent data from several groups have suggested a role for E1^E4 in genome amplification and life cycle completion. However, there appear to be significant differences between its roles in different HPV types and when function is examined in different cell backgrounds. Whereas in HPV18 and 31, loss of full length E1^E4 protein led to impaired genome amplification when analysed in human primary foreskin keratinocytes [20, 21], in HPV 11, the truncated E1^E4 has been reported not to compromise life cycle completion, which was examined in N-Tert cells, a human foreskin keratinocyte line immortalized by hTERT (the catalytic subunit of human telomerase) [22]. In HPV16, our previous work showed different effects depending on the length of the E4 protein following expression in NIKS, a spontaneously immortalised human keratinocyte cell line [13]. Taken together, the precise contribution of E1^E4, and the mechanism by which it acts to optimize genome amplification have yet to be firmly established. These varying results do however suggest that E4 may not have precisely the same function in all papillomaviruses, and that this may differ even between members of the high-risk HPV group, mirroring the differences in the functions of other high-risk HPV gene products such as E6 and E7.
To define E4’s role in the virus life cycle more methodically, we have examined its function in the context of the two most important HPV types (HPV16 and 18) during their full productive life cycle. We have taken care to overcome issues relating to keratinocyte batch-variation and/or protocols by using a common cell background (i.e. the isogenic NIKS keratinocyte cell-line), and an optimized organotypic raft culture approach that robustly support the life cycle of a diverse range of HPV types including HPV16 [23, 24]. NIKS (Near-diploid Immortalised KeratinicyteS) is a pathogen-free, immortal human keratinocyte progenitor, that provides a reliable and reproducible model system for the study of epithelial keratinocyte growth and differentiation [25], and the papillomavirus life cycle [26]. Indeed, NIKS epithelium ‘STRATAGRAFT’, is currently being developed by STRATATECH, and is being tested in Phase 3 trials as a temporary skin replacement for burns victims (http://www.stratatechcorp.com/). Based on these systems, we have carried out a comparative functional analysis using genomic mutants, and have in addition made extensive use of time-course experiments to discriminate between loss of function and functional delay in the organotypic raft system. It appears from this that E1^E4 contributes to virus replication efficiency and life cycle completion, rather than being essential for these events, with HPV16 E1^E4 having a much more dramatic contribution when direct comparisons are made in common genetic background. This may be linked to the greater propensity of HPV16 to drive neoplasia at stratified epithelial sites, and thus the greater role for E1^E4 in arresting cell cycle progression in a G2-like state during the late stages of infection. The additional novel observation, that 16E1^E4 prominently sequesters active JNK kinase, and other members of the MAPK family in the cytoplasm during the productive phases of the virus life cycle, provides an additional role of E4 in enhancing the accumulation of the viral E1 helicase in the nucleus. It is clear from this, that 16 E4 contributes to genome amplification success through several molecular mechanisms, and that the extent of E4’s contribution differs even amongst the high-risk HPV group.
The collection of clinical material used in this study complied with the Helsinki Declaration of 1975, as revised in 1983 as described previously [29]. All tissue sections were obtained as part of a study with Jagiellonian University, Krakow Poland [29], with sample collection being approved by the relevant Ethical Review Board (IRB nr KBET/2/B/2010 dated 07.01.2010). All patients participating in that study were adults and gave written informed consent prior to surgery. All experiments using human tissue are licenced at University of Cambridge by Human Tissue Authority (Licensing number 12196).
C33a (ATCC), SiHa (ATCC, a cervical carcinoma cell line that constitutively expresses HPV16 E6 and E7), 293T (ATCC) were maintained in Dulbecco’s Modified Eagle’s Medium (DMEM, SIGMA) supplemented with 10% fetal calf serum (FCS, HyClone) and 1% penicillin and streptomycin. NIKS (a gift from Paul Lambert, McArdle Laboratory for Cancer Research, University of Wisconsin), a HPV-negative spontaneously immortalised human keratinocyte cell line, was maintained at sub-confluence on γ-Irradiated J2 3T3 feeder cells (a gift from Paul Lambert) in F medium with all supplements as previously described [27, 28]. All cells were incubated at 37°C in a 5% CO2 environment.
To construct the HPV18 E4 genomic mutant, a subgenomic HPV18 fragment was excised from plasmid pBSSK-HPV18 (pBSSK-HPV18 was a gift from Craig Meyers, the Pennsylvania State University College of Medicine [30]) by digestion with ApaI, which cleaves at position 663 on pBSSK, and at nucleotide 4535 in the HPV18 genome. This fragment, which contains the E4 ORF, was subcloned into pBC (Stratagene, The Netherlands) and used as a template for PCR-mediated, site-directed mutagenesis (QuickChange XL Site-Direct Mutagenesis system, Stratagene, The Netherlands). Two nucleotide substitutions (T3467 to A and T3470 to A) were inserted into the E4 ORF using the forward primer 5’-TAT CCG CTA CTC AGC TAG TAA AAC AGC TAC AGC ACA CCC-3’ and reverse primer 5’GGG TGT GCT GTA GCT GTT TTA CTA GCT GAG TAG CGG ATA-3’. These changes introduced a translational stop codon into E4, with only a silent change in the overlapping E2 coding sequence. After mutagenesis, the subgenomic fragment was cloned back into pBS-HPV18, and the entire HPV18 genome was sequenced to confirm that the genome contained only the desired mutations. The construction of E4 mutant HPV16 genome (16st15) was described previously [13]. To generate the G2 arrest mutant of HPV16E1^E4, the same W12 strain of the HPV genome (originally cloned into the pSp64 vector, a kind gift from Prof. Margaret Stanley, Department of Virology, University of Cambridge) was cloned into the BamHI site of the modified plasmid pTZh19U (pTZhW12 HPV16 [31].Mutagenesis was performed by PCR-based site-directed mutagenesis as described above (QuickChange XL Site-Directed Mutagenesis system, Stratagene, The Netherlands) using primers AGGCAGCACTTGGCCAATCATTCCGCCGCG and CGCGGCGGAATGATTGGCCAAGTGCTGCCT in order to generate a mutant genome with changes in the 16E1^E4 region necessary for G2-arrest motif. Both wild type (WT) and mutant (T22I, T23I) genomes were sequenced to ensure that no additional base changes were present. The pDrive-GAPDH plasmid used for copy-number evaluation was prepared by cloning a 361 bp fragment of human GAPDH (Genbank accession NM-002046) into pDrive (Qiagen Ltd, UK) after amplification by RT-PCR from total RNA extracted from NIKS (forward primer GCCTCCCGCTTCGCTCTC and reverse primer GCCAGCATCGCCCCACTTG). RNA isolation and RT-PCR were described previously [18]. To construct the E1/E2/E4 expression vector (pIRES-E1E2E4), the region of HPV16 containing the E1, E2 and E4 open-reading frames (ORFs) was amplified from the HPV16 WT and E4KO genome by PCR amplification using primers ACG TGA ATT CTA ATC TAC CAT GGC TGA TCC TGC AGG TAC CAA T and ACG TTC TAG ACG CGG ATC CTC ATA TAG ACA TAA ATC CAG TAG A. The amplified DNA fragments were cloned into the pIRESeGFP (pIRES) vector (Clontech) utilising the EcoRI and BamHI restriction enzyme sites to create plasmids pIRES-E1E2E4 and pIRES-E1E2. The HPV16 origin-containing reporter plasmid (p16Ori-CMV-Gluc) was prepared from pCMV-Gluc2 (New England Biolabs, NEB) by cloning a PCR-amplified HPV16 LCR fragment (nt positions 7267–103) in reverse orientation as shown diagrammatically in S6 Fig (primer sequences available upon request). Ligation was carried out using an In-Fusion Cloning Kit (Clontech) according to the manufacturer’s instructions.
Plasmids containing HPV16 and HPV18 genomes (wild type or mutant) were digested with BamHI (HPV16) or EcoRI (HPV18) to release the whole viral genome. The linearized HPV16 or HPV18 genomes were then re-circularized and purified as described previously, before being co-transfected with a plasmid encoding blastocidin into NIKS cells [32]. Twenty-four hours later, NIKS cells were selected with 6 μg/ml blasticidin for 5–7 days, and after screening by PCR and Southern blotting, isogenic WT and E4KO pairs of HPV16 or HPV18 populations with similar viral genomic copy number/cell were used for comparative analysis in organotypic raft cultures as described by Isaacson et al [23, 24, 32].
Prior to differentiation, NIKS cells were prepared from monolayer cultures following trypsinization, and suspended in 1.5% methylcellulose as previously described [33]. Approximately 4 x 106 keratinocytes were plated into a total of 6 ml methylcellulose medium in 6 wells of a 24-well plate and incubated at 37°C in a CO2-containing humidified incubator for 24–72 h. Cells were washed 5 times by centrifugation in ice-cold PBS and re-suspended in 100 μl Qiagen ATL buffer before DNA preparation using the QIAamp DNA Micro Kit (Qiagen, UK).
Real-time PCR was used for the detection and quantification of HPV DNA and human GAPDH using the primers and TaqMan probes shown in Table 1. Reactions were prepared in a volume of 25μl containing either 1×TaqMan Universal PCR Mastermix (Applied Biosystems Ltd, UK), 2.5mM of each primer and 6.25mM TaqMan probe, or containing 1× Absolute SYBR Green QPCR Rox Mix (Applied Biosystems Ltd, UK) with 70nM of each primer. PCR was performed using an ABIprism 7500 system with 15 min denaturation at 95°C followed by 40 cycles of 95°C for 15s and 60°C for 60s. For each real-time PCR assay, a standard curve experiment was performed to allow absolute quantification of DNA. For HPV genomic DNA, pTZhW12HPV16 or pBSHPV18 were used as templates, and for human GAPDH, pDriveGAPDH was used as template. Serial dilutions of plasmid DNA were prepared to allow DNA copy number to be plotted against cycle threshold value. All real-time PCR reactions were run in triplicate alongside no-template controls. To quantitate HPV DNA levels in HPV genome-containing cell populations and cell lines, HPV copy number per cell was expressed relative to GAPDH copy number. The human genome blast using designed primers and probe for human GAPDH shown in Table 1 indicated that there are 4 copies, comprising 2 copies of GAPDH and 2 pseudogene copies of GAPDH per NIKS cell. This was confirmed by determination of GAPDH copy number against known NIKS cell number.
Slides of formalin fixed, paraffin embedded raft tissue were baked and de-paraffinised by washing three times in xylene followed by two washes in 100% ethanol. The slides were then rehydrated in 80%, 50% and 30% ethanol for two minutes each, and then finally in PBS for 5 minutes. Sections were then heated in 0.01M citric acid (pH 6.0) in microwave for 2 minutes, prior to protein digestion with Digest All 3 pepsin (Zymed) at 37°C for 10 seconds. HPV16 or 18 DNA was labelled with digoxigenin by use of the DIG DNA labeling Kit (Roche Applied Science Gmbh, Germany) according to the manufacturer’s instructions. DIG-Labeled HPV DNA was added to slides, denatured for 5 min at 76°C, and incubated overnight at 42°C. After washing, the DIG-positive HPV probe was detected with an anti-digoxigenin-POD antibody at 1:400 (Roche Applied Science Gmbh, Germany) and the signal amplified using a Tyramide Signal Amplification Kit (Perkin-Elmer Ltd, UK) according to the manufacturer's instructions. E1^E4 protein and nuclei were identified with anti-16E1^E4 antibody TVG 405 conjugated to Alexa 488 and DAPI (4', 6-diamidino-2-phenylindole).
Immunofluorescence and immunohistochemistry were performed as described previously [10]. For activated MAPK detection, the formalin fixed, paraffin embedded tissue sections were incubated in solution D pH 9.0 (Dako, Glostrup, Denmark) for 10 min at room temperature prior to autoclaving for 2 min at 121°C. The antibodies used were anti-phospho-ERK1/2 (Thr202/Tyr204) rabbit mAb (Cell Signaling Technology), anti-phospho-p38MAPK (Thr180/Tyr182) rabbit mAb (Cell Signaling Technology), anti-phospho JNK rabbit polyclonal antibody (abcam, UK), anti-MCM2 rabbit polyclonal antibody (ab31159, abcam, UK), anti-HA (16B12; Covance, Princeton, NJ) and anti-16E1^E4 antibody TVG 405 [34] conjugated to Alexa 488 and an anti-E2 rabbit polyclonal antisera ([35]; a gift from Dr Yuezhen Xue, Institute of Medical Biology, Singapore). To measure the intensity and distribution of the stained protein by ‘cross-sectional imaging’, digital images of immuno-fluorescently stained raft tissues were captured with a DeltaVision microscope and imaging system (GE Healthcare Life Sciences) fitted with a 10x objective or using a Panoramic Slide Scanner, 3D Histotech, UK. The captured images were then analysed using ImageJ in order to establish the intensity of staining through the epithelial layers, from the basal layer to the cornified layer. For each immunofluorescent stain, these ‘top to bottom’ scans were collected across 10 areas of equal thickness that were derived from at least 3 different raft sections. Each of the graphs shown (e.g. in Fig 1D and 1E) represents an average of at least 100 individual cross-sectional profiles therefore, and is in general more representative of the raft phenotype than the individual IF image. Where the tissue architecture of the raft was disrupted during the tissue sectioning (e.g. image shown in Fig 1C (NIKS/HPV18 WT day 14), gaps in the continuity of the scan were compensated for during the analysis. In most cases, such disruption was minimised by positioning blocks in the microtome so that the sectioning blade traversed the block from the epithelial surface towards the epithelial basal layer. Because E4 can disrupt cell-cell contacts, such disruption to raft structure was more obvious with the WT rafts and also at late time points.
Proteins were extracted from tissue sections as previously described [36] and quantified using the BCA protein assay kit (Pierce, UK), before being separated on 4–12% gradient polyacrylamide-SDS-Tris-Tricine denaturing gel (Invitrogen, UK) and transferred onto PVDF membranes (Bio-Rad). After transfer, membranes were blocked for 1 hour at room temperature in 1% milk in PBS-T (PBS, 0.1% tween20). Blots were then incubated overnight at 4°C with the appropriate primary antibody diluted in 1% milk PBS-T. Primary antibodies used were anti-E2 rabbit polyclonal antisera [35], anti-tubulin clone B512 (Sigma, UK), anti-HPV16L1 (CAMVIR-1, Santa Cruz, USA), anti-GFP clone B-2 (Santa, Cruz), anti-16E1^E4 antibody TVG 405 or anti-GAPDH clone 374 (Chemicon, UK), followed by the appropriate HRP-conjugated secondary antibody (GE Healthcare, UK), and detection using ECL, or ECL plus kits (GE Healthcare, UK) or by the appropriate IRDye 800CW fluorescent secondary antibody (Licor, UK) followed by detection using an Odissey imaging system (Licor, UK).
The production and infection of recombinant lentiviruses and retroviruses were accomplished as previously described [37]. Lentivirus or retrovirus vectors CSII-TRE-Tight-HA16E1, in which the N-terminal hemagglutinin (HA)-tagged codon-optimized HPV16 E1 was inserted under a tetracycline-responsive promoter, and pQCXIZeo-tetON ADV was described previously (the virus vectors were a gift from Tohru Kiyono, National Cancer Center Research Institute, Tokyo, Japan [38, 39]). To introduce mutations into the HPV16E1 MAPK phosphorylation sites, PCR-based mutagenesis was used with the KOD-plus mutagenesis kit (Toyobo, Japan), with all mutants being completely sequenced to ensure that no additional changes had been introduced. To generate SiHa cells expressing HA-HPV16E1 in a doxycycline-inducible manner, SiHa cells were seeded 1 day before and inoculated with pQCXIZeo-tetON ADV at MOI of 3 followed by zeocin selection (2 μg/ml). The generated SiHa-tetON cells were then infected with CSII-TRE-Tight-HA16E1 (SiHa-tetON/CSII-TRE-tight-HA16E1). The expression of E1 was induced in the presence of 1 μg/ml of Doxycycline.
Recombinant adenoviruses (rAd) expressing 16E1^E4 (rAd16E1^E4) or β-galactosidase (rAdβ-Gal) have been described previously [15, 40]. For infection experiments, 1x106 SiHa cells were seeded into 90mm tissue culture dishes, or 1x104 SiHa-tetON/CSII-TRE-tight-HA16E1 cells were seeded into Falcon 4 well culture slides (BD Biosciences, USA). 24h after seeding, either rAd16E1^E4 or rAdβ-Gal was added to the media at a multiplicity of infection (MOI) of 100. Cells were subsequently harvested at 24 or 48 h post-infection unless otherwise stated. Transfection was performed using the Effectene Transfection Kit (Qiagen, UK) according to the manufacturer’s protocol. The HPV16 E5-expressing SiHa cell line was established as described previously [18].
Cells were seeded at a density of 4x105 cell per 60mm dish, and were transfected with pIRES-E1E2 or pIRES-E1E2E4, and an HPV16 origin-containing plasmid (p16Ori), a kind gift from Peter Howley (University of Harvard, USA) [41], using the Effectene transfection reagent (Qiagen) according to the manufacturer’s instructions. The cells were harvested as for Western blotting (above), and low molecular weight DNA was extracted as previously described [42] and purified using Wizard spin columns (Promega). To distinguish replicated DNA from input DNA, the extracted plasmid samples were compared before and after DpnI digestion by Southern blotting as described by Del Vecchio et al [41]. To linearize p16Ori, the sample DNA was digested with XmnI. The digested samples were separated by 1% agarose gel electrophoresis and then analysed by Southern blotting essentially as described previously [41]. To measure the relative replication activity, the amount of amplified DNA was normalized to the signal of the input p16Ori DNA for each experiment. Hybridisation signals were quantitated using a Phosphor Imager (Storm) and Image Quant 5.0 software.
For the Luciferase DNA replication assays, NIKS cells were seeded at a density of 2x105 cells per well into 6-well plates without feeder cells, before being transfected using FuGENE HD (Fisher Scientific, UK) with 2ug of pIRES-E1/E2 or pIRES-E1/E2/E4, along with 15ng of the HPV16 origin-containing reporter plasmid (p16Ori-CMV-Gluc) and 15ng of pCMV-Cluc2. 12 hours post-transfection, the media was replaced with F medium with all supplements and feeder cells were added. 48 hours post-transfection, the media was replaced. 24 hours later, culture media was collected and luciferase activity measured using a BioLux Gaussia Luciferase Assay Kit and a BioLux Cypridina Luciferase Assay Kit (NEB, UK). Gaussia Luciferase activity was normalized against Cypridina Luciferase activity.
Total genomic DNAs were isolated using QIAamp DNA blood mini kit. Total genomic DNAs isolated from each population were digested with either BamHI or HindIII for at least 8h at 37°C and electrophoresed on a 0.8% agarose gel. Total genomic DNA isolated from untransfected NIKS cells was used as a negative control. Following electrophoresis, DNAs were transferred to a nylon membrane (Hybond-N; Amersham, Piscatway, NJ) and hybridized to [α-32P]dCTP-radiolabeled DNA probe for full-length HPV16 prepared with Ready-to-Go DNA labelling system (GE Healthcare Life science). Radioactive bands were detected using a Storm imaging system (Amersham).
Total RNAs from cells (C33a, SiHa) transfected with pIRES-16E1/E2/E4, pIRES-16E1/E2 or pIRES (empty plasmid) were extracted using RNAeasy mini Kit (Qiagen). RNA samples were electrophoresed on a 1.2% formaldehyde agarose gel. Following electrophoresis, RNAs were transferred to a nylon membrane (Hybond-N; Amersham, Piscatway, NJ) and hybridized to [α-32P]dCTP-radiolabeled DNA probe for E1E2E4 DNA fragment of HPV16 prepared with Ready-to-Go DNA labelling system (GE Healthcare Life science). Radioactive bands were detected using a Storm imaging system (Amersham). APOT assay using RNA sample of NIKS/HPV16 was carried out following the protocol described previously [43].
Total RNAs from NIKS containing HPV genome cell were extracted using RNAeasy mini Kit (Qiagen). Total RNA (1 μg) was reverse transcribed with 100 U of SuperScript III Reverse Transcriptase (ThermoFisher scientific) using 100 μM random hexamer primers according to the manufacturer’s instructions. Reactions were prepared in a volume of 20μl containing either 1×T ABsolute qPCR SYBR Green Mixes (ThermoFisher scientific) with 70nM of each primer. Each primer set were designed to detect the viral gene transcripts specifically (the primer sequences are available upon request). PCR was performed using a ViiA 7 system (ThermoFisher scientific) with 15 min denaturation at 95°C followed by 40 cycles of 95°C for 15s, and 60°C for 60s. For each real-time PCR assay, a standard curve experiment was performed to allow absolute quantification of cDNA.
The role of E1^E4 in the virus life cycle has been examined in different HPV types by different groups using different cell backgrounds and protocols [13, 20–22], prompting us to carry out a comparative analysis in the same cell background. To do this, a ‘generic’ keratinocyte cell line that supports the life cycle of a range of different HR HPV types was used [23, 24], in conjunction with a common protocol for the transfection and maintenance of genome-containing cells. As reported previously, this approach allows the generation of cell lines harboring HPV genomes as viral episomes, with no obvious signs of viral genome integration, either by Southern blotting or APOT analysis (see [23] and S1 Fig). Blotting was carried out after blastocidin selection and the expansion of HPV-containing cell populations, or after the isolation and expansion of HPV-containing cell clones. In all cases, supercoiled, relaxed (open) circular, and linear HPV genomes were detected, with a prominent 8Kb band being apparent after restriction enzyme digestion using a single-cut restriction enzyme. Results from the analysis of three clonal populations (T1, T2 and T3) are shown in S1A Fig as an example. The E6/E7 mRNA expression pattern visualized in these cell lines by APOT assay indicates that these genes are expressed from episomal rather than integrated viral genomes and is shown in S1B Fig, (tracks T1, T2 and T3). The aberrant pattern seen in a cell line harboring integrated HPV genomes is shown in the lane marked ‘IN’, with the range of episomal copy numbers seen in individual clones being shown for HPV16 and 18 in S1C and S1D Fig. Although individual clones, (which represent expansion from single cells in the cell population) contained a range of genomic copy numbers, viral episomes were always apparent. When making comparisons, we were careful to adjust our transfection to achieve the same average copy number in the cell population, or to use copy-number matched clones. The HPV16 E4KO mutant used in this study contained a substitution of T3384 to A, in order to introduce a stop codon at amino acid position 15 in the HPV16 E4 protein sequence. In the context of the HPV16 genome, this is the shortest E4 truncation that is silent in the E2 reading frame. Although close to the E1^E4 slice acceptor site at nucleotide 3357, the T3384 to A base substitution did not affect the relative abundance of viral transcripts spanning E2 or E1, or the use of the E1^E4 splice site (S2A and S2B Fig). Cell lines harbouring this mutant genome as a viral episome were successfully established, as described previously by Nakahara and colleagues [13]. In HPV18, the E4KO mutant was made by double point mutations of T3467 to A and T3470 to A, which introduces two stop codons at position 17 and 18 of the E4 coding sequence without changing the coding capacity of the E2 ORF. Comparable with the results from the HPV16 analysis, no changes in patterns of transcription were seen. NIKS populations and clonal cell lines containing the wild type (WT) and E4KO genomes of both HPV types were subsequently prepared. Interestingly, the average genome copy in transfected NIKS population was often higher for HPV18 (approximately 200 copies/cell) than HPV16 (approximately 100 copies/cell), and while individual clones derived from each transfected NIKS population exhibited a range of episomal copy numbers (see S1C and S1D Fig), no significant differences were seen between NIKS cells harbouring WT or E4KO genomes, allowing isogenic WT or E4KO cell line pairs with similar copy numbers to be selected for comparative analysis. To generate cell populations with matched copy number, we found that a lower level of HPV18 genomic DNA was generally required at the transfection stage.
Previous studies have suggested that loss of full length HPV18 E1^E4 can confer a growth advantage on undifferentiated primary human foreskin keratinocytes [21]. To look at this further, HPV16 or 18, copy-number matched WT or E4KO-containing NIKS were plated, and grown in monolayer culture, from day 1 to day 10 (Fig 1A). NIKS were harvested after removal of feeder cells, and counted using a Z1 Coulter Particle Counter (Beckman Coulter, California, United States). Monolayer tissue culture approximates the epithelial basal layer, especially at confluence when cells become contact inhibited but have not begun to differentiate [23]. Such cells lack detectable E4 by immunostaining. As shown in Fig 1, the WT and E4KO genomes of both HPV16 and 18 confer a growth advantage on NIKS keratinocytes when compared to parental cells. However, no significant and/or reproducible difference was apparent between WT and E4KO genomes of either HPV16 or 18. When taken together, these results indicate that the E1^E4 protein does not obviously contribute to early events in the HPV life cycle, such as the initial amplification following infection, the maintenance replication of the viral genome or the regulation of the viral proteins that drive cell proliferation in the basal layer.
We next considered whether this lack of an obvious E4 function in basal cells extended to the basal and parabasal layers when cells were propagated in organotypic raft culture. Raft culture provides a spatial separation of cells at various stages of differentiation, and closely mimics the differentiation and organization of normal epithelial tissues. Rafts were harvested at 8, 10, 12 and 14 days after placing the cell monolayer at the air/liquid interface in order to examine the consequences of E4 loss as the extent of differentiation increases. As expected, E4 was not detectable in the lower epithelial layers by immunofluorescence staining at any of the time points, but was apparent in the upper epithelial layers in the WT rafts (Fig 1B and 1C, right panels), becoming more extensive as the extent of differentiation increased from day 8 to 14. The extent of cell cycle entry in basal and parabasal layers was examined by staining for the cellular MCM protein. MCM can be regarded as a surrogate marker of E7 expression when present above the basal layer in HPV-driven rafts [23], with expression declining as the extent of differentiation increased from day 8 to 14. Because of this, raft tissues of similar thickness, and which exhibited a similar degree of differentiation were chosen for comparison. As shown in Fig 1B and 1C, no obvious differences in MCM staining were apparent between WT and E4KO rafts at any of the time points examined, with broadly similar results for HPV16 and 18. To rule out the possibility that E4 may affect either the density of MCM-positive cells in the lower epithelial layers, or the intensity of staining in each cell, HPV 18 WT and E4KO rafts were digitally scanned from basal layer to the top and analyzed by cross-sectional imaging analysis (Fig 1D and 1E). Again, no obvious difference was apparent in the MCM expression profile in the lower epithelial layers, supporting the conclusion that the early stages of the virus life cycle are not dependent on HPVs ability to express E4. A similar observation was previously reported using an in vivo model of E4 function and neoplastic progression in the CRPV/domestic rabbit system [44].
During the HPV life cycle, viral genome amplification begins during host epithelial differentiation, and is coincident with E1^E4 expression [6, 31]. Indeed, the role of E1^E4 in viral genome amplification has been examined using a number of HPV genomic mutants in different cell backgrounds using a range of differentiation protocols [13, 20–22]. Despite differences in the extent of productive infection in these systems, an almost complete failure to support genome amplification was reported for HPV18 and 31 [20, 21], with no effect on genome amplification reported for HPV11 [22]. Here we have focused specifically on the two most important high-risk HPV types (HPV16 and 18) using a common isogenic cell background, with differentiation being carried out over an 8–14 day time course. We conclude from this more extensive analysis, that E4 is not in fact essential for viral genome amplification, as in situ hybridization signals were apparent in the middle and upper layers of both WT and E4KO mutants in both the HPV16 and 18 rafts (Fig 2A and 2C). While viral genome amplification can thus be triggered without E4 proteins, its detection by in situ hybridization was however much reduced in HPV16 E4KO raft tissues at all time-points (Fig 2A), and was much more apparent when the extent of differentiation in the raft was limited (i.e. day 8 and 10). The loss of E4 had no apparent effect on differentiation however, as revealed by hemotoxylin and eosin staining (H&E), and the detection of early and late differentiation markers (keratin 10 and filaggrin) (S3 Fig). In fact the raft time course experiments described here may explain the heterogeneity of results reported previously, with E4 effects becoming less apparent as the extent of epithelial differentiation increases over time. To examine this more precisely, the 10 and 14 day raft time-points were digitally scanned (Panoramic Slide Scanner, 3D Histotech, UK), and the number and intensity of the HPV in situ-positive cells averaged across 10 raft sections of comparable thickness. Using this approach, which overcomes some of the limitations of showing just selected images, the extent of genome amplification was found to be around three fold higher in the NIKS/HPV16 rafts when compared to the NIKS/HPV16 E4KO rafts at both the 10 and 14 day time points. Comparable raft cultures prepared with the HPV18 E4KO NIKS showed a much less dramatic reduction (Fig 2C). At day 10 and 12, viral genome amplification was lower with the E4KO mutant than with the WT genome, but by day 14, viral genome amplification was similarly extensive in both the WT and E4KO genomes (Fig 2D). Although such in situ methodologies are particularly useful in understanding events in heterogeneous tissues such as rafts (Fig 2B and 2D), we went on to also examined raft sections by qPCR and western blotting following laser capture microscopy (LCM) and/or epithelial microdissection (Fig 2E and 2F). Similar to what was seen in the in situ hybridisation studies (Fig 2A–2D), the effect of E4KO on genome amplification (Fig 2E (and L1 accumulation (Fig 2F)) was again seen much more clearly with HPV16 than HPV18. When taken together, these results indicate that for both HPV16 and 18, genome amplification is restricted rather than abolished in the E4KO cells upon differentiation, with apparently similar levels of genome amplification being seen irrespective of E4 presence in the mature HPV18 rafts.
The final stages of the HPV productive cycle involve expression of the viral capsid proteins (L1 and L2), to allow genome packaging. To address whether the absence of E1^E4 directly compromises the expression of capsid proteins per se, we compared the extent to which L1 accumulation is dependent on E4 in HPV16 and HPV18 rafts during our time course. Indeed, this was initially suggested from the analysis of total protein extracts prepared from SDS-solubilised, microdissected day 14 raft epithelium shown in Fig 2F, where the HPV16 E4KO was found to support L1 accumulation only poorly. When L1 distribution was examined, the extent of L1 expression was found to be lower than WT in the HPV18 E4KO rafts at day 10 and 12, equivalent by day 14, when differentiation was more extensive (Fig 3A). This pattern is very similar to that seen for viral genome amplification, suggesting that genome amplification and capsid synthesis are coordinated events. The similarities seen between the day 14 rafts suggest that 18 E1^E4 does not play an essential role in either the regulation of genome amplification or the expression of virion structural proteins (Fig 2C, Fig 3A). By contrast, no detectable L1 was found in raft tissues containing HPV16 E4KO mutants by immunofluorescence (Fig 3B), only barely detectable levels of L1 transcripts and L1 protein detectable by Western blotting (Fig 2G). During the course of our experiments, three different HPV16 E4KO populations were examined, as well as one clonal cell line harboring the HPV16 E4KO genome, but on no occasion did we detect any expression of 16L1 by immunofluorescence staining. In our experience, HPV16 NIKS are always less proficient in supporting the full productive life cycle than HPV18 NIKS, with the absence of 16L1 expression possibly reflecting the lower base levels of genome amplification seen with HPV16 (Fig 2A). No obvious differences were noticed however when patterns of differentiation-dependent transcription were compared [13]. We suspect in fact that the different contributions of E4 to the life cycle of HPV16 and HPV18 may reflect differences in the ability of each HPV type to complete their life cycle in the common ‘generic’ keratinocyte background used here, and maybe also to differences in the in vivo disease-associations of these two HPV types.
Since loss of E4 affected HPV16 genome amplification and L1 expression more dramatically than HPV18, we focused our subsequent analysis on HPV16, and in particular, the role of E4 in this process. A well-established function of 16E4 when expressed at high-level, is its ability to cause cell cycle arrest in a G2-like phase, where viral DNA replication may proceed in the absence of cellular replication [14, 15, 45]. It has been hypothesised that this may facilitate the timely onset of viral genome amplification during epithelial differentiation by antagonizing the proliferative functions of E6 and E7 [45]. Given the clear effect of E4 loss on HPV16 genome amplification, we were prompted to further explore this hypothesis. Our previous studies have shown that HPV16 E1^E4-mediated G2 arrest depends on a PTTP motif at amino acid position 21–24. A mutant 16E1^E4 in which these two threonines were mutated to alanines (i.e. T22A, T23A (PTTP motif changed to PAAP)) no longer inhibited mitotic entry [15]. Since the E4 ORF lies entirely within the E2 gene, the preparation of a G2 arrest-mutant in the context of the full length HPV16 genome could only be made by mutating the E4 PTTP motif to PIIP (E1^E4 T22I, T23I), a modification that introduces only silent changes into the E2 coding sequence. To confirm that the 16E4 PIIP mutation (i.e. T22I, T23I in 16E1^E4) abolishes the G2-arrest function of E4, WT E4 and the 16E1^E4 PIIP mutant were first expressed in Cos-7 cells as described previously [15]. 24 hours after transfection, cells were treated with nocodazole to disrupt mitotic spindle formation, before being left for an additional 24 hours to allow cell-cycle progression. Any cells passing from G2 into M, that are not arrested by E4, will subsequently be arrested in mitosis as a result of the nocodazole treatment. These cells were subsequently visualized by double staining using an anti-16E1^E4 antibody conjugated to Alexa-fluor 488 in conjunction with an antibody to the mitotic marker, phospho-histone H3 and visualization by microscopy. As shown in Fig 4A, the 16E1^E4 PIIP (T22I, T23I) mutant protein no longer inhibited cell cycle progression in G2, with the number of cells passing from G2 to M being broadly similar to what is seen with the GFP control. WT E4 by contrast, reduced the number of cells progressing from G2 to M, which has been shown previously to result from its G2-arrest function. We next prepared the E4 PIIP mutant in the context of the HPV16 genome, before introducing both the WT and mutant genomes into NIKS cells. Both WT and the E4 PIIP mutant genomes were maintained at similar copy number prior to differentiation. To establish how loss of E4’s G2-arrest function affected HPV16 genome amplification, we used the organotypic raft culture system described above, in parallel with a methylcellulose-based differentiation system. Although differentiation in methylcellulose does not perfectly mimic epithelial differentiation in vivo, the approach does facilitate quantitation of viral genome copy number. In our experiments, NIKS cells harbouring either the WT or the E4 PIIP genomic mutant were suspended in 1.5% methylcellulose [46], and harvested at 0h, 24h, 48h and 72h. Viral DNA copy number was measured by qPCR, and normalized to cell number using primers directed to the cellular GAPDH gene (see materials and methods). The qPCR results showed that HPV16 DNA copy numbers per cell were slightly reduced in the E4 PIIP genomic mutant, as compared to the WT HPV16 genome at all time-points (Fig 4B), with the trend suggesting statistical significance (P<0.05). To examine the role of the G2 arrest motif more thoroughly, the NIKS cell populations harbouring either the WT or E4 PIIP HPV16 genomes were grown in organotypic raft culture. By in situ hybridization and immunofluorescence, less genome amplification and L1 expression were apparent with the 16E4 PIIP mutant genome tissues (Fig 4C and 4D). The finding that L1 levels were reduced but not absent contrasts sharply with the complete loss of capsid protein expression seen with the 16E4 KO genome (see Figs 3B and 4D). As seen for HPV16 E4KO, the E4PIIP mutant genome did not show any obvious differences in the relative abundance of transcripts spanning the early region (S2 Fig), with the reduction in late transcripts being expected because of the lower levels of genome amplification. When taken together, the organotypic raft results appear in broad agreement with the data obtained from the methylcellulose experiments, and indicate that the G2 arrest function of 16E4 contributes importantly to successful life cycle completion. Interestingly, previous work with HPV18 has suggested that the 18E4 G2 arrest function does not contribute in a similar way to the life cycle of HPV18 [47]. We suspect that this may be linked to differences in patterns of E6/E7 expression between the HPV16 and HPV18 rafts, with HPV18 producing a robust and well-ordered productive life cycle in organotypic raft systems. Our previous work has shown that HPV16 produces a range of phenotypes using this model, including those that resemble the CIN 1 and 2 seen in patients [23]. In this setting, it is perhaps not surprising that the anti-proliferative effects of E4 may contribute positively to genome amplification success.
MAPK (mitogen-activated protein kinase) is an important regulator of viral protein function and viral genome amplification [48], which occurs in the upper layers of HPV-infected epithelial tissue as cells are driven through S-phase and into G2. In addition, it has also been suggested that 16E4 may stimulate p38 MAPK as part of a cellular stress response [49], prompting us to examine whether this additional E4 function may also play a role in HPV genome amplification. To examine this, we first examined HPV16 WT and E4KO day 14 rafts by immunostaining and ImageJ analysis (Fig 5A and 5B). Although p-p38MAPK could be detected in the basal and parabasal layers in both WT and E4KO rafts, the WT rafts showed a marked elevated p-p38 MAPK in the middle and upper layers where E4 accumulation was seen (Fig 5B). Previous studies using undifferentiated monolayer cell culture systems have shown that both E4 and E5 can activate p38 MAPK [49, 50], which is compatible with our observations (S4 Fig). Interestingly, loss of 16E4’s keratin-binding motif (Δ16N), as occurs in the upper epithelial layers where p38 MAPK activation was observed (Fig 5A), did not significantly compromise E4s ability to activate p38 MAPK in this system (S4 Fig), although as expected, N-terminal deletion led to E4 accumulation because of its increased ability to assemble into amyloid-like fibrils [19, 51]. In addition to p38 MAPK [49], HPV16 E4 can stimulate other members of the MAPK family in monolayer cell culture [18], with ERK1/2 activity triggering the initial association of 16E4 with the cellular cytokeratin network [10]. Interestingly, p-ERK1/2 had a similar distribution in both the WT and KO HPV16 rafts, and was confined to the mid epithelial layers close to the point where E4 levels start to accumulate (Fig 5C and 5D). A difference in the intensity of p-ERK1/2 staining was however apparent, with a generally lower signal in the HPV16 E4KO rafts (Fig 5D), with the profiles revealing a different distribution of p-ERK1/2 and p-p38 MAPK in the mid and upper epithelial layers (Fig 5B and 5D). For HPV18, this prominent difference was less apparent when p-p38 levels were compared (S5A and S5B Fig). No apparent difference was seen in either the abundance or the distribution of p-ERK1/2 between HPV18 WT and E4KO rafts (S5C and S5D Fig).
Several studies have implicated elevated MAPK activity and transition into the G2 phase of the cell cycle as being important for HPV genome amplification, and it is notable from the above, that the activity of one member of the MAPK family (i.e. p-ERK1/2), rises transiently in both the HPV16 and 18 rafts in the mid epithelial layers where genome amplification occurs. The HPV E1 protein, which is a DNA helicase/ATPase required for viral genome amplification, is a target for both p-ERK1/2 and p-JNK MAPK as well as CDK1 (a kinase activated in G2), with phosphorylation of E1 affecting its nuclear/cytoplasmic localization [48]. The reduced viral genome amplification seen in 16E4 KO mutants, and the apparent dependency on E4 for HPV16-mediated MAPK elevation, prompted us to extend our analysis of E4s effects to include p-JNK. To our surprise, p-JNK showed by far the most dramatic change in the presence of 16E4, with p-JNK and E4 co-localizing precisely in the cytoplasm of cells in the mid epithelial layers and above (Fig 6A). Cytoplasmic p-JNK was however seen sporadically prior to the onset of E4 expression in the middle layers of the epithelium, but was only ever sustained from the mid to upper epithelial layers when 16E4 was expressed (compare HPV16 E4KO and WT panels in Fig 6A). This sporadic expression of cytoplasmic p-JNK in the basal and parabasal layers prior to its accumulation in E4-positive cells was a direct consequence of HPV presence in the cell, and was never seen in the NIKS parental rafts where the p-JNK patterns were almost exclusively nuclear (Fig 6A, bottom panels). Because the p-JNK/E4 co-localisation pattern was so striking, we next went on to establish whether a similar pattern could be seen in vivo in low-grade cervical neoplasia caused by HPV16. A noticeable similarity was apparent when p-JNK and E4 staining in CIN1 was compared to the WT HPV16 rafts, and similarly when uninfected cervix was compared to uninfected NIKS rafts (Fig 6B, left panel). In both instances, 16E4 and p-JNK localize closely as E4 accumulation first begins in the mid epithelial layers, with staining eventually declining as the cells reach the epithelial surface. Although the Cytoplasmic p-JNK staining was more prominent, in both the 16 WT rafts and the HPV16 CIN1 lesion, p-JNK was also apparent in the nucleus of the E4-positive cells (see Fig 6A). Cytoplasmic activated JNK was never observed in normal uninfected cervix (Fig 6B, right panel), which showed a distribution similar to that seen in the NIKS-only rafts. Interestingly, the dramatic E4/p-JNK co-localization seen with HPV16, was not seen in organotypic rafts generated using HPV18 WT or E4KO genomes (Fig 6C), and in this context, it is notable that the other MAPK members were also not notably affected by 18E4 (S5 Fig). Although examined extensively in repeat experiments, the prominent effects seen with HPV16, did not extend to either HPV45 or HPV31 (Fig 6D and 6E), although HPV18, 31 and 45 all showed some evidence of nuclear p-JNK, which tended to decline as E4 levels increased (Fig 6C, 6D and 6E). These results suggest that cytoplasmic pJNK-sequestration during the papillomavirus life cycle may be a HPV16-specific function, rather than a general characteristic of high-risk HPV types. As discussed above, the pJNK pattern seen in HPV16 NIKS rafts was broadly comparable to what was seen in HPV16-associated CIN despite the mucosal epithelial origin of the latter (Fig 6A and 6B).
Although use of the organotypic raft culture system, when combined with the analysis of clinically-derived biopsy material remains central to our understanding of HPV biology, the use of undifferentiated cells in cultured monolayer can be used to provide valuable additional mechanistic information. To examine the different abilities of the HPV16 and 18 E4 proteins to modulate p-JNK activity and to consider the consequence of this, the E1^E4 proteins of these viruses were first examined in isolation using undifferentiated epithelial cells grown in monolayer culture. As seen in the organotypic raft system, the HPV16 E1^E4 protein (Fig 7A, 7B and 7C), but not that of HPV18 (Fig 7D) lead to the cytoplasmic accumulation of p-JNK in SiHa cells (Fig 7B and 7C), as well as producing this phenotype in monolayer NIKS cells (Fig 7A). To do this, cells were infected with the recombinant adenovirus expression vectors, Ad.16E1^E4, Ad.18E1^E4 or Ad.βGal, as described previously [15] with E4, β–gal and p-JNK being localized by indirect immunofluorescence as described in the Materials and Methods [49]. These results suggest that 16 E1^E4, but not 18 E1^E4, can sequester active p-JNK in the cytoplasm in absence of other viral proteins, as seen in organotypic rafts. It has previously been shown that 16 E1^E4 associates with the cellular cytokeratin network in both NIKS and SiHa cells, and in SiHa cells promotes network collapse to the nuclear periphery [52]. The characteristic perinuclear E4 distribution is apparent in the enlarged image shown in Fig 7C, and although there is some variation in the p-JNK levels between cells, the co-localisation with 16E4 is reproducibly seen, with some cells showing the characteristic E4/keratin perinuclear bundles (arrowed) (Fig 7C, and [49, 53]).
The various effects seen with 16E1^E4 on members of the MAPK family, coupled with the ability of 16E1^E4 to arrest cells in G2, suggests that 16E4 may be contributing to genome amplification-success by modulating the intracellular environment to favour the phosphorylation of viral and cellular proteins by these kinases. Although such effects on the cell are likely to be complex and to involve multiple targets, we decided to make a first investigation of this by examining whether E4 expression had any effect on the location of the HPV16 E1 helicase, which is required for genome amplification, and which has been reported to shuttle between the nucleus and the cytoplasm depending on the extent of phosphorylation by members of the MAPK and Cyclin-Dependent Kinase (CDK) families. Amino acid sequence alignment revealed that these phosphorylation sites are conserved amongst the E1 proteins of HPV16, HPV31 and HPV11, with the latter containing an additional upstream phosphorylation site (see Fig 8A and [48, 54, 55]). Our initial E1/E1^E4 co-transfection experiments resulted in only a small number of cells co-expressing the two proteins. To overcome this difficulty, we established an inducible cell line in which a HA-tagged 16E1 protein is expressed from a doxycycline inducible promoter using a lentivirus CSII-TRE-tight-HA16E1 construct. A SiHa background was used for these experiments, as SiHa cells constitutively express the HPV16 E6 and E7 proteins, that would be expressed along with E1^E4 and E1 in cells supporting viral genome amplification in vivo. We were unable to generate cell lines constitutively expressing the E1 protein however, presumably because of E1-mediated cell toxicity following overexpression. Lentivirus infection was carried out into Tet-On SiHa cells prepared using retrovirus pQCXIzeo-tetON ADV followed by selection in zeocin as described in the Materials and Methods. In agreement with the hypothesis that E4 expression may modulate the function of other viral gene products, a very clear increase in E1 nuclear localization was apparent when expression was carried out along with E4 in the same cell (Fig 8B), with E4 expression leading to a statistically significant change in intracellular E1 distribution (Fig 8D). Statistical analysis was facilitated by combining the inducible E1 expression system with use of the rAd16E1^E4 expression vector, which ensured E1/E4 double-positivity in the majority of cells expressing E1 in the dish. In contrast to the wt 16E1 protein, the phospholyation site-deficient E1 protein (E1S93A.S107A) was found primarily in the cytoplasm, with the presence or absence of 16E4 having little effect on its intracellular location (Fig 8B and 8C). The % of cells showing nuclear E1 in the presence or absence of E4 was calculated from the analysis of 3000 E1-positive cells is shown in Fig 8D.
Given that the presence of E4 can facilitate E1 nuclear accumulation, we next wanted to establish whether E4 could also enhance E1/E2-mediated viral genome replication in a functional assay. Because E1, E2 and E4 are not expressed at equivalent levels during genome amplification, a HPV16 viral DNA fragment encoding just the E1/E2/E4 region was first isolated from the HPV16 WT or E4KO genome, and cloned downstream of a CMV promoter positioned immediately in front of E1 (S6A Fig). With respect to the E1, E2 and E4 open reading frames, the position of the CMV promoter in these clones (pIRES-16E1/E2/E4 or pIRES-16E1/E2) mimics that of the HPV16 late (p670) promoter in the viral genome, which drives the co-ordinated expression of E1, E2 and E1^E4 during viral genome amplification. In agreement with data shown in S2 Fig, the pIRES plasmids exhibited similar transcription profiles across the E1, E2 and E1^E4 regions to those seen with the full-length viral genome (S6B Fig), with E2 and E4 also being detectable by western blotting (S6C Fig). As expected, E4 was not apparent in cells transfected with the E4KO plasmid, (pIRES-16E1/E2), and in the absence of a HA tag, E1 protein could not be reliably visualized using available antibodies (S6C Fig). In fact E1 transcript levels were tightly regulated in this system, just as they are during the HPV16 life cycle [56–60]. Interestingly, and as reported previously, total E2 levels were always elevated in the presence of E4, in agreement with previous studies showing that E2 can become sequestered and stabilized in the cytoplasm in undifferentiated cells expressing both proteins together [61]. In addition to E1 nuclear localization therefore, it remains a possibility that E2 cellular accumulation may also contribute to E4’s overall effects on genome amplification. The precise mechanism by which E4 affects E2 abundance, and the possible effects of MAPK phosphorylation are not however known, although a direct interaction between E1^E4 and E2 in the cytoplasm has been previously reported [61]. As with most HPV early proteins [62], the detection of E1 has never been shown in raft tissue, and the detection of HPV16 E2 during productive infection has only been convincingly reported with a single purified rabbit polyclonal antiserum [35], with other antibodies to the HPV16 E2 protein failing to detect the protein during productive infection using immunofluorescence or immunohistochemistry protocols. Unfortunately, this purified antibody is no longer available from the researchers who originally produced it (Dr Yuezhen Xue, Institute of Medical Biology, Singapore, personal communication), although we were able to obtain a small aliquot of the unpurified rabbit serum. Curiously, our optimized staining protocol using this antibody did appear to show an elevated in E2 signal in the middle layer of HPV16 WT rafts, when compared with E4KO rafts (S7A and S7B Fig), with nuclear E2 staining apparent at around the time of E4 elevation (S7B Fig). Although background staining was higher than we would have liked using this antibody, the conclusion was supported by digital scanning across the length of the raft tissue (S7C Fig). In contrast to the situation in undifferentiated monolayer cells however, the detectable E2 protein seen in the differentiated raft cultures was predominantly nuclear in our hands, although previous studies have reported both nuclear and cytoplasmic E2 protein using this antibody [63]. In an attempt to substantiate these observations, we next went on to examine E2 by western blotting (S7D, S7E and S7F Fig), using the microdissection approach used previously to assess L1 accumulation. E2 antibody specificity was first assessed against 16E2 expressed from endogenously expressed E2 in 293T cells (S7D Fig), before being examined in organotypic raft extracts. The results were broadly compatible with the immunofluorescence staining, and was reproducibly seen in the replicate rafts prepared here (S7E and S7F Fig). Although we suspect that 16E4 most likely affects E2 function as well as E1, in the absence of more robust E2 detection reagents and a better understanding of E2 MAPK phosphorylation, a contribution to the E1 recruitment to the viral origin of replication may be expected.
Given the above effects of 16E1^E4 on the viral replication proteins, and in particular its effect on E1 nuclear localization, we next examined the contribution of 16E1^E4 on E1/E2-mediated replication using the modified in vitro replication assay. In the first instance, pIRES-16E1/E2/E4 and pIRES-16E1/E2 were transfected into monolayer C33a or SiHa cells along with a HPV16 replication origin-containing plasmid (p16Ori [41]). Despite variation in pOri input levels, E4 (pIRES-16E1/E2/E4) appeared to enhance E1/E2-mediated replication efficiency (Fig 8E and 8F). To confirm these results, and to avoid the inaccuracies that can result from the measurement of band intensities (see Fig 8E), a modified reporter plasmid was prepared (p16Ori-CMV-GLuc (S6 Fig)), to allow replication efficiency to be estimated from the Gluc/Cluc ratio after co-transfection with a pCMV-CLuc reporter plasmid. In this system, the presence of 16 E4 reliably enhanced E1/E2-mediated replication success in repeat (6x) experiments (Fig 8G).
When taken together, the above results suggest a mechanism by which 16E4-mediated MAP kinase modulation facilitates replication success, and also predicts that genome amplification, the activation of MAP kinase, and the accumulation of 16E4 should occur simultaneously during the late stages of the HPV16 life cycle. To establish whether this is indeed the case, a quadruple stain was carried out to localize the sites of viral genome amplification (as determined by fluorescence in situ hybridisation (FISH)) with E4, p-JNK and DAPI. In the HPV16 WT rafts, a remarkable co-localisation of all three markers was apparent, with genome amplification occurring only in cells showing positivity for cytoplasmic p-JNK that were also positive for 16E4 (Fig 9, upper panel, arrowed). A much lower level of genome amplification was seen in the E4KO rafts, in the absence of E4 in cells that were always negative for activated cytoplasmic p-JNK (Fig 9, lower panel, arrowed). As expected, and in line with other reports, it was not possible to detect E1 during the productive life cycle with available antibodies [64]. When taken together with our E4 in vitro expression studies, this data strongly supports the idea that 16E4 acts to modulate the cellular environment during the late stage of infection via the MAPK activation, and that this influence the activity and/or localisation of viral proteins more directly involved in viral genome amplification.
Conflicting data have been reported as to the role of E4 during the virus life cycle, especially in its contribution to viral genome amplification amongst different HPV types [13, 20–22]. To address this issue, and to examine the mechanistic roles of E4 during productive infection, we have carried out a careful comparative analysis of E4’s role in genome amplification between the two most important high-risk HPV types, HPV16 and 18. To do this, we made use of WT and E4KO genomes maintained in an isogenic keratinocyte cell background in order to eliminate variation between different keratinocyte pools and differentiation protocols. Experiments were carried out using matched cell population harboring WT and E4KO genomes of either HPV16 or 18 at similar copy number and at similar passage. Finally, because the HPV life cycle is critically dependent on the differentiation status of the infected host cell, effects on viral genome amplification were examined using raft culture time course experiments rather than at a single raft time point, with digital imaging being used to help determine the significance of changes in patterns of gene expression. Using these approaches, we conclude that while E4 can contribute to virus replication efficiency and life cycle completion, it is not essential for these events. In the isogenic NIKS background, loss of 16E4 was found to have a much more marked effect on life-cycle-success than 18E4, allowing us to identify a role for 16E4’s G2 arrest function as well as its ability to modulate members of the MAPK family, most notably JNK, which we show to be a prominent modification seen during natural HPV16 infection in vivo. The elevation of p-JNK, which was not seen in HPV18, once again highlights differences in the life cycles of high-risk viruses. For HPV16, it appears that E4, and its effects on MAPK affect E1 nuclear localisation, and that in both transient replication assays and during epithelial differentiation, that 16 enhances replication success as a result.
Previous reports have suggested that the loss of full length E4 may enhance the growth rate in undifferentiated HPV18-containing primary foreskin keratinocytes [21], possibly because the over-expression of 18E4 in monolayer culture can cause cell cycle arrest at G2/M. The growth rate of WT or E4KO genome-containing NIKS populations in our system were comparable (Fig 1), which fitted with our observation that E4 is not detectable in undifferentiated monolayer cell culture, either by immunostaining or western blotting, and is unlikely to disrupt cell-cycle progression at G2. By contrast, high-level expression of the E4 protein of several types, including HPV16 and 18 induces cell cycle arrest at G2/M in monolayer-cultured cells [15, 16]. During the HPV life cycle, the expression of the E4 protein is controlled by the HPV late promoter (p670 in HPV16, p811 in HPV18), which is driven by the epithelial differentiation machinery. Although not always apparent, differentiation can occur to some extent in keratinocyte monolayer culture post confluence, which may have contributed to the previous observation [21]. In our hands using NIKS, we find that the presence of HPV16 or 18 episomal genomes is a strong suppressor of differentiation in monolayer culture, with E4 expression not apparent over the time course of our experimental growth assays. Given that we also saw no effect of E4-loss on either epithelial thickness or the extent of cell cycle entry in raft culture, we do not feel that E4 has any obvious role during the early stages of either the HPV16 or 18 life cycles. A similar conclusion was reached when E4 loss was assessed in the CRPV model of infection in vivo [44].
Although the loss of E4 delays viral genome amplification in both HPV16 and HPV18, the effect was much more dramatically seen with HPV16. This allowed us to examine specific E4 functions in the context of the HPV16 life cycle, especially in viral genome amplification. In repeat experiments, our stringent analysis showed that the loss of the G2-arrest motif in HPV16 leads to a noticeable and consistent reduction in genome amplification success, suggesting that G2 arrest function of E4 in HPV16 contributes to viral genome amplification efficiency. In HPV-induced lesions or in HPV containing raft tissues, E4 expression starts before E6/E7 levels decline [29], with the expression of these proteins overlapping in the middle epithelial layers where viral genome amplification occurs [7, 29]. The G2 arrest function of E4 is thought to suppress E7-mediated cell proliferation in the upper epithelial layers and to prolong the G2-like state that is required for efficient viral genome amplification (reviewed in [45]). From our results, this hypothesis remains reasonable for HPV16. There is however some uncertainty regarding the role of E4 mediated G2-arrest in the life cycle of HPV18. Disruption of a motif required for G2/M arrest in the E4 ORF of HPV18 genomes did not affect HPV18 genome amplification [47]. Although we have not tested the effect of the G2 arrest function of E4 on genome amplification in HPV18 in our system, in our raft culture time course experiment, a reduction in genome amplification was seen only at early time points with the HPV18 E4KO mutant, and was not apparent at late time points. The expression pattern of E4 is similar between the HPV16 and 18 WT, which may suggest that the G2 arrest function of HPV18 E4 may sometimes contribute to HPV18 genome amplification efficiency as seen here with HPV16. From our previous analysis of HPV gene expression in model system of disease [23] and in clinical material [24, 29], we suspect that this may be apparent at epithelial sites where there is some deregulation of E6/E7 expression rather than in the raft model, which for HPV18 (as well as HPV31 and 45) appears to be a model of well-ordered productive infection rather than neoplasia. Raft culture does not however mimic the precise characteristics found during in vivo infection, and we are not yet able to model the different epithelial sites where HPV18 naturally causes disease. The different abilities of HPV16 and HPV18 to complete their life cycle in organotypic rafts does however suggest important differences in tropism, and that a common cell type is unlikely to support the life cycles of different HPV types equally. Indeed, this was an important consideration when deciding to carry out our evaluation of two HPV types in an isogenic cellular background. Interestingly, although we were able to match genomic copy number in the WT and E4KO cell populations and lines, the copy number of established HPV-containing NIKS was always different in cell populations prepared using HPV16 and 18. It is clear from this that HPV16 and 18 have different regulation of maintenance replication in the isogenic NIKS background, again reflecting differences between these two viruses.
It appears however that other functions of E4 can contribute to the enhancement of viral genome amplification efficiency. This study suggests that 16E4’s contribution to viral genome amplification efficiency may also involve the activation of MAPK (mitogen-activated protein kinase). The MAPK super-family is composed of ERKs, JNK/SAPKs and p38 MAPKs [65]. In HPV11 and HPV31, the E1 protein, a DNA helicase/ATPase involved in initiating viral DNA replication, shuttles from the cytoplasm into the nucleus after phosphorylation by ERK and/or JNK, and remains in the nucleus because of the presence of activated CDK1/2, which can phosphorylate a CRM-1 dependent NES which inhibits E1 nuclear export, allowing an enhancement of viral DNA replication with appropriate timing during productive infection (reviewed in [55]). The work presented here reveals that E4 loss in the context of the HPV16 genome results in a significant reduction in sustained ERK, JNK and p38 MAPK activity in the upper layers of the epithelium, with these proteins being absent in genome amplifying cells in the HPV16 E4KO raft tissues. In addition to this, activated JNK was found to be prominently associated with HPV16 E1^E4 in the cytoplasm of cells supporting HPV genome amplification, both in the WT HPV16 rafts, but also in low-grade cervical disease caused by HPV16. In fact the patterns of E4 and JNK expression, and their correlation with vegetative viral genome amplification were very similar in the HPV16 WT rafts and in the patient biopsies, which strengthen both the significance of our observations, and our confidence in the NIKS organotypic raft system. Interestingly, our previous work indicated that HPV16 E1^E4 stimulated JNK and CDK activity when expressed in SiHa cells [14, 49] and retained CDK1/cyclin complex in the cytoplasm and induced G2 arrest. As mentioned above, both ERK and JNK activation are important for E1 phosphorylation, with E1 phosphorylation being important for viral DNA replication. Like HPV11 E1, the HPV16 E1 protein sequence contains a consensus MAPK docking domain that is essential for ERK, JNK and p38 MAPK binding [48, 66], and a consensus cyclin binding motif (RxL). A plausible hypothesis therefore is that HPV16 E1^E4 may activate MAPK and CDKs to modulate E1 nuclear localization in order to facilitate viral genome replication. The observations here, that HPV16 E1^E4 expression facilitates HPV16 E1 nuclear localization and that HPV16 E1 lacking conserved putative phosphorylation sites failed to localize in the nucleus, support the idea that 16E1^E4 can contribute to viral genome amplification by facilitating the unclear localization of E1 as a result of phosphorylation by kinases such as MAPKs and CDKs activation. Interestingly, HPV16 E4 is also a target for self-phosphorylation by activated MAPK, which occurs during viral genome amplification [18], and which enhances the association with cytokeratin and E4 protein stability. This is thought to trigger E4 protein accumulation, and prolong the G2 arrest period to facilitate optimal viral genome amplification. Interestingly, a recent report has suggested that the phosphorylation of E5 by activated p38 MAPK in HPV18 was important for viral genome replication (Tom Broker, University of Birmingham, Alabama, USA; personal communication). Our previous studies have shown that HPV16 E4 activates p38 MAPK directly as a result of an E4 triggered cell stress response [49], and the results shown here, as well as those presented previously, suggest E4 may also augment E5 and E2 function to some extent [61]. Interestingly, 16 E4 has been reported to be able to bind and stabilise E2 in the cytoplasm [61], and just as with E1, the E2 protein can also be regulated by the cellular CDK1/2 kinases, which lead to an increase in E2 stability during S-phase [67]. In addition, E1^E4 and E5 are expressed together from an abundant bicistronic viral transcript during the late stages of the virus life cycle, a pattern of expression that fits well with their possible functional association in stimulating and sustaining the activity of MAPK. Clearly E4 has other roles during the late stages of infection, and our data suggest that N-terminally cleaved E4, which assembles into E4 amyloid-like fibers, may activate p38 MAPK in the later stage of viral life cycle, with a potential role in virus transmission and release (data presented here & [19], [49]). The extended p38 MAPK activation apparent in the upper layers of HPV16 WT raft tissues supports this hypothesis. How E4 is associated with ERK activation is not currently answered. Unlike pJNK, the level of phosphorylated pERK was not changed when HPV16 E4 was expressed in SiHa cells by transfection with a 16E1^E4 expression vector or following rAd.16E1^E4 infection (data not shown). Indeed, ERK activation is thought to be mediated primarily by E7 and E5 [18]. Similarly, the co-expression of E2 and E4 did not alter ERK activation (data not shown), although the E4/E2 association reported previously led to E2 stabilization [61], at least in the SiHa model. For JNK MAPK, we suspect that 16E4 role may be to sequester the kinase onto the abundant E4 amyloid structures, and that in the upper epithelial layers, the role of E6, E7 and E5 are critical for its presence. In this case, some similarity is apparent with the prominent E4-mediated sequestration of CyclinB/Cdk1 seen in lesions caused by Mu PV types such as HPV1, and Lambda papillomaviruses such as the Canine or Rabbit Oral Papillomaviruses [14].
A key reason for initiating this study arose from the need to clarify whether E4 loss abolishes genome amplification, or whether E4 acts to modulate amplification-efficiency but is not essential for genome amplification per se. Our data demonstrate that loss of full length E1^E4 expression in HPV18 acts to delay differentiation-dependent viral DNA amplification and L1 expression. These data are slightly different from the previous report, which showed impaired or abolished genome amplification and late gene expression in HPV18 E4KO. These differences may come from the different approaches used, and indeed our raft time course experiments show that both conclusions could be reached depending on the time point analyzed. Indeed, the culture of primary foreskin keratinocytes containing HPV18 WT and HPV18 E4KO genomes using the organotypic raft system, leads to similar results to those shown here as culture times are extended, with both WT and mutant genomes producing similar infectious titres (Craig Meyers, Penn State College of Medicine (personal communication)). Our conclusions were facilitated by carrying out raft culture time course coupled with digital scanning, which can reveal the different degree of epithelial differentiation and the relative changes in protein expression. Poor genome amplification is always more apparent at early stages of differentiation in the E4KO raft tissues, with raft time courses and differences in the time of analysis explaining many of the previous discrepancies. With HPV16, an abolition of genome amplification is apparent at the early time point, whereas with the 18 rafts, it would looks as though E4 has no effect on genome amplification if only the last time point is evaluated. These are the two extremes. By looking at all the time points together we can get the clearest conclusion that E4 acts generally to optimize replication efficiency, but is not essential for viral genome amplification. To some extent, this conclusion may apply to all high risk HPV types, although clearly the contribution of E4 is not uniform, and the sequestration of JNK in particular appears a characteristic of the HPV16 productive cycle. A different result was previously reported during the low risk HPV11 life cycle, which may reflect the very different functions of the low risk E6 and E7 genes, and the different ability of these proteins to drive cell proliferation. Interestingly, the ability of HPV18 to complete its life cycle in the absence of E4 in the NIKS model, will allow us in future to consider the role of E4 in virus assembly, virus maturation, and transmission, where the E4 proteins, including those of HPV16 and 18 almost certainly have their primary function.
When the data is taken together, we conclude that loss of full length E4 in HPV16 and 18 only delays viral genome amplification and L1 expression, but does not abolish these events in the virus life cycle. For HPV16, this delay appears to be mediated by several of E4’s activities including its G2 arrest function, and its role in activating members of the MAPK pathway, such as ERK, JNK and p38, with sustained JNK activation being a HPV16-specific characteristic that affects E1 nuclear localisation, and as a result of this, replication efficiency. Previous studies showing effects of 16E4 on E2 stability and cytoplasmic accumulation [61], and the cross-talk between 16E4 and E5-mediated kinase stabilisation most likely represent additional E4-associated modifications involved in this process. The massive accumulation of E4 in the upper epithelial layers, coupled with its ability to assemble into amyloid-like fibres, must however underlie an additional and perhaps more critical role for E4 post-genome amplification.
|
10.1371/journal.pgen.1007557 | Baldspot/ELOVL6 is a conserved modifier of disease and the ER stress response | Endoplasmic reticulum (ER) stress is an important modifier of human disease. Genetic variation in response genes is linked to inter-individual differences in the ER stress response. However, the mechanisms and pathways by which genetic modifiers are acting on the ER stress response remain unclear. In this study, we characterize the role of the long chain fatty acid elongase Baldspot (ELOVL6) in modifying the ER stress response and disease. We demonstrate that loss of Baldspot rescues degeneration and reduces IRE1 and PERK signaling and cell death in a Drosophila model of retinitis pigmentosa and ER stress (Rh1G69D). Dietary supplementation of stearate bypasses the need for Baldspot activity. Finally, we demonstrate that Baldspot regulates the ER stress response across different tissues and induction methods. Our findings suggest that ELOVL6 is a promising target in the treatment of not only retinitis pigmentosa, but a number of different ER stress-related disorders.
| Differences in genetic background drives disease variability, even among individuals with identical, causative mutations. Identifying and understanding how genetic variation impacts disease expression could improve diagnosis and treatment of patients. Previous work has linked the endoplasmic reticulum (ER) stress response pathway to disease variability. When misfolded proteins accumulate in the ER, the ER stress response returns the cell to its normal state. Chronic ER stress leads to massive amounts of cell death and tissue degeneration. Limiting tissue loss by regulating the ER stress response has been a major focus of therapeutic development. In this study, we characterize a novel regulator of the ER stress response, the long chain fatty acid elongase Baldspot/ELOVL6. In the absence of this enzyme, cells undergoing ER stress display reduced cell death, and degeneration in a Drosophila disease model. Feeding of excess fatty acids increases degeneration to original disease levels, linking the regulatory activity of Baldspot to its enzymatic activity. Finally, we demonstrate that Baldspot can alter the ER stress response under a variety of other ER stress conditions. Our studies demonstrate that Baldspot/ELOVL6 is a ubiquitous regulator of the ER stress response and is a good candidate therapeutic target.
| Phenotypic heterogeneity is common in simple and complex diseases and is the driving force behind the Precision Medicine Initiative [1–4]. Genetic variation among individuals accounts for much of this heterogeneity, but the identity and nature of modifying genes or variants is largely unknown [4, 5]. These modifying variants are often cryptic and may not influence the physiology or visible phenotypes of healthy individuals, but can alter the expression of disease phenotypes. Understanding the role of modifiers and the pathways in which they function will enable the development of patient-specific therapeutic approaches.
One such process is the response to endoplasmic reticulum (ER) stress, which occurs when misfolded proteins accumulate in the lumen of the ER. This activates the Unfolded Protein Response (UPR) in an attempt to return the cell to homeostasis. Failure to do so will eventually result in apoptosis [6]. The UPR is controlled by the activation of three sensors located in the ER membrane: IRE1, PERK, and ATF6 [7]. IRE1 is the most highly conserved of these sensors and contains an endonuclease domain responsible for the non-canonical splicing of Xbp1. The spliced Xbp1 transcript is translated and the protein translocates to the nucleus, where it activates the expression of chaperones and other genes involved in resolving ER stress [8, 9]. IRE1 also cleaves a number of additional mRNA targets, which reduces the quantity of newly translated proteins into the ER. This process is known as Regulated IRE1 Dependent mRNA Decay (RIDD) and is commonly linked with increased cell death [7, 10, 11]. When PERK is activated, it phosphorylates the translation initiation factor eIF2α, which decreases the translation of most mRNA transcripts with canonical translation initiation mechanisms [7]. This reduces the protein-folding load of the ER and allows for the upregulation of select transcripts involved specifically in the UPR, such as ATF4 [7]. ATF6, the third transmembrane activator of the UPR, is transported to the Golgi upon the sensing of misfolded proteins and cleaved, whereupon the cytosolic portion of ATF6 travels to the nucleus to act as a transcription factor, binding ER stress response elements to further upregulate key players in the UPR [7].
Chronic ER stress leads to apoptosis, tissue degeneration, and dysfunction. In rare cases, primary mutations of key ER stress pathway components cause Mendelian syndromes [12–16]. More commonly, ER stress can exacerbate disease, including obesity [17], neurological disease [18], retinal degeneration [7], and some cancers [19]. Experimental genetic or pharmacological manipulation of ER stress levels can alter disease phenotypes, suggesting that inter-individual differences in ER stress responses may influence disease severity. Indeed, human [20], mouse [21], and Drosophila [22, 23] show extensive genetic variation in their response to ER stress. Understanding the role of genetic variation in modulating ER stress pathways may provide more therapeutic targets and improve the accurate identification of high risk patients.
In a previous study, we crossed the Rh1G69D model of retinitis pigmentosa (RP) into the Drosophila Genetic Reference Panel (DGRP) to identify genetic modifiers of this disease [23]. In this model, retinal degeneration is induced by overexpression of misfolded rhodopsin in the developing larval eye disc [24], resulting in chronic ER stress and apoptosis. The DGRP is used to study the genetic architecture underlying complex traits. There are approximately 200 inbred DGRP strains, each strain representing a single, wild-derived genome. Thus, the DGRP captures genetic variation that is present in a natural, wild population. Importantly, we have the whole genome sequence of each strain, allowing phenotype-genotype studies [25]. The degree of degeneration induced by this model varied substantially across the strains of the DGRP. Using an association analysis, we identified a number of promising candidates that modify this RP phenotype and have putative roles in the ER stress response [23].
In this study, we demonstrate that Drosophila Baldspot, one of the candidate RP modifiers identified in our genetic variation screen, is both a disease modifier of ER stress-induced retinal degeneration and a more general modifier of the ER stress response across different physiological contexts. Baldspot, which is orthologous to mammalian ELOVL6, is an ER-associated long chain fatty acid elongase which catalyzes the extension of palmitate to stearate [26, 27]. We show that loss of Baldspot rescues retinal degeneration by reducing ER stress signaling. Loss of Baldspot elongase activity reduces IRE1 and PERK activity and downstream apoptosis, without affecting the misfolded protein load. Strikingly, in the absence of Baldspot activity, dietary supplementation of stearate bypasses the need for Baldspot to increase UPR activation. Finally, we show that Baldspot regulates the ER stress response in a number of different genetic and chemically-induced ER stress models across multiple tissue types, making Baldspot a potentially wide-reaching modifier of ER stress-associated diseases. Understanding how modifier genes alter disease outcomes is the first step to developing personalized therapies.
In previous work, we overexpressed a misfolded rhodopsin protein in the developing eye disc (GMR-GAL4>UAS-Rh1G69D), the larval precursor to the Drosophila retina and eye, in the 200 strains from the DGRP [23]. Using a genome-wide association approach, we identified SNPs that are associated with degree of retinal degeneration. One of these candidate SNPs, Chr3L:16644000 (BDGP R5/dm3), is located in intron four of Baldspot. Baldspot is the only Drosophila orthologue of mammalian ELOVL6, a member of the very long chain fatty acid elongase family of enzymes (Fig 1A) [28]. Strains carrying the minor allele (C, 18880 ± 3010 pixels) exhibit more severe Rh1G69D–induced retinal degeneration as compared to strains carrying the major allele (T, 21822 ± 2294 pixels; P < 0.005) (Fig 1B). This is also apparent when examining the qualitative range of degeneration in strains containing each of these alleles (Fig 1C). Based on the position of this SNP in an intron, we predict that it will affect the timing or levels of Baldspot expression. We also predict that it is exerting its effect in the small subset of cells that are GMR-expressing. This is very difficult to assess, as these cells make up a very small proportion of the eye disc. It is also difficult to know when, during development, this SNP is affecting expression. Nevertheless, we examined the effect of this SNP on Baldspot expression in adult flies and in the larval brain/eye-imaginal disc complex. Expression of Baldspot is unaffected by variation at this SNP in whole adult flies (not containing the Rh1G69D transgene) (S1 Fig). In larval brain-imaginal disc complexes, there is not a significant increase in Baldspot expression from strains carrying the C allele and expressing the Rh1G69D transgene (C, 1.15 ± 0.03 as compared to T, 1.01 ± 0.09;) (S1 Fig). While there is not a detectable substantial difference in expression between the alleles, the possibility remains that altering the expression or activity of Baldspot may impact degeneration and the causative pathways.
To test whether eliminating Baldspot expression can modify Rh1G69D-induced retinal degeneration, we expressed an RNAi construct targeted against Baldspot in the GMR-GAL4>UAS-Rh1G69D background. Expression of this construct results in a strong, significant reduction in Baldspot mRNA (~7% of controls; P = 3.5 x 10−3) (S2 Fig). We found that loss of Baldspot expression results in partial rescue of Rh1G69D-induced retinal degeneration (Rh1G69D/Baldspoti) as compared to controls expressing only the Rh1G69D misfolded protein and no RNAi (Rh1G69D controls) (Fig 2A). Quantification of eye size demonstrates a significantly larger, less degenerate eye in Rh1G69D/Baldspoti flies (13742 ± 913 pixels, n = 20) as compared to Rh1G69D controls (11432 ± 688 pixels, n = 20) (P = 5.3 x 10−11). We found a similar effect using another, independent RNAi strain (S2 Fig). Loss of Baldspot on a wild-type background (26478 ± 1191 pixels, n = 20) resulted in no significant change in eye size or phenotype as compared to genetically matched controls (26582 ± 1110 pixels n = 20) (Fig 2B). We conclude that loss of Baldspot expression provides significant rescue of Rh1G69D-induced retinal degeneration.
Rescue of eye size in the Rh1G69D model is often accompanied by reduction in apoptosis [29, 30]. Based on this and the rescue effect we observed, we hypothesized that apoptosis in the Rh1G69D eye discs would also be reduced upon loss of Baldspot expression. Indeed, we found that apoptosis, as measured by terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) staining, is significantly reduced in Rh1G69D/Baldspoti eye discs as compared to Rh1G69D controls (101 ± 95 TUNEL-positive cells in Rh1G69D/Baldspoti and 226 ± 69 TUNEL-positive cells in controls; P = 1.8 x 10−3) (Fig 2C and 2D).
Excessive ER stress in the Rh1G69D eye discs leads to apoptosis and degeneration and abnormal, degenerate adult eyes. To determine if the decrease in apoptosis we observed was due to changes in the ER stress response, we measured the activity of IRE1 and PERK, two of the transmembrane sensors of misfolded proteins in the ER. Because the ATF6 branch of the UPR is mostly unstudied in Drosophila, we are unable to draw any conclusions about the activity of this branch of the UPR. Under conditions that induce ER stress, Xbp1 transcript is spliced by IRE1. The in-frame Xbp1 transcript is translated and acts as a transcription factor, inducing genes involved in the UPR [8, 9]. The level of spliced Xbp1 transcript is directly proportional to the degree of activation of the ER stress response. To measure IRE1 signaling and Xbp1 splicing in the absence of Baldspot expression, we used a previously characterized transgenic Xbp1-EGFP marker [24, 31–33]. In this transgene, the 5’ end of the Xbp1 mRNA is expressed under the control of the UAS promoter. The 3’ end of the gene, downstream of the IRE1 splice site, has been replaced with sequence encoding EGFP. As with the endogenous Xbp1 transcript, the in-frame, spliced Xbp1-EGFP transcript, and subsequently the Xbp1-EGFP fusion protein, increases as ER stress signaling increases. Thus, EGFP signal increases as ER stress signaling increases. As previously reported, expression of UAS-Xbp1-EGFP in Rh1G69D controls results in increased EGFP, indicating high levels of ER stress (Fig 2E) [24, 31–33]. When UAS-Xbp1-EGFP is expressed in Rh1G69D/Baldspoti larvae eye discs, we see a significant reduction in EGFP signal (0.59 ± 0.13 in Rh1G69D/Baldspoti relative to controls, 1.00 ± 0.18; P = 9.6 x 10−4), indicating reduced ER stress signaling (Fig 2E and 2F). However, Rh1G69D/Baldspoti eye discs display no significant change in rhodopsin protein levels (0.78 ± 0.33 in Rh1G69D/Baldspoti relative to controls, 1.00 ± 0.44 in controls; P = 0.286) (Fig 2E and 2F), suggesting that loss of Baldspot expression reduces IRE1/Xbp1 signaling without affecting the accumulation or degradation of misfolded Rh1 protein.
IRE1-dependent cell death is primarily triggered by the Jun Kinase (JNK) signaling cascade when IRE1 has been chronically and strongly activated [7]. To measure activation of JNK signaling, we monitored expression of puckered (puc), a well characterized JNK target, using a LacZ-tagged allele (puc-LacZ). While puc-LacZ is induced in Rh1G69D controls (1.00 ± 0.41), it is significantly reduced in Rh1G69D/Baldspoti (0.16 ± 0.19 relative to controls; P = 9.5 x 10−3) (Fig 2G and 2H). This reduction in JNK signaling correlates with a reduction in cell death in the absence of Baldspot and could be due to reduced signaling through the IRE1-JNK signaling axis. However, changes in JNK signaling can also be achieved through a number of alternative pathways and influence other downstream processes besides apoptosis [34, 35]. Our current analyses do not distinguish between these possibilities.
We also examined the effect of Baldspot loss on the PERK branch of the UPR. When PERK is activated, it phosphorylates the translation initiation factor eIF2α, which decreases the translation of most mRNA transcripts with canonical translation initiation sequences [7]. We monitored P-eif2α levels by Western blot in brain/eye-imaginal disc complexes isolated from Rh1G69D control and Rh1G69D/Baldspoti larvae. P-eif2α levels were substantially reduced in samples isolated from Rh1G69D/Baldspoti compared with Rh1G69D controls, indicating that PERK activity is reduced in the absence of Baldspot (Fig 2I).
Our data demonstrate that loss of Baldspot expression reduces IRE1 and PERK signaling without altering misfolded protein levels. This is consistent with previous reports demonstrating that the cellular concentration of stearate is associated with increased activation of the ER stress response through direct activation of IRE1 and PERK, independently from misfolded proteins [36–38]. IRE1 and PERK both contain a conserved domain that is embedded in the ER membrane and can detect and respond to changes in membrane lipid composition, leading to increased activation of these sensors [39]. Together, this suggests that the modifying effect of Baldspot is linked to its fatty acid elongation activity.
Baldspot converts palmitate to stearate [26, 27] and the absence of Baldspot function should result in lower levels of stearate. We hypothesized that lower levels of stearate underlies the reduced ER stress response and retinal degeneration we observed, and that this signaling could be restored if stearate were supplemented during development. To test this, we raised Rh1G69D/Baldspoti and Rh1G69D control flies on media with or without 10% stearate and measured eye size in adults. This diet has no effect on eye size in wild-type individuals (Fig 3A). Stearate supplementation resulted in increased degeneration and reduced eye size in Rh1G69D/Baldspoti flies (14033 ± 1553 pixels) as compared to those raised on standard media with no stearate (16846 ± 1699 pixels; P = 1 x 10−7) (Fig 3B). This increase in degeneration in Rh1G69D/Baldspoti flies was accompanied by an increase in IRE1 activity and spliced Xbp1, as evidenced by an increase in Xbp1-EGFP expression in eye discs (1.26 ± 0.08 on stearate relative to standard media, 1.00 ± 0.10; P = 3.4 x 10−3) (Fig 3C and 3D). Stearate supplementation had no effect on eye size, retinal degeneration, or Xbp1-EGFP signal in Rh1G69D control flies (13478 ± 1197 pixels on stearate and 12873 ± 1006 pixels on standard media; P = 0.502; 0.83 ± 0.15 on stearate relative to standard media, 1.00 ± 0.21; P = 0.171) (Fig 3B–3D), likely because of the high level of ER stress signaling already present. High stearate levels are sufficient to induce ER stress signaling and bypasses the need for Baldspot activity.
Because Baldspot is ubiquitously expressed (www.flyatlas2.org/), it is possible that Baldspot’s modifying role in the ER stress response is generalizable across different contexts and tissues. We first tested whether Baldspot modifies ER stress in another tissue. Because wing size is a visible and easily scorable trait, we used the MS1096-GAL4 driver to express Rh1G69D in the developing larval wing disc (wing-Rh1G69D). The Rh1G69D protein acts as a misfolded protein, inducing ER stress only in that tissue. The wing disc gives rise to the adult wing, and perturbations in cellular homeostasis in this tissue result in small, misshapen wings [40]. Expression of wing-Rh1G69D produces a phenotype similar to a vestigial wing (Fig 4A). When the RNAi construct targeted against Baldspot was concurrently expressed in the wing disc (wing-Rh1G69D/Baldspoti), this vestigial phenotype was partially rescued. The adult wings showed an increase in size and an increase in unfolding upon eclosion (Fig 4A). While this phenotype is variable in the degree of degeneration and unfolding of the wing, we do see a significant increase in 2D wing area and wing length in the absence of Baldspot (area: 100769 ± 9903 pixels in wing-Rh1G69D/Baldspoti and 57681 ± 30912 pixels in controls; P = 0.043; length: 633 ± 44 pixels in wing-Rh1G69D/Baldspoti and 378 ± 183 pixels in controls; P = 0.039) (S3 Fig). Similar to what we observed in the eye discs, this partial rescue was associated with a decrease in the level of Xbp1-EGFP in the wing disc (0.76 ± 0.17 in wing-Rh1G69D/Baldspoti relative to controls, 1.00 ± 0.14; P = 6.0 x 10−3) (Fig 4B and 4C). We also measured JNK activation in these wing discs by monitoring the puc-LacZ transgene. Again, as we saw in eye discs, LacZ expression is significantly reduced in wing discs lacking Baldspot (0.64 ± 0.12 in wing-Rh1G69D/Baldspoti relative to controls, 1.00 ± 0.12; P = 4.7 x 10−3) (Fig 4D and 4E). It appears that loss of Baldspot modifies UPR signaling and degeneration through the same mechanisms across different tissues and is not eye disc-specific.
To test whether Baldspot modifies ER stress signaling by pharmacological induction, we used tunicamycin to induce ER stress. Tunicamycin inhibits N-linked glycosylation in the ER and results in massive misfolding, ER stress, and a robust UPR [41]. Treatment with tunicamycin in Drosophila larvae results in a robust activation of the ER stress response [33, 42]. We acutely exposed 2nd instar control larvae (Tub-GAL4, no RNAi) and larvae with ubiquitous knockdown of Baldspot (Tub-GAL4>Baldspot RNAi) to tunicamycin or control DMSO and scored survival to pupation as a proxy for the physiological ability to recover from ER stress. To account for any larval lethality due to loss of Baldspot, we normalized the survival of larvae on tunicamycin to the DMSO treatment. Strikingly, larvae lacking Baldspot were less susceptible to tunicmycin-induced lethality than control larvae (Fig 4F, S1 Table), as evidenced by increased survival to pupation (P = 9.2 x 10−4). This was accompanied by reduced Xbp1 splicing (S4 Fig). We conclude that the modifying effect of Baldspot expression on the ER stress response is generalizable across different tissues and methods of ER stress induction.
Because our data indicates that Baldspot regulates Ire1/Xbp1 signaling and IRE1-dependent JNK signaling, we next tested whether Baldspot regulates another key IRE1 function: Regulated IRE1 Dependent Degradation of mRNAs (RIDD) [10, 11, 43]. In addition to Xbp1 splicing, the RNase activity of activated IRE1 is also important for the degradation of a number of ER-associated transcripts through RIDD. To test if RIDD activity is affected by loss of Baldspot expression, we measured levels of known RIDD target mRNAs, under ER stress, in control cells and tissues with or without Baldspot expression. We employed S2 cell culture and treated S2 cells with dsRNA targeting either Baldspot or EGFP as a control. This treatment resulted in a near complete reduction in Baldspot transcript levels (~7.8% of control; P = 1.3 x 10−9) (Fig 5A).
We used Dithiothreitol (DTT) to induce ER stress in S2 cells. DTT is a reducing agent that breaks disulfide bonds in proteins, resulting in misfolded proteins and a strong ER stress response [44]. As expected, treatment of cells with DTT increased Xbp1 splicing ~3-fold, indicative of ER stress. In agreement with our in vivo studies in the Rh1G69D fly, this splicing was significantly reduced in the absence of Baldspot, with only ~2-fold increase (P = 0.023), indicating that IRE1 signaling is disrupted (Fig 5B). We also monitored PERK activity by measuring levels of P-eif2α, which increase in control cells upon treatment with DTT (Fig 5C), as expected (P = 0.021; N = 3). In cells treated with Baldspot DsRNA, there was no increase in P-eif2α levels upon DTT exposure (P = 0.954; N = 3) (Fig 5C). Due to a lack of antibody, we could not monitor ATF6 activity in S2 cell culture. It appears that loss of Baldspot expression in S2 cells has a similar effect on ER stress response as observed in our in vivo model.
We next measured IRE1 RIDD activity in the absence of Baldspot in S2 cells. When RIDD is activated, ER-localized mRNAs are cleaved by the RNase domain of IRE1, leading to a reduction in protein translation into the ER [10, 43]. A number of these targets have been well-characterized in Drosophila, including the transcripts sparc and hydr2. Both of these mRNAs are degraded upon the initiation of ER stress [43, 45]. We tested the effect of Baldspot knockdown on these targets after treatment with DTT. Hydr2 expression is significantly reduced in control cells treated with DTT (0.71 ± 0.07; P = 2.0 x 10−3), but not in those treated with DsRNA against Baldspot (0.91 ± 0.16; P = 0.41; P = 0.021) (Fig 5D). However, we found that sparc is equally degraded upon DTT treatment in cells treated with DsRNA against either Baldspot (0.74 ± 0.09; P = 2.0 x 10−5) or EGFP (0.71 ± 0.06; P = 1.9 x 10−5) (Fig 5D). Non-RIDD targets, such as Actin5C, were unaffected (Fig 5D). The fact that only some RIDD targets are affected is in agreement with our other results: loss of Baldspot activity partially reduces IRE1 functions, including RIDD.
The ER stress response has the potential to be an important modifier of human disease. ER stress is activated in a wide variety of diseases ranging from Mendelian degenerative diseases to complex metabolic disorders [7]. The role of the ER stress response is complex and in some cases, a stronger response is beneficial and in other cases, a strong response underlies the disease. In some forms of retinitis pigmentosa, like the focus of this study, loss of vision is a direct result of the ER stress-induced cell death [46]. Neurodegenerative diseases such as Parkinson’s, Alzheimer’s, Amyotrophic Lateral Sclerosis (ALS), and prion disease are all associated with the accumulation of cytoplasmic misfolded proteins that can activate a secondary ER stress response and subsequent cell death [19]. In diabetes and obesity, misfolded proteins and the ER stress response are at least partially responsible for the cellular dysfunction that is a hallmark of the disease state [47]. In some cancers, however, cells selectively upregulate ER stress pathways associated with survival, while inhibiting those associated with cell death [48]. Genetic variation that alters the activation of the ER stress response may modify disease in the population.
We and others have shown that genetic variation can influence how and to what extent the ER stress response is activated. Genetic variation in Drosophila [22, 23], mouse [21], and human cells [20] can impact the expression of genes involved in ER stress pathways. Furthermore, this variation in expression has been linked to differences in the UPR and the downstream activation of cell death [21–23]. It is apparent that cryptic genetic variation drives these extensive inter-individual differences in the ER stress response [21]. Stress or disease conditions reveal the effect of this otherwise silent genetic variation. The same variation that drives differences in this important stress response might also influence disease heterogeneity. Indeed, modulating the ER stress response is an effective way of modulating disease outcome in models of human disease. [49–54].
In this study we show that loss of Baldspot/ELOVL6 activity can affect the outcome of retinal degeneration and has the potential to be a broader regulator of ER stress-associated diseases. As a 16-to-18 carbon fatty acid elongase in the ER membrane, this enzyme is responsible for increasing the concentration of C18 stearate, a saturated long chain fatty acid [28, 37]. Increases in stearate concentrations in the ER membrane can induce and enhance ER stress, and is associated with disease outcomes in mammalian systems [36–38]. Here, we show that reducing the activity of Baldspot/ELOVL6 reduces the degree to which the UPR and apoptosis are induced in models of ER stress. Our study shows that Baldspot/ELOVL6 has a modifying effect on the IRE1 and PERK branches of the ER stress response. This is consistent with observations that increased long chain fatty acids in the ER membrane impacts IRE1 and PERK signaling [38, 39]. Future research will explore the Baldspot/ELOVL6 interaction with ER membrane composition and how this alters IRE1 and PERK oligomerization and function. It is particularly interesting to note that RIDD is only partially affected in the absence of Baldspot. In Drosophila, the RIDD activity of IRE1 targets most ER-localized mRNAs [43]. In contrast, mammalian RIDD is much more selective, and only mRNAs with specific signal sequences and RNA secondary structures are degraded by IRE1 [55]. More studies on how these processes are regulated in Drosophila and mammals is necessary to understand the differences in RIDD between species as well as the conservation of Baldspot/ELOVL6 function. These studies will enhance our understanding of how these pathways are regulated in different disease contexts.
Therapeutics that reduce ELOVL6 activity could conceivably be used to treat or delay progression in a variety of ER stress-associated diseases. In RP patients with mutations that cause the accumulation of misfolded rhodopsin, similar to our Drosophila model, reduction of ELOVL6 expression might reduce cell death, delaying vision loss. We present evidence that the modulation of the ER stress response through Baldspot/ELOVL6 activity may impact the severity of diseases beyond RP. Particularly intriguing is that Baldspot/ELOVL6 can alter the ER stress response when stress is induced by different conditions. Regardless of the mechanism by which ER stress is induced, inhibiting ELOVL6 activity reduces IRE1 and PERK signaling. In fact, in a recent study of the db/db mouse model of diabetes, loss of ELOVL6 alters insulin sensitivity and the expression of ER stress genes [26]. It is likely that reducing Baldspot/ELOVL6 expression would also alter complex diseases where ER stress contributes to degeneration and cell death. In some cases, even a small delay in degeneration could dramatically improve the quality of life.
We have identified Baldspot/ELOVL6 in Drosophila as a novel modifier of the ER stress response. Its potential as a therapeutic target extends beyond RP to any number of ER stress-associated disorders. The identification of Baldspot in a natural genetic variation screen [23] suggests that variation in expression of this gene is well tolerated under healthy conditions and might be amenable to modulation. Identifying variable elements of the ER stress response pathway may prove to be a successful strategy to nominating new modifiers with broad applications to many ER stress-associated diseases.
Flies were raised at room temperature on standard diet based on the Bloomington Stock Center standard medium with malt. The strain containing GMR-GAL4 and UAS-Rh1G69D on the second chromosome (GMR>Rh1G69D) has been previously described [23]. The following strains are from the Bloomington Stock Center: MS1096-GAL4 (8696), Baldspot RNAi (44101), control attP40 (36304). The puc-LacZ enhancer trap is available from Kyoto (109029). A second Baldspot RNAi line (Vienna Drosophila Resource Center: 101557KK) showed similar rescue of the original Rh1G69D phenotype. Most analysis was performed in the Bloomington Stock Center RNAi line. The strains containing the UAS-Xbp1-EGFP transgenes were a gift from Don Ryoo (NYU) [24].
For eye and wing images, adult females were collected under CO2 anesthesia and aged to 2–7 days, then flash frozen on dry ice. Eyes were imaged at 3X magnification using a Leica EC3 camera. Wings were dissected away from the body, then imaged at 4.5X magnification using the same camera. Eye area was measured in ImageJ as previously described [23].
Eye discs and wing discs were dissected from wandering L3 larvae in cold 1X PBS, then immediately transferred to cold 4% PFA on ice. Tissues were fixed in 4% PFA for 15–20 min, then washed in 1XPAT (0.1% TritonX100) prior to blocking with 5% normal donkey serum. Tissues were stained with primary antibodies for rhodopsin (1:50, Developmental Studies Hybridoma Bank #4C5), GFP (1:2000, Thermo-Fisher #A6455), and LacZ (1:20, Developmental Studies Hydbridoma Bank #40-1a). Apoptosis was monitored using the ApopTag Red In Situ Apoptosis Detection Kit (Millipore #S7165). Tissues were mounted in Slowfad Diamond Antifade Mountant (ThermoFisher #S36967) and imaged with an Olympus FV1000 confocal microscope.
For GFP blots, 5 L2 larvae were homogenized in 1X Laemmli/RIPA buffer containing 1X protease inhibitors. For P-eif2α blots, protein from S2 cells or brain/eye-imaginal discs was isolated in 1X Laemmli/RIPA buffer containing 1X protease inhibitors (Roche cOmplete Mini EDTA-free protease inhibitor tablets), as well as the phosphatase inhibitors Calyculin A and okadaic acid.
Equivalent amounts of protein were resolved by SDS-PAGE (10% acrylamide) and transferred to PVDF membrane by semi-dry transfer. Membranes were then treated with either 5% BSA or 5% milk protein block in 1XTBST prior to immunoblotting. Blots were probed with antibodies for GFP (1:5000, Thermo-Fisher #A6455), tubulin (1:2000, Developmental Studies Hybridoma Bank #12G10), P-eif2α (1:1000, abcam #32157), and Pan-eif2α (1:500, abcam #26197). Blots shown are representative of at least three biological replicates, and quantification was performed using Image J software.
Crosses were set up on the control, standard media alone, or standard media with 10% stearate supplementation. To make stearate supplemented media, standard media was melted in the microwave and kept at 98°C on a stir plate. Stearate (10% final concentration) was added to this media and mixed until lipids were homogenous with the media. The standard control diet received the same treatment, but no additional stearate was added. To determine retinal degeneration, flies were aged and imaged as described above. Antibody staining was performed on the eye imaginal discs of L3 larvae, also as described above.
Crosses to generate the indicated genotypes were set up on egg caps containing yeast paste. L2 larvae were then treated with either 10 μg/mL Tunicamycin (diluted 1:1000 from a 10 mg/mL stock solution) or 1:1000 DMSO in Schneider’s media for 30 minutes at room temperature. The larvae were then washed in 1XPBS twice and placed on solid media containing 6% yeast and 1% agar. Viability was determined by survival to pupation. Survival for each genotype was normalized to the DMSO-treated control condition. Each replicate represents ~25–60 larvae per genotype; a total 180 to 250 larvae per genotype were analyzed across all replicates. Protein was isolated from five L2 larvae expressing Xbp1-EGFP, one hour after treatment was concluded. A Western blot was performed to analyze Xbp1-EGFP levels.
DsRNA was generated using the MEGAscript T7 Transcription kit (ThermoFisher #AM1334), with primers for EGFP (F: TTAATACGACTCACTATAGGGAGACCACAAGTTCAGCGTGTCC and R: TTAATACGACTCACTATAGGGAGAGGGGTGTTCTGCTGGTAGTG) and Baldspot (F: TTAATACGACTCACTATAGGGAGAATCCGCCCAGGTTCATCTCG and R: TTAATACGACTCACTATAGGGAGAGTCACATCTCGCAGCGCAAC). S2 cells were treated with DsRNA against EGFP (as a control) or against Baldspot at a density of approximately 2 x 106 cells/mL in a 24-well plate. Cells were incubated with DsRNA for 4–7 days before being split and treated with either DMSO as a control or DTT. Concentrations of DTT varied with experiments: 0.5 mM for 1 hour (Xbp1 splicing), 2.0 mM for 3 hours (RIDD target analysis), or 2.0 mM for 4 hours (P-eif2α detection). RNA was isolated from cells using either Trizol/chloroform extraction (Xbp1 splicing) or the Direct-zol RNA Miniprep Kit (qPCR) and used to generate cDNA (Protoscript II, NEB). Protein was isolated from cells as described above and levels of P-eif2α normalized to Pan-eif2α as a loading control compared between matched DMSO or DTT-treated S2 cells.
Knockdown of Baldspot was confirmed using qPCR (primers: F: TGCTGGTCATCTTCGGTGGTC and R: ACGCAGACGGAGTGGAAGAG). Xbp1 splicing was evaluated from the cDNA using PCR (primers for Xbp1: F: TCAGCCAATCCAACGCCAG and R: TGTTGTATACCCTGCGGCAG). The spliced and unspliced bands were separated on a 12% acrylamide gel and the proportion of these bands quantified using ImageJ software. RIDD target levels were analyzed by qPCR (sparc primers: F: AAAATGGGCTGTGTCCTAACC and R: TGCAGCACAATCTACTCAATCC; hydr2 primers: F: CGCATACACGACTATTTAACGC and R: TTTGGTTTCTCTTTGATTTCCG; Actin5C primers: F: ATGTGTGACGAAGAAGTTGCT and R: GAAGCACTTGCGGTGCACAAT; rpl19 primers: F: AGGTCGGACTGCTTAGTGACC and R: CGCAAGCTTATCAAGGATGG). Transcript levels were normalized to rpl19 and compared between matched DMSO or DTT-treated S2 cells.
Statistics were calculated using R software. P-values were determined using ANOVA for eye size, fluorescence levels, transcript levels in qPCR, and protein levels by Western blot. Tukey multiple testing correction was used for the fatty acid feeding experiment and RIDD target analysis. Bonferroni correction was used for Xbp1 splicing in S2 cells. A pairwise T-test was performed for larval tunicamycin treatment and P-eif2α response to DTT in S2 cells. A cutoff of P = 0.05 was used for significance.
|
10.1371/journal.ppat.1005520 | Optimal Combinations of Broadly Neutralizing Antibodies for Prevention and Treatment of HIV-1 Clade C Infection | The identification of a new generation of potent broadly neutralizing HIV-1 antibodies (bnAbs) has generated substantial interest in their potential use for the prevention and/or treatment of HIV-1 infection. While combinations of bnAbs targeting distinct epitopes on the viral envelope (Env) will likely be required to overcome the extraordinary diversity of HIV-1, a key outstanding question is which bnAbs, and how many, will be needed to achieve optimal clinical benefit. We assessed the neutralizing activity of 15 bnAbs targeting four distinct epitopes of Env, including the CD4-binding site (CD4bs), the V1/V2-glycan region, the V3-glycan region, and the gp41 membrane proximal external region (MPER), against a panel of 200 acute/early clade C HIV-1 Env pseudoviruses. A mathematical model was developed that predicted neutralization by a subset of experimentally evaluated bnAb combinations with high accuracy. Using this model, we performed a comprehensive and systematic comparison of the predicted neutralizing activity of over 1,600 possible double, triple, and quadruple bnAb combinations. The most promising bnAb combinations were identified based not only on breadth and potency of neutralization, but also other relevant measures, such as the extent of complete neutralization and instantaneous inhibitory potential (IIP). By this set of criteria, triple and quadruple combinations of bnAbs were identified that were significantly more effective than the best double combinations, and further improved the probability of having multiple bnAbs simultaneously active against a given virus, a requirement that may be critical for countering escape in vivo. These results provide a rationale for advancing bnAb combinations with the best in vitro predictors of success into clinical trials for both the prevention and treatment of HIV-1 infection.
| In recent years, a new generation of monoclonal antibodies has been isolated from HIV-1 infected individuals that exhibit broad and potent neutralizing activity when tested against diverse strains of virus. There is a high level of interest in the field in determining if these antibodies can be used to prevent or treat HIV-1 infection. Because HIV-1 is adept at escaping from immune recognition, it is generally thought that combinations of multiple antibodies targeting different sites will be required for efficacy, much the same as seen for conventional antiretroviral drugs. How many and which antibodies to include in such combinations is not known. In this study, a new mathematical model was developed and used to accurately predict various measures of neutralizing activity for all possible combinations having a total of 2, 3, or 4 of the most promising antibodies. Through a systematic and comprehensive comparison, we identified optimal combinations of antibodies that best complement one another for enhanced anti-viral activity, and therefore may be most effective for the prevention or treatment of HIV-1 infection. These results provide important parameters that inform the selection of antibodies to develop for clinical use.
| The ability to elicit potent broadly neutralizing antibodies through immunization remains an elusive goal in the development of an effective HIV-1 vaccine [1]. This has motivated major efforts over the past 6 years to isolate and characterize Env-specific antibodies from HIV-1-infected individuals who exhibit broad and potent serum neutralizing activity [2–4]. Through technological advances in single cell sorting of antigen-specific memory B cells [5–11], high-throughput antibody cloning and screening methods, numerous novel monoclonal antibodies have since been isolated, some of which exhibit exceptional neutralization breadth and potency when tested in vitro against large panels of diverse HIV-1 isolates [7, 9–20]. Identification of the epitope targets of these bnAbs has dramatically expanded our knowledge regarding sites of common vulnerability on the Env spike [21]. Major epitope targets include the CD4bs [5, 11, 16, 19, 22–27], a glycan-dependent site in variable region 3 (V3) of gp120 [9, 17, 28–31], a V1/V2 glycan-dependent quaternary site on the apex of the Env trimer [9, 10, 12, 32–37], the MPER [15, 38–41], and epitopes bridging both gp120 and gp41 [13, 14, 18, 42]. The hope remains that characterization of these epitope targets and efforts to elucidate the pathways of bnAb development in vivo will eventually result in the rational design of novel immunogens and immunization strategies for eliciting such antibodies through vaccination [12, 16, 24, 43–46]. However, a more immediate potential exists for using bnAbs in clinical settings of passive transfer for the prevention and/or treatment of HIV-1 infection.
In support of preventative modalities, pre-clinical studies in non-human primates (NHP) have demonstrated that passive transfer of bnAbs can confer sterilizing protection against high dose mucosal challenges with chimeric simian-human immunodeficiency viruses (SHIVs) [23, 47–53]. Studies in NHP and humanized mice have further investigated the therapeutic potential of bnAb infusion in the setting of established viral infection, and demonstrated that transfer of single bnAbs can result in a transient decline in plasma viremia, reduction of proviral DNA, and in some cases extended control of viral replication [53–56]. However, viral rebound generally occurs once the concentration of transferred antibody decays below the therapeutic range, and the emergence of neutralization resistant escape variants is often observed. Similar observations were recently described in a phase I clinical study evaluating passive infusion of the CD4bs bnAb 3BNC117 in HIV-1 infected individuals [57]. While escape from antibody monotherapy remains a concern, additional data from animal model studies have shown that therapeutic strategies employing combinations of bnAbs to simultaneously target different epitopes on the Env spike can impede viral rebound and escape, and exert sustained control of viral replication [53–55]. Thus, for bnAbs to be effectively employed for treatment of HIV-1 infection, combinations of multiple antibodies will likely be required to confront the extraordinary diversity of the virus and its ability to escape from selective immune pressure.
Recent studies of in vitro neutralization have established that combinations of bnAbs targeting distinct epitopes can act in a complementary and additive manner, and exhibit improved neutralization breadth and potency compared to single bnAbs [58–60]. In the study by Kong et al., it was shown that the breadth and potency of bnAb combinations could be reliably predicted using an additive model, with consistent patterns of minor non-additive interactions for particular bnAb combinations, either antagonistic or synergistic [60]. Certain double, triple and quadruple bnAb combinations were found to achieve 89 to 100% coverage when tested against a large diverse multiclade virus panel. However, due to the complementary nature of the bnAb combinations, in many cases increased breadth was due to only a single bnAb in the mixture exhibiting neutralizing activity against a given virus. In a clinical setting, such a bnAb combination would in essence be the equivalent of single antibody monotherapy against a substantial fraction of viruses, which would have a greater opportunity for escape. Thus, for treatment of HIV-1 infection, it may be advantageous to use bnAb combinations that offer the best potential for active coverage of most viruses by two or more antibodies.
For bnAb immunotherapy in the setting of chronic infection, viral clearance is the most desirable outcome, albeit challenging to achieve. Thus, more complex options are being considered, such as including combinations of the most potent bnAbs together with latency reversing agents (LRAs) and standard antiretroviral drug treatment [61–63]. For such strategies to be beneficial, bnAbs will need to be effective at three levels. First, they will need to neutralize the diversity of viruses circulating in the population targeted for treatment. Second, they will need to effectively neutralize the complex within-host quasispecies that develop during chronic HIV-1 infection. And finally, they should be effective against the full spectrum of expressed forms of Env on any given virion. It has been observed that some bnAbs exhibit neutralization curves that plateau well below 100% when tested against particular Env pseudoviruses in vitro [10, 13, 64, 65]. This well-established behavior is surprising given the genetically clonal nature of viruses used in these assays, and could possibly stem from post-translational variation in the glycosylation patterns or alternate variable loop and structural configurations of expressed Env [13, 65–68]. It is a concern that such incomplete neutralization may pose a severe limitation for achieving the desired therapeutic efficacy in vivo. Thus, an ideal immunotherapy candidate antibody combination should maximize the genetic and antigenic spectrum of viruses that are potently neutralized, while minimizing the impact of incomplete neutralization.
A key question that remains is how many bnAbs will be required for long term beneficial effects in a preventative or therapeutic setting, and which combinations of bnAbs will provide the most potent and active coverage for testing in human clinical trials. Over the past several years, multiple bnAbs for each major epitope have emerged as viable candidates based on extensive in vitro and pre-clinical animal model testing. Given the tremendous resources required to move even a single candidate bnAb forward into human clinical trials, rational decisions must be made to select single antibodies, bivalent antibodies, or components of bnAb combinations that will theoretically provide the highest potency and coverage against the diversity of circulating HIV-1. As bnAb clinical efficacy studies are currently being planned for conduct in southern Africa, coverage and potency of bnAbs against the HIV-1 clade C viruses that dominate the epidemic in that region is of considerable interest.
Here we utilized a newly described panel of 200 acute/early clade C HIV-1 Env pseudoviruses to assess the breadth and potency of 15 of the most promising bnAb candidates targeting four major epitopes of HIV-1 Env. A mathematical modeling approach was developed that increased the accuracy in predicting neutralization titers of bnAb combinations. We experimentally validated the improved accuracy of this model, and then used it to predict the behavior of all possible 2, 3, and 4 bnAb combinations using data derived from single bnAb testing. Using these predictions, we compared the performance of a comprehensive spectrum of potential bnAb combinations, and identified those that provide the most optimal potency, breadth, complete neutralization, and active coverage.
A panel of bnAbs targeting HIV-1 Env was used to assess and compare the breadth and potency of neutralization against acute/early clade C Envs. Fifteen bnAbs were selected that target four distinct epitope regions: the CD4 binding site (CD4bs: 3BNC117, VRC01, VRC07, VRC07-523, VRC13) [11, 19, 23, 69, 70], the V3-glycan supersite (V3g: 10–1074, 10-1074V, PGT121, PGT128) [9, 17], the V1/V2-glycan site (V2g: PG9, PGT145, PGDM1400, CAP256-VRC26.08, CAP256-VRC26.25) [9, 10, 12, 20, 32], and the gp41 MPER epitope (10E8) [15]. BnAbs were tested against a panel of 200 clade C HIV-1 Env pseudoviruses using the validated luciferase-based TZM-bl assay. This virus panel consists of viruses isolated from individuals in the acute/early stages of infection from five southern African countries, including South Africa, Tanzania, Malawi, Zambia, and Botswana. Serial dilutions of individual bnAbs were tested against each virus using a starting concentration that ranged from 10–50 μg/ml, depending on sample availability at the time of testing. Neutralizing activities were evaluated using potency-breadth curves (the percentage of viruses neutralized versus an IC50 or IC80 cutoff, Fig 1A and 1B), scatter plots (Fig 1C and 1D) and heatmaps (Fig 1E and 1F). The 5 bnAbs targeting the V1/V2-glycan region neutralized between 67–75% of viruses with positive IC50 titers, and the 4 bnAbs targeting V3-glycan neutralized 54–68%. When positive, these glycan-dependent bnAbs were strikingly potent. Using the more stringent IC80 measure, median IC80 titers ranged from 0.003–1.274 μg/ml for V1/V2-glycan and 0.073–0.203 μg/ml for V3-glycan bnAbs (Table A in S1 Text). CD4bs bnAbs tended to exhibit greater breadth (71–96% at IC50), but were generally less potent than V1/V2-glycan or V3-glycan antibodies (median IC80 titers 0.30–1.58 μg/ml). The MPER directed bnAb 10E8 exhibited lower overall potency (median IC80 3.399 μg/ml), yet had exceptional IC50 breadth, neutralizing 98% of viruses. Even the most resistant isolates were sensitive to at least 3 bnAbs, which most often targeted the CD4bs or MPER. Overall, clear differences in potency and/or breadth were observed among bnAbs of the same class (defined here as bnAbs that target the same epitope region). Based on IC50 and IC80 titers, best-in-class bnAbs were CAP256-VRC26.25 (V2-glycan), 10-1074V (V3-glycan), VRC07-523 (CD4bs), and 10E8 (MPER).
As visualized in heat maps (Fig 1E and 1F), and by hierarchical clustering (Fig A in S1 Text), bnAbs targeting the same epitope region exhibit similar patterns of neutralizing activity, with clear patterns of complementarity between epitope classes. For example, distinct clusters of viruses were resistant to V1/V2-glycan antibodies but sensitive to V3-glycan antibodies, whereas other virus clusters exhibit the opposite phenotype. These data illustrate how different combinations of bnAbs targeting distinct epitopes can complement one another for enhanced coverage against clade C viruses.
Because it is not practical to assay all combinations of bnAbs against a large panel of viruses, a new method to accurately predict combination bnAb neutralization efficacy using the available large-scale single bnAb neutralization data was developed to facilitate rational decisions for selection of the best bnAb combinations for clinical testing.
In a previous study by Kong et al., the additive model worked well in predicting potency of bnAb combinations using experimental data from single bnAbs [60]. They also found that the experimental bnAb combination data deviated slightly from model predictions. Most combinations performed slightly better than predicted, while a few combinations that included a V3-glycan bnAb performed slightly worse than predicted. The additive model derives from an application of equilibrium mass action kinetics to simplified in vitro antibody-virus interactions (S1 Text). This theoretical treatment assumes that single bnAb neutralization curves follow Hill curves with Hill exponents equal to one, and that antibodies act independently with little possibility of multiple antibodies inhibiting the same virion. The first assumption of a unit Hill exponent is largely valid for CD4bs and V3-glycan bnAbs, however, bnAbs targeting the V2-glycan and MPER epitopes frequently exhibit Hill exponents of less than 1 [65, 71, 72].
To overcome these limitations of the additive model, we developed a new model, the “Bliss-Hill model” (BH model). This model combines single bnAb Hill curves (with arbitrary slopes) within the framework of the Bliss independence model for the binding of multiple species of ligands to a substrate [72, 73], and incorporates a correction for multiple ligands independently attaching to the substrate (S1 Text). We tested the BH model by using experimental data from combination bnAb neutralization assays. The assays comprised 10 combinations of 2, 3 and 4 bnAbs (including 2-bnAb combinations with both antibodies targeting similar epitopes, Fig B in S1 Text) assayed against a smaller panel of 20 viruses. The 20 viruses were chosen because they are sensitive to almost all bnAbs tested and comprise a maximized range of IC80 titers for the bnAb combinations. The BH model proved highly accurate in explaining the clade C panel bnAb combination data (Fig 2A, R2 = 0.9154, Pearson r = 0.9584). Moreover, the BH predictions were closer to the observed data than the additive model for 9 of the 10 combinations tested (Fig 2B, p = 0.021 using Binomial Test), with the only exception being the combination VRC07-523 + 10-1074V. Thus the BH model offered a significant, though modest in magnitude, improvement in prediction accuracy over the additive model. We confirmed this by reanalyzing a larger dataset from Kong et al., and again found the BH model predictions to be highly accurate (R2 = 0.9655, Pearson r = 0.9862, Fig C in S1 Text). The BH model performed slightly better than the additive model in all cases, and the difference reached high levels of statistical significance for most of the 2, 3, and 4 bnAb combinations tested. This improvement was due to the systematic trend of BH predictions being more potent than the additive model predictions (Figs D and E in S1 Text), and thus closer to the observed titers since additive model predictions were found to be less potent than the observed titers for most combinations [60].
Nonetheless, for some antibody combinations, experimentally measured IC80 titers still showed minor deviations from the BH model predictions (Fig 2, Figs C-G in S1 Text). For a few viruses, the combination IC80 titers were 3-fold higher than the most potent bnAb in the combination (Fig D in S1 Text), which is counter-intuitive since both the additive and BH models predict greater potency for combinations relative to the component bnAbs. In such cases we find that the very potent neutralization of a virus by an antibody (particularly CAP256-VRC26.25, Fig D in S1 Text) is somewhat inhibited by the presence of additional antibodies, albeit still resulting in potent neutralization by the combination. Models that incorporated additional parameters based on observed deviations could further improve predictions in some cases (S1 Text, Figs F and G in S1 Text), but the magnitude of deviations were small for most viruses. Furthermore, using deviation modeling with BH model (Fig H in S1 Text) or using additive model (Fig I in S1 Text) did not affect the conclusions below, as the best combinations selected were robust using either model.
Passive and active immunization strategies that aim to protect against the acquisition of HIV-1 infection would benefit from information regarding how many and which bnAb combinations provide optimal coverage and potency. An antibody that may have the best characteristics when considered alone may not have the optimal complementarity when considered for combination bnAb regimens. We predicted the combination scores for all potential 2, 3 and 4 bnAb combinations using the BH model on single bnAb neutralization data for 15 bnAbs against 200 clade C viruses, thus enabling direct comparisons of bnAb combinations. For 2 bnAb combinations, only combinations consisting of bnAbs targeting different epitopes were considered, while for 3 and 4 bnAb combinations, multiple bnAbs targeting the same epitope region were also considered. Predicted potency-breadth curves for all of the 2, 3 and 4 bnAb combinations (1,622 combinations total) are shown in Fig 3.
The combinations were stratified by the number of bnAbs targeting different epitopes (referred to as “categories”, e.g., CD4bs+V2g is a combination of a CD4bs and a V2-glycan bnAb, and V2g(2x)+V3g has two V2-glycan and one V3-glycan bnAbs). Within each category, multiple combinations were possible due to multiple bnAbs targeting the same epitope. Best-in-category bnAb combinations were identified as those with the lowest geometric mean IC80 values for the 200 viruses (highlighted in Fig 3 by dark, bold lines). Of note, the area under the IC80 potency-breadth curve is negatively, but linearly, and almost perfectly correlated to the Log10 geometric mean IC80. Thus using either measure gives identical results. The best-in-category combinations were not always clear, as second best combinations were very comparable (e.g. CAP256-VRC26.25 + 10-1074V + PGT128 or PGT121 with geometric mean IC80 of 0.007 and 0.0071μg/ml, respectively). Comparisons of best-in-category combinations having the same number of bnAbs are shown in Fig 4.
Best-in-category 2 bnAb combinations had significantly better predicted potency (geometric mean IC80 range = 0.02–0.29 μg/ml) and breadth (88.5–97.5% of viruses with IC80 < 10 μg/ml), than single bnAbs (geometric mean IC80 = 0.17–5.91 μg/ml and breadth = 44–92.5%). The two best-in-category 2 bnAb combinations, CAP256-VRC26.25 (V2-g) with either 10-1074V (V3-g) (geometric mean IC80 = 0.020 μg/ml) or VRC07-523 (CD4bs) (geometric mean IC80 = 0.021 μg/ml) were significantly better than the other best-in-category 2 bnAb combinations (p < 0.01 and q-value < 0.02) (Fig 4A, 4B and 4C). However, it was unclear which of these two combinations was better, because each pairing had different advantages. While CAP256-VRC26.25 and 10-1074V alone are more potent than VRC07-523 when active (Table A in S1 Text), they have more limited breadth, each neutralizing ~60% viruses at IC80 < 10 μg/ml as compared to 92.5% for VRC07-523. Consistent with this, we found that the combination of CAP256-VRC26.25 + 10-1074V missed ~13% of viruses at IC80 < 10 μg/ml, while CAP256-VRC26.25 + VRC07-523 missed only ~3%. Thus, while CAP256-VRC26.25 + VRC07-523 was slightly less potent than CAP256-VRC26.25 + 10-1074V, it provides ~10% better coverage.
For 3 bnAb combinations, the best breadth and potency was seen with CAP256-VRC26.25 + 10-1074V + VRC07-523 (Fig 4D, 4E and 4F). This combination, which targets 3 separate epitopes, neutralized 99.5% viruses (all but one in the panel) at IC80 < 10 μg/ml, with a geometric mean IC80 of 0.0083 μg/ml. The superior performance of this combination draws from the complementary neutralization profiles of the most potent panel bnAbs, CAP256-VRC26.25 and 10-1074V, combined with the broad and potent profile of VRC07-523 (Fig 1). This combination was significantly more potent than most other best-in-category 3bnAb combinations (p < 0.02, q < 0.03). Replacing VRC07-523 with either PGDM1400 or 10E8 in combinations containing CAP256-VRC26.25 + 10-1074V resulted in a small loss of potency and breadth that was not statistically significant. Overall, 3 bnAb combinations showed improved breadth (89 to 99.5% at IC80 < 10 μg/ml) and markedly improved potency (geometric mean IC80 of 0.008–0.060 μg/ml) than 2 bnAb combinations, with 6 out of 7 best-in-category 3 bnAb combinations predicted to have better geometric mean IC80 than the best 2 bnAb combinations.
The two best-in-category 4 bnAb combinations, one targeting 3 epitopes and another targeting 4 epitopes, had comparable potency (geometric mean IC80 ~ 0.007 μg/ml) and breadth (99.5% at IC80 < 10 μg/ml) (Fig 4G, 4H and 4I), and were more potent and broadly active than 4 bnAb combinations targeting only 2 epitopes (geometric mean IC80 of 0.01 to 0.05 μg/ml and breadth 92–98.5% at IC80 < 10 μg/ml). Thus bnAb combinations targeting three epitopes showed a significant gain in breadth and potency compared to those targeting two, but the further gain in targeting all four major epitopes, for this panel is negligible. This information is useful to efforts that aim to achieve optimal coverage and potency to protect against the acquisition of infection in passive or active vaccination settings, but does not take into account ease of escape in the setting of passive immunotherapy for active infection.
Combinations of bnAbs are likely to be advantageous in a therapeutic setting not only to maximize potency and breadth but also to minimize the potential for viral escape by targeting multiple epitopes simultaneously [55]. Thus, we investigated the extent of simultaneous neutralization by two or more bnAbs in the best-in-category bnAb combinations at different activity thresholds.
First we quantified the percent of panel viruses actively neutralized by at least 2, 3 or 4 bnAbs in all best-in-category 2, 3 and 4 bnAb combinations at physiologically relevant concentrations. We used IC80 thresholds of 1, 5 and 10 μg/ml, which fall in the range of bnAb serum concentrations in HIV-1 infected patients administered a single dose of 1–30 mg/kg of 3BNC117 [57]. For combinations with multiple bnAbs targeting the same epitope class, a modified counting procedure was employed that accounted for overlap in escape-associated mutations (S1 Text). The percent of viruses neutralized by the best bnAb combinations at different thresholds of activity are shown in Table B in S1 Text. We modified the potency-breadth curves for best-in-category bnAb combinations to highlight cases where multiple bnAbs in a combination were simultaneously active (Fig 5). These curves show cumulative coverage of the 200 panel viruses at a given predicted combination IC80 value limited by counting only those viruses that were simultaneously sensitive to 2, 3 or 4 bnAbs at single bnAb IC80 < 1, 5, or 10 μg/ml.
When the percentage of viruses neutralized by at least 2 bnAbs was considered, the best coverage at our least restrictive threshold within the experimental assay range of IC50 <10 μg/ml was 92.5%, 97.5% and 100% for 2, 3 and 4 bnAb combinations, respectively (Table B in S1 Text, Fig 5). This coverage decreased, as expected, to 80%, 91% and 95.5%, respectively, when a more stringent IC80 <10 μg/ml threshold was used, and continued to decrease until only 44%, 67.5% and 73.5% coverage was seen, respectively, at our most stringent threshold of IC80 <1 μg/ml. The percentage of viruses neutralized when requiring at least three bnAbs in the best-in-category 3 and 4 bnAb combinations to be active was of course even lower at each of these thresholds. Here, the best coverage at the less restrictive threshold of IC50 <10 μg/ml was 66.5% and 89% for 3 and 4 bnAb combinations, respectively, and progressively decreased to only 19.5% and 26.5% coverage at the most stringent IC80 <1 μg/ml threshold. Poor coverage was seen at all thresholds when all 4 bnAbs in the best-in-category 4 bnAb combinations were required to be active.
Using extrapolated single bnAb neutralization curves (see “BnAb combinations reduce levels of incomplete neutralization” below), we also investigated coverage with multiple active bnAbs using single bnAb IC80 < 50 μg/ml and < 100 μg/ml (Fig J in S1 Text). These concentrations roughly approximate the 28 day trough plasma concentrations of passively-administered VRC01 and 3BNC117 in human trials [57, 74] and more closely approximate the range of plasma concentrations that resulted in transient reductions in plasma viremia in patients who received 3BNC117 [57]. We found that the best coverage with 2 bnAbs active at IC80 <50–100 μg/ml was 93–100% for 2, 3 and 4 bnAb combinations, and with 3 bnAbs active was 68–92.5% (Fig J and Table B in S1 Text). The overall most potent and broad 2, 3, and 4 bnAb combinations (Fig 4), also had best or close to best coverage with multiple bnAbs active (Fig 5). However, best-in-category combinations that included the exceptionally broad but less potent 10E8 showed superior coverage with multiple bnAbs active at less restrictive thresholds.
Neutralization curves for some bnAb/virus pairings can show incomplete neutralization of the genetically clonal virus population [65]. This suggests that a sub-population of virus is resistant to neutralization by the bnAb even at the highest concentrations tested. Given the importance of carbohydrates for many bnAb epitopes, post-translational glycan heterogeneity resulting from incomplete carbohydrate addition or modification may be an important contributing factor to such resistant sub-populations [68]. The inability to neutralize all variants would compromise the use of bnAbs for immunotherapy and may also impede the ability of bnAbs to protect against HIV acquisition. Hence, we investigated the extent of incomplete neutralization of clonal viruses by various bnAb combinations.
We first analyzed neutralization curves for single bnAbs and bnAb combinations that were experimentally measured in the study by Kong et al. [60]. We could accurately predict the combination maximum percent inhibition (MPI) using the Bliss independence model on single bnAb MPI values (Methods, Fig K in S1 Text, Pearson r = 0.9904, difference between observed and predicted MPI: median = 0.1%, 95% CI = 0–4.5%). Using this model, we then predicted the MPI values for the 2, 3 and 4 bnAb combinations composed of the best single bnAbs against the clade C panel. Experimental MPI values for single bnAbs are shown in Fig 6A (see S1 Text for discussion on different assay starting concentrations for panel bnAbs), and the predicted MPI values for 2, 3 and 4 bnAb combinations are shown in Fig 6B, 6C and 6D, respectively.
Incomplete neutralization was observed against several viruses for all single bnAbs and was frequent for the V2- and V3-glycan bnAbs CAP256-VRC26.25 and 10-1074V, (56% and 44% viruses with MPI < 95%, respectively). A lower frequency of incomplete neutralization was observed with VRC07-523 (11% viruses with MPI < 95%) and 10E8 (16.5%). Encouragingly, the fraction of resistant variants within a single virus preparation was predicted to decrease with increasing number of bnAbs in a combination, indicating that bnAbs tend to be complementary not only in terms of viral sensitivity at the population level, but in terms of the resistant subpopulations of post-translational Env variants. The 2 bnAb combination with the least fraction of viruses incompletely neutralized was VRC07-523 + 10E8 (2%), while VRC07-523 + CAP256-VRC26.25, which had one of the best potency and breadth profiles, had 4% viruses with MPI < 95%. Consistent with the high levels of incomplete neutralization seen with the V2- and V3-glycan bnAbs, a higher extent of incomplete neutralization was predicted for CAP256-VRC26.25 + 10-1074V, where MPI <95% was seen for 18% of viruses. Strikingly, the 3 bnAb combinations had MPI < 95% for only 0.5–1% viruses (n = 1–2 out of 200), and the 4 bnAb combination never had MPI < 95% for any virus. The analysis of experimentally measured MPI from the Kong et al. study also showed similar patterns (Fig L in S1 Text).
Studies of passive bnAbs in humans aim to achieve plasma concentrations that for periods of time exceed 25 μg/ml, a dose commonly tested in our neutralization assays [60]. We therefore experimentally tested the extent of incomplete neutralization at concentrations of up to 100–200 μg/ml against a subset of 24 viruses that were selected based on incomplete neutralization at the lower doses tested (Fig M in S1 Text). Most of these viruses were still incompletely neutralized at the highest concentrations tested (only 1 out of 24 showed 95% or higher neutralization). We then estimated the best-fit Hill curves using data points below 25 μg/ml (Methods, S1 Text) and used these to predict neutralization at the highest concentrations tested for each of these high-concentration assays. The predictions were quite accurate (average root mean square error = 6%, Kendall Tau p = 3.7 x 10−5, Fig N in S1 Text). Thus, using this approach, we predicted the MPI at 100 μg/ml for all best-in-class bnAbs (Fig N in S1 Text) and their combinations (Fig O in S1 Text) for all 200 clade C panel viruses. As expected, the fraction of viruses with predicted neutralization less than 95% at 100 μg/ml was reduced compared to the values at 25 μg/ml. Still, we found substantial levels of incomplete neutralization at 100 μg/ml and these results qualitatively recapitulated the above patterns of MPI at 25 μg/ml for single bnAbs and for bnAb combinations.
The metric instantaneous inhibitory potential (IIP) measures the log10 reduction in a single round of infection events in the presence of a drug. This metric correlates with clinical success of antiretroviral drug combinations, and can be used to characterize their efficacy [75]. Jilek et al. found that IIPave values (average IIP during the dosing interval, given drug pharmacokinetics) of 5–8 logs were necessary for successful antiretroviral therapy. Drug combinations in this range showed a reduction of viral load to <50 RNA copies/ml at 48 weeks in 70% or more of infected individuals. Applying their approach, we calculated the IIP values for the best-in-class single bnAbs and best bnAb combinations for the clade C panel.
IIP values for single bnAbs were calculated using either the best-fit Hill curves of experimental neutralization data for the best-in-class bnAbs (Fig 7, S1 Text), or estimated Hill curves using IC50 and IC80 values (Fig P in S1 Text) (with the former expected to yield more accurate predictions since IIP values are critically sensitive to neutralization close to 100%). Using BH model, we calculated the IIP values (Methods) for 2, 3 and 4 bnAb combinations of the best-in-class bnAbs (Fig 7). Since IIP values depend on bnAb concentration, and precise doses and pharmacokinetics of bnAbs are still being established, we analyzed IIP at bnAb concentrations of 1, 10 and 100 μg/ml. The 1 and 10 μg/ml concentrations are within the experimental assay range, whereas results for the 100 μg/ml dose are estimates obtained by extrapolation.
The best-in-class single bnAbs had median IIPs of 0.4–2.8 across viruses, depending on the bnAb and concentration, with CD4bs bnAb VRC07-523 giving the highest value, followed by V3-glycan bnAb 10-1074V (Fig 7, Fig P in S1 Text). The best-in-category bnAb combinations showed higher median IIP values of 1.2–5.0, 2.3–6.6, and 3.5–8.1 for 2, 3 and 4 bnAb combinations, respectively. The 2 bnAb combinations with highest IIP values consisted of VRC07-523 with either CAP256-VRC26.25 or 10-1074V, depending on the concentration. The 3 bnAb combinations with the highest IIP values were VRC07-523 + 10-1074V with either CAP256-VRC26.25 or 10E8, with the latter combination having a slightly better median IIP at 100 μg/ml (median IIP of 6.2 and 6.6, respectively).
Single bnAbs rarely had IIP > 5, the level found to be critical for clinical success of antiretroviral drug combinations [75], while 2, 3 and 4 bnAb combinations had IIP > 5 for 0–50%, 1.5–79%, and 15–92% of viruses, respectively, depending on concentration. The median IIP of the best 3 bnAb combinations exceeded 5 only at 100 μg/ml, while the best 4 bnAb combination had median IIP > 5 at a lower concentration threshold of 10 μg/ml. The range of median IIP values for the best 4 bnAb combination (3.5–8.1) is comparable to the average IIP for some of the currently prescribed antiretroviral triple-drug combinations (IIP ~ 3.5–12) [75].
We next systematically compared the best-in-category 2, 3, and 4 bnAb combinations to evaluate the benefit of having combinations with more total antibodies on overall performance using the metrics described above; namely the overall potency-breadth curves (Figs 3 and 4), the number of active bnAbs in the combination (Fig 5), the extent of incomplete neutralization (Fig 6), and IIP values (Fig 7). The relative impact of these metrics on clinical success is unknown and the relevance of each metric might differ for prevention versus treatment of HIV-1 infection, e.g. neutralization by multiple active bnAbs and IIP may be more relevant for latter. Working under the a priori hypothesis that an ideal combination should maximize performance using all four metrics, we chose VRC07-523 + CAP256-VRC26.25, VRC07-523 + CAP256-VRC26.25 + 10-1074V and VRC07-523 + CAP256-VRC26.25 + 10-1074V + 10E8 as the best 2, 3, and 4 bnAb combinations for comparison, respectively. These combinations showed best or near best performance using all four metrics when compared with other combinations with same number of bnAbs.
Using overall potency and breadth profiles, the best 3 and 4 bnAb combinations were significantly more potent than the best 2 bnAb combination, with a 2.6–3.1-fold more potent geometric mean IC80 (Fig 8A, p < 0.0014), and showed higher breadth of 97–99% versus 87% viruses neutralized at IC80 < 10 μg/ml, respectively. The best 3 and 4 bnAb combinations also demonstrated superior performance over the best 2 bnAb combination in limiting the extent of incomplete neutralization (Fig 8B). The fraction of viruses predicted to have < 95% neutralization at 10 μg/ml for 3 (1.5% viruses) and 4 bnAbs (0.5% viruses) was significantly lower than that for 2 bnAbs (8% viruses, p < 0.0036). Similarly, IIP for 3 and 4 bnAb combinations were significantly higher than the 2 bnAb combination (Fig 8C, p < 2.5 x 10−16), and showed significantly higher fraction of viruses above the clinically relevant threshold of 5 (p < 1.2 x 10−10, Fisher’s exact test). The best 3 and 4 bnAb combinations also showed significant improvement of coverage with at least 2 bnAbs active (Fig 8D and 8G, 28–42% improvement in coverage, p < 7.7 x 10−10). The main reason behind the poor coverage of viruses neutralized by 2 bnAbs in the best 2-bnAb combination was the limited breadth of CAP256-VRC26.25, which was included for its potency when positive (Fig 8E–8G).
Four bnAbs were predicted to be similar to 3 bnAbs by some metrics, and significantly better by others. The best 3 and 4 bnAb combinations showed nearly identical distributions of IC80 values (Fig 8A), and levels of incomplete neutralization (Fig 8B). In contrast, the best 4 bnAb combination showed significantly higher coverage than the best 3 bnAb combination for both neutralization by at least 2 active bnAbs (improvement in coverage 9.5% using activity threshold of IC80 < 10 μg/ml, p = 0.0001), and by at least 3 active bnAbs (improvement in coverage 47%, p = 1.9 x 10−21, Fisher’s exact test) (Fig 8D and 8G). Also potentially relevant for success in therapeutic settings, the best 4 bnAb combination showed significantly higher IIP scores (Fig 8C, p = 8.9 x 10−9) and significantly higher number of viruses with IIP > 5 than the best 3 bnAb combination (25% more viruses, p = 8.5 x 10−7). These results indicate that 4 bnAb combinations may be more effective in preventing viral escape compared to 3 bnAb combinations.
The exceptional breadth and potency of a new generation of bnAbs offers new clinical opportunities for the prevention and/or treatment of HIV-1 infection. Two CD4bs bnAbs, VRC01 and 3BNC117, have already initiated phase I clinical testing in infected subjects, and efficacy studies for the prevention of HIV-1 infection are planned [57, 74]. The most effective approaches will likely employ combinations of bnAbs targeting multiple epitopes on HIV-1 Env to maximize potency and coverage and to impede escape, which may be particularly important in the case of immunotherapy. Prevention of sexual transmission of HIV-1 may represent a relatively easier target for success, as bnAbs at mucosal surfaces at the time of exposure need only to block the infecting virus, while therapeutic approaches need to contend with high levels of replicating virus, complex within-host viral diversity, and established latent viral reservoirs. Given the large number of bnAbs now available against multiple epitope regions of HIV-1 Env, it is of great interest to have experimental measures and predictive models that can be used for evaluating and selecting optimal combinations of bnAbs for clinical development for the prevention and/or treatment of HIV-1 infection.
Among the bnAbs tested here, the best-in-class single bnAbs for potency and breadth against our panel of 200 clade C viruses were CAP256-VRC26.25 (V2-glycan), 10-1074V (V3-glycan) and VRC07-523 (CD4bs) (Fig 1). While 10E8 was the only MPER-directed bnAb tested, it was previously shown to be the most broadly reactive and potent of the known MPER bnAbs against other virus panels [15]. To evaluate various combinations of bnAbs we developed a new model, the Bliss Hill (BH) model, and found it to more accurately predict the breadth and potency of antibody combinations than the additive model (Fig 2, Fig C in S1 Text). We applied the BH model to predict neutralization profiles of over 1,600 possible 2, 3, and 4 bnAb combinations against the 200 clade C viruses using experimental data from the testing of single bnAbs alone (Figs 3 and 4). These predictions allowed us to identify and compare best-in-category bnAb combinations. The overall potency and breadth of neutralizing activity significantly improved as the total number of bnAbs in the combination was increased from 2 to 3, but not from 3 to 4 (Figs 4 and 8). Two best 2 bnAb combinations were identified that demonstrate superior performance in overall potency and breadth. While CAP256-VRC26.25 + 10-1074V was slightly more potent than CAP256-VRC26.25 + VRC07-523, the latter combination exhibited better breadth, and thus may be preferred. The best 3 bnAb combination (CAP256-VRC26.25 + VRC07-523 + 10-1074V) benefitted from combining the complementary potent profiles of CAP256-VRC26.25 and 10-1074V, with the added potency and breadth of VRC07-523. The best 4 bnAb combinations were significantly better than the best 2 but not 3 bnAb combinations. Together, these results demonstrate the substantial benefit bnAb combinations afford when selected to complement and optimize target epitopes, potency, and breadth of coverage. These parameters will be important to consider when selecting bnAb combinations for both prevention and immunotherapy of HIV-1 clade C infection.
We note that 8 of the 15 bnAbs tested here did not show up as a component of best combinations. In most cases these bnAbs exhibited weaker potency and breadth of neutralization than bnAbs in the corresponding epitope class that did show up (Fig 1A). An exception is VRC07, which had a better profile than 3BNC117, yet 3BNC117 and not VRC07 showed up as a component of best combinations. Another exception is PGT121, which was marginally better than PGT128, yet PGT128 and not PGT121 showed up in best combinations. In both of these cases the bnAb in best combinations (3BNC117 and PGT128) had slightly greater potency against sensitive viruses (Table A in S1 Text).
Our analyses further highlight that bnAb combinations, especially those to be used for treating established HIV-1 infection, can be selected to increase the probability of having at least two antibodies in the mixture active against a patient’s virus. While having an increased number of active bnAbs in a combination is desirable, our results illustrate the sobering limitations with even the best bnAbs currently available (Figs 5 and 8). For IC80 thresholds of 1–10 μg/ml, the percentage of clade C viruses neutralized was reduced to 44–95.5% when requiring a minimum of 2 bnAbs in the combination to be active. This coverage substantially increased when IC80 thresholds of 50 μg/ml or higher were considered (Fig J in S1 Text). Therefore maintaining high in vivo antibody concentrations, in plasma and especially in infected tissues, may be key in therapeutic settings, and thus the tissue distribution and in vivo pharmacokinetics of individual bnAbs will be critical factors. The coverage of viruses by active antibodies naturally increased with the total number of bnAbs included in a combination, yet even for the best 4 bnAb combination, only 73.5%, 26.5%, and 2.5% of viruses would have either 2, 3, or all 4 antibodies active at a threshold IC80 titer of < 1.0 μg/ml, respectively. From these analyses, it becomes apparent that inclusion of a bnAb with better overall breadth (such as 10E8, Figs 1 and 5, Fig J in S1 Text) in a combination may be more advantageous than choosing the most complementary bnAbs with the highest potency. By further analogy to antiretroviral therapy, it is possible that at least 3 agents simultaneously active against the virus will be critical to avoid escape. For the prevention of HIV-1 infection, it may not be quite as critical to have multiple antibodies simultaneously active, as bnAbs at mucosal surfaces need only to block the transmitting virus at the time of exposure. Nonetheless, combinations of at least 2 or 3 bnAbs may provide an advantage for breadth and potency in preventing infection, and should enhance coverage against viral quasispecies from a chronically infected donor.
We also considered the impact of bnAb combinations on limiting the extent of incomplete neutralization of HIV-1 Env pseudoviruses. Combinations with a higher number of bnAbs, in addition to improving breadth and potency across different viruses, also improved the capacity to completely neutralize the expressed forms of an Env within a genetically clonal virus population (Fig 6, Figs L and O in S1 Text). The experimental data suggests that the resistant sub-populations of virions for different bnAbs do not overlap substantially. This complementarity reduces the extent of incomplete neutralization shown by combinations with higher number of bnAbs, an important consideration when selecting optimal bnAb combinations for both prevention and treatment of HIV-1 infection. It should be noted that the pseudoviruses utilized in our study were produced in 293T cell lines, and thus may differ in glycan heterogeneity and susceptibility to incomplete neutralization compared to viruses derived from primary PBMC. However, a recent study comparing clonal viruses grown in either 293T or human PBMC found overall similar trends in levels of incomplete neutralization for individual bnAbs [65]. These data suggest that the complementarity of bnAbs to limit incomplete neutralization will likely prove to be effective for primary PBMC grown viruses as well.
The slopes of in vitro neutralization curves for individual bnAbs have been shown to exhibit inherent variability, with bnAbs exhibiting slopes >1.0 predicted to have greater in vivo efficacy than classes of bnAbs having slopes ≤1.0 [71]. The metric instantaneous inhibitory potential (IIP), which measures the Log10 reduction in infectious events in the presence of drugs/antibodies, is positively correlated with neutralization curve slopes, in that bnAbs with higher slopes are predicted to have IIP values that increase faster with concentration [76]. Here we calculated IIP values for best-in-category bnAb combinations as an opportunity to quantitatively compare their efficacy based on what is seen with antiretroviral drug combinations [75]. Such a comparison between bnAbs and standard antiretroviral drugs comes with several caveats. First, Env is much more variable than the targets of most antiretroviral drugs, making it essential to measure bnAb activity against a large panel of virus variants, whereas IIP values in the Jilek et al. study were calculated for a single virus. Second, because bnAbs can engage in Fc receptor-mediated effector functions [77, 78], the overall in vivo efficacy of bnAb combinations might be greater than the neutralization measured in vitro. Third, since IIP values depend on the concentration of drug, tissue-wide heterogeneity and pharmacokinetic profiles of bnAbs will be needed for accurate prediction. With these caveats in mind, we found that IIP values for the best 3 and 4 bnAb combinations compare favorably with those of several available antiretroviral drug combinations, for which an IIP threshold of 5–8 was found to correlate with clinical success [75]. While single bnAbs and 2 bnAb combinations had IIP < 5 for most viruses, we found that the best 3 and 4 bnAb combinations had median IIP values > 5 at concentration thresholds of 100 μg/ml and 10 μg/ml, respectively (Fig 7). Thus, using the Jilek et al. criterion, the 3 and 4 bnAb combinations could lead to favorable clinical outcomes, while single and 2 bnAb combinations are less likely to succeed.
It must be emphasized that the results from our analyses do not imply that other bnAb candidates should not be further considered for inclusion in combinations for clinical testing. In fact V3-glycan bnAbs 10–1074 and PGT121 have either started or will soon initiate phase I clinical testing, respectively. Our results do, however, suggest favorable bnAb combinations for future studies, and provide a reasoned way to narrow the otherwise vast array of possible bnAb combinations. We provide modeling strategies that enable quantitative assessment of the neutralization patterns of combinations of bnAbs using several metrics, to better inform selection for clinical use. Yet these in vitro measures and modeling results are just a few of the parameters that must be considered when selecting optimal bnAb candidates. The stability, manufacturability, and in vivo pharmacokinetics, tissue distribution, and safety profiles are just a few additional key parameters that must also be evaluated when moving bnAb candidates forward in the clinical pipeline.
Our study focused on HIV-1 clade C viruses as the predominant subtype in sub-Saharan Africa where bnAb clinical efficacy studies will likely be conducted, and is a dominant subtype globally. Some bnAb combinations may be more effective against other genetic subtypes, as bnAbs can exhibit variable levels of neutralization breadth among different clades of virus (e.g. many V3-glycan antibodies exhibit more limited breadth against CRF01_AE viruses, and CAP256-VRC26.25 has limited breadth against clade B viruses) [17, 32]. Extensive data sets are available from the testing of individual bnAbs against large standardized panels of viruses from multiple subtypes, and the BH-model presented here may be utilized to thoroughly investigate the question of how viral clade impacts optimal bnAb combinations. We are developing a web-tool, CombiNaber, which will available on the Los Alamos HIV Immunology Database (http://www.hiv.lanl.gov/content/sequence/COMBINABER/combinaber.html). This tool will predict bnAb combination neutralization results from single bnAb neutralization data using either BH or additive models and perform systematic analysis to provide the user with the best candidate combinations for their panel (S1 Text).
In summary, we have assessed optimal bnAb combinations predicted to have greatest success in the prevention and treatment of infection by HIV-1 clade C, taking into account multiple metrics. In addition to evaluating overall potency and breadth, we have also taken into account the number of active bnAbs within a given combination, the impact of combinations in limiting the extent of incomplete neutralization, and to calculate the IIP of bnAb combinations. These latter metrics may be of critical importance when considering the use of bnAbs for the treatment of HIV-1 infection, as they directly relate to confronting the ability of virus to escape from selective immune pressure. Our results indicate that for both the prevention and treatment of HIV-1 infection, combinations with higher numbers of bnAbs are advantageous in providing increased potency, breadth, complete neutralization, and active coverage. Given the tremendous resources required to take each single bnAb forward into clinical testing, our results outline important parameters that can inform the selection of bnAbs with the best indicators of success for clinical development, and stresses the importance of considering the behavior of bnAb combinations early in planning stages.
This was a non-randomized laboratory study designed to investigate the breadth and potency of HIV-1 bnAbs against a panel of 200 clade C HIV-1 Env pseudoviruses, and to develop mathematical models to predict combinations of 2, 3, or 4 bnAbs that would exhibit enhanced breadth, potency, extent of complete neutralization, and IIP relative to single bnAbs. Fifteen recently described bnAbs targeting four distinct epitopes on HIV-1 Env were each tested against the panel pseudoviruses in vitro to determine IC50 and IC80 titers and MPI. All neutralization assays were performed in duplicate and without blinding.
Neutralizing antibody titers of bnAbs were determined using a luciferase-based assay in TZM.bl cells (NIH AIDS Research and Reference Reagent Program) as previously described [79, 80]. Unless stated otherwise, starting concentrations of individual bnAbs ranged from 10–50 μg/ml depending on the available supply at the time of testing. BnAbs were serially diluted seven times using 5-fold titration series. The concentration range tested for each bnAb is indicated in Table A in S1 Text. All assays were performed in a laboratory meeting GCLP standards.
A panel of 200 clade C HIV-1 Env pseudoviruses was utilized to assess the potency and breadth of bnAb neutralization activity. Functional Env clones were derived from individuals in acute/early stages of infection from South Africa (65%), Tanzania (14%), Malawi (11.5%), Zambia (6.5%), and Botswana (3%) collected over 12 years (1998–2010). All Envs were from heterosexual transmissions except for a single case of breastfeeding transmission. The majority of Envs exhibit a Tier 2 neutralization phenotype (75%, n = 150), with 1% classified as Tier 1A, 8.5% classified as Tier 1B, and 15.5% classified as Tier 3 [81]. Pseudovirus stocks were generated via transfection in 293T/17 cells (ATCC, Manassas, VA) and titrated using TZM.bl cells as previously described [82].
A panel of 15 particularly broad and potent human monoclonal antibodies was selected based on prior data from testing against large multiclade panels of HIV-1 pseudoviruses. In some cases we included somatic variants or newly engineered variants that exhibited enhanced activity over parental wildtype bnAbs (i.e. 10-1074V, VRC07-523, CAP256-VRC26.25). Importantly, we included bnAbs that are currently in human clinical trials (VRC01, 3BNC117, 10–1074) or are advanced candidates for clinical testing (PGT121, 10E8, PGDM1400, CAP256-VRC26.25). Antibodies were generated in the laboratories of D. Burton at The Scripps Research Institute (PGT145, PGMD1400, PG9, PGT121, PGT128), M. Nussenzweig at The Rockefeller University (10–1074, 10-1074V, 3BNC117), or the NIH Vaccine Research Center (CAP256-VRC26.08, CAP256-VRC26.25, VRC01, VRC07, VRC07-523, VRC13, 10E8). VRC01 and VRC07 are CD4bs bnAbs of the same lineage [69]. VRC07-523 is an engineered clonal variant of VRC07 with increased potency and breadth [23]. Of note, VRC07-523 was made with a two amino acid mutation in the Fc domain (M428L/N424S) to increase affinity for the FcRn and therefore increase circulating in vivo half-life [83]; these mutations do not affect antibody-mediated neutralization. VRC13 is a CD4bs antibody that is distinct from the VRC01-class of antibodies in that it contacts gp120 primarily via CDR binding loops [70]. 10–1074 and PGT121 are clonal variants from the same donor [17]. 10-1074V is a variant of parental 10–1074 in which six complex-type glycan-contacting residues in IgH have been substituted with those from bnAb PGT121.
For theoretical derivations of models, see S1 Text. The additive model [60] predicts combination IC80 as
IC80comb=1/(1/IC80A+1/IC80B+…)
, where
IC80A,IC80B,…
are the single bnAb scores. The equation for combination IC50 is similar using single bnAb IC50.
The Bliss-Hill model involves estimating single bnAb neutralization curves using Hill functions, f(c) = cm/(km + cm), where c = bnAb concentration, k = IC50, and m = log(4)/[log(IC80)–log(IC50)]. The combination neutralization, using the Bliss Independence model, is f = 1 − (1 − fA)(1 − fB)(1 − fC) … where fA(c),fB(c),fC(c), … are the single bnAb neutralization functions and c is the bnAb concentration. This equation is solved for the combination IC50/IC80 (we use Brent algorithm[84, 85] implemented in Scipy [86]). Treatment of single bnAb IC50/IC80 values above or below experimental thresholds is detailed in the S1 Text.
For combinations with multiple bnAbs targeting the same epitope, the combined neutralization function of such bnAbs is calculated using
fA(c)=(gA1(c)+gA2(c)+…)/(1+gA1(c)+gA2(c)+…)
, where
gAi(c)=fAi(c)/(1−fAi(c))
and
fAi(c)
are Hill curves for each of the bnAbs A1, A2, … The Bliss independence model equation is used with the neutralization functions fA(c),fB(c),fC(c), … for each epitope to get neutralization by the combination.
Given the experimental or predicted MPI values for single bnAbs at a given concentration, fA,fB,fC,…, the combination MPI value was predicted as f = 1 − (1 − fA)(1 − fB)(1 − fC) …
IIP is defined as IIP = −Log10(1 − f(c)), where f(c) is the neutralization by a single bnAb or bnAb combination at concentration c. The Hill functions for neutralization by single bnAbs were calculated from IC50 and IC80 values, or by fitting experimental neutralization curves (S1 Text). For IIP of combinations, best-fit single bnAb neutralization functions together with BH model were used.
All statistical analyses were performed using the Stats module in Scipy [86]. Non-parametric tests were preferred and two sided p-values are reported. False discovery rates (q-values) were calculated by using qvalue package for Python (https://github.com/nfusi/qvalue), based on the calculations outlined in reference [87].
|
10.1371/journal.pntd.0006299 | Why snakebite patients in Myanmar seek traditional healers despite availability of biomedical care at hospitals? Community perspectives on reasons | Snakebite is a major public health problem in many developing countries. Farmers are particularly exposed to snakes, and due to their rural location often experience delays in accessing formal healthcare. The reasons to use traditional healers may include difficulties in accessing formal healthcare, certain beliefs about snakes and snake venom, tradition, and trust in the capacity of traditional healers. Traditional healing, however, may have serious consequences in terms of delays or added complications. There is little in-depth current information about the reasons for its continued use for snakebite. As part of a health services development project to improve health outcomes for snakebite patients, community attitudes to the use of traditional healers were explored in the Mandalay region of Myanmar.
With the objective of learning from local communities, information was generated in three communities using participatory appraisal methods with the communities, and focus group discussions with the local healthcare staff. Many snakebite victims in these communities use traditional healing. Reasons include transport difficulties, low cost for traditional healing, inadequacy of anti-snake venom in the formal healthcare sector, and traditional beliefs, as traditional healing practices are rooted in many cultural and traditional factors. The communities reported that even if access to medical care were improved, traditional healing would continue to be used.
These findings point to the need for working with traditional healers for prevention, appropriate first aid and timely access to effective treatment for snakebite.
| Snakebite is a major health issue that affects many people, particularly young poor farmers in developing countries in the tropics. Many patients suffer poor outcomes due to inadequate or delayed access to effective treatment. A large number still use traditional healers. Often patients visit traditional healers before accessing formal health care; thus incurring delays in receiving antivenom (AV) needed to treat envenoming. In other cases, traditional healing methods may themselves cause complications. In Myanmar, while most patients now access formal medical care, many also use traditional healers. We consulted communities in three rural areas in the Mandalay region and found that the reasons for using traditional healers include difficulties with transportation, cost, inadequacy of AV in the formal health sector, and trust in traditional healing within the context of longstanding tradition. These findings point to the need for working with the traditional healers as they could be effective agents to encourage prompt use of formal healthcare.
| Snakebite is a neglected tropical disease, affecting disempowered rural communities in developing countries. It has been difficult to identify the exact incidence due to inadequate health statistics and the fact that some patients do not seek medical care. In 2008, global annual incidence was estimated as 1.8 million bites [1], whereas in 1998 Chippaux estimated the incidence as high as 5.4 million bites [2]. The global mortality from snakebite likely exceeds 125,000 deaths annually. However, a comprehensive community survey indicated that in India alone, the annual mortality from snakebite exceeded 45,000 [3]. Considering that many patients may not use health services and die before accessing care, the actual number may be much higher [1, 3, 4].
The burden of snakebite is highest in rural areas of the tropics and subtropics of South/ Southeast Asia and sub-Saharan Africa, mainly due to the density and species of venomous snakes present, population density, agriculture base, inadequate public health programs and lack of mechanised farming practices [1, 5, 6].
The only specific treatment for snakebite envenoming is antivenom (AV; “anti-snake-venom”, “ASV”) [7]. However, with less than adequate health literacy, inadequate access to AV and treatment facilities and other reasons, traditional healing continues to be used by a large number of people. Traditional medicine and healing are based on the communities’ past experiences and observations, passed on through generations verbally or in writing [8], and is defined as “the sum total of the knowledge, skill, and practices based on the theories, beliefs, and experiences indigenous to different cultures, whether explicable or not, used in the maintenance of health as well as in the prevention, diagnosis, improvement or treatment of physical and mental illness”[9]. Traditional medicine is used, whether alone or in conjunction with biomedicine (medical care and system based on the principles of Western science), by many people in both developing and developed countries; 80% of people in Africa reported to be users, 40% of all health care in China is reported as traditional care, and 38% to 75% of people in developed countries such as Australia, Canada and France are said to access complementary and alternative medicine [10].
Globally, traditional methods such as tattooing and herbal remedies and other methods including electric shock and suction are still used for snakebite [7, 11, 12]. The reasons listed in the literature for the continuing use of traditional healing include affordability, availability, and cultural familiarity [13, 14]. Unfortunately, a significant number of people continue to die after snakebite. This is often due to severe envenoming, made worse in many cases by delays in obtaining effective medical care. These factors may generate a misperception that formal biomedicine (also known as Western or allopathic medicine) is ineffective [15, 16]. A proportion of snakebites by venomous species are ‘dry bites’, where the bite fails to inject enough venom to cause perceptible clinical effects. Further, many snake species are either non-venomous, or minimally venomous and so unable to cause envenoming. Should the patient seek help from a traditional healer after such a dry bite or non-venomous bite, the patient and the community are likely to mistakenly attribute recovery to the use of traditional medicine.
Snakebite incidence is historically high in Myanmar (15.4/100,000/yr) [17]. 70% of Myanmar’s population resides in rural areas with heavy reliance on subsistence agriculture [18]. Agriculture is a major occupational risk exposing farmers to snakebites. Most venomous bites in Myanmar are attributed to Russell’s Viper, envenoming by which can cause local pain and swelling, coagulopathy, life-threatening hemorrhage, shock, and acute renal failure. Overall, the annual number of snakebite cases, as reported in the national data of snakebite victims who seek care at the government hospitals or health centres fluctuates between 15,000 and 20,000 [19]. A large proportion of these snakebites occur in six high incidence regions i.e. Mandalay, Sagaing, Bago, Magwe, Ayeyarwady and Yangon. According to these health services data, in 2016 there were 16,767 snakebites in Myanmar, out of which 2,566 bites were in Mandalay region. These numbers are probably an underestimation of the magnitude of this important public health issue, as they do not include those victims who use traditional healers only and those who die before seeking care at the government health care centres or hospitals. Myanmar Australia Snakebite project for improved health outcomes for snakebite patients worked closely with the main tertiary hospital in Mandalay region. The hospital admission records and clinical information informed that 965 snakebite victims were admitted in 2016; 68.5% of these 965 suffered from coagulopathy, 63.2% suffered acute kidney injury, 31.5% required dialysis and 12.4% died. These figures point to the significance of this neglected public health issue.
Snakebite treatment according to the biomedical management protocol includes AV, which is essential and the only antidote for envenoming, and supportive treatment such as airway management, treatment of hypotension and shock, treatment of acute kidney injury, management of hemostatic shock and treatment of the bitten site with antibiotics if needed [20]. In Myanmar, government doctors and paramedical staff are trained to provide treatment with AV and supportive treatment.
In Myanmar, treatment of snakebite is impaired by problems with the supply of AV, and by shortage of adequately trained staff, particularly in rural areas. These limitations may contribute to the persisting use of traditional methods. Use of the healthcare system in Myanmar is dependent on several factors, including cost, previous experience, fear of surgery, and belief in religious or spiritual healers. Additionally, in some cases AV causes adverse reactions [7, 21]. Hence, the communities may also harbour fears of biomedical treatment.
Many traditional healing methods, such as local incision, herb ingestion, application of snake stones, and tattooing, are ineffective, and in some cases, harmful [7, 10]. Their use can cause infection, bleeding, gangrene and other problems. In this way, the use of traditional healing may further delay or complicate necessary biomedical treatment.
With the continued use of traditional healing practices, it is important to develop a better understanding of the nature of healing practices, the communities’ reasons for and views about its use, and the interface between traditional and biomedical components of the health system. In many snakebite-affected countries, an envenomed victim may need to walk (or be carried) for many miles to reach a primary health post. Gutiérrez and colleagues assert that ‘studies of the circumstances that delay the access of people bitten by a snake to health centres are of great value …… [and that the studies] should include in-depth analyses of the cultural characteristics of the communities, the way snakes and snakebites are perceived, the cultural background of local healers…..’ [22].
As part of a larger community and health services development project in Myanmar, the aim of this participatory action research was to engage with rural communities to learn from their perspectives, their health knowledge, and reasons for healthcare-seeking practices.
Using participatory methods, traditional healing for snakebite was studied through the lens of community knowledge, experiences and traditions. Gutiérrez et al. note that modern health programmes in rural communities are often culturally biased and paternalistic, lacking participation of the community in question [23]. Participatory methods acknowledge that local communities have valuable stores of knowledge which can guide development [24, 25]. Learning from communities through participatory approaches is even more important for issues which affect impoverished people. Snakebite is a problem that mainly affects impoverished rural people, and its neglect at the global level is largely due to the fact that the affected populations lack political voice [26].
Participatory rural appraisal sessions (PRA) were organised in three communities in villages in Kyaukse and Madaya townships of Mandalay Division. They included the creation of ‘problem walls’ to reflect what the community saw as problems they faced. Focus Group Discussions with three groups of health care providers were conducted in the same settings. The primary care workers, which included Health Assistants, public health staff and midwives, are responsible for basic curative care at community health centres. They provide vaccinations, outreach, public health, preventative and health promotional activities in community and home settings.
The three communities and the health care providers in those areas were selected considering representation of various areas of the township, distance to services in the city, access to care, and logistics and feasibility. Three participatory appraisal sessions took place, in 2016, in public local community meeting places, ensuring an appropriate environment that fostered maximum comfort and interaction between the participants. 135 participants took part in PRA sessions that were held between 10 am and 2 pm.
The majority of participants attended for the whole session, whilst others joined late or left early due to obligations such as work and family. This flexibility is part of the participatory and empowerment process, and contributions were respected and considered even if community members were unable to participate for the whole duration of the session. The communities were approached through primary health care workers, who, after permission from village leaders, invited individuals to participate. No community members or primary health care workers refused to take part and no one withdrew from the research. Table 1 informs about the selection process.
After introducing the learning aims of the sessions, community members were encouraged to decide among themselves the most important aspects of the snakebite problem and issues with health care for snakebite in their own and surrounding communities. Throughout this process, participants were also encouraged to share their personal and family experiences. Community members discussed, defined and wrote key issues and problems on the flip charts in the shape of bricks (problems). They then discussed the solutions that would be needed to ‘break the wall down’.
In the same villages where the PRA took place, FGDs were conducted at the government rural health centres with 23 primary care workers, with a focus on the extent of snakebite problem in the area, use of traditional and biomedicine health care by the locals and the reasons for such use. These health staff had experiences of snakebite either personally or through their work.
In addition to the communities noting their issues on flip charts, a scribe took notes of the discussions. Each of the participatory appraisals and FGDs yielded several pages of raw narrative data. The data were then analysed to identify themes. The thematic analysis consisted of 6 phases [27], the first step of which was familiarisation with the data through reviews of the detailed notes taken at PRA and FGDs as well as the visual data such as problem walls. The next stages in this analysis involved generating initial codes and interpretative analysis of these codes, leading to searching for themes i.e. important patterns and concepts relevant to the research question about healing methods and the reasons for their use. Themes were then reviewed and refined by carefully considering relevance to the main research questions, whether identified themes were backed by sufficient data, and whether there was clear distinction between the identified themes. The themes were then defined and reported for discussion. This analysis focused on patterns of use, type of traditional methods and the reasons rather than on prevalence. Therefore, such thematic analysis did not include counts or statistical analysis.
The research was conducted with ethics approval from the Human Research Ethics Committee at the University of Adelaide and Ethics Committee at the Department of Medical Research at the Ministry of Health in Myanmar.
Kyaukse and Madaya townships are farming communities. All community members who participated in the discussions and appraisals acknowledged that snakebite was a problem in their communities. Working on farms and or walking to or from farms were the activities associated with snakebite, particularly early in the morning and in the evening. People informed that the harvest times in this region, June-July and October-November, were associated with higher incidence of snakebite. The community members considered snakebite a major health issue, and emphasised the need for further inputs for prevention and improved curative care. They informed about the inadequate health awareness and difficulties in accessing transport. They had good knowledge about preventive methods, particularly the need to wear boots when working in the fields and to have a torch while working at night. However, they informed that many do not practice these preventive methods for cost and convenience reasons. For example, it was mentioned that the boots are costly, hot to work in, and get stuck in the mud causing the work to slow down. The solutions by the community members included the need for further health education, access to less costly or subsidised appropriate boots, adequate supplies of AV at health centres closer to their villages, and better access to transport. In fact, groups of local volunteers across many communities facilitate transport for patients from their villages. For example, one of the communities where a PRA session was held had a community car that the locals use to transfer patients to hospitals. However, that car was in need of repair when PRA took place.
Healthcare providers considered that with increased use of mechanised farming, the snakebite problem was decreasing. Most bites were by Russell’s vipers, and a few by cobras and a variety of other species. Healthcare providers informed us that access to health services had improved in the last few years as a result of better transport.
Firsthand accounts by those who had been bitten by snakes informed us that the use of traditional healing in these communities was either as a stand-alone treatment without using biomedical care, or in conjunction with biomedical care. Those who used traditional healing in conjunction with biomedicine did so before or after the biomedical care. Seven of the PRA participants shared some information about their experience and use of health services and traditional healing. This information is summarised in Table 2.
Two victims informed us that they had visited a traditional healer after presenting to a hospital. Another two said they had been to a hospital but did not use the services of any traditional healer, and the other three said that they had been to a traditional healer but not to a hospital or rural health centre. One woman reported being bitten on her finger while she was picking betel leaves when she was 18 or 19 years old. She went straight to the monk for traditional treatments in the form of herbal medicine and tattooing. She said that she was fully healed after a month of that treatment. One man working in his turmeric plantation at dusk was bitten by a snake and fainted. He was driven to Kyaukse hospital where he was admitted, received AV and discharged after two months. Details of his hospital treatment were not discussed as the focus of discussions was traditional healing. After discharge from the hospital he went to see the monk for further treatment.
With regards to types of traditional healing, the community members did not distinguish between traditional healing as a practice and traditional healers as the practitioners. Traditional healing was seen both as a profession and a tradition. However, in reference to the methods being used, community members made a clear distinction between monks and other traditional healers as two separate types of traditional healer. This distinction appeared to be due to the spiritual healing aspect of the care provided by the monks.
A range of methods of healing were used by both monks and other traditional healers. Both practiced physical and spiritual methods. The physical techniques included asking the patient to chew the root of a particular plant to diagnose what type of snake had bitten the patient. The diagnosis was based on taste; whether it was bitter or not. Other practices included making incisions with a razor blade, tattooing with either ink or herbal medicines, use of a syringe to suck out venom, and rolling a heated glass bottle on the bite site to draw snake venom out. Tying a rope or a piece of cloth above the bite site as a tourniquet is practiced by many snakebite victims and the community as a first aid method, noted through observation of patients at the Project’s site hospitals. Communities did not discuss the tourniquet at the PRA sessions as a traditional method; probably because of the fact that health care providers at health centres and hospitals now advise communities not to use a tourniquet (a recommended first aid method for snakebite about a decade ago).
Faith-based spiritual techniques used by the monks and by the other healers included use of holy water, chants, prayers and astrology. Community members mentioned that “herbal concoctions” are used by both monks and other traditional healers. The following is a description, as reported by a community member, of one of the methods used by a monk:
“Using a razor blade, the monk makes 10 parallel surface cuts around the wound. He then takes a 20cc plastic syringe and cuts off the top, placing it on the bite site. Using a second syringe and a thin tube he draws out the poison, which can be seen being removed in thick clots (Fig 1). After the poison is removed, blood starts to come out of the syringe. If the blood is not clotting, the monk knows [that he needs] to refer patients to hospital. The monk uses a bowl of bottled water to flush out the syringe throughout the procedure, and has the patient consume some traditional medicine [recipe unknown to the community members] to increase urine output. If the patient urinates 3 times [after the treatment], they are said to be cured”.
How the traditions are passed on, and the methods used, was elaborated on by a 58 year old local traditional healer who lives in the village and participated in the participatory discussions and analysis:
The staff who participated in FGDs were also aware of these treatment types, including tattooing, incision, and using a glass bottle to remove venom.
The quality of traditional healing was perceived to be good and was trusted by the participants. The community members reported that visiting a monk or other traditional healer was common and most of the snakebite victims in these communities access such services at some point following the bite. Many participants were of the view that even if the local health centres were adequately stocked with AV and if the treatment was available close to the village(s), people would still choose to see a traditional healer in addition to using biomedical care. In contrast to the community members’ views, the healthcare providers said that if there was enough medicine available at the centres, most people would report straight there and not to a traditional practitioner, despite strong community belief in traditional healing for snakebite.
According to community members, one reason for trust in and use of traditional healers is that deaths from snakebite after treatment by traditional healers are rarely seen by the communities. They reported success rates of up to 95% among those who use the local traditional treatments. Communities also held the monks in high regard due to the spiritual and community service aspects of the healing, and lacked access to formal biomedical care because of high cost, distance from health care facilities, and difficulties with transport. For community members, the cost of care was a major factor. Most pointed out that the cost of hospital treatment for snakebite, including car hire, food for carers, accommodation for carers and some out of pocket medicine was around 300,000 Kyat (about 220USD). In contrast, treatment from a monk or traditional healer was often free, for a voluntary donation, or minimal fee. One person informed that they paid approximately 30,000 Kyat (22USD) for their visit to the monk. One village had no car and for any emergency the villagers needed to rent a car at a high cost. In the two villages that did have community cars, one had broken down and the community lacked the funds to repair it.
Another reason community members gave for using traditional healing first was the desire to avoid visiting hospitals, with some speaking unfavourably of the treatment in hospital settings. They cited factors such as unkind treatment and being afraid of the staff. In contrast, others said that they liked the care provided at the hospital. In fact, one of the community members bitten was very satisfied with the treatment he received at hospital, and as a result, did not seek any traditional treatment.
The misconception that snakebite victims cannot be treated at a hospital or health centre without bringing the snake involved with them was voiced by two of the community members who had been bitten, one of whom had tried to catch the snake after being bitten. It is true that accurate identification of the offending snake can be useful, particularly if the snake is a non-venomous or minimally venomous species, thus allowing the snake-bitten person to be reassured and discharged from treatment in most cases. At some FGD sessions, staff noted that if the snake was clearly a non-venomous or minimally venomous species, they could then discharge the patient without need for referral on to a larger health facility. These staff appeared confident in their ability to identify such non-risk snakes.
The healthcare staff mentioned transportation, low cost of traditional healing and strong traditional beliefs as reasons for the continued use of traditional healers. Some staff emphasised that traditional beliefs, not cost, were the major reason. The healthcare staff themselves did not appear to believe in traditional healing, and stressed that they had no contact with the traditional healers or monks about the treatment of snakebite patients. The healthcare staff also considered inadequate staffing and AV supplies at the health centres as contributing to the use of traditional healing. At one of the sessions, the staff said that the local monk only treated people because he wanted what was best for the community and that he knew that the local rural health centres were not adequately stocked with AV. Adding to this, some staff suggested that monks and other traditional healers could possibly be incorporated into the modern healthcare system.
Traditional beliefs, proximity to care, low cost, and perceived or actual inadequacies of the formal biomedical healthcare system emerged as the main factors associated with the local communities’ use of traditional healers. Spirituality also plays a key role in influencing decisions about treatment, and services by monks are seen as associating care with spirituality and community service. These beliefs appear to have a major influence on the use of traditional service despite awareness among community members that treatment is available in the formal biomedical sector, that the formal care is heavily subsidised and that many patients indeed have used formal biomedical care. The trust in and use of traditional care for snakebite is continuing despite expansion and better access to and use of biomedical care [28]. In fact, around the globe, the overall use of traditional and alternate medicine has increased in the last decade within the context of escalating costs of care and an increasing emphasis on patient-centred care [29].
Traditional healing plays a positive role in terms of social connectedness, harmony, services and support for fellow community members. However, there exists a knowledge gap which needs to be addressed to facilitate better health and wellbeing. Results of this participatory assessment indicated that one of the reasons communities trust the capacity of traditional healers is a perceived better success rate of the treatment in the traditional sector. In fact, a significant number of snake bites are either dry bites by venomous snake species with no systemic envenoming, or bites by one of the many species of non-venomous or minimally venomous snakes. Many who visit traditional healers as their first point of care therefore may view their traditional treatment as having been successful after suffering only from a dry bite or non-venomous bite. Additionally, as traditional healers often refer patients to hospital only when the symptoms are severe, hospitals end up seeing more patients with clinical complications, many of whom deteriorate and die [30]. These factors combine to generate a misperception about success of traditional healing and ineffectiveness of services in the hospitals. This suggests a need for health education for communities to bring about informed choice. Another misconception that community health education programs should address is that of the need to bring the dead snake to the hospital for identification, in order for diagnosis and treatment of the patient. Efforts by the patient or community members to kill and bring the snake could lead to further harm, either by the snake, or through further delays in necessary AV treatment. However, as noted earlier, identification of the snake can be beneficial in some circumstances. The increasing availability of mobile phones with inbuilt cameras, even in rural communities, might provide a future avenue for snake identification without a need to capture/kill the snake.
Some felt that the traditional health practitioners simply want what is best for their patients, and would happily forego their practice if formal services were more available. In fact, they informed that healers and monks do already refer severe snakebite cases to formal biomedical services. At the same time, we received the impression that some monks and traditional healers relied on snakebite treatment as a form of income, which would mean that foregoing treatment of snakebite victims could affect their livelihood. One strategy could be to integrate traditional medicine into the national health systems and define how it might support disease prevention, promotion and treatment [29]. This research informed us about the communities’ trust in the service provided by the traditional healers, and for that reason we believe that the traditional healers could be engaged to provide community health education and appropriate first aid and facilitate early transfer of patients to nearby healthcare facilities. As some of the traditional healers provide care on a fee for service basis, identifying mechanisms to financially compensate traditional healers could facilitate their engagement for community health education, appropriate first aid and timely referrals. The role of traditional practitioners in working with skilled care providers and facilitating referrals to hospital is well investigated for other highly important public health issues such as safe motherhood services. For economic, access or cultural reasons, many women receive care from traditional birth attendants. As deliveries are safer if conducted by a skilled attendant or at a facility providing quality care, integration of traditional birth attendants with the formal system has been promoted. One of the barriers to this integration is the potential financial implication for the traditional care providers [31]. Research, however, has highlighted that the traditional care providers are willing to work with the formal health services [32]. As the use of services provided by traditional healers is common among snakebite victims, it is important that efforts are made to engage the traditional healers; and it is anticipated that they could be willing. However, as provision of care to snakebite victims is a source of income for traditional healers, a strategy that does not address financial concerns may fail to create an effective linkage between snakebite traditional healers and the formal health sector.
While it could be argued that some traditional healing practices should, at the very least, be discouraged as being harmful and delaying referral to medical care, this participatory analysis of the local situation suggested that ‘phasing out’ of the whole concept of traditional healing for snakebite would not be an easy or advisable option. Policies that intend to force such moves may be challenging due to deeply held beliefs and a myriad of factors influencing peoples’ attitudes and practices, but more importantly, may also create tension and conflict between the formal health sector, communities, and traditional healers. Conversely, the concomitant use of biomedical care and traditional care by these communities offers an opportunity to facilitate the involvement of traditional healers for improved preventative practices, correct first aid procedures, and timely access to AV and other biomedical care as needed. In Nepal, for example, it was found that the traditional healers could be successfully trained to perform critical roles in primary prevention, first aid, and referrals [13].
Though disregarding traditional healthcare altogether may seem logical from a clinical point of view, it fails to take into account complex societal and health system factors. Even worse, such an approach tends to undermine cultural and traditional beliefs, and could cause further alienation and disempowerment. Within a context of deeply embedded beliefs, more credible are those suggestions which promote a dialogue between traditional healers and modern medical practitioners [8, 2, 33]. A plan to combat the snakebite problem needs to acknowledge traditional healers as an important stakeholder with potential to act as partners in prevention, appropriate first aid and prompt referral to effective treatment at health facilities. With monks providing traditional healing to the communities in Myanmar, they could be engaged in a similar way to the Maw Phra, or Doctor Monk, program which was implemented by Dr Prawase Wasi during the 1970s and early 80s. The program involved Buddhist monks, who are highly respected in Myanmar, and hold a strong association between Sadha, education and care for patients. It provided refresher courses in herbal medicine, as well as some basic skills of biomedicine [34].
Limitations: This research was able to yield some valuable data, particularly about the methods of and reasons for the use of services by traditional healers. Nevertheless, it has some important limitations. First, while local staff were actively involved in facilitating these sessions, interpreters were required for the researchers who didn’t understand Myanmar language. The local staff interpreted and translated; however, some interpretations may have been lost in translation. Secondly, this research was limited to three communities only, and the perspectives that we have gained about cost, transport, effectiveness of biomedical care could be area specific. Thirdly, the participatory sessions took place during the middle of the day when some young farmers had to be working in the fields. Since they are a demographic at such high risk of snakebite their participation would have been valuable. Fourthly, some of the healthcare staff had interacted separately with the project team members as a part of the wider project, and their views, particularly about the need for AV to treat snakebite patients, might have been influenced by that interaction. Despite these limitations, the purpose of this research was to highlight the snakebite phenomenon from the communities’ point of view. Public health and health system managers should take into account the valuable perspective that was gained, and would stand to benefit from discussions with their local communities on how the biomedical and traditional systems might operate side by side.
|
10.1371/journal.pgen.1005003 | The Bicoid Class Homeodomain Factors ceh-36/OTX and unc-30/PITX Cooperate in C. elegans Embryonic Progenitor Cells to Regulate Robust Development | While many transcriptional regulators of pluripotent and terminally differentiated states have been identified, regulation of intermediate progenitor states is less well understood. Previous high throughput cellular resolution expression studies identified dozens of transcription factors with lineage-specific expression patterns in C. elegans embryos that could regulate progenitor identity. In this study we identified a broad embryonic role for the C. elegans OTX transcription factor ceh-36, which was previously shown to be required for the terminal specification of four neurons. ceh-36 is expressed in progenitors of over 30% of embryonic cells, yet is not required for embryonic viability. Quantitative phenotyping by computational analysis of time-lapse movies of ceh-36 mutant embryos identified cell cycle or cell migration defects in over 100 of these cells, but most defects were low-penetrance, suggesting redundancy. Expression of ceh-36 partially overlaps with that of the PITX transcription factor unc-30. unc-30 single mutants are viable but loss of both ceh-36 and unc-30 causes 100% lethality, and double mutants have significantly higher frequencies of cellular developmental defects in the cells where their expression normally overlaps. These factors are also required for robust expression of the downstream developmental regulator mls-2/HMX. This work provides the first example of genetic redundancy between the related yet evolutionarily distant OTX and PITX families of bicoid class homeodomain factors and demonstrates the power of quantitative developmental phenotyping in C. elegans to identify developmental regulators acting in progenitor cells.
| Animals develop as one initial cell, the fertilized egg, repeatedly divides and its progeny differentiate, ultimately producing diverse cell types. This occurs in large part by the expression of unique combinations of regulatory genes, such as transcription factors, in precursors of each cell type. These early factors are typically reused in precursors of different cell types. The nematode worm Caenorhabditis elegans is a powerful system in which to identify developmental regulators because it has a rapid and reproducible development, yet it shares most of its developmental regulators with more complex organisms such as humans. We used state-of-the-art microscopy and computer-aided cell tracking methods to identify the developmental role of worm homologs of the OTX and PITX genes, whose human homologs play a role in the development of the brain, eye, and pituitary among other tissues. We identified broad roles for OTX in regulating development for many distinct cell types including muscles, neurons and skin, and found a redundant role for both OTX and PITX in a subset of cells. Future studies of these genes should address whether these genes also act redundantly in mammals.
| Identifying regulators of the intermediate steps that link pluripotency and terminal differentiation is a fundamental challenge in developmental biology. These regulators are comparatively poorly understood for most tissues due to the difficulty of recognizing and isolating cells in these transient intermediate states (“progenitors”) and their complex combinatorial logic. Individual transcription factors (TFs) acting at these stages often have broad and diverse expression domains that don’t correlate well with specific tissue or cell types [1], with multiple TFs typically acting together to specify any given intermediate progenitor. Therefore, loss of function can lead to pleiotropic phenotypes, while partial redundancy between regulators can lead to reduced penetrance, making it hard to determine the relationship between expression and biological function. Large-scale screens for gene pairs with synthetic phenotypes, as has been done for yeast [2] can identify genes acting in parallel, but screening at that scale is not feasible in animals. We are overcoming these challenges with a systematic approach to define pleiotropic and redundant progenitor TFs in Caenorhabditis elegans, a simple model organism where lineage relationships are already understood, large-scale gene expression resources allow rapid identify patterns of TF overlap, and powerful tools exist for characterizing mutant phenotypes across all embryonic cells. Previous studies of genetic redundancy in C. elegans have prioritized gene pairs for synthetic lethality testing based on similar functional interactions [3,4], expression patterns [5] and homology or conservation [6,7].
Progenitor cells are easily identified in C. elegans because the relationship between cell lineage and fate is known and invariant[8,9]. The first several embryonic divisions give rise to founder cells, some of which have clonal or partially clonal cell fates. Most cells, however, retain a multipotent state until the final round of embryonic cell divisions, when two daughters adopt such different fates as a neuron and an epithelial tube or neuron and hypodermal (skin) cell. Thus, any TF expressed in a non-clonal progenitor cell or group of lineally related cells (i.e. lineage) at any time after the earliest cell divisions but prior to the final round could play a role in progenitor identity. Despite this potential, genetic studies have identified numerous regulators of both early founder cell identity [10–16] and of terminal fate[17–19], but fewer regulators of intermediate progenitor identity. Automated methods to track cell lineages from confocal microscopy image series have allowed quantitative expression measurements for over 200 transcription factors across every cell of C. elegans embryos [1,20–22], and this EPIC (Expression Patterns In Caenorhabditis) dataset suggests many candidate regulators of progenitor identity [1,23]. Computer-aided cell tracking of mutant embryos can confirm these regulators by identifying a wide range of pleiotropic defects, from wholesale fate transformations to subtle defects in cell migration or division timing [10,14,24–27].
Many previous studies of TF function relied on reporter gene expression to infer developmental defects. We reasoned that the complex patterns of cell cycle length asynchrony and cell migration that occur in later embryos may allow identification of defects at single cell resolution without such reporters. We used this approach to characterize the developmental role of the candidate progenitor regulator ceh-36, which encodes an orthodenticle/OTX homeodomain family transcription factor orthologous to mammalian OTX1, OTX2 and CRX proteins. A ceh-36 reporter is expressed in multiple progenitor cells, encompassing the precursors of 248 terminal cells with diverse fates including neurons, glia, the excretory (renal) system, visceral and body muscles, epidermal and rectal epithelial cells[1]. Vertebrate OTX factors are similarly expressed and required in precursors of diverse tissues [28–37], suggesting these factors could be conserved regulators of progenitor identity. However, previous studies of ceh-36 mutants identified defects only in the embryonic specification of four neurons [38–40]. The large number of expressing cells combined with the small number of cells known to require ceh-36 raises the question of whether ceh-36 is required across most expressing cells or only a minority of these cells.
We found that ceh-36 null mutants are viable embryonically, with partially penetrant larval lethality and superficially normal morphology. Cell lineage tracing of ceh-36(-) embryos revealed variably penetrant defects in cell division patterns or cell migration in over 100 cells that normally express ceh-36. Double mutants lacking both ceh-36 and the coexpressed PITX-family homeobox gene unc-30 exhibited 100% synthetic lethality and severe morphological defects. These double mutants have dramatically increased rates of defective cell division and migration in coexpressing cells, indicating ceh-36 and unc-30 act in parallel to regulate the development of these cells. This provides the first evidence for genetic redundancy between OTX and PITX homeodomain factors, two bicoid class TFs that are predicted to bind similar sequences, yet diverged prior to the radiation of metazoan species.
A ceh-36 deletion allele that removes the majority of coding regions, including the homeodomain, was annotated as embryonic lethal in WormBase based on limited previous characterization [40,41]. After outcrossing, we found that nearly all embryos homozygous for this allele hatched, while ~60% of animals arrest as larvae (Table 1, Fig. 1A, B, S1 Table). The remainder of ceh-36(ok795) animals survived to adulthood and were fertile. Most arrested larvae had normal body morphology, with 5.6% of L1s containing a small bubble-like “vacuole” at the tip of the head (Table 2, Fig. 1C). Two other ceh-36 alleles predicted to eliminate or alter the homeodomain displayed similar rates of larval arrest (Table 1), suggesting this is the null phenotype. A fourth allele, ky640, which truncates the protein but is predicted to encode a complete homeodomain, displayed lower lethality rates, suggesting it leads to partial loss of function. An extrachromosomal genomic fosmid transgene containing CEH-36::GFP (+) rescued ceh-36(ok795) larval lethality; the 85% survival in this strain corresponds to nearly 100% after accounting for the 25% rate of transgene loss (Table 1). Consistent with this, 95% of CEH-36::GFP-positive L1s survived. Ectopic expression of CEH-36::GFP under the control of a heat-shock promoter caused extensive lethality when induced prior to the 50-cell stage, while later induction had little effect (Fig. 1D), indicating that CEH-36 is toxic when expressed in these early embryonic cells, but not in later cells. We conclude that ceh-36 is required for robust larval viability but not for gross morphology or embryonic viability.
We previously analyzed expression of a 5-kb ceh-36 promoter fusion reporter and identified expression in several major lineages (Fig. 2, S1 Fig) [1]. Since this reporter may not contain all relevant regulatory sequences, we generated transgenic strains using a fosmid clone from the “Transgeneome” project [22] where CEH-36 protein is fused to GFP in the context of the endogenous locus (Fig. 2A). This transgene rescues the higher-penetrance ceh-36 mutant lineage defects and larval arrest phenotype described below (Table 1). Using lineage analysis, we identified all CEH-36::GFP expressing cells through the comma stage, at which point the embryo starts to move. CEH-36::GFP is expressed in progenitors of 248 terminal cells from six lineages that together produce a mix of diverse cell types including pharyngeal cells, muscles, neurons, glia and specialized cell types, and programmed cell deaths (Fig. 2B, C). CEH-36::GFP is predominantly (>90%) expressed symmetrically between left and right symmetric lineages, despite left-right asymmetric expression and function for two of the four neurons previously shown to require ceh-36 [38,40]. The spatial expression pattern is similar to the previously analyzed ceh-36 promoter fusion (Fig. 2B, S2 Fig), but includes additional expression in the ABara lineage. We also analyzed a previously published 2-kb promoter fusion reporter [38,40,42] that we found is expressed in the MSa, MSp and ABalpa lineages but not ABara, ABplp or ABprp, indicating the existence of multiple regulatory elements for ceh-36 in different lineages (S3 Fig).
The CEH-36 protein fusion reporter exhibits complex dynamics that we confirmed by single molecule RNA-FISH (smFISH) [43] of endogenous ceh-36 mRNA (see below). Expression in the ABpxp, MSaa, and MSpa lineages begins between the 50-cell and 100-cell stages and decreases in most cells after 2–3 cell cycles, prior to morphogenesis (Fig. 2B). However, a few cells maintain stable expression much longer, up to at least comma stage. The CEH-36::GFP expressing cells include progenitors of three neurons previously shown to require ceh-36 (MI, AWCL and AWCR), with additional stronger expression in AWCL, AWCR and the fourth ceh-36-requiring cell, ASEL, beginning after the worm begins to elongate and twitch [38–40]. In total, we found expression of ceh-36 in progenitors of over 30% of embryonic cells suggesting it could play a broad role in embryonic patterning. Its early and transient expression in progenitor cells suggested that ceh-36 might be an important regulator of progenitor identity or function.
The lack of obvious morphological defects in ceh-36 mutants suggests that ceh-36 might play a minimal role in the development of most expressing cells. To test this, we searched for defects in lineage patterns and cell migrations in mutant embryos using automated cell tracking. We examined quantitative features of embryonic development, including timing and patterns of cell division, division orientation, and positions in eight ceh-36(ok795) embryos through the comma stage (~400 minutes after fertilization, when nearly all cell divisions have occurred) and compared these phenotypes to a wild-type reference set [27] (see Materials and Methods) and to three embryos expressing a rescuing CEH-36::GFP transgene. We also examined one embryo carrying a second predicted ceh-36 null mutation (ky646). As detailed below, we found that many cells in ceh-36(-) embryos have partially penetrant defects in both cell cycle timing and cell position (Figs. 3,4, S2 Table). In total, 5.1% (495/9636) of cells in ceh-36(ok795) embryos were defective in cell division or position, compared with 0.3% (85/26171) of cells in wild-type control embryos (p < 10-220; chi-squared test). This suggests that ceh-36 is broadly important for robust development across its expressing cells.
The C. elegans lineage is composed of an invariant pattern of cell divisions and deaths. In wild-type embryos, the division timing is highly stereotyped, with most cells having variability in cell cycle length of less than 5% [27] [44]. We identified 49 cells with cell cycle or lineage timing defects in at least one ceh-36 mutant embryo (Figs. 3A, D, 4, S2 Table), defined as cells dividing both three standard deviations and at least five minutes earlier or later than expected, not dividing at all, or dividing inappropriately. In addition, three cells failed to undergo programmed cell death when expected, as recognized by the characteristic pattern of chromatin compaction observed for histone-mCherry. For example, in three embryos, MSpaapp, which normally is the first embryonic cell to undergo apoptosis, instead survived and divided, with both sisters migrating into the pharynx to adopt unknown fates (Fig. 3C). In some cases, cells not passing our threshold for defect calling appeared to have different mean cell cycles or positions. For example 35 of 49 cells with cell cycle defects in one or more embryos also had a nominally significant difference in mean cell cycles (p < 0.1; FDR < 0.15; S2 Table). The CEH-36::GFP fusion protein is expressed in precursors of 86%(12/14) of cells with cell division timing defects in two or more ceh-36(-) embryos, and 60% (21/35) of cells with defects in one embryo. This is significantly more than the 30% of all cells that express CEH-36::GFP (chi squared p < 2 × 10-6). CEH-36::GFP is also expressed in all of the cells with supernumerary divisions or failed cell death.
Cell positions are also highly consistent between wild-type embryos, allowing us to identify cell migration defects by comparing cell positions between ceh-36 mutant and wild-type embryos. We identified 124 cells whose deviation from expected position was at least 3.5 standard deviations greater than in the wild-type set and that had aberrant neighbors as defined by an empirical neighbor-distance score (S2 Table; see Methods). Position defects were strongly enriched in expressing cells; 81% (55/68) of cells with position defects in two or more embryos normally express CEH-36::GFP. By comparison, in 22 wild-type embryos examined, only 13 cells had defective positions, in one embryo each.
A cell could be misplaced because of a defective migration, in which case it would have both different position and different neighbors than in the wild type. Alternatively, a cell could be misplaced because its normal position is occupied by another cell that migrated inappropriately, in which case its position relative to its normal neighbors would be unchanged. We used these criteria to classify 50 cells with position defects by examining their position and neighbors in 3D visualizations (Fig. 3E, F). We scored 82% (41/50) of cells as likely defective migrations, while 9/50 (18%) defects could be explained by defective migration of other cells (S3 Table). 100% (18/18) of higher-penetrance (seen in at least three of eight ok795 embryos) position defects examined were scored as likely migration defects. The migration defects include both cells that undergo novel migrations in the mutant (Fig. 3E) as well as cells that fail to undergo their expected migrations (Fig. 3F). The cells scored as possible secondary defects were less penetrant, with each identified as defective in one or two embryos. Still, most low-penetrance defects (23/32) were scored as likely migration defects
We observed dramatic defects in eight laterally positioned cells that were born in the correct position but subsequently migrated across the midline to the opposite lateral side of the embryo, sometimes displacing the position of their bilateral counterpart (e.g. Fig. 3E). These lateral migration defects occurred on both sides of the embryo (3 L→R, 5 R→L) and include diverse cell types: neurons (I1R and I2L), pharyngeal cells (pm3R and mc1DR), rectal cells (left intestinal muscle and anal depressor muscle), and tail cells (Hyp10 and tail spike). These defects were all low penetrance (seen in one or two of eight ok795 embryos), but we saw no defects of this class in the 22 wild-type control embryos, and all eight of these cells normally express ceh-36. This indicates that C. elegans cells’ lateral position is not merely a result of their birth position but is regulated by factors that include ceh-36.
We determined that lateralization defects are maintained through embryonic elongation and not corrected by subsequent cell movements by examining worms expressing FKH-4::GFP, a marker of three visceral muscles (left and right intestinal muscles and anal depressor; Fig. 3E, S4 Fig). 100% of both wild-type and ceh-36 mutant elongated (pretzel-stage) embryos have three FKH-4(+) cells, indicating that ceh-36 is not necessary for FKH-4 expression. However, one FKH-4(+) cell is laterally mispositioned in 14% of ceh-36(-) embryos (Table 3; Fig. 3E). This is consistent with the left-right migration phenotype and low penetrance observed in our lineage data (1/8 ok795 embryos, 12.5%), increasing confidence in the low-penetrance defects identified by lineage analysis.
Multiple pharyngeal gland cell precursors had cell cycle and position defects in ceh-36 mutants. For example, the daughters of the MSaapapa cell normally produce a pharyngeal gland cell and a programmed cell death and the early division of this cell was the largest division-timing defect we observed in ceh-36 mutants (Fig. 3A). Precursors of four of the five pharyngeal gland cells express CEH-36::GFP and all four of these had partial penetrance defects in cell cycle or position (Figs. 3A, 4). Since pharyngeal gland cells are known to be required for feeding and viability [45], we examined them for additional defects by examining expression of the pharyngeal gland marker hlh-6::GFP in elongated ceh-36(ok795) embryos. We observed altered pharyngeal gland morphology in 20% of ceh-36(ok795) elongated embryos. An additional 9% of embryos were missing one or more hlh-6::GFP-positive cells (Fig. 3B, Table 2), suggesting that ceh-36 regulates not only gland cell cycle patterns and morphology but also terminal fate. While only 41% (23/56) of larvae with normal gland morphology arrested prior to the L4 stage, 92% (46/50) of larvae with abnormal gland morphology arrested. Thus, defects in pharyngeal gland morphology predict larval arrest in ceh-36 mutants.
Defects occurred in 223 unique cells, typically with low penetrance; only 82/223 (37%) cells were defective in two or more (of eight) ok795 embryos. Most of the defective cells normally express CEH-36::GFP (77%), significantly more than the 30% fraction of all cells that express ceh-36 (p < 10-90, chi-squared test). Most of the defective cells that do not normally express ceh-36 were only called as defective in one embryo. Still, even defects seen in a single embryo were enriched in expressing cells (59% of such cells express CEH-36::GFP). While cells with prior cell cycle defects were 2.9-fold more likely to have position defects (p < 10-9), 90% of cells with position defects had no detectable cell cycle defect. Only 22 expressing cells had defects in at least 50% of analyzed embryos (e.g. Fig. 4, S2 Table). The low penetrance of most individual defects may explain the viability of ceh-36(-) embryos.
We determined whether cells with low penetrance defects have noticeable defective terminal positions or numbers by examining several fluorescent markers expressed in these cells (Table 2). Reporters for two cells previously reported as requiring ceh-36 (MI(sams-5) and ASEL(gcy-5)) showed the expected terminal defect frequencies in the ceh-36 deletion. As described above, the visceral muscle reporter FKH-4::GFP and the pharyngeal gland reporter hlh-6p::GFP also showed terminal position defect frequencies consistent with the observed embryonic defects. Finally, a FLP-1::GFP reporter reported as expressed in the AVK neuron (which expresses ceh-36 but was not identified as defective in our analysis) showed little or no terminal defects (~2%). Given that the mutant embryos hatch without major morphological defects despite an average 40 cells with position defects and 10 cells with altered division timing, development of C. elegans embryos must be robust to a substantial amount of developmental error.
The rescuing CEH-36::GFP transgene expression is typically strongest several divisions before the birth of the terminal cells where most defects were identified, suggesting that defects in ceh-36(-) may result from regulatory events occurring in mitotic progenitor cells. If this is true, partially penetrant defects should preferentially co-occur in closely related cells within a given embryo. We identified 71 examples of defective sister cell pairs in ceh-36-expressing cells. We found preferential co-occurrence of defects in sisters for seven embryos (p < 0.001) by using a bootstrap evaluation, and this co-occurrence was only significant in cells expressing ceh-36. This along with the early and dynamic CEH-36::GFP expression suggests that ceh-36 regulates development in part through its activity in progenitor cells, rather than the terminal cells that exhibit the defects.
To confirm that most defects identified in ceh-36(ok795) embryos result from loss of ceh-36, we specifically examined high-penetrance (≥6 of 8 ok795 embryos) position defects in an embryo carrying a second predicted ceh-36 null mutation (ky646). We found four of the five cells examined had similar defects in this embryo. We examined these cells in two ceh-36(ok795) embryos expressing CEH-36::GFP, and one embryo with mosaic CEH-36::GFP expression, and found that these defects were rescued in all CEH-36::GFP expressing cells. Taken together, these results show that ceh-36 regulates the robustness of cell cycle and migration patterns in many cells. Our analysis did not explicitly test for changes in cell fate, but given the known role of ceh-36 in fate specification [38–40], there may be additional unidentified cells with defects in fate, but not position or cell cycle timing.
Most defects in ceh-36(ok795) have low penetrance, so other transcriptional regulators likely function in parallel with ceh-36 to ensure robust development. Therefore, we searched for transcription factors that might act redundantly with ceh-36 (Fig. 5). Previous work demonstrated that the three OTX family members ceh-36, ceh-37, and ttx-1 can rescue the others’ mutant phenotypes when expressed in the appropriate cells [39]. We asked if these genes’ early embryonic expression overlaps with that of ceh-36 by lineage analysis of fluorescent reporters and single molecule (sm)RNA-FISH [43]. Lineage analysis of a ceh-37 promoter-fusion reporter[46] identified ten cells where its expression overlaps spatially but not temporally with ceh-36; the ceh-37 reporter is expressed after CEH-36::GFP in these cells (Fig. 5A). The ceh-37 reporter is also expressed in several lineages that do not express CEH-36::GFP. ceh-37 transcripts identified by smRNA-FISH did not overlap with positions of ceh-36 transcripts prior to the 50-cell stage and there was only a small amount of overlap between the 50 and 200 cell stages (Fig. 5B, S5 Fig). We could detect no embryonic expression of a ttx-1 promoter reporter prior to morphogenesis and little or no overlap between ttx-1 and ceh-36 transcripts by smRNA-FISH (Fig. 5B). We examined these genes’ expression in ceh-36(ok795) by smRNA-FISH and observed no ceh-36 transcripts and no changes in ceh-37 or ttx-1 expression. We also observed no substantial increase in ceh-36(ok795) lethality after ttx-1 or ceh-37 RNAi. This indicates that most ceh-36-expressing cells do not express other OTX homologs in wild-type or ceh-36 mutant embryos, and redundancy with these factors is unlikely to explain the low penetrance of most ceh-36 mutant defects.
We mined the EPIC database of embryonic expression patterns [1,23] for additional factors coexpressed with ceh-36, and identified substantial coexpression with the PITX homolog unc-30. An UNC-30::GFP fosmid “Transgeneome” reporter [22] was transiently expressed at the same time as CEH-36::GFP in the descendants of the ABplp and ABprp progenitor cells (together “ABpxp”), which give rise to diverse cell types, but not in other CEH-36-expressing lineages. We confirmed this expression overlap between endogenous ceh-36 and unc-30 transcripts in ABpxp-derived cells by lineage tracing (Fig. 5C) and observed significant overlap of these genes’ endogenous transcripts by smRNA-FISH between the 28-cell and 50-cell stages (Fig. 5D). ceh-36/OTX and unc-30/PITX both encode bicoid-type homeodomains that are predicted to bind similar target sequences [47,48]. In addition, the combined frequency of position and cell cycle defects in ceh-36(ok795) embryos was lower (3%) in the ABpxp lineages than in other CEH-36::GFP-expressing cells (12%) suggesting that ceh-36 may have more redundancy in ABpxp than in other lineages. This suggested the possibility that these two factors might act redundantly to regulate the development of the ABpxp lineages.
In addition to its early ABpxp expression, we observed UNC-30::GFP in the six embryonic type D GABA-ergic motorneurons as well as a few other neurons (PVP, AWA, ASG, AIB, ASI and GLR) at morphogenesis (bean stage), consistent with the known role of unc-30 in the terminal differentiation of type D neurons [49] (S6 Fig). Consistent with the phenotypes of other unc-30 alleles, the deletion allele unc-30(ok613) is uncoordinated yet fully viable, with no embryonic or larval arrest (Table 1).
We tested for redundancy between unc-30 and ceh-36 by examining the progeny of a strain homozygous for both unc-30(ok613); ceh-36(ok795) and carrying the rescuing extrachromosomal CEH-36::GFP fosmid. Animals that had lost the rescuing transgene displayed 100% lethality (54% embryonic, 46% larval), while embryos expressing CEH-36::GFP had no embryonic lethality and low larval arrest rates (Table 1), with 75% progressing to L4. The residual larval arrest rate could result from transgene mosaicism or incomplete rescue by the CEH-36::GFP transgene. This indicates that ceh-36 and unc-30 are redundantly required for viability.
The unc-30(ok613); ceh-36(ok795) double mutants displayed visible phenotypes characteristic of defects in ABpxp-derived cells not observed in either single mutant (Table 3, Fig. 6). These included variable abnormalities in body morphology (Vab) defects, which are also seen when ABpxp-derived cells fail to act as a substrate for hypodermal enclosure [50,51], “no backend” (Nob) tail defects characteristic of severe defects in patterning posterior cells including many derived from ABpxp [52], and a “rod-like” arrest posture and large edemas near the pharynx characteristic of defects in the excretory system [53], which is formed by descendants of ABpxp cells. Double mutants did not contain the more anterior head “vacuoles” we saw in ceh-36(ok795) single mutants; however this phenotype could be masked by the more severe Vab and excretory phenotypes. Taken together, ceh-36 and unc-30 are redundantly required for viability and for aspects of normal development associated with cells produced by ABpxp.
The highly penetrant viability and morphological phenotypes of unc-30(ok613); ceh-36(ok795) double mutants led us to hypothesize that these animals would have more frequent cell lineage and position defects in the cells that normally coexpress both factors. We tested this by automated lineage analysis of six unc-30(ok613); ceh-36(ok795) embryos that had lost the rescuing CEH-36::GFP transgene (Fig. 7, S2 Table). We observed a significant increase in cell cycle and cell position defects in the ABpxp lineages of unc-30; ceh-36 double mutants as compared to ceh-36 alone. Double mutant embryos averaged 15.5 cell cycle defects and 94.7 position defects per embryo in ABpxp compared with 2.25 and 11 in ceh-36 single mutants. We also saw a smaller increase in position defects for cells that do not normally express either CEH-36::GFP or UNC-30::GFP (64 vs 26.1), consistent for a role of the ABpxp cells in migration of cells from other lineages. In contrast, we saw no corresponding increase in cell cycle defects in double mutants for nonexpressing lineages (Fig. 7), and no corresponding defects in UNC-30 single mutants (S7 Fig).
We observed an increased co-occurrence of defects in sister cell pairs in the ABpxp lineages of double mutants (27.2 sisters pairs per embryo) compared with ceh-36(ok795) alone (3.8 sister pairs per embryo). Because unc-30 expression in the ABpxp lineage is even more transient than that of ceh-36, the increased co-occurrence of defective sister cells likely results from primary defects in progenitor cells. We individually examined ceh-36(ok795) single mutant and unc-30(ok613); ceh-36(ok795) embryos for “defect trios” of a defective mother with two defective daughter cells. While ceh-36(ok795) embryos have on average 1.9 of these defect trios (none in ABpxp), double mutant embryos average 13.5 defect trios (10.8 in ABpxp). We observed only one defect trio in three unc-30(ok613); ceh-36(ok795) embryos expressing rescuing CEH-36::GFP, indicating that these defects do not result from unc-30(ok613). Together this suggests that unc-30 and ceh-36 cooperate to regulate robust lineage patterning of ABpxp-derived progenitor cells.
Two cell divisions (ABplpappa and ABplppaa) each displayed a novel type of defect we term “anterior-posterior reversals” in 2 of 6 double mutant embryos. In this class of defect, the anterior daughter of a division adopts the division pattern of the posterior daughter and vice versa. This was evident in the patterns of asymmetric division timing, cell death and migration (Fig. 7B-D). For example, ABplpappap, which undergoes cell death in the wild type, survives and divides in the double mutant. Meanwhile, that cell’s anterior sister ABplpappaa, which should generate RMEV and the excretory canal cell, instead undergoes programmed cell death. Consistent with this being a fate reversal, the division occurs with normal orientation and the daughters of the cell that should have died go on to adopt positions characteristic of RMEV and the excretory canal cell. Defects in these fate-reversed cells or their failure to function correctly in their new location could explain some of the excretory system edema observed in the double mutants.
We identified numerous defects in the organization of the ventral midline in the double mutants. Several ABpxp-derived cells failed to respect the midline and crossed to the opposite side; these were distinct from those seen in ceh-36 single mutants (Fig. 7E). Also, in contrast with ceh-36 single mutants, the double mutants had a much larger number of (presumably nonautonomous) defects in cells that normally express neither ceh-36 nor unc-30. Most of these defects (43/65) were in cells derived from the ABpxa lineages in cells that should form the ventral epidermis. Previous work showed that in the process of ventral enclosure, the epidermal cells migrate over ABpxp-derived substrate cells, some of which are mispositioned in the double mutants. The “leading cells” hyp6/ABpxaappap and hyp7/ABpxaappaa, which initiate ventral enclosure, along with adjacent migrating epidermal cells hyp4, G2, and W, had the largest magnitude defects in cell position of nonexpressing cells (Fig. 7E). This suggests that ceh-36 and unc-30 regulate development of the ABpxp-derived substrate for normal epidermal migration and morphogenesis.
Several ABpxp-derived cells had position defects of much larger magnitude than we observed in ceh-36(-) alone (Fig. 7F). Among the cells with the largest defects (average 8.9 micron (>2 cell diameters) deviation from expected position compared with 1.9 microns in ceh-36 alone and 1.6 microns in wild type) were two sister pairs that normally produce two DB neurons and the excretory duct and G1 pore cells (Fig. 8A, B). In wild-type embryos, these cells migrate from anterior lateral positions to the ventral midline where the duct and pore cells connect with the excretory canal cell to form a continuous three-celled tube (Fig. 8C) [54].
Previous work showed that the HMX homeodomain transcription factor mls-2 is required for robust development of the excretory duct and pore [55]. MLS-2::GFP is expressed in several lineages including the precursors of the excretory duct and pore (Fig. 8A, C). To determine if loss of mls-2 leads to cell migration defects similar to those seen in unc-30;ceh-36 double mutants, we traced the lineages and cell positions of the excretory system cells in 23 mls-2 mutant embryos (Fig. 8D). We found migration failure or inappropriate migration into the head in 43% (10/23) of duct cells and 57% (13/23) of pore cells indicating that mls-2 is required for robust migration of these cells. Consistent with this, previous work [55] found that 5 of 25 mls-2 mutant larvae were missing an excretory tube cell; the difference in rates suggests that in some cases the misplaced cells may eventually migrate to the correct position, or that another cell may sometimes adopt a duct or pore fate.
We tested whether mls-2 expression depends on ceh-36 and unc-30 by measuring the expression of a genomically integrated rescuing MLS-2::GFP reporter [55] in unc-30; ceh-36 double mutant embryos. MLS-2::GFP was expressed normally in ceh-36(ok795) single mutant embryos (n = 6). Since MLS-2::GFP is expressed later and ~10-fold more strongly than CEH-36::GFP in the excretory duct and pore lineages (Fig. 8A), we were able to compare MLS-2::GFP expression between double-mutant embryos carrying the rescuing CEH-36::GFP with unrescued embryos. We found that in all (7/7) embryos carrying CEH-36::GFP, MLS-2::GFP was robustly expressed in both the duct and pore precursors, and the duct and pore cells migrated normally. In contrast in six embryos (12 duct/pore lineages) that had lost the rescuing transgene and expressed no CEH-36::GFP, 58% of duct/pore lineages (7/12) had no MLS-2::GFP expression, with the remaining lineages expressing MLS-2::GFP at lower levels than in wild-type or rescued embryos. Absence of MLS-2::GFP expression predicted migration defects; all seven duct or pore cells with no MLS-2::GFP expression had severe migration defects, while three of five MLS-2::GFP-expressing duct/pore cells migrated normally, sometimes with moderate delays. MLS-2::GFP expression in other lineages that don’t normally express ceh-36 or unc-30 was unaffected.
Discontinuities in the excretory tube are associated with formation of edemas and eventual lethality with a characteristic rod-like posture [53–56]. Thus the edemas (Table 3) and rod-like lethality we see in unc-30; ceh-36 double mutants could be explained by the duct and pore migration defects or defects in specification of these cells or the canal cell. We conclude that ceh-36 and unc-30 are required for robust mls-2 expression in ABpxp descendants that give rise to the excretory system, and that misregulation of mls-2 may account for the observed phenotypes in those cells.
Our analysis of ceh-36 and unc-30 function across all embryonic cells highlights the complex biology of transcriptional regulation during development that would not have been discovered using traditional approaches. We showed these factors regulate distinct processes including the cell cycle, lineage patterning, cell position, and cell fate specification in many embryonic cells that go on to adopt diverse fates. These factors likely function together to regulate progenitor identity in the ABpxp lineage and ceh-36 likely works with other unknown factors in progenitors from other lineages.
The lineage-specific cellular phenotypes and defect penetrance in ceh-36(-) and the ABpxp-specific functional interaction between unc-30 and ceh-36 are consistent with context-dependent roles for these factors. Each factor has distinct expression outside of the early ABpxp coexpression, suggesting that each may work with other factors in these other lineages; indeed, unc-30 is a well-established regulator of motor neuron differentiation later in development [49], and ceh-36 mutants have partially penetrant defects in lineages where unc-30 is not expressed. Even within ABpxp, most defects were still partially penetrant even in unc-30; ceh-36 double mutants, consistent with the existence of additional redundant factors. One role of these factors is to directly or indirectly regulate the expression of mls-2. Intriguingly, mls-2 may itself act as a progenitor identity factor, as it regulates the development of lineally-related embryonic cells including glial, excretory and neuronal cells [42,55,57] and is expressed in these cells’ progenitors. In fact, mls-2 is required for expression of ceh-36 in the AWC neurons [42], suggesting that ceh-36 (with unc-30) indirectly regulates its own expression later in development. Similarly, ceh-36 and unc-30 can bind to the unc-30 promoter [4], which is intriguing given that the later expression of unc-30 in GABA-ergic motor neurons occurs in ABpxp-derived cells. We suggest a model in which C. elegans develops robustly with an invariant lineage because each of many lineage-specific TFs [1], provides a small amount of information to each cell about its lineage history. Combining this information from many TFs allows cells to robustly adopt a fate appropriate to their lineage history (Fig. 9). This model suggests that the expression of each individual factor could be regulated by lineage mechanisms (e.g. [1]) in parallel rather than hierarchically. Another intriguing possibility is raised by our observation of cell cycle and migration defects in cells that nonetheless express appropriate terminal fate markers. This suggests that distinct regulators may modularly control different aspects of each cell’s developmental phenotype (i.e. one factor regulates fate, another, cell cycle, and yet another, migration).
Our data suggest that ceh-36 and unc-30 act in embryonic progenitor cells to regulate development, which is distinct from their previously characterized role in neuronal terminal differentiation. They are expressed early and transiently in progenitor cells from multiple lineages and these progenitors give rise to varied cell types, similar to multipotent progenitors in other organisms. The migration and cell division defects that we observe occur across these distinct cell types, and while most defects were observed in terminal cells, they were clustered in the lineage suggesting an underlying defect in the common ancestor. Together this strongly suggests that defects occurred in progenitor cells, although it does not rule out additional roles in the subset of terminal cells where ceh-36 expression persists. Early progenitor factors such as ceh-36 and unc-30 may regulate factors important in later progenitor cells, but they could also directly regulate genes expressed in terminal cells by creating stable chromatin alterations, as was recently demonstrated for another factor [58]. Cell division and migration patterns in unc-30; ceh-36 double mutant embryos do not, however, suggest a switch in fate from ABpxp to its sister ABpxa or any other recognizable sublineage. Thus, other ABpxp factors remain to be discovered or other factors are required to specify alternative progenitor fates.
Gene regulatory networks are generally robust against biological noise and often employ transcription factors (TFs) with overlapping or redundant functions to decrease transcriptional and phenotypic variability [59]. For example, in C. elegans, redundant pairs of GATA factors regulate intestine development [60], and similar redundancy exists for T-box factors [6,61] and HLH factors [62]. Despite the superficial redundancy of these factors, in some cases the single mutants exhibit decreased robustness in fate determination and partial penetrance phenotypes [60,63]. Our finding of similar redundancy between the more divergent homeodomain factors from the PITX and OTX classes indicates that redundancy can occur between factors with ancient divergence. Worms, insects, and vertebrates all have PITX and OTX homologs, indicating these factors diverged prior to the common ancestor of these phyla. This is the first demonstration of a genetic interaction between these factors that could reflect functional redundancy. Since PITX and OTX factors can bind the same sequence motif this redundancy could reflect regulation of shared targets; consistent with this, a large-scale study of TF binding by yeast 1-hybrid analysis identified binding of ceh-36 and unc-30 to highly overlapping sets of promoters [4]. However it is also possible that they work through independent parallel mechanisms. Intriguingly, vertebrate PITX and OTX homologs have some expression overlap in the pituitary and nervous system, and it will be interesting to determine whether they act together in vertebrates. Our approach of studying robustness across an entire organism at a single-cell level provides the opportunity to sensitively identify cells where each factor or combination of factors plays a role. For example, the overlapping functions of ceh-36 and unc-30 in the ABpxp sublineage allowed us to identify their role in regulating mls-2 expression in the developing excretory system.
Previous studies identified ceh-36 as a regulator of lateral asymmetry for the MI [40] and ASE [38] neurons. The pharyngeal MI neuron is derived from a right lineage, and the left equivalent lineage produces seemingly equivalent cells except for an epithelial cell, e3D, in place of the MI neuron. Mutations in ceh-36 transform MI into an e3D-like cell, and this asymmetry is driven by asymmetric ceh-36 expression in the MI progenitors and not those of e3D [40]. Surprisingly, the same phenotype occurs in a truncating mutant in histone H3, likely acting downstream of ceh-36 [64]. The fact that loss of either an asymmetrically expressed factor (ceh-36) or a symmetrically expressed factor (histone H3) leads to the same phenotype underscores that asymmetry in regulatory networks can influence which cells have phenotypes. While we do observe asymmetric CEH-36::GFP expression in MI, we found that most expression in other lineages is L-R symmetric and most penetrant defects were seen in both symmetric pairs. However we did identify defects in lateral identity, such as the migrations of the left intestinal muscle and anal depressor, in cells where ceh-36 expression is normally L-R symmetric. This suggests that ceh-36 contributes to the regulation of lateral identity even in cells where it is symmetrically expressed.
Although our approach improved the sensitivity for detection of cellular phenotypes compared to previous studies and methods, it is likely that additional cellular defects remain unidentified. For example we identified many defects in only one or two embryos; further improvements to automated cell tracking methods to increase accuracy and reduce curation time would allow analysis of higher numbers of embryos and more sensitive and reliable identification of lower penetrance defects, In the absence of markers for terminal differentiation, a cell with normal migration and division patterns but altered terminal fate cannot be detected. Repeating lineage tracing with a panel of strains expressing distinct fate markers can increase the power to detect lineage transformations [25], but this approach is labor-intensive. On the other hand, some of the cell position defects we identified were apparent only by lineage tracing and not when scored using a terminal fluorescent marker in larvae, which reflects the high sensitivity of the quantitative methods and possibly the correction of some position defects later in development. The power of future applications of lineage-based phenotyping methods would be increased by new methods to directly assay fate transformation while maintaining throughput; such as by analyzing multiple fate markers simultaneously in different colors.
ceh-36(ks86) X
ceh-36(ky640) X
ceh-36(ky646) X
ceh-36(ok795) X
ceh-37(ok272) X
ceh-37(ok642) X
unc-30(ok613) IV
unc-119(tm4063) III
mls-2(cs71) X
bwIs2[flp-1::GFP + pRF4(rol-6(su1006))][65]
nsIs396[sams-5 3′::4xNLS-GFP + lin-15(+)] V [40]
ntIs1[lin-15(+); gcy-5::GFP][38]
sEx14784[ceh-37::GFP][46]
ujEx173[CEH-36::GFP + unc-119(+)]
ujEx130[CEH-36::GFP + myo-2::mCherry + myo-3::mCherry]
oyIs48[ceh-36 2KB promoter::GFP][39]
stIs10501[ceh-36 5KB promoter::HIS-24-mCherry][1]
ujIs113[pie-1::mCherry::H2B + unc-119(+); Pnhr-2::mCherry::histone + unc-119(+)] II
wgIs108[FKH-4::GFP+ unc-119(+)] I [22]
wwIs19[hlh-6::GFP + unc-119(+)] X [66]
csIs55[MLS-2::GFP] X [55]
wgIs395[UNC-30::GFP+unc-119(+)]
All strains were grown as previously described [67]. N2 was used as the wild-type reference strain. All manipulations were performed at room temperature (21°C).
Knockout consortium alleles ceh-36(ok795) and unc-30(ok613) were outcrossed three times. VC579 ceh-36(ok795)/szT1 hermaphrodites were mated with males carrying an extrachromosomal copy of ceh-36(+)::GFP (ujEx173), and F2 progeny were tested for ok795, which deletes 406 base pairs of ceh-36, by PCR. Additional outcrossing of ceh-36(ok795) was with N2 males. unc-30(ok613) was outcrossed by mating unc-30(ok613) hermaphrodites with N2 males and picking F2 Unc progeny. Combinations of reporters with ceh-36(-) were created using a mating strategy that did not produce heterozygous ceh-36(-) hermaphrodites at any step or else were verified using PCR.
Combinations of unc-30(ok613) and ceh-36(ok795) were created using nT1[qIs51](IV;V) to balance unc-30(ok613) while testing for ceh-36(ok795) by PCR. unc-30(ok613)/nT1[qIs51](IV;V); ceh-36(ok795) males were mated with unc-119(tm4063); ceh-36(ok795); ujEx173[ceh-36::GFP + unc-119(+)] hermaphrodites, and F2 Unc progeny with the genotype unc-30(ok613); ceh-36(ok795); ujEx173 were isolated. ujEx173 was generated by microparticle bombardment of the CEH-36::GFP Transgeneome fosmid [22] into unc-119(tm4063) using methods previously described [22,68]. ujIs113 was generated by co-bombardment of pAA64H2B (pie-1::mCherry-H2B::pie-1UTR)[69] and pJIM20_nhr-2 (nhr-2promoter::HIS-24-mCherry::let-858YTR) into unc-119(tm4063). ujEx130 was generated by injection of the CEH-36::GFP transgeneome fosmid into ceh-36(ok795) worms.
All strains were grown at 20°C for over two generations before scoring. Young adult hermaphrodites were dissected at room temperature in egg buffer (118mM NaCl, 48mM KCl, 2mM CaCl2, 2mM MgCl2, 25mM HEPES) [70], and embryos with four or more cells were transferred onto NGM plates. Embryos were counted and replaced in the 20°C incubator. Embryonic lethality was determined by counting unhatched embryos on the subsequent two days. Due to a variable rate of larval development for ceh-36(-) mutants, L4 hermaphrodites were picked off the NGM plates and counted as survivors for one week following dissection. We observed no L4 lethality or adult sterility. Similar rates of lethality for ceh-36(ok795) were obtained by counting eggs laid by free moving ceh-36(ok795) hermaphrodites and following their progeny to the L4 stage. To track the presence of the fosmid in rescued animals, we generated unc-119(tm4063); ceh-36(ok795) worms that were doubly rescued by the presence of the fosmid and reduced the larval lethality of ceh-36(ok795). This allowed us to score absence of the fosmid by the presence of the Unc phenotype.
Lethality checks of unc-30(ok613); ceh-36(ok795) double mutants followed a similar protocol. unc-30(ok613); ceh-36(ok795); ujEx173[ceh-36::GFP + unc-119(+)] young adult hermaphrodites were dissected and embryos counted as described above. Embryonic lethality was scored the next morning. Unhatched embryos were mounted in 20μm beads in egg buffer/methyl cellulose [71] and scored for CEH-36::GFP expression in ASE and AWC neurons. All hatched L1s were examined using a fluorescent dissecting microscope for CEH-36::GFP expression in ASE and AWC neurons (Leica M205FA, Leica Microsystems). CEH-36::GFP expressing and non-expressing L1s were transferred to separate plates, and several larvae were found and transferred the following day. L4 survivors were picked off the NGM plates and counted as survivors for one week following dissection. A similar procedure was used to score survival of ceh-36(ok795) worms carrying wwIs19(hlh-6::GFP).
All strains were grown at 20°C for over two generations before young adult hermaphrodites were dissected at room temperature in egg buffer and embryos with four or more cells were mounted into a solution of 20μm beads in egg buffer/methyl cellulose. Sealed slides containing 10–15 embryos were incubated overnight at 20°C and scored the following morning.
Examination of unc-30(ok613); ceh-36(ok795) double mutant phenotypes followed the above protocol except that embryos were also scored for CEH-36::GFP expression in ASE and AWC neurons following DIC examination to exclude rescued animals.
We acquired confocal images with a Leica TCS SP5 resonance scanning confocal microscope (67 z planes at 0.5 μm spacing and 1.5 minute time spacing) and generated lineages using StarryNite and AceTree as previously described [20,21,72–75]. Embryos were mounted in egg buffer/methyl cellulose with 20μm beads as spacers [71] and imaged at 22°C using a stage temperature controller (Brook Industries, Lake Villa, IL).
We updated the 4D reference model of wild-type C. elegans embryogenesis through the 600-cell stage using eighteen embryos expressing fluorescently tagged histone by tracing four embryos to the comma stage. Deviation of cell-cycle length, division orientation, and anterior-posterior position for eight ujIs113; ceh-36(ok795) embryos, one ujIs113; ceh-36(ky646) and five ujIs113; unc-30(ok613); ceh-36(ok795); ujEx173(CEH-36(+)::GFP) embryos was calculated as previously described [27]. Deviant cell cycle length was defined as beyond three standard deviations and five minutes of the average wild-type cell-cycle length. For position defects, we calculated the expected position of each cell in the embryo based on the overall rotation of the embryo and the wild-type model and scored the distance from the expected position. Cells were considered mispositioned if their mean or maximum distance was more than 3.5 standard deviations greater than the wild-type mean. We also developed a heuristic “neighbor distance” score, consisting of the mean distance of the cell to the 10 cells that are closest to that cell in wild-type embryos, and required 3.5 standard deviation defects in this score as well. Deviant cell position was confirmed by comparison of time-lapse 3D-models for both mutant and wildtype embryos. Defects in all cells identified through statistical analysis mentioned in the text were confirmed by manual retracing of curated lineages. For bootstrap analysis of defective sister pairs, the number of total defective cells (X) and defective sister pairs (Y) were separately counted for each embryo as well as for defined subgroups (e.g. the ABpxp lineage or ceh-36 expressing versus non-expressing). Cells born before the onset of ceh-36 expression were not considered. The number of defective sister cells expected by chance was determined by 100,000 iterations of counting sister pairs from (X) randomly picked cells from a defined subgroup. A p-value was calculated by dividing the total number of iterations equal to or greater than the observed value (Y) with 100,000.
Mixed-stage embryos were picked into a solution of 10mM sodium azide and 1% methyl cellulose in egg buffer with 25μm beads on top of a glass slide. Coverslips were sealed using petroleum jelly, and embryos became immobilized due to azide and hypoxia. All fluorescent reporters were scored by analyzing confocal GFP and DIC z-stacks of pretzel-stage embryos, which provided a more discrete developmental stage than possible in larvae due to the larval arrest of ceh-36(-) mutants. Positional defects and wild-type variation of fluorescent reporters were measured using LASAF software. Single-molecule RNA FISH was performed as previously described [43,76].
|
10.1371/journal.pgen.1002430 | Azospirillum Genomes Reveal Transition of Bacteria from Aquatic to Terrestrial Environments | Fossil records indicate that life appeared in marine environments ∼3.5 billion years ago (Gyr) and transitioned to terrestrial ecosystems nearly 2.5 Gyr. Sequence analysis suggests that “hydrobacteria” and “terrabacteria” might have diverged as early as 3 Gyr. Bacteria of the genus Azospirillum are associated with roots of terrestrial plants; however, virtually all their close relatives are aquatic. We obtained genome sequences of two Azospirillum species and analyzed their gene origins. While most Azospirillum house-keeping genes have orthologs in its close aquatic relatives, this lineage has obtained nearly half of its genome from terrestrial organisms. The majority of genes encoding functions critical for association with plants are among horizontally transferred genes. Our results show that transition of some aquatic bacteria to terrestrial habitats occurred much later than the suggested initial divergence of hydro- and terrabacterial clades. The birth of the genus Azospirillum approximately coincided with the emergence of vascular plants on land.
| Genome sequencing and analysis of plant-associated beneficial soil bacteria Azospirillum spp. reveals that these organisms transitioned from aquatic to terrestrial environments significantly later than the suggested major Precambrian divergence of aquatic and terrestrial bacteria. Separation of Azospirillum from their close aquatic relatives coincided with the emergence of vascular plants on land. Nearly half of the Azospirillum genome has been acquired horizontally, from distantly related terrestrial bacteria. The majority of horizontally acquired genes encode functions that are critical for adaptation to the rhizosphere and interaction with host plants.
| Fossil records indicate that life appeared in marine environments ∼3.5–3.8 billion years ago (Gyr) [1] and transitioned to terrestrial ecosystems ∼2.6 Gyr [2]. The lack of fossil records for bacteria makes it difficult to assess the timing of their transition to terrestrial environments; however sequence analysis suggests that a large clade of prokaryotic phyla (termed “terrabacteria”) might have evolved on land as early as 3 Gyr, with some lineages later reinvading marine habitats [3]. For example, cyanobacteria belong to the terrabacterial clade, but one of its well-studied representatives, Prochlorococcus, is the dominant primary producer in the oceans [4].
Bacteria of the genus Azospirillum are found primarily in terrestrial habitats, where they colonize roots of important cereals and other grasses and promote plant growth by several mechanisms including nitrogen fixation and phytohormone secretion [5], [6]. Azospirillum belong to proteobacteria, one of the largest groups of “hydrobacteria”, a clade of prokaryotes that originated in marine environments [3]. Nearly all known representatives of its family Rhodospirillaceae are found in aquatic habitats (Figure 1 and Table S1) suggesting that Azospirillum represents a lineage which might have transitioned to terrestrial environments much later than the Precambrian split of “hydrobacteria” and “terrabacteria”. To obtain insight into how bacteria transitioned from marine to terrestrial environments, we sequenced two well studied species, A. brasilense and A. lipoferum, and a third genome of an undefined Azospirillum species became available while we were carrying out this work [7].
In contrast to the genomes of their closest relatives (other Rhodospirillaceae), the three Azospirillum genomes are larger and are comprised of not one, but seven replicons each (Figure S1 and Table 1). Multiple replicons have been previously suggested for various Azospirillum strains [8]. The largest replicon in each genome has all characteristics of a bacterial chromosome, whereas the smallest is a plasmid. Five replicons in the genomes of A. lipoferum and Azospirillum Sp. 510 can be defined as “chromids” (intermediates between chromosomes and plasmids [9]), whereas in A. brasilense only three replicons are “chromids” (Tables S2 and S3). While multiple replicons, and chromids specifically, are not unusual in proteobacteria [9], [10], Azospirillum lipoferum has the largest number of chromids among all prokaryotes sequenced to date [9] indicating a potential for genome plasticity.
Comparisons among the three genomes reveal further evidence of extraordinary genome plasticity in Azospirillum, a feature that has also been suggested by some experimental data [11]. We found very little synteny between replicons of Azospirillum species. The genetic relatedness among Azospirillum strains is comparable to that of rhizobia, other multi-replicon alpha-proteobacteria (Table S4). Surprisingly, we found substantially more genomic rearrangement within Azospirillum genomes than within rhizobial genomes (Figure 2) that are suggested to exemplify genome plasticity in prokaryotes [10]. This could be a consequence of many repetitive sequences and other recombination hotspots (Tables S4 and S5), although the detailed mechanisms underlying such extraordinary genome plasticity remain incompletely understood.
Which genes does Azospirillum share with its aquatic relatives, and what is the origin of its additional genes? To answer this question, we developed a robust scheme for detecting ancestral and horizontally transferred (HGT) genes (Figure 3) using bioinformatics tools, then classified most protein coding genes in the Azospirillum genomes as ancestral or horizontally transferred with quantified degrees of confidence (Figure 4A and Table S6). Remarkably, nearly half of the genes in each Azospirillum genome whose origins can be resolved appeared to be horizontally transferred. As a control, we subjected the genomes of other Rhodospirillaceae to the same analysis, finding a substantially lower HGT level in aquatic species, while the number of ancestral genes in all organisms was comparable (Figure 4B). Horizontally transferred genes are frequently expendable, whereas ancestral genes usually serve ‘house-keeping’ functions and are conserved over long evolutionary distances [12]. To further validate our classifications, we determined functional assignments of genes in each of the two categories using the COG database [13]. The ‘ancestral’ set primarily contained genes involved in ‘house-keeping’ functions such as translation, posttranslational modification, cell division, and nucleotide and coenzyme metabolism (Figure 5). The HGT set contained a large proportion of genes involved in specific dispensable functions, such as defense mechanisms, cell wall biogenesis, transport and metabolism of amino acids, carbohydrates, inorganic ions and secondary metabolites (Figure 5 and Table S6). This is consistent with the role of HGT in adaptation to the rhizosphere, an environment rich in amino acids, carbohydrates, inorganic ions and secondary metabolites excreted by plant roots [14].
Such an extraordinary high level of HGT in Azospirillum genomes leads us to hypothesize that it was a major driving force in the transition of these bacteria from aquatic to terrestrial environments and adaptation to plant hosts. This process was likely promoted by conjugation and transduction, as Azospirillum hosts phages and notably a Gene Transfer Agent [15]; this should have also resulted in loss of ancestral ‘aquatic’ genes that are not useful in the new habitat. Indeed, one of the defining features of Rhodospirillaceae, photosynthesis (responsible for the original taxonomic naming of these organisms – purple bacteria) is completely absent from Azospirillum. We have analyzed origins of genes that are proposed to be important for adaptation to the rhizosphere and interactions with the host plant [6], [16]. Consistent with our hypothesis, the majority of these genes were predicted to be horizontally transferred (Figure 6 and Table S7). It is important however to stress that plant-microbe interactions involve a complex interplay of many functions that are determined by both ancestral and horizontally acquired genes.
What was the source of horizontally transferred genes? A large proportion of genes that we assigned as HGT show relatedness to terrestrial proteobacteria, including representatives of Rhizobiales (distantly related alpha-proteobacteria) and Burkholderiales (beta-proteobacteria) (Figure 7) that are soil and plant-associated organisms. In the absence of fossil data, it is nearly impossible to determine the time of divergence for a specific bacterial lineage; however, a rough approximation (1–2% divergence in the 16S rRNA gene equals 50 Myr [17]) suggests that azospirilla might have diverged from their aquatic Rhodospirillaceae relatives 200–400 Myr (Table S8). This upper time limit coincides with the initial major radiation of vascular plants on land and evolution of plant roots, to 400 Myr [18], [19]. Grasses, the main plant host for Azospirillum, appeared much later, about 65–80 Myr [20], which is consistent with reports that azospirilla can also colonize plants other than grasses.
Using a global proteomics approach we have found that many HGT genes including nearly 1/3 of those that are common to all three Azospirillum genomes were expressed under standard experimental conditions and under nitrogen limitation, a condition usually encountered in the rhizosphere of natural ecosystems (Figure 8 and Table S9).
Genes that differentiated the Azospirillum species from one another and from their closest relatives are implicated in specializations, such as plant colonization. Azospirillum and closely related Rhodospirillum centenum both possess multiple chemotaxis operons and are model organisms to study chemotaxis [21], [22]. Interestingly, operon 1, which controls chemotaxis in R. centenum [22], plays only a minor role in chemotaxis of A. brasilense [23]. All three Azospirillum species possess three chemotaxis operons that are orthologous to those in R. centenum; however, they also have additional chemotaxis operons that are absent from their close aquatic relative (Figure S2 and Tables S6 and S10). Additional chemotaxis operons have been acquired by azospirilla prior to each speciation event yielding 4, 5 and 6 chemotaxis systems in A. brasilense Sp245, A. lipoferum 4B and Azospirillum sp. 510, respectively. These stepwise acquisitions have made the latter organism an absolute “chemotaxis champion”, with 128 chemotaxis genes, more than any other prokaryote sequenced to date (data from MiST database [24]). Recent analysis showed the prevalence of chemotaxis genes in the rhizosphere [25]. We have determined that the dominant chemotaxis genes in this dataset belong to a specific chemotaxis class F7 [26] (Figure S3 and Table S11). Strikingly, it is this F7 system that is shared by all Azospirillum and is predicted to have been transferred horizontally into their common ancestor.
Cellulolytic activity may be crucial to the ability of some azospirilla to penetrate plant roots [27]. All Azospirillum genomes encode a substantial number of glycosyl hydrolases that are essential for decomposition of plant cell walls (Figure 9). The total number of putative cellulases and hemicellulases in azospirilla is comparable to that in soil cellulolytic bacteria (Table S12) and most of them are predicted to be acquired horizontally (Table S6). We tested three Azospirillum species and found detectable cellulolytic activity in A. brasilense Sp245 (Figure 10). The A. brasilense Sp245 genome contains three enzymes encoded by AZOBR_p470008, AZOBR_p1110164 and AZOBR_150049 (Figure 11) that are orthologous to biochemically verified cellulases. We propose that these and other horizontally transferred genes (e.g. glucuronate isomerase, which is involved in pectin decomposition) contributed to establishing A. brasilense Sp245 as a successful endophyte [27]. Interestingly, another successful endophytic bacterium, Herbaspirillum seropedicae, lacks the genes coding for plant cell wall degradation enzymes [28] indicating that endophytes may use very different strategies for penetrating the plant.
Attachment, another function important for plant association by Azospirillum, was also acquired horizontally. Type IV pili is a universal feature for initiating and maintaining contact with the plant host [29], [30]. The genome of A. brasilense Sp245 lacks genes coding for Type IV pili, but encodes a set of genes for TAD (tight adhesion) pili that are known to be HGT prone [31]. In our analysis, TAD pili were confidently predicted to be a result of HGT (Table S6). We show that a mutant deficient in TAD pili had a severe defect in attachment and biofilm formation (Figure 12) suggesting a role for TAD in plant-microbe association.
Horizontal gene transfer has been long recognized as a major evolutionary force in prokaryotes [12]. Its role in the emergence of new pathogens and adaptation to environmental changes is well documented [32]. While other recent studies indicate that HGT levels in natural environments may reach as much as 20% of a bacterial genome [33], our data suggest that HGT has affected nearly 50% of the Azospirillum genomes, in close association with dramatic changes in lifestyle necessary for transition from aquatic to terrestrial environments and association with plants. Emergence of these globally distributed plant-associated bacteria, which appear to coincide with radiation of land plants and root development, likely has dramatically changed the soil ecosystem.
The genome of Azospirillum lipoferum 4B was sequenced by the whole random shotgun method with a mixture of ∼12X coverage of Sanger reads, obtained from three different libraries, and ∼18X coverage of 454 reads. Two plasmid libraries of 3 kb (A) and 10 kb (B), obtained by mechanical shearing with a Hydroshear device (GeneMachines, San Carlos, California, USA), were constructed at Genoscope (Evry, France) into pcDNA2.1 (Invitrogen) and into the pCNS home vector (pSU18 modified, Bartolome et al.[34]), respectively. Large inserts (40 kb) (C) were introduced into the PmlI site of pCC1FOS. Sequencing with vector-based primers was carried out using the ABI 3730 Applera Sequencer. A total of 95904 (A), 35520 (B) and 15360 (C) reads were analysed and assembled with 504591 reads obtained with Genome Sequencer FLX (Roche Applied Science). The Arachne “HybridAssemble” version (Broad institute, MA) combining 454 contigs with Sanger reads was used for assembly. To validate the assembly, the Mekano interface (Genoscope), based on visualization of clone links inside and between contigs, was used to check the clones coverage and misassemblies. In addition, the consensus was confirmed using Consed functionalities (www.phrap.org), notably the consensus quality and the high quality discrepancies. The finishing step was achieved by PCR, primer walks and transposon bomb libraries and a total of 5460 sequences (58, 602 and 4800 respectively) were needed for gap closure and quality assessment.
The genome of strain Azospirillum brasilense Sp245 was sequenced by the whole random shotgun method with a mixture of ∼10X coverage of Sanger reads obtained from three different libraries and ∼25X coverage of 454 reads. A plasmid library of 3 kb, obtained by mechanical shearing with a Hydroshear device (GeneMachines, San Carlos, California, USA), were constructed at Plant Genome Mapping Laboratory (University of Georgia, USA) into pcDNA2.1 vector (Invitrogen). Large inserts (40 kb) were introduced into the PmlI site of pCC1FOS. Sequencing with vector-based primers was carried out using the ABI 3730 Applera Sequencer. The Arachne “HybridAssemble” version combining 454 contigs with Sanger reads was used for assembly. Contig scaffolds were created using Sequencher (Gene Codes) and validated using clone link inside and between contigs.
AMIGene software [35] was used to predict coding sequences (CDSs) that were submitted to automatic functional annotation [36]. The resulting 6233 A. lipoferum 4B CDSs and 7848 A. brasilense Sp245 CDSs were assigned a unique identifier prefixed with “AZOLI” or “AZOBR” according to their respective genomes. Putative orthologs and synteny groups were computed between the sequenced genomes and 650 other complete genomes downloaded from the RefSeq database (NCBI) using the procedure described in Vallenet et al. [36]. Manual validation of the automatic annotation was performed using the MaGe (Magnifying Genomes) interface. IS finder (www-is.biotoul.fr) was used to annotate insertion sequences [37]. The A. lipoferum 4B nucleotide sequence and annotation data have been deposited to EMBL databank under accession numbers: FQ311868 (chromosome), FQ311869 (p1), FQ311870 (p2), FQ311871 (p3), FQ311872 (p4), FQ311873 (p5), FQ311874 (p6). The A. brasilense Sp245 nucleotide sequence and annotation data have been deposited at EMBL databank under accession numbers: HE577327 (chromosome), HE577328 (p1), HE577329 (p2), HE577330 (p3), HE577331 (p4), HE577332 (p5), HE577333 (p6). In addition, all the data (i.e., syntactic and functional annotations, and results of comparative analysis) were stored in a relational database, called AzospirilluScope [36], which is publicly available at http://www.genoscope.cns.fr/agc/mage/microscope/about/collabprojects.php?P_id=39.
BLAST searches were performed using NCBI toolkit version 2.2.24+ [38]. Multiple sequence alignments were built using the L-INS-i algorithm of MAFFT [39] with default parameters. Phylogenetic tree construction was performed using PhyML [40] with default parameters unless otherwise specified. 16S rRNA sequences were retrieved from the Ribosomal Database Project [41].
A concatenated ribosomal protein tree was constructed from sequenced members of alpha-proteobacteria with a 98% 16S rRNA sequence identity cutoff to limit overrepresentation. The following ribosomal proteins were used: L3, L5, L11, L13, L14, S3, S7, S9, S11, and S17. The proteins were identified using corresponding Pfam models and HMMER [42] searches against the genomes of sequenced alpha-proteobacteria selected above. The sequences were aligned and concatenated. GBlocks [43] with default parameters was used to reduce the number of low information columns. The tree was constructed using PhyML with the following options: empirical amino acid frequencies, 4 substitution categories, estimated gamma distribution parameter, and NNI tree topology search.
Protein sequences queries from all 3 Azospirillum genomes were used in BLAST searches against the non-redundant microbial genome set constructed by Wuichet and Zhulin [26] supplemented with sequenced members of Rhodospirillales absent in the original set (Acetobacter pasteurianus IFO 3283-01, alpha proteobacterium BAL199, Magnetospirillum gryphiswaldense MSR-1, and Magnetospirillum magnetotacticum MS-1). E-value cutoff of 10∧−4 was used.
Only the first occurrence of each species was used in ancestry assignment. Proteins were assigned as being ancestral or horizontally transferred, with varying degrees of confidence, based on the presence of members of Rhodospirillales and Rhodospirillaceae in the top eight BLAST hits. Ancestral assignment was based on the top 8 hits, based on the number of Rhodospirillaceae genomes in the database: 2 Azospirillum, 3 Magnetospirillum, 2 Rhodospirillum, and Nisaea sp. BAL199, excluding the organism on which ancestry assignment is being performed. High confidence ancestral proteins have at least 6 of the top 8 species belonging to Rhodospirillales or all but 1, if the BLAST result had less than 8 species. This rule allows for 1–2 independent events of HGT from Rhodospirillales to other distantly related species. Medium confidence ancestral proteins have at least 4 Rhodospirillaceae in the top 8. Low confidence ancestral proteins have at least 1 Rhodospirillaceae in the top 8, excluding hits to other Azospirillum genomes. High confidence horizontally transferred proteins have 0 hits to Rhodospirillales in the top 10, excluding hits to other Azospirillum genomes. Medium confidence horizontally transferred proteins have 0 hits to Rhodospirillales in the top 5, excluding hits to other Azospirillum genomes. Low confidence horizontally transferred proteins have 0 hits to Rhodospirillaceae in the top 8, excluding hits to other Azospirillum genomes. Unassigned proteins either have no BLAST hits outside Azospirillum, or simultaneously classify as medium confidence horizontally transferred and medium or low confidence ancestral.
Bidirectional BLAST was used to identify orthologs of the putative glycoside hydrolase (GH) genes. Phyml package was used to confirm evolutionary relationships and visualize the results. Domain architectures were obtained through Pfam [53] search for each protein. Then information from CAZy [54] and recent analysis [55] was used to assign putative activities of the predicted GHs.
Chemotaxis proteins were identified in genomic datasets as previously described [56]. Using CheA sequences from a recent chemotaxis system classification analysis [26], alignments of the P3–P5 regions of CheA were built for each class and for the entire set of CheA sequences. Each alignment was made non-redundant so that no pair of sequences shared more than 80% sequence identity. Hidden Markov Models (HMMs) were built from each non-redundant alignment and used to create library via the HMMER3 software package (version HMMER 3.0b3) [42] and default parameters.
The rhizosphere CheA sequences from a recent study [25] were run against the CheA HMM library. Unclassified sequences (Unc) are those with top hits to the full CheA set HMM rather than a class-specific HMM. The remaining sequences were assigned to the class of the top scoring HMM.
Azospirillum strains and control strains (Dickeya dadantii 3937 as a positive control, A. tumefaciens NT1 as a negative control) were cultured for 16 h in liquid AB minimal medium [57] containing 0.2% malate and 1 mg/L biotin. An aliquot of 107 cells (for Dickeya dadantii 3937) or 2.107 cells (for all other strains) was deposited on top of AB plates containing 0.1% carboxymethylcellulose instead of malate. Plates were incubated for 5 days before being stained as previously described [58].
A 211-bp cpaB (AZOBR_p460079) internal fragment was amplified by PCR with primers F6678 (GCGTGGACCTGATCCTGAC) and F6679 (GTGACCGTCTCGCTCTGAC) and subcloned into pGEM-T easy (Promega). White colonies were screened by PCR with primers F6678 and F6679 for correct insertion in pGEM-T easy, resulting in pR3.37. The insert of plasmid pR3.37 was digested with NotI and cloned into the NotI site of pKNOCK-Km [59], resulting in pR3.39 after transfer into chemically-competent cells of E. coli S17.1 λpir. pR3.39 was introduced into A. brasilense Sp245 by biparental mating. Transconjugants resulting from a single recombination event of pR3.39 were selected on AB medium containing 0.2% malate, ampicillin (100 mg/mL) and kanamycin (40 mg/mL). The correct insertion of pKNOCK into cpaB was confirmed by PCR with primers (F6678 and F5595 TGTCCAGATAGCCCAGTAGC, located on pKNOCK) and sequencing of the PCR amplicon.
Sp245 and Sp245cpaB were labelled with pMP2444 [60] allowing the constitutive expression of EGFP. The strains were grown in NFB* (Nitrogen free broth containing 0.025% of LB) with appropriate antibiotics in glass tubes containing a cover-slide, under a mild lateral agitation for 6 days. After the incubation, the liquid and the cover-slide were removed from the tubes and the biofilm formed at the air/liquid interface was colored by 0.1% crystal violet. After two washings with distilled water, crystal violet was solubilized by ethanol and quantified by spectrophotometry at 570 nm. The experiment was performed twice in triplicate. In parallel, the colonization of the glass cover-slide was monitored by confocal laser scanning microscopy (510 Meta microscope; Carl Zeiss S.A.S.) equipped with an argon-krypton laser, detectors, and filter sets for green fluorescence (i.e., 488 nm for excitation and 510 to 531 nm for detection). Series of horizontal (x-y) optical sections with a thickness of 1 µm were taken throughout the full length of the Sp245 and Sp245cpaB biofilms. Three dimensional reconstructions of biofilms were performed using LSM software release 3.5 (Carl Zeiss S.A.S.).
|
10.1371/journal.pgen.1002402 | Genome-Wide Meta-Analysis of Five Asian Cohorts Identifies PDGFRA as a Susceptibility Locus for Corneal Astigmatism | Corneal astigmatism refers to refractive abnormalities and irregularities in the curvature of the cornea, and this interferes with light being accurately focused at a single point in the eye. This ametropic condition is highly prevalent, influences visual acuity, and is a highly heritable trait. There is currently a paucity of research in the genetic etiology of corneal astigmatism. Here we report the results from five genome-wide association studies of corneal astigmatism across three Asian populations, with an initial discovery set of 4,254 Chinese and Malay individuals consisting of 2,249 cases and 2,005 controls. Replication was obtained from three surveys comprising of 2,139 Indians, an additional 929 Chinese children, and an independent 397 Chinese family trios. Variants in PDGFRA on chromosome 4q12 (lead SNP: rs7677751, allelic odds ratio = 1.26 (95% CI: 1.16–1.36), Pmeta = 7.87×10−9) were identified to be significantly associated with corneal astigmatism, exhibiting consistent effect sizes across all five cohorts. This highlights the potential role of variants in PDGFRA in the genetic etiology of corneal astigmatism across diverse Asian populations.
| Corneal astigmatism is associated with reduced visual acuity and an increased risk of developing refractive amblyopia. Although it is highly heritable, there is no prior study on the genetic etiology of corneal astigmatism. Our genome-wide meta-analysis across 8,513 individuals in five genome-wide surveys from three genetically diverse populations in Asia reveals that genetic variants in the PDGFRA gene on chromosome 4q12 is significantly associated with corneal astigmatism. These polymorphisms in the PDGFRA gene exhibit strong and consistent effects over all five Asian cohorts. PDGFRA is a receptor for platelet-derived growth factor, which is expressed in many retinal tissues in the eyes and appears to contribute to ocular development. Results from our study further suggest the potential role of PDGFRA in the regulation of corneal biometrics.
| Astigmatism is a condition where light rays are prevented from focusing at a single point in the eye, resulting in blurred vision at any near or far distance. While astigmatism comprises cornea and non-corneal components, it typically results from the unequal curvature of two principle meridians in the anterior surface of the cornea known as corneal astigmatism [1], [2]. The presence of a high degree of astigmatism during early development is believed to be associated with refractive amblyopia [3], [4], [5], as evidenced by decreased best-corrected visual acuity which cannot be remedied by external corrective lenses. Early abnormal visual input caused by uncorrected astigmatism can lead to orientation-dependent visual deficits, despite optical correction of visual acuity later in life [6]. In addition, it has been suggested that optical blurring by astigmatism may predispose the development of myopia, commonly known as nearsightedness [7], [8], [9], [10].
Astigmatism is highly prevalent across most populations and poses a significant burden to global public health with at least 1 in 3 adults above 30 years of age suffering from astigmatism of −0.5 diopters (D) or less [11]. The reported age-adjusted prevalence of astigmatism was 37.8% for Chinese adults [12], 54.8% in rural Asian Indians [13], 37% (≤−0.75D) for Caucasian in Australia [14] and 36.2% in the US [15]. The prevalence of astigmatism in children varies considerably across different studies and ethnic groups. For instance, the prevalence of astigmatism (≤−0.75D) in school-children ranges from 13.6% in Australia [16], 20% in Northern Ireland [17], 28.4% for Singapore school children [8], to 42.7% for Chinese children in urban China [18].
Although the precise cause of astigmatism is unknown, genetic factors have been implicated in the etiology of corneal astigmatism. Studies have reported a higher risk of developing astigmatism in individuals whose sibling or parents have astigmatism [11]. Evidence from twin studies suggests a genetic etiology in astigmatism development, with the estimated heritability ranging from 30% to 60% [19], [20], [21], [22], [23]. For instance, Hammond and colleagues [21] investigated the inheritance of astigmatism for 226 monzygotic (MZ) and 280 dizygotic (DZ) twins in the United Kingdom and found genetic effects accounted for 42% to 61% of the variation in corneal astigmatism. While most of the twin studies have been conducted in Caucasian populations, a study on Chinese twins in Taiwan reported a heritability estimate of 46% for corneal astigmatism, suggesting that genetic factors account for a similar extent in the etiology of astigmatism for Asian populations [22]. However, no genetic loci have been systematically and consistently identified to be implicated in the development of corneal astigmatism.
Here we report the findings from the meta-analyses of five genome-wide association studies (GWAS) performed in 8,513 individuals from three Asian populations. The discovery phase of our study comprises 4,254 individuals from two population-based GWAS performed in adults of Chinese and Malay ethnicities from the Singapore Prospective Study Program (SP2) and the Singapore Malay Eye Study (SiMES) respectively. The replication phase comprises of data from three other genome-wide surveys of: (i) 2,139 Indian adults from the Singapore Indian Eye Study (SINDI); (ii) 929 Chinese school children from the Singapore Cohort Study of the Risk Factors for Myopia (SCORM); and (iii) 397 Chinese trios of parents and astigmatic offsprings from the Singaporean Chinese in the Strabismus, Amblyopia and Refractive Error Study (STARS).
The characteristics of the post-QC samples from the five studies are summarized in Table 1. The post-QC SP2 dataset comprised 2,016 adults, of which 1,231 individuals had corneal astigmatism (≤−0.75 D) and 785 subjects were defined as non-astigmatic controls. The post-QC SiMES dataset comprised 2,238 adults (1,018 cases and 1,220 controls). In total, 462,518 and 515,712 autosomal genotyped SNPs passed stringent quality control criteria for SP2 and SiMES respectively and the genome-wide meta-analysis was conducted on 460,528 SNPs present in both studies.
There was no evidence of over-inflation of statistical significances due to population structure in either of the discovery cohorts (SP2 λGC = 1.006, SiMES λGC = 1.007) or in the meta-analysis of both studies (overall λGC = 1.007). Suggestive evidence of association (defined as 10−6<P-value<10−5) were seen in each of SP2 and SiMES (Figure S1A and S1B), as well as in the meta-analysis of SP2 and SiMES where a collection of SNPs deviated from their expected distributions in the quantile-quantile plots of the P-values (Figure S1C).
None of the SNPs in the discovery meta-analysis attained genome-wide significance of P-value<5×10−8. Seven SNPs exhibited evidence stronger than P-value<1.0×10−5 and these were found to cluster in the platelet-derived growth factor receptor alpha (PDGFRA) gene on chromosome 4q12 (lowest P = 9.44×10−7 at rs7677751; Table 2; Figure S2). Interestingly, these SNPs are located within the MYP9 region identified previously as a candidate locus for myopia through linkage scans [24].
In the replication phase with the three additional GWAS cohorts, three SNPs in PDGFRA (rs7677751, rs2307049 and rs7660560) attained genome-wide significance in the combined analysis (Table 2) with the lead SNP rs7677751 from the discovery phase remaining as the strongest signal in the combined analysis (P = 7.87×10−9; Figure 1). All seven SNPs from the discovery phase exhibited P-values<0.05 in SINDI but not in SCORM or STARS. However the direction and magnitude of the effect sizes at these seven SNPs in all three replication cohorts were highly similar to those seen in the discovery populations of SP2 and SiMES (Table 2, Figure 2). No significant evidence of effect size heterogeneity was detected across the SNPs (heterogeneity I2 P-value≥0.246), and the minor allele frequencies of these SNPs are consistently similar across all five studies (Table S1). A genome-wide meta-analysis of the combined five cohorts did not yield any additional locus with genome-wide significance (see Figure S3 for QQ and Manhattan plots, λGC = 1.002; Table S2).
At the lead SNP rs7677751 in PDGFRA, the frequency of the risk T-allele ranged from 0.19 to 0.26 in the five cohorts and conferred a 26% higher risk of corneal astigmatism than the C allele (OR = 1.26, 95% CI = 1.16–1.36) in the meta-analysis across all five studies. This SNP alone explains 0.41% of the variation in corneal cylinder power. In addition, a general genetic model identified that the 5.5% of the individuals in the combined cohorts that carry the TT genotype at rs7677751 had a 1.65-fold (95%CI = 1.33–2.06, P-value = 6.23×10−6) increase in the risk of developing corneal astigmatism compared to those that are not carrying any copies of the risk allele (Figure S4). All of the associated SNPs spanned 10 kb within PDGFRA at 4q12 (Figure 2), and a high degree of linkage disequilibrium is seen at this locus in all three Asian populations (Chinese, Malays and Asian Indians; Figure S5).
We have performed a genome-wide survey for corneal astigmatism across 8,513 individuals, where the discovery phase combines the data from two GWAS performed in Chinese and Malay adults, and the replication phase included Asian Indian adults, Chinese children and family trios. We observed a strong and consistent association with the onset of corneal astigmatism at the PDGFRA gene locus on chromosome 4q12 across all five Asian cohorts, with three SNPs in this locus exhibiting evidence stronger than genome-wide significance in the meta-analysis. To the best of our knowledge, this is the first GWAS to investigate the genetic etiology of corneal astigmatism in a genome-wide fashion.
The PDGFRA gene spans 69 kb with 23 coding exons and resides on chromosome 4q12. The receptor for platelet-derived growth factor (PDGF) contains two types of subunit, a- and β- PDGFRA, which are differentially expressed on the cell surface [25]. PDGFR-a binds to three forms of PDGF (PDGF-AA, AB and BB) and mediates many biological process including embryonic development, angiogenesis, cell proliferation and differentiation. The role of PDGFRA in cellular growth and proliferation is underlined by its contribution to the pathogenesis of gastrointestinal stromal tumours [26]. A large body of evidence has shown that both PDGF and its receptors are expressed in corneal epithelium, stromal fibroblasts and endothelium [27], [28] as well as proliferative retinal tissues in eyes [29], [30], [31]. Along with other cytokines (epidermal growth factor, transforming growth factor-a,-β etc), studies have further suggested that PDGF and its receptors can mediate corneal fibroblast migration, matrix remodeling and play an important role in corneal wound healing [28], [32], [33], [34]. The corneal stroma comprises a large portion of the cornea; the sensitivity of stromal tissue to various growth factors is well described [35]. The administration of PDGF resulted in keratinocyte elongation using rabbit corneal stroma tissue [36]. In light of this, a role for PDGFRA in the regulation of ocular development and parameters cannot be excluded, and our study suggests that genetic polymorphisms within PDGFRA may be involved in the regulation of corneal biometrics resulting in the occurrence of corneal astigmatism.
In addition, Hammond et al. reported that 4q12 (MYP9; LOD 3.3) was significant linked with myopia from a genome-wide linkage study of 221 dizygotic twin pairs [24], and subsequent replication revealed nominal significance of 4q12 (P = 0.065) for refractive error in African-American families [37]. We thus undertook a candidate SNP approach with the identified SNPs to investigate the possible association between PDGFRA and (i) the onset of high myopia; (ii) the refractive error as a quantitative trait. We did not observe any striking association between the identified variants with either outcomes, suggesting that the association of PDGFRA with corneal astigmatism is probably not through any shared etiology with myopia.
The lead SNP in our analyses rs7677751 is located in the intro 1 of PDGFRA. Interestingly, among the SNPs identified, rs2228230 is coding-synonymous (valine:GTC>valine:GTT) and resides in exon 18, while rs3690 is within the untranslated-3′ region. These three SNPs (rs7677751, rs2228230 and rs3690) are strongly correlated with each other (pair-wise Pearson correlation coefficient r ranging from 0.77 to 0.81), although the association evidence at the latter two SNPs did not reach genome-wide significance. As the next closest gene (GSX2) from the 5′ end of PDGFRA is 127 kb away and is not within the LD block with our identified SNPs (Figure 1), it is unlikely that the signals observed in our study are attributed to functional variants located beyond PDGFRA.
Our group recently reported a strong association between variants in PDGFRA with corneal curvature [38]. Corneal curvature is an ocular dimension defined as the average of the radius of corneal curvature at the horizontal and vertical meridians. Myopic eyes have been found to have steeper corneas (reduced radius of curvature), but the significant correlation between corneal curvature and refractive error was not consistently observed [39], [40], [41]. Excessively flatter cornea is associated with cornea plana, producing high hyprotropia and likely resulting in angle-closure glaucoma [42], [43]. On the other hand, corneal astigmatism is an eye-disorder, where the cornea is more curved in one meridian direction compared to the other. This fragmentizes the light rays entering the eye, leading to the inability to focus onto a single point in the eye [1]. It is thus interesting that the same PDGFRA gene has been identified in two ocular outcomes that are biologically different, given the presence of a weak correlation between corneal astigmatism and corneal curvature (Spearman correlation coefficient r between 0.088 and 0.192 in our cohorts; Figure S6), pointing to a possible pleiotropic contribution of PDGFRA.
Our study has adopted a binary definition of corneal astigmatism (affected and unaffected) that is commonly adopted in clinical practice and eye-trait epidemiology [16], [44]. One caveat of this definition is the potential for misclassifying the affected status, particularly for samples with cylinder power around the cutoff threshold of −0.75D. To evaluate the robustness of our findings to the choice of threshold used, we additionally performed the association analysis at the identified SNPs with four different combinations of the thresholds used to define cases and controls. We observed that the odds ratios were highly similar across all four scenarios (Table S3), with the combined evidence at rs7677751 ranging from Pmeta of 1.5×10−4 to 6.7×10−8. Unsurprisingly, the association evidence was weakest in the scenario with the most stringent thresholding (≤−1.5D for cases and >−0.5 for controls), given this stringency comes at the expense of decreasing the number of individuals in each study. We additionally performed a secondary analysis treating corneal cylinder power as a quantitative trait. Strong statistical evidence was consistently observed at the three leading SNPs (rs7677751, P = 1.76×10−7; rs2307049, P = 3.41×10−7 and rs7660560, P = 4.41×10−7; Table S4), indicating that our findings are robust to the definition of the phenotype.
Owing to the relatively small sample sizes within each of the five GWAS studies, we have chosen to prioritize our survey to identify genetic variants that contribute to the etiology of corneal astigmatism in multiple Asian populations. While Malays have been observed to be genetically closer to the Chinese, the Asian Indians tend to be genetically closer to the Caucasians [45]. Our discovery at PDGFRA thus suggests that part of the underlying biological pathway responsible for astigmatism development is common to multiple populations, although there may be population-specific genetic variants that our current study is not sufficiently powered to identify.
Our study has included two pediatric Chinese populations (SCORM and STARS) with school or pre-school children who are still progressing to their final phenotype. It was documented that a high degree of astigmatism occurs during infancy and the early childhood [46]. The prevalence rates remain stable during young adulthood (18 to 40 years), but increase consistently during late adulthood at aged 40 years or older [1], [12]. Studies have also indicated that the age-related change in astigmatism is associated with meridians changes in the cornea [11]. Children and adolescents have a predominance of “within-the-rule” corneal astigmatism in general, where the vertical curve is greater than the horizontal (axis of 1° to 15°); while in older adults, it shifts to “against-the-rule” astigmatism (axis of 75°–105°) [47], [48]. However, our study considers corneal astigmatism without reference to the axis nor the longitudinal changes from children to adults. Whether PDGFRA plays the same role in pediatric and adults populations will however need further investigation.
This study adhere to the Declaration of Helsinki. Ethics approvals have been obtained from the Institutional Review Boards of the Singapore Eye Research Institute, Singapore General hospital, National University of Singapore and National Healthcare Group, Singapore. In all cohorts, participants provided written, informed consent at the recruitment into the studies. For studies involving children who were still minors (SCORM and STARS), written informed consent was obtained from the children's parents.
All studies used a similar protocol to measure ocular phenotypes including corneal curvature, autorefraction and cylinder power by a team of eye care professionals. Participants in SP2, SIMES and SINDI underwent non-cycloplegic automated refractive assessments using the autorefractor (Canon RK-5, Tokyo, Japan). For SCORM and STARS, cycloplegic measurements (Canon RK-F1, Tokyo, Japan) were performed 30 minutes after three drops of 1% cyclopentolate which were administered 5 minutes apart.
Corneal curvature radii in the horizontal and vertical meridian were determined with keratometry in millimeters [60]. The keratometer measured the anterior corneal surface and used a refractive index of 1.3375 to account for the contribution from the posterior corneal surface to derive the corneal refractive power in diopters. Corneal cylinder power was calculated as the difference between the flattest and steepest meridian of the keratometry readings in diopters of power.
As the corneal cylinder power between the right and left eyes are strongly correlated across all five cohorts (Pearson's correlation coefficient r ranging from 0.51 to 0.79; P<2.2×10−16), the mean corneal cylinder power over both eyes was used to define corneal astigmatism. Averaging ocular measurements between two eyes in genetic studies has been suggested to be statistically more powerful than using the information from only one eye [61], and this approach has been consistently adopted in genome-wide studies of myopia [62], [63]. As with previous studies [16], [44], we have defined individuals with average corneal cylinder power ≤−0.75D as cases, and those with average corneal power between −0.75D and 0D as controls.
For SP2, a total of 2,867 blood-derived DNA samples were genotyped using the Illumina Human610 Quad and 1Mduov3 Beadchips. For the samples that were genotyped on the two platforms, the genotypes from the denser platform were used in our study. For SiMES (n = 3,072), SINDI (n = 2,593) and STARS (n = 1,451), the Illumina Human610 Quad Beadchips was used for genotyping all DNA samples. For the 1,116 SCORM children, DNA samples were genotyped on the Illumina HumanHap 550 Duo Beadchips.
Detailed data quality control (QC) procedures for each study were provided in the supplementary information (Text S1). In brief, for case-control study design, QC criteria included a first round for autosomes SNP QC to obtain a cleaned set of genotypes for sample QC, by excluding SNPs with: (i) missingness (per-SNP call rate) >5%; (ii) minor allele frequency (MAF) <1%; and (iii) HWE p-value<10−7. Using the subset of SNPs passing the first round QC, samples were then excluded based on the following conditions: (i) per-sample call rates of less than 95%; (ii) excessive heterozygosity (defined as the sample heterozygosity to be beyond 3 standard deviations from the mean sample heterozygosity); (iii) cryptic relatedness; (iv) gender discrepancies; and (v) deviation in population membership from population structure analysis. A second round of SNP QC was then applied to the remaining samples passing quality checks. We excluded SNPs with missingness >5%, gross departure from HWE (P value<10−6), MAF<1% and low concordance between duplicate samples on different genotype platforms (relevant to SP2 samples only).
Population structure was ascertained using principal components analyses (PCA) with the EIGENSTRAT program [64]. Population substructure of SP2 and SiMES was examined by PCA with respect to three population panels in the HapMap samples (Figure S7). Due to the presence of population structure within the Malay and Indian samples (Figures S8 and S9 respectively), we adjusted for the top 5 principal components in the association analyses for the SiMES and SINDI datasets.
For the STARS trios, we additionally excluded samples and trio-sets on the basis of excessive Mendelian inconsistencies defined as having >1% of the post-QC SNPs exhibiting Mendelian errors. SNPs with more than 10% Mendelian errors are excluded from the association analyses, and the genotypes leading to Mendelian errors in all other remaining SNPs are coded as missing. As family trios are more robust to the presence of population structure, we did not exclude any samples due to population structure.
The genome-wide association tests were performed using PLlNK (version 1.07; http://pngu.mgh.harvard.edu/~purcell/plink/) as the primary analysis tool. A logistic regression adjusted for age and gender is used to model the association of genetic markers with corneal astigmatism. For each of SiMES and SINDI, the top 5 principal components of genetic ancestry from the EIGENSTART PCA were also included as covariates to adjust for population stratification in these populations. We assumed an additive genetic model where the genotypes of each SNP is coded as 0, 1, and 2 for the number of minor alleles carried, in keeping with increments in allelic dosage. For family GWAS association tests in STARS, a transmission disequilibrium test (TDT) is used to measure significant distortions in transmission of an allele from heterozygous parents to the affected offspring under the condition of Mendel's law [65].
We also performed a quantitative trait analysis with the average corneal cylinder power as the outcome. This is performed in PLINK for the unrelated samples, and in FBAT (http://www.hsph.harvard.edu/~fbat/fbat.htm) for the family trios. As the distribution of the quantitative trait of corneal cylinder power is skewed (Figure S10), we performed a normal quantile transformation [66] prior to the association analysis for unrelated samples. For family-based data, no transformation was conducted since the FBAT does not require normal trait [67]
Meta-analyses are performed using weighted-inverse variance estimated from fixed-effect modeling in METAL (http://www.sph.umich.edu/csg/abecasis/metal/). We adopt the method by Kazeem and Farrall [65] to pool the evidence from the case-control analyses and the family trio TDT. For the quantitative trait analysis, the overall z statistics is calculated as a weighted sum of the z-statistics from the linear regressions in the non-familial data and FBAT analysis for the family-based data, weighted by number of unrelated individuals or trios in the respective studies [68].
Results from a genome-wide meta-analysis of the SNPs common to SP2 and SiMES are used in the discovery phase to identify putative variants that are associated with the onset of corneal astigmatism, defined as a P-value<10−5. The remaining three cohorts (SINDI, SCORM and STARS) are used to validate the putative findings. In addition, a genome-wide meta-analysis of all five datasets is also performed. Genotyping quality of all reported SNPs in this paper have been visually assessed by checking the intensity clusterplots.
|
10.1371/journal.ppat.1000276 | Trypanosome Lytic Factor, an Antimicrobial High-Density Lipoprotein, Ameliorates Leishmania Infection | Innate immunity is the first line of defense against invading microorganisms. Trypanosome Lytic Factor (TLF) is a minor sub-fraction of human high-density lipoprotein that provides innate immunity by completely protecting humans from infection by most species of African trypanosomes, which belong to the Kinetoplastida order. Herein, we demonstrate the broader protective effects of human TLF, which inhibits intracellular infection by Leishmania, a kinetoplastid that replicates in phagolysosomes of macrophages. We show that TLF accumulates within the parasitophorous vacuole of macrophages in vitro and reduces the number of Leishmania metacyclic promastigotes, but not amastigotes. We do not detect any activation of the macrophages by TLF in the presence or absence of Leishmania, and therefore propose that TLF directly damages the parasite in the acidic parasitophorous vacuole. To investigate the physiological relevance of this observation, we have reconstituted lytic activity in vivo by generating mice that express the two main protein components of TLFs: human apolipoprotein L-I and haptoglobin-related protein. Both proteins are expressed in mice at levels equivalent to those found in humans and circulate within high-density lipoproteins. We find that TLF mice can ameliorate an infection with Leishmania by significantly reducing the pathogen burden. In contrast, TLF mice were not protected against infection by the kinetoplastid Trypanosoma cruzi, which infects many cell types and transiently passes through a phagolysosome. We conclude that TLF not only determines species specificity for African trypanosomes, but can also ameliorate an infection with Leishmania, while having no effect on T. cruzi. We propose that TLFs are a component of the innate immune system that can limit infections by their ability to selectively damage pathogens in phagolysosomes within the reticuloendothelial system.
| Innate immunity (present from birth) is the first line of defense against microorganisms and provides an initial barrier against disease. Here we show that a minor sub-fraction of human high-density lipoprotein (the good cholesterol), known as Trypanosome Lytic Factor (TLF), not only kills the parasite Trypanosoma brucei, but is also a more broadly acting antimicrobial component of the innate immune system in humans. As TLF is activated under acidic conditions, we evaluated the activity of TLF against the intracellular parasite Leishmania, which infects and grows within acidic compartments of macrophages, cells in our blood that normally destroy invading microorganisms. Here we show that TLF acts directly on Leishmania parasites, causing them to swell, thereby decreasing their infectivity. Furthermore, microscopy of macrophages infected with Leishmania reveal that TLF is taken up and delivered to the same compartment as Leishmania, concomitant with a reduction in the intracellular parasite number. Finally, we made mice that expressed the genes for human TLF; these mice reduced the pathogen burden and thereby controlled the Leishmania infection better than unmodified mice. In contrast, TLF mice were not protected from infection by Trypanosoma cruzi, a related parasite, which transiently passes through acidic compartments within cells.
| Human blood is a potentially hostile environment to colonizing pathogens due in part to effectors of innate immunity. Trypanosome Lytic Factors (TLFs) are a subset of high-density lipoproteins (HDLs) that protect against infection by many but not all species of the African trypanosome. Two TLFs have been characterized in human blood: TLF1 and TLF2. TLF1 is a large (500 kDa) lipid rich HDL composed predominantly of apolipoprotein A-I (apoA-I), haptoglobin-related protein (Hpr), and apolipoprotein L-I (apoL-I) [1],[2]. TLF2 is a 1000 kDa lipid-poor HDL, which is an immunocomplex composed of apoA-I, Hpr, apoL-I, and IgM [1],[3]. Hpr and apoL-I are the two unique protein components of TLFs that are required to give optimal trypanolytic activity.
African trypanosomes are single cell eukaryotes (from the order Kinetoplastida) that live extracellularly in the bloodstream and tissue spaces of their host, from which they endocytose transferrin and lipoproteins for growth. TLF, a lipoprotein, is endocytosed by trypanosomes and trafficked to the lysosome, wherein the acidic pH activates TLF [4]–[7]. TLF forms ion selective pores in trypanosome membranes, which leads to the loss of osmoregulation allowing water influx, swelling and lysis of the trypanosomes [8],[9]. The pore forming activity has been assigned to apoL-I because a purified recombinant preparation can kill trypanosomes [9],[10]. However, in vitro experiments show that the association of Hpr and apoL-I in the same HDL particle is necessary to achieve optimal TLF activity, because reconstitution of individual components reveal that the combination of Hpr and apoL-I are ten-fold more lytic than either component alone and native HDLs with either Hpr or apoL-I alone have levels of activity several hundred fold lower than HDL with both Hpr and apoL-I [11]. Hpr promotes the efficient uptake of TLFs via a putative trypanosome receptor [12],[13]. The presence of an Hp (Hpr) receptor was initially reported by Drain et al. [13]. Recent data indicates that the trypanosome receptor ligand is in fact the complex of Hpr bound to hemoglobin (Hpr-Hb) [13],[14] and/or haptoglobin bound to hemoglobin (Hp-Hb) [15].
There are two other parasites from the order Kinetoplastida, Leishmania sp. and Trypanosoma cruzi, which represent important human pathogens. These parasites, which are primarily intracellular, do not have an ortholog of the Hpr-Hb receptor identified in African trypanosomes [15]. However, they do reside in an acidic parisitophorous vacuole (PV) (permanently or transiently), where TLF could be delivered, activated and act against them. We hypothesize that TLF may function more broadly as a reservoir of antimicrobial proteins such as apoL-I and Hpr-Hb that could be released from the carrier HDL and activated, in the case of Leishmania within the intracellular acidic PV of macrophages.
Leishmania is the causative agent of leishmaniasis, a disease whose manifestations in humans range from mild cutaneous and mucocutaneous lesions to fatal visceral infections. Leishmania undergoes a complex life cycle; human infection initiates with the deposition of non-dividing metacyclic promastigotes by sand flies biting the host skin. The parasites are then taken up by professional phagocytes [16],[17]. The major host cell is the macrophage in which the parasite resides within the PV, a phagosome that ultimately fuses with endosomes and lysosomes, forming an organelle with an acidic pH. Inside the PV, the parasites differentiate to amastigotes, multiply, and eventually rupture the cell and spread to uninfected cells [18].
T. cruzi is the causative agent of Chagas' disease in humans. An infected triatomine insect vector feeds on blood and deposits metacyclic trypomastigote forms of the parasite in its feces, which can enter the host through breaches in the skin or through intact mucosal membranes. T. cruzi trypomastigotes circulate and are disseminated to the heart and other organs through the blood, where they could encounter innate effectors. The parasite is capable of invading and replicating in a wide variety of nucleated cells in the vertebrate host. The trypomastigote form of the parasite enters a host cell and is enclosed within a membrane-delimited vacuole that rapidly fuses with lysosomes [19],[20] providing the acidic environment that is essential for vacuole disruption and parasite replication in the cytosol [21], a critical step in the T. cruzi life cycle.
In the present study we evaluated the effect of TLF on the intracellular growth of Leishmania sp. and T. cruzi, because they are parasites that traffic to acidified PVs to which TLF may be delivered and activated. We find that in axenic acidic conditions TLF damages metacyclic promastigotes externally and reduces their infectivity. Furthermore, TLF ameliorates infection by Leishmania by accumulating within the PVs of macrophages, thereby reducing the pathogen number. In vivo, TLF reduces the pathogen burden of Leishmania in mice, whereas TLF has no measurable effect on T. cruzi infection.
The infective form of Leishmania, metacyclic promastigotes do not divide. They are covered in a dense glycocalyx of lipophosphoglycan, which contributes to their resistance to complement killing. After deposition in the skin by the bite of a sandfly, metacyclic promastigotes are rapidly opsonized and phagocytosed by macrophages and gradually fuse their PV with endosomes and lysosomes. The vacuoles containing Leishmania are acidified and eventually reach pH 5 [22],[23].
We tested the effect of lytic HDL on L. major and L. amazonensis purified metacyclic promastigotes under neutral conditions (such as those encountered in the tissues spaces and blood) and acidic conditions (such as those ultimately encountered in the PV). After 24 hours of co-incubation with a physiological concentration of lytic HDL (1.5 mg/ml), which contains TLF at physiological concentrations (∼10–15 µg/ml) at 27°C in acidic media (pH 5.2) the L. major metacyclic promastigotes became swollen but remained motile (Figure 1A), we could not detect any uptake of propidium iodide indicating that the parasites are still viable (data not shown). In contrast there was no visible effect of lytic HDL in neutral pH media (Figure 1B). TLF binds to the parasites independently of the pH. Incubation with Alexa Fluor-488 labeled human TLF (10 µg/ml) (pH 5.2, Figure 1I, and pH 7.5, Figure 1J) reveals a net shift in fluorescence of the whole population of parasites. Bovine HDL, which does not kill trypanosomes and does not contain TLF, was used as a non-lytic HDL control at an equivalent concentration. The parasites remained motile and elongated in acidic or neutral media in the presence of bovine HDL (data not shown).
The pretreatment of L. major metacyclic promastigotes with lytic HDL in acidic media substantially reduced their infectivity (p<0.01 compared to bovine HDL, Figure 1C), measured by their ability to infect BALB/c bone-marrow derived macrophages. In contrast, there was no change in infectivity after pretreatment of metacyclic promastigotes with lytic or non-lytic HDL in neutral media (Figure 1D). We observed the same outcome after pretreatment with lytic HDL of L. amazonensis parasites before infection of BALB/c bone-marrow derived macrophages (Figure 1E and 1F). Pretreatment of promastigotes with lytic HDL in acidic media significantly reduced their infectivity (p<0.01 compared to bovine HDL, Figure 1E). There was no change in infectivity after pretreatment in neutral media (Figure 1F). In contrast pretreatment of amastigote-like forms (day 13 of axenic transformation) with lytic HDL in acidic or neutral media did not reduce their infectivity for macrophages (Figure 1G and 1H).
We conclude that lytic HDL (which contains TLF) can damage L. major and L. amazonensis promastigotes under acidic conditions thereby affecting their shape and infectivity. In contrast, amastigote like forms are apparently resistant to lytic HDL.
Inside the macrophages, the Leishmania parasite resides in an acidic vesicular compartment, the PV, which has phago-endosomal/lysosomal properties. The fusion properties of the PV are dependant upon the life cycle stage used for infection in vitro i.e. the use of purified metacyclic promastigotes versus a heterogeneous promastigote population and the source and activation status of the host cells [23]–[28]. TLFs are a subset of HDLs and macrophages have receptors for binding and endocytosing HDLs [29]–[31] and haptoglobin [32],[33], 1% of which circulates bound to HDLs [34]. We therefore reasoned that TLF might bind to one or all of these macrophage receptors, be endocytosed, traffic to PVs and exert lytic activity against Leishmania at acidic pH.
We used confocal fluorescent microscopy to visualize the potential uptake and colocalization of TLF with L. major within macrophages. BALB/c bone-marrow derived macrophages were infected with L. major parasites for 2 hours and physiological concentrations of lytic human TLF (10 µg/ml) labeled with Alexa Fluor-594 (Figure 2). After 2 hours incubation TLF (red image) and parasites (small blue dots) were found within the phagolysosome delineated by Lamp-1 antibodies, which label all lysosomal compartments within the macrophages (green image). When all three images were merged, we observed that the parasites and TLF are found within the PV of the macrophage (Figure 2, merged panel, solid arrows). To determine whether the parasites endocytosed the TLF or were coated by the TLF within the parasitophorous vacuole we used GFP-L. major, which express GFP in the entire cytoplasm of the parasite. BALB/c bone-marrow derived macrophages were infected with GFP-L. major parasites for 2 hours. After 2 and 24 hours incubation with Alexa Fluor 594 labeled TLF there was no detectible colocalization of the two dyes, as revealed by the 2D cytofluorograms (Figure 3B, 2 h; and 3D, 24 h), which represent the data collected from 25 individual z-stacks of the two maximum projection images (Figure 3A, 2 hours; and 3C, 24 hours). Therefore, we conclude that TLF is taken up by the macrophages and surrounds the parasites within the PV but may not be endocytosed by the parasite.
The fusion properties of the PV vary with the infecting species of Leishmania. Initially the PVs fuse with the late endosomes/lysosomes of the macrophage and eventually become fully acidified [22]–[28],[35]. Within 24 hours, Leishmania differentiation into amastigotes begins. L. major (organism of the Old World) and L. amazonensis/L. mexicana (of the New World) cause cutaneous leishmaniasis but diverged from each other 40–80 million years ago. Consequently, significant differences in host-parasite interactions have evolved, including differences in the PV. For example, the PVs that harbor L. amazonensis or L. mexicana (large communal PVs) versus those that harbor L. major or L. donovani (small individual PVs) indicate that the fusion/fission processes occurring at the level of these organelles differ mechanistically or kinetically in macrophages infected with these different species [35].
In order to assess the effect of lytic HDL on intracellular Old World L. major parasites within macrophages we added different concentrations of human lytic HDL two hours post-infection of peritoneal macrophages from Swiss-Webster mice with purified metacyclic promastigotes. Bovine HDL, which does not kill trypanosomes and does not contain TLF, was used as a non-lytic HDL control. Two hours post-infection we observe an equivalent infection rate of all macrophages (Figure 4A). However, in the presence of lytic HDL the initial parasite burden of ∼11 parasites/100 macrophages was reduced to ∼5 parasites/100 macrophages (Figure 4A) after 24 hours. To evaluate the lytic capacity of HDL in large communal PVs generated by New World Leishmania, we repeated the 2 and 24 hours incubation with lytic HDL using BALB/c mice bone-marrow derived macrophages infected with L. amazonensis. Two hours post-infection we observe an equivalent infection rate of all macrophages (Figure 4B). At 24 hours L. amazonensis was also killed intracellularly by lytic HDL, reducing the parasite burden by ∼65% (p<0.05 compared to bovine HDL, Figure 4B). At 72 hours post-infection the parasites begin to divide within the macrophages. Of note the Leishmania with macrophages incubated with lytic HDL are growing at 72 hours, which suggests that the parasites have escaped the effect of lytic HDL (TLF), either by transforming or remodeling their PV or both.
Once inside macrophages metacyclics differentiate into amastigotes and begin to divide, this takes 1–3 days. We tested the susceptibility of axenically cultivated amastigote-like forms within macrophages to lytic HDL. BALB/c bone-marrow derived macrophages were infected with promastigotes (Figure 5A) or amastigote-like forms (Figure 5B) of L. amazonensis before treating with lytic HDL for 24 hours. There was no reduction in amastigote numbers within macrophages (Figure 5B). We conclude that amastigote-like forms of L. amazonensis are resistant to lytic HDL (TLF) in macrophages.
Binding and endocytosis of lytic human HDL (TLF) does not activate BALB/c mice bone-marrow derived macrophages infected with metacyclic promastigotes of L. major (Figure 6A). There was no measurable increase in nitrite oxide (NO) production unless the macrophages were treated with IFNγ and LPS. Furthermore, lytic HDL effectively reduced the parasite number in murine bone-marrow macrophages harvested from inducible NO synthase knock-out mice (iNOS−/−), which are unable to make NO (p<0.01, Figure 6B). Taken together the data indicate that lytic HDL does not require nor generate NO to exert anti-leishmanial activity within infected macrophages. In addition, lytic HDL effectively reduced the parasite number in murine bone-marrow macrophages harvested from NAD(P)H oxidase knock-out mice (gp91phox−/−) (p<0.001, Figure 6C), indicating that reactive oxygen species are not required for the anti-leishmanial activity of lytic HDL within macrophages. Lytic HDL effectively reduced the parasite number in bone-marrow macrophages harvested from the parental wild-type mice (C57BL/6 mice, p<0.01, Figure 6D). The magnitude of parasite killing in the presence of human HDL inside macrophages harvested from all three murine strains was the same (∼50%). Overall the data show that macrophages are not activated by lytic HDL and do not require activation for lytic HDL to reduce the parasite burden.
We next examined whether TLF can ameliorate an infection with intracellular L. major in vivo. Previously in our laboratory human TLF was reconstituted in transgenic mice, by generating human HDL particles that contain both apoL-I and Hpr, in vivo [36]. This was achieved using hydrodynamics-based gene delivery (HDG), by which single or multiple transgenes can attain a significantly high level of expression within days of DNA injection [37]. The main organs that are transfected by this technique are the liver and lungs [38]. As the liver is the main tissue that expresses the genes that encode Hpr and apoL-I (and lung), we found sufficient production and correct processing of Hpr and apoL-I occurs by this in vivo transfection technique [36].
To test the effect of reconstituted lytic HDL (TLF) on L. major in vivo, we transfected mice with a single plasmid that contains either apoL-I or Hpr. We also transfected mice with a single plasmid which contains both apoL-I and Hpr (apoL-I∶Hpr) under the control of individual promoters, which results in HDL particles that contain both apoL-I and Hpr [36]. We used C57BL/6 mice, which have the capacity to resolve a leishmanial footpad infection within 8–12 weeks and best “mimic” a human course of infection. ApoL-I, Hpr and apoL-I∶Hpr plasmids were injected 3 days before an L. major footpad infection and protein levels in the plasma and footpad size were monitored. Serial dilution of transgenic-murine plasma revealed that the level of apoL-I was approximately equivalent to that found in human plasma (Figure 7A), while Hpr was expressed at a lower level in the dual plasmid (apoL-I∶Hpr) than in the single Hpr plasmid (Figure 7B). Within 2–3 weeks post infection we observed a 50% reduction in the size of the lesion in mice expressing TLF (apoL-I∶Hpr) (p<0.05; Figure 8A), which translates into a significant three-fold reduction in parasite burden 3 weeks post-infection (p<0.05; Figure 8B).
Although HDL particles that contain both apoL-I and Hpr are more robust and have greater lytic capacity than apoL-I alone [11],[12], we have found that apoL-I is necessary and sufficient to control a trypanosome infection in vivo [36]. Therefore, we next investigated the individual contributions of apoL-I and Hpr toward controlling the L. major infection in vivo. In order to decrease the burden of disease and maximize the effectiveness of human TLF, the mice expressing different human TLF genes were infected with 50% fewer parasites, and the L. major isolate was slightly decreased in virulence by passaging one additional time in vitro. We found that human apoL-I (closed squares) exerted an anti-leishmanial effect that was measurable by a reduction in the footpad lesion size (p = 0.004), while the effect of Hpr (open triangle) was not significant. When Hpr and apoL-I were both expressed (closed inverted triangles) the anti-leishmanial effect appeared to be co-operative (p<0.001; Figure 8C). These results suggest that both apoL-I and Hpr are required to attain the optimal effect against L. major infection. Whether mice, were transfected with a single plasmid that expresses both apoL-I and Hpr (apoL-I∶Hpr), which allows for synthesis in the same transfected cell or transfected with two individual plasmids encoding apoL-I and Hpr (apoL-I+Hpr), we found that both methods of gene delivery and protein expression afford protection compared to control (Figure 8D; apoL-I∶Hpr, p = 0.045; apoL-I+Hpr, p = 0.006). Notably, complete resolution of the lesion follows a similar time course irrespective of the innate immune modulator (apoL-I alone or apoL-I and Hpr), indicating that adaptive immunity plays a key role in the resolution of the disease.
In order to assess the contribution of Hpr to lytic HDL (TLF) activity on intracellular L. major parasites in vivo, we evaluated the role of Hpr as a potential ligand that facilitates the uptake and thus activity of lytic HDL in macrophages. Hp is an abundant serum protein, which when complexed with heamoglobin (Hp-Hb) is an effective inhibitor of lytic HDL (TLF) uptake into African trypanosomes [14],[15]. Therefore, we incubated BALB/c bone-marrow derived macrophages with human lytic HDL (1.5 mg/ml) for 24 hours with or without the addition of Hp (1 mg/ml) two hours post-infection with purified metacyclic promastigotes. Hp prevented lytic HDL from killing the intracellular parasites (Figure 9).
In order to determine if TLF would have an effect on a pathogen that transiently localizes within a phagolysosomal vacuole we compared the kinetics of infection with T. cruzi in wild-type mice to our TLF expressing mice. T. cruzi is another member of the Kinetoplastida that invades cells (including macrophages, smooth and striated muscle cells, and fibroblasts) passing transiently through lysosomes before escaping to the cytosol to replicate. The acute phase of infection is characterized by high blood parasitemia and tissue parasitism. Mice injected with either apoL-I or Hpr plasmid alone or both were infected three days later with T. cruzi trypomastigotes intraperitoneally. Expression of the apoL-I and Hpr proteins were confirmed by western blot (data not shown). The acute phase of the infection was followed by monitoring blood parasitemia daily (Figure 10). No difference in parasitemia was observed between the control mice and mice expressing apoL-I or Hpr, alone or in combination. This suggests that TLF does not have an effect on the acute stage of T. cruzi infection.
Our data shows that TLF has broad anti-microbial properties, with the ability to kill other organisms beyond trypanosomes. Because TLF requires an obligate acidic environment to become activated for pore-forming activity, we have focused on microbes that reside in an acidic environment. Leishmania metacyclic promastigotes are phagocytosed by macrophages wherein they transform into amastigotes within membrane-bound organelles of the endocytic pathway, progressively acquiring late endosomal/lysosomal characteristics. The phagosome acidification and fusion with the late endosomes/lysosomes is variable [22]–[28]. The differentiation to amastigotes starts in the hours following phagocytosis and takes 1 to 3 days to complete [39]. During this differentiation the parasite may be vulnerable to attack because HDL and TLF can be endocytosed and delivered to acidic endo/lysosomes in cells that have an appropriate lipoprotein scavenger receptor, such as SRB-I [29],[31] or SR-BII [30], or Hp receptors that are expressed on macrophages [32],[33].
Our results show that L. major parasites pretreated with lytic HDL in acidic media have a drastic change in morphology whereas in neutral media they maintained normal morphology (Figure 1A and 1B). We observed that TLF bound equally well to the parasites irrespective of the pH (Figure 1I and 1J) and that propidium iodide was excluded from the treated parasites, which suggests that the parasites remain “viable” (data not shown). However, the pretreatment of L. major or L. amazonensis promastigotes with lytic HDL in acidic media substantially reduced their infectivity; whereas, there was no change in infectivity after pretreatment in neutral media (Figure 1C–1F). We interpret this data as follows; TLF increases susceptibility to host macrophage microbicidal processes by damaging the parasites.
African trypanosomes, are killed in neutral media, because lytic HDL (TLF) is endocytosed by the parasites via a Hp-Hb receptor and activated within the acidified lysosome of the parasite, wherein it forms pores [7]–[12],[14],[15]. Leishmania do not have a homologue of the trypanosome Hp-Hb receptor [15] and may not be able to accumulate sufficient lytic HDL (TLF) within 24 hours. Given that the binding of TLF is equivalent in neutral or acidic media, the data suggest either (1) lytic HDL (TLF) may interact with the surface of Leishmania promastigotes and damage the plasma membrane when activated under acidic conditions, possibly by forming pores; or (2) TLF was endocytosed by Leishmania parasites and the promastigote lysosome is weakly acidified in neutral media, as it stains poorly with the pH sensitive probe, lysotracker [40], but in acidic media the parasite lysosome will be fully acidified, allowing the activation of the TLF.
TLF accumulates within the PV of macrophages (Figure 2). The observation that all PVs contain TLF, which surrounds the parasites but does not appear to be endocytosed by the parasites (Figure 3), concurs with the axenic data; TLF may act directly at the parasite plasma membrane within the PV, though we cannot rule out that some TLF may be endocytosed by the parasite. We find that the number of parasites within macrophages decreased by 24 hours post-addition of lytic HDL in a dose-dependent manner (Figure 4). However, the clearance of the parasites was not complete, which may be due to individual differences each in PVs acidification process. Overall these data indicate that addition of lytic HDL (TLF) decreases the number of metacyclic promastigotes in vitro in macrophages. In contrast, we find that amastigotes are resistant to lytic HDL (TLF) axenically (Figure 1G and 1H) and within macrophages (Figure 5). Therefore, we conclude that the window of Leishmania susceptibility to lytic HDL is after phagocytosis of the metacyclic promastigotes during acidification of the PV and before transformation into amastigotes (Figure 11). Our data also show that the effect of lytic HDL on Leishmania is independent of macrophage activation (Figure 6).
In vivo infections with L. major lead to the development of cutaneous lesions, which are considered to arise from growth within tissue macrophages. In our in vivo system, apoL-I and Hpr were both required to maximally reduce the Leishmania lesion (Figure 8). The reduction of the Leishmania lesion by apoL-I was statistically significant (p = 0.004; Figure 8C). However, the dual expression of Hpr and apoL-I reduced lesion size significantly compared to apoL-I alone (p = 0.001). The reduction in lesion size was effective whether the genes were expressed from individual plasmids such that expression is from different transfected cells (p = 0.006) or the same plasmid, which allows for expression from the same transfected cell (p = 0.045; Figure 8D). Therefore, the two proteins appear to be acting co-operatively. ApoL-I likely forms a pore in the membrane of the Leishmania parasite directly at the plasma membrane and/or at the lysosomal membrane (Figure 1). Hpr appears to be a ligand, which binds to a putative receptor on macrophages and enhances the uptake of TLF into macrophages. We draw this conclusion from the in vitro competition data in Figure 9, which showed that Hp prevented the lytic HDL from killing the intracellular parasites. Recent studies demonstrate that neutrophils are the initial host cell that phagocytose a substantial fraction of L. major parasites after sandfly transmission [17]. Neutrophils can bind and endocytose HDLs particles [41] and Hp [42] 1% of which circulates bound to HDLs [34]. It is plausible that in vivo, in addition to macrophages, TLF might be endocytosed and traffic to PVs within neutrophils and exert lytic activity against Leishmania at acidic pH.
The co-operative effect of Hpr and apoL-I may also require Hb as proposed for African trypanosomes [14],[15]. Hb (from the FBS in culture media and in murine blood) can be bound to TLF via Hpr, and thereby be taken up by infected macrophages. The Hpr-Hb complex may be the ligand that facilitates uptake of apoL-I (in TLF complexes) into macrophages. It has been proposed that Hpr-Hb complexes may generate free radicals by reacting with hydrogen peroxide within the acidified lysosomes [14]. Although free radicals could contribute to the damage of Leishmania parasite membranes, we find that macrophages devoid of any NAD(P)H oxidase, which generates superoxide that can dismutate to hydrogen peroxide, are able to kill Leishmania (Figure 6C) as effectively as wild-type macrophages. Furthermore, Hpr-mice did not change the lesion size significantly in vivo (Figure 8C).
While the transgenic-TLF mice do not completely clear the Leishmania infection, they substantially reduce the parasitemia. The fact that Leishmania infection is not eliminated in the presence of TLF is not unexpected since humans have TLF but remain susceptible to Leishmania infection. Thus, TLF may serve to reduce the initial pathogen numbers and limit dissemination of the parasite until adaptive immunity takes effect. Other possible explanations for partial parasite clearance may be that TLF is less abundant in tissue spaces (∼25%) than in blood, and therefore TLF levels may not be optimal at the footpad lesion derma in order to act against the parasite. In addition, the hydrodynamic gene delivery system allows maximal expression of the proteins for ∼10 days. Indeed, ∼2 weeks post-injection of the plasmid, the protein expression in plasma drops below the limit of detection. Nevertheless, some low level of protein is maintained for months since mice can be infected with T. brucei several months post-injection of apoL-I plasmid, and resist the infection (data not shown).
In contrast to L. major, we could not detect any effect of transgenic-TLF against T. cruzi parasites (Figure 10). This finding suggests that bloodstream trypomastigotes, which accumulate to high numbers in the circulation during the acute stage of infection and invade both phagocytic and non-phagocytic cell types, are refractory to TLF. This may reflect the ability of the parasite to infect non-phagocytic cells that may not take up HDL efficiently. Additionally, it is possible that because T. cruzi resides transiently (8–16 hours) within acidified vacuoles, the parasites are not exposed to active TLF for a sufficient period of time.
These new findings support the hypothesis that TLF not only kills African trypanosomes, but also contributes to the innate immunity against other pathogens, such as Leishmania. The efficiency of killing other pathogens by TLF may depend on both a physical interaction as well as an extended period of contact between the susceptible pathogen and TLF. African trypanosomes grow in the blood and tissues spaces of the human host and constantly endocytose TLF, whereas Leishmania parasites grow within phagocytic cells in fully acidified PVs to which TLF may be delivered, but then transform to evade TLF action. In contrast, T. cruzi parasites infect non-phagocytic cells as well as professional phagocytes, and are only transiently localized within acidified vacuoles, such that constant exposure to active TLF is unlikely. We conclude that TLFs are a component of the innate immune system, which can limit infections by their ability to selectively damage pathogens such as Leishmania, that reside within the reticuloendothelial system.
HDL was purified from normal human serum by adjusting to a density of 1.25 g/ml with potassium bromide (KBr) and ultracentrifuged at 49,000 rpm (NVTi 65; Beckman) for 16 hours at 10°C. The lipoprotein fraction was collected and the density of this fraction was adjusted to 1.3 g/ml with KBr and 4 ml aliquots were layered under 8 ml of 0.9% NaCl. The lipoproteins were then centrifuged at 49,000 rpm for 3 hours at 10°C (NVTi 65 rotor; Beckman). HDL was harvested and dialyzed against Tris-buffered saline (TBS; 50 mM Tris-HCl, 150 mM NaCl (pH 7.5) at 4°C and then concentrated by ultrafiltration (XM300 filter membrane; Amicon). HDL was concentrated to about 50 mg of protein/ml. TLF was obtained by affinity purification of human HDL using a mouse anti-human Hp monoclonal (H6395, Sigma) coupled to a HiTrap column (Amersham Biosciences). The fractions containing Hpr (TLF) were pooled and concentrated.
HDL and LDL from bovine serum cannot be efficiently separated by density, unlike human HDL and LDL. Therefore when bovine HDL was used as a control the lipoproteins were purified by adjusting their density to 1.25 g/ml with KBr and ultracentrifuged for 16 hours at 49,000 rpm, 10°C. The lipoprotein fraction (density 1–1.25 g/ml) was then collected, and size-fractionated on a Superdex 200 HR 10/30 column (Amersham) equilibrated with TBS [1]. Fractions containing apoA-I, the canonical HDL apolipoprotein, were pooled and concentrated.
L. major strain Friedlin V1 (MHOM/JL/80/Friedlin) and L. major FV1 SSU: GFP+(b)-SAT promastigotes were grown as previously described in medium M199 [43] (neutral medium 1), and infective-stage metacyclic promastigotes were isolated from stationary cultures (5-days old) by density centrifugation on a Ficoll gradient [44].
L. amazonensis IFLA/BR/67/PH8 strain promastigotes were maintained in vitro as previously described [45] (neutral medium 2). L. amazonensis axenic amastigote-like forms were cultured at 32°C in the same medium supplemented with 0.25% glucose, 0.5% trypticase, and 40 mM Na succinate (acidic medium) [45].
L. major and L. amazonensis metacyclics and amastigote-like forms were incubated for 24 hours at 27°C and 32°C respectively in corresponding neutral medium or in amastigote acidic medium in the presence or the absence of HDL. They were washed and checked for integrity under the microscope. Thereafter they were allowed to invade macrophages in DMEM containing 10% heat-inactivated FBS, 5% penicillin-streptomycin, 5 mM L-glutamine (DMEM culture medium), at a multiplicity of infection of 3 to 6 parasites per macrophage for 24 hours at 33°C (5% CO2, 95% air humidity). Intracellular parasites were assessed after staining with DAPI (3 µmol/L) by fluorescence microscopy.
Bone marrow-derived macrophages were prepared as described previously [46]. Cells were prepared from femurs of BALB/c mice (Taconic), B6;129P2-Nos2tm1Lau/J or B6.129S6-Cybbtm1Din/J, C57BL/6/J and after 3 days in culture, non-adherent progenitor cells were taken and cultured for an additional 7 days in culture medium supplemented with 30% (v/v) L cell-conditioned medium as a source of CSF-1. Adherent cells were harvested with cold DMEM+0.5 mM EDTA and seeded into an 8-well Lab-Tek II (Nalge Nunc International, Naperville, IL) chambered coverglass at a concentration of 50,000 cells/chamber and allowed to adhere for 24 hours (37°C, 5% CO2, 95% air humidity) before being used for infections.
Unactivated intraperitoneal macrophages were isolated by lavage of the intraperitoneal cavity of Swiss-Webster Mice (Taconic). The cells were resuspended in DMEM culture medium, seeded into an 8-well Lab-Tek II (Nalge Nunc International, Naperville, IL) chambered coverglass (50,000 cells/chamber), and allowed to adhere for 24 hours (37°C, 5% CO2, 95% air humidity). Thereafter, non-adherent cells were removed by three extensive washings with culture medium before being used for infections.
L. major metacyclics and L. amazonensis promastigotes or amastigotes were opsonized by 30 min incubation in DMEM medium containing 4% BALB/c or Swiss-Webster mouse serum and allowed to invade strain matched macrophages in DMEM culture medium, at a multiplicity of infection of 3 parasites per macrophage for 2 hours at 33°C (5% CO2, 95% air humidity). Thereafter, non-phagocytosed parasites were washed off, and the cultures were further incubated in the presence or the absence of HDL with or without Hp (H3536, Sigma) for indicated times. Intracellular parasites were assessed after staining with DAPI (3 µmol/L) by fluorescence microscopy.
Bone marrow-derived macrophages BALB/c mice were seeded into an 8-well Lab-Tek II chambered coverglass at a concentration of 150,000 cells/chamber before being used for infections with L. major at a multiplicity of infection of 3 parasites per macrophage for 2 hours at 33°C (5% CO2, 95% air humidity). Thereafter, non-phagocytosed parasites were washed off, and the cultures were further incubated in the presence or the absence of HDL (1.5 mg/ml) with or without murine IFNγ (5 µg/µl, 315-05, Preprotech and LPS (100 µg/µl, L6511, Sigma) for 24 hours. Nitrite quantification was measured by the Griess reaction according to the manufacturer (G7921, Molecular Probes).
Metacyclic promastigotes (1×106) were inoculated intradermally into the right hind footpad of C57BL/6 mice (Taconic) in a volume of 50 µl using a 28.5-gauge needle (5 mice per group). The evolution of the lesion was monitored by measuring the lesion thickness with a direct-reading Vernier caliper. A non-parametric approach for several independent groups, Kuskal Wallis test, was used to analyze the data. For post-hoc comparisons, Mann Whitney tests were used with a Bonferroni correction. Parasite titrations were performed with footpad tissue homogenates obtained from individual mice and serially diluted. Each dilution was dispensed into 36 wells to give sufficient data for Poisson distribution. After 10 days, the growth of parasites was determined microscopically. The number of viable parasites in each sample was determined from the highest dilution at which promastigotes could be grown out after 7 days of incubation at 27°C. For treatment comparisons Mann Whitney tests were used.
TLFs were labeled with Alexa Fluor-594 or Alexa Fluor-488 protein labeling kit (Molecular Probes) according to the manufacturer's instructions.
L. major metacyclics FV1 or FV1 SUU: GFP+(b)-SAT purified metacyclics were opsonized by 30 min incubation in DMEM medium containing 4% serum from BALB/c mice and allowed to invade BALB/c bone-marrow derived macrophages for 2 hours at 33°C (5% CO2, 95% air humidity). Thereafter, non-phagocytosed parasites were washed off, and the cultures were further incubated in the presence of Alexa labeled TLF for 2 or 24 hours. Live parasites within macrophages were fixed with 2% paraformaldehyde. Cells were permeabilized with 0.05% saponin. Lamp-1 staining was performed using a rat monoclonal antibody to mouse Lamp-1 (1∶100, 1D4B; Developmental Studies Hybridoma Bank, Iowa City, IA), followed by goat anti-rat IgG conjugated to FITC antibodies (1∶200, Sigma). Intracellular parasites were observed by staining with DAPI (3 µmol/L) or direct GFP fluorescence of parasites. The samples were visualized and analyzed with a Leica TCS SP2 AOBS confocal laser scanning microscope.
For flow cytometry on live Leishmania, purified L. major metacyclics were washed twice in PBS and incubated (2×107/ml) with 10 µg/ml Alexa Fluor-488 labeled TLF in bicine-buffered saline with glucose (pH 5 or 7.5) for 30 min. Cells were washed twice in FACS buffer (PBS, 5% FBS, and 0.1% sodium azide) before being analyzed. Flow cytometry was performed with a Becton Dickinson FACSCalibur system.
Expression of human Hpr and apoL-I in plasma of mice was achieved using hydrodynamics-based gene delivery [38]. Briefly, 20 g male C57BL/6 mice (for L. major experiments), and Swiss-Webster (for T. cruzi) were injected IV, in less than 10 seconds with 2 ml of sterile 0.9% NaCl solution containing 50–100 µg of plasmids [36]. Three days after injections and every other day thereafter, blood samples (20 µl) were taken from the animals via tail bleeds and expression of the human proteins was evaluated by western blotting.
Plasma samples were separated on 7.5% Tris-glycine PAGER Gold precast Gels (Cambrex Bio Science Rockland, Inc. ME). Gels were transferred onto PDVF membranes (GE Healthcare Bio-Sciences, Uppsala, Sweden). For western blot analysis membranes were blocked with 5% skimmed milk and 0.1% Tween-20 in TBS and probed for 1 hour with the following antibodies: mouse monoclonal anti-Hpr (1∶5000); mouse monoclonal anti-apoL-I (1∶10,000; kindly provided by Dr. Stephen Hajduk). The secondary antibodies were conjugated to horseradish peroxidase, and used at the following dilutions: anti mouse IgG (1∶50,000; Promega, Madison, WI). Primary and secondary antibodies were diluted into 2.5% skimmed milk and 0.1% Tween–20 in TBS. Bound antibodies were detected by chemiluminescence using ECL (GE Healthcare Bio-Sciences, Uppsala, Sweden).
Tissue culture-derived T. cruzi trypomastigotes (Y strain) were generated by weekly passage in confluent monolayers of LLcMK2 cells in DMEM containing 2% FBS as described previously [47]. Trypomastigotes harvested from culture supernatants were washed three times in serum free DMEM prior to use. T. cruzi trypomastigotes (106) were injected intraperitoneally into Swiss-Webster mice (Taconic) three days after transfection (3 mice per group). Parasitemia was monitored in peripheral blood of infected mice by microscopic examination of non-fixed blood.
Apolipoprotein L-I, NM_003661; Haptoglobin-related protein, NM_020995.
|
10.1371/journal.pntd.0006262 | Comparison of three data mining models for prediction of advanced schistosomiasis prognosis in the Hubei province | In order to better assist medical professionals, this study aimed to develop and compare the performance of three models—a multivariate logistic regression (LR) model, an artificial neural network (ANN) model, and a decision tree (DT) model—to predict the prognosis of patients with advanced schistosomiasis residing in the Hubei province.
Schistosomiasis surveillance data were collected from a previous study based on a Hubei population sample including 4136 advanced schistosomiasis cases. The predictive models use LR, ANN, and DT methods. From each of the three groups, 70% of the cases (2896 cases) were used as training data for the predictive models. The remaining 30% of the cases (1240 cases) were used as validation groups for performance comparisons between the three models. Prediction performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Univariate analysis indicated that 16 risk factors were significantly associated with a patient’s outcome of prognosis. In the training group, the mean AUC was 0.8276 for LR, 0.9267 for ANN, and 0.8229 for DT. In the validation group, the mean AUC was 0.8349 for LR, 0.8318 for ANN, and 0.8148 for DT. The three models yielded similar results in terms of accuracy, sensitivity, and specificity.
Predictive models for advanced schistosomiasis prognosis, respectively using LR, ANN and DT models were proved to be effective approaches based on our dataset. The ANN model outperformed the LR and DT models in terms of AUC.
| Worldwide, approximately 240 million individuals are infected with schistosomiasis, a parasitic neglected tropical disease that continues to be a significant cause of morbidity and mortality, especially in China. Effective tools that can accurately predict the prognosis of patients with advanced schistosomiasis would aid in the treatment and management of the disease. To this end, we constructed and compared the performance of three predictive models—an artificial neural network (ANN) model, a logistic regression (LR) model and a decision tree (DT) model—in their ability to predict the prognosis of patients with advanced schistosomiasis. We found that while all three models proved effective, the ANN model outperformed the LR and DT models in terms of AUC and sensitivity. Yet, to achieve the highest level of prediction accuracy and to better assist medical professionals, we recommend comparing the performance of the three predictive models to select the optimal one, which will be better than select a model at random. The findings of this study not only provide valuable information on the construction of effective predictive models for the prognosis of advanced schistosomiasis, but also offer new methodology for clinically determining patient diagnosis and prognosis.
| Approximately 240 million individuals are infected worldwide by schistosomiasis, with an estimated 3.31 million disability-adjusted life years lost as a result of the disease [1–4].Further, one meta-analysis and several scientific reports have suggested that global burden caused by schistosomiasis may be several times higher[3]. This concern mainly comes from the following reasons. The first reason is that the low sensitivity schistosomiasis diagnostic methods and insufficient investment of health resources may result in underdiagnosis of schistosomiasis in epidemic areas. The second reason is that the value of disability weight (DW) of schistosomiasis might be set too low (0.005–0.006) in the calculation of DALY value, which is similar to those for disorders such as moderate discolouration of the face (facial vitiligo)[5]. The third reason is that whether infected with schistosomiasis was set as the only healthy outcome in the estimation of DALY value rather than considering disparity in different clinical stages of schistosomiasis (acute, chronic, advanced). Fourth, the disparity in different schistosome germline was also not taken into account for the pathological process varies greatly among Schistosoma mansoni, Schistosoma haematobium and Schistosoma japonicum. Nevertheless, Schistosomiasis was still regarded as one of the most important neglected tropical diseases worldwide.
In China, schistosomiasis has been endemic in 12 provinces and municipalities [6]. Currently, the prevailing regions endemic for schistosomiasis are located in the lake and marshland regions, such as Hunan, Hubei, Jiangxi, Jiangsu, and Anhui, and in the hilly and mountainous regions, such as in Yunnan and Sichuan. However, other regions, such as Fujian, Guangdong, Shanghai, Zhejiang, and Guangxi have successfully fulfilled the criteria for interrupting schistosomiasis transmission since 1985[7]. Hubei province is one of the five lake and marshland schistosomiasis endemic regions which located in the middle and lower regions of the Yangtze River [8]. In addition, Hubei has the largest area of the freshwater snail Oncomelania hupensis, which is the only intermediate host of Schistosoma japonicum. Moreover, Hubei has the highest rates of schistosomiasis transmission in China [9].
By 2015, Hubei had 9098 (29.50% of China’s total cases) documented cases of advanced schistosomiasis, ranking Hubei first in all schistosomiasis endemic provinces in China [10]. Advanced, or late-stage schistosomiasis japonica can be regarded as an extreme form of chronic schistosomiasis, which is more serious than the advanced hepatosplenic disease of Schistosoma mansoni infection found in Africa and the Americas [11]. According to ‘Diagnostic Criteria for Schistosomiasis’ (WS261-2006), one of health industry standards in People's Republic of China provided by National Ministry of Health, the advanced schistosomiasis case is defined as a patient with schistosomiasis who develops portal hypertensive syndromes of liver fibrosis, severe growth disorders or significant colon granulomatous hyperplasia. Due to repeated or mass infection of schistosome cercariae, without thorough and timely treatment, patients can evolve into advanced schistosomiasis usually after 2 to 10 years of pathological development process. Clinical symptoms of advanced schistosomiasis include ascites, splenomegaly, portal hypertension, gastro-esophageal variceal bleeding, granulomatous lesions of the large intestine, and serious growth retardation [12, 13]. Advanced schistosomiasis japonica is much more common in highly endemic areas, because repeated, heavy exposure to cercariae means that early-stage chronic cases may not be effectively treated in routine control programs. The eggs of S. japonicum retained in the intestine and liver tissue stimulate a granulomatous response, leading to continuous fibrosis of the periportal tissue and developing a pipestem fibrosis. Although down-modulation of the granulomatous response, which could prevent further chronic morbidity after 2–5 years or more, parasite-induced periportal fibrosis may progress to cause obstruction of the portal vessels and damage to the liver parenchyma, leading to development of advanced schistosomiasis. Mortality eventually results from bleeding of the upper gastrointestinal tract, spontaneous bacterial peritonitis, and hepatic failure, among other factors. Based on its major symptoms, advanced schistosomiasis japonica in China represents a widespread, serious health burden, and has been classified into four clinical sub-types, namely ascites, megalosplenia, colonic tumorous proliferation, and dwarfism [14, 15].
Predictive models used in disease prognosis studies can answer the following questions such as the seriousness of the patient’s condition and whether can be cured. Also it can be used to guide clinical treatment and help to select the right medical decision-making. Therefore, the predictive model is of great significance. Specifically, the predictive model can be used to understand the trends and consequences of a disease and help clinicians make treatment decisions and determine the urgency of treatment. The model can be applied to study the various influencing factors that affect the prognosis of the disease and assess the effectiveness of a treatment.
Logistic regression (LR) model is a probabilistic non-linear regression model. As a popular multivariate analysis method, it is widely used to study the relationship between dichotomous observations and some influencing factors. In epidemiology, LR model is always used to explore the risk factors of a disease, predict the probability of a disease occurring based on risk factors and so on. For example, to explore the risk factors for gastric cancer (GC), you can choose two groups of people, a GC group and a non-GC group with different signs and lifestyles. The dependent variable here is gastric cancer ("yes" or "no"), while independent variables can covers a lot, such as age, gender, eating habits, Helicobacter pylori infection. The arguments in the model can be either continuous or categorized. By logistic regression analysis, we can get a general understanding of which factors are risk factors for GC.
ANN model is a mathematical model that simulates the structure of the human brain and the way of information transmission. It consists of a set of interconnected “neurons” linked with weighted connections. The model was constructed by an input layer, a hidden layer and an output layer. The input layer contains neurons that receive input data available for analysis (e.g. various demographical, clinical or laboratory data), and output layer contains neurons that export different values.ANN can learn through examples and associate each input with the corresponding output by modifying the weight of the connections between neurons. The output value is compared with the expected output. If there is a discrepancy between these two values, an error signal is generated and then a back propagation (BP) method is applied to alter the weight of the connections between neurons to decrease the overall error of the network. As learning proceeds, the error between the ANN output and the expected output decreases until a minimum is reached. The process was called convergence of the network. After these two training processes, the ANN can generate outputs (prognosis) from new input data based on the knowledge accumulated during training, which is regarded as inference process. Thus, after training, the ANN can make predictions on data sets never seen before or identify patterns.
There are some similar studies. A study demonstrated that the ANN model is a more powerful tool in determining the significant prognostic variables for gastric cancer (GC) patients, compared to the Cox proportional hazard regression (CPH) model [16]. In another study, the ANN model was shown to be more accurate in predicting 3-month mortality of acute-on-chronic hepatitis B liver failure (ACHBLF) than Model for end-stage liver disease (MELD) based scoring systems [17]. In addition to these examples, a trained ANN performs at least as well as physicians in assessments of visual fields for the diagnosis of glaucoma in a ophthalmology research [18].
The decision tree (DT) is a machine learning model, composed of decision rules based on optimal feature cutoff values that recursively split independent variables into different groups to predict an outcome in a hierarchical manner. The principle of DT is similar to that of variance discomposition in ANOVA. The basic purpose is to divide the research population into several relatively homogeneous subgroups through some attribute values. The values of internal variables in each subgroup are highly consistent, and the corresponding variations (impurities) fall in different subgroups as far as possible. All DT model algorithms follow this principle, which is different from ANOVA by definition of variation (impurity), such as P values, variance in ANOVA and information entropy, G1NI coefficients, deviance in DT.
Some examples are also provided. A simple, clinically relevant DT model was developed and validated to reliably discriminate patients at high and low risk of death using routinely available variables from the time of diagnosis in unselected populations of patients with malignant pleural mesothelioma (MPM) [19]. Another simple decision tree can provide a quick assessment of the severity of the chronic obstructive pulmonary disease (COPD) by using variables commonly gathered by physicians, as measured by the risk of 5-yr mortality [20]. The DT modeling based on C4.5 algorithm which was applied to predict prostate cancer risk in another study showed different interaction profiles by race [21].
Traditional LR model is the most popular predictive among different classification methods because the effects of each factors in LR model could be quantitatively explained and an approximately estimate of the relative risk (OR) could be derived easily. However, whether the data could fit the model requires that the data satisfy a given condition and the collinearity and interaction between the variables cannot be solved. ANN model possesses strong ability to solve such problems and has no limitation on the distribution of data. It is generally believed that the ANN model is better than LR model for the disease with many pathogenic factors and complicated relationships among these factors. The DT model also generally considers the interaction between the variables, and it shows a clear screening process in the form of a tree. Compared with the OR value of LR model, the DT model is more conducive for clinicians' understanding. Therefore, the aim of this study was to compare the performance of three predictive models (ANN, LR and DT) for the prognosis of advanced schistosomiasis cases, along with a 10-fold cross-validation technique. The performance of the predictive models was evaluated according to the area under the receiver operating characteristic curves (AUC), accuracy, sensitivity, and specificity.
The study was approved by Research Ethics Committee in Tongji Medical College of Huazhong University of Science and Technology. The methods of the present study were put into effect according to the approved protocols. All participants in this study were adults. Note: Though a child from Xingzi county, Jiangxi Province had ever been reported to diagnosed as advanced schistosomiasis[22], such case is exceedingly rare and never been reported in Hubei. In general, all the advanced schistosomiasis patients are adults.
The participants read the investigation purpose statement and signed informed consents. All data were anonymized and handled confidentially.
Schistosomiasis surveillance data was collected from a previously constructed database of advanced schistosomiasis cases in the Hubei province from a study conducted by the Hubei Institute of Schistosomiasis Prevention and Control. The information was obtained by a standard sociodemographic and epidemiological questionnaire for patients in Hubei with advanced schistosomiasis. Participants were recruited from schistosomiasis epidemic areas all over the province, primarily along the Yangtze River regions. The treatment methods of advanced schistosomiasis patients vary with different disease conditions. Liver protection and symptomatic treatment was applied for ascites type patients. Splenectomy was needed to be done in splenomegaly patients if there is hypersplenism symptom existed. The praziquantel (PZD) treatment can be utilized after six months of stable period in which the general situation of the patient is fine (e.g. no ascites or hemorrhage symptoms).
The medical records of the patients with advanced schistosomiasis were reviewed by attending physicians. Criteria of cases inclusion are as follows:
To avoid the confounding effect of other diseases on the prediction of advanced schistosomiasis prognosis, the patients with following diseases were excluded from the study.
A total of 4136 cases were included in the study which consisted of 2674 men and 1462 women and were divided into two groups: favorable prognosis and poor prognosis. Favorable prognosis referred to cases of recovery and improved disease outcomes while poor prognosis referred to cases of deterioration and death. The presence of the event (dead or deterioration) was coded as 1 and the absence of the event (recovery or improved) was coded as 0. The death of advanced schistosomiasis patients was mainly due to schistosomiasis and schistosomiasis-induced complications, such as upper gastrointestinal hemorrhage, hepatorenal syndrome (HRS), hepatic coma and liver cancer. Therefore, the death outcome that appears in this article refers to all-cause death. The deterioration outcome means that the primary symptoms persist (e.g. no ascites regression sign) or patients in splenomegaly type have no surgical indications.
Data collection included demographical data, hospitalization costs, clinical features, surgical procedures, and outcomes. This study was entirely retrospective which was utilizing records from the hospitals specializing in schitosomiasis of various epidemic counties, Hubei province.
In the first step, the continuous explanatory variables were transformed into categorized variables to decrease the effect of extreme values and enhance the computational efficiency of the ANN. The cutoff points of these variables were set as 0.5. The variables included occupation, annual income, body mass index (BMI) and so on. The sociodemographic and epidemiological characteristics of the 4136 advanced schistosomiasis cases are presented in Table 1. The criterion used for the histopathologic diagnosis of advanced schistosomiasis was the national standardized diagnostic criteria for schistosomiasis (WS261-2006). In the second step, a univariate Cox proportional hazard model was used to improve the computational efficiency and prediction performance of the ANN model by testing the potential relationships between independent variables. Variables with statistically significant differences (log-rank test, P<0.05) were reserved to build the ANN model (Table 1). In total, 16 variables were selected to build the ANN model.
Patients were randomly assigned to the training group (70% of the total cases) for the development of the ANN, DT, and LR models. The rest of the patients (30% of the total cases) were assigned to the validation groups for the assessment of model performance. Of the 4136 patients with advanced schistosomiasis, 2896 were assigned to the training group and1240 were assigned to the validation group. As listed in Table 1, the effects of the input variables did not significantly differ between the training group and the validation group of all three models (P>0.05), indicating the reliability of the data partition.
The data mining software package MATLAB (Matrix Laboratory, Math Works Company, USA, R2014a software) was used to run ANN and C4.5 DT models.
SPSS 19.0 (IBM Corp, Armonk, NY, USA) was used to establish the LR model.
For all comparisons, differences were tested with two-tailed tests and P values less than 0.05 were considered statistically significant.
An ANN is one of the most widely applied models in the medical domain, such as for the interpretation of imaging techniques, prognosis, diagnosis, or diagnostic tests. ANN differs from other conventional statistical models in that ANN usually has more parameters. This study used an ANN model with a standard feed-forward back propagation (BP) network structure, including an input layer of 16 neurons, a hidden layer of 20 neurons, and an output layer of 2 neurons, to predict the prognosis of patients with advanced schistosomiasis. Sigmoid transfer functions were applied to the hidden and output layers. Gradient descent was used to calculate the synaptic weights. The initial learning rate was defined as 0.07 and the momentum was 0.95. The batch size was defined as 256 and the number of iterations was 200. Ten-fold cross-validation was employed. Fig 1 shows the structure of the ANN model. As there is currently no accepted theory that predetermines the optimal number of hidden layer neurons, the number of hidden layer neurons was determined by repeated trial and error test until the best sensitivity and specificity was achieved.
For the categorical dependent variables, a LR model was conducted to identify the risk factors of various diseases by using patient demographic characteristics and other disease parameters. The LR model formula calculates the probability of a given disease, y (y = 1 if the selected case suffers from the disease, otherwise, y = 0). If the subject suffers from the disease, the conditional probability is represented as p(y = 1∣X) = p(X), and the formula of the LR model is expressed as log [(p(x) ∣1− p(x)] = β0+β1x1+β2x2+…+βkxk],where X = (x1, x2,…, xk) denotes the vector of independent variables. An ‘entry’ approach was used to construct the LR model using the 16 variables. The LR model was built using the training dataset and tested using the validation data.
The model-based clinical data interpretation system C4.5 algorithm for the prognosis of advanced schistosomiasis is shown in Fig 2.C4.5 was used as the multiclass classification algorithm, which was a development of the DT algorithm ID3. The algorithm contained the same working principle, but calculated information gain differently. In the ID3 algorithm, the learning process is conducted in reference to the gain calculation, which is the same gain calculation in the feature selection process of the information gain, as shown in Eqs (1) and (2). In the C4.5 algorithm, the learning process uses the ID3 normalized gain, as shown in Eqs (3) and (4):
Entropy(S)=∑tc−pilog2(pi)
(1)
Gain(S,A)=Entropy(S)−∑v∈Values(A)SvSEntropy(Sv)
(2)
GainRation(S,A)=Gain(S,A)/SplitInfo(S,A)
(3)
SplitInfo(S,A)=∑t=1cSvSlog2(SvS)
(4)
The AUC was used to compare the prediction performance of the three data mining models. The classification accuracy referred to the fraction of cases classified correctly. Sensitivity referred to the proportion of positive cases that were classified as positive. Specificity referred to the proportion of negative cases that were classified as negative. The formulas are shown as follows, where TP, FP, TN, FN represent true positives, false positives, true negatives, and false negatives, respectively. The AUC value of ANN can be interpreted
Accuracy=(TP+TN)/(TP+FP+TN+FN)
Sensitivity=TP/(TP+FN)
Specificity=TN/(FP+TN)
For the training and validation group, the ROC curves for the ANN, LR, and DT models are shown in Figs 3 and 4. In the training group, the AUC value for the prognosis of patients with advanced schistosomiasis was 0.927 for the ANN model, 0.828 for the LR model, and 0.823 for the DT model. The AUC values of the ANN model were superior to those of the DT and LR models. In the validation group, the AUC value for the prognosis of patients with advanced schistosomiasis was 0.832 for the ANN model, 0.835 for the LR model, and 0.815 for the DT model. The AUC values of the ANN, DT, and LR models were approximate.
The performance comparison of the three models in the two groups is listed in Table 2. We evaluate the differences in order to see whether there was significance. AUC value could be shown as the normalized Mann–Whitney U statistics. Concerning the normalization denominator is universal for all models, we could thus show the superiority by the AUC value from nonparametric test perspective. Specifically, given the true label of each sample, the larger AUC value, the lager Mann–Whitney U statistics, the better classified capability of the model. We additionally conduct two pairwise tests for AUC values to substantiate the superiority.
For ANN and DT, the result shows the difference is significant. (Z = 15.742,P = 0.000).For ANN and LR, we obtain the similar result as following.(Z = 15.117,P = 0.000)
Advanced schistosomiasis, resulting from either repeated infection or acute infection without chemotherapy, is the most severe form of schistosomiasis and clinically presents with portal hypertension [23], periportal liver fibrosis, spleen enlargement, congestion, and other serious conditions [24–26].
Data mining systems aim to extract implicit, previously unknown and potentially valuable relationships and patterns from large amounts of data to provide clear and useful information through advanced processes of selecting, exploring, and modeling [27, 28]. Recent years have seen a rapid development of data mining technology [29, 30].Currently, predictive models are being used in the clinical setting to improve diagnostic and prognostic accuracy and enhance clinical decision-making [28, 31]. Of these predictive models, LR, ANN, and DT models are among the most widely used models for predicting a patients’ prognosis [14, 32–34]. However, little research has been conducted on the use of data mining methods to establish predictive models for prognosis of advanced schistosomiasis. Thus, the current study used data from the Hubei Institute of Schistosomiasis Prevention and Control to develop and compare three predictive models in their ability to predict the prognosis of patients with advanced schistosomiasis.
One of the most attractive features of ANN is the system’s ability to apply machine learning, also referred to as training. ANNs can continuously adjust parameters, such as connection weights, and store the sample set as a connection weight matrix under circumstance of external environment stimulation, such as the input of the sample set. When the ANN accepts the input again, the system can provide the appropriate output. In the present study, there were many neurons in the model and the sample size had rigorous requirements. Therefore, only the variables that were selected by single factor analysis and closely related to the prognosis of advanced schistosomiasis were used as input variables. A good predictive model can distinguish population at high risk from the one at low risk, which is so called discrimination. Discrimination is generally expressed as the area under the ROC curve, referred to as AUC. The higher the AUC value, the better the model can discriminate between high and low risk groups. Due to the serious adverse prognosis of advanced schistosomiasis patients, the sensitivity of the predictive model should be as high as possible in order to avoid false negatives on condition that the discrimination of the model is fine (e.g. AUC≥0.75).
Data from the designated training set was then used to evaluate the ANN model, and the prediction accuracy of the ANN model was 0.8660, which was better than the LR model (0.7990,) and the DT model (0.8194). The AUC of the ANN, LR, and DT models was 0.9267, 0.8276, and 0.8229, respectively, which indicates that the ANN model had the best prediction performance by Mann–Whitney U test.
In comparison to the LR and DT models, the ANN model had the best fitting effect for the relationship between advanced schistosomiasis and pathogenic factors. Schistosomiasis’ pathogenesis of disease is a complicated process influenced by multiple factors; thus, the use of traditional LR models to predict the development of disease is significantly limited by the inability to determine effects of multiple co-linearity between the independent variables. DT models can be easily applied to discrete values, but when there are more attribute values, the effect may be poor [35]. While ANN models can handle more attribute values, they have the potential to over-fit effects and their network training speed can decrease when there are more independent variables [36].
Despite its limitations, the LR model has been widely adopted because it offers other advantages [37, 38]. LR models have the function of discrimination and prediction and LR models are suitable for qualitative and semi-quantitative indicators [39]. In addition, LR models can use log transformation to convert nonlinear relationships between dependent variables and independent variables into linear relationships, which has less restriction conditions and a relatively low requirement of data types. To build predictive models, LR frameworks can automatically select highly correlated indices to be included as independent variables in the equation, which makes LR models convenient, feasible, and easy to popularize[40, 41]. It should be noted that once we develop a LR model in medical practice, it always means the LR model for every disease itself rather than for any disease.
In comparison to LR models, DT models can not only detect statistically significant risk factors, the model can also intuitively compare the intensity of various risk factors on the prognosis of patients with advanced schistosomiasis [42, 43]. The DT algorithm can simultaneously handle diverse types of data and missing data values without having to address the parameters in advance. DT models have a fast training speed, high classification efficiency, and ability to handle large sets of complex non-linear data [44–46].
ANN simulates the function and structure of biological neural network to establish non-linear mathematical models with strong fault tolerance, adaptiveness, nonlinear comprehensive reasoning ability, and the powerful ability to solve co-linearity and interactions between variables [47, 48]. Although complex relationships often exist between output and input factors in the medical field, ANNs have been used in clinical settings to effectively solve this issue and successfully applied to large and complex sample statistics.[49–51]. ANN models can not only realize the objective detection and classification of disease, but they can also improve the efficiency of disease prognosis and differential diagnosis. While the predictive ability of ANNs has many advantages, the model still has several limitations. First, the network changes with the setting of parameters, functions, and initial values. The correctness of these settings lack a theoretical basis, as the settings can only be determined by experience and repeated tests. Second, unlike the LR model, the ANN model does not have a recognized model of input variable access and elimination. Third, as a result of their structure, ANN models do not provide any medical explanation pertaining to each independent variable; thus, the hypothesis test methods, confidence intervals, and other issues require additional research [52, 53].
The advantages and disadvantages between these models on the implementation of them in the medical practice are noteworthy. A study that used ANN models and generalized additive models (GAM) to estimate glomerular filtration rate (GFR) in patients with chronic kidney disease found that the advantage of ANN is obvious only when multiple variables added to the model, especially the multicollinearity existed [54]. ANN is difficult to solve the problem of internal authenticity (repeatability) within the model due to the single data set source. However, the advantages of the ANN model over LR were also demonstrated: dealing with noise and incomplete input variables, high fault tolerance and good generalizability. LR model still plays an important role in the study of prognosis of disease due to its better interpretability. In a study that used large national samples to find the cause of arthritis pain, the DT model incorporated more than 200 variables with a high accuracy of 85.68% [55]. In the era of big data, the DT model facilitates algorithms transforming from hypothesis-driven to data-driven. Like ANN model, the robustness of DT model is better when there are more covariables [56]. Tree models can produce visual classification rules which are closer to people's way of thinking. However, DT model also has its disadvantages such as potentially introducing bias due to division of the tree every time, with the other drawbacks of high variance and instability.
The present study constructed three predictive models—the ANN model, the LR model, and the DT model—to predict advanced schistosomiasis prognosis. While each of the predictive models proved effective and had their own advantages, the ANN model outperformed the LR and DT models in terms of AUC and sensitivity. However, to achieve the highest level of prediction accuracy and better assist medical professionals, the three predictive models should be applied after model comparison.
|
10.1371/journal.pgen.1002175 | Histone Crosstalk Directed by H2B Ubiquitination Is Required for Chromatin Boundary Integrity | Genomic maps of chromatin modifications have provided evidence for the partitioning of genomes into domains of distinct chromatin states, which assist coordinated gene regulation. The maintenance of chromatin domain integrity can require the setting of boundaries. The HS4 insulator element marks the 3′ boundary of a heterochromatin region located upstream of the chicken β-globin gene cluster. Here we show that HS4 recruits the E3 ligase RNF20/BRE1A to mediate H2B mono-ubiquitination (H2Bub1) at this insulator. Knockdown experiments show that RNF20 is required for H2Bub1 and processive H3K4 methylation. Depletion of RNF20 results in a collapse of the active histone modification signature at the HS4 chromatin boundary, where H2Bub1, H3K4 methylation, and hyperacetylation of H3, H4, and H2A.Z are rapidly lost. A remarkably similar set of events occurs at the HSA/HSB regulatory elements of the FOLR1 gene, which mark the 5′ boundary of the same heterochromatin region. We find that persistent H2Bub1 at the HSA/HSB and HS4 elements is required for chromatin boundary integrity. The loss of boundary function leads to the sequential spreading of H3K9me2, H3K9me3, and H4K20me3 over the entire 50 kb FOLR1 and β-globin region and silencing of FOLR1 expression. These findings show that the HSA/HSB and HS4 boundary elements direct a cascade of active histone modifications that defend the FOLR1 and β-globin gene loci from the pervasive encroachment of an adjacent heterochromatin domain. We propose that many gene loci employ H2Bub1-dependent boundaries to prevent heterochromatin spreading.
| The transcription of genes in eukaryotes occurs within the context of chromatin, a complex of DNA, histone proteins, and regulatory factors. Whole-genome profiling of chromatin proteins and histones that are post-translationally modified has revealed that genomes are organized into domains of distinct chromatin states that coordinate gene regulation. The integrity of chromatin domains can require the setting of their boundaries. DNA sequences known as chromatin insulator or boundary elements can establish boundaries between transcriptionally permissive and repressive chromatin domains. We have studied two chromatin boundary elements that flank a condensed chromatin region located between the chicken FOLR1 and β-globin genes, respectively. These elements recruit enzymes that mediate the ubiquitination of histone H2B. Histone H2B ubiquitination directs a cascade of so-called “active” histone modification events that favor chromatin accessibility. We observe a striking collapse of the active histone modification signature at both chromatin boundaries following the depletion of ubiquitinated H2B. This loss of boundary function leads to the comprehensive spreading of repressive chromatin over the entire FOLR1 and β-globin gene region, resulting in gene silencing. We propose that chromatin boundaries at many gene loci employ H2B ubiquitination to restrict the encroachment of repressive chromatin.
| There is growing consensus that the non-random chromosomal arrangement of genes in higher eukaryotes enables the sharing of specific chromatin environments that facilitate co-regulation.
Recent genomic profiling of histone modifications, chromatin factors and nuclear proximity in Drosophila and mammalian cells have revealed prevalent organization of genes into domains, or neighborhoods, of common chromatin state [1]–[5]. Genes taken out of their natural chromosomal environment become deregulated in a variety of human genetic diseases [6]. This so-called chromosomal position effect also underlies the variable expression of transgenes depending on their site of integration [7].
The maintenance of chromatin domain integrity can require the setting of boundaries. Boundaries not only allow the partitioning of gene regulation, but also may also maintain the concentration of factors required for heterochromatin structures and normal genome homeostasis [8]. Fixed chromatin boundaries can be established by DNA sequence elements called insulators, which function to protect genes from inappropriate signals emanating from their surrounding environment [9]–[12]. HS4 is a well characterized element that has served as a paradigm for the study of insulators in vertebrates. HS4 lies at a boundary between the chicken β-globin gene cluster and upstream region of condensed chromatin that is enriched in the epigenetic hallmarks of heterochromatin [13]–[15]. A 275 bp core of the HS4 element has two separable activities that functionally define insulators: it can block the action of an enhancer element on a linked promoter when positioned between the two and it can act as a barrier to chromosomal position effect silencing [16]–[18]. The enhancer blocking and barrier activities of HS4 involve different proteins and mechanisms and are separable in assay systems. The CTCF binding site footprint II (FII) is necessary and sufficient for enhancer blocking, but can be deleted from HS4 without affecting barrier activity [18]–[20]. HS4 requires a USF1/USF2 binding site (FIV) and three VEZF1 binding sites (FI, FIII and FV) for its barrier activity, which control histone modifications and DNA methylation, respectively [21]–[24].
HS4 manipulates histone modification signatures to counteract gene silencing [22], [24]. HS4 has been found to be persistently enriched in high levels of H3 and H4 acetylation, H3-lysine 4 methylation, H4-arginine 3 methylation and acetylated histone variant H2A.Z regardless of neighboring gene expression [13]–[14], [23], [25]. We proposed that the active histone modifications at HS4 collectively act as a chain terminator to heterochromatin assembly by interfering with the propagation of repressive histone modifications [24].
Given that chromosomal silencing has been shown to be processive and stable, we reasoned that the HS4 element needs to act as a constitutive barrier if it is to effectively shield the locus. In this study, we address a hypothesis that HS4 might recruit histone modifications that act as master controllers of the active chromatin state to facilitate barrier stability. Intense study in recent years has begun to unravel the complex language of crosstalk between histone modifications during the establishment of different chromatin states [26]. Principal among the active histone modifications is the monoubiquitination of H2BK120 (H2BK123 in S. cerevisiae), which is required for the tri-methylation of H3K4 [27]–[30]. H3K4me3 is a pivotal mark of the active chromatin state, by acting as a platform for the binding of multiple histone acetyltransferase, histone demethylase and nucleosome remodelling complexes [31]–[33]. We therefore investigate whether i) H2B ubiquitination directs a cascade of active histone modifications at the HS4 insulator, ii) this modification is required for its barrier activity, and iii) the integrity of the 3′ chromatin boundary of the condensed chromatin located upstream of the β-globin locus. We also extend our analysis to look at the 5′ chromatin boundary of the same condensed chromatin and its role in shielding the FOLR1 gene locus.
We sought to address whether histone H2B ubiquitination plays a key role in establishing and maintaining the boundaries of a condensed heterochromatin-like domain that separates the FOLR1 and β-globin gene loci. Firstly, we mapped the presence of ubiquitinated nucleosomes across 50 kb encompassing the chicken β-globin locus (Figure 1). We established native chromatin immunoprecipitation (N-ChIP) assays using nucleosomes prepared by micrococcal nuclease (MNase) digestion of chromatin in low salt conditions to ensure the retention of potentially unstable variant nucleosomes found at this locus [34]. We prepared di- and tri-nucleosomes using a range of MNase concentrations which ensured that they were representative of open and condensed chromosomal regions (Figure 1A, data not shown). The N-ChIP method strips away non-nucleosomal proteins (Figure 1B), which allows the analysis of ubiquitinated histones using anti-ubiquitin antibodies. The enrichment of nucleosomes containing the 25 kDa monoubiquitinated form of H2B was confirmed by western blotting (Figure 1C).
N-ChIP analysis of histone ubiquitination was performed on primary red blood cells (RBC) from 10 day chick embryos, in which the β-globin locus is highly transcriptionally active, but the 5′ folate receptor (FOLR1) locus is silent [15], [35]. We observe a striking enrichment of ubiquitinated histones specifically at the core HS4 insulator element (Figure 1D; p-value from student's t-test of enrichment = 2e−6). Perhaps surprisingly, no enrichment of histone ubiquitination was observed at the promoters or enhancers of the highly active β-globin genes in RBCs. We also mapped histone ubiquitination in 10 day embryo whole brain tissue (Figure 1E), where both the FOLR1 and β-globin genes are reported to be silent [15]. We also observe a specific enrichment of ubiquitinated histones specifically at the core HS4 insulator element in brain tissues (Figure 1E, p-value = 6e−5). We also observe significant levels of histone ubiquitination at the FOLR1 gene regulatory elements HSA and HSB in both 10 day embryo RBCs (Figure 1D, p-value = 0.007) and whole brain (Figure 1E, p-value = 2e−5). These elements are situated between the FOLR1 gene and the condensed region and may harbor chromatin boundary activity.
We sought to determine which of the well characterized activities of the 275 bp core HS4 element are required for the recruitment of histone ubiquitination. We performed N-ChIP analyses of histone ubiquitination at HS4 insulators present on single copy transgenes stably integrated into the early erythroid CFU-E stage cell line 6C2 [18]. We find that transgenic HS4 insulators are enriched in histone ubiquitination at a level equivalent to the endogenous HS4 element (WT, Figure 1G). Histone ubiquitination is therefore likely to be recruited by one of the factors that mediate the insulator functions of the core HS4 element. We performed N-ChIP analysis of single copy transgenic HS4 elements that are mutated at the CTCF (FII), VEZF1 (FIII) or USF1 (FIV) binding sites. These mutations have been extensively characterized and disrupt HS4's ability to mediate enhancer blocking, protection from DNA methylation or active histone modification, respectively [19], [21], [24]. We find that the USF1/USF2 binding site, footprint IV, is required for the recruitment of histone ubiquitination (Figure 1G). This correlates with our previous finding that the USF site was also required for H3K4 methylation of the HS4 insulator [24].
The histone ubiquitination that is enriched at the HS4 element may occur on any of the core histones. We anticipated that histone H2B is subject to this modification as HS4 is constantly enriched in methylated H3K4 [13], [24], and H2BK120 mono-ubiquitination is required for proper H3K4 methylation [27]–[30]. Using crosslinking ChIP with recently developed antibodies, we confirmed that the core HS4 insulator was enriched in H2BK120ub1 in the early erythroid CFU-E stage cell line 6C2 (Figure 2A). We were unable to detect any enrichment of H2AK119ub1 at HS4 or other β-globin sequences (not shown).
We sought to identify the E3 ligase responsible for H2Bub1 at the HS4 element so that the effects of depleting H2Bub1 could be studied. We used crosslinking ChIP analysis to show that RNF20 interacts with the core HS4 insulator element in erythroid cells (Figure 2B). Chicken RNF20 (BRE1A), is 90% identical to human RNF20/BRE1A, which is an E3 ligase responsible for efficient H2B ubiquitination (Figure S1A) [36]–[37]. The presence of RNF20 therefore suggests that this enzyme is required for the enrichment of H2BK120ub1 at the HS4 insulator.
Next, we investigated whether H2Bub1 levels can be depleted following RNAi of RNF20. It was important that we were able to knockdown RNF20 levels for prolonged periods as this would allow the study of progressive repression of the β-globin locus and insulated transgenes. This was achieved in 6C2 cells using a lentiviral vector system for doxycycline-regulated expression of miRNA-shRNA (Materials and methods, Figure S1). After four days of shRNA expression, RNF20 protein levels were reduced to ∼20% of wild type (Figure 2C). We saw little change in the whole cell protein levels of the HS4-binding proteins CTCF, VEZF1 or USF1. We studied the effect of this short term knockdown of RNF20 expression on a panel of histone modifications in total chromatin. We found that H2BK120ub1 levels in chromatin were reduced by ∼80% (Figure 2D). RNF20 knockdown did not affect H2AK119ub1 levels. This confirmed that chicken RNF20 is a H2B-specific ubiquitin E3 ligase like its Bre1 orthologs. The reduction of H2B ubiquitination resulted in substantial reductions in H3K4me3 (70%) and H3K79me2 (53%) levels, and a minor reduction in H3K4me2 (20%) (Figure 2D). This demonstrates that chickens also employ the same trans-histone crosstalk pathways observed in yeast and mammals [26]. H3K9acK14ac was slightly reduced (7%), but the levels of other modifications associated with active or repressive chromatin remained largely unchanged (Figure 2D).
To determine whether RNF20 was responsible for all the histone ubiquitination observed at the HS4 and FOLR1 elements, we performed N-ChIP analysis across the FOLR1 and β-globin region before and after RNF20 knockdown in 6C2 cells. Consistent with our observations in primary 10 day embryo tissues, we find that the HS4 insulator and the FOLR1 HSA/HSB elements are substantially enriched in histone ubiquitination in 6C2 cells (Figure 2E). In addition, there is elevated histone ubiquitination across the FOLR1 gene, which is highly active in 6C2 cells, consistent with co-transcriptional deposition (Figure 2E). We observed a substantial depletion of histone ubiquitination at the HS4 insulator and FOLR1 HSA/HSB elements following four days of RNF20 knockdown (Figure 2E; p-values from student's t-test of difference between WT and RNF20kd are 2e−5 and 0.002, respectively). We observe similar profiles of RNF20-dependent H2Bub1 in 6C2 cells (Figure S2). The histone ubiquitination observed at HS4 and the FOLR1 regulatory elements is therefore RNF20-dependent H2B monoubiquitination.
The HS4 insulator is marked by an assemblage of histone modifications and variants typically associated with transcriptionally permissive open chromatin; H3K9acK14ac, H4K5acK8acK12acK16ac, H3K4me2, H3K4me3, H4R3me2as, H2A.ZK4acK7acK11ac, and H2BK120ub1 [13]–[14], [23] and this study). This active modification signature is a constant feature of HS4 in a variety of cell types irrespective of local gene expression. A very similar chromatin signature is observed at the FOLR1 HSA/HSB regulatory elements in 6C2 cells. We hypothesize that H2Bub1 may be the keystone for the deposition of the active histone signature at these elements. We therefore performed N-ChIP analysis of active and repressive modifications across the 50 kb β-globin gene neighborhood following short term knockdown of H2B ubiquitination.
We found that H3K4me2 and H3K4me3 enrichments at the HS4 insulator element were reduced by 40% and 70%, respectively, following RNF20 knockdown (21.540, Figure 3A and 3B). This is consistent with trans H2Bub1-H3K4me3 cross-talk occurring at HS4 nucleosomes. The depletion of H3K4me at HS4 is specific as the levels observed at the active FOLR1 gene promoter remain unchanged (5.613, Figure 3A and 3B). Strikingly, the loss of H2Bub1 also considerably impacts the hyperacetylation of multiple histones at the HS4 insulator, with H3ac, H4ac and H2A.Zac reduced by 55%, 60% and 70%, respectively (21.540, Figure 3C, 3D and 3F). The depletion of histone acetylation at the HS4 insulator is in contrast to the relatively unchanged levels of histone acetylation in bulk chromatin (Figure 2D). We note that very similar depletions in active modifications are also observed at the FOLR1 gene regulatory elements HSA/HSB. These regulatory elements may harbor functional properties similar to those of the HS4 insulator element. H2A.Z incorporation was mostly unaffected, but there was a 50% reduction in H2A.Z levels specifically at the core of the HS4 insulator (21.540, Figure 3E).
We investigated whether the depletion of H2Bub1 and the resulting loss of the active histone signatures at the HSA/HSB and HS4 elements affected the containment of the intervening condensed region. We determined that H3K9me2 and H3K9me3 are restricted to the condensed region upstream of HS4 in wild type 6C2 cells (8.9 to 17.7, Figure 3G and 3H, not shown). We find that after only four days of H2Bub1 depletion there is marked encroachment of H3K9me2 beyond the HS4 insulator. Significant H3K9me2 spreading into the β-globin locus is observed at all sites from the condensed region to the ρ-globin gene promoter (Figure 3G). Significant H3K9me2 spreading is also observed in the other direction, encompassing the FOLR1 promoter and gene body. No encroachment of H3K9me3 is observed after short term depletion of H2Bub1, but there is considerable consolidation of this mark at the edges of the condensed region (Figure 3H). The heterochromatin associated mark H4K20me3 is also enriched in the condensed region and at the 3′ end of the FOLR1 gene, but did not spread upon four days of RNF20 knockdown (Figure S3). Finally, we found that the gene silencing mark H3K27me3 was present at comparably low levels across the condensed region and β-globin locus in 6C2 cells, which did not alter upon RNF20 knockdown (data not shown). In summary, short term depletion of H2Bub1 is sufficient to disrupt H3K4me3 at the HSA/HSB and HS4 elements, which results in a rapid loss of multiple histone acetylation and chromatin boundary integrity. H3K9me2 appears as the first repressive mark to spread beyond the defective boundaries of the condensed heterochromatin region.
The comprehensive loss of active histone modifications at the HS4 boundary following RNF20 knockdown may be due to reduced binding of the insulator proteins that recruit histone modifying enzymes. We showed above that the expression of the insulator proteins is unaffected following RNF20 knockdown (Figure 2C). We therefore determined the binding of the insulator factors USF1, CTCF and VEZF1 to HS4 using crosslinking ChIP analysis before and after the loss of active modifications following RNF20 knockdown. We find that the binding of each factor is unaffected following RNF20 knockdown (Figure 4A–4C). We also discovered that the heterochromatin barrier factors VEZF1 and USF1 are also stably bound at the FOLR1 HSA and HSB elements, which contain binding motifs for both factors (Figure 4B and 4C). The FOLR1 region is not bound by the enhancer blocking and chromatin looping factor CTCF (Figure 4A). In summary, the disruption of insulator protein binding is not responsible for the comprehensive loss of active histone modifications at the HSA/HSB and HS4 elements.
We also addressed whether the depletion of H2Bub1 prevented the stable recruitment of histone methyltransferase (HMT) complexes that target H3-lysine 4. Existing models used to explain trans-tail crosstalk between H2Bub1 and H3K4me3 propose that the ubiquitination of H2B either regulates HMT residence by controlling nucleosome stability or creates a binding interface for HMT binding to chromatin (see discussion). We performed crosslinking ChIP analysis for RBBP5, a structural component of the SET1/COMPASS complex that interacts with USF1 [22]. We find that RBBP5 interacts with the HS4 insulator and the FOLR1 regulatory elements, all of which are sites of H3K4me3. The binding of RBBP5 to the HS4 or HSA/HSB boundary elements is not significantly affected by RNF20 knockdown (Figure 4D). The loss of H3K4me3 upon H2Bub1 depletion is therefore not due to the decreased residence of the core SET1 complex at HS4.
The ability of the HS4 element to shield genes from chromosomal position effect silencing in a wide variety of systems is well established [38]. This so-called barrier activity can be scored using a well established reporter transgene assay in erythroid cells [17]–[18], [24]. We used this assay to monitor the expression of a human IL-2R fragment from stably integrated transgenes (Figure 5A) using flow cytometry over time in culture. Non-insulated transgenes typically succumb to chromosomal silencing by 40–60 days of culture, whereas transgenes insulated by HS4 elements are able to maintain original levels of expression for 80 days and beyond [18]. We took extensively characterized stable lines that each contain a single copy of the IL-2R transgene flanked by paired 275 bp core HS4 insulators. It was previously determined that transgene expression from these cells remains constant beyond 80 days of culture, the transgenic HS4 insulators are bound by CTCF, USF1 and VEZF1 [21], and they are enriched in H3ac, H4ac, H3K4me [18], [24] and H2Bub1 (Figure 1G).
We transduced early passage IL-2R transgenic cells with lentiviruses that express RNF20 shRNA. The lentiviral miRNA-shRNA system we employed allowed the stable knockdown of RNF20 for at least sixty days (validated by Western blotting). We observed no change in 6C2 cell morphology and only a minimal reduction in cell doubling during this period (not shown). We found that four days of RNF20 knockdown had no effect on transgene expression (day 4, Figure 5B, 5D). The depletion of H2Bub1 therefore has little direct effect on the transcription rate of the transgene. However, transgene expression became progressively silenced with continued depletion of H2Bub1, with the IL-2R expression levels in independent transgenic lines falling by 50–60% after long term depletion (Figure 5B, 5D). This level of silencing is less than that observed when flanking insulators are absent or mutated [18], but is comparable to that observed in cells transfected with AUSF, a truncated form of USF1 that dominantly inhibits USF1 function [22]. Thus, constant H2B ubiquitination is required for HS4 to act as a stable barrier to chromosomal silencing.
It has been postulated that the HSA/HSB regulatory region and the HS4 insulator might form chromatin boundaries that protect the FOLR1 and β-globin genes from the encroachment of the potentially repressive condensed chromatin that separates these loci [15], [24]. We have therefore studied how long term depletion of H2Bub1 impacts on the containment of heterochromatin associated marks at these loci. We maintained the induction of RNF20 knockdown for forty days, which reduced proteins levels to 9% of wild type, compared to 19% seen after four days of knockdown (Figure 6A, Figure 2C). The prolonged RNF20 knockdown resulted in the depletion of H2BK120ub1 in total chromatin to 13% of wild type levels (Figure 6B). This in turn, resulted in considerable reductions in total H3K4me2 and H3K4me3, reduced by 78% and 77%, respectively (Figure 6B). Conversely, we observe 43% and 39% increases in the heterochromatin marks H3K9me3 and H4K20me3 in total chromatin (Figure 6B). This is in clear contrast to the unchanged levels of heterochromatin marks after short term knockdown (Figure 2D). Interestingly, the incorporation of the variant histone H2A.Z in total chromatin also increased by 24% after prolonged RNF20 depletion, perhaps to compensate for the gross shift from active to repressive chromatin across the genome (Figure 6B).
We performed N-ChIP analyses of histone modifications across the FOLR1 and β-globin loci to determine the effects of long term RNF20 knockdown on chromatin domain integrity. Firstly, we confirmed that H2BK120ub1 was depleted from the HS4 insulator (Figure 6C). The levels of H3K4me2 and H3K4me3 at HS4 were greatly depleted (by 80% and 65%, respectively) as a result of the long term depletion of H2BK120ub1 (Figure 6D, 6E). The loss of active modifications at HS4 for a prolonged period results in extensive encroachment of the heterochromatin associated marks H3K9me3 and H4K20me3, which are normally restricted to the condensed region between the FOLR1 and β-globin loci. Strikingly, H3K9me3 spreads beyond HS4 to encompass the entire 33 kb β-globin locus (Figure 6F). H3K9me3 spreading is likely to have occurred in the majority of cells in the population as the enrichment levels over the β-globin locus are comparable to those in the upstream condensed region. H3K9me3 spreading is also observed in the opposite direction, with significant increases in this mark over the FOLR1 promoter and gene body (Figure 6F). Furthermore, H4K20me3 is also observed to spread from the upstream condensed region to cover the FOLR1 gene in one direction and as far as the ρ-globin promoter in the other (Figure 6G).
H3K9me2, H3K9me3 and H4K20me3 are widely associated with gene silencing and heterochromatin formation. The encroachment of these marks over the FOLR1 and β-globin genes following RNF20 depletion may result in the silencing of their transcription. While the β-globin locus is becoming primed for expression at the CFU-E progenitor stage represented by 6C2 cells, the β-globin genes themselves are not expressed until terminal differentiation [15], [35]. 6C2 cells cannot be induced to terminally differentiate, so we are unable to study the impact of heterochromatin spreading on the activation of β-globin gene transcription in this system. We therefore focused our attention on the expression of the FOLR1 gene, which is active in 6C2 cells [15]. RT-PCR analysis shows that FOLR1 expression is not affected by four days of RNF20 knockdown (Figure 7). This is despite the depletion of H2Bub1, H3K4me2/3, H3ac, H4ac, H2A.Zac at the HSA/HSB regulatory region (Figure 3A–3D, 3F) and the encroachment of H3K9me2 across the FOLR1 promoter and gene body (Figure 3G). Closer inspection shows that H3K4me2/3 and H4ac of the FOLR1 promoter are unaffected following short term RNF20 depletion. FOLR1 gene transcription is therefore not directly dependent upon RNF20 or on maximal active histone modifications at the HSA/HSB elements. However, we find that FOLR1 gene transcription is progressively silenced with prolonged RNF20 knockdown, with 94% repression observed after sixty days of knockdown (Figure 7). The silencing of FOLR1 coincides with both the loss of H3K4me2/3 at its promoter (Figure 6D and 6E) and the accumulation of H3K9me3 and H4K20me3 over its promoter and gene body upon heterochromatin spreading (Figure 6F and 6G).
Taken together, these findings demonstrate that the elements HSA/HSB and HS4 form the boundaries of the condensed chromatin region between the FOLR1 and β-globin gene loci. They employ an H2Bub-dependent active chromatin signature that protects these genes from the encroachment of multiple heterochromatin associated marks. The spreading of H3K9me3 and H4K20me3 coincides with the silencing of the FOLR1 gene.
The first high resolution maps of histone modifications across gene loci during vertebrate development revealed that the well characterized chromatin boundary marked by the HS4 insulator is constitutively enriched with histone modifications associated with open chromatin [13]–[14]. Here we show that the HS4 insulator is also constitutively marked by H2BK120 mono-ubiquitination. We show that RNF20-dependent H2Bub1 is required not only for H3K4me2/3 at HS4, but also for multiple acetylation of H3, H4 and H2A.Z at this element (Figure 8A). A very similar H2Bub1-dependent active histone signature is also found at the HSA/HSB elements upstream of the FOLR1 gene. To our knowledge, this is the first example of H2Bub1 directing such an extensive cascade of trans histone tail modifications at specific gene regulatory elements.
HSA/HSB and HS4 mark the 5′ and 3′ flanks of the condensed chromatin region between the FOLR1 and β-globin loci, which is enriched in the epigenetic hallmarks of heterochromatin ([13]–[15], this study) (Figure 8B). The loss of the active histone modification signature at these elements following the depletion of H2Bub1 in erythroid cells results in the progressive spreading of multiple repressive histone marks across the entire FOLR1 and β-globin loci. These findings clearly demonstrate that the elements HSA/HSB and HS4 form the boundaries of the condensed chromatin region between the FOLR1 and β-globin gene loci. The ability of the HS4 insulator to shield transgenes from chromosomal silencing in a wide variety of systems is well established [38], but this study provides firm evidence that HS4 functions as a chromatin boundary element in its endogenous context. Both the endogenous and transgenic HS4 elements require continued deposition of H2Bub1 to maintain chromatin boundary integrity and chromosomal position effect protection, respectively.
It has been unclear for some years how the FOLR1 gene locus is defended from heterochromatin spreading. An earlier study demonstrated that a 3.7 kb region that encompasses the FOLR1 promoter and upstream regulatory elements is capable of directing strong copy number-dependent expression of randomly integrated transgenes in chicken erythroid cells [15]. This fragment contains the major promoter-proximal element HSA and an additional DHS, which we have named HSB (Figure 1F). The elements may harbor locus control region (LCR)-like enhancer and/or chromatin boundary activities. Our observations are consistent with the latter. We find that the HSA and HSB elements are bound by the HS4 barrier proteins USF1 and VEZF1, they recruit RNF20 and the SET1 complex and establish an H2B-dependent active histone modification signature. These molecular features mirror those at the HS4 element. A key different between the HSA/HSB and HS4 boundary elements is the absence of CTCF binding at the FOLR1 boundary. This indicates that CTCF is not required to act a barrier to the spreading of heterochromatin from the condensed region. This is consistent with our previous findings that the CTCF binding site of the HS4 insulator is dispensable for its ability to act as a barrier to chromosomal silencing in different assay contexts [18], [20].
The modification of histones at chromatin boundaries is conserved across eukaryotes. It is well established that several histone acetyltransferases (HATs) are required for heterochromatin boundary integrity in budding yeast [10], [39]–[40]. Indeed, artificial tethering of HAT chimeras is sufficient to create synthetic barriers to heterochromatin-mediated gene silencing [41]. It has also recently been found that the ILB barrier element at the Drosophila reaper locus also recruits histone acetylation [42]. Our observations that the depletion of multiple histone acetylation marks results in chromatin boundary failure at the chicken FOLR1 and β-globin loci adds further support for a conserved role for active histone modification in chromatin boundary formation. The finding that multiple active histone modifications at the HSA/HSB and HS4 elements are directly or indirectly dependent upon prior H2B ubiquitination is particularly striking. Given the conservation of the factors that mediate H2B ubiquitination and the trans-histone H2Bub1-H3K4me3 pathway, we anticipate that this modification will be employed at boundaries across eukaryotes. The finding that artificial tethering of Lge1, a factor required for H2B ubiquitination and H3K4/K79 methylation, is sufficient to create a synthetic barrier to heterochromatin-mediated gene silencing in budding yeast supports this view [41], [43].
A number of budding yeast boundary elements are also associated with regions of nucleosome depletion and elevated histone turnover [40], [44]–[45]. This may be related to the incorporation of the histone variant H2A.Z, which supports heterochromatin boundary integrity [10], [46]. However, we did not observe any extensive depletion in histone density at the chicken HSA/HSB or HS4 chromatin boundaries (not shown). Furthermore, we found that the incorporation of H2A.Z at these boundaries remains intact following RNF20 knockdown and the loss of active modifications. Further studies are required to determine the role of H2A.Z at these elements, but it is clear that H2A.Z incorporation is not sufficient to prevent the spread of heterochromatin into the FOLR1 and β-globin loci.
We have shown that the trans histone modification pathway from H2Bub1 to H3K4me3 reported in yeast and man is also conserved in chicken. How the mono-ubiquitination of H2B facilitates H3K4me3 has been subject to intense study over the last few years. Three models have arisen to explain this pathway. The ‘wedge’ model postulated that the bulky ubiquitin moiety would increase the access of H3K4 methyltransferases by non-specifically disrupting chromatin fiber packing in some way [47]–[48]. This simple mechanism appears improbable as substitution of ubiquitin with the bulkier SUMO moiety at the equivalent residue of H2B does not recapitulate H2Bub1-directed trans tail crosstalk in S. cerevisiae [49]. In contrast, a ‘stability’ model was recently put forward in response to findings in S. cerevisiae that H2Bub1 promotes nucleosome reassembly following RNA polymerase II transcription and enhances global nucleosome stability [49]–[51]. It is proposed that H2Bub1 may restrict the eviction of the H2A/H2B dimer from nucleosomes, thereby increasing the nucleosomal residence of the SET1/COMPASS methyltransferase complex which interacts with basic and acidic patches on H2A and H2B, respectively [52]–[53]. In this study, we found that the interaction of RBBP5 (SWD1), a core component of the SET1 complex remains bound at the HS4 insulator following the depletion of H2Bub1. The loss of H3K4me2/3 cannot be explained by the decreased residence of SET1 complexes.
Recent studies provide compelling evidence that H2Bub1 acts as ‘bridge’ to facilitate H3K4me3. The core SET1 complex can interact with chromatin and mediate H3K4 mono-methylation in the absence of H2B ubiquitination [54]. However, it has been found that the accessory COMPASS subunit Cps35/Swd2 in yeast (WDR82 in humans) interacts with H2Bub1 and activates the processive H3K4 methyltransferase activity of the SET1 complex [55]–[56]. While the composition of SET1 complexes in chickens remains to be determined, our data are consistent with a mechanism of H2Bub1-directed activation of pre-loaded SET1 complexes to facilitate processive H3K4 methylation. The loss of H3K4me2/3 upon the depletion of H2Bub1 is likely to be the primary reason for the subsequent losses of multiple histone acetylation at the HS4 insulator. Methylated H3K4 is a pivotal recognition site required by multiple histone acetyltransferase complexes [57]–[60]. H3K4me3 also facilitates the recruitment of the NURF chromatin remodelling complex via its BPTF subunit [22], [61].
The mono-ubiquitination of H2B is broadly recognized as a mark of transcriptional activity [30]. H2Bub1 is enriched in the bodies of expressed genes throughout yeast and mammalian genomes [62]–[63], and the bulk of H2Bub1 requires many factors involved in the early steps of transcription elongation [30]. While the HS4 insulator has the epigenetic chromatin signature of a housekeeping promoter, it lacks either promoter or enhancer activity [64]. In addition, HS4 is not bound by RNA polymerase II (Figure S4) and is not a source of transcripts [22], [65]. It therefore appears most likely that HS4 recruits H2Bub1 through a process that is not linked to transcription. We found that the recruitment of H2Bub1 to HS4 is dependent upon the USF1/USF2 binding site. While we have been unable to detect RNF20 in stable complexes with USF1 or USF2 (data not shown) [22], this is reminiscent of activator-dependent recruitment of Bre1/RNF20 to yeast and human promoters [66].
In addition to the recruitment of the E3 ubiquitin ligase RNF20, the HSA/HSB and HS4 boundary elements also require sufficient activity levels of the E2 conjugase RAD6 to enable sufficient levels of H2Bub1 for chromatin boundary stability. There are two broad mechanisms that could result in the persistent H2B ubiquitination of the HS4 insulator. Firstly, the HSA/HSB and HS4 elements might not be subject to the rapid turnover of H2Bub1 associated with promoter clearance and transcription elongation [66]–[67]. Such a scenario would negate the need for the co-transcriptional stimulation of RAD6 conjugase activity [68], as low efficiency H2B ubiquitination may be sufficient for high steady state levels of H2Bub1 at HS4. Alternatively, HS4 may recruit factors that mediate RAD6 phosphorylation in the absence of RNA polymerase to stimulate efficient H2B ubiquitination of this element.
Depletion of H2Bub1 disrupts the assembly of the active histone modification signatures at the HSA/HSB and HS4 boundary elements. This results in progressive spreading of heterochromatin-associated histone marks into the FOLR1 and β-globin loci either side of the condensed region. We find that the heterochromatin-associated marks H3K9me2, H3K9me3 and H4K20me3 are propagated in a continuous manner from the upstream condensed region into the FOLR1 and β-globin loci. The heterochromatin domain expands from a ∼10 kb domain of the condensed region to cover the entire ∼50 kb FOLR1 and β-globin region given time (Figure 8B).
Intriguingly, each of the three repressive marks at the β-globin locus spreads in a different temporal manner, suggesting that different enzyme complexes are involved in propagating these marks. The first repressive mark to spread is H3K9me2, which propagates over the entire FOLR1 locus and extends 14 kb into the β-globin locus after only four days of H2Bub1 depletion (Figure 8B). Conversely, H3K9me3 and H4K20me3 do not extend beyond the upstream condensed region at this early stage, but H3K9me3 appears to consolidate at the borders of the condensed region. However, both H3K9me3 and H4K20me3 spread into the FOLR1 and β-globin loci upon longer periods of H2Bub1 depletion. H3K9me3 uniformly spreads to encompass the entire ∼50 kb FOLR1 and β-globin region, while H4K20me3 spreads into a ∼ 30 kb region covering the entire FOLR1 locus to the rho gene promoter (Figure 8B).
Several mechanisms have been proposed to explain the spreading of repressive chromatin [69]. Given that the de novo repressive marks in the FOLR1 and β-globin loci manifest as continuous domains with consistent modification levels throughout, we speculate that the marks are propagated via linear cis-spreading mechanisms. The simplest way to rationalize all our findings is that the spreading occurs using a classical stepwise assembly mechanism, where sequential iterations of repressor protein binding and methyltransferase recruitment propagate the repressive methyl mark onto neighboring nucleosomes. The ability of HP1 adaptor proteins to recognize H3K9me3, interact with H3K9 and H4K20 methyltransferases and spread from sites of recruitment is a potential example of the self-reinforcing repressor interactions that may occur at the FOLR1 and β-globin loci [70]–[72]. Stepwise assembly mechanisms are consistent with the observed sequential pathway of repressive chromatin modification. It is possible that the propagation of H4K20me3 might require prior H3K9me3, which requires prior H3K9me2 at this locus. While further investigations will be required to define the exact pathway of repressive mark assembly, it is clear that HS4 acts a chain terminator to heterochromatin spreading by using a panel of active histone modifications, which collaborate to block and inhibit repressive histone methylation.
There may be a role for RNA-directed heterochromatin assembly at the FOLR1 and β-globin loci. It was recently shown that the maintenance of heterochromatin region's condensed conformation requires RNAi factors [65]. This suggests conservation with mechanisms employed in fission yeast where RNAi factors work together with heterochromatin proteins including the HP1 homolog Swi6 to mediate heterochromatin establishment [8]. However, RNAi factors are dispensable for the maintenance of heterochromatin [73]. It remains to be investigated whether RNA and RNAi factors play a role in heterochromatin spreading in higher eukaryotes.
It is also conceivable that the continuous spreading of heterochromatin is dictated by the three dimensional organization of the FOLR1 and β-globin loci. If these gene loci are insulated from the upstream condensed region by positioning into different nuclear compartments, disruption of the active signature at the HSA/HSB and HS4 boundaries may result in the transfer of most or all the FOLR1 and β-globin loci into a repressive compartment. Such a scenario appears complex, however, as it would require the sequential transfer of these loci into different compartments rich in H3K9me2, H3K9me3 and H4K20me3 methyltransferases.
There is a paucity of information about the elements that form chromatin boundaries in vertebrates. Our finding that the HSA/HSB and HS4 boundary elements employ H2B ubiquitination to direct a cascade of active histone modifications suggests that genomic profiling of chromatin signatures will be a useful approach to identifying boundary elements. We note that there was a gross increase in total heterochromatin marks results when H2Bub1 is depleted for long periods. This observation suggests that many loci employ H2Bub1-dependent boundaries to heterochromatin spreading. It will be interesting to see whether other chromatin boundaries in vertebrate genomes also require such a large complement of active marks or employ a more restricted palette to deal with locus-specific threats.
Antibodies against H3K4me2 (07-030), H3K4me3 (05-745R), H3K9me3 (07-523), H3K27me3 (07-449), H3 (07-690), H3K9acK14ac (06-599), H4K5acK8acK12acK16ac (06-598), H2AK119ub (05-678), H2A.Z (07-594) and CTCF (06-917) were obtained from Millipore. Antibodies against H3K9me2 (ab1220), H3K79me2 (ab3594), H3K79me3 (ab2621), H2A.ZK4acK7acK11ac (ab18262), PAF1 (ab20662), RPB1 (ab5408) and TBP (ab51841) were obtained from Abcam. Antibodies against H4K20me3 were a kind gift from Judd Rice [74]. Antibodies against ubiquitin (sc-8017), (BML-PW8805) and USF1 (H00007391-A01) were obtained from Santa Cruz, Enzo Life Sciences and Abnova, respectively. Antibodies against RNF20 (A300-715A) and RBBP5 (A300-109A) were obtained from Bethyl Laboratories. Anti-H2BK120ub1 antibodies were initially a kind gift from Moshe Oren [62] then purchased from Médimabs (MM-0029) or Millipore (17-650). Anti-VEZF1 antibodies were raised as described [21]. PE conjugated anti-CD25 (IL-2R) was obtained from Dako.
Chicken 6C2 erythroleukaemia cells were grown in αMEM supplemented with 10% FCS, 2% chicken serum, 1 mM HEPES, 25 µM β-mercaptoethanol and 1% Penicillin/Streptomycin solution. Crosslinking chromatin immunoprecipitation was performed as described previously [Litt et al, 2001]. Briefly, 6C2 cells (2×107 cells/ml) were crosslinked in fresh growth medium with 1% formaldehyde at room temperature for 20 minutes (RNF20, PAF1 and RBBP5), 10 minutes (CTCF, USF1 and VEZF1) or 2 minutes (H2Bub). Reactions were quenched by adding glycine to a final concentration 0.125 M. The crosslinked cells were washed by PBS twice and then lysed (0.25% Triton X-100, 10 mM EDTA, 0.5 mM EGTA and 10 mM Tris pH 8). Cell nuclei collected by centrifugation were washed (0.2 M NaCl, 1 mM EDTA, 0.5 mM EGTA and 10 mM Tris pH 8) followed by chromatin solubilization (0.5% SDS, 10 mM EDTA and 50 mM Tris pH 8). Chromatin was fragmented by sonication (Misonix) for a total time of 10 minutes in regular 10 second pulses. Insoluble material was removed by centrifugation at 15,000 g for 10 minutes at 4°C. Sizes of chromatin fragments were ∼500 bp on average.
Soluble chromatin was diluted by X-ChIP buffer (1.1% Triton X-100, 1.2 mM EDTA, 167 mM NaCl, 0.01% SDS and 16.7 mM Tris pH8) to obtain chromatin from 1×107 cells per ml. 1 ml of chromatin was pre-cleared with 5 µg of non-immune IgG and 100 µl (50% slurry in X-ChIP) of protein A/G agarose at 4°C for 3 hours. 10 µg of specific antibody was incubated with pre-cleared chromatin at 4°C with agitation overnight. Binding of protein A/G agarose was carried out at 4°C for 2 hours. The agarose was washed extensively with buffer 1 (1% Triton X-100, 0.1% SDS, 2 mM EDTA, 150 mM NaCl and 20 mM Tris pH 8), buffer 2 (1% Triton X-100, 0.1% SDS, 2 mM EDTA, 500 mM NaCl and 20 mM Tris pH 8), buffer 3 (0.25 M LiCl, 1% NP-40, 0.5% sodium deoxycholate, 1 mM EDTA, 10 mM Tris pH 8) and twice with TE buffer (10 mM Tris pH 8, 1 mM EDTA). The bound chromatin was eluted into elution buffer (1% SDS, 0.1 M NaHCO3), crosslinks reversed and protein digested. DNA was extracted by phenol/chloroform and ethanol precipitated in the presence of 10 µg of glycogen for quantitative PCR (qPCR) analysis.
Circulating red blood cells and whole brain tissue (without major vasculature) were collected from 10 day fertilized chick embryos kindly provided by Aviagen, Ltd. Native nucleosomes were prepared in low salt conditions to ensure retention of all nucleosomes, as described [34]. In brief, cells were collected in the presence of inhibitors (25 µg/ml AEBSF, 0.5 µg/ml Leupeptin and 0.7 µg/ml Pepstatin, 10 mM N-ethylmaleimide and 10 mM sodium butyrate) and nuclei were isolated by lysis buffer (10 mM NaCl, 3 mM MgCl2, 0.4% NP-40 and 10 mM Tris pH 7.5) for MNase (Sigma) digestion in the presence of 1 mM CaCl2. The MNase concentration (X) required to yield mostly di- and tri-nucleosomes was firstly determined. For ChIP experiments, three equal aliquots of nuclei were incubated with ½X, 1X and 2X MNase at 37°C for 17 minutes to obtain representative di- and tri-nucleosomes [14]. Digestion was stopped with 10 mM EDTA. Soluble chromatin was collected by centrifugation at 2,500 g for 5 minutes. The three supernatants were combined (S1). The remaining pellets were combined and resuspended in lysis buffer supplemented with 10 mM EDTA and left on ice for 15 minutes. Chromatin was released by passing through 20 then 25 gauge needles, and collected by centrifugation at 10,000 g for 10 minutes. The supernatant (S2) was combined with S1 for sucrose gradient fractionation. ∼1.5 mg of S1–S2 chromatin was fractionated on 13.5 ml 5∼25% linear sucrose gradients (Biocomp gradient master) in a SW40Ti rotor at 31,000 rpm for 14 hours at 4°C. 1 ml fractions were collected and 10 µl aliquots were extracted for checking DNA fragment sizes. Fractions containing di- and tri-nucleosomes were pooled and fixed with 0.1% formaldehyde at room temperature for 10 minutes. The crosslinking reaction was stopped with 0.125 M glycine. Nucleosomes were exchanged into N-ChIP buffer (50 mM NaCl, 5 mM EDTA, 10 mM Tris pH 7.5) buffer using P-6DG Bio-Gel (BioRad).
50 µg of nucleosomes were pre-cleared with 5 µg of non-immune IgG and 100 µl (50% slurry in N-ChIP buffer) of protein A/G agarose at 4°C for 3 hours. 10 µg of specific antibody was incubated with pre-cleared chromatin at 4°C with agitation overnight. Binding of protein A/G agarose was carried out at 4°C for 2 hours. Immunoprecipitated chromatin was collected and washed 5 times with 1 ml N-ChIP wash buffer (150 mM NaCl, 0.2 mM EDTA, 0.1% Tween-20 and 20 mM Tris pH 7.4). Chromatin was eluted with N-ChIP buffer supplemented with 1% SDS followed by 0.5% SDS. Eluates were digested with Proteinase K at 45°C for 2 hours and DNA extracted by phenol/chloroform and precipitated for qPCR analysis.
Relative DNA enrichments were quantified in triplicate by TaqMan real-time PCR on a Roche 480 Lightcycler. The primers used in this study were described previously [14], [21]. The comparative Ct method (with correction for primer efficiencies) was used to calculate fold enrichments and their standard deviations, as described previously [24]. Two sample equal variance Student's t-tests using a two-tailed distribution were applied to ChIP enrichment values to assess the significance of enrichments over controls, or changes following RNF20 knockdown. The calculated p-value ranges for enriched sites are indicated in the figure legends.
The pSLIK micro RNA-based lentiviral expression system was used to mediate long term conditional knockdown of RNF20 in chicken cells [75]. Gene-specific shRNAs are embedded into the primary transcript of human miR30, which is located in the 3′UTR of a doxycycline-regulated GFP transgene. The psm2 shRNA design tool was used to identify 20 potential shRNA targets (http://hannonlab.cshl.edu). These were scored and four targets were cloned into pEN_hUmiRc2, packaged into lentivirus particles and tested for performance as previously described [21]. GgRNF20-2628, which targets CAGAGTAACTAGAGAGAAA, was the most potent miR-shRNA and was used throughout this study. Wild type or 8103 (containing a single copy HS4 flanked IL-2R reporter transgene, [24]) 6C2 cells were transduced with GgRNF20-2628 lentiviral particles and cloned following flow sorting or serial dilution. GFP-miRNA expression was induced with 2 µg/ml doxycycline. GFP expression was monitored by FACS analysis to confirm expression of the miRNA cassette and RNF20 protein levels were monitored by Western blotting during prolonged knockdown time courses.
Nuclear extracts were prepared from cells harvested and washed with PBS and then lysed with hypotonic buffer (0.2% NP-40, 0.1 mM EDTA and 20 mM HEPES pH 8). Cell nuclei were collected by centrifugation at 2,500 g for 5 minutes. Nuclear proteins were extracted by incubation in high salt buffer (0.2% NP-40, 0.4 M NaCl, 13.3% glycerol and 20 mM HEPES pH 8) at 4°C for 30 minutes. Insoluble debris was removed by centrifugation at 16,000 g for 10 minutes at 4°C. Soluble nuclear extract was quantified by Bradford assay (Bio-Rad). 25 µg of nuclear extract or 7 µg of native S1–S2 nucleosome preparations were used for separation by SDS-PAGE. Proteins were then transferred to a PVDF membrane and imaged with HRP-conjugated secondary antibodies on a LAS-3000 imager (Fujifilm). Band intensities corrected for background were quantified using AIDA software.
106 cells were harvested by centrifugation at 1,000 g for 5 minutes. Cells were washed twice in 1 ml of Hank's buffered saline solution (Sigma) supplemented with 0.1% BSA and 0.1% NaN3 (HBSS+). Cells were resuspended in 100 µl of HBSS+ and incubated with 10 µl of anti-CD25-PE (Dako) antibody in the dark at 4°C for 30 minutes. Excess antibody was removed by washing twice with 1 ml of HBSS+. After the last wash, cells were resuspended in 500 µl of HBSS+ for FACS analysis. Flow cytometry was performed on a FACSCalibur flow cytometer (BD Biosciences) using CELLQuest software. RNF20 knockdown cells express GFP upon induction with doxycycline. Color compensation was used to correct for GFP fluorescence in the FL2 (585 nm) filter; it was set as FL1 – 1% FL2 and FL2–20% FL1. Data was acquired for 10,000 viable or GFP-expressing cells (FL1 = 530 nm). Histograms were generated using FlowJo software (Tree Star, Inc). Mean IL-2R fluorescence intensities of RNF20 knockdown cells were determined and normalized to those of parental reporter transgene cells with wild type RNF20 levels.
Total RNA was isolated from 6C2 cells using TRI reagent (Sigma) and cDNA prepared using SuperScript III (Invitrogen) and random hexamers. The following primers were used SYBR green real time PCR assays:
5′ TGCTGCGCTCGTTGTTGA
5′ CATCGTCCCCGGCGA
5′ ATGCGTCATCTCATCAGCAG
5′ TTGGGAAGAAGGGTCATCAG
5′ GATTCTGCATGTGCCACTGT
5′ AAGACCTGGGTGAAGGGTCT
|
10.1371/journal.pntd.0000266 | The Interplay between Entamoeba and Enteropathogenic Bacteria Modulates Epithelial Cell Damage | Mixed intestinal infections with Entamoeba histolytica, Entamoeba dispar and bacteria with exacerbated manifestations of disease are common in regions where amoebiasis is endemic. However, amoeba–bacteria interactions remain largely unexamined.
Trophozoites of E. histolytica and E. dispar were co-cultured with enteropathogenic bacteria strains Escherichia coli (EPEC), Shigella dysenteriae and a commensal Escherichia coli. Amoebae that phagocytosed bacteria were tested for a cytopathic effect on epithelial cell monolayers. Cysteine proteinase activity, adhesion and cell surface concentration of Gal/GalNAc lectin were analyzed in amoebae showing increased virulence. Structural and functional changes and induction of IL-8 expression were determined in epithelial cells before and after exposure to bacteria. Chemotaxis of amoebae and neutrophils to human IL-8 and conditioned culture media from epithelial cells exposed to bacteria was quantified.
E. histolytica digested phagocytosed bacteria, although S. dysenteriae retained 70% viability after ingestion. Phagocytosis of pathogenic bacteria augmented the cytopathic effect of E. histolytica and increased expression of Gal/GalNAc lectin on the amoebic surface and increased cysteine proteinase activity. E. dispar remained avirulent. Adhesion of amoebae and damage to cells exposed to bacteria were increased. Additional increases were observed if amoebae had phagocytosed bacteria. Co-culture of epithelial cells with enteropathogenic bacteria disrupted monolayer permeability and induced expression of IL-8. Media from these co-cultures and human recombinant IL-8 were similarly chemotactic for neutrophils and E. histolytica.
Epithelial monolayers exposed to enteropathogenic bacteria become more susceptible to E. histolytica damage. At the same time, phagocytosis of pathogenic bacteria by amoebae further increased epithelial cell damage.
The in vitro system presented here provides evidence that the Entamoeba/enteropathogenic bacteria interplay modulates epithelial cell responses to the pathogens. In mixed intestinal infections, where such interactions are possible, they could influence the outcome of disease. The results offer insights to continue research on this phenomenon.
| In amoebiasis, a human disease that is a serious health problem in many developing countries, efforts have been made to identify responsible factors for the tissue damage inflicted by the parasite Entamoeba histolytica. This amoeba lives in the lumen of the colon without causing damage to the intestinal mucosa, but under unknown circumstances becomes invasive, destroying the intestinal tissue. Bacteria in the intestinal flora have been proposed as inducers of higher amoebic virulence, but the causes or mechanisms responsible for the induction are still undetermined. Mixed intestinal infections with Entamoeba histolytica and enteropathogenic bacteria, showing exacerbated manifestations of disease, are common in endemic countries. We implemented an experimental system to study amoebic virulence in the presence of pathogenic bacteria and its consequences on epithelial cells. Results showed that amoebae that ingested enteropathogenic bacteria became more virulent, causing more damage to epithelial cells. Bacteria induced release of inflammatory proteins by the epithelial cells that attracted amoebae, facilitating amoebic contact to the epithelial cells and higher damage. Our results, although a first approach to this complex problem, provide insights into amoebic infections, as interplay with other pathogens apparently influences the intestinal environment, the behavior of cells involved and the manifestations of the disease.
| Once trophozoites of Entamoeba histolytica reach the host intestine, they can damage the mucosa epithelial layer and spread through the submucosa and the lamina propia and other tissues. Neutrophils and other cells infiltrate the tissue in the vicinity of amoebic lesions increasing the inflammatory response and tissue damage [1],[2]. In contrast, Entamoeba dispar, an amoeba that colonizes the human intestine together with E. histolytica and that is morphologically indistinguishable and genetically very similar to the latter, is not invasive and does not produce the clinical manifestations of an E. histolytica intestinal infection [3],[4],[5].
Search for expression of genes that could be correlated with the difference in pathogenicity between E. histolytica and E. dispar has mainly revealed higher expression in the former, of molecules involved in lysis of target cells , such as the amoebapore and specific cysteine proteinases [6],[7],[8]. Nonetheless, E. histolytica trophozoites can also remain as commensals in the intestinal lumen without causing manifestations of disease [9],[10].
It has been proposed that E. histolytica pathogenicity could be induced by ingestion of bacteria present in the host intestine. In vitro experiments have shown that after phagocytosis of an E. coli non-pathogenic laboratory strain (Ec346), trophozoites of E. histolytica increased their virulence together with their adhesive properties to target cells [11],[12],[13]. However, the same authors reported that long time cultivation with this bacteria strain rendered the amoebae less virulent [13]. In spite of the possible important role of intestinal bacteria in amoebic behavior in their natural habitat, little has been explored or elucidated about responses triggered by bacteria/amoeba interplay that could be important in the induction of tissue invasion and disease.
Pro-inflammatory cytokines released by cultured epithelial and endothelial cells after viral or bacterial infections induce structural and functional alterations in non-infected cells. These alterations lead to increased monolayer permeability and disarray of intercellular junctions and the cortical cytoskeleton which would facilitate passage of pathogens [14],[15],[16],[17]. Experimental infection of intestinal animal models with E. histolytica trophozoites has shown that pro-inflammatory cytokines released by epithelial cells are activated by amoebic cysteine proteinases. It has been proposed that activated cytokines would then recruit neutrophils and other inflammatory cells to the sites of infection, suggesting an important role of the host inflammatory response in tissue damage [17],[18],[19],[20].
In regions where amoebiasis is endemic mixed intestinal infections with E. histolytica/enteropathogenic bacteria are common [21],[22],[23],[24],[25],[26]. It is also well established that co-infection with the non-pathogenic E. dispar is prevalent in these regions [9],[10],[27]. How amoeba/bacteria interplay in these infections could modify disease manifestations by modulating pathogen virulence and the host response, has not been determined.
We approached this problem analyzing the interaction of E. histolytica and E. dispar with two pathogenic enterobacteria strains frequently associated with mixed infections, EPEC and Shigella dysenteriae, isolated from infected individuals. The results were compared with those obtained with amoebae that were not interacted with bacteria and with amoebae that phagocytosed a commensal E. coli strain. The response of epithelial cells to bacteria exposure and the effects of this exposure on cell damage by amoebae were then investigated. E. histolytica/enteropathogenic bacteria interactions induced higher virulence of amoebae. Enteropathogenic bacteria altered the epithelial barrier and induced release of chemoattractant molecules for both neutrophils and E. histolytica and better adhesion of amoebae to epithelial cells with subsequent increases of the cytopathic effect. In vivo, one could hypothesize that such conditions might confer higher susceptibility to pathogen invasion and severe disease manifestations. Our observations about survival and escape of infectious bacteria from amoebae, although preliminary, could be an interesting factor to consider when studying the intestinal environment of mixed infections.
E. histolytica HM1-IMSS trophozoites were cultured in TYI-S-33 medium as indicated [28] after their recovery from hamster liver passage and determination of virulence by production of liver abscesses. E. dispar SAW 760 RR, clone 2, trophozoites were cultured in axenic medium LY-S-2 as reported [29]. The bacteria utilized corresponded to clinical isolates of the commensal Escherichia coli 086:H18 and the pathogenic EPEC B171-0111: NM, kindly donated by Dr. Teresa Estrada (CINVESTAV, México) and Shigella dysenteriae kindly provided by Dr. Celia Alpuche (Pediatrics Hospital, National Institutes of Health, Mexico).
MDCK (NBL-2), dog kidney epithelial cells, passage 72, were grown to form confluent polarized monolayers as previously reported [30]. Cells were seeded on 24-well culture dishes for interaction with amoebae or bacteria, or on Millicel filters for transepithelial resistance measurements. Monolayers grown on glass cover slips were used for fluorescence microscopy observations.
To standardize the conditions to measure phagocytosis of bacteria by amoebae and bacteria viability inside amoebae, bacteria were transfected with vector pd2EGFP (Clontech Laboratories, Palo Alto, CA) to express green fluorescent protein, as reported [31]. EGFP-expressing bacteria were co-cultured with amoebae in amoeba/bacteria ratios of 1∶100 for different periods of time. Amoebae were then freed of non-phagocytosed bacteria by extensively rinsing the culture wells with PBS containing 5 mM sodium azide and 50 µM gentamycin. Attached amoebae were gently detached with PBS-gentamycin solution and residual extracellular bacteria removed by centrifugation-resuspension cycles in the same solution. Aliquots of the pelleted amoebae were further checked by fluorescence microscopy for absence of bacteria. Pellets were then fixed with 3.7 % formaldehyde, washed with PBS and resuspended in 300 µl of PBS. Intracellular EGFP-fluorescence was determined by flow cytometry. The highest fluorescence values corresponding to the highest number of bacteria ingested by amoebae were registered at 2 and 3 h after the interaction.We chose two and a half hours as the optimal time for phagocytosis. Amoebae that phagocytosed bacteria (not expressing EGFP) were utilized for all the following experiments. Cell cultures were freed of bacteria following the techniques applied to fluorescent bacteria and proven to be effective for this purpose. For the viability assays, bacteria expressing EGFP were interacted with amoebae in the above conditions. The capacity of bacteria, recovered from amoebae at different times after been phagocytosed, to form colonies on LB-agar plates was measured by CFU assays. At the indicated times, amoeba cultures were freed of non-phagocyted bacteria as described above. Amoebae were lysed with 0.12% Triton X-100 in LB medium and 100 µl of serial dilutions (up to 10−4) of the lysate added to LB agar plates and incubated at 37°C. Colonies formed in each plate after 24 h were quantified. One hundred per cent viability corresponds to the number of bacteria forming colonies 2.5 h after been phagocytosed by amoebae. Quantification of colonies formed by bacteria strains not expressing EGFP and recovered from amoebae in the above conditions revealed similar viability to that registered for EGFP-expressing bacteria.
E. histolytica trophozoites (2×105) were co-cultured with bacteria strains not expressing EGFP for 2.5 h in amoeba/bacteria ratios of 1∶100. After removal of non-phagocytosed bacteria, as indicated above, amoebae were deposited on confluent MDCK cell monolayers. After one hour of co-cultivation at 37°C, amoebae were removed by keeping the culture dishes in ice for 5 min and extensive rinsing with ice-cold PBS. The remaining intact epithelial cells were quantified by staining with the methylene blue method [32]. The number of intact cells in MDCK cell monolayers not exposed to amoebae was the control for 0 % cytopathic effect. Inhibition by 100 mM galactose or 250 µM E-64 was measured in amoebae incubated for 30 min before their addition to the epithelial monolayers.
E. histolytica trophozoites were labeled with 1.0 µl of calcein AM (Molecular Probes, Eugene OR) incubating at 37°C for 30 min. After washing with PBS, amoebae were checked for viability with trypan blue and co-cultured with each of the bacteria strains for 2.5 h. Non-phagocytosed bacteria were removed as described above and amoebae tested for adhesion to formaldehyde-fixed MDCK confluent monolayers for 20 min, as reported [13]. Adhered amoebae were detached by rinsing with ice-cold PBS, fixed with 3.7 % formaldehyde and their fluorescence measured at 517 nm by flow cytometry. Adhesion indexes of calcein-labeled amoebae to MDCK monolayers, previously exposed to bacteria for 4 h, as well as those of amoebae pretreated with 100 mM galactose or the Gal/GalNac lectin polyclonal antibody (10 µg/10,000 amoebae) were estimated in the same way.
After interaction of bacteria as indicated above, trophozoites were fixed with 2% paraformaldehyde for 20 min and rinsed 3X with PBS. A polyclonal antibody to the Gal/GalNac 170 kDa subunit (H5), kindly donated by Dr. Barbara Mann, was added at 1∶50 dilution in PBS/2% FBS to a suspension of 2.5×105 trophozoites and these incubated at 4°C for 40 min. A secondary antibody (Goat anti-Rabbit IgG tagged with FITC) was added to the cells at a dilution of 1∶400 and incubated for 45 min at 4°C. Cells were rinsed with PBS and resuspended in 300 µl of PBS/2% paraformaldehyde. FITC fluorescence was measured in a FACSCalibur flow cytometer at emission peak of 520 nm.
Ezymatic activity of cysteine proteinases (CP) in E. histolytica lysates and in culture medium was analyzed in control amoebae and amoebae co-cultured for 2.5 h with the bacteria strains. After removing non-phagocytosed bacteria as indicated above, trophozoites were cultured for 2 h in culture medium without serum and lysed by freeze-thaw cycles in 50 mM Tris-HCl, pH 7.2, 150 mM NaCl, 1.0 mM CaCl2. Two micrograms of each of the trophozoite lysates and 5 µl of their respective culture medium (freed of debris by centrifugation at 15,000 xg) were loaded in 1% gelatin, 10% polyacrylamide gels using reported conditions to analyze individual CP activities [32]. The clear areas in the gels revealed cysteine proteinase activity by digestion of the gelatin. Enzymatic activity areas were scanned with the SigmaGel Program in gel from three independent experiments. Lysates and culture media were also separated by electrophoresis in Laemmli's 10% SDS-polyacrylamide gels and silver-stained as controls for protein loading.
MDCK cells were plated on polycarbonate filters (1.2 cm diameter, Millipore Co, Bedford, MA) previously coated with a solution containing 30 mg/ml of rat Type I collagen. After cells reached confluence, approximately after 48 h in culture, bacteria were added in a ratio of 100∶1. TER was registered as described [30], before adding bacteria and at different times of co-culture, previous removal of bacteria and addition of fresh culture medium. FITC-labeled annexin V, Rhodamine-phalloidin and DAPI (Molecular Probes, Eugene, OR) were utilized to stain cells and monitor apoptosis, cell morphology and organization of the cytoskeleton in MDCK monolayers exposed to bacteria. For this, cells grown on cover slips were fixed with 3.7 % paraformaldehyde and stained by standard fluorescence microscopy methods recommended for these indicators. A monolayer irradiated with UV light was the positive control for apoptosis
Chemotaxis was assayed in Transwell chambers as previously reported [33], loading 150,000 calcein-labeled amoebae in the upper chambers resuspended in migration buffer (50 mM Tris-HCl pH 7.2, 150 mM NaCl, 1 mM Ca Cl2, 0.01% BSA). Amoebae that had not ingested bacteria as well as amoebae that phagocytosed bacteria for 2.5 h were tested. The lower chambers contained solutions containing 100 ng/ml in migration buffer of recombinant human cytokine IL-8 (Preprotech Inc., Rocky Hill, NJ), conditioned culture media obtained from MDCK monolayers co-cultured with each of the bacteria strains in the absence of serum or media from overnight bacteria cultures. Bacteria were removed by centrifugation before adding the culture media to the lower chambers. The number of calcein-labeled amoebae that migrated to the lower chambers was determined by flow cytometry. Culture media from monolayers not exposed to bacteria were used as control. Chemotactic index (CI) was calculated considering CI = % chemotactic migration/ % random migration. CI = 1.0 corresponds to migration to control media. Purified canine neutrophils were obtained from citrate-treated dog blood, separated in 20 % Ficoll gradients. The layer containing neutrophils was further purified by resuspension/centrifugation cycles in PBS and labeled with calcein. Eighty thousand cells were loaded in the upper chambers and chemotaxis assessed as indicated for amoebae.
Total RNA was obtained from control MDCK monolayers, monolayers exposed to E. coli strains, exposed to S. dysenteriae or from monolayers incubated with 10 µM of BAY11-7085 (inhibitor of NFκB activation, Calbiochem, La Jolla, CA), previous exposure to Shigella. RNA was extracted with TRIzol (Invitrogen, Rockville, MD) following the specifications of the manufacturer. cDNA was synthesized from 5 µg of DNAse I-treated RNA (DNA-free™, Ambion Inc.) in a reaction mixture containing 5 mM Mg Cl2, 50 mM KCl, 10 mM Tris-HCl, pH 8.3, 0.25 mM of each dNTP, 40U of RNAse inhibitor, 0.5 µM of oligo-dT-primers and 50U of Superscript II (Invitrogen). The reactions were allowed to proceed for 45 min at 42°C and inactivated for 5 min at 65°C. Amplification of IL-8 cDNA was done by mixing 1 µl of cDNA with 50 µl of PCR buffer supplemented with 2.5 mM MgCl2, 0.5 µM each of sense (5′ATGACTTCCAAGCTGGCTG3′) and antisense (5′TCTGAGTTTTCACAATGTGG3′) primers (designed accordingly to the IL-8 mRNA canine sequence, Canis familiaris) and 1U of Taq-polymerase (Invitrogen). PCR cycle conditions were 30 s at 94°C, 20 s at 45°C and 1 min at 72°C for 32 cycles. Sense (5′ATGGATGATGATATCGCCGC3′) and antisense (5′TTGGGGTTCAGGGGGGC3′) primers were utilized for amplification of the canine β-actin cDNA. The resulting RT-PCR products were analyzed in 1% agarose gels stained with a 2 µg/ml solution of ethidium bromide to monitor the presence of the expected size bands corresponding to IL-8 (194 pb) and β-actin (338 bp).
Data are presented as means±standard deviation. The significance of the results was calculated by t–Student test utilizing the program Sigma Stat. *p values≤0.05 were considered significant respect to controls. n values correspond to at least 3 independent experiments done in duplicate.
Unless otherwise specified, all reagents were obtained from Sigma Chemical (St. Louis, MO).
After a 2.5 h incubation of E. histolytica and E. dispar trophozoites with E. coli 086∶H18 (Ec), EPEC, or Shigella dysenteriae (Ed), the cytopathic effect of amoebae on MDCK cell monolayers was quantified and expressed as percentage of cell damage (Figure 1). Amoebae that were not incubated with bacteria (Eh or Ed) were controls for damage inflicted by amoebae that phagocytosed bacteria. Phagocytosis of E. coli 086∶H18 by E. histolytica (Eh/Ec) increased its cytopathic effect, but the increase was not significantly higher than that caused by control amoebae (49.6±0.95 versus 46.0±1.08). However, after phagocytosis of EPEC (Eh/EPEC) or Shigella (Eh/Sd), the cytopathic effect of E. histolytica increased to 64.6±2.63 and 77.6±1.24, respectively. In contrast to what was observed with E. histolytica, phagocytosis of any of the bacterial strains by E. dispar did not induce cytopathic behavior (Ed/Ec, Ed/EPEC, Ed/Sd). Figure 1 also shows that 100 mM Galactose, a known ligand of the amoebic surface Gal/GalNAc lectin, drastically reduced the cytopathic effect of control amoebae to less than 1.0 %. The same concentration of the sugar inhibited cell damage 5 %, 13 % and 29 % in amoebae that phagocytosed the commensal E. coli, EPEC or Shigella, respectively. Furthermore, incubation of amoebae with E-64, a specific inhibitor of cysteine proteinase activity at concentrations not deleterious to amoebic viability [7],[32], also caused significant inhibition of the cytopathic effect: 82 % in control amoebae, 76 %, 66 % and 55 % in E. coli, EPEC and Shigella, respect to their cytopathic effect in absence of E-64.
These results showed that phagocytosis of bacteria did not induce virulence in E. dispar, while in E. histolytica it induced an increase of the cytopathic effect particularly after phagocytosis of pathogenic bacteria. Galactose and E-64 inhibition suggest a role for the lectin and cysteine proteinase in the induction of enhanced virulence. As E. dispar did not show induction of virulence in any of the conditions tested, the following results refer only to E. histolytica.
Since amoebic adhesion to target cells depends on the activity of the Gal/GalNAc lectin, we analyzed its concentration in E. histolytica trophozoites before and after phagocytosis of bacteria. Amoebae were fixed and labeled with a polyclonal antibody directed to the 170 kDa heavy subunit of the lectin that contains the galactose-binding domain, and a secondary antibody tagged with FITC. Fluorescence intensity on the surface was measured in 10,000 cells for each condition. Figure 2A shows the mean index of fluorescence (MIF) in a representative experiment. Amoebae that were not exposed to bacteria (Eh) showed a MIF = 24.7±4.61. Amoebae that phagocytosed E. coli (Eh/Ec) showed a MIF = 64.42±6.0. Amoebae that phagocytosed EPEC (Eh/EPEC) showed a MIF = 73.65±5.60 and those that phagocytosed Shigella (Eh/Sd) showed a MIF = 94.42±7.34. The average of MIF values, obtained in three independent experiments, indicated 3.0-fold and 3.9-fold increases in amoebae after phagocytosis of EPEC or Shigella. Phagocytosis of the commensal E. coli only induced a 2.0-fold increase. Treating amoebae with only the secondary antibody did not increased MIF values.
Augmented Gal/GalNAc lectin concentration on the surface of amoebae could result in better adhesion and higher cytopathic effect, thereby adhesion of amoebae that phagocytosed bacteria to MDCK cells was quantified (Figure 2B). The adhesion index for control amoebae (not incubated with bacteria, Eh) was given value 1.00. After phagocytosis of the commensal E. coli, the adhesion index of trophozoites was 2.17±0.12. After phagocytosis of EPEC, the adhesion index was further increased to 2.52±0.21 and for amoebae that phagocytosed Shigella it reached values of 3.65±0.18. The specificity of the adhesion was corroborated in assays where amoebae (control as well as those that phagocytosed bacteria) were preincubated with the polyclonal antibody to the amoebic Gal/GalNAc lectin before interaction with the cells. The competition with the antibody reduced adhesion indexes in all the cases to 50% of the control value. An irrelevant antibody of the same isotype did not compete the adhesion of amoebae.
These results showed that phagocytosis of bacteria, but particularly pathogenic bacteria, induced a higher concentration of the Gal/GalNAc lectin on the surface of amoebae that seems correlated with a lectin-mediated increase of amoebic adhesion to MDCK cells. However, the antibody could only decrease binding in all the cases to the same level, suggesting that increased adhesion of amoebae after phagocytosis of bacteria is mainly, but not completely lectin-dependent.
The results in Figure 1 showing that the increase in cytopathic effect of amoebae that phagocytosed bacteria could be inhibited by E-64, led us to analyze cysteine proteinase (CP) activitiy in these amoebae. Figure 3 shows representative gelatin zymograms of lysates and culture media of E. histolytica trophozoites after phagocytosis of bacteria. The enzymatic activity of each proteinase, corresponding to the area of gelatin digested, was quantified by densitometry (Figure 3A and 3B). The bars in Figure 3C and 3D, express the fold-increase over value 1.0, given to each of the digested areas in lysates and culture media from control amoebae not exposed to bacteria (Eh). Proteinase activity bands corresponding to 48, 35, 29 and 27 kDa have been identified in lysates of trophozoites as the major proteinases CP1, CP2, and CP5 [6]. Figure 3C shows the values obtained from 3 separate gels where the enzymatic activities of CP1 and CP2 increased above two and three-fold in lysates of amoebae that phagocytosed EPEC or Shigella. The bands of 29 and 27 kDa corresponding to CP5 showed a lower but significant increase above control amoebae that however, was not significant between EPEC and Shigella for the 29 kDa band. After phagocytosis of the commensal E. coli the increase in proteinase activities was not significant. The highest fold-increase for all the activities was observed after phagocytosis of Shigella. The enzymatic activity band of 70 kDa, present in both zymograms, may represent an aggregate of some of the major proteinase activities, as it does not correspond to characerized proteinases.
Cysteine proteinases released to the medium of amoebae cultured in axenic conditions have been identified as CP1, CP3 and CP5 [6]. Zymograms of culture media without serum in which the amoebae were kept for 2 hours after phagocytosis of bacteria (Figure 3B), showed an increase close to 2-fold or higher for all the bands in Shigella. For EPEC the increase was lower, but still significant for bands of 48 kDa, 35 kDa and 27 kDa (Figure 3D). Parallel 10% SDS-polyacrylamide gels of lysates and culture media were silver stained to visualize all the protein bands in the gels (Figure S2). The gels show that the same protein concentration was loaded in all lanes and no particular difference was observed in particular bands. Therefore, the differences in enzymatic activity in the zymograms are real, indicating specific activation of some proteinases in the amoebae that phagocytosed bacteria. The observation that E-64 substantially inhibited the increase in cytopathic effect induced by phagocytosis of bacteria, shown in Figure 1, supports that these enzymes could have an important role in the induction of higher virulence.
From the results above, it is possible to think that in mixed amoeba/bacteria infections, the interplay of pathogens might modulate damage to the epithelial cells, up-regulating expression of specific pathogenic molecules in the amoebae. However, it is known that epithelial cells exposed to pathogens respond in different ways to their presence. To investigate this, we measured damage of MDCK cells by E. histolytica trophozoites when the monolayers had been previously exposed to the enteropathogenic bacteria used in this study. We analyzed the interaction with amoebae that were not exposed to bacteria as well as with amoebae that had phagocytosed bacteria. As shown in the first set of bars in Figure 4A, damage to control monolayers by amoebae that phagocytosed bacteria was increased (compare Eh/MDCK with bars in the same set) corroborating results shown in Figure 1. The second set of bars shows that amoebae that had not ingested bacteria, but were co-cultured with epithelial cells exposed to bacteria, increased their cytopathic effect (compare values in this set of bars with those in the left). The third set of bars shows that amoebae that had phagocytosed bacteria, when co-cultured with monolayers exposed to pathogenic bacteria, greatly increased cell damage. In this case, there was not only more damage, but it occurred faster, as the monolayers were completely destroyed in 45 min by amoebae that had phagocytosed EPEC or Shigella.
The increased damage to monolayers exposed to bacteria suggested that the presence of bacteria could be inducing changes in the epithelial cells that facilitated adhesion of amoebae. Figure 4B shows in the first set of bars, the adhesion of amoebae that had phagocytosed bacteria to control MDCK cells. As also shown in Figure 2, these amoebae showed higher adhesion than amoebae that had not phagocytosed bacteria. The second set of bars shows that amoebae not incubated with bacteria adhered better if monolayers had been exposed to bacteria, especially Shigella. The third set of bars shows that after exposure of monolayers to bacteria, adhesion of amoebae that had phagocytosed bacteria reached the highest adhesion index, particularly after interaction with Shigella.
These results showed that MDCK cell monolayers exposed to bacteria, but particularly to pathogenic bacteria, are better targets for adhesion and damage by amoebae than unexposed monolayers. These two processes can be further enhanced if these monolayers were incubated with amoebae that had phagocytosed bacteria, in particular the pathogenic strains that, as shown above, also increased amoebic virulence.
Neutrophils and macrophages are not the only cells attracted by pro-inflammatory cytokines to participate in an inflammatory response of epithelia. E. histolytica trophozoites are also attracted to intestinal epithelial tissue when inflammatory cells are present [18],[34]. Recent reports have shown that E. histolytica trophozoites migrate in response to human TNFα and IL-1β [35],[36], suggesting a possible role for cytokines in the migration of amoebae to sites where bacteria are present and have initiated an inflammatory response. As shown in Figure 5A, amoebae and neutrophils were induced to migrate by culture media from MDCK cells exposed to bacteria. Culture media from MDCK monolayers exposed to E. coli induced a slight but not significant increase of migration of trophozoites and neutrophils, while culture media from EPEC or Shigella-exposed cells induced 2 times and almost 3 times higher migration of amoebae. Neutrophils were particularly attracted to culture medium from cells exposed to Shigella. Lack of response of amoebae and neutrophils to overnight medium of Shigella ruled out that migration could had been induced by bacterial products in the culture media. It has been reported that MDCK cells release IL-8 when subjected to Salmonella typhimurium infection [37]. The presence of IL-8 in the culture media of MDCK cells exposed to Shigella was corroborated by ELISA assays utilizing a monoclonal antibody to canine IL-8 (clone 258901, RD Systems Inc, Minneapolis, MN, kindly donated by Dr. A. Castillo, CINVESTAV). The results showed concentrations of the cytokine in three different culture media in the range of 180–200 pg/ml.
At the same time, it was found that human IL-8 induced migration of both neutrophils and amoebae, providing support to the chemoattractant role of this chemokine when present in culture media. Furthermore, we found that culture media from cells exposed to Shigella in the presence of the inhibitor of IL-8 mRNA expression, BAY11-7085 [38], reduced migration of both neutrophils and amoebae by 50%. If the inhibitor was added to culture media of cells after their exposure to Shigella, it had not effect on the migration of the amoebae (data not shown).
It has been shown that epithelial cells in culture release pro-inflammatory cytokines, such as IL-8 as a defense mechanism when infected by bacteria [39],[40]. Shigella infection of intestinal cells activates NFκB through a polysaccharide-dependent innate intracellular response leading to the expression of this cytokine [39],[41]. Other enteropathogenic bacteria also activate release of cytokines [37],[39],[40],[41]. Changes in transcription patterns induced by pathogens could provide a clear indication of the response mechanisms of infected cells. Our data above showed that enterobacteria induced important changes in amoebae and epithelial cells. The changes induced by pathogenic bacteria and especially by Shigella were always more pronounced. Thus, it was clear that the presence of bacteria was affecting the interaction between amoebae and epithelial cells. We analyzed the expression of IL-8 mRNA in cells exposed to E. coli 086∶H18, EPEC and Shigella by RT-PCR assays using specific primers for MDCK cell IL-8 mRNA designed for this purpose. We thought that it was very interesting that IL-8 was expressed in cells exposed to bacteria as a defense response, but at the same time this chemokine was acting as chemoattractant for the amoebae. Figure 5B shows the results of a representative experiment (out of three) where after 4 h of exposure of MDCK cells to E. coli 086∶H18, EPEC or Shigella, the expression of IL-8mRNA was differentially induced in cells exposed to enteropathogenic bacteria. The highest expression corresponded to cells exposed to Shigella. The figure also shows that the expression of IL-8 mRNA was inhibited 93% in MDCK cells treated with 10 µM BAY11-7085 before exposure to Shigella. These results revealed that exposure of MDCK cells to bacteria, but particularly to invasive bacteria like Shigella, can induce a signaling process that activates NFκB pathways and expression of IL-8 mRNA.
We then analyzed if the induction of IL-8 and its release induced structural or functional damage to the bacteria-exposed cells and how this response might be modulated by the presence of another pathogen. Co-culture of pathogenic enterobacteria with epithelial cells is reported to induce alterations of epithelial organization [16],[42]. Reorganization of the actin cortical cytoskeleton is closely associated with altered permeability of epithelia and endothelia elicited by the infection [15],[37],[40],[43],[44].
Signaling pathways and mechanisms activated by pathogens to disrupt the structural organization of target cells or to induce a cell response are relatively well known with pathogenic bacteria. In contrast, these aspects are only beginning to be explored in the case of pathogenic amoebae [45]. Thus, we determined if exposure of MDCK cells to enterobacteria modified structural and functional features of the MDCK monolayers that could explain increased amoebic damage. Figure 6A, shows that exposure of monolayers to all the bacteria strains caused gradual decrease of transepithelial resistance (TER), leading to higher permeability. The initial steady state TER values of approximately 528±33 ohm.cm2 dropped to 380±20 ohm.cm2 in monolayers exposed for 5 h to the commensal E. coli and to196±33 ohm.cm2 after exposure to EPEC or Shigella. Monolayers not exposed to bacteria maintained the initial steady state TER. Changes in the organization of the actin cytoskeleton, known to regulate opening of the tight junctions [30],[46], were assessed by staining cells exposed to bacteria with Rhodamine-phalloidin to visualize polymerized actin. As shown in Figure 6B, actin in control monolayers was forming juxtaposed cortical rings and fine actin filaments on the basolateral side of the cells and microvilli on the apical side, all of them characteristic features of confluent polarized MDCK monolayers (Figure 6B, a). In contrast, exposure to EPEC or Shigella (Figure 6B, b, c) induced separation of the cell borders, loss of the cortical actin ring and a striking reorganization of actin into thicker filaments. This rearrangement of actin filaments could be explained by bacteria-initiated disruption of the tight junction components and signaling to activate release of pro-inflammatory cytokines. These changes, not necessarily conducive to cell death (see Figure 6C, a, b) initiate the response to the presence of the parasite and its control by the release of pro-inflammatory cytokines.
When trophozoites of Entamoeba histolytica invade the host intestinal mucosa, they can cause inflammatory colitis. However, trophozoites can remain in the colon without causing tissue damage. Phagocytosis of bacteria, regularly present in the colonic flora, has been considered a possible stimulus to induce amoebic invasive behavior. In regions were E. histolytica and E. dispar are endemic, it is common that intestinal infections caused by enteropathogenic bacteria occur simultaneously with the presence of amoebae [21],[22],[23],[24],[25],[26]. In these conditions, it is also common to find exacerbated manifestations of the infection. It is then feasible that the interplay between pathogens modulates amoebic virulence and the response of the intestinal epithelial cells.
To approach this important aspect of mixed infections, so far poorly examined, we utilized an in vitro system where we could test virulence of E. histolytica and E. dispar trophozoites after phagocytosis of enteropathogenic bacteria and, at the same time, assess if interaction of enteropathogenic bacteria with epithelial cells elicited responses that could modify amoebic damage. The E. coli, commensal 086∶H18 strain, the pathogenic EPEC, as well as an invasive isolate of S. dysenteriae were chosen for this study. Although the natural habitat of EPEC is the not the colon, this non-invasive pathogen was readily phagocytosed by amoebae and it is often present in mixed intestinal infections and has been utilized for in vitro bacterial infection of MDCK monolayers [14],[30],[37],[47]. Additionally, MDCK monolayers are a well characterized epithelial system, known to retain structural, functional and molecular features of polarized epithelia.
A surprising first result was that only E. histolytica trophozoites digested the phagocytosed bacteria, although Shigella retained 70% viability for more than 12 h while, all the bacteria phagocytosed by E. dispar were fully viable after 24 h (Figure S1, Text S1). Survival mechanisms of Salmonella and Shigella to escape digestion in highly phagocytic cells are well characterized [47],[48],[49]. Although a similar situation may prevent digestion of Shigella by E. histolytica, no studies have been done respect to the mechanisms utilized by pathogenic bacteria to survive in Entamoeba. To our knowledge, this is the first report to address this interesting phenomenon, at the moment beyond the scope of this study.
In spite of their different digestive capacity, trophozoites of E. histolytica that had phagocytosed EPEC, and particularly Shigella, caused higher damage to MDCK cells. In contrast, E. dispar, under the same conditions, did not cause any noticeable cytopathic effect. E. histolytica trophozoites have on their surface a protein complex with cell adhesive properties, the Gal/GalNAc lectin [13],[50]. So, increased adhesion to cells could reflect higher levels of the lectin on the trophozoite surface. Indeed, higher levels of the lectin were found in amoebae that showed increased adherence and caused more damage to target cells. Competition experiments showed that although increased adhesion was related to increased lectin expression on the surface, binding was also due to other molecules on the amoebae. Competition of the binding by galactose, the main ligand of the lectin, supported this conclusion. Expression of cysteine proteinases CP2 and CP5 in trophozoites is also a factor in cell damage [7],[51],[52], although with exception of CP5, other major CPs are also expressed in the non-pathogenic E. dispar [6]. Analysis of cysteine proteinase activities in lysates and culture media of amoebae that phagocytosed bacteria showed a selective increase of some of the major activities, both in amoebic lysates and culture media. Therefore, one possibility could be that the increased cell damage inflicted by amoebae that phagocytosed EPEC and Shigella is due to higher CP activities released during increased adhesion to target cells. Specific inhibition of cysteine proteinases in co-cultures of amoebae and epithelial cells blocked the increase of cytopathic activities shown by amoebae after ingestion of pathogenic bacteria. In contrast, no significant increases in CP activities were found in E. dispar which could not digest phagocytosed bacteria (data not shown). These data are very suggestive of an induction of the activity of these proteins after phagocytosis and digestion of enteropathogenic bacteria. To this moment, analyses of amoebic microarrays and phagosomes have not provided any clues for the presence of molecules that could explain the increase in amoebic virulence and the survival of Shigella in the E. histolytica [47],[53].
It has been shown that pro-inflammatory cytokines are released by intestinal cells exposed to E. histolytica infection [18] and that recombinant amoebic proteinases are capable of cleaving pro-inflammatory cytokine precursors to their active form in vitro [34]. It is possible then that increased adhesion of amoebae to epithelial cells together with higher release of CP into the medium, induced by phagocytosis of bacteria and particularly by pathogenic bacteria, would lead to higher concentrations of active cytokines at the sites of contact between amoebae and epithelial cells. The presence of activated inflammatory cytokines would then attract neutrophils and other cells to the sites where amoebae concentrate.
Our in vitro model allowed testing the above mentioned possibilities. It has been shown that several pathogens increase permeability of epithelial cells. E. coli ETEC and EPEC strains diminish the barrier functions of cultured epithelial monolayers [16],[17]. Salmonella and Shigella disrupt the intercellular junctions by alteration of the normal distribution of molecules that associate with them and disrupt the organization of the cytoskeleton [37],[43]. Virus entry into cells also results in this type of cellular disruption [44]. Alteration of epithelial barriers allows penetration of pathogens into the paracellular space and their dissemination into lower cell layers, as well as migration of inflammatory cells to the luminal side [14],[16],[17],[47],[54]. E. histolytica trophozoites can also produce decrease of TER in cultured epithelial cells [45].
We tested the effect of enterobacteria on monolayer permeability. TER registers indicated a small decrease in monolayers exposed to the commensal E. coli, and a gradual, but more accentuated drop in monolayers exposed to EPEC or Shigella. After 5 h, TER values had decreased to levels indicative of complete opening of the intercellular junctions. Unsealed monolayers allow passage not only of ions and big molecules, but even of neutrophils and other cells of the immune system. Opening of the tight junctions also allows exposure of receptors for these cells and for pathogens [47]. For example, H. pylori and L. monocytogenes have proteins that bind to E-cadherin once the tight junctions of the epithelial cells are opened, so they can enter cells or epithelial layers [14],[55]. We found that trophozoites increased their adhesion to MDCK cells that had been exposed to bacteria, ie: with their membrane junctions unsealed. Adhesion was even higher if trophozoites had phagocytosed bacteria. Higher adhesion could be related to the higher cell damage by trophozoites observed in epithelial cells exposed to bacteria, in particular to EPEC or Shigella.
These results corroborate that an increase of adhesion to cells by amoebae results in more cell damage. The increase was induced in trophozoites after phagocytosis of bacteria or after exposure of epithelial cells to the same bacteria. Moreover, the highest adherence and cell damage were observed when both, trophozoites and epithelial cells were incubated with bacteria. As shown above, phagocytosis of bacteria induced higher levels of the Gal/GalNAc lectin on the trophozoite's surface, which would facilitate adhesion of the amoebae. However, this would require higher number of receptors on the target cells for a better interaction. Our data suggest that other molecules on the amoebic surface could be participating in the interaction. Preliminary results, currently investigated in our laboratory, have shown that amoebae can induce exposure of TLRs on the surface of intestinal epithelial cells.
The gradual drop of TER in monolayers exposed to EPEC or Shigella, suggests a gradual effect on the disorganization of the intercellular junctions that was not apoptotic, but capable of inducing a marked reorganization of the actin cytoskeleton. Disruption of the cortical actin circumferential ring, loss of microvilli and rearrangement of the basolateral filaments, have been correlated with the opening of the tight junctions [30],[46], normally a reversible process that allows epithelial cells to adapt to conditions in the medium. However, disorganization of intercellular junctions by enteropathogenic bacteria and other pathogens leads to release of pro-inflammatory cytokines [17],[18],[19],[40]. Our data have shown that MDCK epithelial cells co-cultured with enteropathogenic bacteria suffered functional alterations and released, into the medium, molecules capable of activating chemotaxis of amoebae and neutrophils. Both types of cells were specifically attracted to pro-inflammatory cytokine IL-8. IL-8 is released by epithelial cells during the interaction with pathogens and acts as a chemokine, playing an important role in the migration of neutrophils to sites of infection [37],[39],[40]. We found that MDCK cells exposed to bacteria induced expression of IL-8 mRNA and release of this chemokine to the culture media, corroborating previous results by other authors [39],[40],[41],[44]. Induction was low after exposure to the commensal E. coli, but increased markedly after exposure to Shigella. The induction was almost completely inhibited by inactivation of NFκB. Although this transcription factor activates transcription of different genes [39], the fact that IL-8 mRNA expression could be differentially induced by the exposure of MDCK cells to different bacteria, and was blocked by the inhibitor of NFκB activation, strongly suggest that bacteria, and particularly Shigella, can activate signaling pathways leading to expression of IL-8 mRNA [15],[39],[40],[47]. We showed here, that canine neutrophils and amoeba migrated in a similar way to human IL-8 and to culture media of MDCK cells previously exposed to bacteria. The highest migration was registered for the culture media from MDCK cells exposed to Shigella where the presence of IL-8 was corroborated by ELISA. Moreover, these cells showed the highest induction of IL-8 mRNA. Previous experiments from our group have shown that amoebae also respond to IL-1β [36] and recently, it has been reported that E. histolytica trophozoites respond to human TNFα gradients by chemotactic sliding [35]. The chemotactic effect of inflammatory cytokines on amoebae supports the idea that amoebae can reach sites in the epithelia where an inflammatory response has been started by bacteria. At the same time, amoebae present in the same milieu can increase their virulence by phagocytosis of bacteria and cells that have been altered by the presence of bacteria are more susceptible to adherence and damage by amoebae. It is possible that all of these phenomena contribute to the pathogen's ability to penetrate epithelial layers. What could be the role of E. dispar in this situation? The avirulence of E. dispar suggests a non-aggressive participation of this amoeba in mixed infections. However, its ubiquitous presence in samples from patients, its inability to digest ingested bacteria (Text S1) and isolated reports of lesions produced by some amoebic isolates, make study of this amoeba species also worth pursuing.
A limitation of our in vitro cell system is the fact that we are observing phenomena outside of the intestine. However, with this approach to mixed amoeba/bacteria infections we have obtained results that could not have been monitored in vivo. We have now a better insight into the role played by the participating elements in the organism. A molecular approach to understand better the signaling processes and the molecules involved at different stages of the infection is now feasible. We hope that our findings encourage research on a health problem still prevalent and neglected in developing countries.
|
10.1371/journal.pcbi.1005890 | A systematic atlas of chaperome deregulation topologies across the human cancer landscape | Proteome balance is safeguarded by the proteostasis network (PN), an intricately regulated network of conserved processes that evolved to maintain native function of the diverse ensemble of protein species, ensuring cellular and organismal health. Proteostasis imbalances and collapse are implicated in a spectrum of human diseases, from neurodegeneration to cancer. The characteristics of PN disease alterations however have not been assessed in a systematic way. Since the chaperome is among the central components of the PN, we focused on the chaperome in our study by utilizing a curated functional ontology of the human chaperome that we connect in a high-confidence physical protein-protein interaction network. Challenged by the lack of a systems-level understanding of proteostasis alterations in the heterogeneous spectrum of human cancers, we assessed gene expression across more than 10,000 patient biopsies covering 22 solid cancers. We derived a novel customized Meta-PCA dimension reduction approach yielding M-scores as quantitative indicators of disease expression changes to condense the complexity of cancer transcriptomics datasets into quantitative functional network topographies. We confirm upregulation of the HSP90 family and also highlight HSP60s, Prefoldins, HSP100s, ER- and mitochondria-specific chaperones as pan-cancer enriched. Our analysis also reveals a surprisingly consistent strong downregulation of small heat shock proteins (sHSPs) and we stratify two cancer groups based on the preferential upregulation of ATP-dependent chaperones. Strikingly, our analyses highlight similarities between stem cell and cancer proteostasis, and diametrically opposed chaperome deregulation between cancers and neurodegenerative diseases. We developed a web-based Proteostasis Profiler tool (Pro2) enabling intuitive analysis and visual exploration of proteostasis disease alterations using gene expression data. Our study showcases a comprehensive profiling of chaperome shifts in human cancers and sets the stage for a systematic global analysis of PN alterations across the human diseasome towards novel hypotheses for therapeutic network re-adjustment in proteostasis disorders.
| Protein homeostasis, or proteostasis, is maintained by the proteostasis network (PN), an intricately regulated modular network of interacting processes that evolved to balance the native proteome, supporting cellular and organismal health throughout lifespan. Imbalances and collapse of cellular proteostasis capacity, the capacity to buffer against cytotoxic damage and stress, is increasingly implicated in some of the most challenging diseases of our time, including neurodegeneration and cancers. The systems-level PN alterations in these diseases are not understood to date. Here, we address this challenge, focussing on the human chaperome, the ensemble of chaperones and co-chaperones, which represents a central conserved PN functional arm. We devised a novel data dimensionality reduction approach enabling quantitative contextual visualization of chaperome alterations in the heterogeneous spectrum of cancers based on gene expression data from thousands of patient biopsies. We developed Proteostasis Profiler (Pro2), a new web-tool enabling intuitive visualisation of cancer chaperome deregulation maps. We stratify two cancer groups based on diverging chaperome deregulation and highlight similarities between cancer and stem cell proteostasis. Our study also exposes drastically opposed shifts between cancers and neurodegenerative diseases. Collectively, this study sets the stage for a systematic global analysis of PN alterations across the human diseasome.
| Eukaryotic proteomes comprise a complex repertoire of diverse protein species that are organized in a modular interactome network in order to execute native function in support of proteostasis and a healthy cellular phenotype. Proteome balance is safeguarded by the proteostasis network (PN), an intricately regulated network of conserved processes that have evolved to safeguard the healthy folded proteome [1]. Cellular proteostasis capacity is limited within the constraints of each cell’s proteostasis boundary [2]. Proteostasis imbalances, deficiency and functional collapse are implicated in a broad and increasing spectrum of protein conformational diseases with loss of native function or gain of toxic function, ranging from metabolic and neurodegenerative diseases to cancer [3, 4]. Increasing awareness of the fundamental role of the PN in cellular health, its relevance in diseases and potential as a therapeutic target of proteostasis regulator (PR) drugs call for a systematic and systems-level assessment of PN deregulation throughout the human diseasome, towards improved understanding of diseases of proteostasis deficiency and rationalized network-informed approaches to therapeutic proteostasis re-adjustment.
Important progress has been made in our understanding of proteostasis biology, building on fundamental insights on conserved proteostasis processes and their role in disease, such as chaperone-assisted protein folding and quality control [5–8], clearance through autophagy [9–13] and the ubiquitin-proteasome system (UPS) [14–16], followed by the appreciation of their concerted action within a conserved tightly regulated PN [1, 17]. The identification, development and first clinical evaluations of small molecule PR drugs for therapeutic re-adjustment of proteostasis diseases such as cystic fibrosis represents a novel and powerful therapeutic paradigm [1, 2, 18–23]. First investigations have started to explore systems-level quantitative and functional approaches to assess the implications of PN functional arms such as the chaperome in human tissue aging and disease [4, 24–26]. A precise understanding of the molecular mechanisms by which PN alterations contribute to disease could open novel therapeutic intervention strategies in a wide spectrum of proteostasis-related diseases. Still, to date, there has been no systematic study addressing the characteristics and extent of PN alterations in human diseases at a systems-level.
The folding functional arm, the human chaperome, is highly conserved and of central importance in the PN, responsible for maintaining the native folded proteome. In cancers, mutations and genomic instability inevitably entail alterations of proteome composition and balance that are far less well explored than the consequences of nucleic acid sequence alterations. Post-translational alterations at the proteomic level are beyond the reach of DNA repair mechanisms and cancer cells are constantly challenged by the need to accommodate large amounts of proteotoxic stress in consequence of increased translational flux and proliferation as well as proteotoxic stressors. Proteome instability and pathological alterations in the abundance of key signalling or housekeeping molecules such as kinases, metabolic enzymes or molecular transporters have to be buffered by the PN to ensure cellular survival. The cancerous state poses characteristic requirements on the PN, such as high chaperone levels and elevated proteasome activity in order to ensure for sufficient correction or elimination of aberrant protein species in light of increased translational flux and metabolic stress [27]. This chronic challenge ultimately drives cancer cells into a dependency on quality control and stress response mechanisms, a phenomenon previously described as non-oncogene addiction [28, 29]. Several individual chaperones and heat shock proteins such as HSP90 have consistently been found upregulated in cancers [27]. However, the profile and extent of chaperome differential expression has not been assessed systematically across the human cancer landscape.
Challenged by the genetic complexity and heterogeneity, collective prevalence and unmet medical need of the wide spectrum of human cancers as well as the lack of a systems-level understanding of proteostasis alterations during carcinogenic transformation, we developed a novel integrated analytical pipeline and software toolkit for the quantitative profiling of chaperome changes across the human cancer landscape (Fig 1). We utilized an expert-curated functional chaperome ontology comprising the ensemble of 332 human chaperone and co-chaperone genes [4] (Fig 1A). In order to apply our analytical workflow on a recent and comprehensive cancer gene expression dataset with clinical relevance, we turned to The Cancer Genome Atlas (TCGA) compendium [30]. We started with a customized genomic analysis pipeline in order to map chaperome functional family expression changes across TCGA solid cancers compared to matching normal tissue (Fig 1B). The resulting top-level view on cancer chaperome deregulation revealed a broad chaperome upregulation throughout the majority of cancers. This consistent and high overall chaperome upregulation prompted us to zoom in on functional sub-families. This analysis surfaced clusters of chaperome functional family up- and downregulation signatures that enabled further stratification of cancers. In summary, our analysis of the 10 major chaperome functional families reveals pronounced tissue differences of cancer chaperome deregulation. The preferential upregulation of ATP-dependent chaperone families such as HSP90s and HSP60s, while ATP-independent chaperones, co-chaperones, and small heat shock proteins (sHSPs) are consistently downregulated, is opposed to chaperome alteration patterns observed in brain tissues during aging and in neurodegenerative diseases [4]. These characteristic chaperome-wide differences further justify our approach and need for systematic maps of PN deregulation across the human diseasome.
In order to enable comprehensive, contextual, and quantitative representations of the complexity of chaperome alterations across a large number of patient biopsy disease datasets, we developed a new custom data dimensionality reduction and visualisation approach. Combining Meta-PCA, a novel principal component analysis (PCA) based two-step dimension reduction algorithm and its resulting quantitative M-scores of chaperome functional family disease alteration of gene expression with contextual polar plot visualisations, we provide intuitive quantitative maps of cancer chaperome gene expression changes (Fig 1B).
The mechanistic understanding of genotype-phenotype relationships in complex genetically heterogeneous diseases such as cancers requires the consideration of the cellular interactome network [31]. To reduce complexity and to highlight contextual changes of chaperome functional families, we first generated a custom curated high-confidence physical protein-protein interaction (PPI) chaperome network. We then collapsed proteins (nodes) and physical PPIs (edges) into a meta-network, where meta-nodes represent respective functional family members and meta-edges bundle the interactions between families (Fig 1C). This curated high-quality chaperome meta-network base-grid enabled the contextual projection of cancer-specific chaperome functional family differential gene expression (M-scores) onto the underlying interactome. We integrated these dimensions into interactome-guided, three-dimensional topographic maps visualising chaperome functional family cancer differential gene expression changes in the context of interactome network proximity, intuitively providing quantitative views of cancer chaperome deregulation (Fig 1D).
To make these resources easily available to the community, we developed Proteostasis Profiler (Pro2), an integrated web-based suite of applications enabling intuitive quantitative analyses and comparative visualisation of differential expression of complex PN alterations across large disease dataset compendia such as the TCGA. Visualisation and analysis features include heat map clustering and polar plot display. Integrated meta-networks and interactome-guided 3D topographic maps ease comparative exploration of cancer chaperome deregulation in the context of interactome network wiring. Pro2 is designed to serve the scientific community as a user-friendly application for systems-level exploration of PN disease alterations, at reduced complexity.
Overall, this study represents a systematically derived systems-level atlas of chaperome deregulation maps in cancers and neurodegenerative diseases, with a detailed focus on chaperome functional family alterations. The integrated genomic analysis workflow, built into the Pro2 suite of visualisation tools, provides a resource and analytical platform for future characterisation and exploration of PN deregulation patterns across the human diseasome, and as a readout interface for network shifts induced by therapeutic regulation.
Homeostasis of the cellular proteome, or proteostasis, is fine-tuned by the proteostasis network (PN), an intricately regulated network of conserved processes that have evolved to safeguard the native functional proteome and cellular health. The human chaperome, an ensemble of 332 chaperones and co-chaperones, represents a central functional arm within the PN in charge of maintaining the cellular folding landscape (S1A Table) [4]. Motivated by the genetic heterogeneity of cancers, their prevalence and associated medical need as well as the lack of a systems-level understanding of the role of proteostasis genomic alterations during carcinogenesis, we systematically assessed chaperome gene expression changes across the diverse spectrum of human cancers. We focused on an established resource of human cancer patient biopsy RNA-seq datasets provided through The Cancer Genome Atlas (TCGA) [30, 32]. We considered 22 human solid cancers with available corresponding healthy counterpart tissue biopsy data. To obtain global views on chaperome commonalities or differences between cancers, we applied Gene Set Analysis (GSA) in order to quantify gene expression changes of the chaperome and its functional families. GSA is an advanced derivative of Gene Set Enrichment Analysis (GSEA) that methodologically differs primarily through its use of the maxmean statistic, the mean of the positive or negative gene scores in each gene set, whichever is larger in absolute value, that has proven superior to the modified Kolmogorov-Smirnov statistic used in GSEA [33]. Secondly, GSA uses a different null distribution for false discovery rate (FDR) estimations, through a restandardization of genes in addition to sample permutation in GSEA. This step is crucial, as it allows assessing statistical robustness of the expert-curated chaperome functional ontology gene family groups. We obtained the GSA derived probability (p values) for each functional gene group to be significantly up- or downregulated in cancer as ∆GSA values in the interval [-1, +1] according to ((1—upregulation p value)—(1—downregulation p value)).
Notably, the human chaperome is predominantly upregulated across the majority of TCGA solid cancers with a ∆GSA group mean change of +0.50 as compared to 100 random sets of non-chaperome genes (Fig 2A). This overall chaperome upregulation highlights cellular non-oncogene addiction to chaperone-assisted folding and protein quality control mechanisms in consequence of increased client load, further challenging cellular proteostasis and driving “proteostasis addiction” in cancers [34]. Despite the diverse established knowledge about the role of chaperone upregulation in cancer, the deregulation of the human chaperome has not been assessed at a systems-level throughout the human cancer landscape. To functionally resolve the general chaperome upregulation across cancers, we zoomed in on functional family gene expression alterations. GSA followed by Euclidean clustering of chaperome functional families revealed characteristic cancer differences. We found the key ATP-dependent HSP90 and HSP60 families, of which selected members have previously been shown to be upregulated in cancers, amongst the most highly upregulated functional families with ∆GSA group mean changes of +0.55 and +0.51, respectively, alongside ER-specific chaperone factors (+0.53), followed by Prefoldins (PFDs, +0.40), HSP100 AAA+ ATPases (+0.32), and mitochondria-specific chaperones (MITOs, +0.09) (Fig 2B). These six functional families of predominantly ATP-dependent chaperones represent an upregulation cluster with an overall group mean change of +0.40. Intriguingly, the HSP70-HSP40 system and the large family of TPR-domain containing co-chaperones are overall repressed, with less consistent and largely cancer-specific alterations. HSP40 co-chaperones (-0.09) cluster closest with HSP70s (-0.10), indicative of the functional relationship they engage in during the HSP70 chaperone cycle. While HSP40 co-chaperones are overall weakly downregulated (-0.09), also the second group of co-chaperones, the TPR-domain containing proteins, clustered with the HSP70-HSP40 system and were overall downregulated (-0.18). Strikingly, sHSPs (-0.74) were overall very consistently and strongly downregulated. Overall, sHSPs, TPRs, and the HSP70-HSP40 system clustered in a downregulation cluster with an overall group mean change of -0.28 across cancers.
Besides marked differences in the pattern of cancer functional family changes, Euclidean clustering of cancer groups (rows) revealed two major clusters (Fig 2B)(with a p-value equal to 0.016 via multiscale bootstrap resampling). The vast majority of cancers is characterised by the consistent upregulation of HSP90s, ER-specific chaperones, HSP60s, PFDs, HSP100s and MITOs, opposed by a very consistent downregulation of sHSPs and a more cancer-specific overall downregulation of the HSP70-HSP40 system an TPR-domain co-chaperones. This group comprises Cluster I, representing ~91% of cancers, while Cluster II comprises ~9% of cancers with largely opposed chaperome deregulation signatures, in this set namely skin cutaneous melanoma (SKCM), and pheochromocytoma and paraganglioma (PCPG).
In summary, systematically assessing gene expression data derived from a total of 10,456 patient samples uncovers broad differences in chaperome-scale deregulation across the variety of human solid cancers. While the vast majority of cancers shows consistent and strong upregulation of chaperome genes, this analysis reveals marked clusters of chaperome functional family expression signatures that further stratify cancers by differential chaperome expression.
The major ATP-dependent chaperone functional families are consistently upregulated across a majority of cancers, while co-chaperones and sHSPs are consistently repressed (Fig 2B). In order to quantify this trend, we assessed the 22 differentially regulated cancer chaperomes for functional characteristics.
First, we compared expression of 88 chaperones against 244 co-chaperones represented in the human chaperome [4]. Projecting TCGA cancer groups by their chaperone and co-chaperone differential expression highlights a significant preponderance of cancer chaperome upregulation, including both chaperones and co-chaperones, while only a minor fraction of each is downregulated (S2A and S2B Fig). Overall, chaperones tend to be more upregulated than co-chaperones (S2B Fig). Consistently, within the group of chaperones, we find an overall preponderance of upregulation of both ATP-dependent and ATP-independent chaperones, while only small fractions each are downregulated (S2C and S2D Fig). The 50 ATP-dependent chaperones are more upregulated than the 38 ATP-independent chaperones, while ATP-independent chaperones are more downregulated than ATP-dependent chaperones (S2D Fig).
This analysis exposes a sub-group of cancers as notable exceptions to these trends, suggesting fundamental differences in chaperome deregulation. Projection of TCGA cancer groups by chaperone and co-chaperone up- and downregulation lends support for two groups of cancers, Group 1 and Group 2 (Fig 3A). These groups are recapitulated when projecting cancers by up- and downregulation of ATP-dependent versus ATP-independent chaperones (Fig 3C). K-means clustering confirms the significant separation of the Group 2 cancers pheochromocytoma and paraganglioma (PCPG), thyroid carcinoma (THCA), and the three kidney cancers kidney chromophobe (KICH), kidney renal papillary cell carcinoma (KIRP), and kidney renal clear cell carcinoma (KIRC) from Group 1 cancers, with a median silhouette width of s = 0.63 (Fig 3A) and s = 0.68 (Fig 3C). Group 1 cancers (red) represent the majority of cancers, characterized by strong overall chaperome upregulation, with low chaperone and co-chaperone repression, and a trend for upregulation of ATP-dependent chaperones (Fig 3B and 3D). Five Group 2 cancers however partition more distantly, with a lack of chaperome upregulation (Fig 3A and 3C). These cancers include three different kidney cancers, KICH, KIRP, and KIRC, which consistently lack chaperome upregulation (Fig 3A and 3B). Also, ATP-dependent chaperones are not preferentially upregulated in kidney cancers. Rather, an inverse trend is observed with increased downregulation of ATP-dependent chaperones (Fig 3C and 3D). Notably also, pheochromocytoma and paraganglioma (PCPG), rare related tumors of orthosympathetic origin, similarly show an even more prominent inverse alteration, with a preponderance of overall chaperome downregulation and preferential downregulation of ATP-dependent chaperones. Pheochromocytomas originate in the adrenal medulla, with close spatial association to the kidney, whose cancers are also Group 2 cancers. THCA is similar to PCPG, with a preferential downregulation of the chaperome and both PCPG and THCA represent tumors forming from cells of neuroendocrine origin.
Collectively, the data point to a preferential upregulation of ATP-dependent chaperones in the majority of cancers, which we refer to as Group 1 cancers, with general differences in chaperome deregulation in Group 2 cancers, comprising kidney cancers and cancers of neuroendocrine origin, such as PCPG and THCA.
Human embryonic stem cells (hESCs) are characterized by their capacity to replicate infinitely in culture, while maintaining a pluripotent state [35]. This immortal, undifferentiated phenotype resembles hallmark features of cancer cells such as an elevated global translational rate [36] and is expected to demand increased PN capacity capable of buffering imbalances to maintain proteostasis. Given the “stemness” phenotype of cancer cells and their resemblances with pluripotent stem cells we hypothesized that the consistent chaperome upregulation in cancers acts to mimic an enhanced stem cell PN setup. Increased proteasome activity [37] and elevated overall levels of the TRiC/CCT complex [38], representatives of the clearance and folding functional arms of the PN, respectively, have recently been associated with the intrinsic PN of pluripotent stem cells that acts to support their identity and immortality. It can be hypothesized that increased levels of central PN processes in stem cells exemplify characteristics of an enhanced PN setup. We thus assessed to which extent this stem cell PN setup is recapitulated in cancers.
First, we assessed differential changes of the proteasome across TCGA cancers and observed an overall consistent upregulation of the 43 proteasomal genes (HGNC Family ID 690) in > 70% of cancers (Fig 4A, S1B Table) [39], matching the role of increased proteasomal activity for proteome maintenance in stem cells [37]. Notably, Group 1 and Group 2 cancers, which are specifically defined based on chaperome differential expression signatures (Fig 3), do not co-partition with cancer clusters obtained by proteasome differential expression (Fig 4A). Next, we assessed cancer differential expression of the eukaryotic chaperonin TRiC/CCT, a hetero-oligomeric complex of two stacked rings with each eight paralogous subunits representing the cytoplasmic ATP-driven HSP60 chaperones in charge of folding approximately 10% of the proteome [40]. TRiC/CCT is highly conserved and essential for cell viability [40]. Loss of complex subunits induces cell death and a decline of pluripotency of hESCs and induced pluripotent stem cells (iPSCs) [38]. Within the PN, TRiC/CCT mediated folding and autophagic clearance act in concert to prevent aggregation [41]. TRiC/CCT levels decline during stem cell differentiation, and CCT8 acts as complex assembly factor [38]. Intrigued by the finding that CCT8 is the most highly elevated subunit in stem cells and likely acting as assembly factor [38], we assessed differential expression of individual TRiC/CCT subunits across cancers. Hierarchical clustering of subunit expression across solid cancers highlighted CCT8 as highly consistently upregulated across all cancers (mean change = 0.76, t test), and as overall second most highly upregulated subunit besides CCT6A (mean change = 0.786, t test) (Fig 4B). Consistent with the overall preferential downregulation of ATP-dependent chaperones observed in Group 2 cancers (Fig 3), we found these cancers to cluster together with overall lowest TRiC/CCT expression, where PCPG stands out with a consistent downregulation of all subunits (Fig 4B).
Together these findings suggest that proteostasis shifts in cancer cells add to an altered, enhanced PN state that mimics the immortal and resilient stem cell phenotype, buffering genome instability and ensuing proteomic imbalances in support of sustained and increased cellular proliferation throughout cancerogenesis.
Overall chaperome upregulation in cancers, with preferential enrichment for upregulation of ATP-dependent chaperones, alongside consistent downregulation of sHSPs, is diametrically opposed to chaperome deregulation trends previously observed in a study of chaperome alterations in human aging brains and in patient brains with age-onset neurodegenerative diseases [4]. While sHSPs were the only chaperome family found significantly induced in brain aging and the age-onset neurodegenerative diseases Alzheimer’s (AD), Huntington’s (HD), and Parkinson’s (PD) disease, this family is consistently downregulated across cancers (Fig 2B). This opposed chaperome deregulation points towards characteristic and fundamental differences in PN deregulation between disease families.
To investigate this disease group difference further, we applied the analysis outlined for cancers above also on the gene expression datasets that had earlier revealed global chaperome repression in AD, HD, and PD [4]. Our analysis reproduced the human chaperome as overall downregulated across AD, HD, and PD as compared to random permutations of non-chaperome genes (-0.20 ∆GSA group mean change, Fig 5A). Delving deeper into chaperome functional subfamilies, we reproduce earlier findings reporting broad repression of the major chaperome functional families except for sHSPs, the only family found strongly upregulated (+0.76) accompanied by slight upregulation of ER-specific chaperones (+0.21) and TPRs (+0.13) (Fig 5B). Thereby, our analytical workflow reproduces previously observed trends obtained in independent analyses, with different methods. With strong sHSP repression and upregulation of the HSP90, ER, HSP60, PFD, HSP100 and MITO chaperone families, cancers and neurodegenerative diseases display markedly diametrically opposed chaperome deregulation, not only at the overall chaperome-level (Fig 5C), but also with respect to alteration trends of chaperome functional families, where 70% of functional groups are altered in opposite directions (Fig 5D).
These opposing chaperome deregulation signatures are in line with differing implications of proteostasis alterations in these diseases. While broad chaperome repression and proteostasis functional collapse is associated with aggregation and cytotoxicity of chronically expressed misfolding-prone proteins in neurodegenerative diseases [4], enhanced proteostasis buffering capacity is associated with “stemness”, immortality and proliferative potential of both stem and cancer cells [27]. Indeed, epidemiological evidence suggests an inverse correlation between cancers and neurodegenerative diseases [45–48], supportive of a potential mechanistic link between opposed chaperome deregulation and the molecular underpinnings of the two disease groups. These global differences in chaperome deregulation call for a systematic and quantitative assessment of PN deregulation dynamics in human diseases.
In light of the diverse signatures of differential chaperome deregulation observed across cancers (Fig 2), and motivated by the increasing amount of genomics datasets available for cancers and other human diseases, we aimed at reducing data complexity by extracting quantitative indicators of chaperome differential cancer gene expression alterations, in order to gain insights through reduced complexity while retaining maximum information content.
We devised Meta-PCA, a novel principal component analysis (PCA) based semi-supervised two-step dimension reduction approach that facilitates stratification of cancer patient and normal control samples within heterogeneous gene expression datasets (Fig 1B). Based on previous work on dimensionality reduction of heterogeneous gene expression datasets [49, 50], we hypothesized that the underlying information contained in each chaperome functional group has a low dimensionality (meaning the functionality of each chaperome group can be quantified using only few, possibly one, variable) and could be surfaced using PCA, if there were sufficient samples available to represent the complete heterogeneity of chaperome alterations in cancers. Compared to conventional PCA, our method can deal with the effect of different group sizes, which as a confounder would negatively affect PCA results. Compared to simply calculating mean expression values of each functional group’s genes, Meta-PCA considers a wider range of gene expression information inherent to each functional group, resulting in scores with higher resolution. This could also be achieved by fully supervised methods such as Linear Discriminant Analysis. However, this case requires use of a method, which is blind to sample annotations so that these can later be used for validation, such as the unsupervised classification of cancer tissue type. Meta-PCA first uses tissue-wise PCA analyses to separate cancerous from control samples for individual tissues and maps functional family group gene expression changes to the global, or “meta”, mean expression change across cancers in order to obtain M-scores as quantitative indices of relative disease gene expression change (Eq 1). Inherent to the Meta-PCA method, patient-specific genomic variability is averaged out through the use of Meta-PCs derived from the TCGA collection of cancer biopsy samples, yielding mean reference boundaries. Assessing quantitative M-scores obtained through Meta-PCA against differential gene expression (∆GSA) obtained via GSA (Fig 2), we observed an overall significant Pearson correlation of 0.61 (S1 Fig). In addition to general comparability between results obtained through both methods, Meta-PCA analysis reduces complexity while retaining genomic information. Therefore, we focus on Meta-PCA for quantitative representation of proteostasis alterations in human diseases. We plot differential chaperome changes as M-scores for all cancer samples and chaperome functional families simultaneously using polar plots, such that axes represent functional families or sub-groups (Fig 2B). We obtain the mean of all biopsy samples as reference boundaries for healthy (blue line) and cancer groups (red line) and include the 90% confidence interval (CI) (red and blue halos) (Fig 6). This quantitative visualisation reduces complexity and highlights relativity of disease gene expression changes at a chaperome-scale. The polar maps recapitulate characteristic chaperome deregulation signatures in GSA-derived clusters of functional family upregulation and downregulation signatures, for instance in Cluster I cancers such as lung adenocarcinoma (LUAD) (Figs 2 and 6A), or the inverse trend with overall chaperome downregulation in Cluster II cancers such as pheochromocytoma and paraganglioma (PCPG) (Figs 2 and 6B). Concordantly, chaperome polar maps reveal characteristic patterns of cancer groups stratified based on differential expression of ATP-dependent chaperones versus ATP-independent co-chaperones, Group 1 versus Group 2 cancers (Figs 3 and 6). In LUAD, representative of Group 1 cancers, most functional chaperome families, with a preferential enrichment of ATP-dependent chaperones, are upregulated, while sHSPs are reduced (Fig 6A). On the contrary, in Group 2 cancers such as PCPG, gene expression of most chaperome functional families is downregulated (Fig 6B). Inconsistencies between these broad clusters exist, suggesting differences in tissue of origin and molecular underpinnings of respective cancers. However, broad commonalities between distinct cancers originating from the same organ are revealed. For instance, lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) share overall similarity, revealing only subtle differences, for instance in HSP40 expression (Figs 2 and S3). The kidney cancers KICH, KIRP, and KIRC also show similar patterns. As Group 2 cancers, they share and stand out against other cancers with a lack of preferential upregulation of ATP-dependent chaperones (Fig 3), and overall reduced upregulation, or downregulation, of HSP60s (Figs 2, 3, 4B and S3). A recent study indeed implicated HSP60 downregulation in tumorigenesis and progression of clear cell renal cell carcinoma (KIRC) by disrupting mitochondrial proteostasis [51].
Overall, these contextual quantitative representations enable an appreciation of the complex chaperome shifts in different cancer tissues derived from > 10,000 patient biopsy samples. The resulting compendium of differential cancer chaperome polar plots (S3 Fig) is also available online through the Proteostasis Profiler (Pro2) tool associated to this study.
The integration of disease-related differential transcriptomic changes with the cellular protein interactome network, or the edgotype, is instrumental to our understanding of genotype—phenotype relationships [31]. Towards integrated quantitative views of cancer chaperome deregulation, we curated a high-confidence physical chaperome protein-protein interactome network (CHAP-PPI) to serve as coordinate base grid layout for the analysis of differential chaperome topographies (Fig 1C).
We started with 328,244 unique human PPIs (edges) between 16,995 proteins (nodes) downloaded from the BioGRID, IntAct, DIP, and MINT databases [52],[53],[54],[55]. Zooming in on cancer chaperome alterations in the context of physical interactome wiring, we extracted the CHAP-PPI considering the 332 human chaperome genes as previously described [4]. Considering edges with the PSI-MI annotation 'physical association' we obtained 272,367 unique physical edges, of which 666 unique edges connect 220 chaperome nodes. We developed a custom script to curate the high-confidence physical CHAP-PPI, considering edges with multiple pieces of evidence, either experimental methods or publications (PMIDs), as more reliable than those supported by only a single piece of evidence. The curation script resolves ambiguous database annotation of methods terms through up-propagation within the PSI-MI ontology tree, only accepting uniquely different or rejecting identical experimental evidence. Automated interactome curation results in eight curation levels (L1—L8), through which we obtain three interactomes of increasing confidence level (see Methods). All 666 unique physical edges between 220 chaperome nodes, without curation for type or number of evidence, represent the single evidence chaperome interactome (SE-CHAP). Curating for high-confidence interactions, we obtained a multiple evidence chaperome (ME-CHAP) comprised of 222 unique physical chaperome edges between 128 chaperome nodes, of which a subset of 132 interactions between 96 nodes is supported by multiple different experimental methods (MM-CHAP) (S2 Table).
In order to enable focussed views on transcriptomic alterations of top-level chaperome functional families, we collapsed individual nodes onto functional family meta-nodes, and edges shared between families were collapsed as meta-edges such that meta-node sizes correspond to the number of family members and meta-edge thickness represents the number of shared interactions between families. We considered the meta-interactome derived from the curated high-confidence ME-CHAP interactome (S2 Table), where all meta-nodes corresponding to the 10 functional chaperome families are fully inter-connected in a single network component. We set node colour to visualize cancer gene expression changes (M-scores) and applied a force-directed spring layout algorithm to optimize graph layout [56]. The resulting integrated cancer chaperome meta-interactomes visualize relative chaperome differential changes at reduced complexity across diverse human cancers in the context of physical interactome connectivity (Figs 7A and S4). Next, we extract x-y coordinates of the chaperome meta-nodes in the optimized meta-network graph to serve as a 2-dimensional base grid (x-y coordinates) guiding the spatial layout of 3-dimensional chaperome topographic maps of differential chaperome gene expression changes (M-scores) between cancerous and healthy biopsies (z coordinate) (Figs 7B and S5).
This interactome-guided topographic display of differential chaperome alterations enables dimensionality and complexity reduction for the coherent display and comparative analysis of functional network shifts that can serve to compare differential changes i) in disease versus controls, ii) between diseases and disease classes, and iii) between perturbed or unperturbed states across large numbers of heterogeneous genomic datasets. Furthermore, this visualization lends itself for a systems-level assessment of PN deregulation topologies and their readjustment in human disease and therapeutic intervention. We implemented topographic map visualisations into the Proteostasis Profiler (Pro2) suite of tools, to improve accessibility and applicability by the scientific community.
Here we exemplify a systematic analysis of differential chaperome gene expression alterations in cancers and neurodegenerative diseases. We reduce complexity through the focus on top-level chaperome functional families. The challenge in this analysis is in the complexity and heterogeneity of available samples for disease groups such as cancers, combined with the multitude of diverse biological processes interconnected within the PN and within its functional processes, as highlighted here at hands of the human chaperome.
To date, there has been no systematic interactome-guided analysis of the implications and alterations of cellular proteostasis biology at a systems-level, in a comprehensive set of diseases, such as cancers. Here, we showcase an integrated analytical workflow for the dimension reduction, analysis and visualization of chaperome differential alterations in a representative set of human solid cancers. Our approach focuses on the visualisation of a confined set of Meta-PCA derived quantitative M-scores as descriptors of top-level chaperome functional families. We have developed “Proteostasis Profiler” (Pro2) as an integrated web-based resource and suite of tools, for interactive dimensionality-reduction, analysis and visualisation of disease-specific alterations of proteostasis functional arms, such as the chaperome, in the context of the interactome network. In this study we highlight Pro2 use-cases for the human chaperome across TCGA solid cancers in comparison to the three major neurodegenerative diseases using differential gene expression heat maps (∆GSA) (Figs 2 and 5), Meta-PCA derived quantitative polar plots (M-scores) (Figs 6 and S3), meta-interactomes and interactome-guided 3-D topographic maps (Figs 7, S4 and S5). Pro2 provides an integrated online suite for the application of the underlying algorithms. Pro2 is accessible directly at http://www.proteostasys.org.
Cancer prevalence, genetic complexity and heterogeneity represent unmet medical need and a significant challenge to personalized medicine, calling for genome-informed therapeutic intervention strategies [57]. While important progress has been made in the elucidation of proteostasis alterations in human diseases, revealing numerous alterations of PN functional processes not only in neurodegenerative or metabolic diseases but also in cancers, paradoxically the characteristics and extent of PN alterations in cancers are largely unexplored and not understood at a systems-level. Cancer cell line global transcriptional characteristics have been extensively studied [58] and numerous individual studies have assessed alterations of various chaperone and co-chaperone expression levels in specific cancers [27, 59]. In light of limitations in the clinical translation of hypotheses derived from cell lines and the lack of a systems-level understanding of proteostasis alterations in human disease, we argue that precise quantitative maps of proteostasis deregulation in human disease derived directly from clinical biopsy data will enable precise understanding of the role of PN alterations in pathogenesis towards testable hypotheses and rationalised approaches of PR therapy [1, 60].
Here, we focused on the human chaperome, a central PN component, and highly conserved facilitator and safeguard of the healthy folded proteome using an expert-curated human chaperome functional gene ontology comprising an ensemble of 332 chaperone and co-chaperone genes [4] to systematically characterize chaperome alterations in a representative clinically relevant dataset of 22 human solid cancers with matching healthy tissue, corresponding to over 10,000 patient biopsy samples provided through the TCGA consortium [30]. We found the human chaperome to be consistently highly upregulated across the vast majority of cancers assessed. While numerous individual chaperones and co-chaperones have previously been found upregulated in individual cancers [27, 59], this knowledge has not been coherently derived from consistent data resources or systematic genome-wide analyses in biopsy tissue before. Here, we provide systematic quantitative maps of chaperome deregulation in cancers that highlight the relevance, characteristics and extent of chaperome upregulation in cancers. Our analysis revealed chaperome deregulation signatures that not only feature broad upregulation of ATP-dependent chaperones but also consistent repression of ER-specific chaperones and the ATP-independent sHSPs. The data also suggest two cancer groups that can be stratified specifically by their chaperome deregulation patterns. Overall chaperome upregulation across cancers is in agreement with existing evidence on individual chaperones that has been previously reviewed [59, 61, 62]. For instance, elevated heat shock protein expression levels have been reported for HSP90 in breast and lung cancers [63, 64], HSP70 was found increased in breast, oral, cervical and renal cancers [65–68], and HSP60 showed increased expression in Hodgkin’s disease [69]. The cellular safeguarding functions of chaperones are subverted during oncogenesis to facilitate malignant transformation in light of increased translational flux and aberrant protein species in cancer cells [27]. Increased chaperone levels have previously been correlated with poor prognosis and cancer survival [59, 63, 70]. Chronic dependency on stress response and quality control mechanisms drives cancer cells into a phenotype of non-oncogene addiction [28]. The observed extent of chaperome alterations suggests a broader state of cancer “chaperome addiction”, beyond the dependency on individual chaperones.
Evidence points towards functional associations between increased proteostasis buffering capacity and maintenance of “stemness”, immortality and proliferative potential in both cancer cells and pluripotent stem cells [27]. For instance, autophagy was found to maintain “stemness” by preventing senescence through sustained proteostasis [71]. Increased proteasomal activity and elevated levels of the HSP60 chaperonin complex TRiC/CCT have recently been linked to stem cell identity by conferring proteostasis robustness [37, 38]. Fundamental similarities between stem cells and cancer raise the question to the extent of similarity between cancer and stem cell PN states and capacity. Our data suggest that cancers consistently display signatures of elevated proteostasis functional processes such as the chaperome and proteasome-mediated clearance, and are in agreement with the hypothesis that upregulated clearance mechanisms such as the proteasome and increased chaperome topologies, particularly increases in ATP-driven chaperones such as the HSP60 chaperonin complex TRiC/CCT, confer increased proteostasis capacity and survival benefits to cancer cells just like they are essential to stem cell biology. Precise knowledge of systems-level network deregulation therefore sheds light on fundamental processes at play from stem cell biology to cancerogenesis. Chaperone upregulation is largely regulated through heat shock factor 1 (HSF1) [29]. Overexpression of the TRiC/CCT subunit CCT8 protects against hsf-1 knockdown in C. elegans [38], consistent with a regulatory connection between TRiC/CCT and HSF1 [72]. Connecting processes at the PN level, this evidence suggests a connection between TRiC/CCT and HSF1 stress response signalling also in cancers [38, 73]. While increased expression of TRiC/CCT subunits has been observed in cancer cell lines [74], and increases in CCT8 expression are linked to individual cancers [75, 76], we describe consistent TRiC/CCT upregulation within global cancer chaperome signatures throughout the majority of TCGA solid cancers, or Group 1 cancers, whereas Group 2 cancers lack chaperome and, to large extent, TRiC/CCT upregulation.
Contrary to stem cell proteostasis, which is set up to maintain pluripotency and proliferative capacity, neurodegenerative diseases such as Alzheimer’s (AD), Huntington’s (HD), and Parkinson’s disease (PD) display signs of proteostasis functional collapse. Misfolding diseases feature overexpression of aggregation—prone proteins such as Aβ in AD, α-synuclein in PD, or huntingtin in HD that entail a “toxic-gain-of-function” resulting in chaperome overload, gradually exceeding proteostasis capacity [77], while “loss-of-function” misfolding diseases feature specific perturbations such as dysfunctional ∆F508-CFTR in cystic fibrosis [78]. Functionally deficient steady-state dynamics of the folding environment affect cellular protein repair capacity and proteome maintenance [79]. Most cancer cells however harbour manifold genetic aberrations even at the karyotype level that likely entail dramatic effects on proteome balance [80]. The collective damage caused by oncoprotein expression, compromised DNA repair, genomic instability, reactive oxygen species (ROS), elevated global translation and chaperome overload triggers stress response mechanisms in light of a challenged cellular proteostasis capacity [81]. Chaperome deregulation dynamics observed in cancers indeed display concordantly opposed trends as compared to alterations in the major neurodegenerative diseases.
A recent study linked repression of ~30% of the human chaperome in aging brains and in neurodegenerative diseases to proteostasis functional collapse and pointed to the role of a chaperome sub-network as a conserved proteostasis safeguard [4]. Intriguingly, while only ~8% of the human orthologous chaperome had protective phenotypes upon functional perturbation in C. elegans models of amyloid β (Aβ) and polyQ proteotoxicity, chaperones and co-chaperones far less well studied than HSP90 had equally strong protective effects [4]. Similarly, an overlap between the chaperome and the “essentialome” set of 1,658 core fitness genes in K562 leukemia cells [82] found only 55 overlapping with the 332 chaperome genes [83]. Interestingly, HSP60s showed the highest fraction of essential chaperones in agreement with their function as a highly conserved folding complex that hosts ~10% of the proteome’s clients [40]. Collectively, these findings suggest a highly functionally redundant and robust role of the central conserved chaperome within the PN [84].
In summary, our study showcases a systematic profiling of the extent of chaperome deregulation, as a central PN functional arm, in a panel of human cancers and three major neurodegenerative disorders, accompanied by a resource of quantitative multi-dimensional maps with reduced complexity. Therapeutic PN regulation for increased or restored proteostasis capacity may be beneficial in both loss-of-function and gain-of-toxic-function diseases of protein misfolding [2]. Attenuating the PN on the other hand, such as inhibiting chaperones like HSP70 and HSP90 or the UPS clearance machinery, are widely acknowledged as promising therapeutic avenues in cancers [27, 83, 85, 86]. While this manuscript was in preparation, Rodina and co-workers reported findings on a highly integrated chaperome subnetwork, or ‘epichaperome’, as a classifier of cancers with high sensitivity to HSP90 inhibition, while cancers with a less interconnected chaperome are less vulnerable by HSP90 inhibition [26]. Several HSP90 inhibitors have shown encouraging results in clinical trials [87]. Our study further supports the central role of the chaperome in PN biology, justifying particular focus on understanding chaperome alterations in human diseases at a systems level. The characteristic signatures of cancer chaperome alterations revealed in this study suggest broad commonalities and differences that could serve as testable hypotheses for therapeutic chaperome targeting strategies in cancer. Our results underline the value of charting quantitative systems-level maps and provide a resource towards an improved functional understanding of proteostasis biology in health and disease. A systems-level understanding of contextual PN alterations throughout the human diseasome will be instrumental for charting a clearer picture of the PN as a therapeutic target space, and as a resource for clinical biomarkers, including the chaperome. In face of increasing amounts of genome-scale disease data we are confronted with tremendous challenges of data complexity. Therefore, our study provides Proteostasis Profiler (Pro2), an integrated web-based suite of tools enabling processing, analysis and visualisation of proteostasis alterations in human diseases at reduced dimensionality, towards hypotheses-building for mechanistic understanding and clinical translation.
Focussing on pan-cancer analysis of the human chaperome, we chose The Cancer Genome Atlas (TCGA) as the main source for our analyses, as an established dataset that is widely used and adopted by the scientific community. The Broad Institute TCGA GDAC Firehose was accessed to download TCGA RNAseqv2 raw counts data followed by application of the voom method for the transformation of count data to normalized log2-counts per million (logCPM) [88]. Each of these logCPM values were centered gene-wise for sample normalization and comparability and used for all analyses. Considering TCGA clinical data annotation, we extracted those 22 tissue biopsy group datasets that provide both “primary solid tumor” and “solid tissue normal” sample type annotations.
We applied Gene Set Analysis (GSA) [33], an advanced derivative of Gene Set Enrichment Analysis (GSEA) [89], in order to assess chaperome gene family expression changes between cancerous and corresponding healthy tissue samples. When applying GSA, we implemented 100 permutations of chaperome genes contained in each functional family in order to allow for statistical assessment of differential expression upon re-standardization of gene groups for more accurate inferences. When applying GSA to the chaperome as one set in comparison to the whole genome (non-chap set), we randomly sampled 332 genes from the whole genome, excluding chaperome genes, and compared them to the 332 chaperome genes in order to exclude bias on group sizes in the comparisons. We applied this random sampling process 100 times in addition to 100 permutations we had on each GSA calculation. We calculated the mean value of all results as a robust measure of chaperome changes with respect to the genome. Results are displayed as heatmaps indicating significance of up or down-regulation of gene expression as ∆GSA values derived from the difference of (1—upregulation p value)—(1—downregulation p value) in disease compared to matching healthy tissue for TCGA cancer datasets, or control patient biopsies for neurodegenerative disease datasets (AD, HD, PD). ∆GSA values are normalized within the interval [-1, +1], where ‘+1’ indicates significant upregulation (upregulation p value = 0), while ‘-1’ indicates significant down-regulation (downregulation p value = 0), accordingly. Bar graphs represent group mean changes of each chaperome functional family gene group over all diseases.
We subdivided the human chaperome into functional subsets of chaperones and co-chaperones, and further divided chaperones into two sets of ATP-dependent and ATP-independent chaperones according to the annotations provided by Brehme et al. 2014 [4]. We performed linear modelling using the Limma package in R. Genes with p values < 0.05 following Benjamini-Hochberg correction are considered in the fraction of differentially expressed genes corresponding to each functional subset.
Gene Set Analysis (GSA) is a statistical hypothesis testing method that is by definition limited to confirmatory data analysis with respect to pre-existing hypotheses. In order to serve the goal of quantitative exploratory pan-cancer chaperome analysis, while retaining a maximum information content during model reduction, we devised Meta-PCA, a novel quantitative multi-step dimension reduction model fitting strategy based on principal component analysis (PCA). Principal component analysis (PCA) uses orthogonal transformation to convert a set of variables to linearly uncorrelated variables, such that they are ordered by their information content, which allows for removal of dimensions with lowest information content for dimensionality reduction in complex heterogeneous datasets. In order to stratify cancer from healthy biopsy gene expression samples based on chaperome functional family gene expression in highly convoluted datasets comprising multiple different cancer types, we designed Meta-PCA as a novel two-step method capable of handling this type of heterogeneous data. We hypothesized that each chaperome functional family or process can be described by a low number of variable dimensions, considering that genes within each group are either related or act together in molecular complexes. Therefore, we used a PCA-based approach for quantitative assessment and dimensionality reduction of functional chaperome alterations based on disease gene expression data. Challenged by highly varying sample counts in the different TCGA cancer group datasets, where datasets (tissues) with high sample numbers are at risk of dominating PCA results as compared to cancer groups with low sample numbers, we developed a custom approach that is not limited by a lack of underlying models for interpolation or undesirable loss of information, such as in up- or down-sampling, respectively, allowing us to consider all samples in the included TCGA cancer groups. Assuming distinct roles for each chaperome functional group we define
MCHAPx=Fx(Gx)
(1)
where M denotes the M-score of chaperome (CHAP) family x, Gx is the vector of gene expression values corresponding to genes in CHAP family x, and Fx is the function we want to fit. For simplicity, we considered a linear first degree model as follows: Fx is a vector of weights Wx with identical length as the vector Gx, and we aim to find Wx for all x using PCA. Assuming equivalent biological function of each CHAPx among all tissues, we first calculate Fx for each tissue in order to separate disease from healthy samples for each tissue, and then combine all “relevant” PCs in order to obtain the main underlying PC, or ‘Meta-PC’, of the corresponding CHAP group. We outline the ‘Meta-PCA’ algorithm as follows:
Step 0: For each CHAP group and tissue we assume a model
MCHAPxt=Fxt(Gxt)
(2)
Where MCHAPxt is the M-score of CHAP group x in tissue t, Fxt is the unknown function mapping gene expression values for CHAP group x in tissue t to an M-score value, and Gxt is the gene expression vector of all genes in CHAP group x in tissue t.
Step 1: Assuming MCHAPxt can be approximated using PC1, we assume Fxt is equal to W tx, which is the vector of weights for CHAP group x and tissue t. Then we calculate PCA on the gene expression matrix (GEX) comprising all genes in CHAP group x, and all samples of tissue t, including ‘solid tissue normal’ and ‘primary solid tumors’. So in this step we have Fxt≃Wxt as loadings of PC1.
Step 2: The Fxt≃Wxt assumption in Step 1 is not necessarily true; PCA extracts the most variable direction in GEX, but in case CHAP group x does not change drastically between healthy and cancer, PC1 will represent an unwanted variable or even noise. So we have to filter out the Fxt that did not fit well to the data. For this we use Student’s t-test. For each tissue, we test the separation of MCHAPxt between ‘solid tissue normal’ and ‘primary solid tumor’ samples, and discard all Fxt with p values > 10−4.
Step 3: We combine all Wxt to obtain Wx, which is the universal mapping of gene expressions in CHAP group x to its corresponding M-score, regardless of tissue type. Therefore, we calculate Wx as
Wx≃M¯xt
(3)
where the loading of each gene in the universal mapping is the mean value of all the loadings of the same gene on different tissues. Importantly, prior to calculating mean loadings, we set all Wxt to be uni-directed in order preserve directionality of change from healthy to cancer, yielding final Meta-PCs. Wx can be used as the universal function Fx (Eq 1) in order to map a query sample to the corresponding M-score of CHAP group x.
Step 3’: In order to validate Fx and resulting M-scores we performed random forest regression using 80% of M-scores and their annotation labels as training set and 20% as test set.
In order to visually represent quantifications of chaperome functional family differential cancer gene expression, we used Meta-PCA fitted functions in order to calculate disease-specific M-scores for each chaperome functional gene group as described. We then plotted relevant M-scores using polar plots, such that radial axes represent functional processes.
Human physical protein—protein interactions (PPIs), hereafter referred to as ‘edges’, were downloaded on 23 Dec 2016 from the BioGRID [52], IntAct [53], DIP [54], and MINT [55] databases. In order to obtain a high confidence chaperome physical protein—protein interactome network, we developed a custom Python script to curate raw interactome pairs, or edges, as downloaded from the above databases, considering edges detected by multiple experimental methods as more reliable than those detected by only a single method. Similarly, edges supported by multiple publications are considered at higher confidence than edges supported by only one study. Edges supported by multiple methods and / or multiple studies are collectively referred to as ‘multiple evidence’ (ME), of which those identified by multiple different methodologies represent a subset of highest confidence (MM). The Python script processes the interactome raw data as follows: UniProt IDs are mapped to NCBI Entrez Gene IDs and for each human PPI between any two chaperome members (nodes), interacting partners are mapped to Gene IDs. Only edges annotated with PSI-MI term 'physical association' type are considered. Eight different curation levels exist:
Considering these curation levels, three physical chaperome (CHAP) interactomes of increasing confidence level are obtained (S2 Table):
Different PPI source databases may annotate an identical reported PPI to different PSI-MI terms situated at different depth of the same branch within the PSI-MI ontology tree. In these cases, PPIs that are actually only supported by one piece of evidence can unintentionally be mislabelled as multiple evidence PPIs. Our automated quality curation script resolves this problem through up-propagation within the PSI-MI—ontology tree. Assume one PPI is annotated with two different interaction detection methods, A and B, then 1) if PSI-MI ontology tree levels of method A and method B are identical but their PSI-MI terms (IDs) are different, then the methods are considered as different, otherwise A and B are considered the same and the interaction is eliminated from the MM-CHAP interactome, 2) if the level of method A is higher (deeper in the ontology tree) than the level of method B, then the code searches for its parent situated at the same level as method B and compares the parent method ID with B to determine if the methods are identical or different.
We considered 2-dimensional physical interactome information to guide the spatial layout (x-y coordinates) of human chaperome functional ontology families in a 3-dimensional (x-y-z coordinates) topographic representation of chaperome M-score changes between disease and healthy tissue (z coordinate). Physical chaperome protein-protein interactome network data (PPIs) was obtained and curated as described above. We considered a network involving only high quality curated interactions supported by multiple pieces of evidence (ME-CHAP). We used the R package iGraph in order to collapse nodes corresponding to each level 1 functional ontology family into meta-nodes, and edges shared between all members of any two different level 1 functional families into meta-edges, such that meta-node size corresponds to the number of family members and meta-edge thickness represents the number of shared interactions between two families. Meta-node colour is set to reflect gene expression changes of each respective functional family in disease. We then applied a force-directed network graph layout algorithm to the meta-network according to Kamada and Kawai [56] and extracted resulting x-y coordinates of each family meta-node in the network. We used Python to draw the meta-network according to the parameters obtained in iGraph to serve as interactome-guided base grid for disease-specific quantitative 3-dimensional topographic network representations. To this end we expanded the 2-dimensional network landscape with Meta-PCA derived chaperome M-score values (z coordinate).
We designed a web-based Proteostasis Profiler (Pro2) in order to enable visual exploration of the data and results described in this manuscript, obtained through our algorithms and visualisation tools. Pro2 is accessible directly at http://www.proteostasys.org. Pro2 is implemented using Django (https://www.djangoproject.com/))), a web framework written in Python language (https://www.python.org). All the charts in the tool are generated using the plotly platform (https://plot.ly). The Pro2 tool itself is hosted on the Heroku platform (https://www.heroku.com).
All R and Python scripts and code related to this manuscript are accessible through the Proteostasis Profiler (Pro2) Github repository at https://github.com/brehmelab/Pro2.
|
10.1371/journal.pntd.0002665 | Spatial Analysis Spotlighting Early Childhood Leprosy Transmission in a Hyperendemic Municipality of the Brazilian Amazon Region | More than 200,000 new cases of leprosy were reported by 105 countries in 2011. The disease is a public health problem in Brazil, particularly within high-burden pockets in the Amazon region where leprosy is hyperendemic among children.
We applied geographic information systems and spatial analysis to determine the spatio-temporal pattern of leprosy cases in a hyperendemic municipality of the Brazilian Amazon region (Castanhal). Moreover, we performed active surveillance to collect clinical, epidemiological and serological data of the household contacts of people affected by leprosy and school children in the general population. The occurrence of subclinical infection and overt disease among the evaluated individuals was correlated with the spatio-temporal pattern of leprosy.
The pattern of leprosy cases showed significant spatio-temporal heterogeneity (p<0.01). Considering 499 mapped cases, we found spatial clusters of high and low detection rates and spatial autocorrelation of individual cases at fine spatio-temporal scales. The relative risk of contracting leprosy in one specific cluster with a high detection rate is almost four times the risk in the areas of low detection rate (RR = 3.86; 95% CI = 2.26–6.59; p<0.0001). Eight new cases were detected among 302 evaluated household contacts: two living in areas of clusters of high detection rate and six in hyperendemic census tracts. Of 188 examined students, 134 (71.3%) lived in hyperendemic areas, 120 (63.8%) were dwelling less than 100 meters of at least one reported leprosy case, 125 (66.5%) showed immunological evidence (positive anti-PGL-I IgM titer) of subclinical infection, and 9 (4.8%) were diagnosed with leprosy (8 within 200 meters of a case living in the same area).
Spatial analysis provided a better understanding of the high rate of early childhood leprosy transmission in this region. These findings can be applied to guide leprosy control programs to target intervention to high risk areas.
| Leprosy can lead to physical disabilities and deformities if not diagnosed and treated early. Even today, the disease affects more than 200,000 people per year, particularly the poorest people from developing countries, such as India, Brazil and Indonesia. Cases among children <15 years old have been used as an important indicator of recent transmission in the community. Recently, geographic information systems and spatial analysis have become important tools for epidemiology, helping to understand the transmission dynamics of several diseases. In this work, we determined the spatial and temporal distribution of leprosy in a hyperendemic municipality of the Brazilian Amazon region. In association with clinical, epidemiological and serological data of household contacts and school children in the general population, we further correlated the occurrence of subclinical infection and overt disease with the distribution of reported cases. We identified heterogeneity in the distribution of leprosy, with significant clusters of high and low detection rates. Our analysis revealed that children with leprosy or those harboring subclinical infection were in close proximity to spatial and temporal clusters of leprosy cases. These findings can be applied to guide leprosy control programs to target intervention more systematically to areas where the risk of leprosy is high.
| Leprosy is a chronic granulomatous infectious disease caused by the obligate intracellular organism Mycobacterium leprae that affects mainly the skin and peripheral nerves, which can lead to severe physical disabilities and deformities if not diagnosed and appropriately treated with multidrug therapy (MDT) in its early stages. Evidences suggest that M. leprae can spread from person to person through nasal and oral droplets and this is considered to be the main route of transmission, especially among household contacts of untreated multibacillary (MB) patients. M. leprae multiplies very slowly (12–14 days) and the mean incubation period of the disease is about three to five years, but symptoms can take as long as 30 years to appear. Early detection and properly MDT treatment are the key elements of leprosy control strategy [1].
Although leprosy has been successfully suppressed in developed countries, 219,075 new cases in 105 countries were detected in 2011, as reported to the World Health Organization (WHO), with India, Brazil and Indonesia contributing 83% of all new cases [2]. Brazil, with 33,955 new cases detected in 2011 (according to the official numbers of the Brazilian Ministry of Health), has one of the highest annual case detection rates in the world (17.65/100,000 people), and the prevalence rate has yet to be reduced below the threshold of 1/10,000 people – the level at which leprosy would be considered “eliminated” as a public health problem [2].
The spatial distribution of leprosy in Brazil is heterogeneous: the more socioeconomically developed states in the south have achieved the elimination target, though high-disease burden pockets still remain in North, Central-West and Northeast Brazil [3]. These high-burden areas encompass 1,173 municipalities (21% of all Brazilian municipalities), approximately 17% of the total national population and 53.5% of all Brazilian leprosy cases detected between 2005 and 2007 [4]. Most of the areas with spatial clusters of cases are in the Brazilian Amazon, long recognized as a highly endemic leprosy area [3]–[6].
More than 7.5 million people live in the state of Pará, located in the Amazon region. This state is hyperendemic for leprosy both among the general population (51.1/100,000 people) and among children <15 years old (18.3/100,000 people). These annual detection rates are much higher than the Brazilian averages of 17.6 and 5.2 per 100,000, respectively, in 2011 [7]. Moreover, these rates can be considered an underestimation of the real situation in Pará because only 42% of the population is covered by the primary health care service, responsible for leprosy control implementation and active case finding [8].
Leprosy in children is strongly correlated with recent disease and active foci of transmission in the community, particularly within families living in the same household, reflecting the inefficiency of local control programs for the timely detection of new cases and prompt MDT treatment, which would break the continuous spread of the disease [9]. Furthermore, the prevalence of undiagnosed leprosy in the general population has been estimated to be much more in highly endemic areas, ranging from two to eight times higher than the registered prevalence [10]–[13]. A recent cross-sectional study of 1,592 randomly selected school children from 8 hyperendemic municipalities in Pará revealed that 4% were diagnosed with leprosy based on clinical signs and symptoms [14]. By means of an ELISA test to determine the serological titer of IgM anti-PGL-I (the M. leprae-specific phenolic glycolipid-I antigen), 48.8% of the students were positive, indicating immunological evidence of subclinical infection. Indeed, it was estimated that there may be as many as 80,000 undiagnosed leprosy cases among Pará students [14]. Moreover, it was demonstrated that 2.6% of the household contacts of those people affected by leprosy during the last 5 years in Pará also have leprosy and that 39% of them have a subclinical infection of M. leprae [15]. Individuals who have a positive antibody titer to PGL-I have an estimated 8.6-fold higher risk of developing leprosy than those who are seronegative [16].This scenario of a high hidden prevalence and of subclinical infection urges new studies and innovative interventional approaches.
Geographic information system (GIS) technology and spatial analysis have been applied to identify the distribution of leprosy at national, regional and local levels [4], [17]–[19]. These new analytical tools are used to monitor epidemiological indicators over time, to identify risk factors and clusters of high endemicity and to indicate where additional resources should be targeted. The findings obtained by these methods are useful to increase the effectiveness of control programs, targeting areas of higher risk [20], which is particularly important in regions where available public health resources are scarce. GIS technology can also help to monitor the extent of MDT coverage and, as in the case of other classical tropical diseases or diseases of poverty, could play a major role in vaccine-efficacy or chemoprophylaxis trials [21].
In a previous cross-sectional study performed in June 2010 [15], we described the prevalence of undiagnosed leprosy and of subclinical infection with M. leprae among household contacts and school children in the municipality of Castanhal, located in the Brazilian Amazon region. In the present study, we applied spatial analysis techniques to identify the distribution of leprosy in this hyperendemic municipality. We describe the spatio-temporal distribution of reported cases and its correlation with the occurrence of new cases or subclinical infection among household contacts and school children of public schools.
This study conforms to the Declaration of Helsinki and was approved by the Institute of Health Sciences Research Ethics Committee from the Federal University of Pará (protocol number 197/07 CEP-ICS/UFPA). All data analyzed were anonymized.
Our study was performed in Castanhal (1.29°S; 47.92°W), located 68 kilometers NE of Belém, the capital of the Brazilian State of Pará. The population size was 173,149 inhabitants in 2010, with 88.5% living in the urban area [22]. According to the municipal Secretary of Health, there were 633 newly detected leprosy cases from January 2004 to February 2010 and 132 in 2012 (24.2% among children <15 years old). The annual case-detection rate in the general population was 73.7/100,000 inhabitants in 2012 (roughly four times the rate for Brazil as a whole); such a rate ranks the municipality as hyperendemic according to the parameters designated by the Brazilian Ministry of Health (≥40/100,000) and significantly higher than Pará's average (51.1/100,000) [7].
The residences of people affected by leprosy in the urban area of Castanhal and reported during the period of 2004 to February 2010 were georeferenced to produce detailed maps of the leprosy distribution. Additionally, spatial statistical methods were applied to identify patterns and possible risk factors associated with M. leprae infection.
Leprosy is a compulsory notifiable disease in Brazil; thus, all patients that are detected through clinic-based passive demand, active surveillance and so on have their clinical data and addresses registered in the national notifiable diseases information system (SINAN). A random sample of 90 subjects from 11 urban neighborhoods, identified as leprosy cases from 2004 to February 2010, were electronically selected. These individuals were visited at their homes by a team of health care professionals with experience in treating leprosy patients. Their household contacts were clinically assessed for signs and symptoms of leprosy, and a sample of peripheral blood from each person was collected to identify the prevalence of IgM antibodies against PGL-I [15].
The residential addresses and demographic and epidemiological variables (age, gender, year of notification and operational classification of all cases notified during the defined period) were collected from SINAN. The exact location of each residence in the urban area was then georeferenced using a handheld GPS receptor (Garmin eTrex H, Olathe, KS, USA). However, not all addresses were mapped with a GPS because many areas of Castanhal are difficult to reach and unsafe. Those that could not be reached were geocoded using the Brazilian national address file for statistical purposes (http://www.censo2010.ibge.gov.br/cnefe/) provided by the Brazilian Institute of Geography and Statistics (IBGE); this database comprises all regular street addresses and its respective census tract identification around the country. In association with a high-resolution satellite imagery base map (World Imagery, ESRI, Redlands, CA, USA), we identified the street location inside the specific census tract. This alternative mapping method can result in a loss of positional accuracy of up to 100 meters but allows matching a street address with its respective census tract (the spatial unit of analysis). IBGE was also the source for the base map of the 163 urban census tracts for this city and for the last Brazilian demographic census conducted in 2010.
Combining information from SINAN, IBGE and field-work mapping, it was possible to draw point pattern and kernel case density maps, calculate the number of cases and the annual case detection rate per census tract and identify areas with the highest risk of leprosy. Clinical, epidemiological and serological data from the evaluated household contacts and school children were obtained. The subjects were clinically assessed by an experienced leprologist to detect new cases, and their antibody titers of IgM anti-PGL-I were determined by ELISA as described previously [15]. We established an ELISA optical density of 0.295 as the cutoff for being considered seropositive. The subjects were also interviewed to identify their demographic and socio-economic characteristics. Detailed information about sampling and eligibility criteria can be found in Barreto et al. [15]. All maps were produced with the spatial reference SIRGAS 2000 UTM Zone 23S using ArcGIS 10 (ESRI, Redlands, CA, USA).
We performed spatial analyses by either grouping leprosy cases per census tract or using the georeferenced position. To minimize the effects of small numbers statistical instability, in addition to the calculation of the raw annual detection rate per census tract, we also calculated a spatially empirical Bayes (SEB) detection rate (based on a queen spatial weight matrix) to smooth the differences between contiguous areas, thereby increasing the stability of the data [23]. Global Moran's I spatial autocorrelation [24] was used to investigate the spatial clustering of the raw annual detection rate per census tract. The statistical significance was evaluated by comparing the observed values with the expected values under the complete spatial randomness assumption based on 999 Monte Carlo permutations for a significance level of 0.001. A Global Moran's I correlogram, a global index of spatial autocorrelation, was calculated to identify the range within which autocorrelation is significant and the distance at which it is highest. Local Moran's I [24], as a local indicator of spatial association (LISA), was applied to identify the position of significant clusters of higher and lower detection rates.
Additionally, a Kulldorff's spatial scan statistic [24], [25] was applied to detect the most likely cluster of cases per census tract considering the population at risk per area. The main goal of this analysis was to identify a collection of adjacent census tracts that were least consistent with the hypothesis of constant risk. This method defines circles, with radii ranging from the smallest distance between two tracts to one-half of the width of the study area. The method identifies a region formed by all tracts with respective centroids that fall within the circle and tests the null hypothesis of constant risk versus the specific alternative that the risks within and outside this region are different [19], [24].
Leprosy transmission has been described as following a pattern called “stone-in-the-pond principle”, whereby not only the household contacts of a leprosy case have an increased risk of infection but also the neighbors and the neighbors of neighbors are at higher risk when compared to the general population, with risk inversely decreasing with increasing distance [18], [26], [27]. Given that association among cases is considered to be a fine-scale process, we used areas with radii of 50, 100 and 200 meters around each of the cases detected during the study period to identify the spatial proximity of leprosy cases and students examined during the school-based surveillance.
Furthermore, a multi-distance global spatial cluster analysis (Ripley's global k-function) [28] was used to identify the spatial clusters of individual leprosy cases considering a range of distance from 50 m to 3,000 m, with distance lags of 50 m. This method considers all combinations of pairs of points and compares the number of observed pairs with the number expected at all distances, assuming a random distribution and taking into account the density of points, borders of the study area and sample size [29], [30].
A local Knox test [31] to detect the spatio-temporal interaction of individual cases considering space lags of 50, 100 and 200 meters and time lags from 1 to 5 years was also applied. This method tests for possible interaction between the distance and time separating individual cases based on the number of case pairs found within a particular time-space window [32]. In our study we chose the space and time lags described above based on the average leprosy incubation period (3 to 5 years) and distances at which most of the houses of contacts are located [33]. The expected values of the test under a null hypothesis of random case occurrence (in space and time) were estimated by performing 999 Monte Carlo simulations.
Nonspatial statistics, such as Chi-squared (χ2) [34] and Mann-Whitney U tests [35], were applied to compare the proportion of seropositivity and the titers of IgM anti-PGL-I, respectively, among household contacts and school children according to the different levels of proximity to leprosy cases or hyperendemic areas. The relative risk of leprosy as a ratio of the probability of developing the disease based on exposure was also calculated for specific areas of the city according to the level of endemicity and compared to the risk in the general population (2×2 contingency table) [36].
The following software were used for the statistical analyses: Opengeoda 1.0 (GeoDa Center for Geospatial Analysis and Computation, Tempe, AZ, USA) to calculate the spatial weight matrix, spatially empirical Bayes detection rate per census tract and Local Moran's I (LISA); Clusterseer 2.3 (Biomedware, Ann Arbor, MI, USA) to perform the global Moran's I test, Kulldorff's spatial scan statistics and Knox space-time clustering test; Point Pattern Analysis (PPA) (San Diego State University, San Diego, CA, USA) to obtain the Global Moran's I correlograms; ArcGIS to calculate Ripley's K-function and BioEstat 5.0 (Sociedade Civil Mamirauá, Amazonas, Brazil) to perform the nonspatial statistics.
According to the SINAN database, of the 633 newly detected leprosy cases in Castanhal between January 2004 and February 2010, 570 (90.0%) lived in the urban area and 46 (7.3%) in rural areas; residential addresses were unavailable (missing information) for 17 (2.7%), and these were not included in the analysis. Of those living in the urban area, 499 (87.5%) were mapped, half of them directly in the field using GPS and half via remote geocoding. The other 71 urban cases were not georeferenced due to inconsistent information regarding their residential addresses. Seventy-one percent of all cases were classified as MB.
Figure 1 illustrates the population density and spatial distribution of leprosy cases in the urban area of Castanhal and classifies the census tracts according to the level of endemicity, from low to hyperendemic, following the official parameters for the annual detection rate. The smoothed detection rate (Figure 1D) produced a more refined map of leprosy compared to the raw rate (Figure 1C), decreasing the differences between the contiguous census tracts. A correlogram of the global Moran's I test showing the significant (p<0.01) spatial autocorrelation of the census tracts with the high or low raw detection rate of leprosy per 100,000 people is shown in Figure S1. Taking into account the location of the census tract centroids, the most significant (p<0.01) clustering distance was between 1 and 2 km (peaking at 1.5 km).
The kernel density estimation indicated large differences in the number of cases in different areas, ranging from 0 to 191 per square kilometer (Figure 2A). The highest case densities overlap the census tracts with high population densities, as shown in Figure 1A. Spatial statistics (LISA) detected a significant local spatial association (i.e., association between similar values) between the census tracts with high detection rates (high-high) and between areas with low detection rates (low-low) (Figure 2B). Kulldorff's spatial scan statistics also indicated the most likely cluster of leprosy cases in a specific area of the city (Figure 2C). Both statistics showed similarity in the clustering results in one of the areas but not in the others. Table 1 presents more detailed data regarding the specific regions represented in Figures 1 and 2, including the number of census tracts, population, mean individuals per house and relative risk of leprosy compared to the general population.
Based on our analyses, approximately 88,000 people, 57% of the total urban population of Castanhal, lived in census tracts classified as hyperendemic for leprosy based on the raw detection rate. The population density per square kilometer in areas of clustered high detection rates (Figure 2C, detected by Kulldorff's spatial scan statistics) was more than 2-fold higher than in areas with lower detection rates, and the risk of contracting leprosy in that cluster was almost four times the rate in the low-low areas indicated by LISA (RR = 3.86; 95% CI = 2.26–6.59; p<0.0001). Using a Mann-Whitney test, we also observed that the household density (number of individuals per house) was significantly higher (p<0.0001) in those residences with individuals affected by leprosy (mean = 5.0; SD = 2.6) than the city average (mean = 3.8; SD = 3.2). Hyperendemic census tract (raw detection rate) showed the highest relative risk (RR = 3.69; 95% CI = 2.91–4.67) when compared to the other urban areas of the city, whereas in the low-low areas (LISA test) we observed a decrease of 54% in the risk (RR = 0.46; 95% CI = 0.28–0.74). The Spatial Bayesian Smoothing of detection rates increased the number of census tracts classified as hyperendemic from 93 to 114. Using the raw and smoothed rates, we calculated the number of people whom we need to follow to detect one new case of leprosy in a cohort, and we found that the number of those individuals nearly triples when the smoothed rate was used instead of the raw detection rate (Table 1).
A total of 302 household contacts were evaluated during previous visits to 88 residences of people affected by leprosy [15]. Sixty-three examined contacts (20.9%) lived in areas of clustered high detection rates of leprosy based on LISA and Kulldorff's spatial scan statistics. However, there were no significant differences in the serological titer of IgM anti-PGL-I (p = 0.481) or in the percentage of seropositivity (p = 0.471). Of the 8 new cases detected among household contacts, 2 lived in areas of clusters of high detection rate and 6 in hyperendemic census tracts outside the clusters.
Approximately 10% of the cases from 2004 to 2010 in Castanhal involved children <15 years old. Of the 499 mapped cases, 44 were children, with 36 (82%) living in hyperendemic areas of the city. Four public schools (two elementary and two high schools) located in different peripheral neighborhoods were also visited to evaluate a randomly selected sample of students (n = 188) for the clinical signs and symptoms of leprosy and also for subclinical infection by serological assessment of anti-PGL-I titer by ELISA assay. All four schools visited were in the hyperendemic census tracts: 134 of 188 (71.3%) examined students lived in hyperendemic areas (Figure 3); 41 (21.8%) were residing within 50 meters of at least one leprosy case; and 120 (63.8%) and 178 (94.7%) were dwelling less than 100 or 200 meters, respectively, from a known case. We did not observe significant differences in the levels of IgM anti-PGL-I (p = 0.894) or in the seropositivity between these three levels of proximity (p = 0.455). One hundred and twenty five students (66.5%) were seropositive; 9 (4.8%) were diagnosed with leprosy (8 within 200 meters of a case, 7 within 100 meters and 2 within 50 meters). Additionally, when the students diagnosed with leprosy were visited at home, 3 more cases were detected among their relatives, and 7 tested positive for anti-PGL-I.
Multi-distance point pattern analysis (Ripley's k-function) identified a significant clustering of reported individual cases, starting at a distance of 50 meters (Figure S2). To assure that the remotely mapped leprosy cases (geocoded) did not affect the results of the point pattern analysis as a function of the potential loss of accuracy of this method (up to 100 m), we also performed a multi-distance point pattern analysis (Ripley's global k-function) considering only the cases mapped using GPS directly in the field, revealing the same significant pattern of spatial clustering. Additionally, using the Gi*(d) test, we observed no significant clustering pattern in the underlying population considering the variables: total population per census tract, mean people per house and density of people per square kilometer.
Using the Knox test, we determine that the reported cases were also clustered in space and time and, as expected, frequently among household contacts, as was observed in 21 houses in which more than one case (2 or 3) shared the same residence. Table 2 displays the results of the Knox space-time clustering analysis for the leprosy cases based on different space-time lags. We identified up to 406 of 499 (81.3%) mapped cases that were near other cases in both space and time, summarizing 663 space-time links in 63 clusters. Figure 4 is an expanded view of a specific region identified as a cluster of leprosy and surrounding area, showing the space-time links among cases (100 meters over a 3 year period) and the spatial relationship with a surveyed school and seropositive students. All 6 school children (3.2%) with no clinical manifestations of leprosy who tested strongly positive for anti-PGL-I (ELISA optical density >1.000), similar to that observed in multibacillary patients, were dwelling within 100 meters of at least one leprosy case, consistent with the uncovered and upcoming spatio-temporal associations.
The pattern of leprosy cases reported from 2004 to 2010 in Castanhal showed significant spatio-temporal heterogeneity, and we found spatial clusters of high and low detection rates in the urban area. Using spatial global tests, we were also able to determine that the spatial autocorrelation of both the raw detection rate at the census tract level and of individual cases occurred at fine temporal and spatial scales. According to an analysis of the spatial pattern of serological data obtained by testing students, we ascertained that children with a high serological titer of anti-PGL-I were in close proximity to spatial-temporal clusters of leprosy cases. These findings can be applied to guide leprosy control programs to target intervention to locations with the highest risk of leprosy. De Souza Dias and colleagues [20] described the first application of GIS tools to direct active case-finding campaigns at a fine geographic scale in Brazil [20] and were able to target hot spots, resulting in the enhanced detection of new cases in addition to realizing important cost reductions for leprosy control activities.
The surprisingly high previously undiagnosed prevalence of leprosy and of subclinical infection with M. leprae among school children can be explained by the close proximity of these students' homes to detected cases. It has been shown that, in addition to household contacts, people living in the vicinity of a leprosy case and their social contacts have a higher risk of infection [18], [26], [37]. In fact, because M. leprae is highly infective but has a low pathogenicity, most people who harbor a subclinical infection will never develop clinical signs and symptoms of leprosy; indeed, only about 10% of all infected individuals eventually develop leprosy symptoms [38]. Due to the slow doubling time (13 days) and long incubation period prior to the onset of frank disease symptoms (3–5 years or longer), it is likely that many hidden cases exist, although serological responses to some protein antigens have been shown to predict disease progression up to a year prior to diagnosis [39]–[43]. It has been well-established that the titer of anti-PGL-I IgM antibody is directly correlated to the bacillary index, and that very high titers to PGL-I and certain protein antigens, such as LID-1 and Ag85B (ML2028) indicate a greater risk of developing disease [27], [40], [43]. The main challenge is to discover which biomarkers of infection serve as the best predictors of who will succumb to disease. Accordingly, performing targeted surveillance on individuals living in high endemic areas and following individuals with a high titer of anti-PGL-I is a strategy that must be implemented to perform early diagnosis, prevent physical disabilities and break the chain of transmission.
A number of serological surveys have shown that the rate of anti-PGL-I seropositivity in endemic settings correlates well with leprosy incidence in the community [44], [45]. All of the surveyed schools in this study were located in the hyperendemic census tracts of the city. This finding explains the absence of significant differences in the seroprevalence or in the titer of antibodies in the students based on a geographic location, given that nearly all (95%) of them were living within 200 meters of a detected leprosy case.
As observed for the students, there were no differences in the titer of anti-PGL-I or seroprevalence among the household contacts living inside or outside a cluster of cases. This is also not surprising, given that, even outside a cluster, all household contacts were living in very high or hyperendemic areas and that the most likely source of M. leprae is a close contact that shares the same house or room. Indeed, when 942 students and 58 teachers from Castanhal were asked if they knew a person affected by leprosy, 17.7% of the students and 53.4% of the teachers answered in the affirmative. In addition to this proximity, those harboring a subclinical infection could be a potential source of contamination to others [46], rendering such frequent-, intensive- and close-social-contact environments, such as households and schools, as locations that are favorable for M. leprae transmission.
Considering its total area, the Brazilian Amazon region has the lowest population density (4.12 individuals/km2) in the country but the highest number of people per household (3.97). This is a direct result of poverty, which compels relatives and others to live together for long periods of time, especially young married couples and their children, typically under precarious sanitation conditions. Furthermore, the average household density was even higher in the residences with a leprosy case (5.0), and, for purpose of comparison, this population density per square kilometer within the cluster of leprosy (9,536/km2 – Figure 2C) was as high as New York City (10,429/km2 - http://www.census.gov). Within the context of the wide recognition that high levels of crowding facilitate the transmission of infectious disease [47], it is reasonable to suggest that improvements in the socioeconomic status and living conditions should be part of the overall leprosy control strategy.
The introduction of GIS to leprosy epidemiology brought new insight to the concept of defining contacts based on relative distance. The importance of performing periodic surveillance among household contacts and including different classes of social and neighboring contacts has been highlighted by several authors [33], [37], [48]. Bakker and colleagues [18] observed increased subclinical infection for contact groups living ≤75 meters of anti-PGL-I-positive leprosy patients. Another report described that 92% of the dwellings of contacts were within a distance of 100 meters of the index patient [33]. For this study, we selected radii of 50, 100 and 200 meters and observed significant space-time clusters within all of these distances. Leprosy was also found to exhibit a clustered spatio-temporal pattern in an analysis of more than 11,000 cases for a period of 15 years in Bangladesh [49], with most clusters having a duration of 1 or 2 years and one cluster a 4-year time span. In our study, we observed significant spatio-temporal clustering, even within a very fine geographic scale, which is compatible with direct human-to-human transmission. Most of the students diagnosed with leprosy (8 of 9) lived in close proximity to previously detected cases.
A spatially empirical Bayes smoothed case detection rate has been used in leprosy studies to smooth the random variations in small areas with few people (where small variations in the number of cases results in dramatic changes in disease rates) and to enhance the visualization of spatial patterns [17], [50]–[52]. Smoothing is also a way to estimate uncertain values for areas with no registered cases, areas where disease is not necessarily absent but may not have been detected due to operational limitations. Smoothing produced a clearer map of leprosy in Castanhal but increased the estimate of the number of people to be followed to detect one case. We agree with Odoi and colleagues [23] that the results obtained using spatial smoothing need to be treated with caution because they can mask large differences between neighboring regions.
Given that 71 (12.5%) cases in the urban area were not mapped and analyzed in this study and considering the high prevalence of undiagnosed cases in Castanhal, our data strongly supports the notion that many more individuals than those presented here, including many children <15 years old, are currently infected with M. leprae.
In the last decade, spatial analysis and GIS have become important tools for understanding leprosy transmission dynamics in resource-poor countries. Different spatial statistical methods have been applied, including Kulldorff's spatial scan statistics [53] and global and local Moran's I indices of spatial autocorrelation [54]. However, because all spatial statistics have advantages and disadvantages, more than one method may be necessary to analyze the data and to enable decision makers to determine the priority areas for targeting control activities. Overlaying individual case point maps over high-resolution satellite images from high-risk areas (not shown here to protect the individual addresses) provides a clear visualization of the leprosy problem and can help to optimize active case-finding strategies and plan further clinical, epidemiological and prophylactic studies. Additionally, combining clinical, epidemiological, serological and spatial data provided a better understanding of the transmission dynamics of leprosy at fine spatial scales and indicated high rates of childhood leprosy transmission within hyperendemic cities of the Brazilian Amazon region.
|
10.1371/journal.ppat.1003496 | Evolutionary Toggling of Vpx/Vpr Specificity Results in Divergent Recognition of the Restriction Factor SAMHD1 | SAMHD1 is a host restriction factor that blocks the ability of lentiviruses such as HIV-1 to undergo reverse transcription in myeloid cells and resting T-cells. This restriction is alleviated by expression of the lentiviral accessory proteins Vpx and Vpr (Vpx/Vpr), which target SAMHD1 for proteasome-mediated degradation. However, the precise determinants within SAMHD1 for recognition by Vpx/Vpr remain unclear. Here we show that evolution of Vpx/Vpr in primate lentiviruses has caused the interface between SAMHD1 and Vpx/Vpr to alter during primate lentiviral evolution. Using multiple HIV-2 and SIV Vpx proteins, we show that Vpx from the HIV-2 and SIVmac lineage, but not Vpx from the SIVmnd2 and SIVrcm lineage, require the C-terminus of SAMHD1 for interaction, ubiquitylation, and degradation. On the other hand, the N-terminus of SAMHD1 governs interactions with Vpx from SIVmnd2 and SIVrcm, but has little effect on Vpx from HIV-2 and SIVmac. Furthermore, we show here that this difference in SAMHD1 recognition is evolutionarily dynamic, with the importance of the N- and C-terminus for interaction of SAMHD1 with Vpx and Vpr toggling during lentiviral evolution. We present a model to explain how the head-to-tail conformation of SAMHD1 proteins favors toggling of the interaction sites by Vpx/Vpr during this virus-host arms race. Such drastic functional divergence within a lentiviral protein highlights a novel plasticity in the evolutionary dynamics of viral antagonists for restriction factors during lentiviral adaptation to its hosts.
| At least 40 primate species in the wild are infected with their own lentivirus. Many of these infections arose from cross-species transmission followed by adaptation to a new host. Host antiviral proteins, called restriction factors, work to defend against both known and novel viruses and are thus engaged in a constant arms race with viral proteins. SAMHD1 is a restriction factor that blocks lentiviral infection of certain immune cells. However, SAMHD1 is counteracted by the lentiviral proteins Vpx and Vpr. Here we show that both Vpx and Vpr have evolved to recognize distinct interfaces of SAMHD1, consistent with the idea that SAMHD1 is rapidly evolving to evade this recognition. Furthermore, we show that while the site of this antagonism has been changing back and forth throughout lentiviral evolution, the mechanism by which Vpx and Vpr antagonize SAMHD1 has remained constant. These data illustrate a novel phenomenon in which there is an evolutionary toggling of divergent recognition between a restriction factor and its viral antagonist.
| HIV-1, HIV-2, and other primate lentiviruses encode accessory virulence factors that serve to enhance viral infectivity. This is largely achieved through increasing virus replication by counteracting host antiviral proteins, known as restriction factors [1]. One such restriction factor, SAMHD1, is a deoxynucleoside triphosphate triphosphohydrolase that suppresses cellular dNTP pools [2], [3]. This prevents efficient infection of monocytes, dendritic cells (DCs), and mature macrophages by reducing the dNTP pools below the levels needed for reverse transcription of viral RNA [4], [5]. SAMHD1 has also been shown to contribute to the restriction of HIV-1 infection in resting T cells [6], [7]. SAMHD1 is composed of two structural domains: a sterile alpha motif (SAM) domain, responsible for protein-protein interactions, and a histidine-aspartic (HD) domain, responsible for the phosphohydrolase activity of the protein.
Vpx is a primate lentiviral accessory protein that antagonizes SAMHD1 function by targeting it for proteasome-mediated degradation [8], [9]. Supplying Vpx in trans to HIV-1 infected monocytes, DCs, macrophages, or resting T-cells alleviates the SAMHD1-mediated block of efficient reverse transcription, resulting in productive infection [6], [8], [9]. Vpx achieves this by directly binding to SAMHD1 and simultaneously to the DCAF1 substrate receptor of the CRL4 E3 ubiquitin ligase (CRL4DCAF1), thereby loading SAMHD1 onto this E3 complex for polyubiquitylation and subsequent degradation [8], [10]. The CRL4DCAF1 E3 complex usurped by Vpx consists of DCAF1, DDB1, CUL4 and RBX1 [8]–[12]. While SAMHD1 alone is unable to interact with DCAF1, Vpx bridges this interaction [8], [10], in such accelerating SAMHD1 protein turnover.
Despite its important role in lentivirus infectivity [13], Vpx is found in only two of the eight major lineages of primate lentiviruses: those of the SIVsmm/SIVmac/HIV-2 (Simian Immunodeficiency Virus of sooty mangabey, rhesus macaque, and Human Immunodeficiency Virus 2, respectively) and SIVrcm/mnd2 (Red Capped Mangabey and Mandrill viruses, respectively) lineages [14], [15]. In contrast, all extant primate lentiviruses encode for Vpr, a paralog of Vpx, which has primarily been shown to cause G2 arrest of infected cells [16]–[19]. Beyond their homology, Vpx and Vpr share functional similarities; namely they are both presumed to target host restriction factor(s) for degradation through their interaction with the CRL4DCAF1 complex (reviewed in [20]). Previously we showed that a subset of Vpr proteins from lentiviruses lacking Vpx are also able to degrade their host SAMHD1, and that this activity arose prior to the genesis of Vpx through recombination/duplication [21]. This indicates that the conflict between Vpx/Vpr and host SAMHD1 has been ongoing for much of primate lentiviral evolution.
One defining characteristic of restriction factors is their engagement in an evolutionary arms-race with the viruses that they inhibit [22]–[24]. This evolutionary dynamic can be identified as a signature of positive selection on host genes, such that codons accumulate mutations that result in amino acid changes at a higher frequency than expected by neutral drift. Due to the strong selective pressure imposed directly by the virus, amino acids important for the interaction between the restriction factor and its viral antagonist are often the same residues displaying the strongest signatures of positive selection [21], [24]–[26]. Indeed, SAMHD1 has undergone bursts of positive selection in the Cercopithecinae branch of Old World monkeys (OWM), and the amino acids in the N-terminal SAM domain of SAMHD1 that show the most significant signatures of positive selection are directly involved in the species-specific interaction of SAMHD1 with SIVmnd2 Vpx and SIVagm Vpr [21]. On the other hand, when the evolution of SAMHD1 is analyzed through a broader window of primates (including prosimians and New World monkeys) residues in the C-terminus of SAMHD1 are also found to be evolving under positive selection [27]. Furthermore, it has been shown that the C-terminus of SAMHD1 is critical for binding and degradation of SAMHD1 by Vpx from HIV-2 and SIVmac [10], [27], [28]. Thus, there is an apparent discrepancy between the sites of positive selection in SAMHD1 in OWM, where most of the primate lentiviruses exist, and the specificity of Vpx.
Here we resolve the complex specificity of Vpx/Vpr for SAMHD1 by showing that different orthologs of Vpx/Vpr have evolved multiple distinct means to target SAMHD1. While some Vpx proteins (those of HIV-2 and SIVmac) require the C-terminus of SAMHD1 for recognition, binding, ubiquitylation, and ultimately degradation of this host restriction factor, others (those of SIVmnd2 and SIVrcm) recognize the N-terminus of SAMHD1. Furthermore, by analyzing more diverse Vpr proteins, we found that this recognition is evolutionarily dynamic, with both N- and C-terminal recognition arising throughout Vpx and Vpr evolution. Thus, this viral antagonist has the capacity to evolve distinct specificities for its host target. We present a model to explain this unique phenomenon based on the head-to-tail orientation of SAMHD1 tetramers. This drastic functional divergence within a lentiviral protein highlights a novel plasticity in the evolutionary dynamics of viral antagonists for restriction factors during lentiviral cross-species transmission and adaptation to a new host.
The finding that amino acids in the N-terminus of SAMHD1 confer the specificity of SIVmnd2 Vpx and SIVagm Vpr for SAMHD1 [21], yet HIV-2 and SIVmac Vpx require the C-terminus of SAMHD1 for binding and degradation [10], [27], [28], suggested a certain complexity to the modes of SAMHD1 recognition by Vpx/Vpr. We hypothesized that different Vpx/Vpr proteins have evolved to interact with different surfaces of SAMHD1. In fact, although Vpx/Vpr proteins are related at the sequence level, phylogenetic analysis shows that there are two distinct Vpx clades that interact with SAMHD1, one containing HIV-2 and SIVmac, and another containing SIVmnd2 and SIVrcm [21] (summarized in Figure 1A). In addition, Vpr from some, but not all primate lentiviruses can degrade SAMHD1 [21] (Figure 1A). In order to assess how distinct Vpx and Vpr proteins recognize SAMHD1, we generated a set of SAMHD1 constructs, either truncated at the C-terminus or chimeric at the N-terminus, and assayed for the ability of multiple Vpx/Vpr proteins to degrade these SAMHD1 proteins. Chimeras and truncations were based on previous results which showed that SAMHD1 is still catalytically active and able to form tetramers when truncated at the C-terminus or the N-terminus [29].
First, we looked at the dependence of Vpx on the C-terminus of SAMHD1. Previous results have shown that truncation of the last 15 amino acids of human SAMHD1 abrogated the interaction of SIVmac and HIV-2 Vpx with SAMHD1 [10], [27]. To test for degradation of a C-terminally truncated version of SAMHD1 (ΔC SAMHD1, truncated from amino acid 612–626) (Figure 1B), we co-transfected human 293T cells with FLAG-tagged Vpx and HA-tagged SAMHD1, and assayed for a decrease of SAMHD1 expression. We initially tested HIV-2 Vpx, SIVmac Vpx, SIVmnd2 Vpx and SIVrcm Vpx for their ability to degrade their own host species SAMHD1. As expected, each Vpx tested was able to degrade its autologous host SAMHD1 (Figure 1C). However, while HIV-2 and SIVmac Vpx were unable to degrade SAMHD1 truncated at the C-terminus (Figure 1C, left panel), neither SIVmnd2 nor SIVrcm Vpx required the C-terminus of SAMHD1 for degradation, as both Vpx proteins were readily able to degrade WT and ΔC constructs (Figure 1C, right panel). This suggests that these Vpx proteins evolved differences in their specificity determinants for degradation of SAMHD1, possibly due to the selective pressure on SAMHD1. We also assayed a panel of phylogenetically distinct HIV-2 Vpx genes representing clades A, B, and H as well as Vpx from SIVsmm (Figure S1A) for degradation of WT and ΔC human SAMHD1. We found that while all but one of the Vpx tested could degrade full-length SAMHD1, none was able to degrade the ΔC construct (Figure S1B). This suggests that the dependence on the C-terminus of SAMHD1 is indeed conserved within the diversity of HIV-2 Vpx and its precursor SIVsmm Vpx, but is different from the requirements of SIVmnd2 and SIVrcm Vpx to degrade SAMHD1.
Because we have previously shown that amino acids under positive selection in the N-terminus of SAMHD1 affect binding and degradation by SIVmnd2 Vpx [21], we next tested the effects of natural variation in the N-terminus of SAMHD1 on Vpx-mediated degradation by diverse Vpx proteins. We therefore generated SAMHD1 chimeras by swapping the first 114 amino acids (including the N-terminus and the SAM domain) of human SAMHD1 with those of either mandrill or RCM SAMHD1 (Figure 2A). Human SAMHD1 was used for these chimeras because it is quite divergent at the N-terminus with 12 or 13 amino acids that differ relative to mandrill or RCM, respectively, three of which show strong signatures of positive selection (Figure 2A and Figure S2). Notably, neither SIVmnd2 Vpx (Figure 2B, top panels) nor SIVrcm Vpx (Figure 2B, lower panels) were able to degrade human SAMHD1, while HIV-2 Vpx was able to degrade both mandrill and RCM SAMHD1. Importantly, when we replaced the first 114 residues of mandrill or RCM SAMHD1 with that of human, neither SIVmnd2 nor SIVrcm Vpx were able to degrade these chimeric SAMHD1 proteins. However, both SIVmnd2 Vpx and SIVrcm Vpx (as well as HIV-2 Vpx) were able to degrade the respective reciprocal SAMHD1 chimeras, where either mandrill or RCM N-terminus was fused with the rest of human SAMHD1 (Figure 2B). Thus, the N-terminus of mandrill and RCM SAMHD1 are sufficient to confer degradation of human SAMHD1 by SIVmnd2 and SIVrcm Vpx, respectively. These results demonstrate that while SIVmnd2 and SIVrcm are insensitive to truncations at the C-terminus of SAMHD1 (Figure 1C), the species-specific sequence differences within the N-terminus of SAMHD1 are necessary and sufficient for degradation by SIVmnd2 and SIVrcm Vpx. On the other hand, HIV-2 Vpx (and SIVmac Vpx, data not shown), which depends on the C-terminus of SAMHD1 for degradation (Figure 1C), is not affected by the natural sequence variation present in the N-terminus of mandrill or RCM SAMHD1 (Figure 2).
We also wanted to determine if a Vpx shows the same requirement for either the C-terminus or changes in the N-terminus regardless of the SAMHD1 it is challenged with. Therefore, we assayed for the ability of the divergent SIVmac or SIVmnd2 Vpx to degrade multiple species' SAMHD1, either C-terminal deleted or chimeric at the N-terminus (Figure 3). Similar to the results with the autologous SAMHD1, we found that SIVmnd2 Vpx degrades both wild type and ΔC SAMHD1 from rhesus, mandrill, and RCM, while SIVmac was unable to degrade ΔC SAMHD1 from any of the four species, though it was able to degrade the wild type protein from all (Figure 3A). Conversely, when SIVmac Vpx was assayed against chimeric SAMHD1 constructs (Figure 3B), it was able to degrade all four constructs, regardless of the sequence at the N-terminus. However, SIVmnd2 Vpx was unable to degrade those with human N-termini (Figure 3B). Together, these data indicate that Vpx from distinct viral lineages have evolved different requirements for degradation of SAMHD1 and that these requirements are intrinsic to Vpx.
SIVmac Vpx has been shown to target human SAMHD1 for proteasome-mediated degradation by interacting directly with SAMHD1 as well as the E3 ubiquitin ligase complex consisting of DCAF1, DDB1, and CUL4-RBX1 [10]. However, with such disparate requirements for SAMHD1 degradation between SIVmac and SIVmnd2/SIVrcm Vpx, we asked if the N-terminal binding SIVmnd2 and SIVrcm Vpx proteins utilize the same DCAF1 substrate receptor and E3 complex as the C-terminal binding SIVmac Vpx. We assayed for interaction of SAMHD1 with Vpx and endogenous DCAF1 and DDB1 via co-immunoprecipitations (co-IP) in 293T cells (Figure 4). The presence of DCAF1 and DDB1 in the co-IP represents formation of a CRL4DCAF1 complex that includes SAMHD1 and Vpx. We found that, similar to human and rhesus SAMHD1, full-length mandrill and RCM SAMHD1 interact with Vpx from their autologous virus, as well as the ubiquitin ligase components DCAF1 and DDB1 (Figure 4B, WT SAMHD1 lanes). Consistent with previous results [10], we found that SAMHD1 does not interact with DDB1 and/or DCAF1 in the absence of Vpx (Figure S3B and S3C). Thus, formation of the CRL4DCAF1 complex is conserved between Vpx proteins, regardless of whether the Vpx protein utilizes the N- or C-terminus of SAMHD1 for binding.
These results allowed us to next ask whether the failure to degrade either N- or C-terminally truncated SAMHD1 is due to the inability of Vpx proteins to recruit the mutant SAMHD1 proteins to the E3 complex. We performed co-IP analyses of both C-terminally truncated (ΔC) and N-terminally truncated (ΔN) SAMHD1 (Figure 4A) in 293T cells, and assayed for interaction with Vpx, DCAF1 and DDB1 (Figure 4B). In agreement with the degradation profiles, both human and rhesus ΔC SAMHD1 were unable to bind to their autologous Vpx and consequently to DDB1 and DCAF1 (Figure 4B, left panel), while the ΔN truncations show a less dramatic effect on binding. Conversely, truncating mandrill or RCM SAMHD1 at the C-terminus had no effect on Vpx, DDB1, or DCAF interaction, yet ΔN SAMHD1 truncations completely blocked these interactions. These data were further confirmed through gel filtration size exclusion chromatography of SAMHD1-Vpx-DDB1-DCAF1c complexes formed with purified recombinant proteins in vitro (Figure S3), such that SIVmac Vpx was able to recruit WT and ΔN SAMHD1 to a DDB1-DCAF1c complex, but not ΔC SAMHD1 (Figure S3D; controls without Vpx-DDB1-DCAF1c, without SAMHD1, and without Vpx in Figure S3A, B, and C, respectively). On the other hand, SIVrcm Vpx was unable to recruit a ΔN SAMHD1 from rhesus, yet it was readily able to interact with WT and ΔC rhesus SAMHD1 (Figure S3E). In agreement with the degradation data, these data suggest that the inability of Vpx to degrade SAMHD1 is due to an inability to recruit SAMHD1 to the DCAF1 subunit of the E3 complex. Furthermore, regardless of the precise requirement for SAMHD1 binding, all Vpx tested can utilize the same CRL4DCAF1 E3 ubiquitin ligase to antagonize SAMHD1.
Finally, to show that these interactions result in the ubiquitin-mediated degradation of SAMHD1, we performed an in vitro ubiquitylation assay [10]. We found that when WT, ΔC, or ΔN rhesus SAMHD1 was incubated with a preformed Cul4-DCAF1c-VpxSIVmac E3 ubiquitin ligase complex, WT and ΔN SAMHD1 were readily ubiquitylated in the presence of Vpx, however ΔC SAMHD1 was not (Figure 4C, left panel). Conversely, when these SAMHD1 proteins were incubated with CRL4-DCAF1c-VpxSIVrcm, WT and ΔC SAMHD1 were ubiquitylated, however ΔN SAMHD1 was not (Figure 4C, right panel). This is in agreement with our previous data and demonstrates a conserved pathway of ubiquitin-mediated degradation despite different N- and C-terminal SAMHD1 binding by different Vpx proteins.
Our data showing that all Vpx-SAMHD1 complexes bind DCAF1 and the DDB1 module of CRL4DCAF1 suggests that the region of Vpx that binds this E3 ubiquitin ligase complex has been conserved throughout lentiviral evolution. However, our observation that different Vpx proteins bind distinct regions of SAMHD1 raised the possibility that the Vpx surface utilized to bind SAMHD1 has changed during primate lentiviral evolution. Thus, we wanted to investigate if N-terminal binding Vpx proteins require the same residues in Vpx to degrade SAMHD1 as C-terminal binding Vpx proteins.
It has been previously shown that amino acids 12, 15, 16, and 17 are necessary for the ability of C-terminally binding SIVmac Vpx to interact with human SAMHD1 in order to load it onto DCAF1 for degradation [10] and to subsequently allow for viral infectivity in macrophage and dendritic cells [30]. However, the amino acids at positions 12, 15, 16, and 17 in SIVmnd2 and SIVrcm Vpx differ from those of HIV-2 and SIVmac Vpx (Figure 5A), suggesting the possibility that differences in this region may be important for N- and C-terminal recognition. Therefore, we attempted to change the specificity of these Vpx proteins for the N- or the C-terminus by swapping amino acids 12/15/16/17 in SIVmnd2 and SIVrcm to those of HIV-2 Vpx and SIVmac Vpx, as well as the reciprocal change of HIV-2 Vpx and SIVmac Vpx to those of SIVmnd2 and SIVrcm. Changing these residues in both HIV-2 and SIVmac Vpx abolished the activity of these proteins against their host SAMHD1 (Figure 5B), indicating that the precise residues are necessary for HIV-2 and SIVmac Vpx to degrade SAMHD1. These changes also rendered SIVmnd2 Vpx unable to degrade mandrill SAMHD1 (Figure 5B), suggesting that though the specific residues differ between HIV-2/SIVmac and SIVmnd2 Vpx, the precise residues at this position are also required for recognition of SAMHD1 by SIVmnd2 Vpx. Interestingly, a SIVrcm Vpx mutant at these positions was still able to degrade RCM SAMHD1, although with slightly decreased activity (Figure 5B). Importantly, these amino acid changes did not change the specificity of SIVrcm Vpx for the N-terminus of SAMHD1 versus the C-terminus of SAMHD1 (Figure S4). Nonetheless, these data suggest that SIVrcm Vpx utilizes a distinct interface to bind to RCM SAMHD1 for recruitment to the CRL4DCAF1 E3 complex.
Our previously published data suggests that the ability of Vpx to degrade SAMHD1 arose first in Vpr, prior to the duplication/recombination event that led to the genesis and subsequent subfunctionalization of Vpx [21]. Given our results that SAMHD1 recognition is distinct between different Vpx lineages, we wished to determine how the specificity for SAMHD1 evolved in Vpr proteins that arose prior to the Vpx duplication (e.g. SIVdeb, SIVsyk, and SIVmus Vpr, Figure 1A). We found that SIVsyk and SIVmus Vpr, like HIV-2 and SIVmac Vpx, were unable to degrade a rhesus SAMHD1 that was truncated at the C-terminus (Figure 6A, left panels), suggesting a C-terminal dependence. We further assayed for the dependence on natural variation at the N-terminus of SAMHD1 by testing each Vpr for their ability to degrade human SAMHD1 and a chimeric rhesus SAMHD1 with human SAMHD1 N-terminus (schematic as in Figure 2A). We found that SIVsyk Vpr cannot degrade human SAMHD1, but can degrade the rhesus SAMHD1 chimera with a human N-terminus (Figure 6A, right panels). This indicates that determinants for SIVsyk Vpr are not at the N-terminus of SAMHD1, and is consistent with the inability of SIVsyk Vpr to degrade C-terminally truncated SAMHD1 (Figure 6A, left panels). SIVmusVpr is able to degrade both human SAMHD1 and the human-rhesus chimeric SAMHD1, which further suggests that SIVmus Vpr is not affected by natural variation in the N-terminus (Figure 6A, right panels). These results suggest that both SIVmus and SIVsyk Vpr depend primarily on the C-terminus of SAMHD1 for degradation.
SIVdeb Vpr is interesting because its ability to degrade SAMHD1 is not affected by truncation of the C-terminus of SAMHD1 (Figure 6A, left panels), yet it is also not affected by natural variation of human SAMHD1 at the N-terminus either (Figure 6A, right panels). This result is consistent with the previous finding that SIVdeb Vpr can degrade SAMHD1 from all primate species tested [21], and suggests that determinants at the N- and C-terminus are partially redundant for recognition by SIVdeb Vpr. SIVagm.gGri Vpr also appears to involve interactions at both N- and C- termini (Figure S5).
To directly test the specificity of SIVdeb Vpr for SAMHD1, we used purified recombinant proteins to assay for the ability of SIVdeb Vpr to in vitro ubiquitylate DeBrazza's SAMHD1 lacking the N-terminus, the C-terminus, or both the N- and C-termini (ΔN/ΔC SAMHD1, Figure 6B). Consistent with our degradation and interaction results, SIVdeb Vpr is able to polyubiquitylate both ΔC and ΔN SAMHD1, although both were reduced relative to WT SAMHD1. Interestingly, SIVdeb Vpr is still able to polyubiquitylate ΔN/ΔC SAMHD1, though less efficiently than ΔC or ΔN SAMHD1. This is in contrast to SIVmac which cannot catalyze the polyubiquitylation of ΔC or ΔN/ΔC SAMHD1, and SIVrcm Vpx which cannot catalyze the polyubiquitylation of ΔN or ΔN/ΔC SAMHD1 (Figure 6B). These data suggest that in addition to contributions of both the N- and the C-terminus, SIVdeb Vpr could have evolved to further recognize the central portion of SAMHD1 for binding and degradation. Together, these data argue that the Vpx/Vpr-SAMHD1 interface is a dynamic interface whose requirements for binding and degradation have toggled back and forth through the evolution of this virus-host arms-race (summarized in Figure 6C).
We show that distinct Vpx/Vpr proteins require disparate termini of SAMHD1 in order to antagonize this host restriction factor. These differences correlate with the phylogenetic separation of Vpx, with the SIVsmm/SIVmac/HIV-2 Vpx proteins requiring the C-terminus of SAMHD1, while the SIVmnd2/SIVrcm Vpx proteins require the N-terminus of SAMHD1 for binding, ubiquitylation, and subsequent degradation. Thus, despite a conserved mechanism of degradation, Vpx from distinct lineages of SIV show both phylogenetic and functional disparity. Most surprisingly, the specific requirements in SAMHD1 for degradation/binding have toggled back and forth during Vpx/Vpr evolution in primate lentiviruses. These results demonstrate that the antagonism between Vpx/Vpr and SAMHD1 is evolutionarily dynamic, with virus and host sides evolving to counteract each other through an evolvable interface.
The arms-race between antiviral host proteins and viral antagonists often depends on sequence variation at a single interface between virus and host proteins that results in a rapid evolution at the point of contact [24], [31]. Our results with Vpx/Vpr and SAMHD1 indicate that more complex evolutionary scenarios can exist in which the mode of recognition between the viral antagonist and host protein can change. While it is possible that recognition of the N- or the C-terminus of SAMHD1 by Vpx/Vpr arose independent of each other, an attractive alternative is that Vpx/Vpr can sample both the N- and the C-terminus of SAMHD1 for optimal binding. This hypothesis is supported by biochemical studies of SAMHD1, which show that these termini of SAMHD1 are likely proximal in three-dimensional space. SAMHD1 was recently crystalized as a head-to-tail dimer [2], and further biochemical evidence indicates that tetramerization of SAMHD1 is necessary for the phosphohydrolase activity of the protein [29]. Thus the proximity of the SAMHD1 N- and C-terminus could allow Vpx/Vpr to shift its binding from one terminus to the other (Figure 7). In this model we suggest that both N- and C-terminal domains in SAMHD1 contribute to overall binding, but that one or the other shows stronger affinity depending on which Vpx/Vpr protein is binding (Figure 7A). Evolutionary pressure to evade this antagonism would create escape mutations in SAMHD1 that would destabilize the interaction with Vpx/Vpr (Figure 7B). However, new Vpx/Vpr variants that re-establish the interaction with SAMHD1 could do so by strengthening the interaction at either the N-terminus or the C-terminus of SAMHD1 (Figure 7C). Thus giving rise to the evolutionary toggling of Vpx/Vpr specificity for SAMHD1.
The model proposed in Figure 7 would hypothesize that some Vpx/Vpr proteins are dependent on both N- and C-terminal sequences for binding and degradation. Indeed, our data on SIVdeb Vpr further supports this hypothesis, as SIVdeb Vpr degradation of SAMHD1 is unaffected by truncations at the C-terminus and occurs independent of variation in the N-terminus. Furthermore, SIVdeb Vpr can polyubiquitylate both N- and C-terminal truncations of SAMHD1, though to a lesser extent then WT (Figure 6). Together, the identification of Vpr proteins that show requirements for multiple domains in SAMHD1 additionally supports our hypothesis that different affinities to either the N- or the C-terminus are what drive the toggling of Vpx/Vpr specificity for SAMHD1.
This evolutionary interaction of Vpx/Vpr with SAMHD1 differs from previously described arms-races between virus and host factors, where the escape and re-establishment of interaction takes place at a discrete interface [31]. For example, the PRYSPRY domain of the antiviral factorTRIM5α is highly variable between primate species [26], and this variation governs the functional recognition of retroviral capsid [32]. Moreover, variation in a single region of the restriction factor APOBEC3G changes specificity for the interaction with the lentiviral Vif protein [33]. When lentiviruses have adapted to changes in the host through alternate binding modes, amino acid deletions (and theoretically alterations in protein structure) were the driving force behind these adaptations. For example, Vif has shifted its recognition of APOBEC3G by three to 15 amino acids in a specific subset of lentiviruses infecting the Colobinae subfamily of Old World Monkeys, presumably due to a deletion in APOBEC3G upstream of the binding site [34]. Additionally, while Nef recognizes the restriction factor BST-2/tetherin in most primates, the Nef binding site in BST-2/tetherin is missing in humans [35]. Instead of merely shifting the region of BST-2/tetherin that Nef binds, HIV-1 adapted a different protein, Vpu, to target and degrade BST-2/tetherin. In contrast, the results shown here illustrate the extent to which a novel interface can evolve in the absence of gross changes in the sequence of the host protein under attack.
Changes in Vpx also accompany changes in SAMHD1 recognition. We found that SIVrcm Vpx does not require the same amino acids to target SAMHD1 as HIV-2/SIVmac Vpx. However, this is not specific to N-terminal binding Vpx proteins, as SIVmnd2 Vpx utilizes the same residues as HIV-2/SIVmac Vpx. How SIVmnd2 Vpx is able to bind to the N-terminus of SAMHD1 while using the same residues as C-terminally binding Vpx proteins will need to be determined. Furthermore, whether SIVmnd2 Vpx or SIVrcm Vpx is the outlier within N-terminally binding Vpx proteins is still unclear, as more N-terminal binding Vpx and Vpr proteins will need to be analyzed for their binding to SAMHD1.
Our results further help to explain the evolutionary signatures observed in SAMHD1. We had previously reported that most of the residues in SAMHD1 that show strong signatures of positive selection in Old World Monkeys were located in the N-terminal SAM domain, and through altering the amino acids evolving under positive selection, we functionally showed the SAM domain as an important interface in this virus-host antagonism [21]. Another group, analyzing a wider range of primate species, observed several sites in the C-terminus of SAMHD1 evolving under positive selection, and this region was further shown to be necessary for degradation by Vpx [10], [27], [28]. Taken together, these results argue that SAMHD1 has faced selective pressure at both the N and C-terminus, with possibly a more recent selective pressure on the N-terminus. Our results in this study indicate that evolution at both of these surfaces could have been driven by lentiviral Vpx/Vpr, highlighting the strong selective pressure that lentiviruses impose on primate evolution. Moreover, our results suggest that adaptive changes in SAMHD1 have been partially responsible for driving the changes in Vpx/Vpr that permit toggling of the region of SAMHD1 that is recognized. Together, our data not only fortifies the importance of the Vpx/Vpr-SAMHD1 interface in the evolutionary history of lentiviral interactions with Old World monkeys, but it further highlights the evolutionary plasticity of an interface between restriction factors and their lentiviral antagonists.
SAMHD1, Vpx, and Vpr primary sequences, as well as LPCX-HA SAMHD1 and pCDNA-3xFLAG Vpx or Vpr constructs were described previously [21]. SIVsyk Vpr was cloned from the proviral plasmid [36]. SAMHD1 ΔC, ΔN, and chimeric constructs were generated by PCR and subcloned into LPCX or pCG vector [8], using standard cloning techniques. Vpx mutants were generated using Quickchange site-directed mutagenesis PCR (Stratagene). The following genes were codon-optimized and synthesized (Genscript): HIV-2 Vpx n019015, n012055, n012002, n004008, n012034, and SIVsm SL2 Vpx.
Transfection in 293T cells was performed as described [21] with minor modifications. Briefly, cells were transfected with 200 ng of LPCX-HA-SAMHD1 with or without 100 ng of Vpx/Vpr constructs using TransIT-LT1 (Mirus Bio). Codon optimized Vpx/Vpr was titrated in order to normalize for similar levels of protein expression. The total amount of DNA in all transfections was maintained constant with appropriate empty vectors. Thirty-six hours post-transfection, cells were harvested for Western blot analysis.
Immunoprecipitations were performed as described previously [8], [10]. For whole cell lysates, 293T cells were lysed in RIPA buffer for 10 minutes on ice, and spun at 15,000 g for 10 minutes to pellet. Lysates were run on NuPAGE Novex 4–12% Bis-Tris gradient gels (Invitrogen). The following antibodies were used: HA-specific antibody (Babco), anti-FLAG M2 antibody (Sigma-Aldrich), anti-tubulin (Sigma-Aldrich), anti-DDB1 (Invitrogen); DCAF1 was detected with rabbit antibody raised to recombinant protein [37]. Primary antibodies were detected with a corresponding horseradish peroxidase-conjugated secondary antibody (Santa Cruz Biotech).
Phylogenetic trees were constructed from amino acid alignments of Vpx sequences obtained from the Los Alamos HIV sequence database [38]. Alignments were performed using fast statistical alignment (FSA) [39] or Muscle [40]. Phylogenies were constructed with PhyML [41] by the maximum-likelihood (ML) method. Support for ML trees was assessed by aLRT [42]. Analyses were performed at least three times.
Recombinant Rhesus SAMHD1 including full-length (WT), 1–595 (SAMHD1-ΔC) and 113–625 (ΔN-SAMHD1) and human DDB1-DCAF1c (1045–1396), and DDB1-DCAF1c in complex with viral accessory proteins, including Vpx (SIVmac) and Vpx (RCM) were expressed and purified as described previously [10]. Typically, 100 µL of protein mixtures as indicated were injected into an analytical size exclusion column (Superdex200, 10×250 mm, 24 mL) equilibrated with a buffer containing 25 mM sodium phosphate, pH 7.5, 150 mM NaCl, 5% glycerol, and 0.02% azide at a flow rate of 0.8 mL/min. The peak fractions (0.5 mL) were concentrated to 20-fold and analyzed by SDS-PAGE. The gel was first developed with Coomassie Blue stain, and subsequently Silver stained.
In vitro ubiquitylation was performed as previously described [10]. Briefly, E1 (UBA1, 0.2 µM), E2 (UbcH5b, 2.5 µM), and E3 complexes (mixtures of DDB1-DCAF1c-Vpx and CUL4A-RBX1 at 0.3 µM, indicated as CRL4-DCAF1-Vpx) were typically incubated with 0.6 µM of T7-tagged SAMHD1 and 2.5 µM of His6-FLAG-tagged ubiquitin in a buffer containing 10 mM Tris-HCl, pH 7.5, 150 mM NaCl, 5% Glycerol, 20 U/mL pyrophosphatase, 2 mM DTT and 5 mM ATP at 37°C for 0, 15 and 30 min. The reaction was quenched with non-reducing SDS-PAGE gel loading buffer and analyzed by Western blotting.
|
10.1371/journal.pcbi.1000146 | Classifying RNA-Binding Proteins Based on Electrostatic Properties | Protein structure can provide new insight into the biological function of a protein and can enable the design of better experiments to learn its biological roles. Moreover, deciphering the interactions of a protein with other molecules can contribute to the understanding of the protein's function within cellular processes. In this study, we apply a machine learning approach for classifying RNA-binding proteins based on their three-dimensional structures. The method is based on characterizing unique properties of electrostatic patches on the protein surface. Using an ensemble of general protein features and specific properties extracted from the electrostatic patches, we have trained a support vector machine (SVM) to distinguish RNA-binding proteins from other positively charged proteins that do not bind nucleic acids. Specifically, the method was applied on proteins possessing the RNA recognition motif (RRM) and successfully classified RNA-binding proteins from RRM domains involved in protein–protein interactions. Overall the method achieves 88% accuracy in classifying RNA-binding proteins, yet it cannot distinguish RNA from DNA binding proteins. Nevertheless, by applying a multiclass SVM approach we were able to classify the RNA-binding proteins based on their RNA targets, specifically, whether they bind a ribosomal RNA (rRNA), a transfer RNA (tRNA), or messenger RNA (mRNA). Finally, we present here an innovative approach that does not rely on sequence or structural homology and could be applied to identify novel RNA-binding proteins with unique folds and/or binding motifs.
| Gene expression in all living organisms is regulated by a complex set of events at both transcriptional and posttranscriptional levels. RNA-binding proteins play a key role in posttranscriptional events including splicing, stability, transport, and translation. Nowadays, there is increasing evidence that many other cellular processes may be mediated by RNA. Identifying new proteins involved in interaction with RNA is thus essential to unraveling the cellular processes in which these interactions are involved. In the current study we present a successful computational approach for classifying RNA-binding proteins and distinguishing them from other proteins based on structural and electrostatic properties. We test the method on a unique protein domain, the RNA recognition motif (RRM), which mediates both RNA and protein interactions. We show that we can discriminate RNA-binding RRMs from protein-binding RRMs. Further, we demonstrate that we can classify known RNA-binding proteins based on their RNA target (mRNA, rRNA, or tRNA). Our method does not rely on any kind of evolutionary information and thus can be applied to identify RNA-binding proteins with novel modes of RNA recognition.
| In recent years, there has been a growing appreciation for the importance of RNA and its interacting proteins. RNA-binding proteins (RBPs) function both in basic cellular processes and as key regulators of gene expression. Genome sequencing and analysis has identified many highly conserved noncoding RNAs [1] as well as numerous RBPs whose biological roles are still unknown. An increasing amount of new evidence on noncoding RNAs suggests that many other cellular processes may be mediated by RNA [2]. In most cases, RNA is found in complexes with proteins, either as large ribonucleoprotein complexes (such as the ribosome) or in more transient interactions (such as the helicase-RNA interactions) [3]. Identification of proteins involved in interaction with RNA is essential to unraveling the cellular processes in which these interactions are involved.
RBPs are characterized by a modular structure and are composed of multiple repeats that are built from a small number of basic domains that are arranged in various ways in order to satisfy their diverse functional requirements [4]. The RBPs can be classified into different families based on their basic binding motifs, for example: the RNA recognition motif (RRM), the KH domain, the double stranded RNA-binding domain (dsRBD), and the zinc finger motif [5]. Based on the first draft of the human genome, it was estimated that there are more than a thousand RBPs with known RNA-binding motifs in the genome. These numbers are expected to increase dramatically when considering all proteins that have RNA-binding capacities [6]. In recent years, new RRMs, such as the PAZ domain and the PIWI motif, which are found in the RNA-induced silencing complex (RISC), have been identified [7], revealing distinct, novel modes of RNA recognition [8]. An increasing amount of evidence on noncoding RNAs suggest that new RNA-binding motifs are yet to be discovered [9].
For many years, computational methods for identifying RNA-binding function based on structural information were not practical, due to the great diversity of the proteins and lack of structural information about them. With the exponential increase in the number of proteins being identified by genomics and proteomics projects, and specifically by structural genomics initiatives, predicting RNA-binding function from structure is now feasible. Since it is impractical to perform a functional assay for every uncharacterized protein, scientists have been turning to sophisticated computational methods for assistance in annotating the huge volume of sequence and structural data being produced. To date, many techniques are available for automatic function prediction, including: homology-based methods, phylogenetic methods, sequence patterns, structural similarity, structural patterns, methods based on genomic context, and microarray expression data [10]. Among these, several computational methods have concentrated specifically on predicting DNA-binding proteins from three-dimensional (3D) structures [11]–[16]. In addition, a couple of successful methods for prediction of RNA-binding function based on primary sequence were recently developed [17],[18].
The structural work of the last decade has elucidated the structures of many major RNA-binding protein families. Furthermore, the structures of proteins in complex with their RNA targets have shed light on how RNA recognition takes place [5]. Recently, several bioinformatics approaches have been applied for identifying RNA-binding sites on RBPs [19]–[22]. Here we present a machine learning approach to classifying RBPs, in an attempt to identify new RBPs with unique binding motifs. The method is based on characterizing the structural and electrostatic properties of the proteins. The electrostatic properties are mainly calculated from patches on the protein surfaces that are automatically extracted using our PatchFinderPlus algorithm [11],[23]. Combining an ensemble of features, we train an SVM system to distinguish RBPs from other non-nucleic-acid binding proteins that are characterized by large positive patches on their surfaces, with a very high accuracy of 88%. Applying a multiclass SVM, we show that we can successfully classify RBPs based on their RNA target (tRNA, rRNA, or mRNA), although we could not distinguish DBPs from RBPs. Interestingly, when tested on a nonredundant set of proteins that possess the RNA recognition motif (RRM), a typical RNA-binding motif known to be also involved in ssDNA binding and protein–protein interactions [24], we could successfully distinguish between RRM motifs involved in RNA-binding and the atypical RRMs involved in protein interactions.
The tremendous increase in structural information on RBPs enabled us to generate a nonredundant dataset of protein structures on which we were able to perform a comprehensive analysis. In the first step, we extracted from the Protein Data Bank (PDB) all RBP structures solved either by X-ray crystallography or by NMR. The original list was cleaned for redundancy by removing all structures that had more than 25% identity (for details see Materials and Methods). Further, the structures were annotated using the SCOP classification [25] and only protein chains including domains from unique families were retained in the final dataset. Overall, the final set included 76 nonredundant structures. As a control, we used a nonredundant database of 246 non-nucleic-acid binding protein chains (NNBP), used previously for nucleic-acid binding (NA-binding) prediction [11].
In order to examine whether the calculated features can be used to distinguish the RBPs from other proteins (specifically NNBPs that possess large positive patches), we applied a machine learning approach, namely, the support vector machine (SVM). SVMs are supervised learning methods; they take as inputs a set of features, called feature vectors, to train a model and output a classification for a query based on the model. After being trained on a set of feature vectors whose expected outputs were already known, SVMs are able to classify new input vectors. Recently, SVMs have been increasingly used in addressing the problems of protein classification, including fold recognition [35] protein structural class prediction [36], protein–protein interaction [37], membrane protein type recognition [38],[39], and G-protein coupled receptors classification [40]. Furthermore, SVMs have been utilized to solve protein classification problems and were shown to complement other methods that are based on sequence similarity [41].
We applied an SVM classifier to distinguish between the nonredundant set of RBPs and the NNBPs, as well as between the RBPs and the subset of NNBPs with large positive patches. For training, we applied a normalized feature vector that included all 40 sequence and structural parameters that were extracted from both the electrostatic patches and from the whole protein. For testing, we applied a cross-validation (leave one out) test, where for each SVM run, one protein was extracted from the training and tested separately. To evaluate the SVM performance, we plotted the ROC curve (receiver operating characteristic) describing the relationship between the false positive rate (FPR) and the true positive rate (TPR). The results of the SVM test are illustrated in Figure 4; overall we could successfully distinguish RBPs from NNBPs and from the subset of large-patch NNBPs with 88% and 86% accuracy, respectively (details in Table 1). The areas under the curve (AUCs) calculated for these experiments were 0.9 and 0.88, for the full and subset, respectively. The high performance achieved for distinguishing RBPs from other protein with large patches is extremely encouraging, since by visual inspection of the physical and electrostatic properties of the proteins one cannot distinguish between the two functionally different groups. Furthermore, when calculating each parameter independently, many of the properties did not show significant differences between the RBPs and NNBPs with large positive patches; only by combining all parameters using an SVM could we clearly distinguish between the groups. These results imply that RBPs have unique properties that can distinguish them from proteins that do not bind nucleic acids. Importantly, the distinctive properties do not relate either to the fold of the protein or to its binding motif.
To ensure that the good performance of the cross-validation test was not due to overfitting of the data, we tested an independent set of hypothetical proteins from the PDB database, which were solved by structural genomics projects and classified as RNA-binding proteins. To prevent circularity, the hypothetical proteins chosen for the test did not share more than 25% identity with any of the proteins in our training set, each representing a different fold and a different RNA-binding motif. Furthermore, since in many cases RNA binding is automatically predicted based on the existence of a known RNA-binding motif or sequence similarity, we included in the testing set only proteins that were verified experimentally to bind RNA (detailed description of the test set is given in Table S3). Overall we tested 13 proteins verified experimentally to bind RNA and 10 (78%) were successfully predicted as RBPs. Interestingly, all three false negative results were annotated to be involved in tRNA binding.
Since RBPs share many common characteristics with DBPs in terms of their electrostatics and structural features, clearly the most challenging goal would be to distinguish between these two groups. Several studies have demonstrated that RNA-protein recognition differs from DNA recognition in several aspects [22],[42],[43]. Since the RNA and the DNA adopt different helical parameters, dsDNA usually adopting a B-form while dsRNA adopts A-form helices frequently interrupted by internal loops and bulges [44], it is expected that the electrostatic patches will differ between the two types of NA binding proteins. As a first step we examined whether the new feature set selected for predicting RBPs would be as efficient for predicting DBPs. To test this, we calculated the 40 features for the set of nonredundant DNA binding proteins and built an SVM classifier for DBPs vs. NNBP. As for the RBP classifier, here too we tested the DBPs against the set of nonredundant NNBPs applied in Stawiski et al. [11],[45]. Overall the SVM performed similarly to the RBP vs. NNBP classifier, though with lower accuracy (85%). Interestingly, the current SVM results were slightly inferior to those previously reported with artificial neural network (ANN) classifiers [11]. These results are as expected, since the feature set we used in the current study was specifically designed for predicting RBPs and excluded the evolutionary information. Nevertheless, the relatively high performance achieved for predicting DBPs reinforces that the two sets of NA binding proteins have much in common. Next, we examined how well the SVM classifier discriminates between RBP and DBPs. Using the set of 40 features we were not able to distinguish RBPs from DBPs (Table 1).
It is well established that certain RNA-binding motifs can also bind DNA and vice versa (e.g., [46]). Furthermore, it is anticipated that nucleic-acid binding proteins have several roles in gene expression pathways and thus potentially have the intrinsic ability to bind both DNA and RNA [47]. Nevertheless, after excluding from our training data all proteins that bind via motifs known to bind both DNA and RNA (e.g., C2H2 zinc finger) and generating two unique data sets, single strand RBPs (ssRBPs) vs. double stranded DBPs (dsDBPs), we still could not distinguish between the RPBs and DBPs based on the above parameters. When testing on 36 dsDNA vs. 40 ssRNA-binding proteins (full list given in the Materials and Methods section), we classified only 19 as DNA-binding and 21 as RNA-binding, achieving a weak overall accuracy of 47%. This suggests that further refinement of nucleic-acid binding function will be required in order to build a classifier to distinguish exclusively RNA-binders from DNA binding proteins.
To further study the role of the electrostatic properties in discriminating RBSs from NNBPs we excluded from the SVM classifier all features related to the protein parameter group (features 19–25 in Dataset S1). Though the SVM performance was evidently reduced upon eliminating these features (Table 1 and Figure 4), we still found that the electrostatic features were sufficient for distinguishing RBPs from NNBPs. Further, to test which of the calculated features contributes most to the RNA-binding prediction, we performed a Recursive Feature Elimination procedure (RFE) (see Materials and Methods). When applying the RFE algorithm to our data, eliminating 50% of the features at each iteration, for the first three rounds of selection we did not observe notable changes in the AUC value. Only in the fourth iteration did the SVM performance decrease dramatically. The lists of the selected features that were retained in the third iteration (both when testing RBPs vs. all proteins and the RBPs versus NNBPs with large-patches) are shown in Table 2. As expected, the majority of features (8/10) selected among the top ten properties in the RBPs vs. NNBPs classifier were electrostatic-related features. Interestingly, there was a large overlap between the top ten parameters that were selected with the RFE algorithm in both classifiers. These results reinforce that the differences between the RBPs and the NNBPs are related to the function of the RBP and not simply to the size of the patch.
To further test the contribution of each one of the top ten parameters to the final SVM performance we conducted a backwards feature selection procedure and eliminated, in turn, each one of the parameters from the feature set and repeated the SVM testing (using the same cross-validation approach). For each test, we calculated the ΔAUC, which is the difference between the AUC achieved when including the feature and the AUC after excluding the feature. When testing on the full dataset of RBPs vs. NNBPs, no notable reduction was observed after eliminating a single parameter from the top ten list. Generally, the ΔAUC analysis suggests that all features that were selected by the RFE contribute equally to the SVM performance. Nevertheless, as shown in Table 1, when including only the top ten features in the RBP vs. NNBP classifier, the SVM achieved the same results as with the full parameter set. However, in the more challenging case of RBPs vs. the large patch NNBP set, all 40 features were needed to achieve the best performance (both in terms of sensitivity and selectivity). Thus for achieving the best performance for RNA-binding classification in general, we consistently use the extended classifier.
Although the 76 RBPs in our positive set were cleaned for redundancy both at the sequence and structural (family) level, within the structural groups we still had representatives of RBPs with a common binding motif (e.g., two proteins with an RRM motif). In order to be confident that the SVM results do not depend on having several proteins sharing the same binding motif within our dataset, we applied a motif-independent test. In this test we withheld, in turn, all proteins sharing a common binding motif and trained the SVM on the remaining proteins (Table S4). We then tested each member of the binding motif family on an SVM classifier from which that group had been completely withheld. As shown, the motif test performed exactly the same as the original test did, with very slight differences in the discriminating values obtained for each tested protein (Table S5). Interestingly, there was one motif group of tRNA-binding proteins which was completely misclassified (seven out of seven proteins) using both the RBP classifiers (leave-one-out vs. leave-family-out).
Overall the SVM results suggested that in the majority of cases RBPs can be uniquely characterized, independent of their binding motif. These results encouraged us to further test whether our method could discriminate RNA from non-RNA-binding proteins that possess a common binding motif. The RRM is one of the most abundant protein domains in eukaryotes. This motif is a classical RNA-binding motif, however it has been found to appear in a few ssDBPs, and most interestingly, in many proteins the RRM motif is involved in protein-protein interactions [24]. While the RRMs that mediate protein interactions commonly interact both with RNA and protein (frequently the protein-protein interactions are between two RRMs), in unique cases the RRM is solely involved in protein–protein interactions [24]. To test whether our method can distinguish between these cases, we obtained from the PDB a nonredundant set of protein chains that possess an RRM domain (Table S6). The structures were extracted automatically from PDB using a 35% sequence identity cutoff. The existence of the RRM motif was further verified against the pfam database [48]. Further, we tested each of the 27 protein chains with our SVM classifier using all 40 features. Consistent with the motif-independent test, the proteins were tested against a classifier in which the two original proteins including an RRM were excluded from the training. Overall, amongst the 27 protein chains, 21 were classified as RBPs, with one marginal prediction and six chains classified as NNBPs (Table S6).
Amid the six protein chains that were classified as NNBPs was the RRM domain of Y14 from the Y14-Magoh complex (PDB code: 1rk8A), which has been confirmed experimentally to be involved only in protein-protein interactions [24],[49]. In addition, the RRM1 domain of the SET1 histone methyltransferase (PDB code: 2j8aA) was classified as NNBP. The latter result is consistent with experimental studies which have shown that the RRM1 of the SET1 protein does not bind RNA in vitro, suggesting that the protein may be involved in RNA binding in vivo only via RRM–RRM interactions [50]. Three other chains that were predicted as NNBPs are the RRM of U2AF 35 (PDB code: 1jmtA) and the atypical RRMs (U2AF-homology motif) of U2AF65 and SFP45 (PDB codes: 1opiA and 2pe8A, respectively); all three were confirmed to be involved in protein–protein interactions in the spliceosome [51]. Interestingly the protein chain of the splicing factor SRp20, including an RRM and a TAP binding motif (PDB code: 2i2yA), was also classified as NNBP. It is plausible that these results are influenced by the existence of the TAP protein binding domain within the protein chain [52]. Notably, among the chains classified as RBPs, only in the case of elF3 (PDB code: 2nlwA) was our classification in contradiction to the experimental data, which suggests that the RRM motif does not bind RNA directly [53]. The elFj is part of a large multiprotein complex involved in initiation of translation in eukaryotes, binding the 40s ribosomal subunit. Recent studies have shown that the RRM of elFj interacts with elFb, which directly binds the ribosome [53]. Interestingly, we found the largest positive patch of the surface of elFj is on the opposite side of the RRM (data not shown), suggesting that the protein might not be interacting with the rRNA via the RRM. Consistent with our previous result, the RRMs of UP1, which binds RNA and ssDNA, was classified as RNA binding.
Overall, our results suggest that we can distinguish between RRM motifs involved in nucleic acid binding from those that are involved in protein–protein interactions. However, since our current method can only distinguish RNA from non-NA binding, in the ambiguous cases where the protein is involved in both RNA and protein interactions (either via the RRM motif or another motif), the SVM results may not be sufficient for prediction. To better understand which of the features used for the SVM training contributed to the ability of the classifier to distinguish the RNA from non-RNA-binding RRMs, we split the data into positive and negative predictions and applied the Mann–Whitney–Wilcoxon test on each one of the 40 parameters. Interestingly, the features that showed the most significant differences between the positive and negative groups were the features related to the electrostatic patches (Table S7). Figure 5 illustrates the largest positive patch in the U2B″–U2A′ complex (PDB code: 1a9nA), including an RRM known to be involved both in RNA and protein interactions, in comparison to the largest electrostatic patch in the Y14 proteins (PDB code: 1rk8A), including an RRM motif which is involved only in protein-protein interactions. In the U2B″–U2A′ complex, the large positive patch (blue) overlaps the RRM (green), which interacts directly with the RNA, while in the Y14 complex the largest positive patch is relatively small and does not overlap with the RRM motif, which is involved in the interaction with the Magoh protein.
A critical step in evaluating the strength of a classifier is to carefully examine the cases were it fails (i.e., the false negatives and the false positives). As mentioned earlier, when we analyzed the results of the SVM, we discovered that amongst the false negative results there were several tRNA-binding proteins. Previous structure analysis of the aminoacyl-tRNA synthetases demonstrated that these proteins bind tRNA via multiple domains, each of which independently recognizes different sites on the RNA [54]. In addition, it has been observed that the aminoacyl tRNA synthetases possess an unexpectedly negatively charged surface [29]. Other RBPs, such as the bacterial release factors that mimic tRNA also have highly negatively charged surfaces [55]. To further explore the unique properties of tRNA-binding proteins, we generated a set of 13 nonredundant tRNA-binding proteins that share not more than 25% sequence identity among them (six of them were in our original dataset). Further, we built a new SVM classifier for the 13 tRNA-binding proteins against all RBPs (excluding the tRNA-binding proteins). Applying a cross validation test, the SVM was able to separate the two data sets with very high accuracy (AUC = 0.94). Interestingly, when testing the misclassified proteins from the hypothetical test (Table S3) against the tRNA vs. RBPs classifier, all three proteins were classified correctly as tRNA-binding. These results are consistent with previous studies on tRNA-binding proteins that showed a very different mode of binding to RNA relative to other RNA-binding proteins [56], and are also consistent with recent sequence-based RNA-binding predictions, which demonstrated high prediction accuracy for tRNA-binding proteins [17],[18].
To test which are the most significant features for distinguishing between the tRNA-binding proteins and all other RBPs, we calculated the Spearman correlation coefficient (CC) of each one of the 40 features. Figure 6 demonstrates the correlation values (ρ) for the 40 features (numbered as in Dataset S1). Interestingly, the features that showed the highest correlations were the molecular weight and surface accessibility of the whole protein (colored in red); both were significantly higher in the tRNA group (p∼10−16), suggesting that tRNA-binding proteins are generally larger than other RBPs in our data. In addition, the roughness of the large positive patch was significantly greater in the tRNA group, while the average surface accessibility was lower in the group of tRNA binders compared to other RBPs. Strikingly, as can be noticed on the right hand side (blue bars) of Figure 6, all the ten features related to the “other patches” (i.e., the size of the negative, second and third patch, distances between the patches, etc.) were among the top ranked features that showed a significant, high CC. These results emphasize that the tRNA-binding proteins have unique electrostatic properties that can be utilized for identifying novel proteins possibly involved in tRNA processing. Moreover, we noticed that the electrostatic properties distinguishing between the tRNA and the other RBPs are mainly related to the secondary patches and not to the largest positive patch.
Following these observations, we were encouraged to test whether we could automatically distinguish between different RNA-binding strategies of known RNA-binding proteins. Previously, a multi-SVM approach was applied for classifying genes involved in different stages of the gene-expression pathway into subclasses based on microarray data [47],[56]. To test whether a multiclass approach could be applied for classifying subsets of RBPs based on the type of RNA they bind, we built three new SVM classifiers, which were trained on experimentally verified RBPs: an rRNA-binding protein classifier, an mRNA-binding protein classifier and a tRNA-binding protein classifier (see Materials and Methods). It is important to note that the groups were not split based on the RNA-binding motif and in several cases the same motif (such as the KH motif or the zinc finger motif) was found in different subsets. The 82 RBPs were tested subsequently on each of the three classifiers (in each case, the tested protein was held out from the training set). Finally, a protein was assigned a value based on the classifier in which it achieved the highest positive discriminating value. The results of the multi-SVM test are shown in Figure 7 and summarized in Table 3 (detailed results are given in Table S8). As demonstrated in Table 3, in all three subclasses the highest number of proteins was correctly assigned to the appropriate subgroup. As expected, the best results were obtained for the tRNA-binding proteins, where 13 of the 13 tRNA-binding proteins were clearly assigned as tRNA-binding. As can be observed in Figure 7C, the majority of tRNA-binding proteins also achieved a positive score in the mRNA classifier, though in all cases the scores were lower than for the tRNA classifier. Different studies have demonstrated that tRNA synthetases are also involved in mRNA-binding; for example, it was recently shown that the Glu-Pro tRNA synthetase has a role in blocking the synthesis of specific proteins by binding to the 3′ UTR of their mRNA [57]. In the rRNA-binding protein group, while the majority of the proteins (70%) scored the highest in the correct rRNA classifier, some proteins were still misclassified. Among the 14 misclassified proteins, nine were classified as mRNA and five as tRNA (Figure 7B and Table S8). These results are consistent with the notion that ribosomal proteins have several other functions in the gene expression pathway [58]. Interestingly, included in the set of rRNA proteins that were misclassified as tRNA, was the ribotoxin restrictocin bound to the sarcin/ricin domain (SRD) from the large ribosome subunit (PDB code 1jbr). This toxin disrupts elongation factor binding to the SRD domain that also binds tRNA [59]. Notably, our classification is purely based on structural information and does not rely on homology information, and thus it is expected to achieve lower performance compared to available sequence-based rRNA classification [17].
Finally, for the mRNA group we collected 23 nonredundant proteins: 13 proteins that bind mRNA at the different stages of the gene expression pathway (transcription, splicing, polyadenylation, etc.) and ten other proteins that bind mRNA such as hydrolases, export factors, viral mRNA, binding, etc. (for details see Table S8). Overall, amongst the 23 mRNA-binding proteins composed of different binding motifs, 73% of the proteins were assigned correctly (Figure 7A). Among the false negatives, five were predicted as rRNA. Notably, the false negative mRNA-binding proteins did not belong to a certain binding motif or fold (2 KH, 1 RRM, 1 LRR, 1 PUF, and 1 Zinc Finger), again reinforcing that our classification is motif-independent.
As noted, the basic assumption behind our algorithm was that the electrostatic patch is related to the nucleic acid binding interface. Thus it is expected that the success of the method would depend on the correlation between the patch residues (identified automatically by our algorithm) and the experimentally defined RNA-binding interfaces. We previously found that in DNA binding proteins the largest positive patch of the protein encompasses, on average 80% of the protein-DNA interface [11]. As demonstrated in Figure 1, the positive patch of the RBPs does not always coincide with the real binding interface. Here we tested the correlation between the patch–interface overlap and the confidence of the RNA-binding classification, as derived from the SVM. Applying an SVM, each tested protein was assigned a discriminating value (generally the distance of the protein from the hyper plane). As illustrated in Figure 8, when applying a Spearman correlation coefficient, we found a significant positive correlation (ρ = 0.64, p<10−8) between the percent overlap of the positive electrostatic patch and RBP interface and the discriminating value obtained by the SVM. These results imply that the success of the method at classifying RBPs from NNBP strongly relies on the degree of overlap between the largest positive patch and the binding interface. The correlation between the patch-interface overlap and the SVM performance is also consistent with the feature selection results that showed that the majority of the features contributing to the performance were associated with the largest positive patch.
In this study we applied a machine learning approach to classify RNA-binding function from the 3D structure of the protein. Using features extracted from the positive electrostatic patches on RNA and non-nucleic-acid binding proteins, we trained an SVM to classify RBPs. We show that our method successfully distinguishes, with relatively high accuracy (88%), the RBPs from other proteins that do not bind nucleic acids. Similar results were achieved both when applying a cross-validation (leave one out) approach and when testing an independent set of proteins solved by a structural genomics initiative and confirmed experimentally to bind RNA. However, our method was not able to distinguish between RNA and DNA binding proteins. Interestingly, although the RBPs were distinguished from non-nucleic acid binding proteins by a combination of properties, we show that the success of the classification strongly depends on the degree of overlap between the largest positive patch and the real binding interface. Furthermore, we could show that the results do not depend on the RNA-binding motif, and correct classification was also achieved when we withheld all proteins that share a similar binding motif. Overall, our method is applicable for classifying RBPs that are generally very diverse in terms of their structure, function, and RNA recognition motifs. Moreover, since the method does not rely on sequence or structure conservation, we suggest that it could be applied to identify novel nucleic acid binding proteins with unique binding motifs.
One of the great challenges in classifying ligand binding proteins (such as RBPs) is to be able to identify to which ligand it will bind. For this purpose, we have applied a multiclass SVM classifier, which was trained on three different groups of known RBPs classified according to their RNA target: tRNA, rRNA, or mRNA. In the majority of cases, given that a protein is a RBP, we could assign it to a specific subgroup. Consistent with sequence-based predictions, we succeeded in correctly predicting all tRNA-binding proteins, whereas only 70–73% of rRNA and mRNA-binding proteins were assigned correctly. Overall, the results we obtained are very encouraging, reinforcing the idea that structural properties of proteins that are not directly related to the protein fold can give clues to the protein's interacting partner. It is important to note that subclassification of the RBPs to the three subgroups (mRNA, rRNA, or tRNA) using our multiclass approach is only possible given the prior knowledge that the protein binds RNA. Finally, consistent with other recent studies, our results suggest that electrostatic features of the protein surface can contribute to fine-tuning predictions of nucleic-acid binding proteins.
A nonredundant set of RBPs was constructed based on the RNA recognition motifs definition in Chen and Varani [5]. Additional proteins have been added to the data set based on manual data mining of the RCSB Protein Data Bank using the SCOP family definition [60]. From each SCOP family, only one representative protein was added to the dataset. From each protein included in our dataset, only the chain or chains containing the RNA-binding domain were analyzed. The chains involved in RNA binding were selected by manual inspection using the PyMOL viewer [61]. All selected chains were further cleaned for redundancy, including only proteins that share less than 25% sequence identity. In addition, the PISCES program [62] was applied to automatically select for proteins with resolution better than 3.5 Å, R-factor ≤0.3, and a sequence length from 40 to 1000 amino acids.
The NNBP data set was constructed from Hobohm and Sander's “pdb select” list of proteins [63] used previously in Stawiski et al. [11], excluding all proteins involved in binding NAs. Similarly to the RBP set, the control data set was further cleaned by excluding sequences with more than 25% identity. The subset of large-patch NNBPs was selected from the control set by sorting the proteins by the size of the largest patch; the top 76 proteins were chosen: 1skf, 1a6oA, 1pbe, 1a17, 1hcl, 1a7s, 1oaa, 1gox, 1ayl, 1uae, 1oyc, 1fnc, 1hcz, 1cpt, 1pda, 1lam, 1frb, 1ido, 1drw, 1fds, 1axn, 1gky, 1opr, 1lfo, 1ciy, 1fmk, 1csn, 1nsj, 1ndh, 1a8p, 1atg, 1bg2, 1csh, 1lit, 1rcb, 1cot, 1lid, 1bdb, 1fit, 1pbv, 1br9, 1ppn, 1a53, 1czj, 1a8e, 1mai, 1dhr, 1lki, 1c52, 1mrp, 1sbp, 1php, 1gnd, 1nfp, 1af7, 1aj2, 1alu, 1rhs, 1ddt, 1amf, 1ng1, 1al3, 1koe, 1mla, 1bhp, 1lbu, 1kte, 1nox, 1amm, 1a6m, 1phd, 1gen, 1b6a, 1gsa, 1ash, 1moq
A nonredundant set of RBPs that bind ssRNA was constructed from the original dataset and includes the following 40 protein chains: 1a1tA, 1a9nB, 1aq3A, 1asyA, 1b23P, 1b34A, 1cx0A, 1ddlA, 1e8Ob, 1ec6A, 1f7uA, 1fjgB, 1fjgC, 1fjgF, 1fjgG, 1fjgI, 1fjgJ, 1fjgK, 1fjgL, 1fjgM, 1fjgN, 1fjgO, 1fjgP, 1fjgR, 1fjgS, 1fjgT, 1gtfA, 1h2cA, 1hq1A, 1i6uA, 1jidA, 1k8wA, 1knzA, 1kq2A, 1m8wA, 1mmsA, 1mzpA, 1rgoA, 1ropA, 2fmtA. The set of dsDNA binding proteins was selected from the DNA binding proteins dataset [11]. The 36 selected protein chains were: 1a02F, 1a31A, 1a3qA, 1a73A, 1aayA, 1am9A, 1b3tA, 1bdtA, 1bnkA, 1cktA, 1cmaA, 1d66A, 1ddnA, 1ecrA, 1fokA, 1hmiA, 1ignA, 1ihfA, 1lmb3, 1mnmA, 1pdnC, 1pnrA, 1sknP, 1tc3C, 1trrA, 1tupA, 1wetA, 1xbrA, 2bopA, 2dgcA, 2hmiA, 2irfG, 2nllA, 3croL 3mhtA, 3pviA
For the independent test set we extracted from PDB RNA-binding proteins that were classified as “hypothetical” or “structure genomics.” The RNA-binding function was defined based on Gene Ontology (GO) terms, considering the molecular function level http://www.geneontology.org/. In cases where GO annotation was not available, we included proteins that were defined as RNA-binding proteins in the primary citation. Further, the list was manually curated, including only proteins that were verified experimentally (based on the literature) to bind RNA. Importantly, proteins which were defined by GO as RBP based on the existence of an RNA-binding domain or on high sequence similarity to a known RBP were not included in the final list. The detailed list of the hypothetical proteins is given in Table S3.
Overall, 40 different input features were calculated; the features can be roughly classified into four major subgroups:
The PatchFinder algorithm[11] was applied to extract all continuous positive patches on the proteins surface with a cutoff of >2kT/e [23]. The patches were sorted based on the number of grid points included within the patch, and the largest three patches were selected. The largest negative patch (<−2kT/e) was extracted as described in Stawiski et al. [11]. The distances between the patches were calculated from the center of mass of each patch. Protein features were calculated as described in [11]. In addition, the dipole and quadrupole moments were calculated using the Protein Dipole Moments Server [64]. Interface residues were calculated using the Intervor web server [65]. Intervor calculated macromolecular interface using the Voronoi cells approach. This approach was shown to be highly compatible with classical surface accessibility calculations [66]. The Voronoi cells represent a convex polyhedron that contains all points of space closer to that atom than to any other atom. Two atoms are in contact if their Voronoi cells have a facet in common [66]. The overlap between the patch and the interface was calculated as the number of patch residues included in the interface divided by the total number of residues in the interface.
The F-test, Student's t-test (assuming equal variance), Mann–Whitney–Wilcoxson, and the Spearman correlation coefficient (CC) were performed using the R Stats package [67]. To account for multiple testing, the P-value was adjusted using the Bonferroni correction.
A standalone package, NAbind, for nucleic-acid binding prediction (suitable for linux OS) is available for download (Dataset S2).
|
10.1371/journal.ppat.1004946 | Virulence Factors of Pseudomonas aeruginosa Induce Both the Unfolded Protein and Integrated Stress Responses in Airway Epithelial Cells | Pseudomonas aeruginosa infection can be disastrous in chronic lung diseases such as cystic fibrosis and chronic obstructive pulmonary disease. Its toxic effects are largely mediated by secreted virulence factors including pyocyanin, elastase and alkaline protease (AprA). Efficient functioning of the endoplasmic reticulum (ER) is crucial for cell survival and appropriate immune responses, while an excess of unfolded proteins within the ER leads to “ER stress” and activation of the “unfolded protein response” (UPR). Bacterial infection and Toll-like receptor activation trigger the UPR most likely due to the increased demand for protein folding of inflammatory mediators. In this study, we show that cell-free conditioned medium of the PAO1 strain of P. aeruginosa, containing secreted virulence factors, induces ER stress in primary bronchial epithelial cells as evidenced by splicing of XBP1 mRNA and induction of CHOP, GRP78 and GADD34 expression. Most aspects of the ER stress response were dependent on TAK1 and p38 MAPK, except for the induction of GADD34 mRNA. Using various mutant strains and purified virulence factors, we identified pyocyanin and AprA as inducers of ER stress. However, the induction of GADD34 was mediated by an ER stress-independent integrated stress response (ISR) which was at least partly dependent on the iron-sensing eIF2α kinase HRI. Our data strongly suggest that this increased GADD34 expression served to protect against Pseudomonas-induced, iron-sensitive cell cytotoxicity. In summary, virulence factors from P. aeruginosa induce ER stress in airway epithelial cells and also trigger the ISR to improve cell survival of the host.
| Pseudomonas aeruginosa causes a devastating infection when it affects patients with cystic fibrosis or other chronic lung diseases. It often causes chronic infection due to its resistance to antibiotic treatment and its ability to form biofilms in these patients. The toxic effects of P. aeruginosa are largely mediated by secreted virulence factors. Efficient functioning of the endoplasmic reticulum is crucial for cell survival and appropriate immune responses, while its dysfunction causes stress and activation of the unfolded protein response. In this study, we found that virulence factors secreted by P. aeruginosa trigger the unfolded protein response in human cells by causing endoplasmic reticulum stress. In addition, secreted virulence factors activate the integrated stress response via a parallel independent pathway. Both stress pathways lead to the induction of the protein GADD34, which appears to provide protection against the toxic effects of the secreted virulence factors.
| The Gram-negative bacterium Pseudomonas aeruginosa is an opportunistic pathogen that increases morbidity and mortality in chronic lung diseases, such as cystic fibrosis (CF) and chronic obstructive pulmonary disease (COPD; GOLD stages III-IV)) [1–3]. P. aeruginosa often causes chronic infection due to its ease of developing antibiotic resistance and its ability to form biofilms in these patients. Furthermore, its survival in the host in the early stages of infection is supported by the secretion of toxins and virulence factors, including pyocyanin and its proteases elastase and alkaline protease (AprA) (reviewed in [4, 5]). Interestingly, their production appears to be lower in the later stages of infection [6, 7]. Therefore, the specific role of these virulence factors in chronic infections is incompletely understood. Pyocyanin is a redox-active toxin that causes cellular senescence [8], ciliary dyskinesia [9], increased expression of IL-8 [10] and disruption of calcium homeostasis [11] in human lung epithelial cells. Pyocyanin inactivates α1-antitrypsin, thereby contributing to the protease-antiprotease imbalance found in CF lungs [12], while P. aeruginosa elastase additionally cleaves many proteins of the extra-cellular matrix, including collagen, fibrinogen and elastin, and opsonin receptors, thus contributing to the invasion of bacteria into the lung parenchyma [13]. AprA is thought to modulate the host response and prevent bacterial clearance by degrading proteins of the host immune system, including TNFα and complement factors [14–16].
P. aeruginosa requires iron both for its respiration and for biofilm formation [17, 18]. Competition with the host is fierce and so P. aeruginosa has evolved specific strategies to obtain iron [19]. It produces redox-active phenazine compounds to turn insoluble Fe3+ to the more soluble Fe2+, siderophores to scavenge iron and receptors for the uptake of iron-siderophore complexes, proteases to degrade host iron-binding proteins, and bacteriocins to eliminate competitors (reviewed in [19]). Moreover, iron availability regulates the production of virulence factors such as pyocyanin, AprA and exotoxin A [20].
The endoplasmic reticulum (ER) functions to fold secretory and membrane proteins and its quality control systems ensure that only properly folded proteins exit the organelle. Accumulation of incompletely folded proteins can impair ER homeostasis and induces “ER stress”, which activates intracellular signal transduction pathways collectively called the “unfolded protein response” (UPR; Fig 1). This response restores ER homeostasis by reducing the influx of new proteins into the lumen of the ER and by enhancing the organelle’s capacity to fold proteins; however, if the stress cannot be resolved then apoptotic cell death pathways are invoked (reviewed in [21]).
Three distinct sensors detect ER stress: protein kinase RNA (PKR)-like ER kinase (PERK), inositol-requiring enzyme 1 (IRE1) and activating transcription factor 6 (ATF6) [21]. Early during ER stress, the kinase PERK phosphorylates eukaryotic translation initiation factor 2 on its alpha subunit (eIF2α) causing the inhibition of protein synthesis and thus preventing the load on the ER from increasing further [22–24]. In addition, this promotes the translation of specific mRNAs, for example that encoding the transcription factor ATF4 [25]. One important target of ATF4 is the transcription factor called C/EBP homologous protein (CHOP), and both individually can trans-activate the GADD34 gene [26]. GADD34 is a phosphatase that selectively dephosphorylates eIF2α, completing a negative feedback loop and enabling the translation of other targets of the UPR [27]. In parallel, IRE1 initiates the unconventional splicing of the mRNA encoding X-box binding protein-1 (XBP-1) [28]. Spliced XBP-1 mRNA encodes an active transcription factor that, in concert with ATF6, induces expression of UPR genes, such as the chaperones GRP78 (also known as BiP) and GRP94 [28–30].
The phosphorylation of eIF2α is a point at which the responses to several forms of stress are integrated [31]. During ER stress, PERK phosphorylates eIF2α, but eIF2a can also be phosphorylated by PKR responding to double-stranded RNA during viral infection [32, 33], by GCN2 during amino acid starvation [25, 34, 35], and by HRI during iron deficiency (reviewed in [31]). For this reason, the events initiated by eIF2α phosphorylation have been termed the “integrated stress response” (ISR; Fig 1 and [36]).
Abnormal function of the ER has been implicated in the pathogenesis of many diseases, including diabetes mellitus, atherosclerosis, Alzheimer’s disease and cancer [21, 37]. Remarkably, the ER also plays an important role during immune responses to infection and malignancy. For example, during bacterial infection, Toll-like receptor (TLR) activation triggers splicing of XBP1 mRNA, possibly in response to the increased biosynthesis of secreted inflammatory mediators, increasing the capacity for protein secretion and thus contributing to an augmented inflammatory response [38–40]. In addition, induction of GADD34 is required for cytokine expression during viral infection; however, in contrast to ER stress, pathogen-induced induction of GADD34 appears to be independent of CHOP [41, 42]. Nevertheless, sustained activation of the UPR can impair the immune response by triggering cell death [26, 43].
Previously, it has been shown that infection of airway epithelia or Caenorhabditis elegans with P. aeruginosa can elicit an UPR [39, 44, 45]. In worms, activation of the IRE1-XBP-1 branch of the UPR was dependent on p38 MAPK-signalling [39], but it is unknown if this signalling response is conserved in humans. Moreover, it is unclear whether living bacteria are required for the induction of ER stress or if unidentified secreted factors are sufficient.
In the present study, we set out to test the hypothesis that virulence factors secreted by P. aeruginosa trigger the UPR in human cells via the p38 MAPK pathway. We found that p38 MAPK signalling was required for the response of human epithelial cultures to bacterial conditioned medium and that the secreted factors pyocyanin and AprA contribute to the induction of ER stress. Furthermore, we showed that induction of the ISR target GADD34 is mediated by the iron-regulated kinase HRI and this induction protects the host against the toxic effects of P. aeruginosa.
Infection with live P. aeruginosa has previously been shown to induce the UPR in mouse macrophages and human immortalized bronchial epithelial cells [40, 45]. To identify whether P. aeruginosa could induce the UPR in primary bronchial epithelial cells (PBEC) and whether living bacteria were necessary for this, we stimulated PBEC with filter-sterilised conditioned medium (CM) from P. aeruginosa strain PAO1 (CM-PAO1), containing secreted virulence factors without living bacteria. Treatment with CM-PAO1 induced ER stress in a time- and dose-dependent manner, as evidenced by a 9.9-fold increase of splicing of XBP1 mRNA (p<0.01), a 12.8-fold increase of CHOP mRNA (p = 0.02) and a 16.2-fold increase of GADD34 mRNA (p<0.05) after 8–12 hours (Fig 2A and 2B). This was accompanied by an increase in phosphorylation of eIF2α and protein expression of GADD34 and GRP78 (Fig 2C). This increase in phosphorylated eIF2α was accompanied by a decrease in global protein translation as assessed by puromycin incorporation in nascent proteins (Fig 2D) [46]. In line with previous reports [47–49], CM-PAO1 gradually impaired epithelial integrity until the monolayer was completely disrupted after 24 hours. Although the epithelial layer was disrupted by CM-PAO1 (as reported by trans-epithelial resistance; S1A Fig and visualised by light microscopy; S1B Fig), the cell membranes themselves remained intact as reported by exclusion of trypan blue stain (S1B Fig).
Infection of C. elegans with P. aeruginosa has been reported to cause splicing of XBP1 mRNA in a p38 MAPK-dependent manner [39]. To exclude the effects of donor variation and complex nutrient/growth factor requirement of primary cells, we tested whether exposure of 16HBE cells, a SV-40 transformed bronchial epithelial cell line, to P. aeruginosa conditioned medium would trigger phosphorylation of p38 MAPK and activate the UPR. We observed that CM-PAO1 caused prolonged phosphorylation of p38 MAPK in 16HBE cells up to 6 hours (Fig 3A). We reasoned that the activation of p38 MAPK after 15 minutes might represent the activation of TLR signalling, since stimulation of HEK-TLR2 or HEK-TLR4 cells [50] with CM-PAO1 demonstrated robust TLR2 and TLR4 activation. The sustained activation was similar to that observed in C. elegans infected with Pseudomonas [39], which suggests the importance of p38 MAPK in the induction of the UPR. To examine if p38 MAPK signalling was required for the ER stress response, we pre-treated 16HBE cells with an inhibitor of p38 MAPK (SB203580) or an inhibitor of TAK1 (5Z-7-oxozeanol, better known as LL-Z1640-2), a kinase upstream of p38 MAPK. We then exposed cells to CM-PAO1 and observed that both compounds markedly reduced activation of p38 by CM-PAO1 (Fig 3B). In addition, both compounds reduced secretion of IL-8 in response to CM-PAO1 treatment (Fig 3C). Of note, these compounds strongly inhibited splicing of XBP1 mRNA and abrogated the induction of CHOP and GRP78 mRNA (Fig 3D). However, the induction of GADD34 was insensitive to the inhibitors (Fig 3D) suggesting the involvement of an additional pathway independent of CHOP.
To prepare P. aeruginosa conditioned medium, cultures were grown for 5 days (see Experimental procedures and [47]) to a high optical density, at which quorum-sensing is activated in this strain, thus triggering the production of a variety of virulence factors among which the cytotoxic exoproduct pyocyanin. When pyocyanin levels in CM-PAO1 were measured, values up to 5.5 μg/ml (26 μM) were detected (Fig 4A), which were similar to values observed in sputum of CF patients colonised with P. aeruginosa [51]. We first wished to determine if pyocyanin was an important mediator of the observed ER stress response by CM-PAO1. To this end, P. aeruginosa bacterial cultures were supplemented with iron to suppress pyocyanin production together with other iron-regulated factors (Fig 4A). The conditioned medium prepared in this manner was significantly less efficient at triggering the splicing of XBP1 mRNA and at increasing expression of GRP78 mRNA (Fig 4B). Surprisingly, CHOP mRNA was not significantly affected (Fig 4B), whereas GADD34 mRNA induction was completely abrogated.
These experiments provided only indirect support for the involvement of pyocyanin, since iron supplementation also affects production of other P. aeruginosa virulence factors and may also affect host cells. We therefore tested whether purified pyocyanin could induce ER stress in 16HBE cells. Treatment with purified pyocyanin caused dose-dependent splicing of XBP1 mRNA, induction of CHOP and GRP78 mRNAs and expression of GRP78 and GRP94 protein (Fig 4C and 4D), maximal at 10 μM (2.1 μg/ml). In contrast, GADD34 mRNA continued to rise up to a maximum at ≥ 30 μM (6.3 μg/ml) of pyocyanin (Fig 4D). Once again, this suggested that induction of GADD34 in this system might not simply reflect activation by ER stress. As expected, pyocyanin potently induced secretion of IL-8 by 16HBE cells (Fig 4E) [10].
Since pyocyanin is a redox active toxin, we tested the effect of co-administration of the anti-oxidants N-acetylcysteine (10 mM) and glutathione reduced ethyl-ester (10 mM) for 24 hours. Both failed to ameliorate the ER stress response suggesting that pyocyanin caused ER dysfunction independent of causing oxidative stress [52, 53] (see online repository).
Taken together, these observations suggested that conditioned medium of P. aeruginosa caused ER stress via multiple virulence factors, including pyocyanin. Furthermore, the induction of GADD34 appeared to involve an additional pathway independent of CHOP.
Having found evidence for the involvement of multiple virulence mechanisms in the induction of ER stress, we next attempted to determine their identities. The P. aeruginosa AB toxin exotoxin A is known to cause translational attenuation by catalysing the ADP-ribosylation of elongation factor 2 (EF2) [54]. We investigated whether purified exotoxin A could also induce ER stress, but detected no increase in spliced XBP1, CHOP, GADD34 or GRP78 mRNA (S2A Fig) nor the phosphorylation of eIF2α (S2B Fig). Next, to more broadly explore the involvement of other potential virulence factors, we made use of strains of P. aeruginosa that lacked specific toxic products: PAN8, a lasB aprE double mutant, which is deficient in the production of elastase [55] and the secretion of AprA; PAN11, an xcpR lasB mutant, which is deficient in the production of elastase and the secretion of all other substrates of the type II protein secretion system but still produces AprA; and PAO25, a leu arg double mutant derivative of PAO1 and the direct parental strain of both mutants (Table 1). CM-PAO25 did not differ from CM-PAO1 in the content of all toxins measured (S3A and S3B Fig) and in inducing spliced XBP1, CHOP, GADD34 and GRP78 mRNA (S3C Fig). In spite of the aprE mutation, still traces of AprA were detected in the culture supernatant of the PAN8 strain (Fig 5A), presumably due to cell lysis during the 5-days growth period.
When 16HBE cells were incubated with CM-PAN8 (lacking elastase and AprA), XBP1 mRNA splicing and induction of GRP78 mRNA were completely abolished, and only low induction of CHOP mRNA remained (Fig 5B). In contrast, the response of 16HBE cells to CM-PAN11 (containing AprA, but no elastase or other substrates of the type 2 secretion system) was much less affected relative to CM-PAO1 treatment (Fig 5B), indicating that the reduced response to CM-PAN8 is primarily due to the absence of AprA in this CM rather than to the absence of elastase. Indeed, stimulating 16HBE cells with purified elastase did not elicit an ER stress response within 24 hours (see online repository). On the other hand, incubation with 10 nM purified AprA induced the splicing of XBP1 mRNA, and up-regulated CHOP and GRP78 mRNA (Fig 5C). These experiments suggested that, in addition to pyocyanin, AprA also contributed to the induction of ER stress in 16HBE cells. We therefore next generated conditioned medium of a series of specific AprA and pyocyanin mutant strains to demonstrate the relative contribution of AprA and pyocyanin to the induction of ER stress. However, these experiments were inconclusive because the corresponding wild type strains did not induce sufficient ER stress (see online repository).
Remarkably, once again the induction of GADD34 mRNA followed a distinct trend from the other markers of ER stress. Particularly a lack of AprA (in CM-PAN8) was correlated with an increased expression of GADD34 (Fig 5B), whilst purified AprA did not induce GADD34 mRNA (Fig 5C). This suggested that an unrelated mechanism regulated GADD34 induction by CM-PAO1 and that this might be independent of ER stress.
To examine the involvement of ER stress-dependent and-independent responses to CM-PAO1, we next made use of the specific inhibitor of IRE1, 4μ8C, which blocks splicing of XBP1 mRNA during ER stress ([56] and Fig 6). Of note, this compound not only attenuated the splicing of XBP1 mRNA elicited by CM-PAO1, but interestingly, it also attenuated the secretion of IL-8 by 16HBE in response to CM-PAO1 (S4A Fig).
During ER stress, the kinase PERK phosphorylates eIF2α, thereby activating the ISR. When Perk-/- mouse embryonic fibroblasts (MEFs) were exposed to CM-PAO1, the induction of Gadd34 mRNA was unaffected, while the response to the ER stress-inducing agent tunicamycin (Tm) was abrogated (Fig 7A). However, phosphorylation of eIF2α was required for the induction of Gadd34 mRNA in response to CM-PAO1 as demonstrated by the failure of the conditioned medium to induce Gadd34 mRNA in fibroblasts homozygous for the eIF2αAA mutation, which renders them insensitive to all eIF2α kinases (Fig 7B). Moreover, ATF4, a transcription factor translationally up-regulated upon phosphorylation of eIF2α, was essential for the induction Gadd34 mRNA by CM-PAO1 (Fig 7C). As we have shown previously [26], CHOP was only partially required for tunicamycin (ER stress)-induced expression of Gadd34 mRNA (S4B Fig). The same was observed for CM-PAO1, although it did not reach statistical significance (S4B Fig). Interestingly, murine fibroblasts stimulated with CM-PAO1 failed to splice Xbp1 mRNA (S4C Fig), suggesting that activation of IRE1 by CM-PAO1 may be less important in this cell type than in human epithelial cells. However, reassuringly, ISR-dependent signalling in response to pseudomonal toxins was preserved in these cells and, once again, expression of Chop mRNA was regulated via eIF2α and ATF4. As had been observed for Gadd34, Chop induction was independent of PERK, suggesting that in MEFs treated with CM-PAO1, Chop was induced by a stimulus other than ER stress (S4D–S4F Fig).
We next examined which eIF2α kinase was responsible for activation of the ISR by CM-PAO1. To this end, we made use of Pkr-/-, Gcn2-/- and Hri-/- MEFs [25, 57, 58] and observed a significant deficit of CM-PAO1 induction of Chop and Gadd34 mRNA in Hri-/- cells, suggesting the involvement of the iron-sensing kinase HRI (Fig 7D–7F and S4G–S4I Fig). In contrast, although it has been suggested previously that GCN2 is involved in the stress response induced by P. aeruginosa in gut epithelial cells [59], we observed no significant effect on the induction of Gadd34 mRNA in Gcn2-/- cells (Fig 7E). We therefore went on to deplete either GCN2 or HRI in HeLa cells using two separate siRNA oligonucleotides for each gene and obtained similar results: whereas both siRNAs directed against HRI decreased induction of Gadd34 mRNA, one siRNA directed against GCN2 had no effect whereas the other even increased Gadd34 mRNA expression (Fig 7H and S4J Fig). Whereas we cannot exclude the possibility that this increasing effect of one siRNA directed against GCN2 may result from putative off-target effects, we conclude that these data support a role for HRI rather than GCN2.
Since RPMI is an iron-poor medium, we reasoned that the CM-PAO1 would limit iron availability to epithelial cells, e.g. by the presence of siderophores [60], which might activate HRI through depletion of iron from the culture medium. We therefore first evaluated the effect of iron depletion of the epithelial cell culture medium using deferoxamine (DFO). DFO treatment resulted in a marked increase in the expression of the ISR and UPR related genes CHOP and GADD34, whereas GRP78 and spliced XBP1 were not affected (Fig 7H). This is line with selective activation of the ISR by iron depletion. We next confirmed the presence of the iron-chelating siderophore pyoverdine in the CM-PAO1 by the bright fluorescence of the medium upon exposure to UV light (see online repository). To test the possible involvement of iron depletion in CM-PAO1-mediated Gadd34 induction, we supplemented the epithelial cell culture medium with iron, which indeed completely suppressed the induction of Gadd34 mRNA (Fig 7I and S4K Fig).
Taken together, these data demonstrate that CM-PAO1 induces splicing of XBP1 mRNA (ER stress) in human bronchial epithelial cells, while induction of GADD34 predominantly reflects an iron-dependent ISR mediated by the eIF2α kinase HRI.
During chronic ER stress in cell and animal models of disease, the induction of GADD34 appears to mediate cellular toxicity [26, 43]. In contrast, during the acute stress of SERCA pump inhibition by thapsigargin, GADD34 has been shown to be protective [61]. To test the role of ER stress-independent induction of GADD34 by exposure to CM-PAO1, we made use of Gadd34ΔC/ΔC MEFs [61], which lack GADD34 phosphatase activity. Cells expressing wild-type GADD34 were more resistant to the cytotoxic effects of CM-PAO1 compared with Gadd34ΔC/ΔC fibroblasts, as reported by the release of lactate dehydrogenase (LDH) (Fig 8A). To confirm these findings, we repeated these experiments in HeLa cells expressing GADD34 from a tetracycline-responsive promoter. The induction of GADD34 with doxycycline significantly increased cell viability upon exposure to CM-PAO1 (Fig 8B). When the cell culture medium of wild-type cells was supplemented with iron, the release of LDH was prevented (Fig 8C, left panel). Iron supplementation was also observed to rescue cell viability reported by MTT assay (Fig 8C, right panel).
Taken together, these data suggest that the toxicity of CM-PAO1 is sensitive to iron and that HRI-mediated induction of GADD34 is protective in this context. Supplementation with iron relieves both the cytotoxicity and the requirement for induction of GADD34.
It is known that a normal response to ER stress is required for an efficient innate immune response to bacterial infection [39], but whether live bacteria are required for this has been unclear. In this study, we have shown that secreted virulence factors of P. aeruginosa cause ER stress in primary bronchial epithelial cells and in a cell line, and that this is mediated by TAK1 and phosphorylated p38 MAPK. In addition, we have identified GADD34 induction via an ER-stress independent ISR. We have demonstrated pyocyanin to be one of the factors eliciting these responses, while AprA contributes to the activation of the UPR. We were however unable to establish the relative contribution of pyocyanin and AprA to the activation of the UPR. In contrast, activation of the ISR with induction of GADD34 mRNA is most likely a response to reduced iron availability and may serve a cytoprotective role during exposure to conditioned medium of P. aeruginosa.
In line with these observations, phosphorylation of p38 MAPK has previously been shown to be involved in the splicing of XBP1 upon infection with P. aeruginosa [39, 45], although the involvement of TAK1 upstream of p38 MAPK and its essential involvement in the activation of CHOP and GRP78 are novel findings. Interestingly, GADD34, classically a downstream target of CHOP, was regulated independently of the TAK1-p38 MAPK pathway. The induction of GADD34 is only partially dependent on CHOP (S4B Fig and [26]), but it is absolutely reliant on phosphorylation of eIF2α and ATF4 [26]. This is concordant with the recent description of a virus-induced “microbial stress response” mediated via the PKR/eIF2α/ATF4 pathway, which fails to induce CHOP, but potently induces GADD34 [41, 42].
In contrast to the response of human airway epithelial cells, P. aeruginosa conditioned medium failed to cause splicing of Xbp1 mRNA in murine fibroblasts, suggesting that ER stress may not be a conserved feature of the cellular response to this insult. This is unsurprising, as induction of ER stress is known to be highly cell-type dependent [40]. In the absence of ER stress in the murine fibroblasts, the induction of Chop and Gadd34 suggests that activation of the ISR by the secreted virulence factors may be a more conserved response. Of note, in human bronchial epithelial cells, the induction of CHOP seems primarily subordinate to an ER stress-induced ISR, rather than the microbial stress response (S7 Fig). Consequently, induction of CHOP was dependent on the TAK1-p38 MAPK pathway in those cells (Fig 3D) and its induction was only partially inhibited when bacterial cultures were supplemental with iron (Fig 4B), in contrast to MEFs where Chop induction was dependent on HRI (S4I Fig).
Recent evidence suggests that bacterial components may function as triggers for the UPR. Flagellin has been shown to induce an atypical ER stress response in CF bronchial epithelial cells during live infection [45], while N-(3-oxo-dodecanoyl) homoserine lactone (C12) has been observed to phosphorylate eIF2α and activate p38 MAPK [62]. We have now shown that at least two secreted virulence factors, pyocyanin and AprA, also contribute to this ER stress response to Pseudomonas. More research has to be done to assess the involvement of (other) individual virulence factors.
High concentrations of pyocyanin also mediated an ER stress-independent, ISR-dependent induction of GADD34 (Fig 4E). We were able to identify a crucial role for iron availability and for the iron-sensing kinase HRI in this response, although we cannot fully exclude a role for the kinase GCN2 that has been previously implicated in responses to Pseudomonas spp [59]. Of note, it is possible that the protective effect of GADD34 is unrelated to its ability to dephosphorylate p-eIF2alpha. Interestingly, AprA was not involved in the induction of the ISR response but rather appeared to dampen it, since considerably higher GADD34 expression was observed when conditioned medium of the aprE mutant PAN8 was used to stimulate the cells (Fig 5B). Among other possibilities, an explanation for this observation could be that AprA present in the conditioned medium of the wild-type strain partially degrades HRI, a possibility that warrants further investigations. The discovery of this ER-independent ISR may plausibly offer novel potential therapeutic targets.
It has been shown recently that spliced XBP1 is required for C12-mediated apoptosis [62]. Remarkably, exposure of cells to C12 does not itself trigger the splicing of XBP1 mRNA suggesting that basal levels of XBP1 splicing are both necessary and sufficient for this response. Moreover, the transcriptional activity of spliced XBP1 does not appear to be required for this cell death, indicating that the spliced XBP1 protein may have additional, as yet unidentified, activities. C12 appears able to trigger the ISR in an ER stress-independent matter, although the mechanism for this remains to be determined. It would be interesting to determine if C12 can activate HRI.
Chronic elevation of GADD34 in ER stress can mediate cellular toxicity [26], but GADD34 has been shown to be protective during the acute stress of SERCA pump inhibition with thapsigargin, which depletes the ER of calcium [61]. As with thapsigargin, P. aeruginosa has been associated with altered ER calcium signalling [38, 44]. It is therefore of interest that expression of GADD34 reduced cell toxicity and increased cell survival upon iron deficiency caused by treatment with conditioned medium from P. aeruginosa. It has been shown that lungs of cystic fibrosis patients lack the ability to induce GADD34 [45], which might plausibly lead to increased cytotoxicity or altered innate immunity due to Pseudomonas infection of the lungs of CF patients. However, future in vivo studies are required to confirm the observed cytoprotective effect of GADD34 induction during Pseudomonas infections.
In summary, secreted virulence factors of the PAO1 strain of P. aeruginosa, including pyocyanin and AprA, are sufficient to elicit an ER stress response but the relative contribution of these virulence factors remains to be investigated. In contrast to these virulence factors, our findings strongly suggest that iron depletion mediated by the secretion of siderophores causes an ER stress-independent ISR. The induction of GADD34 by this may serve to ameliorate the toxic effects of P. aeruginosa conditioned medium.
All strains used in this study are listed in Table 1. CM was prepared as described previously with slight modifications [47]. Briefly, overnight bacterial cultures in Luria Broth were inoculated 1:50 into RPMI 1640 (Gibco, Life Technologies, Breda, the Netherlands) and incubated at 37°C shaking at 200 rpm. After 5 days, the cultures were centrifuged and supernatants were filter-sterilized through 0.22 μm pore-size filter (Whatman, Dassel, Germany) to obtain CM. Pyocyanin and AprA levels in CM were measured as described previously [63, 64].
PBEC were obtained from tumour-free resected lung tissue by enzymatic digestion as described previously [65]. 16HBE cells (passage 4–15; kindly provided by Dr. D.C. Gruenert, University of California, San Francisco, CA, USA) were cultured in MEM (Invitrogen) supplemented with 1 mM HEPES (Invitrogen), 10% (v/v) heat-inactivated FCS (Bodinco, Alkmaar, the Netherlands), 2 mM L-glutamine, 100 U/ml penicillin and 100 μg/ml streptomycin (all from BioWhittaker). All MEFs were maintained as described previously [23, 26, 36, 66, 67]. HEK-TLR2 and HEK-TLR4 [50] were a kind gift from Prof. Dr. M. Yazdanbakhsh (Leiden University Medical Center, the Netherlands). HeLa cells were transfected for 6 hours with two different ON-TARGETplus Human EIF2AK1 siRNA (GCACAAACUUCACGUUACU and GAUUAAGGGUGCAACUAAA) and knockdown was assessed 48 hours after transfection (S5 Fig).
GADD34-N1-eGFP (kind gift form S. Shenolikar, Duke-NUS Graduate Medical School Singapore, Singapore) was excised with BglII and NotI and ligated into pTRE2-hyg plasmid (Clontech Laboratories, Mountain View, CA, USA) digested with BamHI and NotI. HeLa Tet-On advanced cells (Clontech Laboratories) were transfected with the pTRE2-hyg_GADD34-eGFP plasmid and selected with 600 μM hygromycin to generate a stable cell line conditionally expressing GADD34-GFP (S6 Fig). Positive cell clones were visualised by GFP expression in response to 1 μg/ml of doxycycline. Once identified, expanded and characterized, these clones were maintained in DMEM (Sigma) supplemented with 10% FBS and antibiotics (100 U/ml penicillin G, 100 μg/ml streptomycin, 200 μg/ml G418 and 200 μM hygromycin). Expression of GADD34 was typically induced using 1 μg/ml doxycycline (Sigma) for 24 hours.
Cells were exposed at 80–90% confluence for 24 hours (unless stated otherwise) to CM-PAO1 (1 in 5 dilution, unless stated otherwise), pyocyanin (1–30 μM), ammonium iron (III) citrate (100 μM; Fe3+), exotoxin A (1–10 ng/ml), AprA (10 nM), elastase (16–64 μg/ml) and/or DFO (1–100 nM) as indicated (all from Sigma). Puromycin (10 μg/ml; Sigma) was added 30 minutes before harvesting. Thapsigargin (100 nM; Sigma), TNFα and IL-1β (both 20 ng/ml; Peprotech, Rocky Hill, NJ) were used as positive controls. The compounds SB203580 (10 nM; Sigma) and 5Z-7-oxozeanol (also called LL-Z1640-2; 100 nM; TebuBio, Heerhugowaard, the Netherlands) were added 30 minutes before stimulation for the inhibition of p38 MAPK and TAK1, respectively. The specific IRE1-inhibitor 4μ8C (30 μM) [56] was a kind gift from Prof. Dr. D. Ron, University of Cambridge.
Cells were lysed in buffer H (10 mM HEPES, pH 7.9, 50 mM NaCl, 500 mM sucrose, 0.1 mM EDTA, 0.5% (v/v) Triton X-100, 1 mM PMSF, 1X Complete protease inhibitor cocktail (Roche Applied Science, Mannheim, Germany)) supplemented with phosphatase inhibitors (10 mM tetrasodium pyrophosphate, 17.5 mM β-glycerophosphate, and 100 mM NaF [25, 27]) for detection by antibodies directed against phospho-eIF2α (Cell Signaling Technology, Danvers, MA, USA), eIF2α (gift from Prof. Dr. D. Ron), KDEL (Enzo Life Sciences), GADD34 (ProteinTech, Chicago, IL, USA), puromycin (Millipore, Billerica, MA, USA), β-actin and GAPDH (CellSignalling), or in sample buffer (0.2 M Tris-HCl pH 6.8, 16% [v/v] glycerol, 4% [w/v] SDS, 4% [v/v] 2-mercaptoethanol, 0.003% [w/v] bromophenol blue) for detection by antibodies directed against (phospho-) p38 MAPK (both CellSignalling). The proteins in the samples were separated using a 10% SDS-PAGE gel and transferred onto a nitrocellulose membrane. After blocking with PBS containing 0.05% Tween-20 (v/v) and 5% skimmed-milk (w/v), the membrane was incubated overnight with the primary antibody (1:1000) in TBS with 0.05% Tween-20 (v/v) and 5% BSA (w/v) at 4°C. Next, the membrane was incubated with HRP-labelled anti-mouse or anti-rabbit antibody (Sigma) in blocking buffer for 1 hour and developed using ECL (ThermoScientific).
Total RNA was isolated using Qiagen RNeasy mini kit (Qiagen/Westburg, Leusden, the Netherlands). Quantitative reverse-transcriptase polymerase chain reaction (qPCR) was performed as described previously [68] using the primer pairs as defined in Table 2. Relative mRNA concentrations of RPL13A and ATP5B (GeNorm, PrimerDesign Ltd., Southampton, UK) were used as housekeeping genes for human genes and Actb (β-actin) and Sdha for mouse genes to calculate normalized expression.
IL-8 was measured using commercially available ELISA kit according to manufacturer’s instructions (Sanquin, Amsterdam, the Netherlands).
LDH release was measured with a LDH-cytotoxicity colorimetric assay kit following manufacturer’s instructions (Biovision, Milpitas, CA, USA). Thiazolyl blue tetrazolium bromide (MTT; Sigma) was dissolved in a 5 mg/ml stock concentration in sterile water and cells were incubated with a 1:10 dilution for 2 hours at 37°C. Next, the water-insoluble formazan formed from MTT in viable cells was dissolved in isopropanol for 10 min before the absorbance was read at 570 nm wavelength.
Epithelial barrier function was measured using ECIS (Applied Biophysics, Troy, NY, USA) as described previously [69]. Resistance was measured at 1000 Hz and cells were stimulated with CM-PAO1 when the resistance was stable.
Results are expressed as mean ± SEM. Data were analysed using one- or two-way analysis of variance (ANOVA) and corrected with the Bonferroni post-hoc test. Differences with P-values <0.05 were considered to be statistically significant.
|
10.1371/journal.pgen.1003997 | Comprehensive Analysis of Transcriptome Variation Uncovers Known and Novel Driver Events in T-Cell Acute Lymphoblastic Leukemia | RNA-seq is a promising technology to re-sequence protein coding genes for the identification of single nucleotide variants (SNV), while simultaneously obtaining information on structural variations and gene expression perturbations. We asked whether RNA-seq is suitable for the detection of driver mutations in T-cell acute lymphoblastic leukemia (T-ALL). These leukemias are caused by a combination of gene fusions, over-expression of transcription factors and cooperative point mutations in oncogenes and tumor suppressor genes. We analyzed 31 T-ALL patient samples and 18 T-ALL cell lines by high-coverage paired-end RNA-seq. First, we optimized the detection of SNVs in RNA-seq data by comparing the results with exome re-sequencing data. We identified known driver genes with recurrent protein altering variations, as well as several new candidates including H3F3A, PTK2B, and STAT5B. Next, we determined accurate gene expression levels from the RNA-seq data through normalizations and batch effect removal, and used these to classify patients into T-ALL subtypes. Finally, we detected gene fusions, of which several can explain the over-expression of key driver genes such as TLX1, PLAG1, LMO1, or NKX2-1; and others result in novel fusion transcripts encoding activated kinases (SSBP2-FER and TPM3-JAK2) or involving MLLT10. In conclusion, we present novel analysis pipelines for variant calling, variant filtering, and expression normalization on RNA-seq data, and successfully applied these for the detection of translocations, point mutations, INDELs, exon-skipping events, and expression perturbations in T-ALL.
| The quest for somatic mutations underlying oncogenic processes is a central theme in today's cancer research. High-throughput genomics approaches including amplicon re-sequencing, exome re-sequencing, full genome re-sequencing, and SNP arrays have contributed to cataloguing driver genes across cancer types. Thus far transcriptome sequencing by RNA-seq has been mainly used for the detection of fusion genes, while few studies have assessed its value for the combined detection of SNPs, INDELs, fusions, gene expression changes, and alternative transcript events. Here we apply RNA-seq to 49 T-ALL samples and perform a critical assessment of the bioinformatics pipelines and filters to identify each type of aberration. By comparing to exome re-sequencing, and by exploiting the catalogues of known cancer drivers, we identified many known and several novel driver genes in T-ALL. We also determined an optimal normalization strategy to obtain accurate gene expression levels and used these to identify over-expressed transcription factors that characterize different T-ALL subtypes. Finally, by PCR, cloning, and in vitro cellular assays we uncover new fusion genes that have consequences at the level of gene expression, oncogenic chimaeras, and tumor suppressor inactivation. In conclusion, we present the first RNA-seq data set across T-ALL patients and identify new driver events.
| T-cell acute lymphoblastic leukemia (T-ALL) is an aggressive malignancy that accounts for approximately 15% of pediatric and 25% of adult ALL cases. Despite improved outcome over the years, about 25% of children and 50% of adults still fail to respond to intensive chemotherapy protocols or relapse [1]. Improved understanding of T-ALL biology through the identification and characterization of oncogenic lesions is expected to lead to a better prognostic classification and the development of new targeted therapeutic strategies.
T-ALL is caused by the accumulation of multiple oncogenic mutations that have been identified through characterization of chromosomal aberrations and candidate gene sequencing [2]. Chromosomal translocations in T-ALL frequently involve the T-cell receptor (TCR) loci, whereby TCR regulatory elements become juxtaposed to genes that are normally not expressed in T-cells [3], [4]. In this way, a specific set of recurrently over-expressed transcription factors (TFs) have been documented, including TLX1, TLX3, TAL1, LMO1, HOXA, and NKX family members [5]. T-ALL samples expressing each of these transcription factors show a distinctive gene expression signature and as such these transcription factors define distinct molecular subtypes in T-ALL [6]. Chromosomal rearrangements can also lead to large chromosomal deletions and amplifications; to focal gene deletions or amplifications, such as CDKN2A deletion and MYB duplication [7], [8]; and to in-frame fusion genes encoding chimeric proteins with oncogenic properties such as the constitutively active NUP214-ABL1 fusion kinase [9]. In addition, point mutations and small insertions/deletions (INDELs) have also been described leading to oncogenic events, such as mutations activating NOTCH1 that occur in more than 60% of T-ALL cases [10], or mutations in cytokine receptors and tyrosine kinases such as IL7R and JAK3 [11]–[17]. The latter may lead to new opportunities for molecularly tailored therapies with kinase inhibitors [12], [16], [18], [19].
With the advent of next generation sequencing (NGS) technologies, our sequencing capacity has significantly improved in the past five years. It is now possible to apply targeted re-sequencing, exome sequencing (Exome-seq), whole genome sequencing (WGS), whole transcriptome sequencing (RNA-seq) or a combination of these, to investigate individual genomes, especially those related to disease [20]. Also for T-ALL, these NGS approaches have recently proven their value in the discovery of novel driver genes [13], [14], [17], [21]. We previously identified a spectrum of new oncogenic driver genes using Exome-seq on 67 T-ALLs, and described clear differences between pediatric and adult patients [17]. In particular, we identified CNOT3 as a tumor suppressor mutated in 8% of adult T-ALL cases and mutations affecting the ribosomal proteins RPL5 and RPL10 in 10% of pediatric T-ALLs [17]. Similarly, whole genome sequencing of early T-cell precursor ALL cases led to the identification of mutations in several new oncogenes and tumor suppressor genes affecting cytokine signaling, T-cell development and histone-modifying genes [2], [13]. However, the potential of RNA-seq for the discovery of driver genes in T-ALL remains unexplored.
In the present study, we applied paired-end RNA-seq on 49 T-ALL samples (31 patients, 18 cell lines) to gain insights in the transcriptome landscape of T-ALL. First, we show that identification of somatic single nucleotide variants (SNV) and recurrently mutated driver genes is feasible on RNA-seq data, even without matched normal samples (e.g., germlines or remission DNA). We identify STAT5B, H3F3A, and PTK2B as candidate cancer genes in T-ALL. This becomes possible when (1) optimal read mapping and SNV calling procedures are applied; and (2) functional annotation, gene expression, or additional sequencing data from other cohorts is used to prioritize the true driver genes. Next, we optimized gene expression measurements using multiple normalization strategies, and showed that classical gene expression studies (e.g., clustering) are feasible on normalized RNA-seq data. We also detected new fusion genes (SSBP2-FER and TPM3-JAK2) and used gene expression data to determine the consequence of observed chromosomal rearrangements on the over-expression of key driver genes. Finally, we searched for significant alternative transcript events (ATE) but besides one coherent exon-skipping event in SUZ12, we found relatively few candidate ATEs in T-ALL. In conclusion, through a combination of the analysis of gene expression levels, fusion transcripts, SNVs, and INDELs, we could identify known and new driver alterations in T-ALLs and novel potential targets for therapy.
We performed paired-end RNA-seq on 31 T-ALL patients, 18 T-ALL cell lines, and 1 normal thymus sample. We obtained on average ∼110 million reads per sample, leading to an average coverage of ∼88× (Table S1.A). To assess the quality of detecting SNVs from the RNA-seq data, we compared the RNA-seq to Exome-seq data. For 16/18 of the cell lines and for 20/31 patient samples we had exome data available (previously generated [17] or obtained for this study, Table S2). For the exome data analysis, we followed the pipeline of mapping, SNV and somatic mutation detection that we validated previously [17] (using BWA, GATK, SomaticSniper, and Variant Effect Predictor (VEP)) [22]–[25]. For the RNA-seq data we used TopHat2 [26] for mapping, SAMTools [27] for SNV detection, and VEP [25] for variant annotation (Figure 1.A).
By comparing positions that had a coverage of at least 20× in both RNA-seq and Exome-seq, combined with Sanger re-sequencing of a subset of positions, we found that the accuracy of SNV calling in RNA-seq strongly depends on the read mapping, corroborating earlier observations [28], [29] (Figure S1). We found that mapping RNA-seq reads to the genome (as used by TopHat version 1.3.3) is prone to errors when dealing with paralogous genes, as observed by the prediction of false positive SNVs in KIF4A and GLUD1 due to erroneous mapping to KIF4B and GLUD2 (both pseudogenes with no introns) (Figure S1). However, these errors were resolved by mapping to the transcriptome. In the case of the RPMI8402 cell line, 877 SNVs were found by mapping to the genome, while this number was reduced to 283 SNVs when mapping to the transcriptome. Mapping to the transcriptome did not only reduce the number of RNA-seq exclusive calls but also increased the overlap with the Exome-seq calls (Figure 2, Figure S2).
However, transcriptome mapping also has limitations as it relies on current gene and isoform annotation. We observed that a combination of transcriptome and genome mapping provides the best solution. It is important that all reads are mapped twice to the genome, independently of each other; once as entire read and once as split read. This has become possible in TopHat2 by setting the option “read-realign-edit-dist” to zero. Our analysis reveals that this mapping approach results in the best overlap of SNVs compared to exomes (Figure 2, Figure S3). This mapping strategy not only improves the alignment accuracy by preventing misalignment to pseudogenes, but also leads to identification of the most likely isoform structure of a gene by mapping the reads independently both to the transcriptome and to the genome and then selecting the best possible alignment.
Using the optimized mapping and filtering strategy we identified 436,974 SNVs across 49 samples. By using samples for which both the exome and the transcriptome were sequenced several aspects of SNV detection in RNA-seq data can be evaluated, such as sensitivity, specificity, and allelic imbalance. Regarding sensitivity, we found that on average, 32% of the SNVs that are called in Exome-seq were also called by the RNA-seq (Table S3). Similar ratios were observed when comparing validated somatic SNVs from Exome-seq/WGS to RNA-seq SNVs: 36% in a triple negative breast cancer study [30], and 41% in a lymphoma study [31]. We observed that the sensitivity varies considerably between samples, and strongly correlates with the average depth of coverage of the sample (Figure S4). Regarding specificity, we found that the remaining RNA-seq-only and Exome-seq-only SNVs (for positions where both have at least 20× coverage) are found mainly with a low variant allele frequency (VAF) and are therefore likely due to arbitrary VAF and coverage thresholds. For example, on the RPMI8402 and TLE79 samples, many RNA-seq-only SNVs (9/18 and 61/88 respectively) have a VAF below 40%. Regarding allelic imbalance, we found that of all heterozygous Exome SNVs with more than 20× coverage, the majority (2,914/4,043 or 72%) were also heterozygous SNVs in RNA-seq. Of the remaining SNVs, many (988/4,043) are homozygous reference in the RNA-seq (i.e., not detected). A small fraction we can almost certainly attribute to allelic imbalance, namely the 141/4,043 SNVs (3.5%) that are homozygous variant in the RNA-seq, indicating that for those only the variant allele is expressed (or the gene is only expressed in cancer cells that harbor the variant).
Next we asked whether small insertions and deletions (INDELs) can be detected from RNA-seq data. As with the SNVs, we used the Exome-seq data for assessing the quality of our INDEL detection strategy. On average, 47.5% of the INDELs that were detected by RNA-seq were also found in the Exome-seq (unfiltered) INDEL calls. However, only 4% of the Exome-seq INDELs (for which the region containing INDEL is covered by at least 3 reads in RNAseq data) were found back in the RNA-seq calls (Table S3). To investigate this sensitivity issue, we evaluated ten validated INDELs that we previously identified with Exome-Seq [17](Table S4). Three of the ten INDELs were also identified in the RNA-seq data using the default SAMTools parameters (see Materials and Methods). Of the seven missed INDELS, two are found in a gene that is not expressed; another two are clearly present in the RNA-seq data when inspected manually with IGV, but did not reach the default threshold (see Materials and Methods); and the last three are effectively discordant between RNA-seq and Exome-seq, as they show only reads with reference sequence (Figure S5). Re-mapping of the reads with BWA [22] on the transcriptome followed by BLAT [32] on the genome improved the INDEL identification, now revealing the KDM6A INDEL in TLE87 and PTEN INDEL in TLE92, which were previously missed (Figure S6.A–B). It is notable that the combination of TopHat2 (to transcriptome only) and BLAT does not correctly detect these two INDELS (Figure S6.C–D). We conclude that INDEL detection on RNA-seq data is feasible, yet technically challenging and that the fraction of INDELs compared to SNVs is moderate (see also the next Section and Figure 3).
Our next aim was to select candidate driver genes using the collected SNVs and INDELS. To remove germline variants we initially removed all SNPs present in dbSNP [33], 1000genomes [34], the Complete Genomics genomes [35], and those detected in our own exome data from normal samples (39 from our earlier work [17] and 6 from this study). We, however, retained those variants also present in the COSMIC [36] database, since SNP databases are known to contain also some disease-specific SNVs. Some examples of SNVs that are likely driver mutations, but that are also present in polymorphism databases are: JAK3 A572V in R7, and FBXW7 R425C in TUG1. With this filtering, we obtained a final list of 10,403 protein-altering SNVs and 430 protein-altering INDELs, with a median of 63 SNVs and 4 INDELS per sample (Table S1.B). Cell lines harbored significantly more mutations than patient samples (Mann-Whitney test p-value = 1.095E-05), as previously also observed by Exome-seq [17].
As a first approach to identify candidate T-ALL driver genes, we selected all genes that contained a protein-altering mutation in at least two of the 31 patient samples (for recurrence we did not take cell lines into account). This process resulted in the selection of 213 genes (Table S5). We found that this list is strongly enriched for genes related to T-ALL and to cancer in general, with “precursor T-cell lymphoblastic leukemia-lymphoma” as the most highly enriched function (p-value = 1.35E-11 by Ingenuity Pathway Analysis) (Table S6). The list of 213 candidates contained many known T-ALL driver genes (Figure 3), such as NOTCH1, BCL11B, FBXW7, IL7R, JAK1 and JAK3; and it also contained the drivers CNOT3 and RPL10, recently identified in our exome re-sequencing study [17]; and CTCF, which was recently reported to be recurrently mutated in ETP-ALL [13]. In addition, the candidate list contained two established cancer driver genes involved in other cancer types, but not yet reported to be mutated in T-ALL, namely H3F3A and CIC. These genes were reported recently by Vogelstein [37] to be true cancer drivers. We identified two patient samples (TLE76 and TUG6) with H3F3A mutations both on the K28 residue that is a mutational hotspot in glioblastoma [38]. This mutation was confirmed somatic in the TUG6 sample. Sequencing of this hotspot in additional T-ALL samples indicated a low frequency of H3F3A K28 mutation in T-ALL (detected in 3 of 102 cases).
Next we asked if we could identify additional genes in the candidate list that could be linked to T-ALL. We wanted to utilize the genes that are known to be involved in T-ALL as a guide for identifying additional candidates. To this end we used our gene prioritization approach ENDEAVOUR [39], which scores candidate genes based on a set of training genes. It builds a profile based on the training genes (integrating information on protein-protein interactions, genetic interactions, gene expression, text-mining, sequence homology, Gene Ontology, and protein domains) and then prioritizes the candidate genes for their similarity to the derived profile. As training set we used all known drivers, and as test set we used all the 213 candidates with at least two patient mutations (excluding the genes that are in the training set). We reasoned that this would reveal the genes with strong similarity to the known drivers and such genes would be good candidate drivers. We found 45 significantly ranked genes with two interesting genes at the top of the ranking, namely PTK2B and STAT5B that are involved in JAK/STAT signaling (Table S7). Furthermore, the list contained genes for which we had identified single T-ALL cases with a somatic mutation in our previous exome study: ANKRD11, CTCF, DOCK2, H3F3A, and HADHA. We did not select these genes before in our Exome-seq cohort [17] because they were only mutated in one of the 39 samples we analyzed. Now, with the RNA-seq cohort, we thus found additional samples with mutations in these genes.
T-ALL is characterized by the overexpression of transcription factors (TFs), such as TLX1, TLX3, TAL1, and the HOXA family members [6]. Therefore, identifying and analyzing expression perturbations in a T-ALL cohort is highly relevant. To obtain accurate gene expression levels from the mapped RNA-seq reads, we followed the procedure outlined in Figure 1.B, including read aggregation, GC-normalization, length normalization, and between-sample normalization (see Materials and Methods). In addition, we removed a batch effect that was clearly present in the data set using a Generalized Linear Model (GLM, see Materials and Methods) (Figure S7). It is notable that transcript-based expression analysis conducted with cufflinks revealed the same batch effect linked to the origin of the sample, thereby confirming a technical bias in the data set (Figure S7.B, see Materials and Methods).
We next looked at the expression values of TLX1, TLX3, TAL1, and other important TFs in T-ALL. Clustering of TLX1, TLX3, and TAL1 expressing samples confirmed that the correct samples (based on karyotyping and molecular analysis) showed over-expression of the respective TF (Figure 4.A). Indeed, 8 samples that harbored a STIL-TAL1 rearrangement showed high TAL1 expression (Figure 4.D). Note that also other samples with high TAL1 expression were detected. This fits with a previously reported observation of TAL1 over-expression in the absence of a translocation in T-ALL [6], [40].
To assess the accuracy of our expression values obtained after normalization, batch effect removal and clustering, we tested whether previously published gene signatures associated with TAL1, TLX (TLX1 and TLX3) and LYL1 can be detected also in our data set [41]. We used 13 gene signatures obtained by Soulier et al using a microarray study on 92 primary T-ALL samples [41]. Gene set enrichment analysis shows that our TAL1 expressing cases are significantly associated with TAL1 signatures, whereas our TLX over-expressing cases are associated with the TLX signature [7], [8] and the LYL1 cases with the LYL1 signature [10], [11]. This analysis confirms that the obtained expression data represent meaningful values and sample clustering produces gene lists that are biologically meaningful (Figure 4.B).
We next used the gene expression information as a guide to assist in the detection of relevant mutations. We found that the expression profile of PTK2B, a candidate driver identified above by ENDEAVOUR, significantly correlated with the JAK3 expression profile (PTM, with p-value threshold at 1E-05, see Materials and Methods) (Figure 4.C). Indeed, PTK2B was previously implicated in IL-2 mediated signaling and JAK/STAT signaling, and was shown to physically interact with JAK3 [42]. These data warrant further investigation of PTK2B as an important tyrosine kinase in T-ALL case with activated JAK/STAT signaling.
To our knowledge, only very few cancer specific alternative transcript events (ATE) have been described for any cancer type [43]–[45], and no ATE is reported for T-ALL. In contrast to SNVs, INDELS, copy number variations, and fusions, which are all curated and present in large numbers in public cancer mutation databases (e.g., COSMIC [36], CENSUS [46]), we could not find driver ATEs in those databases (although splice sites represent an important class of cancer mutations). If ATEs represent an important, yet underestimated, type of somatic variation in cancer, we would expect at least some of the known cancer driver genes to present a significant ATE. We thus asked whether novel variations could be found in these genes in the form of ATEs. To this end, we applied cufflinks and cuffdiff (see Materials and Methods) and found significant ATEs in 12 of the 47 known driver genes (BCL11B, FLT3, IL7R, LCK, MYB, NKX2-1, SFTA3, RPL10, RUNX1, SETD2, SUZ12, and TAL1) (Table S8). However, when we manually inspected these events in IGV, we found only two interesting cases. One case represents an unambiguous skipping of exon 7 in SUZ12, occurring in several patient samples, but most significant (cuffdiff p-value = 5.10E-05) in the R5 patient sample, and absent in the Thymus (Figure 4.E), and a potential, but less clear, skipping of exon 8 in LCK in three samples (Figure S8). Exon 7 of SUZ12 is a canonical exon (present in all known isoforms) according to RefSeq, Ensembl, and UCSC annotation. The ATE we observe is a heterozygous event with the wild-type junction supported by 90 reads and the novel junction supported by 71 reads. RT-PCR clearly confirmed the exon-skipping event in R5 and to a minor extent in other samples, while being absent in the thymus (Figure 4.F). The functional consequences of these splice variants remain to be determined, but the fact that these variants are both in-frame suggests that these proteins could be functional protein isoforms (Figure S8 and S9). Overall, relatively few significant ATEs are detected, and no obvious ATEs are found with consequences on the protein structure, therefore T-ALL presents robust isoform usage at the current resolution of sequencing and analysis.
Most of the T-ALL cases harbor chromosomal rearrangements that lead to the generation of fusion genes or ectopic expression of genes due to juxtaposition to strong promoters or regulatory sequences. Chromosomal translocations involving the TCR genes are largely underestimated by karyotyping and the TCR partner genes remained unidentified in several cases [4], [47]. On the other hand, a multitude of mechanisms other than translocations could cause ectopic expression of oncogenes [48]. To detect fusion transcripts, we used the defuse algorithm on our entire dataset [49]. Briefly, this method identifies candidate gene fusions by discordant alignments produced by spanning reads (each read in the read pair aligns to a different gene) and by split reads (reads that harbor a fusion boundary). The total number of predicted fusions initially was 1,160 and 1,265 in patient and cell line samples, respectively. Also in normal thymus RNA, 60 fusion transcripts were detected. Next, we implemented additional filters, considering only predictions supported by 8 or more spanning reads and 5 or more split reads. Furthermore, we removed fusions involving ribosomal genes, mitochondrial genes and fusions between adjacent genes, as these could be caused by read-through or trans-splicing [50], [51] (Figure 1.C).
After applying these filters, we obtained an average of 5.5 fusion events per patient sample and 11.1 per cell line (Table S1.C). In total, 397 candidate genes are involved as potential partner in a gene fusion (Table S9). Details on the fusion breakpoints and validation of the novel candidate fusion transcripts are reported in Tables S9 and S12 (see also Materials and Methods: RT-PCR and Sanger Sequencing).
First, to determine the relevance of these predicted fusion transcripts we looked at functional enrichment of these genes. 278 of 397 genes correspond to functionally annotated protein-coding genes according to DAVID functional enrichment [52], [53]. Furthermore, this set is strongly enriched for cancer-related genes, and more specifically for genes involved in Acute Myeloid Leukemia (p-value = 4.48E-10) and T-ALL (p-value = 4.47E-05), including TP53, STAT5B, NOTCH1, IL7R, IKZF1, CDKN2A, MLLT10, ETV6, and ABL1.
Second, we specifically analyzed the 27 in-frame fusions, predicted to encode chimeric proteins (Table S10). This list contained known oncogenic fusion genes, including NUP214-ABL1 (n = 2), MLL-FOXO4 (n = 1), PICALM-MLLT10 (n = 1), ETV6-NCOA2 (n = 1) and SET-NUP214 (n = 1). In addition, we identified 3 novel chimeric transcripts in T-ALL, namely NUP98-PSIP1 (n = 1), TPM3-JAK2 (n = 1) and SSBP2-FER (n = 1) and a novel DDX3X-MLLT10 fusion transcript (n = 1) recently described in a pediatric T-ALL patient [54]. Conventional cytogenetic analysis confirmed the presence of a t(X;10) in the case with the DDX3X-MLLT10 fusion, whereas it failed to detect the chromosomal rearrangements for the TPM3-JAK2, NUP98-PSIP1 and SSBP2-FER fusions, demonstrating the power of RNA-seq to identify cryptic fusion genes and to provide genetic information even in patients with uninformative cytogenetics. Reassuringly, RT-PCR and Sanger sequencing confirmed the presence of these fusion transcripts (Table S12).
The TPM3-JAK2 and SSBP2-FER fusions encode typical tyrosine-kinase fusions that join the tyrosine-kinase domain of JAK2 or FER to the dimerization units of TPM3 or SSBP2, respectively (Figure 5.A). To assess whether the TPM3-JAK2 and SSBP2-FER fusions encode oncogenic proteins, we tested their transforming properties in the IL-3–dependent Ba/F3 cell line [55]. Both TPM3-JAK2 and SSBP2-FER transformed Ba/F3 cells to IL-3–independent growth, with even faster kinetics than the JAK1 A634D mutant, which is a known transforming kinase [18] (Figure 5.B). Western blot analysis confirmed the constitutive auto-phosphorylation of the JAK2 and FER fusion proteins, as well as the downstream STAT proteins (Figure 5.C). Ba/F3 cells transformed by the TPM3-JAK2 fusion were sensitive to a JAK kinase inhibitor, documenting the potential application of JAK2 kinase inhibitors for the treatment of T-ALL cases with JAK2 fusion genes. No specific FER inhibitors were available to test their activity. Both TPM3-JAK2 and SSBP2-FER fusion were screened in 50 additional T-ALL samples, but no additional case with these fusions was found.
Third, we also analyzed the identified fusions that did not seem to encode chimeric proteins (out-of-frame fusions), and which were the majority of fusions detected in T-ALL. These fusion events can be used as surrogate markers for the identification of chromosomal rearrangements, providing accurate information on the precise chromosomal breakpoints. In combination with the gene expression data obtained by RNA-seq, these data can identify genes that are located close to such potential breakpoints and for which the expression is significantly up- or down-regulated. As expected, we identified the STIL-TAL1 fusion in several T-ALL cases (n = 8). We also identified and validated 6 fusion events involving TCR genes. In 4 of these cases, the TCR gene was found to be fused to the potential oncogene (NOTCH1, IL7R, PLAG1, and TLX1). In the two other cases (R4, TLE90), the TCR gene was fused to RIC3 or SFTA3, resulting in the ectopic expression of LMO1 and NKX2-1, respectively, as indicated by RNA-seq gene expression data (Figure 5.D and E). Similarly, we could better characterize the t(10;14) in ALL-SIL cell line that expresses TLX1 at high level.
In addition to the TCR gene rearrangements, also other fusions were associated with overexpression. We detected out-of-frame fusion transcripts that joined exon 4 of CDK6 to exon 2 of HOXA11-AS and exon 5 of CDK6 to sequences downstream of EVX1. In the same patient we also detected a fusion joining DPY19L1 on chromosome 7p14 to HOXA11 on chromosome 7p15. The gene expression analysis documented high expression of genes of the HOXA cluster (i.e. HOXA9, -A5, -A13, -A10, -A11). Moreover, other fusions identified in this study, such as CLINT1-MEF2C, HNRP-ZNF219 (n = 2), ZEB1-BMI1 and AHI1-MYB (n = 2) were also associated with transcriptional activation of MEF2C, ZNF219, BMI1 and MYB as confirmed by the expression data (Table S9 and S12, and Figure S10). Increased MYB expression in T-ALL was previously observed as a consequence of MYB duplication (including in the BE-13 cell line), which may also explain the detected AHI1-MYB fusion [8], [56].
Finally, we also found out-of-frame fusion transcripts leading to the potential inactivation of tumor suppressor genes, such as TP53-TBC1D3F (ALLSIL cell line), PTEN-RNLS (LOUCY cell line), IKZF1-ABCA13 and CDKN2A-miR31HG (R6 case), indicating a third class of fusion events (Figure S10). FISH analysis performed in the R6 case confirmed the p15/p16 deletion. As the genes are in close proximity, the IKZF1-ABCA13 was presumably generated by deletion although no material was available to confirm this hypothesis.
The landscape of genomic variation underlying T-ALL has recently been investigated by sequencing candidate genes [14], [21], whole exomes [17] and whole genomes [13]. The results of these studies, combined with a large body of gene-by-gene evidence collected over the last decade, provide a growing comprehension of the T-ALL genome. The T-ALL genome is mainly characterized by the over-expression of TF, such as TLX1/3 and TAL1, in combination with gain-of-function NOTCH1 mutations, and with additional mutations in chromatin modifiers, cellular signaling factors such as those involved in the JAK-STAT signaling pathway [57], tumor suppressor genes (TP53, PTEN, WT1), or in other genes such as ribosomal genes [17]. Since the majority of observed mutations are point mutations and gene fusions (much more than copy number variations [13]) we reasoned that RNA-seq would be effective to identify many of these mutations, certainly those associated with (over-)expressed oncogenes. Indeed, exome sequencing allows identifying point mutations but not gene fusions; and low coverage whole-genome sequencing allows identifying structural variation (gene fusions) but not point mutations. In this study we present RNA-seq analyses on a heterogeneous group of 31 T-ALL samples and 18 T-ALL cell-lines and demonstrate that RNA-Seq is indeed a very powerful approach to detect gene mutations and fusions as well as expression perturbations.
Our first challenge with regards to the accurate identification of point mutations was finding the optimal analysis pipeline – from read mapping to SNV calling and filtering – to avoid too many false positive SNVs. By exploiting whole-exome sequencing data for a subset of our samples we obtained a recovery ratio of 32% when compared to the exome derived SNVs; a ratio that is comparable with previous RNA-seq studies [30], [31]. However, this concordance could only be achieved by using the optimal read mapping methods and parameters: (1) use of a recent version of TopHat2 (v. 2.0.5. or higher) and (2) forcing this aligner to map all reads twice to the genome (once directly and once using split reads) and once to the transcriptome. Indeed, the computational task of sequence read mapping is more challenging for RNA-seq data because a large fraction of the obtained reads need to be split to allow reads that overlap exon-exon boundaries in the cDNA to be mapped to the genome. In this way, RNA-seq is more prone to the identification of false SNVs due to the erroneous mapping of reads, for example to highly similar non-spliced pseudogenes. For example, in the RPMI8402 cell line, 603 RNA-seq exclusive SNVs were found with the genome mapping strategy, while only 35 when using combined mapping strategy.
Among the previously published large scale RNA-seq cancer studies, only a handful performed variant calling on the RNA-seq data [30], [31], [58], [59]. A combined mapping strategy was followed in all cases either by mapping the reads to a customized genome reference file (by the addition of exon junction segments) or mapping the reads twice (once to the genome and once to the transcriptome). Variant calling pipelines also showed diversity: Morin et al and Shah et al used SNVMix [60] for variant calling, while Seo et al and Berger et al implemented filters based on alignment on the non-reference bases. To our knowledge there is no extensive benchmarking study evaluating aligners and variant callers for RNA-seq data, but a review paper by Quinn et al compared the performance of two variant callers (GATK [23] and SAMTools [27]) with the optional duplicate removal step (pre and post alignment), and concluded that post-alignment duplicate removal and variant calling with SAMTools achieved the best performance in terms of sensitivity and specificity [61]. We have also followed the same strategy in our study and we could achieve a comparable recovery ratio of 32% when compared to Exome-seq calls.
A second challenge in identifying point mutations was the prioritization of candidate driver mutations versus passenger mutations. Due to the lack of matched germline RNA for each patient as control, we used a large cohort of local normal exome datasets, in combination with the commonly used variants from dbSNP and 1000genomes, to distinguish SNPs from candidate somatic mutations. This strategy has been successfully used before on transcriptome sequencing studies [62]. Identifying candidate cancer genes by gene mutation frequency is a frequently used approach [13], [30], [58]. Remarkably, by simply selecting all genes having a candidate somatic mutation in at least two samples (213 genes in total), we already achieved a highly significant enrichment for T-ALL related genes, such as NOTCH1, BCL11B, FBXW7, DNM2, JAK3, JAK1, and IL7R. Among the remaining candidates we searched for additional evidence and we propose seven additional candidate drivers because they are either “functionally similar” to the previously known drivers, or because they were mutated somatically at least once in another T-ALL cohort [17], or both. Six of these genes, namely CIC, H3F3A, PTK2B, STAT5B, ANKRD1 and HADHA have already been implicated in other cancers [63]–[70] while DOCK2 has no association with cancer yet.
We found a remarkable clustering of molecular functions among the identified T-ALL driver genes, with enrichment for functions related to the regulation of gene expression. TFs and their co-factors play a central role in transcriptional regulation and these proteins are often mutated in T-ALL. Also, many of these play important roles in the normal T-cell developmental gene regulatory network [71], such as NOTCH1, TLX1, TLX3, TAL1, BCL11B, CTCF, FOXO4, MYB, and others. Upstream of these activated TFs, multiple kinases and other signaling factors control their activity, and these regulators are also often mutated in T-ALL (for example, JAK1, JAK3, and IL7R). Finally, chromatin modifiers and methylation factors are recurrently mutated and these can have both generally pervasive but also specific effects on the expression of oncogenes, such as MYC [72]. When multiple driver mutations are serially acquired, their combined effect will result in oncogenic expression profiles, whereby genes supporting a growth advantage increase and genes negatively affecting growth advantage (e.g., apoptosis, senescence) decrease in expression. It will be an interesting future challenge to draw the connections between the observed DNA mutations, the oncogenic program, and the final gene expression changes that we and others observe in T-ALL samples. Finally, it is likely that non-coding mutations, such as those in promoters, enhancers, microRNAs, and lncRNAs, add to the cancer-related gene regulatory network changes underlying leukemogenesis.
As mentioned above, only mutations in genes that are actively transcribed are detected, and this likely adds to the specificity of driver gene detection. On the other hand, this could also present a limitation of RNA-seq, because loss-of-function mutations in tumor suppressor genes may lead to nonsense-mediated decay, and as consequence low sequence coverage to call mutations. Based on our data however, this is not the case because we could detect PHF6 mutations in up to 4/31 patient cases (13%), where exome sequencing identified PHF6 mutations in 9/67 cases (13%) [17] and Zhang et al identified PHF6 mutations in 24/106 cases by means of whole genome sequencing and capillary sequencing [13].
Interestingly, the gene expression information used above (i.e., read coverage to identify point mutations) can be further exploited at the quantitative level, similar to gene expression studies performed with microarray technology over the last 15 years. As many leukemia driver genes are characterized by changes in gene expression, this level of information is invaluable, both in research and diagnostic settings. We investigated how accurate gene expression levels can be achieved and we found that multiple normalization steps are required, both within-sample (gene length and gene GC content) and across samples (library size), and that batch effects can be effectively removed using a previously published Generalized Linear Model (GLM) [73]. The gene expression levels of the known drivers (e.g., TLX1/3, TAL1, NOTCH1) are highly representative as driving events and as subtype identifiers. However, to discover driver genes de novo, using only gene expression values, is to our opinion not feasible (data not shown). Alternatively, we attempted to select candidate drivers based on the expression similarity (i.e., co-expression across the cohort) with known drivers. This led to the identification of PTK2B, whose expression strongly correlated with JAK3 and which is known to be implicated in JAK-STAT signaling. The next level of gene expression analysis would preferably be a network-level analysis [74], but this requires a larger sample cohort.
Another kind of information that can be extracted from RNA-seq data, besides point mutations and gene expression changes, are alternative transcript events (ATE) and gene fusions [75]. We found only few significant ATEs but could confirm two exon-skipping events in the known T-ALL oncogenes SUZ12 and LCK. More importantly, we identified (i) known and novel in-frame fusions encoding chimeric proteins, (ii) TCR gene arrangements resulting in over-expression of oncogenes, and (iii) fusions not involving TCR genes but also resulting in over-expression of oncogenic transcription factors. The most recurrent fusion event, observed in 8/31 samples, was the STIL-TAL1 fusion resulting in the ectopic over-expression of the TAL1 gene. We also identified novel gene fusions, including two in-frame fusions, TPM3-JAK2 and SSBP2-FER, producing chimeric oncoproteins; and other fusions resulting in the ectopic expression of transcription factors such as PLAG1, MEF2C, ZNF219, and BMI1. The ectopic expression of these genes is associated with a fusion event and with changed expression, which can both be detected by RNA-seq, making this technology extremely powerful to accurately detect such oncogenic events. Each of these novel events appears to be rare in T-ALL, as we identified at most 2 cases of each fusion. However the evidence of transcriptional activation of the partner genes suggests that further studies are required to establish the recurrence of these lesions and their functional meaning. It is notable that the normal thymus sample also shows four fusion events. However, as these genes are located in close proximity to each other, they may represent unannotated isoforms in the human transcriptome. Despite RNA-seq has offered a deeper insight into the complexity of the transcriptome, several studies have highlighted that the catalogue of all expressed transcripts is still far from complete and it is increasing the number of novel splice junctions connecting novel exon, non-exon regions, or linking independent transcripts [76].
Today, high-quality catalogues of driver genes across cancer types are available, and this influences how and why cancer genomes need to be sequenced. For T-ALL, and for many common cancer types, the objectives of sequencing are shifting from the discovery of cancer genes, to a diagnostic setting in which a list of driver events are a priori known. Targeted re-sequencing provides an interesting route, although this poses technical challenges of amplification or capturing, and perhaps more importantly, is focused on a limited number of genes and on one particular mutation type, namely point mutations and small insertions/deletions. We have shown in this study that, with a list of interesting cancer drivers at hand, and with other datasets being available (e.g., rare variants from local exome studies, 1000 genomes, TCGA data, etc), RNA-sequencing of only the cancer sample provides a technically straightforward approach and delivers at once the point mutations, gene fusions and gene expression changes across the entire transcriptome. And as a corollary, the data analysis strategies provided here would be beneficial for any cancer type as long as a body of knowledge is available for selecting and prioritizing candidate events.
Diagnostic total RNAs from 31 T-ALL patients (20 adults and 11 children) were collected at various institutions. All patients have given their informed consent and all samples were obtained according to the guidelines of the local ethical committees. This study was approved by the ethical committee of the University Hospital Leuven. Diagnosis of T-ALL was based on morphology, cytochemistry and immunophenotyping according to the World Health Organization and European Group for the Immunological Characterization of Leukemia criteria [77]. The clinical and hematologic features of the 31 patients at the diagnosis are summarized in Table S11 Total RNAs from 18 T-ALL cell lines (DSMZ, Braunschweig, Germany) were extracted using QIAGEN RNeasy Mini Kit. A pool of total RNAs from 5 normal human thymuses was purchased from Capital Biosciences.
All the RNA samples showed a high quality RNA Integrative Number (RIN>/ = 7) score on the Bioanalyzer (Agilent Technologies).
Fifty additional RNA samples were used for TPM3-JAK2 and SSBP2-FER analysis.
Genomic DNA from of 71 adult T-ALL patients were used for H3F3A K28 screening.
Next generation sequencing libraries were constructed from 500 ng of total RNA using the Truseq RNA sample prep kit (Illumina). RNA-seq libraries were subjected to 2×100 bp paired-end sequencing on a HiSeq2000 instrument (Illumina). Sequence reads were processed to identify gene fusion transcripts, single nucleotide variants (SNVs) and gene expression levels. For the read mapping, variant calling and transcriptome assembly, we used the infrastructure of the VSC - Flemish Supercomputer Center, funded by the Hercules foundation and the Flemish Government - department EWI.
Fusion transcript discovery was performed using defuse v.0.5.0 [49] with default parameters. The resulting list was filtered as described in [78]. Briefly, fusion transcripts with less than 8 spanning reads and less than 5 split reads were filtered out. In addition, we removed fusion events observed in adjacent genes and fusion events involving ribosomal genes (ribosomal genes were downloaded from Biomart on 24-05-2011 using GO:0005840) and the genes located on chrM. Fusion events were annotated using Pegasus (http://sourceforge.net/projects/pegasus-fus/).
For Gene Expression Profiling analysis, reads were mapped to the human reference genome (assembly GRCh37.68) using TopHat v.2.0.5 [26] with the following parameters: transcriptome-only. Read counts per gene were obtained with the HTSeq package (htseq-count) (http://www-huber.embl.de/users/anders/HTSeq). The aggregated read counts were normalized with EDASeq v1.4.0 [79] and generalized linear model was fitted with edgeR v3.0.4 [73] to remove batch effect originating from the sample collection center. The pathways, and upstream regulators were generated through the use of IPA (Ingenuity Systems, www.ingenuity.com). Expression neighbors were detected with Pavlidis Template Matching (PTM) analysis [80]. Transcript based gene expression values were obtained using cufflinks suite [81], [82]. Transcript assembly was performed with cufflinks v2.1.1 with –g option using assembly GRCh37.68.
Gene set enrichment analysis (GSEA) was performed for TAL1, TLX and LYL1 clusters [83]. We have obtained whole genome rankings for TAL1, TLX (TLX1 and TLX3), and LYL1 simply by calculating the log fold changes between samples expressing the respective gene versus the remaining samples. The gene signatures from Soulier et al were obtained from Table S2 [41].
Tumor patient samples and Thymus RNA-Seq samples were mapped to the Ensembl GRCh37.68 reference genome by Tophat2 [26]. Mapped reads were realigned, and transcript abundance were estimated using cufflinks v2.1.1 [81], [82]. Transcript assembly was reconstructed using the cuffmerge program of the cufflinks package from the realigned transfrags for each of patient RNA-seq samples, merged with the Thymus sample (control), followed by differential expression analysis performed using cuffdiff program. The significant events were extracted from the list of differentially expressed genes, isoforms, primary transcripts and coding sequence and assessed manually with IGV [84]. The mRNA sequences for novel SUZ12 and LCK transcripts were extracted using gffread command of cufflinks, and these sequences were translated using the translate tool of the ExPASy Bioinformatics Resource Portal [85]. The longest ORF sequence was used to verify the domain architecture of the resulting proteins using SMART [86], [87].
The sequence reads were mapped to the human reference genome (assembly GRCh37.68) using TopHat2 setting the option “read-realign-edit-dist” to zero [26]. Duplicate removal process was performed on the aligned reads using Picard v1.74 (http://picard.sourceforge.net). Then SAMTools package v0.1.19+ (pulled from the git repository on 29-07-2013) [27] was used for single nucleotide variant (SNV) and small insertion and deletion (INDEL) detection with minimum mapping quality threshold of 1 and minimum base quality threshold of 13 (-q 1 -Q 13) [27]. The variant calling was done on the coding regions of the genome only (extracted from the transcript definitions in the assembly GRCh37.68). The variant predictions that were supported exclusively by variants located in the beginning or the end of the read were filtered out. Then the SNVs were further filtered with depth of coverage threshold of 20 and minimum variant allele frequency threshold of 0.20. INDELs predictions were filtered with the SAMTools recommended parameters (varFilter -10 -20 -30 -40 -a4 -G90 -S30) and additionally INDELs located in homopolymer stretches longer than 5 bps were filtered. The high quality list of variants was filtered for common population variants using the calls from 1000 genomes, dbSNP, HapMap, and Complete Genomics. Note that, the list of common population variants was cleaned from oncogenic variants using COSMIC listed variants (v66) [36]. Moreover, the variants located in the repeat regions (simple repeat and RepeatMasker) were filtered out. Finally, the variants that are observed in the exomes of remission (i.e. healthy) samples (including the previously published 39 exome remissions [17] and the 6 additional exome remission sequenced) and the variants that are observed in Thymus were also filtered out. The final filtered list of variants was annotated with the Variant Effect Predictor version 2.7 [25] and the protein-altering mutations were selected. The following terms were used for selecting protein-altering SNVs: splice-donor-variant, splice-acceptor-variant, stop-gained, initiator-codon-variant, missense-variant, splice-region-variant. The same terms were used for filtering the INDELs with the addition of the following terms: inframe-insertion, inframe-deletion, frameshift-variant.
The list of candidate genes was created by intersecting the genes with recurrent mutations (SNVs and INDELs) in RNA-seq patient cohort with the somatic mutations in Exome-seq patient cohort [17]. The list of genes that have recurrent mutations in the RNA-seq patient cohort was filtered for mutations observed in chrM.
The list of T-ALL driver genes were curated using the Census database [46] and T-ALL literature and includes the following genes: TLX1, TLX3, PHF6, MYC, BCL11B, HOXA1, SET, MLL, MLLT1, PICALM, MLLT10, WT1, MYB, LEF1, LMO2, LMO1, TAL1, NUP98, NOTCH1, FBXW7, CCND2, PTEN, PTPN2, NF1, FLT3, JAK1, NRAS, LCK, NUP214, ABL1, EZH2, SETD2, SUZ12, JAK3, MEF2C, NKX2-1, NKX2-2, CDKN2A, CDKN2B, RUNX1, KRAS, EED, ETV6, RPL10, DNM2, IL7R, CNOT3.
Somatic mutations from the exome pairs were obtained as described previously [17]. Briefly, the alignment was performed with BWA [22] and post-alignment modifications (duplicate removal, realignment around INDELs and calibration of the quality scores) were done with the Genome Analysis Toolkit (GATK) [23]. Variant calling was performed with GATK using Variant Quality Score Recalibration (VQSR) method. Putative somatic variants were identified by subtracting the mutations observed in the primary samples from the mutations observed in the corresponding remission samples. SomaticSniper score above 70 was used to identify the final list of somatic events [24].
Variant allele frequency (VAF) plots were drawn for the positions that are novel SNVs in either of the RNA-seq or Exome-seq data and covered by at least 20 reads in both datasets.
Novel candidate fusion transcripts were validated by Reverse-Transcription Polymerase-Chain-Reaction (RT-PCR) and Sanger sequencing. In all cases Thymus was used as negative control. cDNA synthesis and PCR amplification were performed using standard protocols that come with Superscript III Reverse Transcriptase (Invitrogen) and GoTaq (Promega). PCR primers were designed to amplify 200–400 bp fragments containing the fusion boundary detected by RNA-seq. The PCR products were analyzed using a QIAxcel automated multicapillary electrophoresis system (QIAGEN). The results were processed and visualized using the BioCalculator Software. PCR products were analyzed by Sanger Sequencing. In cases where multiple PCR products were detected, we performed conventional agarose gel electrophoresis and extraction of specific bands using the gel DNA Recovery Kit (Zymo). Analysis of Sanger chromatograms was performed using CLC Main Workbench 6 (CLC Bio, Aarhus, Denmark). Fusion detection was performed using NCBI Blast alignment. Analysis of the breakpoint was done on the longest isoform reported on the Ensembl genome browser. The tested fusions predictions and the primers used for validations are reported in Table S12.
Validation of SUZ12 exon skipping was performed by RT-PCR, gel extraction and sequencing of the two PCR products (Figure 4.I). The following primers were used for RT-PCR and Sanger sequencing: SUZ12_EX1F (CTGACCACGAGCTTTTCCTC) and SUZ12_EX9R (CCATTTCCTGCATGGCTACT).
The plasmid TPM3-JAK2 pMSCV-GFP was obtained as follows: a DNA fragment containing TPM3 coding region till exon 7 was PCR amplified from thymus cDNA using Phusion High Fidelity DNA Polymerase (Finzyme) and primers containing BglII and XhoI restriction sites. Primers containing XhoI and EcoRI restriction sites were used to amplify JAK2 coding exons 17–25. PCR products were cloned into the BglII and EcoRI sites of the pMSCV-GFP vector after subcloning into the pJET1.2 CloneJET vector (Fermentas). As a final control, plasmid DNA was sequenced by Sanger sequencing.
SSBP2-FER fusion was synthesized by Genscript (Piscataway, NJ, USA) and cloned into pMSCV-GFP by using the unique restriction sites XhoI and EcoRI. The plasmid contained the full length SSBP2-FER fusion including the first 16 coding exons of SSBP2 and the coding exons 14–20 of FER.
Viral supernatants were produced in HEK293T cells using an EcoPack packaging plasmid and TurboFect transfection reagent (Fermentas). Viruses were harvested 48 hours after transfection followed by transduction of the Ba/F3 murine pro-B cells (DSMZ, Braunschweig, Germany) as described previously [88].
Ba/F3 cells were washed twice in PBS to remove all traces of cytokines and were seeded in triplicate in 24-well dishes at 100 000 cells/mL. GFP expression and cell number were measure on a Guava flow cytometer (Millipore). All experiments were terminated at day 8 after cytokine removal and cell lines showing no sign of cell proliferation at that timepoint were declared to be non-transforming.
Total cell lysates were analyzed by standard electrophoresis and western blotting procedures using the following antibodies: anti-phospho-JAK1 (Tyr1022/1023), anti-phospho-STAT1, anti–phospho-STAT5 (Tyr694), anti–phospho-STAT3 (Tyr705), anti-phosphoERK1-2, anti-phospho-SRC families (Tyr416) (from Cell Signaling Technology).
TPM3-JAK2 and SSBP2-FER IL3-independent Ba/F3 cells were seeded in triplicate in 96-well plates at a density of 0.03×106 cells in the presence of JAK inhibitor Ruxolitinib (INCB018424, Azon Medchem). Cell proliferation and viability were assessed on a Guava flow cytometer after 24 hours to determine the IC50, the concentration of inhibitor that gave a 50% inhibition.
Genome data has been deposited at the European Genome-phenome Archive (EGA, http://www.ebi.ac.uk/ega/) which is hosted at the EBI, under accession number EGAS00001000536.
|
10.1371/journal.pcbi.1000926 | Minimal Mesoscale Model for Protein-Mediated Vesiculation in Clathrin-Dependent Endocytosis | In eukaryotic cells, the internalization of extracellular cargo via the endocytic machinery is an important regulatory process required for many essential cellular functions. The role of cooperative protein-protein and protein-membrane interactions in the ubiquitous endocytic pathway in mammalian cells, namely the clathrin-dependent endocytosis, remains unresolved. We employ the Helfrich membrane Hamiltonian together with surface evolution methodology to address how the shapes and energetics of vesicular-bud formation in a planar membrane are stabilized by presence of the clathrin-coat assembly. Our results identify a unique dual role for the tubulating protein epsin: multiple epsins localized spatially and orientationally collectively play the role of a curvature inducing capsid; in addition, epsin serves the role of an adapter in binding the clathrin coat to the membrane. Our results also suggest an important role for the clathrin lattice, namely in the spatial- and orientational-templating of epsins. We suggest that there exists a critical size of the coat above which a vesicular bud with a constricted neck resembling a mature vesicle is stabilized. Based on the observed strong dependence of the vesicle diameter on the bending rigidity, we suggest that the variability in bending stiffness due to variations in membrane composition with cell type can explain the experimentally observed variability on the size of clathrin-coated vesicles, which typically range 50–100 nm. Our model also provides estimates for the number of epsins involved in stabilizing a coated vesicle, and without any direct fitting reproduces the experimentally observed shapes of vesicular intermediates as well as their probability distributions quantitatively, in wildtype as well as CLAP IgG injected neuronal cell experiments. We have presented a minimal mesoscale model which quantitatively explains several experimental observations on the process of vesicle nucleation induced by the clathrin-coated assembly prior to vesicle scission in clathrin dependent endocytosis.
| Cell membranes and membrane-based organelles actively mediate several intracellular signaling and trafficking decisions. A growing number of applications rely on cooperative interactions between molecular assemblies and membranes. Yet, the studies of membrane-based and membrane-mediated signaling are not considered core aspects of systems biology. While a coherent and complete description of cell membrane-mediated signaling is not always possible by experimental methods, multiscale modeling and simulation approaches can provide valuable insights at microscopic and mesoscopic scales. Here, we present a quantitative model for describing how cell-membrane topologies are actively mediated and manipulated by intracellular protein assemblies. Specifically, the model describes a crucial step in the intracellular endocytic trafficking mechanisms, i.e., active transport mechanisms mediated through budding of the cell membrane orchestrated by protein-interaction networks. The proposed theory and modeling approach is expected to create avenues for many novel applications in systems biology, pharmacology, and nanobiotechnology. The particular application to endocytosis explored here can help discern pathological cellular trafficking fates of receptors implicated in a variety of biomedical conditions such as cancer, as well as impact the technology of targeted drug delivery in nanomedicine.
| The cellular process of endocytosis is important in the biological regulation of trafficking in cells, as well as impacts the technology of targeted drug delivery in nanomedicine [1], [2], [3], [4], [5], [6], [7]. In eukaryotic cells, the internalization of extracellular cargo via the endocytic machinery is an important regulatory process required for many essential cellular functions, including nutrient uptake and cell-cell communication. Several experimental [8] as well as theoretical [9], [10], [11] treatments have addressed mechanisms in endocytosis, yet the role of cooperative protein-protein and protein-membrane interactions in the ubiquitous endocytic pathway in mammalian cells, namely clathrin-dependent endocytosis (CDE), remains unresolved. A sequence of molecular events in CDE is responsible for the recruitment of adaptor protein 2 (AP-2), accessory proteins such as epsin, AP180, Eps15, Dynamin, etc., and the scaffolding protein clathrin to the plasma membrane [8]. The accessory proteins such as epsin are implicated in membrane bending [12]. Polymerization of clathrin triskelia in the presence of adaptor proteins such as AP-2 results in the clathrin coat formation, and tubulating proteins such as epsin interact with both the clathrin coat as well as the bilayer [13] to stabilize a clathrin-coated budding vesicle. The involvement of dynamin is believed to be in the vesicle scission step [8]. Even though actin is believed to play an important role in the endocytosis process in S. cerevisiae (yeast), in mammalian cells, actin repression, at best, has a small effect on endocytosis [14].
We focus on the energetic stabilization of a budding vesicle induced by the clathrin-coat assembly. Recent work [15] demonstrates that the membrane invagination only begins in the presence of a growing clathrin coat [16]. Experiments performed by down-regulating AP-2 expression [17], [18] as well as those involving the inhibition of epsin [19] either significantly decrease the number of clathrin-coated pits or alter the distribution of coated-intermediates involved in the vesicle-bud formation. Although the CDE in mammalian cells remains a complex regulatory process, we believe that a critical and self-consistent set of experiments is now emerging which warrants the formulation of physically-based models to quantitatively describe the bioenergetics of protein-induced vesicle formation in CDE [20].
Even though models directly addressing CDE in the experimental (cellular) context have not been proposed, Oster et al. have addressed yeast endocytosis driven by actin [9], [21]. Moreover, Kohyama et al. [22] have shown that model two component membranes bud in response to induced spontaneous curvature or the line tension between the two components of the membrane and Frese et al. have investigated the effect of protein shape and crowding on domain formation and curvature in biological membranes [23]. A recent mini-review examining the current experimental trend by Lundmark and Carlsson on driving membrane curvature in clathrin-dependent and clathrin-independent endocytosis is also available [24]. We formulate a minimal model, by restricting our focus to three proteins in the clathrin-coat assembly (Fig. 1): clathrin, epsin and AP-2, and their role in the stabilization of a budding vesicle on the cell membrane. Mammalian cells have a diverse set of proteins which often serve as surrogates and participate in compensatory mechanisms. In this regard, our choice for the ingredients for the minimal model represents roles for the scaffolding proteins (clathrin), curvature inducing proteins (epsin) and the adaptor proteins (AP-2). Recent experiments [15], [25] have reported characteristics of nucleation and growth of clathrin coat: the initiation was observed to occur randomly, but only within subdomains devoid of cytoskeletal elements. In BSC1 cell lines, such domains appear to be 400 nm in diameter surrounded by a rim of a 200 nm “dead zone”. Notably, the nucleation of clathrin coats was observed only in the 400 nm region [25] with the following salient properties: (a) in the growth phase, the addition of clathrin proceeds at a steady rate of about one triskelion every 2 s, (6s-old coats have 10–20 clathrins). (b) Two fates are possible for a growing coat; they either transform into a vesicle (in 32 s the structure resembles a coated vesicle, 50–100 nm in diameter depending on cell type), or they abort containing about 10–40 triskelia, which suggests that the coat sizes are bounded. While we do not consider the process of nucleation and growth of clathrin, based on the above observations, we study the process of one maturing vesicle in the presence of an assembled clathrin coat of a finite size in a membrane patch free of cytoskeletal elements and subject to a pinned boundary condition at the patch boundary. For our model cell membrane patch not fortified by cytoskeleton, we employ a typical value of bending rigidity of our κ = 20kBT derived from literature [26], [27]; (we also explore the effect of varying κ). In this respect, we describe a mean-field model which characterizes the membrane patch as a homogeneous phase with effective (bulk-like) properties. Our model is also mean-field in the sense that it applies to just one vesicular intermediate and the effect of neighboring coats is not included. As noted earlier, our model does not account for the mechanism of clathrin coat nucleation or that of vesicle scission.
Clathrin triskelia and AP-2 (in a ratio of 1∶1) polymerize to form a coat [28] and the stabilizing interactions in the clathrin coat assembly can be quantified using the free energy of the polymerization process. Based on in vitro equilibrium data of clathrin cage formation, Nossal [29] estimated the energetics of a fully-closed clathrin/AP-2 basket relative to a dissolved coat to be ≈−20 kBT. The inclusion of epsin in the clathrin-coat accounts for −23 kBT of energy per bound epsin: the ENTH domain of epsin binds to the PtdIns(4,5)P2 (or PIP2) lipid head groups on the membrane with a binding energy of −14 kBT per bound epsin [12] and the CLAP domain of epsin interacts with clathrin/AP-2 with an energy of −9 kBT [30]. The ENTH interactions with the membrane require the presence of PIP2, which constitutes about 1% of the total phospholipids on the cell membrane [31]. To produce a coated vesicle d = 50 nm diameter, (based on the empirical scaling relationship, the number of triskelia involved ∼0.031d7/4 is 29 [29]), the area of the clathrin coat required is πd2 = 7850 nm2. Considering the area per lipid head-group to be 0.65 nm2, the number of PIP2 molecules in the membrane spanning the area of the coat is 1% of (7850/0.65) = 185. Hence, we note that the ratio of ENTH binding sites (which correspond to the PIP2 on membrane) to the CLAP binding sites (which correspond to the triskelia) is 185/29≈6, and hence as the clathrin coat grows, we expect sufficient number of the corresponding PIP2 binding sites to be present for the ENTH domain of epsin to bind. For this reason, we are justified in not explicitly considering PIP2 as a necessary/limiting species in our minimal model.
Field-theoretic approaches are popular for studying energetic and entropic contributions in continuum field-based mesoscale models [32], [33] and several successful applications of such mesoscale models for gaining mechanistic insight into cell-membrane mediated processes are available [3], [9], [21], [34], [35], [36]. Here, to model membrane response in CDE, we solve the membrane equations in a curvilinear manifold by assuming an underlying axis-symmetry using the surface evolution formalism outlined by Seifert et al. [37]. We derive the equations governing membrane shapes of minimum energy under imposed curvature fields assuming that curvature fields are additive and that protein insertion does not cause spatial heterogeneities in physical properties of membrane such as bending rigidity and interfacial frame tension. Parameterizing the membrane shape by the angle , where s is the arc-length along the contour, we obtain and , where prime indicates the derivative with respect to arc-length s, (Fig. 2). As described by Safran [38], for topologically invariant membrane shape transformations, the contribution of the Gaussian curvature term to the membrane deformation energy is a constant. Hence, we describe the membrane energy, E using the Helfrich formulation [39]. By including only one of the two principal curvatures, namely the mean curvature:(1)
Here, H is the mean curvature of the membrane, H0 is the imposed (or intrinsic) curvature of the membrane due to curvature-inducing proteins and is a function of arc-length s, is the membrane interfacial frame tension and A is the total membrane area. We express curvature H and the area element dA in terms of . Minimization of this energy functional with respect to leads to (see Text S1):(2)
Here, is a Lagrange multiplier introduced to satisfy the constraint (which defines R). We also impose the boundary condition at R = R0 (or at s = s1) corresponding to the pinning of the membrane by the cytoskeleton at the boundary of the membrane patch. In addition, due to the axis-symmetry, at R = 0, . Since the total arc-length s1 is not known a priori, one additional closure equation is specified, (see Text S1): . We solve the above system of boundary valued differential equations numerically by the shooting and marching technique [40], (see Text S1), yielding membrane profiles for a specified spontaneous curvature function, and pinned at R = R0; in this work, we employ R0 = 500 nm. We also compute the curvature deformation energy of the membrane defined by:(3)
We present our results for the case when interfacial frame tension σ is zero. Results obtained for non-zero σ (not shown) are found to be similar to the σ = 0 case. We also note that in prior work, we showed that the entropic term |TΔS| at T = 300K is small, i.e. ∼5% of the membrane bending energy for κ = 20 kBT [41]. This result justifies the basis for neglecting thermal fluctuations (such an assumption was also employed by Oster et al. for their model for endocytosis in yeast [9]) and is valid except in cases where the vesicle neck region becomes narrow (i.e. same order of magnitude as the bilayer thickness). The situation of a narrow vesicle neck is very pertinent to vesicle scission, where even the continuum treatment of the membrane is subject to approximations and a molecular treatment is necessary as described by Lipowsky et. al, recently [42]. For a given membrane profile, the area of the coat Aa(s0) is computed using the relationship,(4)where, the neck-radius R(s0) is the radius at s0, which marks the coat boundary.
In our model, the dominant factor contributing to the intrinsic curvature H0 in the region where the membrane binds to the clathrin coat is the presence of epsins, bound at the CLAP-binding sites on the coat. In a recent study, [11], we modeled the spontaneous curvature induced by one epsin as a Gaussian function:(5)
That is, for the nature of epsin-induced curvature, we have assumed a form that has a spatial decay. Such a choice of spatially-varying intrinsic curvature function is motivated by recent molecular simulations [35], [36], [43], [44]. We have also employed such models in our earlier work [11], [45]. Similarly, for integral membrane proteins, a local curvature model has been proposed by Goulian et al. [46], Oster et al. [47], and Lubensky et al. [48]. Hence there is a bank of such phenomenological curvature models in use in the literature.
In vitro, Ford et al. [12] observed tubulation of vesicles upon addition of epsin; the observed tubule diameter of 20 nm enables us to estimate C0 = 0.1 nm−1. Using the surface-evolution approach, we calculate the curvature deformation energy of the membrane, Ec (defined in Eq. (3)) when a single epsin interacts with the membrane, i.e. through the curvature function in Eq. (5). Since the energy Ec is stabilized by the negative interaction energy of the ENTH domain of epsin with the membrane (Er), we iteratively determine the value of b in Eq. (5) such that Ec≈|Er|; using Er = −14kBT [12], we obtain b = 8.3 nm for κ = 20 kBT.
The periodicity of clathrin lattice, (from cryo-EM studies [49], the average distance between adjacent vertices of the hexagons in the clathrin cage is 18.5 nm), ensures that epsins are templated to maintain both spatial as well as bond-orientational ordering [50]. Hence, within our axis-symmetric membrane model, we translate the patterning of epsins on the clathrin coat to an intrinsic curvature function H0 of the form:(6)
Here, the index i runs over the number of concentric shells of epsins on the coat separated by a distance of 18.5 nm, the underlying periodicity of the clathrin lattice. Hence, relative to a central epsin bound to the coat at R = 0 and s0,1 = 0, successive shells of epsins are located at s0,2 = 18.5 nm, s0,3 = 37 nm, s0,4 = 55.5 nm, etc. until we reach the periphery of the coat of a prescribed extent (or area Aa); the H0 function is depicted in Fig. S5 and the schematic location of the shells is also depicted in Fig. 2. We note that the coat boundary is prescribed by the value of s0 for the outermost shell and the neck-radius R(s0) is the radius at this value of s0, as described earlier. In Fig. 3a, we depict energy minimized membrane deformation profiles for different values of the clathrin coat area Aa (defined in Eq. (4)) obtained using the surface evolution method and subject to the epsin curvature fields described by Eq. (6); we find that above a critical value of the coat area, the membrane profile develops overhangs, (also evident from the behavior of the neck-radius in Fig. 3b), which when the coat area Aa approaches 6500 nm2, transforms to a mature spherical vesicular bud with a narrow neck. We emphasize the generality of this result, i.e., that there exists a critical coat area above which the membrane deformation develops an over-hang and a constricted neck, by confirming this observed trend using a conceptually simplified “capsid model” in which H0(s) = 0.08 nm−1 if s<s0 and H0(s) = 0 if s≥s0, s0 is the length of the clathrin coat, as described in Text S2 and Fig. S1. In Fig. 3a, we estimate the number of epsins, Nepsins,i in each shell i as:(7)where, R is in nm, and 18.5 (nm) represents the triskelial spacing underlying the clathrin lattice; R(s) is depicted in Fig. S3. The total number of epsins is obtained by summing over the number of shells, which for the mature vesicular bud is estimated to be 23, see (a) in Fig. 3a.
Our results for the epsin shell model assumed a value for the bending rigidity of κ = 20 kBT reported in the literature [26], [27]. However, membrane bending rigidity depends upon multiple factors: membrane lipid and protein composition, anchoring of lipid with cytoskeleton, etc. Hence, a broad range of bending rigidity, 10–400 kBT has been reported in the literature: in particular, there is consensus that cytoskeleton-free membranes have rigidity in the range of 20 kBT and cytoskeleton-fortified membranes can be as stiff as 400 kBT. For this reason, it has indeed been postulated that apparent bending rigidity of the membrane depends on the relevant length scale and lies between 20 kBT (membrane patches below 100 nm) and 500 kBT (membrane patches of 1 µm) [26]. Hence, we have further explored the effect of varying κ in the range κ = 10–50 kBT on the mechanism of epsin-induced vesicular bud formation. In Fig. 4, we plot the membrane profiles for a mature vesicle for different values of κ. We note that, in varying κ, we also self-consistently determined the value of b (the range of epsin curvature) as outlined earlier: the dependence of b on membrane bending rigidity is shown in Fig. S4. For each value of κ, we varied the number of shells i in Eq. (6) to solve for the membrane profiles and determined the number of shells necessary for obtaining a mature vesicle; Nepsins and the diameter of the vesicular bud, d, were also computed as depicted in Fig. 4. The membrane profiles in Fig. 4 suggest that the epsin-shell model is still viable in orchestrating a mature vesicular bud for different values of membrane bending stiffness. However, we note that there is a strong dependence of the bud diameter on the bending rigidity, which suggests that the variations of in the size of the vesicle in CDE across cell types could be due to changes in the effective bending rigidity of the membrane.
The computed deformation energy Ec (defined in Eq. (3)) for the capsid model is plotted in Fig. S2 and is seen to increase linearly with increasing coat area, Aa; we find that the energy Ec required to form a mature spherical bud of diameter 50 nm is estimated to be 25κ = 500 kBT. The estimate is very close to 8πκ, which is the deformation energy of a spherical vesicle of diameter d for which H0 = 4/d (and constant in space). The energy Ec required to deform the membrane can be offset by stabilizing interactions between the proteins in the clathrin coat assembly and between the coat proteins and the membrane. As described in the introduction, the free energy of the clathrin-coat assembly, Ea was estimated by Nossal [29] to be ≈−20 kBT, i.e., |Ec|≫|Ea|. This implies that the curvature induction in the presence of a clathrin-coat is energetically unfavorable in the absence of additional stabilizing interactions. Indeed, as reported in cell-experiments [25], not all growing clathrin coats result in vesiculation events and a commitment step possibly accounting for additional stabilizing interactions (Er which includes those interactions that preferentially stabilize state 2 over state 1 in Fig. 1) is necessary. As noted in earlier 1, inclusion of epsin in the clathrin-coat accounts for εepsin = −23 kBT per bound epsin and hence, within our model, we consider Er(Aa) = Nepsins(Aa)×εepsin. Thus, for a given extent of the coat characterized by its area Aa, the total free energy change of the membrane and clathrin-coat assembly in the curved state (state 2, see Fig. 1) relative to the planar state (state 1, see Fig. 1) is given by: Et(Aa) = Ec(Aa)+Ea(Aa)+Er(Aa).
Recently, Jakobsson et al. [19] studied the role of epsin in synaptic vesicle endocytosis by inhibiting the interactions of epsin with clathrin using a CLAP antibody and those of epsin with PIP2 on membrane using an ENTH antibody. By microinjecting the CLAP antibody into neuronal cells, they observed that while the total extent of clathrin coated regions in the periactive zone on the plasma membrane remained the same, the observed fractions of the coated regions in different stages of coated-vesicle budding prior to scission were altered in a dramatic fashion, (see Fig. 5b): in the control wildtype (WT) cells, coated structures resembling a mature vesicular bud are more probable in comparison to planar structures and early intermediates; however, upon addition of CLAP, the early intermediates are stabilized and become more probable at the expense of the number of mature vesicular buds [19].
By computing Ec and Er for different values of Aa in the capsid model, we determine the energetics of the clathrin coated vesicular bud Et versus coat area, Aa for the capsid model (see Fig. S6). Number of epsins in WT (control) cell = 21: this number differs slightly from 23, the estimate for the epsin shell model, because R(s0) for the capsid model is slightly different from that for the shell model. We also computed probability of observing different coated-intermediates of vesicular structures as P∝exp(−Et(Aa)/kBT) as depicted in Fig. 5a. The predicted distribution of vesicular intermediates (Fig. 5a) closely matches the experimental distribution reported by Jakobsson et al. [19] (see Fig. 5b). For modeling the clathrin-coated vesiculation in CLAP IgG injected cells, we compute the number of epsins as Nepsins(CLAP cells) = Nepsins(WT cells)*Aa(vesicles in CLAP injected cells)/Aa(in WT cells) = 33. The ratio of the respective areas ( = 1.6) is determined based on the experimental observations of increase in the size of the coated intermediates in CLAP injected cells relative to WT cells [19]. Remarkably, with Nepsins = 33 and εepsin = −14 kBT (reduced from −23 kBT due to the abrogation of the CLAP-clathrin/AP-2 interaction), we find not only that Et(Aa) increases monotonically with Aa (a reversal in trend, see Fig. S6) but also the probability P∝exp(−Et(Aa)/kBT) quantitatively matches the experimentally observed distribution in CLAP IgG injected cells, (compare Figs. 5a and 5b). We note that even though Nepsins increase in the CLAP IgG injected cells relative to wildtype, the size of the bud likely increases due to a lack of templating of epsins; arguably, there is lack of bond-orientational order as the CLAP domains of epsin can no-longer bind the periodic clathrin lattice. Corroborating this view, many extended coated structures (cisternae) also appear in the experiments with CLAP IgG injected cells [19]. Furthermore, according to the predictions of our model, disrupting the epsin-membrane interaction (i.e., by targeting the ENTH domain of epsin) completely abrogates Er and should make the coated vesicular bud highly unfavorable. Indeed, consistent with this view, in cells microinjected with ENTH antibodies the extent of clathrin-coated structures decreased by over 90% [19]. Regarding the comparison in Fig. 5, we re-iterate that the fraction (or histogram) is proportional to exponential of the energy. Hence a small error in energy (of the order of kBT which is 0.6 kcal/mol at T = 300 K) can lead to a large change in the fraction [exp(0.6)≈factor of 2]. Hence, an order of magnitude agreement in histograms between theory and experiment in the trends of the intermediate shapes implies that the energetics agree even more closely.
In conclusion, we have presented a minimal mesoscale model which we believe imposes the correct spatial as well as thermodynamic constraints, and quantitatively explains several experimental observations on the process of vesicle nucleation induced by the clathrin-coated assembly prior to vesicle scission in CDE. We re-iterate that the input to our model is the membrane bending rigidity, spacing between epsins bound to the clathrin coat, and the curvature-field imposed by each bound epsin, which have all been determined using independent biophysical experiments. For these choices of input, our calculations then yield the membrane profiles for different sizes of the clathrin coat. Based on the number of shells of epsins accommodated on the clathrin coat (which depends on the size of the coat), and the circumference of each shell (which depends on the coat/membrane deformation), the number of epsins is calculated. Thus, the number of epsins, the membrane profile, and the deformation energy are outputs of our model. While our model does not include nucleation of the clathrin coat or scission of a mature coated vesicular-bud, our results identify a unique dual role for the tubulating protein epsin: multiple epsins localized spatially and orientationally collectively play the central role of a curvature inducing capsid; in addition, epsin serves the role as an adapter in binding the clathrin coat to the membrane. Our results also suggest an important role for the clathrin lattice, namely in the spatial- and orientational-templating of epsins for providing the appropriate curvature field for vesicle budding. We suggest that there exists a critical size (area) of the coat above which a vesicular bud with a constricted neck resembling a mature vesicle is stabilized. Based on the strong dependence of the vesicle diameter on the bending rigidity, we suggest that the variability in bending stiffness due variations in membrane composition with cell type can explain the experimentally observed variability on the size of clathrin-coated vesicles, which typically range 50–100 nm.
Apart from providing a mechanistic description of the budding process in CDE, our model provides estimates for the number of epsins involved in stabilizing a coated vesicle, and without any direct fitting, reproduces the experimentally observed shapes of vesicular intermediates as well as their probability distributions quantitatively in wildtype as well as CLAP IgG injected neuronal cell experiments. We consider such an agreement to be a strong validation for the basis of our model. These model predictions can further be tested by engineering mutations in epsin, clathrin, and AP-2 all of which are predicted to influence the distribution of coated structures. The framework of our approach is generalizable to vesicle nucleation in clathrin-independent endocytosis. Indeed, based on our results we can speculate that alternative mechanisms (such as receptor clustering) which can provide a hexatic bond-orientational templating of epsins on the membrane can facilitate vesicle-bud formation independent of CDE [11]. Future modeling work will address spatial distribution of curvature inducting proteins on vesicle nucleation [20].
|
10.1371/journal.ppat.1002316 | Histone Deacetylase 8 Is Required for Centrosome Cohesion and Influenza A Virus Entry | Influenza A virus (IAV) enters host cells by endocytosis followed by acid-activated penetration from late endosomes (LEs). Using siRNA silencing, we found that histone deacetylase 8 (HDAC8), a cytoplasmic enzyme, efficiently promoted productive entry of IAV into tissue culture cells, whereas HDAC1 suppressed it. HDAC8 enhanced endocytosis, acidification, and penetration of the incoming virus. In contrast, HDAC1 inhibited acidification and penetration. The effects were connected with dramatic alterations in the organization of the microtubule system, and, as a consequence, a change in the behavior of LEs and lysosomes (LYs). Depletion of HDAC8 caused loss of centrosome-associated microtubules and loss of directed centripetal movement of LEs, dispersing LE/LYs to the cell periphery. For HDAC1, the picture was the opposite. To explain these changes, centrosome cohesion emerged as the critical factor. Depletion of HDAC8 caused centrosome splitting, which could also be induced by depleting a centriole-linker protein, rootletin. In both cases, IAV infection was inhibited. HDAC1 depletion reduced the splitting of centrosomes, and enhanced infection. The longer the distance between centrosomes, the lower the level of infection. HDAC8 depletion was also found to inhibit infection of Uukuniemi virus (a bunyavirus) suggesting common requirements among late penetrating enveloped viruses. The results established class I HDACs as powerful regulators of microtubule organization, centrosome function, endosome maturation, and infection by IAV and other late penetrating viruses.
| Histone deacetylases (HDACs) are generally associated with the epigenetic regulation of gene expression in the nucleus, but some have been shown to possess cytoplasmic functions. While analyzing the role of cell factors in influenza A virus entry into host cells, we observed that depletion of members of the class I HDAC family dramatically affected the efficiency of infection. Depletion of HDACs 8 and 3 decreased, and depletion of HDAC1 elevated the efficiency of entry. For HDAC1 and 8, this could be traced back to opposing effects on the architecture of centrosomes and consequences on microtubule organization. HDAC8 depletion caused the centrosomes to split and move away from each other. The microtubules were disorganized, and endosomes failed to move to the perinuclear region of the cell. Endocytosed viruses did not penetrate because the endosomes dispersed throughout the cytoplasm and did not acidify properly. In contrast, when HDAC1 was depleted, fewer centrosomes were split, and endosome transport and acidification became more efficient. Taken together, our results showed for the first time that class I HDACs play a role in the organization of the microtubule network, in endosome maturation, and in the entry of influenza and other late penetrating viruses into host cells.
| To enter their host cells, the majority of animal viruses take advantage of the cell's endocytic machinery. After uptake, penetration of the viruses or their capsids into the cytosol generally occurs from early or late endosomes (EEs or LEs). Since endocytosis and endosome maturation are complex and tightly regulated activities, successful entry and infection relies on numerous cellular factors and processes. This is clearly illustrated by recent high-throughput siRNA screens that have identified hundreds of host cell genes required for infection by different viruses [1], [2]. The starting point for our study was a 7000 druggable-genome RNAi screen performed on the influenza A X31 strain (A/Aichi/2/68) (H3N2) in A549 cells suggesting that histone deacetylases (HDACs) are modulators of early infection.
IAVs are enveloped animal viruses with a segmented, negative-sense RNA genome. Point mutations, reassortment, and interspecies transmission cause recurrent epidemics and global pandemics in humans, birds, and animals [3]. At the cellular level, infection begins by virus binding to sialic acid residues on cell surface glycoproteins and lipids followed by internalization either via clathrin-mediated endocytosis or a clathrin-independent, macropinocytosis-like uptake process [4], [5], [6], [7]. The virus particles are transported into the endosome system. Penetration of the genome into the cytosol is mediated by the hemagglutinin (HA) glycoprotein, an acid-activated membrane fusion factor [8]. The low pH threshold for HA activation (pH 5.4-4.9) dictates that penetration by membrane fusion takes place in LEs or endolysosomes usually in the perinuclear region of the cell [9], [10]. After penetration, the matrix protein (M1) dissociates, and the viral ribonucleoproteins (vRNPs) are imported via nuclear pore complexes into the nucleus where replication and transcription take place [11], [12].
The centrosome is the major microtubule organizing center (MTOC) of animal cells. Centrosomes bind more than 100 regulatory proteins, whose identities suggest roles in a multitude of cellular functions [13]. By nucleating and anchoring microtubules (MTs), centrosomes influence most MT-dependent processes, including organelle transport, cell shape, polarity, adhesion, motility, and division. After duplication of the two centrioles during S phase [14], the two resulting centriole doublets continue to function as a single MTOC until they separate at the onset of the mitosis.
Acetylation is a reversible post-translational modification that neutralizes the positive charge of lysines, changing protein function in diverse ways [15]. It plays a central role in the epigenetic regulation of gene expression through modification of histone tails by histone acetyltransferases (HATs) and histone deacetylases (HDACs) [16]. Acetylated proteins are typically subunits of large macromolecular complexes involved in processes such as chromatin remodeling, cell cycle regulation, splicing, nuclear transport, MT stability, and actin nucleation. With more than 1700 substrate proteins identified by proteomic analysis, the regulatory scope of lysine acetylation is broad [17].
HDACs are classified into three subclasses [18]. In this study, we focused on class I HDACs (HDAC1, 2, 3, 8) and their role in infection. We found that some of them increased or decreased IAV's productive entry by modulating endocytosis, by affecting the properties of the MT network, and by influencing the maturation of endosomes. HDAC8 was found to support infection by promoting LE/LY motility and MT organization at the centrosome, and by increasing centrosome cohesion. In contrast, HDAC1 decreased centrosome cohesion and suppressed IAV infection. The two enzymes thus displayed opposing effects on IAV entry by controlling endosome function through changes in the MT network and the centrosome.
In a high-throughput RNAi screen, HDAC8 emerged as one of the proteins required for efficient infection of influenza A X31 strain (A/Aichi/2/68) (H3N2) in A549 cells. The readout for this screen was the expression of newly synthesized viral nucleoprotein (NP). To validate this finding and expand it to other class I HDACs, we adopted the siRNA silencing approach and tested several oligonucleotides against HDACs 1, 2, 3 and 8 for effects on IAV infection. The specificity and efficacy of the siRNAs were confirmed at the protein level by Western blotting and at the mRNA level by real-time PCR (Figure S1A and B). The efficiency of protein depletion for HDACs 1, 3 and 8 was 90%, 80% and 95%, respectively (Figure S1A). The effects of the siRNAs were specific in that each of them only induced depletion of the intended HDAC (Figure S1B).
Seventy-two hours after transfection with these siRNAs, cells were infected with influenza A X31 strain (A/Aichi/2/68) (H3N2), and 10 h later, the fraction of infected cells was quantified by immunofluorescence staining against the newly synthesized NP. For imaging, automated microscopy was used, and the fraction of infected cells quantified with a MATLAB program using an algorithm developed in our lab [19]. To allow detection of an increase as well as a decrease in the number of infected cells, conditions were adjusted so that 20% of the cells were infected in cells transfected with a control oligonucleotide (All*Neg).
Depletion of HDAC3 and HDAC8 was found to reduce IAV infectivity to 28% and 36% of control, respectively, whereas depletion of HDAC1 increased infectivity more than two-fold (Figure 1A). Depletion of HDAC2 had no effect (not shown). In cells depleted of a subunit of the vATPase (ATP6V1B2) responsible for acidification of endosomes, and CAS (CSE1L), a factor required for nuclear import of influenza vRNPs, infection was almost completely inhibited. In HeLa ATCC cells, HDAC1, 3, and 8 silencing had similar effects on infection as in A549 cells (Figure S1D).
To determine which steps in the infection cycle were affected, we used a series of quantitative, microscopy-based assays that allowed us to follow the progress of incoming virus particles by monitoring binding to the plasma membrane, internalization by endocytosis, conversion of the HA to its acid-induced conformation, and import of vRNPs into the nucleus. The results in figure 1B–G showed that the increase in infection caused by HDAC1 depletion correlated with a doubling in the efficiency of virus exposure to low pH judging by the increased conversion of HA to the acid-induced conformation detected using a specific monoclonal anti-HA antibody (Figure 1D, E). The import of vRNPs to the nucleus was correspondingly increased about 1.3-fold (Figure 1F). Since virus binding and internalization were unaffected (Figure 1B, C), this indicated that the fraction of endocytosed viruses undergoing acidification and productive penetration was elevated.
In cells depleted of HDAC8 or HDAC3, the loss of infectivity was explained by a dual effect on endocytosis and acidification. When measured 30 min after uptake, endocytosis of IAV was reduced to 48 and 41% (Figure 1C). Total acidification was reduced to 30% and 17%, compared with controls, respectively, meaning that roughly half of endocytosed viruses underwent acid exposure (Figure 1E). Consequently, the nuclear import of vRNPs was also reduced (Figure 1F). When dynasore, an inhibitor of dynamin [20] was used, the level of internalization dropped to 29% (Figure 1C). Bafilomycin A (a vATPase inhibitor) blocked acid conversion and nuclear import almost completely (Figure 1E, F). The kinetics of the conversion of HA to the acid-induced conformation was not affected by depletion of any of the HDACs. The acid HA peaked around 1 hour after warming, followed by a decline caused most likely by degradation of the HA in endolysosomes. In HDAC8- and HDAC3-depleted cells, the amount of acidic HA remained low at all time points (Figure 1D).
The assays indicated that the drop in influenza infectivity after HDAC depletion involved pre-penetration steps in the endocytic pathway. This was confirmed for HDAC1 and HDAC8 by inducing infection without the need for the virus to undergo endocytosis. Such by-pass was achieved by adding low pH medium to cells with bound virus, thus inducing fusion of the viral envelope directly with the plasma membrane [21]. In HDAC1-depleted cells, no increase in infectivity was observed compared to control cells, and no loss of infectivity was seen in HDAC8-depleted cells (Figure 1G). Cells depleted of CAS served as a post-penetration control: here the block could not be by-passed. We concluded that the effects of HDAC8 and HDAC1 involved pre-penetration steps in the endocytic pathway. However, the by-pass in HDAC3-depleted cells only partially restored infectivity (55% of control), implying that this enzyme was involved in both pre- and post-penetration steps (Figure 1G). We did not study HDAC3 further.
Taken together, the results indicated that class I HDACs participate in the regulation of the endocytic pathway, especially the pathway from EEs to LYs. Perturbation of HDACs had major consequences on the efficiency of IAV entry. Productive entry was enhanced after HDAC1 depletion and inhibited after HDAC8 or HDAC3 depletion. In HDAC8- and HDAC3-depleted cells, the endocytic internalization and acid-conversion of the virus were inefficient with the outcome that few vRNPs reached the nucleus. In HDAC1-depleted cells, primary endocytosis was normal but acid exposure, penetration, and vRNP delivery to the nucleus were dramatically enhanced.
When we examined the distribution of organelles positive for the LE/LY marker LAMP1 in HDAC8-depleted cells by indirect immunofluorescence microscopy, we observed that instead of clustering mainly in the juxtanuclear region, the vacuoles were distributed throughout the cytoplasm (Figure 2A). The Golgi complex, visualized using giantin as a marker, was also dispersed. In contrast, HDAC1 depletion induced increased clustering of these organelles in the perinuclear region. Some of the vacuoles were larger than in controls. EEs identified by the presence of EEA1 [22] were larger in HDAC1-depleted cells than in controls but their intracellular distribution was normal (not shown).
For quantitation of the dispersion effect, we derived a dispersion index using an algorithm described in materials and methods. With a dispersion index of 1.0 for LAMP1-positive vacuoles in nocodazole-treated cells and zero for unperturbed cells, HDAC8-depleted cells had a positive index of 0.52 (Figure 2B). The depletion of two dynactin subunits (ACTR10 and DCTN2, required for MT- and dynein-dependent retrograde transport of vesicles) [23] resulted in a dispersed phenotype similar to nocodazole. Dispersed vesicles were also observed after depletion of KIFC1 (0.89) and KIFC2 (0.52), which are both minus end-directed kinesin motors [24]. Finally, cells depleted of HDAC1 had a negative index (−0.33) consistent with the increased perinuclear clustering of the vacuoles.
To determine whether the effects were general and affected endogenous cargo, we examined the fate of epidermal growth factor (EGF), which is normally transported to LYs and degraded, and transferrin (TF), which is recycled from EEs to the plasma membrane [25], [26], [27]. We observed that HDAC1 depletion had no effect on either ligand. HDAC8-depleted cells exhibited a partial decrease in the degradation of EGF (p = 0.049, Student's t-test), and intracellular TF increased by 40% suggesting an increase in endocytosis or a decrease in recycling (Figure 2C, D).
The dispersal phenotype of LE/LYs in HDAC8-depleted cells suggested that the HDACs affected the MT system. It was therefore of interest to determine whether MT disruption and stabilization would affect infection. Nocodazole had no effect on IAV endocytosis, but reduced HA acidification by 46% and infection by roughly 50% (Figure 3A, B). Taxol had no effect. Thus by dispersing endosomes, nocodazole had an effect on infection and acidification similar to HDAC8 depletion: it reduced the infectivity of endocytosed IAV to half.
A change in endosome behavior could also be visualized by video microscopy in live cells when Alexa Fluor 647 (AF647)-labeled wheat germ agglutinin (WGA), a lectin that binds to cell surface sialic acids and glycoproteins, was allowed to be endocytosed and routed to LEs [28]. The videos showed clearly that instead of accumulating in the perinuclear region as in control cells, WGA-positive vesicles in HDAC8-depleted cells continued to move throughout the cytoplasm (Video S1). When R18-labeled IAV particles were co-internalized with WGA-AF647, they colocalized with WGA in endosomes soon after uptake. While in control cells the virus- and WGA-containing vacuoles moved into the perinuclear region, they remained peripherally distributed in HDAC8-depleted cells (Figure 3C).
The endosomal compartment in which IAV was acidified could be visualized by indirect immunofluorescence microscopy using the antibody against acidified HA. In cells transiently overexpressing EGFP-Rab7 (a marker for LEs) and LAMP1-mCherry (a marker for LE/LYs), the acidified HA signal was mainly detected in Rab7/LAMP1-positive vesicles, residing in the perinuclear region of the cell. In HDAC8-depleted cells, however, the acidified HA when present was in peripheral, Rab7/LAMP1-positive organelles corresponding most likely to LEs (Figure 3D).
These results indicated that the distribution of LE/LYs was regulated by the HDACs. HDAC8 promoted centripetal movement and perinuclear localization of LE/LYs, HDAC1 opposed the accumulation of LE/LYs in the perinuclear region. It was, moreover, possible that efficient acidification of the vacuoles was somehow linked to their location within the cytoplasm.
To analyze the dynamics of endosome movement in more detail, we used the particle-tracking ImageJ plugin developed by Sbalzarini and coworkers and their trajectory segmentation toolbox [29], [30]. The movement of EGF-AF488-containing vesicles was recorded 15 to 30 min after internalization over 2000 frames (Figure 4A, B). Trajectories were segmented and classified according to their pattern of movement. The frequency of MT-dependent directed motion was 4.5% in control cells and 1.06% in HDAC8-depleted cells. Nocodazole treated cells had only 0.33% directed motion. The velocity of directed motion when it occurred was similar in all three cases suggesting that motor proteins were functional (Figure 4A). Of directed movements in control cells, 10.6% continued for more than 1.5 sec, whereas in HDAC8-depleted cells the value was 3.67% (Figure 4B). This meant that either the motors fell off prematurely in cells lacking HDAC8, or that the MTs were shorter or lacked stability.
Immunofluorescence staining with anti-tubulin antibodies showed that, unlike nocodazole treatment, HDAC8 depletion did not eliminate the MTs, but caused disorganization of the MT system. Instead of radiating from the MTOC, the MTs were randomly oriented, criss-crossing each other in the cytoplasm (Figure 5A). In many cells, the MT network was denser at the cell periphery than in the cell center (Figure S2A).
Moreover, the MTOCs were dramatically altered. Staining for centrosomes with γ-tubulin antibodies showed that instead of exhibiting two closely paired spots in the MTOC, they occurred in similar spots that were far apart, and occasionally on opposite sides of the nucleus (Figure 5A). Centrosome splitting was observed in 66.6% of HDAC8-depleted cells compared to 13.6% in the control cells (n = 500) (Figure 6B). The average distance between the two centrosomes was 5.43±1.30 µm in HDAC8-depleted cells, compared to 1.56±0.25 µm in control cells (n = 250) (Figure 6C). There was also a defect in MT anchoring to the split centrosomes; they appeared to be entirely disconnected (Figure 5A). Judging by the diffuse staining of DNA and the cell shape, these cells were in interphase, during which centrosomes do not normally split. Together with the splitting of the centrosomes, this may explain the gross redistribution and abnormal morphology of MTs.
Tubulin acetylation increases the stability of MTs [31]. To examine whether HDAC8 depletion affected the stability of tubulin, we determined the acetylation level of α-tubulin in control, HDAC1-, 6-, and 8-depleted A549 cells (Figure S3). Depletion of HDAC6, a tubulin deacetylase [32], increased tubulin acetylation by 60%, whereas HDAC8 depletion reduced it to 20% (Figure S3A, B). HDAC8 localized diffusely throughout the cell (Figure S4).
When MTs were depolymerized in the cold in HDAC8-depleted cells and allowed to regrow at 37°C, they were found to polymerize at random sites in the cytoplasm (Figure 5B). Thus, formation of MT asters after 90 sec of regrowth could be observed in only 11–29% of HDAC8-depleted cells compared to 83% of the control cells (Figure 5C). MT nucleation can be affected by the displacement of ninein, a protein normally present in both centrioles with higher affinity to the mother centriole [33]. Ninein localized to centrosomes in HDAC8-depleted cells (Figure 5D) with an efficiency of 92%, compared to control cells (Figure 5E).
Was centrosome splitting the underlying reason for the inhibition of IAV infection? To test this possibility, we used siRNAs to deplete cells of rootletin, a protein present in centriole-associated fibers, the depletion of which causes centrosome splitting [34]. In rootletin-depleted cells, we observed centrosome splitting in 75.3% compared to 13.6% of control cells (n = 500) (Figure 6B). The average distance between the two centrosomes was 4.27±0.37 µm compared to 1.56±0.25 µm in control cells (n = 250) (Figure 6C). The only major difference compared with HDAC8-depleted cells was that the effects on MT organization were less dramatic under rootletin depletion (Figure 5A, S2A). LE/LYs gave an intermediately dispersed phenotype (Figure S2B). The formation of MT asters after 90 sec of regrowth was unaffected (Figure 5C). In other words, the split centrosomes were apparently able to connect with MTs.
When the effect of rootletin depletion on IAV infection was tested, we found that infection was reduced to 20% of controls (Figure 6A). Virus endocytosis was unaffected, but HA acidification was reduced to 40% of normal (not shown). Thus, by causing centrosome splitting by an independent mechanism, it was possible to induce a block in infection similar, but not identical, to that caused by HDAC8 depletion. It was apparent that by inducing the separation of centrosomes, the loss of either HDAC8 or rootletin caused changes in the MT system of the cell that resulted in incomplete endosome maturation and reduced entry of IAV.
The results observed after HDAC1 depletion seemed to support a correlation between centrosome architecture and infection. HDAC1 depletion reduced centrosome splitting to 9.3% compared to 13.6% in control cells (n = 500) (Figure 6B), and shortened the average distance between centrosomes to 1.3±0.09 µm (n = 200) (Figure 6C).
Interestingly, the centrosome splitting caused by HDAC8 depletion could be reduced to 25.6% and the inter-centrosomal distance to 2.4±0.1 µm by co-depletion of HDAC1 (n = 200) (Figure 6B, C). Depletion of both HDAC1 and HDAC8 was sufficient under both single and co-depletion conditions (Figure S5). Co-depletion of HDAC1 with rootletin also reduced splitting of centrosomes from 75.3% to 44.7%, and the average centrosome distance to a mere 2.6 µm (n = 100) (Figure 6B, C). The pan-HDAC inhibitor trychostatin A (TSA) preferentially blocks HDAC1 but does not affect HDAC8 [35]. Of HDAC8-depleted cells that were treated with 5 µM TSA for 12 h, only 28% showed centrosome splitting, compared to 65% of dmso treated cells (Figure 6D). That HDAC8 and HDAC1 had opposing effects on centrosome distance and infection was further supported by double transfection experiments where the two were silenced simultaneously. When the siRNAs were mixed, a linear conversion from inhibition to activation was observed depending on the ratio of siRNAs used (Figure 6E). Finally, compared to control cells, HDAC1 depletion increased the G1 phase population from 57% to 72%, while HDAC8 depletion increased the G2/M phase population from 18% to 25% (Figure 6F). Therefore, a specific block in the cell cycle was unlikely to be the reason for MT disorganization in HDAC8-depleted cells.
We concluded that the effects of HDAC depletion on IAV infection correlated with the architecture of centrosomes. When centrosome splitting was induced, MT organization was disturbed, with the result that endosomes failed to be transported properly into the perinuclear region. Endosome acidification was not properly activated, and virus particles failed to release their vRNPs into the cytosol. In contrast, when the average centrosome distance decreased, as it did in HDAC1-depleted cells, the efficiency of acidification and infection increased. It is of interest to note that infection did not correlate with MT anchoring to centrosomes; whereas the split centrosomes in HDAC8-depleted cells were not anchored, the centrosomes in rootletin-depleted cells remained anchored to MT (Figure 5A, S2A).
Since viruses use different endocytic strategies and different pathways to enter cells [36], it was of interest to test whether HDAC8 or HDAC1 depletion affected other viruses. We tested vesicular stomatitis virus (VSV, a rhabdovirus), Uukuniemi virus (UUKV, a bunyavirus), and mature virions (MVs) of vaccinia virus (VACV, a poxvirus). VSV enters cells by clathrin-mediated endocytosis and the pH threshold for fusion of the viral envelope with endosomal membranes is pH 6.2 [37], [38]. UUKV uses clathrin-independent mechanisms for infectious entry and its threshold for membrane fusion is pH 5.4 [39]. VACV MVs enter cells by macropinocytosis [40].
Depletion of HDAC1 had no effect on any of the viruses tested except IAV. Results were similar when cells were treated with 5 µM TSA for 4 h prior to infection (Figure S6). HDAC8 depletion reduced IAV and UUKV infection to 30% and 52%, respectively (Figure 7). VACV was unaffected. VSV infection increased by 48% confirming once again that there was no defect in clathrin-mediated uptake [37] (Figure 7). Consistent with the increase in TF accumulation, the increase in VSV infection could be due to an expansion of the EE compartment caused by the perturbation of the LE maturation program. Rootletin depletion reduced UUKV infection to 23% (Figure S7). Thus, HDAC8, rootletin and the functionality of centrosomes is critical for the infection of late penetrating viruses.
The most important new insights from our study were that HDAC8 and HDAC1 regulate the properties of the MT system and architecture of centrosomes in interphase cells, and that this influences the motility, distribution, and maturation of LEs and LYs. One consequence is that the infectious entry of IAV and other late penetrating viruses that require acidification to a pH of about 5.5 or lower and penetrate from LEs [41], [42], [43], are affected by the expression of class I HDACs. A role of HDAC1 or 8 in the regulation of the MT cytoskeleton in interphase cells and the maturation of endosomes has not been previously described.
Depletion of HDAC8 and HDAC3 resulted in decreased infection by IAV. HDAC8, which we analyzed more extensively of the two, is localized in the nucleus and in the cytoplasm (Figure S4). Although it is well-characterized at the molecular level, its cellular functions have remained elusive [44], [45], [46], [47]. Unlike other class I HDACs, HDAC8 does not form high-molecular weight multi-molecular complexes [15]. We found that HDAC8 depletion resulted in a two-step reduction in IAV entry. First, endocytic internalization was reduced to less than half compared to control cells. Second, of the internalized viruses only half were acidified and therefore capable of membrane fusion and penetration. Thus, the overall infection rate amounted to 15–30% of controls. That VSV infection and the internalization of TF were not inhibited, ruled out a general defect in clathrin-mediated endocytosis and the cellular translational machinery. Since IAV endocytosis occurs by two parallel pathways, a clathrin-mediated and a recently described macropinocytosis-like mechanism [7], it is possible that it is the latter pathway that is affected. The link between class I HDACs and endosome maturation may be of significance in the differentiation of cells as well as during oncogenesis, which is often associated with elevated HDAC activity. HDACs are also likely to affect the tropism and pathogenesis of influenza and other viruses that use endocytosis for entry. When HDAC inhibitors are deployed in cancer treatment, it will be of importance to consider changes in influenza virus susceptibility.
The second effect of HDAC8 depletion on IAV - the decreased exposure to pH of 5.1 or below - occurred late in the endocytic pathway. A decrease in the degradation of EGF indicated that cargo was not reaching the lysosomal compartments with normal efficiency. In HDAC8-depleted cells, internalized viruses were trapped in Rab7/LAMP1-positive vacuoles, which instead of moving along MTs to the MTOC remained dispersed in the cytoplasm. Particle tracking showed that LEs and LYs were less motile and less tightly connected to MTs, and failed to undergo sustained directional movement. The few vacuoles that contained acidified viruses had a peripheral location.
In normal cells, organized movement and acidification are part of a complex maturation program that prepares LEs for fusion with LYs. Maturation involves conversion of Rab GTPases, a switch in phosphatidylinositides, acidification, formation of intraluminal vesicles, association with dynein and dynactin, and MT-mediated transport to the perinuclear region [36], [48], [49]. These changes are coordinated and interdependent in complex ways, so that when one function is perturbed the whole program may be disrupted. Therefore it is possible that proper acidification failed to occur because the LEs did not move to the MTOC [50]. A similar situation was created when endosome movement was inhibited by disrupting MTs with nocodazole; acidification of the virus and productive infection by internalized viruses also dropped by half. The location of endosomes is determined by many factors including the presence of adaptor proteins that provide a link to the cytoskeleton, and various motors, the size and shape of the organelles, etc. [51]. Our data suggest that the properties of an endosome are affected by their location such that if they cannot move to the perinuclear region of the cell their maturation is not completed.
HDAC8 depletion had a dramatic effect on the MT system. It caused centrosome splitting, loss of tubulin acetylation and MT nucleation/anchoring. In contrast to nocodazole treatment, most MTs were still present, but since the MTOC was disrupted they were disorganized. Disruption of the MTOC may account for the loss of acetylation of tubulin. Long-term displacement of MTs from the mother centrosome may have been the reason why the centrosomes drifted apart, reminiscent of the splitting that can be induced by nocodazole treatment [52]. Aside from being split and unable to support MT anchoring, the centrosomes appeared normal after HDAC8 depletion. Important factors such as γ-tubulin and ninein were present.
The underlying reason for the reduction of IAV infection and infection with another late penetrating virus, UUKV, was most likely the loss of centrosome cohesion. The best evidence was that the induction of centrosome splitting by an independent mechanism, i.e. by the depletion of rootletin, caused dramatic reductions in infection. Rootletin depletion induced an intermediate level of LE/LY dispersal. It is possible that close centrosomes provide a more focused MT network compared to split centrosomes. It has been shown that fewer MTs radiate from centrosomes when they are apart [53].
The effects of HDAC1 depletion also supported centrosome distance as an important factor in IAV infection. HDAC1 is a class I HDAC almost exclusively localized in the nucleus, with important functions in epigenetic regulation of cell fate and cellular processes [15]. When HDAC1 was depleted, infection by IAV more than doubled. The percentage of split centrosomes decreased. Inhibition of HDAC1 function by TSA also increased IAV infection. In fact, we repeatedly observed that the shorter the distance between centrosomes in a cell population, the higher the level of infection. LEs, LYs, and the Golgi complex were tightly bundled around the MTOC, and acidification of HA was twice as efficient as in control levels.
That centrosome splitting was suppressed when HDAC8 was co-depleted with HDAC1, suggested that HDAC1 may have a role in diminishing centrosome cohesion and that HDAC8's role could be to promote MT organization by enhancing the association of MTs with the centrosome. It remains to be proven whether these HDACs work directly on centrosomes, MTs or other factors. It is possible that one or more regulators of centrosome cohesion are acetylated and serve as substrates for deacetylases. C-Nap1, a centriolar linker protein, is an interesting candidate, since it is known to be acetylated [17], and is a well known regulator of centrosome cohesion [54]. Our results show that class I HDACs are important regulators of MT organization, as well as, centrosome architecture and function, and IAV entry.
A549 ATCC and HeLa ATCC cells were propagated according to the ATCC guidelines. Purified influenza A X31 strain (A/Aichi/2/68) (H3N2) was purchased from Virapur (CA, USA). UUKV S23, VSV (wtVSV) (Indiana serotype) and VACV WR-GFP were used as previously described [37], [40], [55].
Purified influenza A X31 strain (A/Aichi/2/68) (H3N2) was purchased from Virapur (CA, USA). Briefly, 60 pathogen-free chicken eggs were inoculated with the virus and incubated for 2 days at 33–37°C. Harvested allantoic fluid was clarified by low speed centrifugation, and concentrated by high-speed centrifugation. The virus was further concentrated by two rounds of 10–40% sucrose gradient centrifugation. Virus bands were harvested, pooled and resuspended in formulation buffer (40% sucrose, 0.02% BSA, 20 mM HEPES pH 7.4, 100 mM NaCl, 2 mM MgCl2), frozen and stored at −80°C until usage (TCID50 in MDCK cells = 3.0×108/ml virus).
Cells were reverse-transfected (final 10 nM) with siRNAs (QIAGEN) in 96-well flat-bottom Matrix plates (Thermo Scientific). siRNAs were mixed with Lipofectamine RNAiMax (Invitrogen) in 30 µl OPTI-MEM (Invitrogen) for 30 min. Cells were trypsinized, counted and plated directly onto the lipofectamine mixture in 70 µl of growth medium.
A549 cells in 96-well Matrix plates were infected with TCID50 = 7500/ml virus in infection medium (D-MEM, 50 mM HEPES pH 6.8, 0.2% BSA) and fixed at 10 h in 4% formaldehyde (FA) in PBS. Cells were permeabilized in permeabilization buffer (0.1% saponin, 1% BSA in PBS) and stained for viral NP by indirect immunofluorescence with monoclonal antibody HB-65 (ATCC) (1∶10 dilution) and secondary anti-mouse antibody labeled with Alexa Fluor 488 (AF488). Nuclei were stained with Hoechst (1∶10000 dilution). Typically, 9 (3×3) or 16 (4×4) images per well were automatically acquired (see Microscopy). Cell numbers and raw infection indices for each well were determined using a MATLAB-based infection scoring procedure (The MathWorks, Inc.). With this method, cells expressing NP were counted as infected. Control cells transfected with All*Neg siRNA exhibited 15–20% infection.
Microscopy-based assays for step-wise dissection of the IAV X31 strain entry pathway were developed in our laboratory. The assays use indirect immunofluorescence for detection and are optimized for high-throughput analyses in a 96-well format (Banerjee I., Horvath P. and Helenius A., manuscript in preparation. Programs used for quantification can be downloaded at www.highcontentanalysis.org). Here we use both confocal and automated fluorescence microscopy to acquire the images and ImageJ to quantify the results.
A549 cells were bound with TCID50 = 2.4×106/ml virus for 1 h at 4°C, washed three times in PBS, fixed in 4% FA in PBS for 20 min, blocked in blocking buffer (1% BSA in PBS), and processed for indirect immunofluorescence with anti-X31 Pinda antibody [56] (1∶2000 dilution), phalloidin-Alexa Fluor 594 (AF594) and DRAQ5 (Biostatus Limited) to stain nuclei. Images were acquired by confocal immunofluorescence microscopy using a 40× objective and quantified using ImageJ. First, a threshold value for the Pinda signal was determined for each experiment in order to eliminate background fluorescence. The area covered by fluorescence signal above this threshold was calculated. Second, phalloidin staining was used to determine cell area. Finally, the total Pinda signal per cell area was quantified for each siRNA treatment and normalized to control (All*Neg) cells.
A549 cells were bound with TCID50 = 2.4×106/ml virus for 1 h at 4°C, washed twice in infection medium (D-MEM, 50 mM HEPES pH 6.8, 0.2% BSA), and warmed to 37°C to allow virus uptake. After 30 min, cells were fixed in 4% FA for 20 min. After blocking, extracellular HA antigens were blocked overnight at 4°C with Pinda antibody (1∶500) and fixed again in 4% FA in PBS for 20 min. The cells were then permeabilized (0.1% saponin, 1% BSA in PBS) and incubated with a monoclonal antibody specific for X31 HA1 [57] (1∶100) for 2 h at room temperature. HA1 was visualized by an AF488-labeled, and Pinda by an AF594-labeled secondary antibody, actin with phalloidin-AF647. This method enables to distinguish between extracellular virus particles and internalized particles. Images were acquired by confocal microscopy and total HA1 signal per cell area was quantified and normalized as in the virus binding assay.
To block dynamin-dependent endocytosis, control (All*Neg) cells were pretreated for 30 min in either dmso or 120 µM dynasore, followed by virus binding and uptake for 30 min in the presence of drug. Cells were fixed, stained and analyzed as above.
A549 cells were bound with TCID50 = 2.4×106/ml virus for 1 h at 4°C, washed twice in infection medium, and warmed to 37°C under the presence of 1 mM cycloheximide to block new protein synthesis. For time course experiments, cells were washed in PBS and fixed in 4% FA in PBS. For a single time point analysis, cells were washed and fixed at 1 h, where the acidification signal peaked. For detection of HA that has undergone acid-induced conformational change, cells were blocked, permeabilized and incubated with anti-A1 monoclonal antibody [58] (1∶2000) for 2 h at room temperature. Acidified HA was visualized by an AF488-labeled secondary antibody and nuclei were stained with DRAQ5. Images were acquired by confocal microscopy or automated microscopy and total signal per nucleus was quantified as above. Cumulative A1 signal over a time course of up to 5 h (t0 = 0, t1 = 30, t2 = 60, t3 = 120, t4 = 210, t5 = 300 min), was obtained by calculating the area under the curve, from the equation {(t1)(y1)+(y1+y2)(t2t1)+…+(y4+y5)(t5−t4)}×0.5.
Binding and uptake of virus was performed as in the HA acidification assay, and the cells were washed and fixed 5 h post warming. Medium containing 1 mM cycloheximide was freshly replaced at 2.5 h post-internalization. Cells were fixed in 4% FA in PBS, blocked, permeabilized and stained for viral NP with the HB-65 monoclonal antibody (1∶10) for 2 h at room temperature. NP (incoming vRNPs) was visualized by an AF488-labeled secondary antibody and nuclei were stained with DRAQ5. Images were acquired by confocal microscopy, and the percentage of cells with nuclear NP signal was counted and normalized to control (All*Neg) cells.
A549 cells were bound with TCID50 = 1×106/ml virus for 30 min at 4°C and washed three times in infection medium. The virus was allowed to fuse in low pH medium (pH 5.0) for 2.5 min on a custom-cut metal plate warmed to 37°C in a water bath. After incubation, the cells were placed on ice, washed three times before incubation at 37°C in Stop medium (D-MEM, 50 mM HEPES pH 7.4, 20 mM NH4Cl) to block acidification of endosomes. The cells were washed and fixed at 12 h, permeabilized, and stained with HB-65 to detect NP. The fraction of infected cells were quantified as in the infection assay.
Cells grown in 96-well Matrix plates were pretreated for 30 min with 30 µM nocodazole, 50 nM taxol or dmso alone, followed by binding TCID50 = 6×105/ml virus for 30 min on ice. The cells were washed to remove unbound virus and warmed to 37°C in the presence of the drug. At given time points, drug-containing medium was washed out and replaced with Stop medium (D-MEM, 50 mM HEPES pH 7.4, 20 mM NH4Cl) to block acidification of endosomes. Cells were fixed at 12 h and analyzed as in the infection assay.
Cells were incubated with dmso or 5 µM trychostatin A (TSA) in normal growth medium for 4 h. The drugs were removed by washing three times in infection medium, followed by an infection assay.
Labeling with R18 was performed as described previously [59]. X31 virus stocks were diluted to TCID50 = 1.5×107/ml in PBS and incubated in the dark with 20 µM R18 (Invitrogen) at room temperature for 1 h with continuous rocking. The mixture was filtered through a 0.2 µm filter to remove aggregated dye.
Confocal fluorescence microscopy was performed with a Zeiss LSM 510 Meta system setup. Cells were either fixed for 20 min in 4% FA in PBS at room temperature for 5 min in −20°C methanol. For indirect immunofluorescence, cells were blocked in blocking buffer (1% BSA in PBS) and permeablized in permeabilization buffer (0.1% saponin, 1% BSA in PBS), followed by incubation with antibodies in permeabilization buffer. Live imaging was done with the Zeiss LSM 510 or Visitech Spinning Disk confocal microscope, using 8-well chamber-slides (Nunc). For imaging, phenol red-free D-MEM (Invitrogen) was used. Image analysis was done with ImageJ or LSM Image Browser. Automated image acquisition of 96-well Matrix plates was performed with a 10× or 20× objective using a BD Pathway 855 Bioimager (BD Biosciences) or a MD Assay Development 2 (Molecular Devices).
A549 cells seeded in 96-well Matrix plates were fixed 72 h after depletion of host cell factors and fixed in 4% FA in PBS for 20 min. After blocking and permeabilization, cells were stained for LAMP1, actin (phalloidin-AF594) and DNA (Hoechst). Sixteen images (4×4) per well were acquired automatically for all three channels using a 20× objective. First, the images were segmented to identify individual cells and to determine their phenotypic properties using the CellProfiler program [60] with custom modifications. Second, supervised machine learning was used to automatically classify cell phenotypes and a regression model was applied to automatically predict the scattering index of the phenotypes. Finally, the effects of different siRNAs or drugs were determined. The scattering index of LAMP1-positive vesicles was set to zero for control (All*Neg) cells, and to 1.0 for control cells treated with 30 µM nocodazole for 1 h. In addition to the cell types described earlier, mitotic cells and segmentation errors were distinguished using a multi-parametric non-linear analysis with the Advanced Cell Classifier program (www.cellclassifier.org). For supervised machine learning, a mixed model of combining neural networks, support vector machines, and logistic classifiers was used.
EGF-AF555 (1 ng/ml) was added to control (All*Neg) A549 cells, HDAC8-depleted (siHDAC8), or control cells treated with 30 µM nocodazole for 45 min (All*Neg+nocod). Live imaging of endocytosed EGF was performed during 15–30 min after its addition with a Visitech Spinning Disk Confocal microscope using a 100× 1.4NA Oil DIC Plan-Apochromat objective. To detect cell boundaries and the nucleus, cells were transiently transfected with NES-2×EGFP 15 h before. For a single video, 2000 frames (Δt = 30.53 msec) of one z-slice were acquired. Endocytosed EGF trajectories were extracted from the recorded videos using the ImageJ implementation of the particle-tracking algorithm [30] with radius = 3 pixel, cutoff = 0, percentile = 0.2, max_displacement = 10 pixel, and linkrange = 1. Only trajectories longer than 50 frames were retained for analysis. Segments of directed motion where identified in the trajectories using the MATLAB trajectory segmentation toolbox as described [29]. Speeds of directed motion were computed by dividing their total length (sum of the lengths of all trajectory segments belonging to the same directed motion) by their duration. Due to noise in the images, this rather over-estimates the true speed of motion. All post-processing was done in MATLAB R2010b (The MathWorks, Inc.).
SYBR green quantitative real-time PCR (qRT-PCR) was performed using the LightCycler 480 SYBR Green I Master Mix (Roche) and the RotorGeneQ thermocycler (QIAGEN). Serial dilutions of the control sample cDNA were also submitted to real-time PCR to generate a standard curve, in order to calculate the efficiency of the primers. The samples were run in triplicates, whereas –RT samples, which do not contain cDNA, due to lack of reverse transcriptase (-RT) at the cDNA synthesis step were run in duplicates. A non-template control was added to assure contamination free PCR reagents. The threshold cycle was assigned by the rotor-gene Q serial software. Quantification of the results was done with the Pfaffl method [61] and the mRNA amounts of target genes were determined relative to the Ct-values of the control sample, and normalized to the reference gene (GAPDH).
The following primers were used; GAPDH (fwd, CTGTTGCTGTAGCCAAATTCGT; rev ACCCACTCCTCCACCTTTGA), HDAC1 (fwd, GGAAATCTATCGCCCTCACA; rev, AACAGGCCATCGAATACTGG), HDAC3 (fwd, ACGTGGGCAACTTCCACTAC; rev, GACTCTTGGTGAAGCCTTGC), HDAC8 (fwd, GGTGACGTGTCTGATGTTGG; rev, AGCTCCCAGCTGTAAGACCA).
A549 cells were depleted of HDACs in 96-well Matrix plates, and 48 h after depletion starved in serum-free medium for 24 h. EGF-AF647 (200 ng/ml) was bound for 30 min on ice. After washing, the cells were warmed in serum-containing medium at 37°C to allow internalization. At 15 min and 4 h post warming, cells were washed in acid buffer (0.1 M Glycine, pH 3) for 2 min on ice to remove non-internalized EGF, washed and fixed in 4% FA in PBS, followed by Hoechst staining. EGF and nuclei were imaged with the BD Pathway 855 Bioimager using a 10× objective. Intensity of EGF signal above background was quantified using ImageJ and the percentage of EGF degradation per nucleus at 4 h (compared to 15 min) was calculated.
1×105 A549 cells were seeded in 12-well plates and depleted of HDAC1 or 8. On the third day of depletion, cells were starved in serum-free medium for 4 h followed by binding transferrin-AF488 (5 µg/ml) for 30 min on ice. Zero and 10 min post warming to 37°C, cells were washed in acid buffer for 2 min on ice. Transferrin signal per cell was analyzed by FACS analysis. In brief, cells were washed with PBS, detached with 200 µl trypsin per well, and collected in 1 ml PBS. The cells were centrifuged at 1,500 rpm for 5 min at 4°C to remove trypsin and re-suspended and fixed in 200 µl 4% FA in PBS for 20 min. After removal of FA by centrifugation, cells were resuspended in 300 µl FACS buffer for analysis with a BD FACSCalibur flow cytometer.
A549 cells in 24-well plates were depleted of HDAC1, 8 for 72 h. The cells were washed in PBS, trypsinized and fixed in cold EtOH for 30 min at 4°C. After centrifugation of cells at 1,500 rpm for 5 min at 4°C, the supernatant was removed and the cells were stained with FACS buffer (PBS, 5 mM EDTA, 2% FCS, 0.02% NaN3) containing 2.5 µM DRAQ5 for 15 min at room temperature. The cells were washed by centrifugation, resuspended in 250 µl FACS buffer for analysis with BD FACSCanto II flow cytometer. Cell cycle analysis was performed with FlowJo.
MTs were depolymerized by incubating A549 cells grown on a 24-well plate on a custom-cut metal plate for 30 min on ice. MTs were repolymerized for 90 sec by incubating in 37°C medium. The cells were fixed immediately in cold methanol for 5 min. Cells were detected for MT asters by indirect immunofluorescence with anti-α-tubulin or anti-EB1 (microtubule plus-end binding protein 1) antibodies.
A549 cells were fixed in cold methanol for 5 min and stained for γ-tubulin by indirect immunofluorescence and nuclei with DRAQ5. Confocal z-stack images were acquired with a 40× objective. The distance between γ-tubulin foci was measured on a maximal projection using ImageJ. Centrosomes were counted as split when the distance was more than 2 µm [52].
Signal intensity of centrosomal ninein was measured from confocal images using ImageJ. Centrosomes were also identified by costaining with γ-tubulin.
Constructs for EGFP-Rab7 and LAMP1-EGFP were provided by J. Gruenberg (University of Geneva, Geneva, Switzerland), NES-2×EGFP was provided by U. Greber (University of Zurich, Zurich, Switzerland) [62], HDAC8-Flag by Ed Seto (H. Lee Moffitt Cancer Center & Research Institute, Tampa, USA). LAMP1-mCherry plasmid was constructed by excising the EGFP cds of the LAMP1-EGFP plasmid with sites AgeI/BsrGI and replacing it with mCherry.
HB-65 was purchased from ATCC. The following antibodies were purchased: HDAC1, HDAC2, HDAC3 (obtained from BioVision), HDAC8, LAMP-1, C-Nap1, EB1 (Santa Cruz), giantin (Covance), EEA1 (Cell Signalling), γ-tubulin, Flag M2 (Sigma), acetylated α-tubulin, α-tubulin, ninein (Abcam). Stocks of nocodazole, taxol, dynasore, trychostatin A (obtained from Sigma), bafilomycin A (Calbiochem) were prepared in dmso (Calbiochem) and stored at −20°C. DRAQ5 was purchased from Biostatus Limited. Hoechst, R18, AF-labeled EGF, TF, WGA, and secondary antibodies were obtained from Invitrogen.
ATP6V1B2 (526); C-Nap1 (11190); EEA1 (8411); Giantin (2804); HDAC1 (3065); HDAC2 (3066); HDAC3 (8841); HDAC6 (10013); HDAC8 (55869); LAMP1 (3916); Ninein (51199); Rab7 (7879); Rootletin (9696); α-tubulin (7846); γ-tubulin (7283).
|
10.1371/journal.pntd.0004837 | A Proteomic Investigation of Hepatic Resistance to Ascaris in a Murine Model | The helminth Ascaris causes ascariasis in both humans and pigs. Humans, especially children, experience significant morbidity including respiratory complications, growth deficits and intestinal obstruction. Given that 800 million people worldwide are infected by Ascaris, this represents a significant global public health concern. The severity of the symptoms and associated morbidity are related to the parasite burden and not all hosts are infected equally. While the pathology of the disease has been extensively examined, our understanding of the molecular mechanisms underlying resistance and susceptibility to this nematode infection is poor. In order to investigate host differences associated with heavy and light parasite burden, an experimental murine model was developed utilising Ascaris-susceptible and -resistant mice strains, C57BL/6J and CBA/Ca, respectively, which experience differential burdens of migratory Ascaris larvae in the host lungs. Previous studies identified the liver as the site where this difference in susceptibility occurs. Using a label free quantitative proteomic approach, we analysed the hepatic proteomes of day four post infection C57BL/6J and CBA/Ca mice with and without Ascaris infection to identify proteins changes potentially linked to both resistance and susceptibility amongst the two strains, respectively. Over 3000 proteins were identified in total and clear intrinsic differences were elucidated between the two strains. These included a higher abundance of mitochondrial proteins, particularly those associated with the oxidative phosphorylation pathway and reactive oxygen species (ROS) production in the relatively resistant CBA/Ca mice. We hypothesise that the increased ROS levels associated with higher levels of mitochondrial activity results in a highly oxidative cellular environment that has a dramatic effect on the nematode’s ability to successfully sustain a parasitic association with its resistant host. Under infection, both strains had increased abundances in proteins associated with the oxidative phosphorylation pathway, as well as the tricarboxylic acid cycle, with respect to their controls, indicating a general stress response to Ascaris infection. Despite the early stage of infection, some immune-associated proteins were identified to be differentially abundant, providing a novel insight into the host response to Ascaris. In general, the susceptible C57BL/6J mice displayed higher abundances in immune-associated proteins, most likely signifying a more active nematode cohort with respect to their CBA/Ca counterparts. The complement component C8a and S100 proteins, S100a8 and S100a9, were highly differentially abundant in both infected strains, signifying a potential innate immune response and the importance of the complement pathway in defence against macroparasite infection. In addition, the signatures of an early adaptive immune response were observed through the presence of proteins, such as plastin-2 and dipeptidyl peptidase 1. A marked decrease in proteins associated with translation was also observed in both C57BL/6J and CBA/Ca mice under infection, indicative of either a general response to Ascaris or a modulatory effect by the nematode itself. Our research provides novel insights into the in vivo host-Ascaris relationship on the molecular level and provides new research perspectives in the development of Ascaris control and treatment strategies.
| Ascaris infection is a significant burden on the people who live in developing countries with infection being linked to poor hygiene and low socio-economic status. The parasite causes a range of symptoms, especially in children, which include both chronic morbidity, such as growth retardation, and acute outcomes, such as intestinal obstruction. Certain people tend to be more heavily infected than others, with those individuals experiencing worse morbidity. The understanding of the difference between susceptible and resistant people is an essential first step in the development of new therapies in order to eliminate this neglected parasitic disease. Using an established mouse model involving a susceptible and resistant strain, we aimed to gain insight into the host-Ascaris interaction at the hepatic interface and elucidate some of the molecular mechanisms potentially involved in resistance. A number of key intrinsic differences were determined between both strains including major differences in mitochondrial and ROS associated processes which may present the nematodes with differing oxidative conditions and explain the failure of the nematode to establish a successful parasitism in the resistant strain. In addition, we resolved signatures of the innate and early adaptive immune response and a major reduction in the proteins associated with translation in both strains under infection. Our findings need to be further explored, but could be the foundation for a better understanding of the mechanisms behind the differential parasite burden and in the future, potential new therapies for control.
| Ascariasis is an important, widespread geohelminth disease of humans and pigs [1]. Over 800 million people are estimated to be infected with the causative agent, the human roundworm, Ascaris lumbricoides [2] and the equivalent in pigs, Ascaris suum, is equally ubiquitious [3]. The impact on children is particularly severe, with chronic morbidity such as growth retardation and diminished cognitive development being exhibited [4]. Transmission is linked to poor disposal of human waste leading to extensive contamination of the environment with long-lived, resistant eggs that embryonate under appropriate conditions of temperature and moisture [4].
The intensity of worm burden in the intestine of the host is a major determinant of the severity of disease [5]. Furthermore, the number of macroparasites a host carries is fundamental to our understanding of helminth parasite epidemiology [5]. Since the seminal work of Crofton [6], that characterized the frequency distribution of macroparasites as clumped or overdispersed, the ubiquity of what is now known as aggregation, has been demonstrated for a wide range of host-macroparasite systems [7]. Longitudinal field-based studies that measured the patterns of helminth re-infection in individual patients after the provision of anthelmintic treatment identified the phenomenon known as predisposition [8], whereby individuals demonstrated a degree of consistency in their patterns of re-infection and this was demonstrated from a range of endemic regions and geohelminth species [5]. Predisposition was also maintained over multiple rounds of chemotherapeutic treatment [9] and many years after a single round of treatment [10]. Aggregation and predisposition were also successfully modelled in outbred pigs infected with A. suum and were described as analogous to those of A. lumbricoides in humans [11].
The basis of this predisposition remains unknown [12] although it has been predicted that both exposure and host susceptibility are likely to influence the observed epidemiological patterns [5]. However, unravelling the relative contributions to aggregation and predisposition and hence susceptibility/resistance remains challenging for both ethical and logistical reasons [13]. As outlined by Keymer and Pagal [14], experimental manipulation utilising appropriate animal models is desirable, in tandem with human studies under field conditions, in order to study the multiple factors likely to be involved in predisposition.
Ascaris is a parasite that not only exists as an adult worm in the host intestine but also has a migratory pathway undertaken by its larvae, known as the hepato-trachaeal migration [15]. Symptoms occur during larval migration due to tissue damage [16] and the resultant pathology has been documented in the liver of both humans [17–19] and pigs [20–24]. Loeffler [25] described a transient or seasonal syndrome of pulmonary infiltrates, mild to marked respiratory symptoms and peripheral eosinophilia that he subsequently attributed to larval Ascaris in the lungs, termed Loeffler’s syndrome [26]. Pulmonary symptoms can be severe and life-threatening [27] and have also been documented in pigs [28].
Despite its global prevalence and sheer numbers of individuals it infects, ascariasis remains a classic neglected disease [1] and part of the explanation for this neglect, is because pigs as animal models are costly and laborious and we lack the versatility of inbred strains. The majority of model organisms that have been infected with Ascaris are so-called abnormal hosts, whereby the parasite does not complete its life-cycle, but manifests itself as the early migratory phase [29]. Furthermore, in the vast majority of such experimental systems, the basis of resistance or susceptibility to infection has not been clearly established [29]. In contrast, a convenient and repeatable mouse model for exploring the susceptibility to Ascaris during the early phase of infection has been developed [30–32].
C57BL/6J are uniquely susceptible to infection with the porcine ascarid, A. suum, and larval burdens recovered from the lungs of this strain, markedly exceed those recovered from similarly infected strains of which the most resistant is the CBA/Ca mouse [30]. These two strains therefore represent opposite response phenotypes [5], mimicking the extremes of predisposition detected in humans and pigs.
Subsequently, this model was utilised to assess the significance of inflammatory processes within the murine lung. Such responses mirrored larval intensity and it was concluded that the pulmonary inflammatory immune response was not prominently involved in primary protection of mice to Ascaris infection in the lungs [31]. The lack of support for a pulmonary mechanism led to the suggestion that a hepatic/post-hepatic factor, which varies between C57BL/6J and CBA/Ca mice, may play a critical role in the successful migration through host tissues [31]. Evidence in the form of a differential histopathological change in the liver between the two strains was observed, whereby the resistant strain, CBA/Ca demonstrated an earlier intense inflammatory response coupled with a more rapid tissue repair in the liver [32].
C57BL/6 and CBA/Ca mice have also been used as model organisms to study susceptibility and resistance to other helminth and protist parasites. For example, in the case of Brugia malayi, the parasites were cleared more rapidly from the blood stream in CBA/Ca mice compared to C57BL/6J mice [33]. Studies involving Leishmania, demonstrated that CBA mice were able to control infection with Leishmania major, but not an infection with Leishmania amazonensis [34]. C57BL/6 mice conversely were found to be resistant to L. major infection, but were, as was the case for the CBA mice, susceptible to L. amazonensis [35]. Schistosoma mansoni caused low grade pathology in C57BL/6 mice, but a more aggravated pathology in their CBA counterparts [36]. In the case of Plasmodium berghei ANKA, both CBA and C57BL/6 were susceptible to the development of cerebral malaria [37]. In short, it is clear that both strains respond differently when challenged by a diversity of parasitic infections. However, depending on the parasite species, the mouse strains will exhibit a difference in resistance and susceptibility.
The aims of the present study were to investigate the differences in the liver proteomes between the two mouse strains, one deemed to be susceptible, the other resistant, with both control and infected groups within each strain. To address these aims we employed high throughput quantitative mass spectrometry (MS) which is routinely used to identify and quantify thousands of proteins from highly complex samples. We specifically used label-free quantitative (LFQ) mass spectrometry [38] on total protein extracts from the left liver lobes of C57BL/6J and CBA/Ca mice with and without Ascaris infection sampled on day 4 post-infection (p.i.). The hepatic lobe and time-point choice were based on the study of Dold et al. [32], who demonstrated relatively high and equivalent larval numbers in C57BL/6J and CBA/Ca left lobes at day 4 p.i., a time point at which they also observed the onset of a differential inflammatory response to infection between the two mouse strains. LFQ-based proteomics is increasingly being used to investigate human pathogens and parasites including Chlamydia trachomatis [39], Acinetobacter baumannii [40] and Plasmodium vivax- infected and uninfected erythrocytes [41] and is providing unprecedented insight into both specific and broad level interactions between the infectious agent and its host.
The samples obtained for this study were part of a project authorised by the Health Products Regulatory Authority (HPRA), the competent authority that regulates scientific animal research in Ireland in accordance with Directive 2010/63/EU and its Irish transposition, SI No 543 of 2012 (Project Authorisation ID AE19136/P008 ID; Case Reference 7015826). In addition, this project was ethically approved by the TCD Animal Research Ethics Committee (AREC) prior to HPRA submission and approval.
Both C57BL/6J and CBA/Ca mice (Harlan laboratories) were infected with 1000 embryonated ova of A. suum and euthanized using cervical dislocation on day four post infection (p.i.). At this time point, five animals were euthanised from all four groups: C57BL/6J infected, C57BL/6J control, CBA/Ca infected and CBA/Ca control [42]. The livers were extracted and each lobe was snap frozen separately using liquid nitrogen and stored at -80°C until required.
To confirm a differential larval burdens in susceptible and resistant strains, post-mortems of 5 infected mice from each strain were conducted on day 7 post-infection. Living Ascaris larvae were recovered from the lungs of each mouse by means of the modified Baermann technique [30]. A pellet of the isolated viable larvae was suspended in a 5 ml solution of 0.9% saline and 6% formalin. Prior to larval counts the 5 ml solutions were agitated to ensure a homogeneous distribution of larvae within the sample. Larval counts were recorded from lung samples by means of pipetting 2 mls of solution into the chamber of a nematode counting slide (Chalex corporation)[32]. The number of larvae in the gridded area, which represented 1 ml, was counted under X40 magnification. The number of larvae in a 1 ml solution was multiplied by the total volume in order to estimate the number of larvae in each lung sample.
The left lobes of day four p.i. were homogenized in 6M urea, 2M thiourea, supplemented with a protease inhibitor cocktail (Roche, Complete Mini). Samples were centrifuged for 5 min at 11,200 × g to pellet any cellular debris. The supernatant was then removed and quantified using the Qubit™ protein quantification system (Invitrogen), following the manufacturer’s instructions. Three independent biological replicates for each group were analysed in this study. 75 μg of each sample was precipitated using the 2D Clean-Up Kit (GE HealthCare), following the manufacturer’s instructions. The resulting protein pellet was resuspended in 6M urea, 2M thiourea, 0.1 M Tris-HCl, pH 8.0. 50mM ammonium bicarbonate was added to each sample and proteins were reduced with 0.5M dithiothreitol (DTT) at 56°C for 20 min and alkylated with 0.55M iodoacetamide (IAA) at room temperature for 15 min, in the dark. 1 μl of a 1% w/v solution of Protease Max Surfactant Trypsin Enhancer (Promega) and 1 μg of Sequence Grade Trypsin (Promega) was added to give a protein:trypsin ratio of 75:1. The protein/trypsin mixture was incubated at 37°C for 18 hours. Digestion was terminated by adding 1 μl of 100% trifluoroacetic acid (Sigma Aldrich) and incubation at room temperature for 5 min. Samples were centrifuged for 10 min at 13,000 × g and a volume equivalent to 40 μg of pre-digested protein was removed and purified for mass spectrometry using C18 Spin Columns (Pierce), following the manufacturer’s instructions. The eluted peptides were dried using a SpeedyVac concentrator (Thermo Scientific Savant DNA120) and resuspended in 2% v/v acetonitrile and 0.05% v/v trifluoroacetic acid (TFA). Samples were sonicated for 5 min to aid peptide resuspension followed by centrifugation for 5 min at 16,000 × g. The supernatant was removed and used for mass spectrometry.
1 μg of each digested sample was loaded onto a QExactive (ThermoFisher Scientific) high-resolution accurate mass spectrometer connected to a Dionex Ultimate 3000 (RSLCnano) chromatography system. The peptides were separated by a 2% to 40% gradient of acetonitrile on a Biobasic C18 Picofrit column (100mm length, 75mm ID), using a 120 minute reverse-phase gradient at a flow rate of 250nL min-1. All data were acquired with the mass spectrometer operating in automatic data dependent switching mode. A full MS scan at 140,000 resolution and a range of 300–1700 m/z was followed by an MS/MS scan, resolution 17,500 and a range of 200–2000 m/z, selecting the 15 most intense ions prior to MS/MS.
Protein identification and LFQ normalisation of MS/MS data was performed using MaxQuant v1.5.0.8 (http://www.maxquant.org) following the general procedures and settings outlined in [43]. The Andromeda search algorithm [44] incorporated in the MaxQuant software was used to correlate MS/MS data against the SWISS-PROT database for Mus musculus [45](16,773 entries, downloaded May 2015) and a contaminant sequence set provided by MaxQuant. The following search parameters were used: first search peptide tolerance of 20 ppm, second search peptide tolerance 4.5 ppm with cysteine carbamidomethylation as a fixed modification and N-acetylation of protein and oxidation of methionine as variable modifications and a maximum of two missed cleavage sites allowed. False Discovery Rates (FDR) were set to 1% for both peptides and proteins and the FDR was estimated following searches against a target-decoy database. LFQ intensities were calculated using the MaxLFQ algorithm [46] from razor and unique peptides with a minimum ratio count of two peptides across samples. Peptides with minimum length of seven amino acids were considered for identification and proteins were only considered identified when more than one unique peptide for each protein was observed.
Perseus v.1.5.0.8 (www.maxquant.org/) was used for data analysis, processing and visualisation. Normalised LFQ intensity values were used as the quantitative measurement of protein abundance for subsequent analysis. The data matrix was first filtered for the removal of contaminants and peptides identified by site. LFQ intensity values were log2 transformed [47] and each sample was assigned to its corresponding group (C57BL/6J control and infected; CBA/Ca control and infected). Proteins not found in two out of three replicates in at least one group were omitted from the analysis. A data-imputation step was conducted to replace missing values with values that simulate signals of low abundant proteins [48] chosen randomly from a distribution specified by a downshift of 2.6 times the mean standard deviation (SD) of all measured values and a width of 0.37 times this SD.
Two sample t-tests were performed for all relevant comparisons using a cut-off of p<0.05 on the post imputated dataset to identify statistically significant differentially abundant (SSDA) proteins. Volcano plots were generated in Perseus by plotting negative log p-values on the y-axis and log2 fold-change values on the x-axis for each pair-wise comparison to visualise changes in protein expression. The ‘categories’ function in Perseus was utilized to highlight and visualise the distribution of various pathways and processes on selected volcano plots. Normalised intensity values were used for a principal component analysis (PCA). Exclusively expressed proteins (those that were uniquely expressed or completely absent in one group) were identified from the pre-imputation dataset and included in subsequent analyses. Hierarchical clustering was performed on Z-score normalised intensity values for all differentially abundant proteins by clustering both samples and proteins using Euclidean distance and complete linkage.
Gene ontology (GO) mapping was also performed in Perseus using the UniProt gene ID for all identified proteins to query the Perseus annotation file (downloaded January 2015) and extract terms for biological process, molecular function, Kyoto Encyclopaedia of Genes and Genomes (KEGG) name, KEGG pathway, protein family (pfam) and InterPro. GO and KEGG term enrichment analysis was performed on the major protein clusters identified by hierarchical clustering using a Fisher’s exact test (a Benjamini-Hochberg corrected FDR of2%) for enrichment in Uniprot Keywords, gene ontology biological process (GOBP), gene ontology cellular component (GOCC) and KEGG (FDR <2%). The Search Tool for the Retrieval of INteracting Genes/Proteins (STRING) [49] v10 (http://string-db.org/) was used to map known and predicted protein:protein interactions. UniProt gene lists (extracted from Perseus) were inputted and analysed in STRING using the medium to high confidence (0.5–0.7) setting to produce interactive protein networks for each group in all comparisons. GO term enrichment analyses for biological process, molecular function and cellular compartment were then conducted to identify potential pathways and processes that warranted further analysis. Such pathways were examined using the KEGG pathway analysis (http://www.kegg.jp/kegg/tool/map_pathway2.html) [50, 51], using the ‘KEGG Mapper—Search&Color Pathway’ tool. The equivalent KEGG identifiers were obtained using the UniProt ‘Retrieve/ID mapping’ function (http://www.uniprot.org/uploadlists/) with the organism set to M. musculus (mmu). Retrieved KEGG IDs were used to identify the most represented pathways. The MS proteomics data and MaxQuant search output files have been deposited to the ProteomeXchange Consortium [52] via the PRIDE partner repository with the dataset identifier PXD003555.
The mean larval burden in the lungs of the C57BL/6J mice (n = 5) was 188 ±SD 78.7 and the mean larval burden for CBA/Ca mice (n = 5) was 12 ±SD 8.5 confirming the resistant and susceptible phenotype in the mice used in our experiment.
LFQ MS was performed on three replicates originating from the liver left lobes of C57BL/6J and CBA/Ca mice with and without Ascaris infection. 3,145 proteins were identified initially of which 2,307 remained after filtering and processing (S1 Dataset). Proteins, exclusive to or absent from an individual sample were included in subsequent analyses as proteins of significant interest (S1 Table). A principal component analysis (PCA) performed on all filtered proteins (Fig 1A) distinguished the C57BL/6J and CBA/Ca samples indicating a clear intrinsic difference between the two strains. Although the C57BL/6J infected and controls are well resolved there is some overlap within CBA/Ca infected and control groups, indicating that the response to Ascaris in C57BL/6J is perhaps more pronounced.
Hierarchical clustering of z-score normalised intensity values for all significantly differentially abundant proteins (n = 194) resolved the three replicates of each sample group (Fig 1B). In addition, one minor and four major protein clusters were identified: the latter comprising C57BL/6J and CBA/Ca abundant proteins (Cluster A), C57BL/6J control and infected proteins (Cluster B); predominantly C57BL/6J and CBA/Ca infected abundant proteins (Cluster D) and CBA/Ca control and infected abundant proteins (Cluster E). GO and KEGG term enrichment analysis identified key biological processes enriched within each cluster (S2 Table; summarised in Fig 1C) and included translation and nucleosome (Cluster A); lysosome and glutathione metabolism (Cluster B); cytoskeleton, regulation of response to stress and extracellular vesicular exosome (Cluster D) and mitochondrial part and oxidation-reduction process (Cluster E).
Performing two sample t-tests resulted in the identification of SSDA proteins both between and within strains (S2 Dataset). These differentially abundant proteins, together with proteins uniquely expressed in each group, were analysed using the interaction network analysis software, STRING, to identify biological pathways and processes over-represented in a particular group. Biological processes and pathways identified by STRING were investigated further for their representation within the differentially abundant protein dataset and displayed on volcano plots using the ‘categories’ function in Perseus to highlight proteins involved in selected biological processes.
Two sample t-tests (p<0.05) on the post-imputation dataset identified 479 SSDA proteins between C57BL/6J infected and control with the log2 fold change ranging from -8.2 to 7.2 (Table 1). The 20 most differentially abundant proteins for both groups are displayed in Fig 2A. The most abundant proteins in C57BL/6J controls were histone H1.4, tyrosine aminotransferase, and histone H1.1. In comparison, for C57BL/6J infected samples, the most abundant proteins consisted of two S100 proteins (protein S100-A8 and protein S100-A9) and cytochrome c oxidase subunit 7C protein. STRING analysis identified enrichment of terms associated with the proteasome, the mitochondrion, RNA splicing and actin cytoskeleton in C57BL/6J infected individuals (Fig 3A) whereas enriched terms were identified for translation in the C57BL/6J controls (Fig 3B).
Two sample t-tests on the post-imputation dataset identified 193 SSDA proteins (log2 fold change range: -6.1 to 5.9; Table 1) with the 20 most differentially abundant proteins displayed in Fig 2B. The most abundant proteins in CBA/Ca infected mice were: protein S100-A8, seminal vesicle secretory protein 4, and dipeptidyl peptidase 1. For CBA/Ca control mice the most SSDA proteins included: 7-alpha-hydroxycholest-4-en-3-one 12-alpha-hydroxylase, costars family protein ABRACL, and collagen alpha-1(I) chain. STRING analysis revealed that CBA/Ca infected mice had an overrepresentation of spliceosome and actin cytoskeleton proteins in comparison to control samples (Fig 3C). As observed in the C57BL/6J mice, terms associated with translational proteins were enriched in addition to RNA processing and glutathione metabolism in CBA/Ca control mice with respect to their infected counterparts (Fig 3D).
Two sample t-tests (p<0.05) on the post-imputation dataset identified 354 SSDA proteins respectively and the Log2 fold changes ranged from -6.7 to 13.6 (Table 1). The top 20 most SSDA proteins are shown in Fig 2C. The most abundantly expressed proteins observed in CBA/Ca control samples are haemoglobin subunit beta-2, S-methylmethionine, and putative hydroxypyruvate isomerase. The log2 fold change for haemoglobin subunit beta-2 was the highest observed in the dataset, being 13.6. For C57BL/6J controls, the most abundantly expressed proteins are aldehyde oxidase 3, ADP-ribosylation factor 5, and H-2 class I histocompatibility antigen K-B alpha chain. Two clusters were observed in the CBA/Ca control samples representing ribosomal and oxidative phosphorylation biological processes (S1A Fig). However STRING analysis failed to identify any well populated clusters in C57BL/6J compared to CBA/Ca control (S1B Fig).
Two sample t-tests (p<0.05) on the post-imputation dataset identified 410 SSDA proteins, respectively and the log2 fold changes ranged from -8.1 to 9.2 (Table 1). The 20 most SSDA proteins are displayed in Fig 2D. The most abundant proteins in CBA/Ca infected compared to C57BL/6J were found to be: putative hydroxypyruvate isomerase, S-methylmethionine, haemoglobin subunit beta-2. The most abundant proteins present in C57BL/6J infected compared to their CBA/Ca counter parts were: murinoglobulin-2, aldehyde oxidase 3, and chitinase-3-like protein 3. STRING analysis resolved clusters with enriched terms for oxidative phosphorylation and translation in CBA/Ca infected mice (S1C Fig). Terms associated with the glutathione metabolism are enriched in C57BL/6J infected mice with respect to CBA/Ca infected mice (S1D Fig).
Proteins involved in the processes and pathways of interest (identified in Perseus and STRING) were displayed on volcano plots arising from the 2-way t-tests to give an indication of expression profile for all significant and non-significant representatives. The five processes that were consistently overrepresented and/or differentially abundant and selected for further analysis were mitochondria, electron transfer chain, immune system, ribosome and iron ion binding (Table 1). KEGG analysis was also used to provide a protein centric view for selected pathways (S2–S4 Figs).
Mitochondrial associated proteins were consistently differentially abundant among all samples. Of the 2,307 proteins identified and reported here, 324 were associated with the mitochondrion. A higher abundance of mitochondrial proteins was generally observed in CBA/Ca mice both with and without infection (Table 1; Fig 1C) in comparison to C57BL/6J mice, although an increase in mitochondrial protein abundances was observed for both strains under infection (Fig 2). GO analysis in Perseus resolved 80 oxidative phosphorylation (OXPHOS) terms, including electron transfer chain (ETC) complex I to V proteins, the most differentially abundant of which were cytochrome c oxidases (cox7c and cox7a2). CBA/Ca control and infected mice had a higher abundance of proteins involved in the OXPHOS pathway, when compared to C57BL/6J (S2A Fig). Although there was a significant upregulation of OXPHOS proteins in C57BL/6J infected mice when compared to their controls, the extent of increase was less than that for CBA/Ca. KEGG analysis also indicated that there was a generally lower abundance of ETC proteins in C57BL/6J infected mice, compared to their CBA/Ca counterparts. However Complex I proteins, were less abundant in C57BL/6J infected mice, when compared to their controls (S2A Fig).
Proteins of the tricarboxylic acid (TCA) cycle were intrinsically more abundant in CBA/Ca control mice, when compared to their C57BL/6J counterparts. This difference between the two strains was observed also under infection. When infected, both mouse strains had increased abundance of these proteins. However, C57BL/6J infected mice had higher abundances for the entire TCA cycle (S3 Fig).
A clear change in the abundance of proteins involved in the translational process during infection was observed. Ribosomal proteins were more abundant in both CBA/Ca controls and infected with respect to their C57BL/6J counterparts (S1A and S1C Fig). Under infection however, both CBA/Ca and C57BL/6J mice had significantly lower abundances for many ribosomal proteins (Fig 2). KEGG analysis further confirmed this finding demonstrating that CBA/Ca had more intrinsically abundant ribosomal proteins than C57BL/6J, a difference that became more pronounced under infection (S2B Fig).
GO and KEGG term enrichment analysis indicated that there were very few immune associated pathways respondent to infection. Of these, 20 immune associated proteins were differentially abundant (relative fold difference > 1.5) between infected and control samples (Table 2). Eight proteins: s100-A8; s100-A9; dipeptidyl peptidase 1; coronin-1A; galectin; vitronectin; moesin and plasminogen were more abundantly expressed in both C57BL/6J and CBA/Ca infected mice with respect to their uninfected counterparts indicating a potential conservation of response to Ascaris in both mice strains. The majority of SSDA immune proteins were observed in C57BL/6J Ascaris-infected mice (Table 2). A number of immune proteins demonstrated higher intrinsic abundances in C57BL/6J compared to CBA/Ca (in both infected and uninfected comparisons) and included dipeptidyl peptidase 1, complement C4-B, galectin-3 and the MHC class I molecules H2-D1 and H2-K1 (Fig 2C and 2D).
16 proteins associated with the biological process term “response to stress” were enriched in the shared “response to infection” cluster (Cluster D on Fig 1B). Many of these proteins are also associated with the cytoskeleton and cytoskeletal organisation, terms that were also shown to be significantly enriched within this “infected” cluster (S2 Table; Fig 1C).
Glutathione is known to reduce the reactive oxygen species (ROS), H2O2 to H2O. Because of its ROS reducing capacities, it was investigated in this study. KEGG analysis identified an increase in glutathione metabolic proteins of C57BL/6J control samples, compared to their CBA/Ca counterparts (S4 Fig). This inherent difference was also observed when comparing both strains under infection. Additionally, there was an increase in abundance of glutathione metabolism proteins in C57BL/6J mice when infected, compared to their controls. This increase was less pronounced in CBA/Ca mice.
With 800 million people worldwide infected with Ascaris [2], improving our understanding of the underlying biological pathways involved with susceptibility to infection is required. Currently, ascariasis control consists primarily of the provision of anthelmintic drugs, such as albendazole or mebendazole [53], however, reinfection can occur rapidly after treatment [54]. Furthermore, the presence of adult worms in the intestine contributes to significant chronic morbidity and, in some cases, more acute complications [4]. Therefore, investigation of the role of the liver—a key organ in Ascaris larval migration, and likely attrition and consequent establishment of adult worms in the intestine could prove fruitful [29]. The liver has previously been identified as the organ of interest to investigate susceptibility and resistance to Ascaris infection [31, 32]. To understand the difference in susceptibility and resistance to heavy infection, we have performed a proteomic analysis of the liver of two different mouse strains, C57BL/6J and CBA/Ca, models for susceptibility (heavy larval burden) and resistance (light larval burden), respectively. The difference in susceptibility was confirmed by assessing larval burdens in the lungs C57BL/6J and CBA/Ca mice on day 7 p.i. Previous studies have demonstrated a distinct, repeatable difference in A. suum larval burden in the lungs of susceptible C57BL/6J and resistant CBA/Ca inbred mice on day 7 post-infection [30, 32, 55]. Although the numbers observed were very similar to those described in these studies, a clearer divergence in larval numbers was observed between both strains in the present experiment.
Using quantitative mass spectrometry, several pathways and individual proteins, which may contribute to these observed differences in the burden of infection, were identified. A clear difference in the proteomes of both strains was observed (Fig 1A) and intrinsic differences identified between the inbred strains, unrelated to Ascaris infection, which may underpin susceptibility and resistance, and are also significant for other branches of biomedical research using these particular mice strains.
Parasite infection can place additional stresses on the physiological processes of the host. One of the major findings of this study was the higher abundance of proteins involved in mitochondrial processes in the CBA/Ca mouse strain, both with and without infection. These processes include OXPHOS and the TCA cycle, which were significantly enriched in the CBA/Ca proteome in comparison to their C57BL/6J counterparts. These intrinsic differences may represent a significant difference between both strains in terms of the susceptibility to Ascaris infection. The mitochondria are the main source of ROS, which are produced as by-products in the OXPHOS pathway as part of cellular respiration [56, 57]. ROS have a vast range of biological functions and play a role in defence against parasites and pathogens, apoptosis, cell survival, cell growth, proliferation, differentiation and many other signalling pathways [57, 58]. Higher ROS production in CBA/Ca mice may generate a prohibitive oxidative environment for Ascaris within the murine liver and explain the decrease in observable migration to the lungs by Ascaris from the CBA/Ca strain [32]. Our results support previous findings that CBA/Ca mice have higher superoxide production and have a higher tolerance to ROS than their C57BL/6 counterparts [59, 60]. Staecker et al. [60] found that the mouse cochlea of the C57BL/6J strain showed an early-onset hearing loss pathology compared to CBA/Ca mice, potentially linked to increased levels of free radicals present in the cochlea. This increase was attributed to the adult hearing loss (Ahl) gene, first identified by Johnson et al. [61], which is thought to cause a decrease in protective antioxidant enzymes in C57BL/6J mice. Taken together, these studies suggest that CBA/Ca mice not only produce higher endogenous levels of ROS and protective antioxidants but can tolerate ROS levels that could potentially affect an invading parasite.
Under infection, both strains exhibited higher abundances of OXPHOS and TCA cycle proteins indicating a general stress response to Ascaris infection. Although the increase in abundance was generally more pronounced in C57BL/6J mice, when compared to their corresponding controls, infected CBA/Ca mice had higher overall abundances representing the higher endogenous levels of mitochondrial proteins in CBA/Ca control mice. Interestingly, C57BL/6J infected mice had lower abundances of complex I proteins specifically in comparison to their controls. Complex I is an important site of superoxide production (together with complex III) [62] and has been linked with a number of diseases spanning early childhood (Leigh disease) [63] to adulthood (Parkinson’s and Alzheimer’s disease) [62]. Additionally, quinone-binding inhibitors, which inhibit complex I, were shown to increase ROS production [62]. TCA cycle proteins showed a similar pattern of abundance changes amongst samples, with higher intrinsic abundances observed in CBA/Ca mice and increased abundances observed under infection. However, C57BL/6J infected mice displayed higher abundances of TCA-associated proteins under infection than CBA/Ca infected mice, compared to their controls. The TCA cycle takes place in the mitochondria [64] and results in the production of the reduced co-enzyme nicotinamide adenine dinucleotide hydride (NADH), which contributes electrons to the OXPHOS pathway. Thus, it seems that in response to infection both mice strains increase OXPHOS activity.
The differences in mitochondrial protein abundance and presumed ROS levels between the susceptible and resistant strains can be explained not only by intrinsic basal differences but also the extent of parasitic interaction within each strain. Based on the significant reduction in nematode numbers that reach the lungs in CBA/Ca mice [31] one can assume that the CBA/Ca-Ascaris interaction represents a far less compatible host-parasite interaction in comparison to the C57BL/6J-Ascaris interaction. Nematodes in C57BL/6J livers seem capable of successful establishment within their hosts and in doing so, modulate host defence pathways and responses to allow for completion of the parasite life-cycle. Ascaris secrete a plethora of proteins in their excretory/secretory fluid with potential immunomodulatory and antioxidant defence functions [65–67]. The first complete predicted secretome was generated during the Ascaris suum genome project identifying 775 predicted secretory proteins with a rich number of peptidases for the migration through host tissues, as well as products potentially involved in the evasion or modulation of host defences [68]. In addition, proteomic analyses of the excretory/secretory products of larval and adult Ascaris have characterised heat shock proteins, ABA-1, proteases, serpins, chitinases and a suite of individual proteins that have been implicated in immune evasion and modulation and have been suggested to play a role in parasite survival [69]. It is not unreasonable to postulate that Ascaris also secrete proteins that reduce host mitochondrial processes and ROS production specifically. Modulation of host ROS production has been widely reported for a number of taxonomically diverse parasites, including the liver fluke, Fasciola hepatica and the causative agent of Chagas disease, Trypanosoma cruzi [70–73]. In addition, antioxidant production in parasites has been linked to improved parasite survival [74, 75] and helminths are known to possess an array of ROS-reducing products, such as superoxide dismutase (SOD) [76] with A. suum also possessing catalase (CAT) [77], and peroxiredoxin (Prx) [76, 78], which offer the nematode mechanisms for ROS contention.
An intrinsic difference between the two mouse strains was also observed for glutathione metabolism, with C57BL/6J mice displaying a higher abundance in proteins involved in this pathway. Under infection, both strains displayed an increase in abundance for these proteins, with C57BL/6J mice presenting a more pronounced upregulation. Most mitochondria lack CAT, therefore, H2O2 conversion is maintained by glutathione (GSH) within these organelles [79]. H2O2 is reduced to H2O by glutathione peroxidase (GPx), using GSH [80] and this reaction produces oxidised glutathione (GSSG), which is then reduced back to GSH by glutathione reductase (GR), consuming one molecule of NADPH. This increase would need to be countered by increased antioxidants, such as GSH.
Our findings clearly indicate that mitochondrial- and ROS-associated proteins are more abundantly expressed in both strains under infection, indicative of the host liver being within a state of stress. In C57BL/6J mice, these increases are less pronounced than in CBA/Ca mice indicating that although a response to Ascaris is mounted, the levels achieved in C57BL/6Js are not equivalent to those in CBA/Ca mice. In addition, because ROS levels are presumably reduced in C57BL/6J mice, the nematodes entering the liver have a less cytotoxic environment to contend with and can potentially mount immuno-modulatory and antioxidative perturbations on their host.
The host immune system represents an important obstacle to the completion of the parasite life-cycle. To circumvent this threat, parasites have evolved strategies to evade or modulate aspects of the host immune response. In relation to Ascaris, previous studies examining the host immune response within the liver have focussed on the role of pro-inflammatory cytokines in co-ordinating defences against the parasite during its migratory phase [31, 81, 82]. Within the present study, gene ontology and KEGG pathway analyses failed to identify any clear process level immune responses to Ascaris infection and few differentially abundant immune proteins were identified either between strains or in response to the presence of Ascaris. The low number of immune-responsive proteins may be explained in two ways. Firstly, the liver is an immune tolerant organ and the hepatic system circumvents the lymphatic system. This ensures that no unwanted immune activation occurs in response to commensal pathogens present in the bowel, which can sometimes enter the blood stream after endothelial damage [83]. This system is thought to be exploited by other parasites, such as malaria-causing Plasmodium species, where the parasite enters the liver in the pre-erythrocytic stage, potentially to evade the host immune system [84]. Secondly, the experimental time point examined within the present study may be too early in the infection to observe an adaptive immune response, as the adaptive immune response takes 4–7 days to engage fully [85]. Of the identified differentially abundant immune proteins, broadly three groups can be defined (see Table 2): proteins observed in both mouse strains, protein abundance only observed in C57BL/6J mice, and proteins observed in CBA/Ca mice only. Overall, there appeared to be more immune-associated proteins that had increased abundance in response to infection, which was to be expected, and C57BL/6J mice displayed a higher abundance in some immune-associated proteins.
The proteins observed to be more abundant in both strains signify a potential innate immune response through the identification of proteins, such as complement component C8a and S100 proteins, S100a8 and S100a9. The alpha polypeptide of complement 8 interacts with other complement proteins to form the membrane attack complex, a cell-killing structure [86]. The complex binds to the cell membrane of target cells, forming transmembrane channels resulting in cell lysis and subsequent death [86]. Although, complement recruitment can occur in response to tissue damage potentially associated with parasite infection, complement components have previously been implicated as factors in mediating the adherence of myeloid cells to nematode parasites, resulting in parasite death although the susceptibility amongst different nematode species is variable [87]. Secretory products released by nematodes, such as Brugia malayi and Trichinella spiralis, have been identified to inhibit the chemotaxis activities of complement component 5a, identifying the complement component as a target for immunomodulation by parasitic nematodes [88].
Proteins S100a8 and S100a9 were among the most abundantly expressed under infection in both mouse strains, and completely absent in the control samples (Fig 1C and 1D). The S100 group is involved in multiple cellular functions, including cellular contraction, motility, cell differentiation and calcium regulation [89]. S100a8 and S100a9 can exist as homodimers, but in the presence of calcium, they will form a heterodimer called calprotectin [90, 91] which is expressed mainly by granulocytes, monocytes, and macrophages [90]. These proteins were recently identified as danger-associated molecular patterns (DAMPs) [92] which have been shown to interact with toll-like receptors (TLRs), such as TLR4 [93]. It is thought that S100a8/S100a9 are actively released by cells sensing danger, rather than passively [90]. Calprotectin has been found to be a chemotactic factor for neutrophils and other mononuclear cells, with the exception of lymphocytes, in mice [94], further confirming its role in the immune system. S100a8 and S100a9 heterodimers are known to be important in Leishmania infection. S100a8 and S100a9 recruitment results in the elimination of Leishmania and knockout mice experiencing a more severe infection. Such proteins probably have a role in the recruitment of neutrophils to the site of infection [91]. Moreover, S100a8 and S100a9 primed macrophages were better at killing Leishmania than non-primed macrophages. In addition, Edgeworth et al. [95], using human biopsies, found that S100a8/S100a9 heterodimer secretion, by macrophages, onto adults of the filarial nematode, Onchocerca volvulus, was part of an early immune response—demonstrating a putative immune role against nematode parasites. A similarly early response may be signified by the presence of these proteins in infected C57BL/6J and CBA/Ca mice suggesting a conserved role in defence against parasitic nematodes. In short, our results suggest the presence of an innate immune response (see Table 2) through the presence of a complement protein (C8a) and several proteins involved in leukocyte chemotaxis (such as S100a8, S100a9, galactin-3). Additionally, there is evidence of an early adaptive immune response through the presence of proteins such as plastin-2 (Lcp1) and dipeptidyl peptidase 1 (Ctsc).
Given that C57BL/6J mice carry more active Ascaris than CBA/Ca mice, it was unsurprising to find that a higher number of immune-associated proteins were observed in the susceptible strain. Of the nineteen proteins with known immune function (Table 2) fifteen had higher fold changes in C57BL/6J compared to CBA/Ca mice or were exclusively more statistically abundant in C57BL/6J mice. The two most abundantly expressed proteins in the ‘C57BL/6J’ only group, were annexin A1 and neutrophil gelatinase-associated lipocalin, both of which are expressed as a cellular response to interleukin-1 (IL-1). It seems therefore that the restrictions encountered by the nematode larva in the resistant strain results in the reduced immune response in comparison to their susceptible counterparts.
Very little is known about how extracellular macroparasites, such as Ascaris, successfully modulate their host. The results of this study suggest that under infection, a down regulation of translational proteins occurs for both strains, in particular S6 ribosomal protein (part of the mammalian/mechanistic target of rapamycin (mTOR) pathway). Additionally, CBA/Ca control mice had, both with and without infection, higher ribosomal protein abundances compared to C57BL/6J mice. Under infection this intrinsic difference, however, became more pronounced. Pathogens are known to interact with translational proteins in their host [96]. Viruses, for example, are dependent on the host translational machinery for their own protein synthesis. Cells in turn can promote gene expression in response to the environmental situation, e.g. hypoxia, glucose deprivation [96] and pathogens, such as certain bacteria, are known to secrete effectors into the cytoplasm which in turn can reduce translation itself [97]. Inactivation of translation machinery is a documented strategy of certain parasites, such as the trypansome, Leishmania major [98]. Once the parasite has invaded a macrophage, the surface protease GP63 cleaves mTOR and mTOR complex 1 (mTORC1), which inhibits translational initiation [98]. It is reasonable to postulate that the widespread lower abundance of translation proteins may indicate direct modulation by Ascaris which is known to secrete a wide range of effector-like molecules [69]. However it must be acknowledged that the down regulation of translation may in fact signify a defensive strategy employed by the host, in response to damage caused by Ascaris, as is seen in bacteria-induced epithelial damage, which results in the triggering of signals that suppress host translation [96].
Aged erythrocytes are broken down by macrophages in the spleen and liver [99]; the liver is therefore an important organ in haemoglobin metabolism. Haemoglobin beta-2 (Hbb2) is the most abundantly expressed protein in CBA/Ca mice, and displayed relative fold changes of over 12,500 between CBA/CA and C57BL/6J controls and a fold chance of over 500 between CBA/CA and C57BL/6J infected mice. The haemoglobin of mice consists of two haemoglobin alpha (Hbba), one haemoglobin beta-1 (Hbb-1, major chain) and one Hbb-2 (minor chain) subunits. There are three different haplotypes for haemoglobin beta (Hbb): HbbS (single, βS), HbbD (diffuse, βD) and HbbP. C57BL/6J mice are homozygous for HbbS [100], whereas CBA mice are heterozygous for HbbD [100–102]. The βS haplotype has one reactive cysteine residue on position 93 (βCys93), whereas βD and βP have an extra reactive cysteine residue: βCys13 [100]. Interestingly, both βCys93 and βCys13 can be modified by GSSG, with βCys13 being more susceptible than βCys93 [100]. Furthermore, GSSG may be reduced to GSH at βCys13, without the need of GR nor NADPH [100]. Hempe et al. [100] postulate that the concentration of these cysteine sulfhydryl groups determines the availability of GSH for enzymatic reactions. Having a higher concentration of haemoglobin could therefore be a mechanism of CBA/Ca mice to establish their ROS tolerance (as observed in the previously mentioned radiation studies), as a higher haemoglobin concentration would coincide with more βCys93, and thus more chances for glutathione to be reduced. The allelic differences in haemoglobin are also thought to confer a differential ability of various mouse strains to contend with different parasites. For example, C57 mice (HbbS) are relatively resistant to Plasmodium infection compared to BALB/c mice (HbbD)[103].
Ascariasis is a debilitating disease affecting an estimated 800 million individuals globally. While the pathology of the disease has been extensively examined, our understanding of the molecular mechanisms underlying resistance and susceptibility to nematode infection is poorly understood. Here we provide a novel insight into the changes in a host liver proteome in response to Ascaris infection in vivo within two murine strains varying in their resistance and susceptibility to infection. Our results provide evidence for significant intrinsic differences in the hepatic proteomes of both mouse strains, potentially associated with resistance to Ascaris infection. Given the higher levels of proteins associated with ROS producing and processing and the general increased tolerance to ROS in CBA/Ca mice [59] in comparison to C57BL/6J mice, in particular, we hypothesise that higher ROS levels and the associated oxidative environment could be involved in the inhibition of Ascaris larval in CBA/Ca mice. Whether the intrinsic differences in mitochondrial protein abundances are due to different levels of mitochondrial biogenesis and number between the two strains has yet to be determined. Our research provides new insights into the intricacies and complexities of the host-parasite relationship of Ascaris. In addition, potential parasite modulation of translational processes by Ascaris were clearly evident in both strains. Our findings also provide a new understanding of previous studies that utilised these two mouse strains for experiments involving early-onset hearing loss, radiation exposure, and several other micro and macroparasite infections. Given our findings and the central role of the liver in the Ascaris migratory pathway, we suggest a potentially novel research direction to develop alternative preventative control strategies for Ascaris. It seems that the key determinant in murine resistance to Ascaris lies in highly oxidative conditions that presumably restricts and arrests successful larval migration within the CBA/Ca hepatic environment. Larval nematodes that enter the C57BL/6J liver seem free to continue their onward progression and through the secretion of their excretory/secretory compounds further sustain their parasitic lifecycle through manipulation and modulation of the host liver. So although defence responses are mounted in C57BL/6J mice it seems that Ascaris is already well-established in its attempts to contend with the host response. However, through the manipulation of hepatic ROS levels in the susceptible mouse strain, we may now be able to determine the importance of intrinsic ROS in conferring resistance to Ascaris. Although significant research is required to fully understand the determinants of resistance to Ascaris in our murine model, it does seem that we have at least been presented with new options in our pursuit of strategies to control a disease that affects an estimated one eighth of our planet’s population.
7-alpha-hydroxycholest-4-en-3-one 12-alpha-hydroxylase: O88962; ADP-ribosylation factor 5: P84084; Aldehyde oxidase 3: G3X982; Annexin A1: P10107; Chitinase-3-like protein 3: O35744; Collagen alpha-1(I) chain: P11087; Complement C4-B: P01029; Complement component C8 alpha chain: Q8K182; Coronin-1A: O89053; Costars family protein ABRACL: Q4KML4; Cytochrome c oxidase subunit 7A2, mitochondrial: P48771; Cytochrome c oxidase subunit 7C, mitochondrial: P17665; Dipeptidyl peptidase 1: P97821; Galectin-3: P16110; H-2 class I histocompatibility antigen K-B alpha chain: P01901; H-2 class I histocompatibility antigen, D-B alpha chain: P01899; H-2 class I histocompatibility antigen, K-B alpha chain: P01901; Haemoglobin alpha: P01942; Haemoglobin beta-1: P02088; Haemoglobin subunit beta-2: P02089; Histone H1.1: P43275; Histone H1.4: P43274; Interleukin-1 beta: P10749; Mitochondrial antiviral-signalling: Q8VCF0; Moesin: P26041; Murinoglobulin-2: P28666; Neutrophil gelatinase-associated lipocalin: P11672; Plasminogen: P20918; Plastin-2: Q61233; Protein S100-A8: P27005; Protein S100-A9: P31725; Putative hydroxypyruvate isomerase: Q8R1F5; S6 ribosomal protein: P62754; Seminal vesicle secretory protein 4: P18419; S-methylmethionine: Q91WS4; Tyrosine aminotransferase: Q8QZR1; Vitronectin: P29788. UniProt accession numbers are provided for all identified proteins in S1 and S2 Datasets.
|
10.1371/journal.pcbi.1006973 | Contextual influence on confidence judgments in human reinforcement learning | The ability to correctly estimate the probability of one’s choices being correct is fundamental to optimally re-evaluate previous choices or to arbitrate between different decision strategies. Experimental evidence nonetheless suggests that this metacognitive process—confidence judgment- is susceptible to numerous biases. Here, we investigate the effect of outcome valence (gains or losses) on confidence while participants learned stimulus-outcome associations by trial-and-error. In two experiments, participants were more confident in their choices when learning to seek gains compared to avoiding losses, despite equal difficulty and performance between those two contexts. Computational modelling revealed that this bias is driven by the context-value, a dynamically updated estimate of the average expected-value of choice options, necessary to explain equal performance in the gain and loss domain. The biasing effect of context-value on confidence, revealed here for the first time in a reinforcement-learning context, is therefore domain-general, with likely important functional consequences. We show that one such consequence emerges in volatile environments, where the (in)flexibility of individuals’ learning strategies differs when outcomes are framed as gains or losses. Despite apparent similar behavior- profound asymmetries might therefore exist between learning to avoid losses and learning to seek gains.
| In order to arbitrate between different decision strategies, as well as to inform future choices, a decision maker needs to estimate the probability of her choices being correct as precisely as possible. Surprisingly, this metacognitive operation, known as confidence judgment, has not been systematically investigated in the context of simple instrumental-learning tasks. Here, we assessed how confident individuals are in their choices when learning stimulus-outcome associations by trial-and-errors to maximize gains or to minimize losses. In two experiments, we show that individuals are more confident in their choices when learning to seek gains compared to avoiding losses, despite equal difficulty and performance between those two contexts. To simultaneously account for this pattern of choices and confidence judgments, we propose that individuals learn context-values, which approximate the average expected-value of choice options. We finally show that, in volatile environments, the biasing effect of context-value on confidence induces difference in learning flexibility when outcomes are framed as gains or losses.
| Simple reinforcement learning algorithms efficiently learn by trial-and-error to implement decision policies that maximize the occurrence of rewards and minimize the occurrence of punishments [1]. Such basic algorithms have been extensively used in experimental psychology, neuroscience and economics, and seem to parsimoniously account for a large amount of experimental data at the behavioral [2,3] and neuronal levels [4–6], as well as for learning abnormalities due to specific pharmacological manipulations [7,8] and neuro-psychiatric disorders [9]. Yet, ecological environments are inherently ever-changing, volatile and complex, such that organisms need to be able to flexibly adjust their learning strategies or to dynamically select among different learning strategies. These more sophisticated behaviors can be implemented by reinforcement-learning algorithms which compute different measures of environmental uncertainty [10–12] or strategy reliability [13–15].
To date, surprisingly little research has investigated if and how individuals engaged in learning by trial-and-error can actually compute such reliability estimates or related proxy variables. One way to experimentally assess such reliability estimates is via eliciting confidence judgments. Confidence is defined as a decision-maker’s estimation of her probability of being correct [16–18]. It results from a meta-cognitive operation [19], which according to recent studies could be performed automatically even when confidence judgments are not explicitly required [20]. In the context of predictive-inference tasks, individuals’ subjective confidence judgments have been shown to track the likelihood of decisions being correct in changing environments with remarkable accuracy [21,22]. Confidence could therefore be employed as a meta-cognitive variable that enables dynamic comparisons of different learning strategies and ultimately, decisions about whether to adjust learning strategies. Despite the recent surge of neural, computational and behavioral models of confidence estimation in decision-making and prediction tasks [17,23,24], how decision-makers estimate their confidence in their choices in reinforcement-learning contexts remains poorly investigated.
Crucially, although confidence judgments have been reported to accurately track decision-makers probability of being correct [18,22], they are also known to be subject to various biases. Notably, it appears that individuals are generally overconfident regarding their own performance [25], and that confidence judgments are modulated by numerous psychological factors including desirability biases [26], arousal [27], mood [28], and emotions [29] such as anxiety [30]. A recent study also revealed that monetary stakes can bias individuals’ confidence in their choice: irrespective of the choice correctness, the prospects of gains and losses bias confidence judgments upwards and downwards, respectively [31]. Given the potential importance of confidence in mediating learning strategies in changing environments, investigating confidence judgments and their biases in reinforcement-learning appears crucial.
Here, we simultaneously investigated the learning behavior and confidence estimations of individuals engaged in a reinforcement-learning task where the valence of the decision outcomes was systematically manipulated (gains versus losses) [8,32]. In this task, young adults have repeatedly been shown to perform equally well in gain-seeking and loss-avoidance learning contexts [32,33]. Yet, in line with the confidence bias induced by monetary stakes [31], we hypothesized that individuals would exhibit lower confidence in their choices while learning to avoid losses compared to seeking gains, despite similar performance and objectively equal difficulty between these two learning contexts. In addition, we anticipated that this bias would be generated by the learned context-value: this latent variable computed in some reinforcement-learning models–see e.g. [32,34]—approximates the overall expected value from available cues on a trial-by-trial basis, hence it could mimic the effects of the monetary stakes observed in [31]. Finally, conditional on those first hypotheses being confirmed, we hypothesized that the valence-induced confidence bias would modulate performance in volatile environments such as reversal tasks.
Our results, which confirm these hypotheses, first illustrate the generalizability of the confidence bias induced by the framing of incentives and outcomes as gains or losses. They also suggest that tracking confidence judgments in reinforcement-learning tasks can provide valuable insight into learning processes. Finally, they reveal that–despite apparent similar behavior- profound asymmetries might exist between learning to avoid losses and learning to seek gains [35], with likely important functional consequences.
We invited 18 participants to partake in our first experiment, and asked them to perform a probabilistic instrumental-learning task adapted from a previous study [32,33]. Participants repeatedly faced pairs of abstract symbols probabilistically associated with monetary outcomes. Symbol pairs were fixed, and associated with two levels of two outcome features, namely valence and information, in a 2×2 factorial design. Therefore, pairs of symbols could be associated with either gains or losses, and with partial or complete feedback (Methods and Fig 1A and 1B). Participants could maximize their payoffs by learning to choose the most advantageous symbol of each pair, i.e., the highest expected gain or the lowest expected loss. At each trial, after their choice but before receiving feedback, participants were also asked to report their confidence in their choice on a Likert scale from 0 to 10. Replicating previous findings [32,33], we found that participants correctly learned by trial-and-error to choose the best outcomes, (average correct choice rate 76.50 ± 2.38, t-test vs chance t17 = 11.16; P = 3.04×10−9), and that learning performance was marginally affected by the information factor, but unaffected by the outcome valence (ANOVA; main effect of information F1,17 = 4.28; P = 0.05; main effect of valence F1,17 = 1.04; P = 0.32; interaction F1,17 = 1.06; P = 0.32; Fig 1C). In other words, participants learned equally well to seek gains and to avoid losses. However, and in line with our hypothesis, the confidence ratings showed a very dissimilar pattern, as they were strongly influenced by the valence of outcomes (ANOVA; main effect of information F1,17 = 2.00; P = 0.17; main effect of valence F1,17 = 33.11; P = 2,33×10−11; interaction F1,17 = 7.58; P = 0.01; Fig 1D). Similar to the valence bias reported in perceptual decision-making tasks [31], these effects were driven by the fact that participants were more confident in the gain than in the loss condition when receiving partial feedback (6.86 ± 0.28 vs 4.66 ± 0.39; t-test t17 = 7.20; P = 1.50×10−6), and that this difference was still very significant although smaller in the complete feedback condition (6.58 ± 0.35 vs 5.24 ± 0.37; t-test t17 = 3.52; P = 2.65×10−3).
While the results of the first experiment are strongly suggestive of an effect of outcome valence on confidence in reinforcement learning, they cannot formally characterize a bias, as the notion of cognitive bias depends on the optimal reward-maximizing strategy [36]. In other terms: does this bias persist in situations where a truthful and accurate confidence report is associated with payoff maximization? We addressed this limitation of experiment 1 by directly incentivizing reports of confidence accuracy in our follow-up experiment. In this new experiment, confidence was formally defined as an estimation of the probability of being correct, and participants could maximize their chance to gain an additional monetary bonus (3×5 euros) by reporting their confidence as accurately and truthfully as possible on a rating scale ranging from 50% to 100% (Fig 2A). Specifically, confidence judgments were incentivized with a Matching Probability (MP) mechanism, a well-validated method from behavioral economics adapted from the Becker-DeGroot-Marschak auction [37,38]. Briefly, the MP mechanism considers participants’ confidence reports as bets on the correctness of their answers, and implements comparisons between these bets and random lotteries (Fig 3A). Under utility maximization assumptions, this guarantees that participants maximize their earnings by reporting their most precise and truthful confidence estimation [39,40]. This mechanism and the dominant strategy were explained to the 18 new participants before the experiment (Methods). In addition, because the neutral and non-informative outcome was more frequently experienced in the punishment partial than in the reward partial context in experiment 1, we replaced the neutral 0€ with a 10c gain or loss (see Methods and Fig 2B).
Replicating the results from the first experiment, we found that learning performance was affected by the information factor, but unaffected by the outcome valence (ANOVA; main effect of information F1,17 = 18.64; P = 4.67×10−4; main effect of valence F1,17 = 1.33×10−3; P = 0.97; interaction F1,17 = 0.77; P = 0.39; Fig 2C). Yet, the confidence ratings were again strongly influenced by the valence of outcomes (ANOVA; main effect of information F1,17 = 4.92; P = 0.04; main effect of valence F1,17 = 15.43; P = 1.08×10−3; interaction F1,17 = 4.25; P = 0.05; Fig 2D). Similar to Experiment 1, these effects were driven by the fact that participants were more confident in the gain than in the loss conditions (85.25 ± 1.23 vs 76.96 ± 2.38 (in %); t-test t17 = 3.93; P = 1.08×10−3).
Importantly, the changes in the experimental design also allowed us to estimate the bias in confidence judgments (sometimes called calibration, or “overconfidence”), by contrasting individuals’ average reported confidence (i.e. estimated probability of being correct) with their actual average probability of being correct. A positive bias therefore indicates that participants are overconfident reporting a higher probability of being correct than their objective average performance. Conversely, a negative bias indicates reporting a lower probability of being correct than the true average (“underconfidence”). These analyses revealed that participants are, in general marginally overconfident (4.07 ± 2.37 (%); t-test vs 0: t17 = 1.72; P = 0.10). This overconfidence, which was maximal in the gain-partial information condition (14.00 ± 3.86 (%)), was nonetheless mitigated by complete information (gain-complete: 2.53 ± 2.77 (%); t-test vs gain-partial: t17 = 2.72; P = 0.01) and losses (loss-partial: 1.56 ± 3.35 (%); t-test vs gain-partial: t17 = 2.76; P = 0.01). These effects of outcome valence and counterfactual feedback information on overconfidence appeared to be simply additive (ANOVA; main effect of information F1,17 = 8.40; P = 0.01; main effect of valence F1,17 = 7.03; P = 0.02; interaction F1,17 = 2.05; P = 0.17; Fig 3B).
While the results from our two first experiments provide convincing support for our hypotheses at the aggregate level (i.e. averaged choice rate and confidence ratings), we aimed at providing a finer description of the dynamical processes at stake, and therefore turned to computational modelling. Standard reinforcement-learning algorithms [1,3] typically give a satisfactory account of learning dynamics in stable contingency tasks as ours, but recent studies [32–34] have demonstrated that human learning is highly context (or reference)-dependent. The specific context-dependent reinforcement-learning algorithm proposed to account for learning and post-learning choices in the present task explicitly computes a context-value, which approximates the average expected value from a specific context [32]. We therefore hypothesized that this latent variable would capture the effects of monetary stakes observed in our previous study [31] and bias confidence. While this hypothesis about confidence will be explicitly tested in the in the next section, we first aim to demonstrate in the present section that context-dependent learning is necessary to explain choices.
Context dependency, by allowing neutral or moderately negative outcomes to be reframed as relative gains, provides an effective and parsimonious solution to the punishment-avoidance paradox. Briefly, this paradox stems from the notion that once a punishment is successfully avoided, the instrumental response is no longer reinforced. Reward learning (in which the extrinsic reinforcements are frequent, because they are sought) should therefore, theoretically, be more efficient than punishment learning (in which the extrinsic reinforcements are infrequent, because they are avoided). Yet, human subjects have repeatedly been shown to learn equally well in both domains, which paradoxically contradicts this prediction [41]. Reframing successful punishment-avoidance as a relative gain in context-dependent learning models solves this punishment-avoidance paradox.
Typically, implementing context dependency during learning generates “irrational” preferences in a transfer task performed after learning: participants express higher preference for mildly unfavorable items to objectively better items, because the former were initially paired with unfavorable items and hence acquired a higher “relative” subjective value [32–34]. As in these previous studies, the participants from our two experiments also performed the transfer task after the learning task (see Methods). The typical behavioral signature of context-dependent learning is a preference reversal in the complete information contexts, where symbols associated with small losses (L25) are preferred to symbols associated with small gains (G25), despite having objectively lower expected value [32–34]. This pattern was present in both of our experiments (% choices; experiment 1: L25: 59.52 ± 4.88, G25: 38.89 ± 5.04; t-test t17 = 2.46; P = 0.02; experiment 2: L25: 67.26 ± 5.35, G25: 28.37 ± 4.46; t-test t17 = 5.27; P = 6.24×10−5, see Fig 4A and 4B, middle panels).
To confirm these observations, we adopted a model-fitting and model-comparison approach, where a standard learning model (ABSOLUTE) was compared to a context-dependent learning model (RELATIVE) in its ability to account for the participants’ choices (Methods). Replicating previous findings [32,33], the context-dependent model provided the best and most parsimonious account of the data collected in our 2 experiments (Table 1), and a satisfactory account of choice patterns in both the learning (average likelihood per trial in experiment 1: 0.72 ± 0.03; in experiment 2: 0.72 ± 0.02; see Fig 4A and 4B, top panels) and transfer tasks (average likelihood per trial; experiment 1: 0.71 ± 0.02; experiment 1: 0.70 ± 0.02; see Fig 4A and 4B, middle panels). Please also note that the model estimated free-parameters (Table 2) are very similar to what was reported in the previous studies [32,33].
We next used latent variables from this computational model, along with other variables known to inform confidence judgments, to inform a descriptive model of confidence formation. We propose confidence to be under the influence of three main variables, entered as explanatory variables in linear mixed-effect regressions (FULL model–see Methods. Confidence Model). The first explanatory variable is choice difficulty, a feature captured in value-based choices by the absolute difference between the expected value of the two choice options [42,43], and indexed by the absolute difference between the option Q-values calculated by the RELATIVE model. The second explanatory variable is the confidence expressed at the preceding trial. Confidence judgments indeed exhibit a strong auto-correlation, even when they relate to decisions made in different tasks [44]. Note that in our task, where the stimuli are presented in an interleaved design, this last term captures the features of confidence which are transversal to different contexts such as aspecific drifts due to attention fluctuation and/or fatigue. The third and final explanatory variable is V(s), the approximation of the average expected-value of a pair of stimuli (i.e., the context value from the RELATIVE model) [32]. The context value, initialized at zero, gradually becomes positive in the reward-seeking conditions and negative in the punishment-avoidance conditions. This variable is central to our hypothesis that the decision frame (gain vs. loss) influences individuals’ estimated confidence about being correct [31]. Crucially, in the FULL model, all included explanatory variables were significant predictors of confidence ratings in both experiments (see Table 3). As a quality check, we also verified that the confidence ratings estimated under the FULL model satisfactorily capture the evolution of observed confidence ratings across the course of our experiments (Fig 4A and 4B, bottom panels).
On the contrary, when attempting to predict the trial-by-trial correct answers (i.e. performance) rather than confidence judgments with the same explanatory variables, the choice difficulty and the confidence expressed at the preceding trial were significant predictors in the two experiments, while the context value was not (Table 4). This again captures the idea that context value might bias confidence judgments above and beyond the variation in performance. Finally, because decision reaction times are known to be (negatively) correlated with subsequent confidence judgments—the more confident individuals are in their choices, the faster their decisions [20,42,45]-, we anticipated and verified that the same explanatory variables which are significant predictors of confidence also predict reaction times (although with opposite signs–see Table 4).
In this paper we investigated the effect of context-value on confidence during reinforcement-learning, by combining well-validated tasks: a probabilistic instrumental task with monetary gains and losses as outcomes [8,32,35], and two variants of a confidence elicitation task [40,47]: a free elicitation of confidence (experiment 1), and an incentivized elicitation of confidence called matching probability (experiment 2). Behavioral results from two experiments consistently show a clear dissociation of the effect of decision frame on learning performance and confidence judgments: while the valence of decision outcomes (gains vs. losses) had no effect on the learning performance, it significantly impacted subjects’ confidence in the very same choices. Specifically, learning to avoid losses generated lower confidence reports than learning to seek gains regardless of the confidence elicitation methods employed. These results extend prior findings [31], by demonstrating a biasing effect of incentive valence in a reinforcement learning context. They are also consistent with other decision-making studies reporting that positive psychological factors and states, such as joy or desirability, bias confidence upwards, while negative ones, such as worry, bias confidence downwards [26,28–30].
Based on the current design and results, we can rule out two potential explanation for the presence of this confidence bias. First, we used both a free confidence elicitation method (experiment 1) and an incentivized method (experiment 2) and clearly replicate our results across these two methods. This indicates that the confidence bias cannot be attributed to the confidence elicitation mechanism. This is also supported by the fact that the confidence bias is observed despite the incentives in the primary task (gain and loss) being orthogonalized from the ones used to elicit confidence judgments (always framed as a gain). Second, an interesting feature of the present experiments is that monetary outcomes are displayed after–rather than before- confidence judgments. At the time of decision and confidence judgments, the value of decision-contexts is implicitly inferred by participants and not explicitly displayed on the screen. Combined with the fact that loss and gain conditions were interleaved and that previous studies indicate that in a similar paradigm subjects remain largely unaware of the contextual manipulations [48], this suggests that the biasing effect of monetary outcomes demonstrated in previous reports [31] is not due to a simple framing effect, created by the display of monetary gains or losses prior to confidence judgments.
Contrary to our previous study [31], the current reinforcement-learning design provides little control on the effect of the experimental manipulations on choice reaction times. Our results show that, like confidence, reaction times are also biased by the context value. Given that some studies have suggested that reaction times could inform confidence judgments [45]–although this has recently been challenged [49]-, the observed confidence bias could be a by-product of a reaction-time bias. However, both our control analysis (Table 5) and our previous study [31] seem to rule out this interpretation and point toward an authentic confidence bias that is at least partially independent of reaction times.
We offer two interpretations for the observed effects of gains versus losses on confidence. In the first interpretation, we propose that loss prospects simply bias confidence downward. In the second interpretation, we propose that loss prospects improve confidence calibration over gain prospects, thereby correcting overconfidence. Following the first interpretation, the apparent improvement in confidence calibration observed in our study does not correspond to a confidence judgment improvement per se, but is a mere consequence of participants being overconfident in this task. Accordingly, in a hypothetical task where participants would be underconfident in the gain domain, while the loss prospects would aggravate this underconfidence under the first interpretation, they would improve confidence calibration (hence correct this underconfidence) under the second interpretation. Future research is needed to distinguish between the two potential mechanisms.
Regardless of the interpretation of the reported effects, we showed that confidence can be modelled as a simple linear and additive combination of three variables: previous confidence rating, choice difficulty and the context value inferred from the context-dependent reinforcement learning model. The critical contribution of the present study is the demonstration that confidence judgments are affected by the value of the decision-context, also referred to as context value. The context value is a subjective estimate of the average expected-value of a pair of stimuli: in our experimental paradigm, the context value is therefore neutral (equal to 0) at the beginning of learning, and gradually becomes positive in the reward-seeking conditions and negative in the punishment-avoidance conditions [32]. The fact that the context-value significantly contributes to confidence judgments therefore complements our model-free results showing that outcome valence impacts confidence, while embedding it in the learning dynamics. The fact that the context value is a significant predictor of confidence judgments also suggests that context-dependency in reinforcement learning is not only critical to account for choice patterns but also to account for additional behavioral manifestations, such as confidence judgments and reaction times. This result therefore provides additional support for the idea that context values are explicitly represented during learning [32]. Crucially, context-dependency has been shown to display locally adaptive (i.e. successful punishment-avoidance in the learning test) and globally maladaptive (i.e. irrational preferences in the transfer test) effects [48]. Whether the context-dependence of confidence judgments is adaptive or maladaptive remains to be elucidated and will require teasing apart the different interpretation of this effect discussed above.
Our findings are also consistent with a growing literature showing that in value-based decision-making, choice-difficulty, as proxyed by the absolute difference in expected subjective value between the available [50–52] is a significant predictor of confidence judgments [42,43]. Finally, the notion that confidence judgments expressed in preceding trials could inform confidence expressed in subsequent trails is relatively recent, but has received both theoretical and experimental support [44,53] and intuitively echoes findings of serial dependence in perceptual decisions [54]. In interleaved experimental designs like ours, successive trials pertain to different learning contexts. Therefore, the significant serial dependence of confidence judgments revealed by our analyses captures a temporal stability of confidence, which is context-independent. This result is highly consistent with the findings reported in Rahnev and colleagues (2015), which show that serial dependence in confidence can even be observed between different tasks.
In the present report, the modelling approach is strongly informed and constrained by our previous studies [31,32]. In this sense, the proposed models are solely meant to provide a parsimonious, descriptive account of the confidence bias observed in the reinforcement-learning task. We acknowledge that other models and model families could provide a better, mechanistic and/or principled account of both learning performance and confidence judgments [1,23,55,56].
Overall, our results outline the importance of investigating confidence biases in reinforcement-learning. As outlined in the introduction, most sophisticated RL algorithms assume representation of uncertainty and/or strategy reliability estimates, which allow them to flexibly adjust learning strategies or to dynamically select among different learning strategies. Yet, despite their fundamental importance in learning, these uncertainty estimates have, so far, mostly emerged as latent variables, computed from individuals’ choices under strong computational assumptions [13,14,57–61]. In the present paper we propose that confidence judgments could be a useful experimental proxy for such estimates in RL. Confidence judgments indeed possess important properties, which suggest that they might be an important variable mitigating learning and decision-making strategies. First, confidence judgments accurately track the probability of being correct in stochastic environments, integrating expected and unexpected uncertainty in a close-to-optimal fashion [21,22]. Second, subjective confidence in one’s choices impacts subsequent decision processes [62] and information seeking strategies [63]. Finally, confidence acts as a common currency and therefore can be used to trade-off between different strategies [64,65].
With this in mind, biases of confidence could have critical consequences on reinforcement learning and reveal important features about the flexibility of learning and decision-making processes in different contexts. Along those lines, our last experiment provides suggestive evidence that, in volatile environments, the valence-induced confidence bias induces differences in learning-flexibility between reward-seeking and loss-avoidance contexts. The fact that such behavioral manifestations were absent in previous experiments—where participants were explicitly told that symbol-outcome association probabilities were stable—suggests that confidence is linked to a higher level of strategic exploration, contingent on the representation of task and environment structure. See also [21] for a similar claim in a sequence learning task.
Considering evolutionary perspectives, future research should investigate whether lower confidence in the loss domain–as demonstrated in the present report—could play an adaptive function, e.g. by allowing rapid behavioral adjustments under threat.
All studies were approved by the local Ethics Committee of the Center for Research in Experimental Economics and political Decision-making (CREED), at the University of Amsterdam. All subjects gave informed consent prior to partaking in the study.
The subjects were recruited from the laboratory's participant database (www.creedexperiment.nl). A total of 84 subjects took part in this study: 18 took part in experiment 1 (8/10 M/F, age = 24.6±8.5), 18 in experiment 2 (8/10 MF, age = 24.6±4.3), and 48 in experiment 3 (26/22 M/F, age = 22.8±4). They were compensated with a combination of a base amount (5€), and additional gains and/or losses depending on their performance during the learning task: experiment 1 had an exchange rate of 1 (in-game euros = payout); experiments 2 and 3 had an exchange rate of 0.3 (in game euros = 0.3 payout euros). In addition, in experiments 2 and 3, three trials (one per session) were randomly selected for a potential 5 euros bonus each, attributed based on the confidence incentivization scheme (see below).
Power analysis were performed with GPower.3.1.9.2 [66]. The sample size for Experiments 1 and 2 was determined prior to the start of the experiments based on the effects of incentives on confidence judgments in Lebreton et al. (2018). Cohen’s d was estimated from a GLM d = .941 t23 = 4.61, P = 1.23e-4). For a similar within-subject design, a sample of N = 17 subjects was required to reach a power of 95% with a two-tailed one-sample t-test.
All model-free statistical analyses were performed using Matlab R2015a. All reported p-values correspond to two-sided tests. T-tests refer to a one sample t-test when comparing experimental data to a reference value (e.g. chance: 0.5), and paired t-tests when comparing experimental data from different conditions. ANOVA are repeated measure ANOVAs.
|
10.1371/journal.pgen.1006470 | Mutations in HYAL2, Encoding Hyaluronidase 2, Cause a Syndrome of Orofacial Clefting and Cor Triatriatum Sinister in Humans and Mice | Orofacial clefting is amongst the most common of birth defects, with both genetic and environmental components. Although numerous studies have been undertaken to investigate the complexities of the genetic etiology of this heterogeneous condition, this factor remains incompletely understood. Here, we describe mutations in the HYAL2 gene as a cause of syndromic orofacial clefting. HYAL2, encoding hyaluronidase 2, degrades extracellular hyaluronan, a critical component of the developing heart and palatal shelf matrix. Transfection assays demonstrated that the gene mutations destabilize the molecule, dramatically reducing HYAL2 protein levels. Consistent with the clinical presentation in affected individuals, investigations of Hyal2-/- mice revealed craniofacial abnormalities, including submucosal cleft palate. In addition, cor triatriatum sinister and hearing loss, identified in a proportion of Hyal2-/- mice, were also found as incompletely penetrant features in affected humans. Taken together our findings identify a new genetic cause of orofacial clefting in humans and mice, and define the first molecular cause of human cor triatriatum sinister, illustrating the fundamental importance of HYAL2 and hyaluronan turnover for normal human and mouse development.
| Cleft lip and palate (CLP) is one of the most common congenital anomalies worldwide. Most of the time the cause is unknown. Affected individuals have multiple problems including difficulties with feeding and speech and have to undergo several operations to correct the orofacial cleft. We describe mutations in an enzyme involved in the development of the palate and heart (Hyaluronidase 2, HYAL2) as the cause of a syndrome of CLP and a congenital heart anomaly (cor triatriatum sinister, CTS) in families from the Amish community and Northern Saudi Arabia of Arabic ethnicity. Mice who lack this enzyme show similar features. HYAL2 is involved in the metabolism of hyaluronic acid, which is important in many tissues of the body. By illustrating a lack of HYAL2 causes CLP and CTS we have demonstrated the importance of HYAL2 for the normal development of the palate and heart, improving our understanding of the complex mechanisms involved in the pathogenesis of CLP and CTS which may ultimately contribute to the development of novel therapies to treat these congenital anomalies.
| Orofacial clefts, which include cleft lip (CL) with or without cleft palate (CLP) and cleft palate alone (CP), collectively referred to as CL/P, are amongst the most common birth defects in all populations worldwide with many syndromic and non-syndromic forms described. The majority of cases are of unknown molecular cause [1]. The average incidence of CL/P is one in 700 newborns, with wide variability across racial and ethnic groups and socioeconomic status [2, 3]. The frequency of CL/P also differs by gender and laterality, there being a 2:1 male to female ratio for CLP, and a 1:2 male to female ratio for CP [2, 4]. Orofacial clefts arise from a failure of the intricate molecular and cellular processes that regulate bilateral fusion of the future lip and palate during craniofacial development [1]. They are clinically categorized dependent upon the absence (non-syndromic CL/P; 70% of cases) or presence (syndromic CL/P; 30% of cases) of additional congenital anomalies [5, 6]. For non-syndromic CL/P, 17 genetic risk loci have been defined and replicated with genome wide association studies worldwide [7–11]. As well as environmental influences, more recent evidence implicates non-coding or regulatory genomic regions in non-syndromic CL/P [12–14], perhaps explaining why re-sequencing of candidate genes or exome sequencing strategies rarely identify mutations in these cases [15]. For syndromic forms of CL/P, >300 different syndromes have been described in which CL/P is a prominent feature [16]. Congenital heart disease (CHD) is commonly associated with syndromic CL/P with a reported incidence of 13%-27% [15, 17–22]. Cor triatriatum is a rare congenital cardiac anomaly reported in 0.1–0.4% of individuals with congenital heart disease [23, 24]. It is characterized by division of either the left (cor triatriatum sinister) or less commonly the right atrium (cor triatriatum dexter) [25, 26]. To date, no molecular cause has been identified for cor triatriatum in humans. In the current study, we identify mutations in the gene HYAL2 as a new genetic cause of orofacial clefting in humans and mice and describe the first molecular basis for cor triatriatum sinister in humans. Our findings illustrate the fundamental importance of hyaluronan (HA) turnover for normal human and mouse development.
In order to map the chromosomal location of the disease gene we undertook a genome-wide SNP study in affected members of Family 1 assuming that a recessive founder mutation was responsible. This identified a single notable shared autozygous region of 10.18Mb on chromosome 3p21.31 (rs6441961-rs2201057; chr3:46310893–56499374 [hg38]) as likely to correspond to the disease locus. In support of this, multipoint linkage analysis with Simwalk2 [30] assuming autosomal recessive inheritance, full penetrance and a disease allele frequency of 0.0001, achieved a LODMAX = 10.37 corresponding to the autozygous interval. To identify the causative mutation, whole exome sequence analysis of a single affected individual (XII:3) was undertaken (Otogenetics Corporation). After filtering for call quality, potential pathogenicity, population frequency and localization within the candidate interval, we identified only a single likely deleterious variant located in the HYAL2 gene (HYAL2 chr3:g.50320047T>C [hg38]; NM_003773.4:c.443A>G; p.K148R; PolyPhen = probably damaging (1); SIFT = deleterious (0.03) MutationTaster = disease causing (probability 0.999997) Fig 1B). The variant was validated by Sanger dideoxy sequencing, which was also used to confirm co-segregation of the variant within the extended Amish pedigree (Fig 1A). Seven heterozygous carriers of the c.443A>G variant were identified in 532 Amish control chromosomes examined, indicating an allele frequency of approximately 0.013 in the population. The variant was not listed in the Exome Variant Server (NHLBI GO Exome Sequencing Project (ESP), Seattle, WA; http://evs.gs.washington.edu/EVS/), 1000 Genomes browser (http://browser.1000genomes.org/index.html) or the Exome Aggregation Consortium (ExAC) database (http://exacbroadinstitute.org/). Interestingly, HYAL2 had recently been identified as a candidate gene as a cause of short stature and facial dysmorphism in a study of multiplex consanguineous families [27]. This study undertook genetic studies in 33 heterogeneous families with dysmorphic and other systemic clinical manifestations, and identified a sequence alteration in HYAL2 as a possible cause of the clinical features in a single family comprising two affected individuals (Fig 1A, Family 2). Although inconclusive, when considered together with the genetic, clinical and functional data from the Amish study, this permits the robust genetic definition and clinical description of a new form of syndromic CLP. Genetic studies in Family 2 comprised whole genome SNP mapping which identified a single autozygous interval of 13.49Mb on chromosome 3p21 (rs7650433-rs9811393; chr3:40857673–54345775 [hg38]) shared by two affected family members, in which a missense HYAL2 variant (chr3:g.50319741G>A [hg38]; NM_003773.4:c.749C>T; p.P250L in Family 2; Fig 1A and 1B), absent in 817 ethnically matched individuals, was identified as the candidate genetic cause. This variant (rs781999115) was not listed in the Exome Variant Server or the 1000 Genomes browser but two carriers were reported on the ExAC database. The genomic autozygous region (chr3:46310893–54379802 [hg38]) common to both family cohorts is 8.0Mb and predicted to contain 270 RefSeq alignment-supported transcripts (S1 Fig).The p.K148R and p.P250L substitutions decrease protein stability.
HYAL2 encodes hyaluronidase 2, a membrane localized protein [33] with weak activity toward HA [34], which is an extracellular matrix glycosaminoglycan that is abundant during development. To assess the impact of p.K148R and p.P250L on the HYAL2 protein, the mutations were each introduced into a human HYAL2 expression construct and transiently transfected into MEFs deficient in HYAL2. Western blot analysis revealed a profound effect of the p.K148R and p.P250L mutations on protein levels, resulting in 11 and 20 fold reductions, respectively, compared to wild type (Fig 1C).
Consistent with the facial dysmorphism in the human affected individuals, Hyal2-/- mice exhibited a short broad nose, increased interorbitary space, and Wormian bones in the interfrontal suture [35]. To determine if the CL/P seen in the human affected individuals was also present in Hyal2-/- mice, we examined the time of pre-weaning lethality described to affect 2/3 of knockouts [35, 36]. Among 201 progeny of Hyal2+/- intercrosses followed from birth to weaning, 16 died at postnatal day (P)1, 16 died between P2-P9, and 16 Hyal2-/- mice survived (S1 Table). Assuming all pups that died were Hyal2-/-, 48 of 201 progeny (24%) would be Hyal2-/-, approximating the Mendelian frequency for an autosomal recessive condition. We also studied ninety E18.5 embryos from Hyal2+/- intercrosses and found two Hyal2-/- and one Hyal2+/- embryo(s) that had died at approximately E15.5. Taking these intrauterine deaths into account, the proportion of Hyal2+/+ mice still exceeded that of Hyal2-/- mice (30% vs 24%, S1 Table). Although this difference did not reach significance, it suggested reduced survival of Hyal2-/- and Hyal2+/- embryos during early gestation.
CL/P was a likely cause of death in Hyal2-/- mice at P1 and possibly between P2-P9, whereas heart defects were the more probable cause at E15.5. Although most deceased pups were cannibalized before their genotype could be verified, four found dead at P1, without milk in their stomachs, were identified as Hyal2-/-. Micro-CT studies of these pups showed an underdeveloped and underossified viscerocranium compared to littermate controls (Fig 2Ai-ii). Several central palate bones were underdeveloped and the vomer did not fuse centrally or form a head that articulated with the maxilla (Fig 2Aiii-vi). The ethmoid bone was nearly absent (Fig 2Ai-ii), or severely reduced in size (S2C and S2D Fig). This underdevelopment of the viscerocranium indicated a defect in intramembranous ossification that was more prominent in the anterior midline and varied in severity among affected skulls (S2A–S2D Fig).
Characterization of the palates of Hyal2-/- E18.5–19.5 embryos under a dissecting microscope revealed partial clefts and/or shortening of the secondary palate as well as abnormally formed rugae in most Hyal2-/- mice and a small number of Hyal2+/- mice (S3A–S3D Fig). Micro-CT confirmed reduced ossification and underdevelopment of the viscerocranial bones, particularly the vomer, consistent with submucosal cleft palate (SMCP) in 15 of 18 (83%; 95% CI 58–96%) Hyal2-/-, 3 of 54 (5.6%; 95% CI 1–16%) Hyal2+/-, and 0 of 23 (0%; 95% CI 0–18%) Hyal2+/+mice. Palatal malformation and clefting are likely to contribute significantly to the pre-weaning lethality in Hyal2-/- mice, indicating further phenotypical overlap with the human condition.
Histological studies of P1 mice confirmed that the viscerocranial bones of all Hyal2-/- mice were underdeveloped (S3E and S3F Fig, and arrows in N, O). Coronal sections showed the failed fusion between the epithelial surface of the vomeronasal organ and the dorsal side of the palate shelf (* in S3L and S3N Fig), leaving an average gap of 472 ± 50μm (n = 3) in Hyal2-/- palates that was absent in control palates (S3M and S3O Fig). HA levels were clearly increased in the Hyal2-/- tissues (S3J and S3N vs S3K and S3O Fig). Excess mesenchymal cells and their surrounding matrix, was also evident, particularly in the anterior head (* in S3H and S3I Fig).
Valvular thickening and atrial dilatation are found in all Hyal2-/- mice, although the severity of the phenotype varies [36]. Cor triatriatum sinister has been detected in 50% of Hyal2-/- mice ([37] and Fig 2B). It is conceivable that deaths in the progeny of Hyal2+/- intercrosses during embryogenesis and between P2-P9 observed are due to more severe (congenital) cardiac anomalies in Hyal2-/- mice. The cardiac findings in the five individuals with the HYAL2 K148R variant display significant similarities including one individual with cor triatriatum sinister and two further individuals with dilated coronary sinuses indicative of a PLSVC (see Table 1 and Fig 2C). This provides convincing evidence of the importance of HYAL2 for normal heart development in humans and mice [36, 37] and identified mutation of HYAL2 as the first described molecular mechanism for cor triatriatum sinister in humans.
Although conductive hearing loss was seen in most of the affected human subjects in this study, one affected individual was found to have profound sensorineural hearing loss. To assess hearing in Hyal2-/- mice (8–12 weeks of age), the auditory brainstem response (ABR) was evaluated (S4 Fig). Hyal2-/- mice exhibited a significantly higher threshold at all frequencies tested, (mean +/- 95% confidence interval for Hyal2-/- and Hyal2+/+ mice were 17.6 +/-7.16 and 5.00 +/-3.88dB respectively; Mann-Whitney U-test U = 1433, p<0.001), indicating hearing loss in 100% of Hyal2-/- mice. The ABR test parameters did not differentiate conductive and sensorineural hearing loss, but because these were performed on adult mice, severe craniofacial defects were excluded, although minor defects may be present. The presence of hearing loss in mice without severe palate defects suggests more studies are needed to determine whether HYAL2 has a direct role in the development and/or maintenance of normal hearing.
Previous studies of Hyal2-/- mice with a complete absence of HYAL2 activity have demonstrated levels of circulating HA that are 19 fold higher than that of control mice at 12 weeks of age and accumulation of HA in the heart and lung [36] as well as other tissues [37]. In the current study, the HA levels were clearly elevated throughout the nasopharynx and in the palate (S3J, S3K, S3N and S3O Fig). It remains unclear to which degree this reflects species differences, but this finding is consistent with our transfection studies which show residual HYAL2 is present. This level of HYAL2 may be sufficient for constitutive turnover, but inadequate during development when rapid turnover of high levels of HA in the provisional matrix is required.
The molecular basis of human CLP is incompletely understood, despite its frequent occurrence and associated morbidity. Our findings demonstrate that a deficiency of the HA-degrading enzyme HYAL2 is a novel cause of syndromic CLP in humans and SMCP in mice, and define the first molecular explanation for cor triatriatum sinister in humans. All but one affected individual exhibited orofacial clefting, and craniofacial dysmorphism was a consistent finding. Myopia and other ocular abnormalities, pectus excavatum, single palmar creases, cor triatriatum sinister, a persistent left superior vena cava and other cardiac features were variably present. Our studies show that Hyal2-/- mice universally exhibit facial dysmorphism, an underdevelopment of the viscerocranium and cardiac valve thickening as well as variably penetrant hearing loss, partial or SMCP (83%), cor triatriatum sinister (50%) and atrial dilatation (50%). The significant parity between the human and mouse phenotypes provides strong evidence that HYAL2-deficiency is the cause of these developmental abnormalities.
Our data indicate that the differences between the human and mouse phenotypes may be largely explained by the complete absence of HYAL2 in Hyal2-/- mice, while the destabilizing and deactivating p.K148R and p.P250L HYAL2 substitutions identified in human affected individuals lead primarily to profound reductions in HYAL2 levels, which may still permit limited residual functionality. In surviving Hyal2-/- mice progressive accumulation of HA in the circulation, pulmonary fibrosis, and cardiac dysfunction lead to premature heart failure [36], while the absence of elevated circulating HA in the affected Amish individuals may partially or completely protect them from these cardiopulmonary complications. The residual activity associated with the amino acid substitutions in human subjects likely supports ongoing constitutive turnover of HA, but is insufficient for the rapid and regulated turnover required during development. Further support for this is provided by our study of Hyal2+/- mice where a small number showed developmental abnormalities in the absence of elevated circulating HA [36]. However, this may be an overly simplistic interpretation as circulating HA levels are approximately 10–20 fold higher in the mouse, despite similar rates of turnover in mice and humans [38]. The identification of additional HYAL2-deficient individuals would undoubtedly provide a more complete understanding of the phenotypic spectrum associated with HYAL2-deficiency.
We identified SMCP as a likely cause of early pre-weaning lethality in Hyal2-/- mice. The underdeveloped viscerocranium likely weakens the palate, and in combination with the facial dysmorphism, presents significant feeding difficulties for Hyal2-/- suckling pups. Neurological abnormalities could also contribute to death, as the cribriform plate of the ethmoid is important in supporting the olfactory bulb and in preventing the leakage of cerebrospinal fluid into the nasal cavity. The bones of the viscerocranium have not yet been examined in affected individuals with the HYAL2 variant, but defects of the vomer are common in affected individuals with CLP [39]. The identification of developmental abnormalities in a small number of heterozygous (Hyal2+/-) mice suggests that haploinsufficiency of HYAL2 may also pose the risk of a similar outcome in humans. Consistent with the notion that altered HYAL2 levels may affect palatal development, decreased HA has previously been associated with increased risk for CP in the Tbx1-/- mouse [40], whereas increased HA has been associated with CP in Sim2-/- mice [41]. Thus, HYAL2 variants may be considered as candidates, even in the heterozygous form, to confer risk for these developmental abnormalities. However screening a cohort of non-syndromic CL/P individuals, provided little evidence for a major role for HYAL2 in this cohort (personal communication- P. Stanier). The cohort consisted of 380 individuals (176 with unilateral or bilateral cleft lip and palate, 38 with cleft lip only, 121 with cleft palate only and 46 with submucous cleft palate) of a mixture of ethnicities the majority of which (72%) were of white European origin recruited from the North East Thames region of the UK. Approximately 22% of cases were familial with affected first or second-degree relatives. While this study cohort is too small to draw definitive genotype-phenotype conclusions, this finding remains consistent with HYAL2 mutation being likely only associated with a rare, autosomal recessive syndrome.
Eloquent studies in mice deficient in HA synthesis (Has2-/-) demonstrated that high molecular mass HA was essential for epithelial to mesenchymal transition (EMT) in the developing heart, whereas HA oligosaccharides inhibited EMT and promoted angiogenesis [42–44]. HYAL2 could modulate the size of HA to inhibit EMT and promote differentiation, in which case its deficiency would result in increased levels of high molecular mass HA, an overproduction of mesenchymal cells, and a decrease in differentiation. Indeed, the thickened cardiac valves and excess mesenchymal cells in some parts of the head of Hyal2-/- mice, suggest that EMT is upregulated in Hyal2-/- mice, and that HYAL2 normally plays a role in inhibiting this process. In histological sections, encompassing the nasopharynx and secondary palate, there is an abundance of cartilage but a deficiency in bone. The deficient ossification in the skull of Hyal2-/- mice, and underdevelopment of the viscerocranium, suggests a failure in the development of osteoblasts, as has been seen in Tbx22-/- mice [45].
In addition to the direct effects of HA on signaling pathways, it is possible that an overabundance of HA in Hyal2-/- tissues leads to alterations in the gradients of morphogens or signal transduction factors, impacting development. Clearly further studies are needed to investigate this possibility. Treatment strategies for this developmental disorder deserve further research, as HA synthesis can be inhibited using 4-methylumelliferone, an FDA-approved drug in some parts of the world [46]. Further, commercial preparations of human hyaluronidase that function at a neutral pH could be used to remove extracellular HA; these enzymes are available for human use [47], and formulations that are stable in the circulation are under development [48]. With the Hyal2-/- mouse model already available, testing of such strategies may enable the development of a pre-natal treatment for this malformation syndrome. Taken together, our findings highlight hyaluronidase enzymes as playing a vital role in both human and mouse development, and in particular have revealed a previously unrecognized pathway involved in the pathogenesis of orofacial clefting, myopia and cor triatriatum sinister in humans, which may ultimately be amenable to treatment therapies.
Signed consent was obtained for all individuals and the study was approved by the institutional review board at the University of Arizona (reference 10-0050-01) and the Great Ormond Street Hospital and the Institute of Child Health Research Ethics Committee (REC No. 08H0713/46). Animals were euthanized by isoflurane overdose and embryos were collected for gross morphological or histological studies. All procedures were in compliance with the Canadian Council on Animal Care and followed a protocol approved by the University of Manitoba Animal Care Committee.
Blood samples for DNA extraction were obtained with informed consent from all participating family members (institutional review board-approved research protocols UoA:10-0050-10 and KFSHRC RAC#2080006). A genome-wide SNP screen was undertaken in the affected siblings (Family 1: XII:3; XIII:5; XII:7; XII:9; XII:12, and Family 2: V:3; VI:2) using Illumina HumanCytoSNP-12 v2.1 or Axiom SNP Chip arrays for autozygosity mapping of regions of >1Mb using AutoSNPa [49]. Whole exome sequencing of genomic DNA from Family 1 was performed at Otogenetics Corporation (Norcross, GA, USA) using the Agilent SureSelect Human All ExonV4 (51Mb) enrichment kit with a paired-end (2 × 100) protocol at a mean coverage of 30X. For Family 2, we used a TruSeq Exome Enrichment kit (Illumina, San Diego, CA) with samples prepared as an Illumina sequencing library enriched for the desired targets using the Illumina Exome Enrichment, with captured libraries sequenced using an Illumina HiSeq 2000 Sequencer. Sequence reads were aligned to the human genome reference sequence [hg19] and read alignment, variant calling, and annotation were performed by DNAnexus (DNAnexus Inc, Mountain View, CA). Intronic variants not predicted to affect splicing or regulation were excluded and SIFT [50], PolyPhen-2 [51], and MutationTaster [52] was used to predict the impact of any identified amino acid substitution on the protein structure and function and to predict and prioritize potential disease causing sequence alterations. The presence of the variants were confirmed in the transcript by bidirectional dideoxy Sanger sequencing performed on an ABI3730 XL capillary sequencer (Applied Biosystems), which was also used to confirm its co-segregation within the respective pedigrees.
Hyal2-/- (null) mice and littermate controls (Hyal2+/- or Hyal2+/+) were obtained from heterozygous intercrosses of mice maintained on an outbred (C129; CD1; C57BL/6) background as viable Hyal2-/- mice have not been obtained on an inbred background [35]. Embryos were considered E0.5 the morning a vaginal plug was discovered. Genotyping was performed as described [36]. Animals were euthanized by isoflurane overdose and embryos were collected for gross morphological or histological studies. All procedures were in compliance with the Canadian Council on Animal Care and followed a protocol approved by the University of Manitoba Animal Care Committee.
Ultrasound imaging was performed on mice anesthetized with 2% isoflurane, using a Visual Sonics Vevo 2100 ultrasound equipped with a 40 MHz transducer. Micro-CT imaging was performed at 9μm resolution with a Skyscan 1176 micro-CT and images were reconstructed and analyzed using Bruker’s NRecon, Data Viewer, CT Analyzer, or CT Vox software. ABR testing was performed as described previously [53]. Skeletal staining of neonatal mice was as described in [54]. For histological studies, tissues were fixed in 1% cetylpyridinium chloride in 10% buffered formalin, decalcified if appropriate, imbedded in paraffin, and 5μm sections were prepared. Sections were stained with H & E for analysis of morphology. HA was detected using the HA binding protein (HABP) following established protocols [36].
HA levels were determined using a DuoSet enzyme linked immunosorbent assay (ELISA) development kit (R & D Systems).
Transfections of HYAL2-expressing vectors were into mouse embryonic fibroblasts (MEFs) derived from a Hyal2-/- embryo. Cells and culture medium were collected 48 hours post-transfection and the β-galactosidase activity was determined using O-nitrophenol-β-D-galactopyranoside as a substrate [55]. Protein concentrations in the lysates were determined using a BioRad protein assay kit based on the Bradford method, and β-galactosidase levels were used to normalize for transfection efficiency in all assays. Western blot analysis was as described previously, but using polyclonal anti-HYAL2 (Proteintech) [56].
|
10.1371/journal.pgen.1002295 | Genomic Analysis of QTLs and Genes Altering Natural Variation in Stochastic Noise | Quantitative genetic analysis has long been used to study how natural variation of genotype can influence an organism's phenotype. While most studies have focused on genetic determinants of phenotypic average, it is rapidly becoming understood that stochastic noise is genetically determined. However, it is not known how many traits display genetic control of stochastic noise nor how broadly these stochastic loci are distributed within the genome. Understanding these questions is critical to our understanding of quantitative traits and how they relate to the underlying causal loci, especially since stochastic noise may be directly influenced by underlying changes in the wiring of regulatory networks. We identified QTLs controlling natural variation in stochastic noise of glucosinolates, plant defense metabolites, as well as QTLs for stochastic noise of related transcripts. These loci included stochastic noise QTLs unique for either transcript or metabolite variation. Validation of these loci showed that genetic polymorphism within the regulatory network alters stochastic noise independent of effects on corresponding average levels. We examined this phenomenon more globally, using transcriptomic datasets, and found that the Arabidopsis transcriptome exhibits significant, heritable differences in stochastic noise. Further analysis allowed us to identify QTLs that control genomic stochastic noise. Some genomic QTL were in common with those altering average transcript abundance, while others were unique to stochastic noise. Using a single isogenic population, we confirmed that natural variation at ELF3 alters stochastic noise in the circadian clock and metabolism. Since polymorphisms controlling stochastic noise in genomic phenotypes exist within wild germplasm for naturally selected phenotypes, this suggests that analysis of Arabidopsis evolution should account for genetic control of stochastic variance and average phenotypes. It remains to be determined if natural genetic variation controlling stochasticity is equally distributed across the genomes of other multi-cellular eukaryotes.
| Understanding how genetic variation controls phenotypic variation is a fundamental goal of biology in both modern medicine and agriculture. Yet, frequently, even a large set of genetic polymorphisms do not fully explain variance of a phenotype within a discrete set of individuals. Numerous mechanistic theories have been proposed, e.g. epigenetics, but we postulated that there may be genome-wide polymorphism controlling phenotype stochastic noise among genotypes. This is similar to what is being found in studies of bet-hedging theory in prokaryotic or single-celled organisms or stability in eukaryotes. Utilizing Arabidopsis, we tested this hypothesis at a genomic level by mapping quantitative trait loci for stochastic noise in global transcriptomics, plant defense metabolism, circadian clock oscillation, and flowering time within a single non-stressful environment. We cloned and validated a set of genes including transcription factors and enzymes that control natural variation in phenotypic noise. These genes provided evidence that stochastic noise can vary independently of average phenotypes. Since population genetic models and quantitative genetic studies focus on natural genetic variations impact upon average phenotypes, these observations suggest that stochastic noise needs to be incorporated to better explain the genotype-to-phenotype link.
| Almost all phenotypes are not fixed within species but instead exhibit significant levels of variation among individuals that is controlled by quantitative genetic loci. The study of such quantitative genetic variation has long been fundamental to evolution and ecology and is rapidly becoming a central focus of numerous other research fields, including breeding for improved crops and individualized medicine for humans. An ultimate goal of research on the basis of quantitative genetic variation is to generate a sufficient level of understanding to be able to predict phenotypic range of a species based on knowledge of that species' genetic variation. These efforts are complicated because phenotypic diversity is typically under polygenic control and can involve complex interactions with numerous factors including, but not limited to, the environment, development, epistatic interactions between genes, and potential higher-order interaction among these factors [1], [2]. Yet even in systems where these are understood to a significant degree, it has been difficult to develop predictive frameworks linking genotype to phenotype. Some of this difficulty has been ascribed to concepts such as epigenetic variance and difficulties in detecting small-effect loci [3], [4]. In this report, we propose that an additional explanation is the presence of numerous polymorphic loci that specify the amount of stochastic noise. If these polymorphisms are frequent in number, heritable and discrete from loci altering mean phenotpyes they can lead to an inability to fully describe the variance within any phenotype using current statistical approaches that focus solely upon the mean phenotype.
The idea that phenotypic variance is genetically determined is supported by a significant amount of research on how cells can limit stochastic noise/variance in genetic, metabolic, and signaling networks through network topology, a characteristic that is known as network robustness [5]–[10]. The specific topology of a network can increase or decrease the robustness of the output, wherein robustness is defined as the inverse of variance. Therefore, the genetic variation for loci within these networks could lead to allele specific changes in robustness/variance of the phenotype. Typically, robustness is thought to be under directional selection pressure to reduce the variance of an output and correspondingly increase network robustness. In evolutionary theory, this is predominantly described as canalization wherein genes function to minimize the variance (maximize the robustness) of a phenotype [11], [12]. In yeast, phenotypic and genetic robustness (i.e. canalization) were shown to correlate using genomic knockout datasets [13]. In plants, loci that control natural variation in canalization of critical developmental processes such as cotyledon opening and leaf formation have been mapped and cloned revealing that canalization genes can be known members of regulatory networks controlling these processes [14], [15]. Additionally, it has been shown that heat-shock protein 90 plays a major role in canalizing existing natural variation possibly as a pool of hidden evolutionary potential [16]-[18]. However it should be noted that the genomic level of, distribution of and importance of naturally variable loci controlling within-genotype variance is currently not fully described in most eukaryotes [19].
While canalization and robustness research focuses on the benefits of decreasing within-genotype variance, there is evidence that increases in per genotype variance can also be beneficial. This is occasionally called the portfolio effect wherein the fitness of a genotype is determined by the portfolio of phenotypes that it can obtain [20]. In some bacterial settings rapid environmental fluctuations have been shown to favor the development of stochastic switching as the optimal means of response [21]–[25]. Similarly in eukaryotes, it has been shown that natural variation can alter stochastic noise of gene expression [19], [26] and that stochastic noise in defense phenotypes could help to delay the evolution of counter-resistance in biotic pests [27], [28]. As such, it is possible that there is wide-spread genetic variation controlling stochastic noise in eukaryotic phenotypes that may play a beneficial role in the evolution of these organisms [16]–[18], [25]. However, little is understood about the genomic distribution of natural quantitative genetic variation for stochastic noise in eukaryotes or about the direction of selection on that natural variation.
The concept that stochastic noise is genetically determined in a quantitative, polygenic manner is supported by the analysis of stochastic variation in expression of a MET17 reporter fusion construct in Saccharomyces cerevisiae [29]. This study identified significant genetic diversity regulating stochastic noise of gene expression and showed that stochastic noise was a complex trait controlled by at least three quantitative trait loci (QTLs) [29]. However, given the nature of these alleles, it is not known if these polymorphisms are present in wild populations or are laboratory derived. Additional evidence comes from the study of the S. cerevisiae galactose regulon where it was found that genetic manipulation of the regulatory feedback loop could lead to increased stochastic noise in the network's output [30], [31]. Genetic control of stochastic noise has also been identified using QTLs for yield stability in crops [32] and gene expression in 18 isogenic mouse lines [19]. Further, it has been shown that HSP90 likely buffers genetic variation which could appear as stochastic noise in fluctuating environments, but little is known about the genomic distribution of natural variation in stochastic noise within a constrained environment [16]–[18]. These studies indicate that there is the genetic variation to regulate stochastic noise in physiology and gene expression suggesting that stochastic noise itself is a phenotype subject to natural selection with potential for pressure in both positive and negative directions.
To begin testing the genomic extent of natural genetic variation in stochastic noise we used the model plant, Arabidopsis thaliana. Arabidopsis is quickly becoming a key organism in the study of complex traits through the use of systems biology and quantitative genomics approaches [33]–[40]. This is due to large repositories of transcriptomic and metabolomic data for homozygous QTL and association mapping populations that, when combined with whole genome sequence of natural accessions, provides the ability to rapidly develop and test hypotheses as well as find causal genes underlying specific loci of interest [41]–[44]. This has enabled the identification and validation of numerous genes and defense pathways under natural selection [45]–[50]. Among these defense mechanisms with known selective consequences are the glucosinolate metabolites, thioglucosides that provide defense against numerous biotic pests and whose accumulation is heritable and under balancing or fluctuating selection in the field [51]–[62]. This makes Arabidopsis an ideal system to search for the genetic and molecular basis of complex phenotypes, such as stochastic noise, in higher organisms.
Using previous datasets, we identified QTLs that control natural variation in stochastic noise of glucosinolate metabolites and related transcripts within a single controlled environment. There were QTLs unique for the different phenotypic levels and we showed that known genes underlying these glucosinolate loci led to altered glucosinolate stochastic noise. We then extended this analysis to show that the Arabidopsis transcriptome shows significantly heritable stochastic noise for expression levels. Further, we were able to identify QTLs that control global stochastic noise in gene expression. Some loci were in common with those altering the average transcript abundance while others appeared unique to controlling transcriptomic stochastic noise. Using an existing single isogenic population, we confirmed that natural variation at the ELF3 locus alters stochastic noise in both physiological and metabolic phenotypes. Given the wide spread genomic variation controlling natural variation in stochastic noise in a single environment that we found within the wild Arabidopsis germplasm, our results suggest that any analysis of Arabidopsis evolution needs to account not only for genetic control of average phenotype value but also for genetic control of stochasticity. It remains to be determined how widely distributed this level of genomic natural variation exists for stochasticity within a wider range of multi-cellular eukaryotes.
To test if there is genetic variation affecting stochastic noise in the higher plant Arabidopsis thaliana we used a previous analysis of quantitative variation in glucosinolate defense metabolites [63]. The glucosinolate biosynthetic, transport and regulatory networks have been highly characterized [64]–[69], providing extensive information about the loci responsible for differences in mean glucosinolates within Arabidopsis thaliana accessions [60], [70]–[72]. Given the extensive knowledge it is possible to use existing glucosinolate data to search for QTLs controlling stochastic noise in glucosinolate accumulation. If stochastic noise QTLs are found they can be compared to existing analyses to determine if the same QTLs control phenotypic mean.
A previous analysis of glucosinolate variation in the Arabidopsis thaliana recombinant inbred line population (RIL) derived from the Bayreuth (Bay) and Shahdara (Sha; syn:Shakdara) accessions [42] reported both the mean glucosinolate accumulation and standard deviation per line for three replicated experiments quantifying concentrations of 62 different glucosinolate phenotypes in 392 RILs [63]. We used this information to obtain the coefficient of variation (CV) of each glucosinolate phenotype for each RIL by dividing the standard deviation of the phenotype by its mean, and used this dimensionless measure of stochastic noise in glucosinolate accumulation to perform QTL analysis [21], [26]. This identified five QTL hotspots controlling differences in glucosinolate CV (Figure 1). The pattern of CV QTL was similar to that found for QTL affecting differences in mean phenotype where GSL.ELONG and GSL.AOP are the major loci followed by two additional hotspots on chromosome 2 that had also been found to affect mean glucosinolates but were less significant than for glucosinolate CV (Figure 1) [63]. Further, we found a new QTL for CV that was not found for the mean phenotype within this population but had previously been found as a glucosinolate QTL in other populations, GSL.MYB2976 (Figure 1) [63]. There were also several QTLs that affected the mean phenotype but did not cause significant differences in glucosinolate CV (Figure 1) [63]. Thus, it is possible to find QTLs controlling CV differences and these are not necessarily the same loci as those that affect the phenotypic mean.
Fortunately, several of the identified QTLs have already been cloned and previously published single gene validation lines exist with published glucosinolate data to allow rapid validation of the CV phenotypes [63], [67], [73]. We have previously shown that the GSL.AOP locus is controlled by differential expression of two enzymes, AOP2 and AOP3, which evolved from a tandem duplication event to control different reactions with the same precursor [63], [74]. Using the same data set which previously showed that the QTL allele for increased glucosinolate accumulation and glucosinolate network transcript abundance was caused by expression of the AOP2 gene [63], we showed that introducing the AOP2 gene into a natural knockout background (Col-0) also significantly increased glucosinolate CV (Figure 1B and C). This increase correlates with the elevated CV found in Sha, which contains the functional AOP2 allele at the GSL.AOP locus (data not shown).
The GSL.MYB2976 locus co-localizes with a previously cloned QTL from a different RIL population (Ler x Cvi) that is controlled by two glucosinolate transcription factors, MYB29 and MYB76 [67], [72], [73]. We used data from previous single gene manipulations of MYB29 and MYB76 as well as the related MYB28, also linked to glucosinolate QTLs in other populations, to test if these genes can influence natural variation in glucosinolate CV [62], [67]–[69], [73], [75]. Interestingly, increasing or decreasing MYB28 expression significantly increases CV for all glucosinolates (Figure 1B and C). This is in contrast to previously published data showing that increasing MYB28 expression increased glucosinolate content while decreasing MYB28 expression correspondingly diminished glucosinolate content. Together this suggests that the effect of genetic variants on CV and mean is not always correlated [67]–[69], [73], [75].
In contrast to MYB28, only increases in MYB29 and MYB76 expression altered metabolite CV while decreased expression at either gene had no impact on glucosinolate CV (Figure 1B and C). This differs from their impact on mean glucosinolate accumulation where increases and decreases in all three gene expression lead to correlated increases and decreases in glucosinolate metabolites [67]–[69], [73], [75]. Interestingly, the natural variation in gene expression of MYB29 and 76 in the Bay-0 x Sha population appears to be a shift from a Col-0 like level in the Bay-0 genotype to elevated expression in the Sha genotype [40], [76] agreeing with the observed introduction of a CV QTL in this position. It is possible absence of a MYB2976 QTL altering the mean phenotypes may be an issue of not having sufficient RILs to identify this locus in the background of the other QTLs showing significant epistatic interactions [63]. To test if the use of CV may be biasing our analysis, we used Levene's F-test to compare variances between the various mutants and WT and obtained similar results (Figure 1). In summary, MYB28, MYB29, MYB76 and AOP2 alter glucosinolate CV, mean and unadjusted variance (Figure 1) [63], [67]–[69], [73], [75]. Since MYB29, MYB76 and AOP2 underlie CV QTLs, they are good candidates to control natural variation in glucosinolate CV within Arabidopsis thaliana. The observation that MYB28 and MYB29 perturbations have similar consequences upon mean glucosinolate accumulation but different influences on glucosinolate CV shows that the CV is not being driven by underlying changes in mean and is a valid approach for this analysis.
The GSL.AOP and GSL.MYB2976 QTLs control differences in both the mean accumulation of glucosinolate metabolites and the relevant biochemical pathway transcripts [63], [67], [73]. Having found that these QTL controlled differences in CV for glucosinolate metabolites, we next tested whether these QTL also control differences in CV for transcripts involved in glucosinolate production. We used pre-existing microarray data [40], [77] and found little evidence for impacts of the GSL.AOP and GSL.MYB2976 loci on the CV of transcript accumulation for individual transcripts in the GLS pathway (Figure S1), in contrast to their effect on CV for glucosinolate metabolites. Hereafter these loci are referred to as CV eQTL (CV eQTL = a QTL altering the coefficient of variation in transcript accumulation) to delineate them from standard eQTL (eQTL = a QTL altering the mean transcript accumulation). Similarly, there was no evidence that these loci impact the GLS related biosynthetic networks (Figure S2). This is in contrast to previous observations showing that AOP2, MYB29 and MYB76 can cause changes in glucosinolate pathway transcription and are known eQTL (expression QTL) hotspots for mean glucosinolate transcript abundance[63]. This is not entirely surprising as glucosinolate regulation shows extensive hallmarks of incoherent feed-forward loops [65], [67], [73] which can cause non-linear relationships in variance at different output levels. Thus the difference in CV partitioning between metabolites and transcripts at these loci is not entirely unexpected. Together, these data suggest that although the GSL.AOP and GSL.MYB2976 QTLs and the underlying causal loci (AOP2, MYB29 and MYB76) affect the mean transcript and metabolite abundance in the GLS pathway, and the CV of metabolite abundance, they don't alter the CV of transcript accumulation in this pathway. Interestingly, a hotspot on chromosome 2, controls the per transcript CV abundance of most genes in the GLS pathway and CV in glucosinolate content (GSL.ELF3, Figure 1). This locus fits the definition of a network CV eQTL as it alters the CV of the glucosinolate transcript network.
While there was no network CV eQTL at the GSL.AOP locus, the AOP2 and AOP3 genes showed evidence for a cis positioned eQTL controlling the CV for transcript accumulation for only these two genes and not the entire pathway (Figures S1 and S2). Interestingly, not all glucosinolate associated transcripts known to have a large effect cis-eQTL also had a cis-CV eQTL. For instance, the GSL.MAM locus contains cis-eQTL for the MAM genes yet there was no corresponding cis-CV eQTL (Figure S1) [63]. If our use of CV was solely tracking changes in mean abundance, the large effect cis-eQTL by default should have large effect cis-CV eQTL. The lack of this absolute relationship suggests that changes in mean are not driving changes in CV and supports the use of CV for mapping stochastic noise QTLs. Additionally supporting this is the fact that we utilized the same threshold estimation approaches for both CV-eQTL and cis-eQTL detection arguing against this being different statistical power issues [40].
The above analysis of existing glucosinolate quantifications suggests that there is significant genetic control of the CV for these defense metabolites. The CV itself may be under selective pressure to generate differences in stochastic variability between different natural populations of Arabidopsis [25], [28]. To query if genetic control of phenotypic CV is a global phenomenon within Arabidopsis, we used a pre-existing dataset consisting of replicated microarray experiments conducted on 211 lines of the Bay x Sha RIL population and the RIL parents [40], [77]. The distribution of CV across the transcripts was similar between Bay and Sha with a statistically significant difference of Bay showing a slight shift of the peak towards a higher CV (Figure 2A). Interestingly, the distribution of CV across the transcripts was more distinctly bimodal within the RILs suggesting significant transgressive segregation in the population only impacting a specific subset of transcripts (Figure 2A). The replicated nature of this experiment allowed us to directly assess the heritability of per line CV differences in both the parents and the RILs across 22,746 different transcripts representing the majority of the genome. The per transcript CV were correlated between Bay and Sha with an average heritability of 17% (Figure 2B and Figure 3A). The average heritability of per transcript CV was much higher in the RILs than the parental genotypes with an average heritability of 57% (Figure 2B). This is similar to the average heritability reported for the mean transcript abundance for the same experiment (∼68%) with the majority of this difference being due to the lack of a high heritability tail for transcript CV as compared to heritability for mean transcript abundance [40], [77]. As found previously for the mean transcript values, there was very little relationship between the heritability as measured in the Bay/Sha parents versus the RIL (Figure 3B). For the mean transcript abundance, this discrepancy was explainable by transgressive segregation due to QTLs of opposing effect and is likely true for CV-eQTLs as well, suggesting that similar levels of robustness in the two parents are obtained via different genetic networks [40], [77]. Supporting this is the observation that the standard deviation of transcript CV across the RILs is significantly greater than would be expected by modeling the expected CV using the parental values. In 1000 models, the maximal standard deviation of CV averaged across the transcripts in the RILs was 0.09 with a mean of 0.08. In contrast, the actual biological values showed an average standard deviation of CV per transcript across the RILs of 0.17 indicating that the RILs show a significantly larger distribution of CVs per transcript per RIL than would be expected given the parental value.
One concern with CV and any other estimate of variance is the potential for a correlation between variance and mean. The above analysis with glucosinolate accumulation did not suggest that this was a concern within Arabidopsis natural variation because we could identify instances where there were QTLs with large effect on the mean but no effect on the CV, even when identical approaches were used to determine significance thresholds. Within the RIL transcriptomic data, we did observe a statistically negative correlation (P<0.001) whereby transcripts with the lowest average abundance had the highest CV and vice versa however this correlation explained only 0.8% of the total variation in CV leaving 99.2% of the variation to be available for genetic control of CV independent of the mean (Figure 3C). This significant but minimal negative correlation likely derives from technical issues in microarrays surrounding the detection of lower expressed transcripts using Affymetrix microarray technology. To test if this technical issue constrains our ability to identify biologically controlled transcript CV, we compared the average per transcript expression to the heritability of per transcript CV within the RILs. This analysis showed that higher expressed genes had only a slightly more reproducible transcript CVs, therefore the technical issues surrounding low expressed genes does not impact our ability to identify biologically controlled CV (Figure 3E). Additionally, the magnitude of per transcript CVs in the RILs showed very little relationship to the heritability of per transcript CV suggesting that any CV/expression level correlation is not creating relationships at higher levels (Figure 3D). Thus, the use of CV to map QTLs for the transcripts appears to be valid. Interestingly, there was a strong negative correlation between the heritability of transcript abundance and the transcript CV, such that transcripts with the lowest CV had the highest heritability (P<0.0001, R2 = 0.59; Figure 3F). To ensure that the relationship between mean transcript abundance and transcript CV was not driving this correlation we repeated the analysis as a partial correlation while controlling for mean transcript abundance, this still showed a highly significant negative relationship between the heritability of transcript abundance and the transcript CV (P<0.0001, R2 = 0.42). Together, this suggests that quantitative genetic control of CV is a genome wide phenomenon within Arabidopsis thaliana that is not limited to defense metabolites and is at least partially independent from genetic variation controlling the mean phenotype.
The high estimated heritability of per transcript CV within the Bay x Sha RIL population suggests that it is possible to map CV eQTL for all transcripts. We used composite interval mapping to map CV eQTL for all 22,746 transcripts within 211 lines of the Bay x Sha population previously used to map eQTL [40], [77]. This identified 98,014 significant CV eQTLs that altered the stochastic noise for 21,974 transcripts for an average of nearly 4 CV eQTL per transcript (Figure S3). This is nearly twice the number of eQTLs per transcript found using the average transcript abundance as a phenotype [40]. This difference may be due to the use of two different experiments in the CV eQTL analysis, whereas the eQTL analysis used just one experiment, reducing its statistical power [40]. Given that we used identical methods to identify global permutation thresholds for both datasets, we do not feel that a higher false positive rate can explain the elevated number of CV eQTLs [40], [78]-[80]. In addition, the elevated number of CV eQTLs is not universal as the glucosinolate transcript measurements actually identified more eQTLs than CV eQTL (Figure S1) [63]. Thus, the elevated CV eQTL level may be more indicative of the specific biological process within which that transcript functions.
An analysis of the distribution of additive effects for the CV eQTL showed a slight bias towards Bay alleles having a negative impact on CV (50429 CV eQTLs with Bay additive effect <0 versus 47585 with Bay additive effect >0)(Figure 4A). The vast majority of CV eQTLs had absolute effects less than 0.1 CV and these were almost entirely acting in trans (Figure 4A and B). In contrast, CV eQTL with absolute effects greater than 0.1 were predominantly acting in cis (Figure 4B). This is similar to eQTL controlling the mean accumulation of a transcript where on average trans-eQTLs have smaller additive effects than cis-eQTLs [38], [40]. This analysis identified 3,720 transcripts as having a cis-CV eQTL, in contrast with the 5,127 transcripts having a cis-eQTL for mean expression level (Figure 4) [40]. While about ¼ of all eQTLs detected were cis, only 1/26th of all CV eQTL were cis, showing that natural variation at trans positions is dramatically more prevalent in controlling transcript CV than average expression (Figure 4) [40]. As expected by the decreased ratio of cis-CV eQTL relative to that found for eQTL, the cis diagonal, while present, was very faint (Figure 5B). Only 1,660 transcripts had both a cis-eQTL and cis-CV eQTL and these included nearly all of the large effect CV eQTLs (Figure 4) [40]. Thus, while a cis-eQTL can be associated with a cis-CV eQTL, it is not a necessity (Figure 4). These results show that stochastic noise measured as CV in transcript abundance is a highly heritable trait suitable for genome wide QTL analysis in multi-cellular eukaryotes. As in eQTL analyses of mean transcript abundance, differences in the CV of transcript accumulation seem to be broadly caused by abundant loci acting in trans, while substantial changes are less frequent and usually associated with variation in cis.
We counted the number of loci per chromosomal position controlling stochastic noise within the Arabidopsis transcriptome to better understand the genomic distribution of CV eQTLs (Figure 5A). This identified a number of locations within the genome that contain trans-hotspots for CV eQTL. Several of these were in common with eQTL trans-hotpots that had previously been identified such as the locations on Chromosome II. However, the relative impact of the trans-hotspots upon the transcriptome was different for the two traits (Figure 5A) [40]. For instance, the trans-hotspots at 12 and 42 cM on chromosome II caused similar numbers of eQTL, yet the hotspot at 42 cM affected many more CV eQTLs than the hotspot at 12cM. Additional hotspots were detected with CV eQTL that were not detected using mean transcript accumulation, most notable is the locus at the bottom of chromosome III that is the highest trans-hotspot for CV eQTL but barely registered for eQTL (Figure 5A) [40]. Two other apparent CV eQTL specific trans-hotspots were peaks over the permutation threshold near the GSL.AOP and GSL.MYB2976 loci on chromosomes IV and V (Figure 5A) [40]. However, none of the glucosinolate transcripts' CVs were regulated by the trans-CV eQTL hotspots near GSL.AOP and GSL.MYB2976 (Figure S1). This raises the question of whether these CV loci near GSL.AOP and GSL.MYB2976 are due to pleiotropic consequences of the metabolic CV controlled by GSL.AOP and GSL.MYB2976 (Figure 1) or if there are additional genes in these regions that alter transcriptomic CV.
The detected CV eQTL hotspots have additive effect biases, with most of the CV eQTLs in one hotspot increasing the CV in the same direction, as noticed before for eQTL hotspots (Figure 5B) [40]. The two major hotspots had opposite effects; with the Sha allele causing increased stochastic noise at the hotspot in chromosome III and decreasing stochastic noise in all hotspots on chromosome II (Figure 5B). = This observation further shows that increased mean abundance does not inherently cause increased CV. Thus, transcript mean abundance and CV are not measures of a single phenotype and instead can involve different genetic mechanisms even when investigating the same locus.
The global effect of trans-CV eQTL hotspots led us to test if we could directly map QTL controlling genome-wide transcriptomic CV (as opposed to per transcript CV). Taking the average CV across all transcripts showed that Bay and Sha have different CV and that the main source of this is the previously identified loci on Chromosome II and III (Figure 6). Thus, these loci appear to have genome wide effects upon stochastic noise of gene expression and likely other traits.
The chromosome II locus found using the global CVs of transcript abundance, glucosinolate accumulation and glucosinolate network expression maps close to the previously identified ELF3 QTL (Figure 1, Figures S1 and S2) [81]. Allelic variation in ELF3 between Bay and Sha has been shown to affect circadian rhythms and shade avoidance responses but not the wave form of the circadian oscillation [81]. We next wanted to test if the ELF3 locus could be the same as the global CV eQTL hotspot. Because of ELF3's involvement in the circadian clock, we first asked whether we could identify stochastic noise QTL for circadian rhythms in the Bay x Sha population and whether these QTL would overlap the ELF3 region. Circadian rhythms in transcript abundance have been measured in this population [82]; we used this same approach to map CV for the expression of circadian clock regulated genes. Briefly, transcripts previously identified as being regulated by the circadian clock were grouped into 24 CT phase groups based upon each transcript's time of peak expression (CT) during the 24 hour photoperiod [82]-[84]. Transcript expression values were then Z normalized and a single expression estimate was independently obtained for each CT phase group for each microarray. These were then used to estimate the variance of the CT phase groups expression as described. Both the ELF3 locus and the chromosome III hotspot were found to alter CV for gene expression across the circadian clock output networks with opposing effects as had been found for general gene expression (Figure 5 and Figure 7). In contrast, the other identified trans-CV eQTL hotspots (Figure 5), do not appear to influence the CV of transcripts regulated by the circadian clock (Figure 7 versus Figure 5).
To test if ELF3 is the causative gene controlling stochastic noise in this region we utilized previously generated Col-0 elf3.1 knockout mutants lines containing a CCR2:luc reporter gene that were rescued with the genomic Bay and Sha ELF3 alleles (elf3:Bay-0 and elf3:Sha) [81]. Since Bay and Sha ELF3 genomic alleles have been shown to affect the period of CCR2:luc oscillations in free running conditions under different light environments [81], we monitored the CV in period in at least 650 T1 plants per transgene distributed in 10 independent experiments performed in constant red or in constant red plus far red light. Independent of light conditions we found that the Sha ELF3 allele reduced stochastic noise in the circadian oscillation period, in agreement with the direction of the global CV eQTL and circadian CT phase group QTL at the ELF3 position (Figure 7 and Figure 8). Although plants in both red and red plus far red light presented lower CV ((P = 0.002 in red light versus P = 0.043 in red plus far red light, via ANOVA), the difference in CV between the two alleles was not significantly affected by the light treatment (Figure 8, P = 0.35 via ANOVA). The Bay and Sha alleles of ELF3 did not affect CV for amplitude, phase or quality of the rhythms (measured as the relative amplitude error) in the transgenic plants (P = 0.10, P = 0.18 and P = 0.50 respectively, data not shown).
To further test if ELF3, could be the gene underlying other the CV QTL identified for other phenotypes at this locus, we tested if the transgenic lines differed in the level of stochastic noise for glucosinolate metabolites (Figure 1). Different alleles of ELF3 led to changes in glucosinolate stochastic noise with the Sha ELF3 allele increasing stochastic noise for the short chain aliphatic glucosinolate 4-methylsulfinylbutyl (4-MSOB) but decreasing stochastic noise for the long chain aliphatic glucosinolate 7-methylsulfinylheptyl (7-MSOH) (Figure 1 and Figure 8). Since the different alleles of ELF3 (Bay v Sha) have also been shown to affect flowering time [81], we measured two traits related to this character in the transgenic lines and found that variation between the ELF3 alleles led to differences in stochastic noise for flowering (Figure 8, Figure S4).
The observation that the Sha allele of ELF3 led to higher stochastic noise in flowering time and 4-MSOB accumulation, whereas it was also associated with lower stochastic noise in circadian periodicity and 7-MSOH accumulation suggests that ELF3 is not simply making the plant more or less robust but instead is partitioning noise between specific phenotypes (Figure 8). Interestingly, this differential effect of ELF3 upon stochastic noise agrees with the observed CV eQTL at this locus.: The Sha allele at the ELF3 QTL was associated with decreased stochastic noise of transcriptional networks for circadian genes and most glucosinolate networks but the Sha allele had increased stochastic noise in the FLC (At5G10140 –Flowering locus C) and GS-OX2 (At1g62540) transcripts (Table S1) [85]–[88]. The increased noise in FLC nicely correlates with the observed flowering time noise. Furthermore, GS-OX2 is required to synthesize 4-MSOB and concordantly links to the increased noise in this metabolite. Interestingly, YUCCA3 (At1g04610) transcript accumulation also shows a CV eQTL at the ELF3 locus suggesting a potential impact on auxin by this locus [89]. In summary, our results show that natural variation in ELF3 leads to changes in stochastic noise in both plant and molecular phenotypes and that the direction of effect depends upon the specific phenotype. Therefore, ELF3 is not a gene leading to plants displaying a general increase in phenotypic noise but instead affects noise in a network specific manner. Finally, it should be noted that there is no measurable difference in gene expression between the Bay and Sha alleles at ELF3 showing that these altered stochastic noise phenotypes in metabolism, transcription and physiology are dependent upon the biochemical differences in the two alleles [40], [90].
The accurate measurement of any phenotype in biology produces two numbers, a measure of central tendency, such as the mean and a measure of variance. However, most genomic studies of quantitative genetics or systems biology in multi-cellular organisms limit the analysis to whether the genetic, environmental or developmental perturbation altered the phenotype's mean and typical do not analyze effects on the stochastic variance. However, numerous microbial and modeling analyses have shown that stochastic noise can be a meaningful phenotypic descriptor that contains information not conveyed by the average [21], [25], [26], [29], [30], [91]. We hypothesized that there may be an unrecognized and broad genomic distribution of natural variation in stochastic noise within higher eukaryotes. The data described in this report shows that there is significant genomic variation in phenotypic stochastic noise within the model plant, Arabidopsis thaliana within a single environment. This genetic variation in stochastic noise, as measured by CV, is highly heritable and influences multiple phenotypic levels ranging from transcripts to metabolites to complex physiology like circadian clock periodicity. We mapped numerous QTLs controlling metabolic and transcriptional CV and demonstrated that specific genes underlying these loci have the ability to influence the phenotypic CV for these traits. Further, phenotypes with higher stochastic noise had lower heritability. As such, it is likely that genetic variation in stochastic noise is widespread with a diverse mechanistic basis, and that to fully understand a quantitative trait both the mean and the stochastic variance of the phenotype need to be investigated.
Our QTL analysis showed that CV and average can be genetically separable measures of a phenotype. QTL mapping using phenotypic CV as the trait identified loci that were not found using the phenotypic average. One example is the transcriptome trans-CV eQTL hotspot on the bottom of chromosome III that did not appear as a major hotspot when using the average transcript accumulation to map eQTL (Figure 5) [40]. Further, natural variation at the ELF3 locus impacts the average circadian clock period and flowering time while only effecting CV of the circadian clock (Figure 8 and Figure S4) [81]. Further, Levene's F-tests of the individual causal genes supports the use of CV to identify genes controlling natural variation in stochastic noise. Thus, directly interrogating stochastic noise as a separate measure of a phenotype can lead to new insights into the biology of a system.
Stochastic noise is frequently divided into that which comes from sources internal to the organism (intrinsic) and environmental sources external to the organism (extrinsic) [26]. In unicellular organisms it is possible to use internal reporters and massive population sizes to begin to partition the two sources. This is much more difficult for multi-cellular organisms. However, in this study, a number of findings support that we are likely measuring largely intrinsic sources of noise rather than purely extrinsic sources. The first is that our measures of noise are correlated with the genotype of the organism, which would not have been true if the variance we were measuring was purely extrinsic/environmental noise. It could be argued that we are mapping genetic variation that leads to differential sensitivity to extrinsic noise. However, each experiment is highly replicated so to be mapping differential sensitivity to extrinsic noise would have required the sources of extrinsic noise to be the same in each experiment and to show similar variations across the experiments. While we can not entirely rule out this possibility, it is much more likely that we have identified loci controlling natural variation in intrinsic stochastic noise within a multi-cellular organism.
An interesting observation in this data is that there is an unexpectedly high genomic level of natural genetic variation controlling stochastic noise in transcripts, metabolites and physiology. The high frequency of trans-CV eQTL rules out the possibility that this is simply the mapping of large effect indel polymorphisms that would be expected to alter transcript CV in cis. Additionally, the finding that the genes underlying trans-CV eQTL also control stochastic noise in metabolites and complex physiology such as the circadian clock shows genetic control of stochastic noise impacts all levels of the plant. Interestingly previous reports have shown that HSP90 could be expected to control stochastic noise in numerous Arabidopsis phenotypes but we did not identify any trans-CV-eQTL hotspots linked to any of the known HSP90 genes [16]–[18], [92]. This suggests that natural variation in HSP90 is not a major driver of stochastic variation within this Arabidopsis population for this environment. It is possible that if we had used multiple environments that natural variation in HSP90 may have been identified but this was not the case.
Our findings raise the question of what genetically variable control of noise means in an ecological and evolutionary context. One possible answer would be that this genetic control of noise is meaningless because stochastic noise may not be under selection. However, this answer runs up against two impediments. The first is that in bacteria, natural variation in phenotypic stochastic noise has been shown to be adaptive under situations where the environment is highly unpredictable [25], [91], similar to that found in plant/herbivore interactions [28]. Additionally, several of the glucosinolate loci, including the GSL.AOP2 locus that we show controls stochastic noise in glucosinolates, have been shown to be under selection in Arabidopsis and other related species [52], [56], [60], [93]–[100]. While these findings do not show that the stochastic noise variation is directly under selection pressure, it is clearly controlled by genetic loci that are themselves likely under selection pressure. Further, this suggests that selection is not solely focused upon decreasing stochastic noise within non-stressful environments especially for defense related traits.
The next question then becomes how natural variation in stochastic noise within environments that are not overtly stressful could benefit a multi-cellular organism. The answer to this might come in the form of a question that is related to the interest in identifying the genetic basis of local adaptation. However, the term local adaptation always engenders the response “what is local?”. It is possible that altering the stochastic noise of a system could alter the range of environments where it can successfully function. For instance, increasing the stochastic noise of the circadian period may enable that particular genotype to occupy more longitudinal niches, albeit at the likely cost of never being the optimal genotype in any specific niche. In contrast, decreasing the stochastic noise of the circadian period would optimize the fitness in a specific niche but likely at the loss of fitness across other niches. In this instance, natural variation in stochastic noise could lead to genetic control over what constitutes local for a specific genotype. As such, it may not be the variance itself that is adaptive, but instead the ability of variance to produce a more flexible network.
In contrast, stochastic noise in defense metabolites, such as glucosinolates, could represent a different benefit of natural variation in CV. Glucosinolates are a major anti-herbivore and anti-pathogen defense of Arabidopsis and relatives [53], [54], [58], [59], [98], [99] and as such could impart a pressure upon these herbivores and pathogens to counter adapt [101]. One mechanism that has been suggested as effective in slowing counter-adaptation is to increase the unpredictability of the defense compound (i.e. stochastic noise) [27]. As such, genetic control on the stochastic noise of defense compounds could in and of itself provide direct benefits to the efficaciousness of the defense. However, the observation that there is natural variation in stochastic noise of defense metabolites would suggest that high levels of noise are not always beneficial, possibly depending upon the ratio of generalist and specialist herbivores in a given genotype's normal locale [98]. Testing these different potential benefits of stochastic noise will require the development of genotypes that differ solely in stochastic noise to allow this effect to be partitioned away from any influence upon the mean phenotype.
A major difficulty in systems biology is the presence of massive datasets that are largely correlative when comparing different transcripts. This has lead to numerous attempts to derive causal information from these correlative datasets. However, even the best approaches are susceptible to a number of systemic errors that deal with predicting regulatory loop structure as well as combinatorial regulation [102]. For regulatory loops, correlative approaches using average responses generate a number of possible network topologies that are similar with respect to regulation of phenotypic average, but that make very different predictions about how perturbations will control the stochastic noise of the system [5], [6], [103]–[105]. Given this, it may be possible to use the presence of genetically controlled stochastic noise to help better refine systems biology models. The mean transcript, protein, or metabolite levels could be used to generate multiple initial models that could then be analyzed by using the stochastic noise in the system to determine which model most accurately predicts the observed stochastic noise. Future work on this approach could be useful but would require true independent replication in systems biology experiments to allow accurate estimations of stochastic noise for each measured phenotype.
The identification that stochastic noise of phenotypes has a level of genetic control that appears to be on par with that observed for the phenotype average suggests that there is a fount of phenotypic information that has largely not been studied in most modern genetic, genomic or systems biology studies. For instance, numerous natural and induced mutant screens and surveys have been conducted in Arabidopsis to determine the genes controlling the phenotypic average [106]–[108]. Similar large scale approaches have been conducted in numerous other organisms focused on phenotypic averages [109]–[111]. While these have provided great advances in our understanding of biology, it raises the question of what would happen if we repeat these screens and surveys to identify genetic variation controlling stochastic noise in phenotypes. Would we identify the same genes or would we begin to identify a large suite of previously unknown genes that control stochastic variation rather than phenotypic average? Experiments focused on the stochastic nature of a phenotype require independent replication but could yield a new view of organismal biology that is currently specified by our focus upon phenotypic averages.
To directly estimate the CV for each individual genes transcript accumulation (22,746 transcripts in total) as a separate phenotype within the Bay x Sha RIL population [42], we obtained two independent microarray experiments (TABM-224 and TABM-518) wherein 211 RILs were each measured in duplicate within each experiment providing four replications [40]. Raw image data from the RIL GeneChips were converted to numeric data via Bioconductor software (www.bioconductor.org). We utilized quantile normalization across all arrays to reduce non-biological variation coming from the technology itself, and when applied at the probe level it has been shown to outperform other normalization methods that are based on what is referred to as a “base-line array” [112]. After quantile normalization, we utilized the absolute expression values to measure the CV for each gene separately for each experiment using σ/µ [21], [26], [113], thus providing two independent biological replicate measures of CV for each gene. The use of CV as a direct phenotype has previously been used in a number of instances. By measuring the within line CV as a phenotype for the Bay x Sha population allows us to then utilize CV as a direct measurement of stochastic variation as a phenotype. The level of per line replication for the array data does not support the use of Levene's variance tests or measures. Additionally, all lines were planted and harvested within a randomized complete block design at all stages thus limiting any potential technical bias to generate these observations [40], [77].
To estimate the CV for specific transcript networks, we utilized a previously published approach whereby we average the expression across a group of genes to provide an estimate of the gene network's expression value [38], [114]. Briefly, this network approach uses any a priori defined group of genes as a network. Every transcript that is defined within a network is z transformed to place them all on the same scale. For every microarray within the dataset, the network expression value is obtained by averaging across the z values for all transcripts within the network. This provides a single network value that can then be utilized for downstream applications. This approach has previously been used to map network QTL controlling the difference in average expression [63] and can be extended to identify differences in network stochasticity using the CV value instead of the average expression. Gene membership within specific circadian networks were defined as previously described [83]. Gene membership within glucosinolate pathways were defined as previously described [63], [67]. This approach was also used to generate a global CV average by averaging the CV across all 22,746 transcripts measured on the ATH1 Affymetrix microarray.
To estimate the CV for specific defense metabolites, we utilized previously published data wherein the µ and σ for a large set of glucosinolates within a Bay x Sha RIL population consisting of 403 lines had been measured [63]. The glucosinolates were measured in a similar growth stage and growth chamber as that for the transcriptomics analysis allowing for better comparison between the datasets [40], [63]. For measuring altered glucosinolate metabolite CV in the independent transgenic lines, we compiled data from multiple independent experiments that had previously been published in separate papers. We analyzed the same lines in at least four independent experiments with replication allowing us to test if the CV differed across these genotypes [58], [63], [67], [73].
Glucosinolate genotype analysis: To test if variation at specific glucosinolate genes could alter the CV of either metabolite or transcripts, we obtained previously published data involving multiple independent biological replicates for the following genotypes all of which are generated within the Arabidopsis Col-0 accession background. To elevate MYB gene expression, we used previous lines where the Arabidopsis Col-0 versions of MYB28, 29 and 76 were separately introduced back into Arabidopsis Col-0 using a 35S promoter to induce their expression – 35S:MYB28, 35S:MYB29 or 35S:MYB76 [73]. To mimic natural variants that have low to no expression of MYB28, 29 or 76, we used previously obtained insertional T-DNA mutants within each of these genes obtained from the Arabidopsis Col-0 accession; myb28-1, myb29-2 and myb76-1 [67], [73]. All insertional T-DNA mutants underwent at least one backcross and had previously been shown to abolish or dramatically diminish MYB gene expression [67], [73]. To mimic the natural variation at the AOP2 locus, we utilized the Arabidopsis Col-0 accessions that contains a natural knockout of AOP2 and introduced the functional enzyme encoding gene back into this natural null background [63], [115]. Thus, all of these lines are single gene manipulations of major glucosinolate loci within a common genomic background, Col-0.
For estimating broad-sense heritability, we utilized the independent measures of CV directly as a phenotypic measure. This allowed us to estimate broad-sense heritability (H) for each CV phenotype as H = σ2g/σ2p, where σ2g is the estimated CV phenotypes genetic variance among different genotypes in this sample of 211 RILs, and σ2p is the CV phenotypic variance for each phenotype [116]. Heritability was estimated for all expression phenotypes. The metabolite phenotypes did not have the individual values from each independent experiment, and therefore, heritability was not measurable.
To map QTL for the CV phenotypes, metabolic, network and individual gene expression, we measured the average CV for each phenotype across all experiments and used the average CV in conjunction with a previously generated map for 211 Bay x Sha RILs ([42], [77]; see also the file “Average CV per transcript per RIL” at http://plantsciences.ucdavis.edu/kliebenstein/TableS1Plosgenetics.txt [note: this file is ∼28 MB]). For glucosinolates, we utilized a larger collection of 400 Bay x Sha RILs [42]. Composite interval mapping (CIM) analysis [117] was employed in conjunction with the 5 cM framework map. The “zmapqtl” CIM module of QTL-Cartographer Version 1.17 [118] with a walking speed of 1 cM and a window size of 10 cM was employed to analyze each phenotype. To obtain a threshold criterion for declaring statistically significant eQTL, a global permutation threshold was obtained by permuting the e-traits while maintaining the genetic information [40]. For each of 100 randomly selected phenotypes, the null distribution of the maximum likelihood ratio test (LRT) statistic was empirically estimated using permutation thresholds based on 1,000 permutations [40], [78]–[80]. We then utilized the 95th percentile permutation threshold across the 100 null distributions [40]. We utilized the resulting output to localize, summarize and count CV QTL using the Eqtl module of QTL-Cartographer in conjunction with the previously optimized 5 cM exclusionary window where no CV QTL can be closer than 5 cM to the nearest QTL (Table S1) [40], [118]. This is an identical approach at all stages to that used to previously map the eQTL for this dataset and as such should increase the direct comparison between datasets [40]. Additionally, we have been able to clone and biologically validate causal loci controlling several of the trans effect loci controlling subtle shifts in physiological networks as identified from the eQTL analysis [90], [119].
We have previously shown that single-feature polymorphisms are not a significant difficulty in this population for this array data when estimating expression values [40], [77]. As such, we did not control for potential single-feature polymorphism issues. The low level of cis-CV-eQTL within our results further supports this observation.
To determine whether a genetic location associated with multiple CV QTLs was a significant cluster or ‘hotspot’, we estimated a significance threshold using permutation as previously described for transcriptomic data [40], [90], [109], [120]–[123]. The positions of the 98,014 CV eQTLs (Table S1) at the marker intervals were permuted across the genome 1,000 times, and the maximal number of CV eQTLs per genetic position per permutation was obtained. Using the distribution of the maximum number of CV eQTLs, the criterion for declaration of a significant eQTL hotspot was 422 CV eQTLs per genetic position at alpha = 0.05. The permutated hotspot approach has been used to identify genes that cause the transcriptional difference for a number of hotspots showing that this approach is identifying biologically validatable effects [39], [63], [120], [124], [125].
elf3-1 null mutants carrying the CCR2::luc reporter gene were obtained from Dr. Stacey Harmer (University of California, Davis). Full genomic clones of ELF3 from Bay and Sha including 1.5 kb of upstream promoter were cloned in pJIHOON212. elf3-1-CCR2::luc plants were transformed with these constructs using Agrobacterium tumefaciens [81], [126]. To account for differences between elf3.1: Bay and elf3.1:Sha due to the transformation protocol, transgenic plants obtained from two independent batches were used, but no effect of the Agrobacterium inoculate was detected (P = 0.31 via ANOVA, data not shown).
elf3:Bay and elf3:Sha transgenic T1 seeds from two different Agrobacterium transformation batches were placed on MS medium with the appropriate antibiotic and stratified for 4 days (4°C, dark). After entrainment under white light in 12∶12 photoperiods for 7 days, resistant plants were transferred to new MS plates and moved to continuous red light or red + far-red light conditions, where luminescence was recorded for 6 to 7 days.
Five independent experiments were conducted in continuous red light (R, total PAR of 64 uE) and 5 experiments in continuous red plus far red light (R+FR, total PAR of 64uE with a R:FR ratio of 0.5) conditions created with LED lights. Plants were monitored using a CCD camera taking pictures every 2 hours. The data collected was analyzed for rhythmicity using the luciferase activity method described in [127]. Only plants showing stable rhythms (Relative Amplitude Error below 0.5) were considered for the analysis. Between 12 and 150 T1 plants (average 75.2, median 86) for each transgene were included in each experiment. Coefficient of variance was calculated as the standard deviation divided by the mean period estimate for each transgenic line in each experiment.
elf3:Bay and elf3:Sha transgenic T1 seeds from three different Agrobacterium tumefaciens transformation batches were planted on soil including elf3.1 mutants and WT Col-0 as a control. The extreme hypocotyl length, flowering time and cotyledon color phenotypes of the elf3.1 mutants were assessed to distinguish transformed from untransformed plants [128]. Transformed plants were grown for 25 days in a 10 hour photoperiod. At 25 days, leaf tissue was harvested from each plant and individually extracted and assayed via HPLC for glucosinolate identity and concentration as previously described [72], [115]. The experiment was replicated 9 times for a total of 106 elf3:Bay and 108 elf3:Sha independent T1 plants. Levene's F-tests were used to compare variance between the two T1 genotype classes.
|
10.1371/journal.ppat.1007144 | Cryptococcus neoformans urease affects the outcome of intracellular pathogenesis by modulating phagolysosomal pH | Cryptococcus neoformans is a facultative intracellular pathogen and its interaction with macrophages is a key event determining the outcome of infection. Urease is a major virulence factor in C. neoformans but its role during macrophage interaction has not been characterized. Consequently, we analyzed the effect of urease on fungal-macrophage interaction using wild-type, urease-deficient and urease-complemented strains of C. neoformans. The frequency of non-lytic exocytosis events was reduced in the absence of urease. Urease-positive C. neoformans manifested reduced and delayed intracellular replication with fewer macrophages displaying phagolysosomal membrane permeabilization. The production of urease was associated with increased phagolysosomal pH, which in turn reduced growth of urease-positive C. neoformans inside macrophages. Interestingly, the ure1 mutant strain grew slower in fungal growth medium which was buffered to neutral pH (pH 7.4). Mice inoculated with macrophages carrying urease-deficient C. neoformans had lower fungal burden in the brain than mice infected with macrophages carrying wild-type strain. In contrast, the absence of urease did not affect survival of yeast when interacting with amoebae. Because of the inability of the urease deletion mutant to grow on urea as a sole nitrogen source, we hypothesize urease plays a nutritional role involved in nitrogen acquisition in the environment. Taken together, our data demonstrate that urease affects fitness within the mammalian phagosome, promoting non-lytic exocytosis while delaying intracellular replication and thus reducing phagolysosomal membrane damage, events that could facilitate cryptococcal dissemination when transported inside macrophages. This system provides an example where an enzyme involved in nutrient acquisition modulates virulence during mammalian infection.
| Cryptococcus neoformans is a relatively frequent cause of life-threatening infection in severely immunocompromised patients, especially those with AIDS. Persistence of infection involves residence within macrophages, where C. neoformans can survive and replicate while residing in the phagolysosome. New treatments may be developed from a better understanding of how this pathogen resists clearance from and adapts for persistence in host phagocytic cells. In this study, we demonstrate a novel role for urease, a major virulence factor of C. neoformans, in its interaction with macrophages. This enzyme is able to break down urea into ammonia, which is a base, thus raising the surrounding pH. In the context of a mammalian infection, we show that cryptococcal urease increases the phagolysosomal pH which delays yeast replication, therefore causing less damage to macrophages and prolongs intracellular residence. Moreover, urease promotes C. neoformans exit from macrophages without killing the host cells. Overall, our data implies that urease also contributes to virulence by allowing the pathogen to persist and disseminate in macrophages.
| C. neoformans, a major life-threatening fungal pathogen predominantly infects severely immunocompromised patients and causes over 180,000 deaths per year worldwide [1]. C. neoformans is ubiquitous, although is most frequently found in soils contaminated with bird excreta or from trees [2–11]. Current treatments for Cryptococcosis often fail, are inadequate and/or unavailable for these infections, especially in developing countries. Therefore, it is important to study the fundamental pathogenic processes of C. neoformans to discover new treatments against this pathogen.
Human infection with C. neoformans follows inhalation of spore or yeast cells. In healthy individuals, pulmonary infections with C. neoformans are normally controlled and macrophages play a central role [12,13]. Soon after phagocytosis, the Cryptococcus-containing phagosome undergoes maturation, acidification and lysosome fusion [14–17]. However, C. neoformans is a facultative intracellular pathogen that it is able to survive and persist in mature phagolysosome, and can become latent and localized within the giant cells or macrophages in granulomas [15,18–22]. Depletion of macrophages is associated with improved survival of infected mice, supporting the notion that yeast cell cells are maintained within macrophages and as such, this host cell can constitute a niche for dissemination and persistence [23]. In the rat, latent infection resides in macrophages [18]. Infection can reactivate in conditions of weakened immunity, with intracellular replication and dissemination [15,17,20,22,24]. Consequently, the ability of C. neoformans to survive and replicate intracellularly contributes to different stages of cryptococcal pathogenesis [25–27]. It has been proposed that this intracellular pathogenic strategy emerged from interactions with amoebae in the environment [28–30]. A recent study reports that C. neoformans spends a relatively short time (~80 min) inside Dictyostelium discoideum and is expulsed before yeast replication occurs [30].
C. neoformans expresses virulence factors that promote its pathogenicity, including formation and enlargement of a polysaccharide capsule, melanin production, extracellular secretion of various enzymes including phospholipase, urease, etc. The role of capsule and melanin in macrophage-pathogen interaction are well understood. The capsule interferes with phagocytosis, by potentially masking macrophage receptor binding sites and polysaccharide shed by the yeast is immunosuppressive [31–34]. Moreover, both capsule and melanin protect C. neoformans from intracellular killing by providing protection against reactive oxygen species (ROS) as well as antimicrobial peptides [35,36]. However, mechanism by which urease contributes to intracellular pathogenesis is unknown.
Urease functions as a general virulence factor for many bacterial pathogens, such as Helicobacter pylori [37], and fungal pathogens Cryptococcus spp. and Coccidioides posadasii [38–40]. Urease catalyzes the hydrolysis of urea into carbon dioxide and ammonia [41,42]. Ammonia generated from ureolytic activity can serve as a nitrogen source. Since urea is evenly distributed throughout the human body it is conceivable it is used as a nutrient by mammalian pathogens [43]. Beyond its nutritional role, ureolytic activity enhances the invasion of C. neoformans to the central nervous system by promoting the yeast sequestration within the microcapillary beds of blood-brain barrier. The underlying mechanism is not known, but was hypothesized that ammonia generated by urease activity was toxic to microvascular endothelial cells [44–46]. Urease-mediated ammonia can also neutralize any acidic microenvironment and thus help pathogens to survive harsh pH of the phagolysosome. The neutralizing effect of H. pylori’s urease is well established, enabling that bacterium to colonize gastric mucosa [47,48]. In addition to its role in gastric colonization, H. pylori urease regulates the host-macrophage interaction by retarding the opsonization of H. pylori [49]. The enzyme can also modulate phagosomal pH and disrupt phagosome maturation to enhance the intracellular survival of H. pylori in macrophages [50]. Furthermore, it induces the expression of inducible NO-synthesizing enzyme (iNOS), a M1 macrophage polarization marker, in mouse macrophages [51]. In contrast, in both C. neoformans and C. posadasii, urease-producing strains promote the polarization of immune responses to a nonprotective Type 2 (T2) rather than a fungicidal Type 1 (T1) immune response [38,52]. Hence, bacterial and fungal urease may have different effects on macrophage activation, and the role of cryptococcal urease during macrophage-pathogen interaction, which may affect the appropriate immune response, is unexplored.
In this paper, we evaluated the role of urease on intracellular pathogenesis of C. neoformans in both amoebae and macrophages. We studied the effect of urease on the macrophage response to C. neoformans, as measured by host cell lysis and non-lytic exocytosis, cryptococcal replication inside macrophages and phagolysosomal pH. The results indicate that C. neoformans urease affects the non-lytic exocytosis and intracellular replication of the yeast by modulating phagolysosomal pH thus illustrating a new mechanism of by which this enzyme contributes to virulence.
Many virulence factors used by C. neoformans for survival in mammalian host such as capsule, melanin and phospholipase B1 are also important for the survival of C. neoformans in its natural environment, where it is subject to predation by amoebae [28,53]. The capacity of C. neoformans for mammalian virulence was proposed to result from the fortuitous selection of traits that allow survival in animal hosts by environmental predators [30,54]. To explore whether urease plays a role in the virulence towards amoebae we examined the viability of A. castellanii during the co-incubation with C. neoformans. Consistent to previous studies [28,55], the percentage of dead A. castellanii cells in the presence of C. neoformans (25–26%) was significantly higher than in PBS alone (14.7%) after 48 h co-incubation (Fig 1A). However, the percentage of A. castellanii cells killed by cryptococcal urease-positive and negative strains was similar (Fig 1A). We also examined the survival of C. neoformans with or without urease during co-incubation with A. castellanii. In addition, the buffer solution was supplemented with 7.5 mM urea to test whether the process of ureolysis could improve the survival of C. neoformans during the interaction with A. castellanii. However, no significant difference in survival between strains with urease and without urease was observed in all the tested conditions (Fig 1B). These results suggest that urease, which is a virulence factor for mammalian hosts, is not necessary for virulence in amoebae.
Macrophages play a central role in host response to cryptococcal infection and harbor the organism as an intracellular pathogen during latent infection. To investigate the effect of urease in intracellular pathogenesis we studied the response of macrophage when they were infected with either C. neoformans ure1 deletion strain, its parental H99 or the URE1 complemented strains. Studies have shown that H. pylori urease can affect phagocytosis [49], modulate the recruitment of lysosomal marker LAMP-1 to phagosome and thus prevent phagosomal maturation [50], and stimulates the expression of iNOS to induce nitric oxide generation production in mouse macrophages [51]. Therefore, we first measured the efficiency of phagocytosis of antibody-opsonized cells for the three strains by calculating the phagocytic index after incubation for 2 h (S1A Fig). We then tested if cryptococcal urease can affect the recruitment of LAMP-1 to phagosomes by examining percentage of LAMP-1 positive phagosomes (S1B Fig). To study if cryptococcal urease affects host iNOS expression, we infected macrophages with urease-positive and negative strains and measured the concentration of nitrite, a stable oxidation product of nitric oxide, in the culture supernatant using Griess assay (S1C Fig). Our data shows that cryptococcal urease does not affect the efficiency of antibody-mediated phagocytosis, phagosomal maturation, and nitric oxide production of macrophages during the infection. We also determined the ability of wild-type, urease deletion and complemented strains to survive intracellularly by enumerating the colony forming units (CFU). Our result show that all strains had similar intracellular survival in macrophages after 2 h phagocytosis (S1D Fig).
We studied host cell outcomes for infected macrophages with regards to non-lytic exocytosis. Three subcategories of non-lytic exocytosis, as defined on a previous study [56] were used: complete non-lytic exocytosis (type I), partial non-lytic exocytosis (type II) and cell-cell transfer (type III) (Fig 2A). Macrophages infected with ure1Δ mutant underwent fewer non-lytic exocytosis than those infected with urease producing C. neoformans, in particular partial non-lytic exocytosis (type II) and cell-cell transfer (type III) (Fig 2B). This result implies that the presence of urease has an effect on non-lytic exocytosis.
We hypothesized that if ureolytic activity of urease influenced non-lytic exocytosis, that the addition of urea to the media would also affect the frequency of these events. Consequently, we adjusted the concentration of urea in cell media to 9 mM, which is the level found in plasma from mouse [57], and studied the interaction of C. neoformans and macrophages. Total non-lytic exocytosis events remained higher with urease-positive strains and increased by approximately 23% with increasing concentration of urea (Fig 3A). However, the increase in exocytosis was also noted when macrophages were infected with urease deleted strain (Fig 3A). Hence, urea appeared to affect the frequency of non-lytic exocytosis independently of any effect related to the urea hydrolysis by urease, precluding definitive conclusions.
To investigate whether urease enzymatic activity affected non-lytic exocytosis, we infected BMDM with H99 in the presence of 5 mM of the urease inhibitor acetohydroxamic acid (AHA), a concentration that inhibits 50% of yeast urease activity while presenting minimal toxicity to murine macrophages (S2 Fig), and measured the frequency of non-lytic exocytosis. Addition of AHA decreased non-lytic exocytosis to a level comparable to that observed with the urease deletion mutant (Fig 3A and 3B). This result is consistent with and supports the notion that urease mediated urea hydrolysis modulates the frequency of non-lytic exocytosis.
We observed no significant difference in the frequency of host cell lysis in the presence versus absence of urease when there is no urea supplementation (Fig 3C). However, after the concentration of urea in the culture medium was adjusted to 9 mM cell lysis events of macrophages infected with urease-positive strain decreased by 38%, whereas the event of cell lysis with urease-negative strain increased by 35% (Fig 3C). That in turn led to the significant difference of cell lysis between macrophages containing urease-positive strain and urease-negative strain (Fig 3C).
We compared the intracellular replication of cryptococcal strains with or without urease. For each strain, we analyzed more than 800 internalized cryptococcal cells in five 24-hour time-lapse movies for their ability to replicate intracellularly. In wildtype C. neoformans, 39.6% of cells underwent replication inside macrophages while 63.5% occurred for ure1Δ cells, i.e., ure1Δ cells replicated nearly twice as more than wildtype (Fig 4A). Adding 9 mM urea to the medium resulted in larger difference in intracellular replication between wild-type and urease deletion mutant, with decreased number of replication on H99 and increased number of replication on ure1Δ (Fig 4A). We also investigated whether the presence of urease affected the onset of intracellular replication, and we measured the time of first budding after the cells were phagocytized by macrophages. Urease-positive strains had more cells that started replication later than urease-negative cells although both of the peaks are at 4–6 h after phagocytosis (Fig 4B). Therefore, even among those cells that replicated inside macrophages, urease-positive strains manifested slightly delayed replication compared to ure1Δ strain. The delay became more pronounced when urea (9 mM) was supplemented to the medium, resulting in a peak shift of urease-positive cells from 4–6 h to 8–10 h, implying that cells took longer to begin to replicate in this condition (Fig 4B). Collectively, these data demonstrate that urease ureolytic activity is strongly linked to the delayed onset of cell replication. However, once the cells started to replicate, the doubling time of all strains was very similar (Fig 4C), although ure1Δ strain had approximately 30 min longer doubling time in a standard laboratory condition (Sabouraud broth with shaking at 30 °C) when comparing to H99 and the complemented strain (Fig 4D).
A prior study had shown that prolonged cell cycle progression resulted in cells with larger capsule, which was associated with protection during phagocytosis and enhance intracellular survival [58]. Since urease-positive strains manifested delayed intracellular replication, we investigated whether urease-positive strains had larger capsules after phagocytosis. Consequently, we measured the capsule size of H99, urease deletion and complemented strains harvested from macrophages after 16 h infection. There was no significant difference in capsule size between urease-positive and negative strains inside macrophages (S3 Fig).
The intracellular replication of C. neoformans is tightly correlated to lysosomal damage, such that macrophages with higher numbers of cryptococcal cells manifest greater lysosomal membrane permeabilization [59]. Consequently, we investigated whether lysosomal membrane permeabilization was associated with the urease activity for C. neoformans strains. We stained macrophages with Lysotracker Deep Red, which localizes to and labels acidic organelle such that loss of fluorescence signal indicates lysosomal damage. The number of cells manifesting loss of Lysotracker fluorescence was then quantified by flow cytometry after 24 h of infection. C. neoformans infected macrophages developed loss of Lysotracker signal (Q1 population in Fig 5A and 5B). As negative control, macrophages infected with heat killed H99 manifested no loss of Lysotracker fluorescence [60]. The percentage of Lysotracker-loss macrophages infected with H99, ure1 deletion and complemented strains was highest for ure1Δ-infected macrophages (Fig 5C). This result suggests that host cells which are infected with ure1 deletion strain undergo significantly more lysosomal membrane damage.
We further investigated if different degrees of lysosomal damage were associated with the presence or absence of urease resulted in different degree of apoptosis given that lysosomal damage can release cathepsins into the cytosol and trigger programmed cell death [61]. Therefore, we stained macrophages with SYTOX as an indicator of death cells and F2N12S (dye that indicates membrane potential) to distinguish apoptotic from healthy cells (Fig 5A). There was no significant difference in the percentage of live cells in the sets infected with H99, ure1 deletion and complemented strains respectively (Fig 5D and 5E). However, there was a slight difference of the proportion of apoptotic and death cells between urease-positive strains and urease deletion mutant (Fig 5D and 5E). These results suggest that the presence or absence of urease translates in different intracellular growth rates with the absence of urease being associated with greater phagolysosomal membrane permeabilization, and a shift on how macrophage death occurs.
C. neoformans urease breaks down urea into ammonia and carbon dioxide, and subsequently ammonia reacts with water to produce hydroxyl ions that increase pH. A previous study shows that non-lytic exocytosis is influenced by phagolysosomal pH [62]. In addition, C. neoformans does not grow well in alkaline pH [14]. We hypothesized that the presence of urease would increase phagolysosomal pH, and the alterations in pH would then increase the frequency of non-lytic exocytosis events and affect cryptococcal growth [62,63]. To test this hypothesis, we measured and compared the phagolysosomal pH with the urease-positive or negative strains. Unlike prior studies, we devised a method to measure pH in specific C. neoformans-containing phagolysosomes by conjugating a pH sensitive probe to 18B7 antibody, which binds to cryptococcal capsule [59]. To validate our methodology, we measured phagolysosomal pH associated with the ingestion of Oregon green conjugated 18B7 labeled polystyrene beads, which have been widely used to study phagocytosis in macrophages. Absolute pH was calculated using a pH standard curve obtained from the measurements of pH of phagosomes containing beads (S4A Fig and Materials and methods). The data showed that the acidification started rapidly and by 2 h phagolysosomes had reached the lowest pH (mean = pH 4.5), which remained constant until 4 h (Fig 6A and S4B Fig). Previous study has shown that the bead-containing phagolysosomes reach pH = 5 in 15 min and up to 30 min [64], which is consistent with our result showing average pH of 4.9 in bead-containing phagolysosomes in the first hour after ingestion. We proceeded to measure the pH of phagolysosome containing wild-type, ure1 deletion and complement strains. Phagolysosomes containing wild-type cells had a pH ranging from 4.6 to 5.1, which was consistently higher than those containing urease deletion mutant cells (pH ranging from 4.2 to 4.7) through all the infection periods measured (Fig 6B, S4C and S4D Fig). Our result is consistent with a prior study showing that the pH of phagolysosomes containing live cryptococcal cells is 4.7 after 3 h infection using a different methodology [14]. The phagolysosomes containing the cells from the urease complement strain had a similar pH (4.7–5.1) as those containing wild-type cells and constantly had higher pH than cells deficient in urease at the first two hours of phagocytosis, but the results were inconsistent with those containing wild-type for longer time intervals (Fig 6B and S4C Fig). This finding could result from differences in urease level expression in the complemented strain relative to wild-type expression and/or other uncharacterized factors affecting the reconstituted strain. We also observed considerable pH variation among individual phagolysosomes, which could reflect many factors including differences in the timing of phagocytosis, heterogeneous microenvironments or the cellular position of the phagolysosomes (peripheral vs juxtanuclear) [65] or plain stochastic variation. Overall, phagolysosomes containing both wild-type and urease complemented strains had higher pH than the bead- and ure1Δ strain-containing phagosomes in the first two hours after phagocytosis, consistent with a mechanism whereby urease increases phagolysosomal pH through hydrolysis of urea, which is present in the system from the metabolism of macrophages and from fetal calf serum in the macrophage media.
We also determined the pH of phagolysosome containing heat-inactivated H99 (50 °C for 30 min). Although we expected that it would behave comparably pH of phagolysosomes containing polystyrene beads, the pH was significantly higher (pH 5.1) from 2–4 h (Fig 6C and S4E Fig). Surprisingly, we found that urease activity was not abolished by heat-inactivating to 50 °C for 30 min, since the plating of heat-inactivated cells on Christensen urea agar turned pink (Fig 6D). Of note, the color effect in urea agar was faster with heat-inactivated cells than with alive cells, suggesting that the heating may have liberated the enzyme. Given this result we repeated the experiment with H99 cells killed by heating to 50 °C for a longer period of time (4 h) which was effective in inactivating the urease activity (Fig 6C). The phagolysosomal pH indeed became lower (pH 4.7) after macrophages ingested H99 heat killed at 50 °C for 4 h when compared to 30 min (Fig 6C and S4E Fig). We note that phagolysosomal pH of heat-killed H99 at 50 °C for 4 h was not as low as the pH of phagolysosome containing beads or the urease deletion mutant. A similar phagosomal pH value was observed after ingestion heat-killed urease deficient cells (50 °C for 4 h) (Fig 6C and S4E Fig), suggesting that part of the increase in pH is independent of the presence of urease. We hypothesize these components could derive from leakage of intracellular contents including proteins with functional groups such as amino acids and carboxylic groups that can absorb hydronium ions, and buffer the acid flux such that it does not reach the low pH observed with polystyrene beads. However, the data with urease deficient mutant and urease inactivation by heat killing demonstrates that cryptococcal urease contributes to neutralize and therefore increase the phagosomal pH after ingestion by murine macrophages.
To confirm whether the increase of pH was a result of urease activity, we supplemented the media with urea at different concentrations (9 mM or 50 mM). To establish that urea can freely pass across cell membrane, we measured urea in macrophages using a colorimetric assay. Incubation of macrophages with urea raised their urea content to physiological urea concentration of 9 mM (S5A Fig). Urea supplementation was associated with a more alkaline phagolysosomal pH in macrophages containing wild-type and urease complemented strains as well as heat-inactivated H99 (50 °C for 30 min), but not urease deletion strain, bead or heat killed H99 (50 °C for 4 h) (Fig 5E and S5B Fig). These results suggest that the increased phagolysosomal pH is associated with urease degradation of available urea.
To investigate how pH could affect the intracellular growth results we studied the growth of C. neoformans as a function of the pH. We used the growth curves to determine three characteristic growth values i.e. growth rate represented by the maximum slope, length of lag phase, and the maximum cell growth and compared them among strains (Fig 7A and 7B). Our experiment displayed an inverse relationship between the growth rate of all tested strains and increasing pH. The growth rate decreased approximately 3-fold from pH 4.2 to 5.6. Hence, C. neoformans grew best in the most acidic milieu, a finding that combined with phagolysosomal pH measured for urease sufficient and deficient strains suggests that the effect of urease on cryptococcal intracellular replication is due to its effect in neutralizing pH since pH affects yeast growth rate.
The strain ure1 Δ had shorter lag phase than H99 and URE1-complemented strains at all pH tested (Fig 7C). One possible explanation is that there is a metabolic cost of producing the enzyme, especially when adjusting to the nutrient-limited medium we used in this particular experiment. However, once the urease-positive strains adapt to the environment, the rate of growth can return to maximum. Interestingly, the maximum cell growth of ure1Δ was similar to that of H99 and URE1-complemented strains in acidic pH, but gradually decreased closer to neutral pH (Fig 7D). Therefore, we questioned if the mammalian physiological pH affected the growth of ure1Δ, so the growth curves of H99, ure1Δ and URE1-complemented strains were determined in minimal medium buffered at pH 7.4. The result showed that ure1Δ had a severe growth defect at pH 7.4 (Fig 7E) while all strains grow equally in unbuffered minimal medium (Fig 8A), suggesting that urease is required for neutral and alkaline tolerance.
To study further the association of phagolysosomal pH and the intracellular replication of C. neoformans, we added the weak base ammonium chloride to the medium, which is known to neutralize phagosomal acidity and inhibit cryptococcal intracellular growth [14,62,66], and measured intracellular replication of urease deletion mutant in a 24 h time-lapse movie. The addition of ammonium chloride retarded intracellular replication of C. neoformans independent of the presence of urease (Fig 7F), providing strong support for the notion that the higher phagosomal pH was the cause of the retarded cryptococcal intracellular replication.
Urease is not only able to elevate the pH of microenvironment, but it also serves as a nitrogen source for pathogenic microbes such as Actinomyces naeslundii and Bacillus cereus during infection by hydrolyzing urea to ammonia [67–69]. A prior study showed that urease activity is required for cryptococcal growth in agar medium with urea as the only nitrogen source [45]. However, the role of cryptococcal urease in nutrition and metabolism has not been fully explored. Consequently, we tested whether C. neoformans could use urea as both a nitrogen and carbon source and whether this was urease-dependent. The growth of the three strains were very similar in minimal medium containing glycine as sole nitrogen source and glucose as a carbon source (Fig 8A). When urea was the sole nitrogen source, there was no growth of urease deletion mutant up to 72 h (Fig 8D), consistent to prior results. The addition of ammonium salt partially complemented the growth defect of urease deletion mutation, suggesting that the growth defect was the result of an inability to produce ammonia (Fig 8E). On the other hand, although urea could serve as a nitrogen source, cryptococcal strains were not able to grow when urea was the sole carbon source (Fig 8F). Taken together, our results show that C. neoformans has an ability to utilize urease to hydrolyze urea into ammonia and use it as nitrogen source, but C. neoformans cannot utilize urea as carbon source for growth.
C. neoformans brain invasion can occur by carriage in macrophages in a Trojan Horse-like mechanism or through transcytosis of endothelial cells [19,23,26,70–77]. Previously, it was reported that urease-negative strains cannot reach the brain from the lungs [44,46]. It is still not clear if this defect was due to ineffective dissemination from the lung or whether this reflected the fact that urease-negative strains are less effective in crossing the BBB [44,46]. Furthermore, it was not clear if Trojan-horse transport inside macrophages was affected by the presence and absence of urease. Since urease retards the intracellular replication, it could promote the coexistence and persistence of C. neoformans within macrophages and thus increase the chance for dissemination by a Trojan Horse-like mechanism. Alternatively, since C. neoformans urease also induces non-lytic exocytosis, this could facilitate escape of C. neoformans from macrophages in the lung or the bloodstream and enhance brain invasion of free yeasts by transcytosis, which is facilitated by urease. Therefore, we hypothesized that macrophages infected with urease-positive strain would be more efficient at the brain invasion than ure1 infected macrophages. To test this, we injected mice with H99- and ure1Δ-infected macrophages and quantified brain and lung fungal burden by CFU at 72 h post-infection. We found lower CFU in the brain of mice infected with BMDM containing ure1 mutant relative to H99-containing BMDM (Fig 9). In contrast, lung CFU were comparable for mice given macrophages containing either strain, suggesting that brain invasion by Trojan-Horse mechanisms depends on active urease.
Urease is an important virulence factors of C. neoformans [40]. However, most studies have focused on urease role in brain invasion and its effects on the host immune response. Cryptococcal urease facilitates transmission of C. neoformans across the blood-brain barrier [44–46] and polarizes the immune system to a Th2 response, which translates into greater fungal burden in lung in mouse model [52]. Yet the effect of urease in macrophage interaction has not been explored despite the fact that the outcome of the interaction of C. neoformans with macrophages is a key determinant of the outcome of infection [13,78,79]. In this study, we analyzed the role of urease in C. neoformans-macrophage interactions. Our results provide new insights on how this enzyme can affect the pathogenesis of Cryptococcus spp since we show that urease influences the intracellular growth of C. neoformans, affects non-lytic exocytosis from macrophages, is critical for growth at mammalian physiological pH and confers upon the yeast the potential for using urea as a nitrogen source in nutrition.
Urease can break down urea to produce ammonia, which in turn raises pH. In our study, the measurements of phagolysosomal pH show that cryptococcal urease contributes to buffering acidic pH in the phagolysosome, which is almost certainly a consequence of the hydrolysis of urea. Urea is present in the fetal bovine serum in the culture medium and easily crosses cell membranes. Urea is also a product of macrophages arginase catalysis, which provides C. neoformans an additional intracellular source of substrate for urease. Moreover, human body fluids normally contain between 2.5 to 7.1 mM urea that is evenly distributed in all body compartments [80–82]. These concentrations make it feasible for C. neoformans to utilize urea in the host and alter its microenvironmental pH such as phagolysosome. Phagolysosomal alkalization has also been observed with other urease-positive microbes. Mycobacterial urease contributes to the alkalization of the phagolysosomal pH in resting macrophages, but its effect is not sufficient to neutralize the pH in the more acidic phagolysosome of activated macrophages [83]. In contrast, the presence of cryptococcal urease was sufficient to raise pH by an average of 0.4 pH units. Helicobacter pylori urease is also known to significantly elevate phagolysosomal pH as well as retard recruitment of Lamp-1, a marker for late endosome and lysosome [50]. However, unlike H. pylori cryptococcal urease did not affect the LAMP-1 acquisition.
Phagosome acidification is usually considered an important component of the antimicrobial machinery [84], but C. neoformans grows best in acidic environments while other pathogens’ growth is inhibited by phagolysosome acidification [63,85–87]. Our analysis shows that even inside the mammalian macrophage C. neoformans grows at a higher rate in pH 5–6. Our results, together with studies of other groups, showed that neutralizing phagolysosomes by treating the macrophages with ammonium chloride inhibited intracellular growth of C. neoformans [14,63]. Therefore, phagolysosomal acidification does not appear to be an important element of in the control of C. neoformans by macrophages. However, we show that urease is required for the optimal growth of C. neoformans at physiological pH. Therefore, we describe a previously unknown role of urease in neutral/alkaline pH tolerance.
The majority of intracellular cryptococcal cells in the urease-positive population did not undergo replication immediately upon entry into macrophages and ure1 deletion strain initiated replication earlier than wild-type counterparts. This effect was pronounced when the culture medium was supplemented with urea. We attribute this effect to an increase in the phagolysosomal pH from 4.4 to 4.8, a pH that reduced the maximum growth rate of C. neoformans. Together, urease retards growth in macrophages in vitro, and at first glance this is the opposite of what would be expected from a virulence factor. However, this effect has to be considered in the context of the larger picture of cryptococcal pathogenesis. Urease-positive strains still possessed the ability of resisting killing by macrophages and delayed replication could promote a quiescent state intracellularly, which may be associated with persistence of infection [88]. A previous study reported a strong correlation between intracellular replication of C. neoformans and lysosomal damage due to the increased number of yeasts in macrophages [59]. Indeed, macrophages containing urease-positive cells manifested less phagolysosomal permeabilization than those containing urease-negative cells. Loss of phagolysosomal membrane integrity could be expected to benefit to C. neoformans by allowing the fungus access to host cytosolic nutrients. However, many bacterial pathogens thrive in vacuoles or phagosome rather than nutrient rich cytoplasm and have developed strategies to maintain phagosomal membrane integrity to avoid immune surveillance pathway and eventually prevent inflammasome-mediated lytic cell death called pyroptosis [89]. A temporary maintenance of membrane integrity could contribute to persistent infections by prolonging the interaction of macrophages and cryptococci. Consistent with this, our results show that when the medium is supplemented with urea, urease-positive cryptococcal cells manifest pronounced growth retardation and cause fewer events of host cell lysis. This correlation suggests that the growth retardation associated with urease mediated alkalization results in fewer yeasts in macrophages, which in turn protects C. neoformans from humoral immune responses and facilitates persistent infection and dissemination by reducing the likelihood of lytic exocytosis.
The cellular and molecular mechanism of the non-lytic exocytosis is poorly understood. It is a highly choreographed process where both host and pathogen factors are involved [17,24]. Previous studies have shown that cryptococcal capsule and phospholipase B1 contribute to non-lytic exocytosis [17,90]. Host cell membrane protein annexin A2 and signaling kinase ERK5 have also been identified to regulate non-lytic exocytosis [27,91]. Repeated cycles of actin polymerization that form around cryptococci-containing phagosome could potentially inhibit non-lytic exocytosis [92]. Moreover, phagolysosome neutralization by the addition of weak base ammonium chloride and chloroquine increased the frequency of non-lytic exocytosis events [62]. Here we identify urease as new fungal factor that modulates non-lytic exocytosis. Moreover, chemical inhibition of urease enzymatic activity decreased the frequency of non-lytic exocytosis, suggesting that the effect was related to urea hydrolysis. The most likely mechanism is that the pH alteration caused by ureolytic reaction contributes to increasing the frequency of non-lytic exocytosis, which is consistent with our observation that urease activity raised phagolysosomal pH. However, the mechanism on how phagolysosomal pH influences the non-lytic exocytosis remains to be elucidated.
Given that phagolysosomal alkalization can increase non-lytic exocytosis [62], we evaluated whether higher concentrations of urea would increase the frequency of this phenomenon. While increase in exogenous urea increased phagolysosomal pH in those containing urease-positive cells, we also observed increased non-lytic exocytosis in macrophages containing urease-negative cryptococcal cells. Therefore, the effect was not entirely dependent on urease-mediated alkalization, but could also be affected by urea, which can promote the fusion of vesicles with bilayer lipid membrane and thus induce exocytosis [93]. Hence, one possible mechanism for the increased non-lytic exocytosis observed for urease-negative cells is that urea encourages phagolysosome-cell membrane fusion to disgorge yeast cells.
The phagolysosomal pH was measured by conjugating a pH sensitive probe to 18B7 antibody, which binds to cryptococcal capsule, a method that can adapted to any system using antibody-mediated phagocytosis. It is noteworthy that we observed considerable pH variation among individual phagolysosomes. This heterogeneity has been observed in many other studies, and could be caused by many factors [65,94–96], such as differences in the timing of phagocytosis. Despite using centrifugation to enhance the yeast cells contact with macrophages, yeast cells might attach to macrophages but not be engulfed synchronously. A recent study also report that the position of lysosomes determines their pH, such that peripheral lysosomes are less acidic than juxtanuclear ones [65]. Phagolysosomal pH variation could also be attributed to the heterogeneity in protein composition of V-ATPases and NADPH oxidase among individual phagolysosomes [95]. Differences in the metabolic state or age of the infecting cells could also contribute to phagolysosomal pH heterogeneity. In this regard, variability in the infecting Salmonella population resulted in heterogeneous macrophage response [97]. Finally, it is possible that the heterogeneity in phagolysosomal pH reflects the outcome of the individual battles between C. neoformans and macrophages that are fought phagolysosome-to-phagolysosome, such that in some phagolysosomes the microbe gains ascendancy while in others it is suppressed.
C. neoformans was able to utilize urea as nitrogen source for growth. This metabolic process was entirely urease dependent since the urease deficient strain was unable to grow in medium where urea was the sole nitrogen source. Supplementation with ammonium salt partially rescued the growth of urease-negative strain, suggesting that ammonia generated from ureolytic activity, not urea itself, is the actual nitrogen source for the growth of C. neoformans. We also showed that urea could not serve as the sole source of carbon for C. neoformans. Therefore, we propose that C. neoformans urease activity may also provide an important nutritional function for in vivo under nitrogen-limited conditions since urea diffuses easily in tissues and macrophages can generate urea. Of note, two conundrums arise from our studies. Firstly, the partial rescue suggests that cryptococcal urease is further involved in the metabolism of ammonia. Secondly, it remains to be investigated why urease is required for growth at physiological pH. In contrast to other virulence factors such as the capsule, melanin, and phospholipase, a deficiency in urease did not increase the vulnerability of C. neoformans for amoebae. Hence, the main role of urease in C. neoformans in its natural environment appears to be nutritional in nature and this enzyme provides an example on how a protein involved in nutrition acquisition can serve a fortuitous role during pathogenesis as a modulator of virulence.
Taken together, we propose the following model for the role of urease in intracellular pathogenesis. Urease secreted by C. neoformans into phagolysosome hydrolyses urea, releases ammonia and increases the pH of this compartment by approximately half of a pH unit, which is sufficiently to retard the replication of C. neoformans inside macrophages. The growth retardation in turn leads to fewer macrophages with phagolysosomal membrane permeabilization and fewer host cell lysis. In parallel, C. neoformans urease induces non-lytic exocytosis events of macrophages. Crossing of the blood brain barrier can be done by yeast cells in a transcytosis event or inside macrophages in a Trojan horse-like mechanism [19,23,26,70–77]. For yeast cells crossing the blood brain barrier alone urease has been shown to promote brain invasion [44,46]. In addition, urease contributes to optimal C. neoformans growth at physiological pH, which may facilitate the its extracellular growth and dissemination to tissues. Urease can also play a role in cryptococcal nitrogen metabolism by providing a source of nitrogen, and that could support the long-term survival of C. neoformans in nitrogen-limited conditions such as macrophages. Our observation shows that the presence of urease promotes non-lytic exocytosis, delays intracellular replication, allows for use of an abundant nitrogen source and facilitates growth at mammalian physiological pH. Therefore, urease can affect all of the types of blood-brain barrier crossing mechanisms by providing more extracellular yeasts and increased pH fitness for transcytosis, as well as increasing residence time in macrophages, with the latter promoting macrophage-associated crossings. This supposition was supported by the observation of higher fungal dissemination to the brain in mice injected with macrophage containing urease producing C. neoformans. Furthermore, it is consistent with the model proposed by others that elevating the frequency of non-lytic exocytosis by altering host cell signaling reduces dissemination, presumably by limiting the opportunity for Trojan horse transport [27]. Overall, we propose that urease helps C. neoformans to both persist in and exit from macrophages, events that could facilitate the dissemination of the pathogen to brain through macrophage-dependent transport mechanisms. We anticipate that the findings here for C. neoformans may also be relevant to other urease-positive fungal pathogens such as Aspergillus fumigatus and Histoplasma capsulatum, two pathogens that modulate phagolysosomal pH and for which acidification is critical for control of infection [87,98,99]. Analysis of urease effects on macrophages is likely to be a fertile area of investigation for the many fungal pathogens that express this enzyme.
All animal procedures were performed with prior approval from Johns Hopkins University (JHU) Animal Care and Use Committee (IACUC), under approved protocol numbers M015H134. Mice were handled and euthanized with CO2 in an appropriate chamber followed by thoracotomy as a secondary means of death in accordance with guidelines on Euthanasia of the American Veterinary Medical Association. JHU is accredited by AAALAC International, in compliance with Animal Welfare Act regulations and Public Health Service (PHS) Policy, and has a PHS Approved Animal Welfare Assurance with the NIH Office of Laboratory Animal Welfare. JHU Animal Welfare Assurance Number is D16-00173 (A3272-01). JHU utilizes the United States Government laws and policies for the utilization and care of vertebrate animals used in testing, research and training guidelines for appropriate animal use in a research and teaching setting.
The C. neoformans strains were used in this study are C. neoformans var. grubii serotype A strain H99, ure1Δ (derived from H99 and lacking urease) and ure1Δ::URE1 (complemented urease mutant). All the strains were kindly provided by Dr. John Perfect (Duke University, USA) and have been described previously (Cox et al. 2000). The urease production phenotype of these strains was validated by using Christensen’s urea agar (2% urea, 1.5% agar, 0.2% KH2PO4, 0.1% peptone, 0.1% dextrose, 0.5% NaCl, 0.0012% phenol red). Cryptococcal cells were cultivated in Sabouraud dextrose broth with shaking (120 rpm) at 30 °C for overnight (16 h). Heat inactivated or killed C. neoformans was prepared for various experiments by incubating the cells at 50 °C for 30 min or 4 h.
To study the effect of pH on cryptococcal growth, the yeast cells were grown in minimal medium (15 mM dextrose, 10 mM MgSO4, 29.4 mM KH2PO4, 13 mM glycine, 3 μM thiamine-HCl) buffered with 100 mM citrate buffer (Sodium citrate and citric acid) at various pH ranging from 4.2 to 5.4 at 0.2-pH unit increments. The pH of the medium was measured using an Accumet Basic AB15 pH meter (Thermo fisher Scientific, Waltham, MA). The pH was further verified by use of MColorpHast pH-indicator strips (EMD Millipore, Jaffrey, NH) before and after the growth assay to ensure the pH keep constant throughout the assay.
To study the utilization of urea in C. neoformans, cells were grown in minimal medium at pH 5.5 with substitution of 7.5 mM urea or ammonium sulfate for glycine as sole nitrogen source and substitution of 7.5 mM urea for dextrose as sole carbon source. Cryptococcal strains were also grown in Sabourand broth to determine their growth and doubling time.
Growth studies were done using a Bioscreen C plate reader (Growth Curves USA) starting the cultures with 105 yeast cells per well in honeycomb plate in different conditions mentioned above at 30 °C and measuring cell density for 72 h.
Bone-marrow derived macrophages (BMDM) were isolated from the marrow of hind leg bones of 5- to 8-wk-old C57BL-6 female mice (Jackson Laboratories, Bar Harbor, ME. For the differentiation, cells were seeded in 100 mm TC-treated cell culture dishes (Corning, Corning, NY) in Dulbecco’s Modified Eagle medium (DMEM; Corning) with 20% L-929 cell-conditioned medium, 10% FBS (Atlanta Biologicals, Flowery Branch, GA), 2mM Glutamax (Gibco, Gaithersburg MD), 1% nonessential amino acid (Cellgro, Manassas, VA), 1% HEPES buffer (Corning), 1% penicillin-streptomycin (Corning) and 0.1% 2-mercaptoethanol (Gibco) for 6–7 days at 37 °C with 9.5% CO2. Fresh media in 3 ml were supplemented on day 3 and the medium were replaced on day 6. Differentiated BMDM were used for experiments within 5 days after completed differentiation. Urea in 9 mM or 50 mM were supplemented in the medium during infection of macrophages with C. neoformans in some of the experiments. The amount of urea inside macrophages were measured using urea colorimetric assay (Sigma-Aldrich, St. Louis, MO) according to the manufacturer’s instruction.
J774.16 cells, which were obtained from the American Type Culture Collection (ATCC), is a murine (BALB c, haplotype H-2d) macrophage-like cell line derived from a reticulum sarcoma. J774.16 cells were maintained in DMEM with 10% NCTC109 medium (Gibco), 10% FBS, 1% nonessential amino acid, 1% penicillin-streptomycin at 37 °C with 9.5% CO2.
Acanthamoeba castellanii strain 30234 was obtained from the American Type Culture Collection (ATCC) was maintained in peptone-yeast extract-glucose (PYG) broth (ATCC medium 712) at 25 °C according to instructions from ATCC.
C. neoformans strains were grown overnight in Sabouraud broth, and diluted into 5 × 107 cells in 2 ml of rapid urea broth (RUH) developed by Roberts [100] and adapted by Kwon-Chung [101]. Different concentrations (1.25–40 mM) of urease inhibitor acetohydroxamic acid (AHA) were added. Cells were incubated at 37 °C for 7 and 24 h. In parallel, H99 and ure1Δ strains were grown without AHA as positive and negative controls. After incubation, cells were collected by centrifugation and 200 μl of supernatant were transferred to 96-well plate. The absorbance of the supernatant was measured at 570 nm using EMax Plus microplate reader (Molecular Devices). The assay was performed in duplicate for each time interval.
BMDM were seeded (5 × 104 cells/well) on poly-D-lysine coated coverslip bottom MatTek petri dishes with 14mm microwell (MatTek Brand Corporation) in medium containing 0.5 μg/ml lipopolysaccharide (LPS; Sigma-Aldrich), 100 U/ml gamma interferon (IFN-γ; Roche). Cells were then incubated at 37 °C with 9.5% CO2 overnight. On the following day, macrophages were infected with cryptococcal cells (1.5 × 105 cells/well) in the presence of 10 μg/ml monoclonal antibody (Mab) 18B7. After 2 h incubation to allow phagocytosis, culture was washed five times with fresh medium to remove extracellular cryptococcal cells. Images were taken every 4 min for 24 h using a Zeiss Axiovert 200M inverted microscope with a 10x phase objective in an enclosed chamber under conditions of 9.5% CO2 and 37 °C. For some of the experiments, 9 mM urea was added into the BMDM culture for overnight incubation, or both BMDM and cryptococcal cells were pretreated with 20 mM ammonium chloride or 5 mM acetohydroxamic acid (AHA) for 30 min before phagocytosis. The chemicals were also present during both the 2 h incubation to permit phagocytosis and the 24 h incubation during time-lapse imaging.
Phagolysosomal pH was measured using ratiometric fluorescence imaging involving the use of pH-sensitive probe Oregon green 488. Oregon green 488 was first conjugated to monoclonal antibody 18B7 using Oregon Green 488 Protein Labeling Kit (Molecular Probes, Eugene, OR). The Oregon Green 488 dye has a succinimidyl ester moiety that reacts with primary amines of proteins to form stable dye-protein conjugates. The labeling procedure is according to the manufacture’s instruction. BMDM were plated (1.25 × 105 cells/well) on 24-well plate with 12 mm circular coverslip. Cells were cultured with completed BMEM medium containing 0.5 μg/ml LPS and 100 U/ml IFN-γ; as well as supplemented with or without urea at 9 mM or 50 mM, and then incubated at 37 °C with 9.5% CO2 overnight. Prior to infection, macrophages were placed at 4 °C for 15 min. In the meanwhile, live, heat inactivated, heat killed cryptococcal strains or anti-mouse IgG coated polystyrene bead (3.75 × 106 cells or beads/ml) were incubated with 10 μg/ml Oregon green conjugated 18B7 Ab for 15 min. Macrophages were then infected with Oregon green conjugated 18B7-opsonized samples in 3.75 × 105 cells or beads per well. Cells were centrifuged immediately at 1200 rpm for 1 min and culture were incubated at 37 °C for 10 min to allow phagocytosis. Extracellular cryptococcal cells or beads were removed by washing three times with fresh medium. Samples on coverslip were collected at 1, 2, 3, 4 h after phagocytosis by washing twice with pre-warmed HBSS and placing upside down on MatTek petri dish (MatTek, Ashland, MA) with HBSS in the microwell. Images were taken by using Olympus AX70 microscopy (Olympus, Center Valley, PA) with objective 40x at dual excitation 440 nm and 488 nm, and emission 520 nm. Images were analyzed using MetaFluor Fluorescence Ratio Imaging Software (Molecular Devices, Downingtown, PA). Relative phagolysosomal pH was determined based on the ratio of 488 nm/440 nm. The relative pH was converted to absolute pH by obtaining the standard curve in which the images are taken as above but intracellular pH of macrophages was equilibrated by adding 10 μM nigericin in pH buffer (140 mM KCl, 1 mM MgCl2, 1 mM CaCl2, 5 mM glucose, and appropriate buffer ≤ pH 5.0: acetate-acetic acid; pH 5.5–6.5: MES; ≥pH 7.0: HEPES. Desired pH values were adjusted H using either 1M KOH or 1M HCl). Buffers were used at pH 3–7.5 using 0.5-pH unit increments.
J774.16 cells were plated (2 × 106 cells/well) on 6-well plate with completed DMEM containing 0.5 μg/ml LPS, 100 U/ml IFN-γ, with or without urea at 9 mM, and incubated at 37 °C with 9.5% CO2 overnight. On the following day, cryptococcal cells were stained with 0.0015% Uvitex 2B (Polysciences) for 1 min and wash once with medium. Macrophages were infected with Uvitex 2B-stained cryptococcal cells (1 × 106 cells/well) in the presence of 10 μg/ml 18B7 for 24 h. After 24 h infection, Lysotracker deep red (Thermo Fisher Scientific) at 1 nM is added to the culture and incubated for 1 h. Cells were harvested from plates and washed once with HBSS. Anti-mouse CD11b-PE (1:1000) (M1/70, eBioscience, San Diego, CA) was added and incubated for 5 min and washed once with HBSS. SYTOX and F2N12S (Thermo Fisher scientific) were added 5 min before flow cytometry analysis. Single color and fluorescence minus one (FMO) controls were used for fluorescence spectral compensation and gating. Flow cytometry analysis were performed by LSRII (BD Biosciences, San Jose, CA). Data were analyzed using FlowJo software (Ashlan, OR).
BMDM cells (5 × 104 cells/well) were seeded in 96-well plates with BMDM containing 0.5 μg/ml LPS and 100 U/ml IFN-γ for overnight. To initiate the phagocytosis, C. neoformans with 1.5 × 104 cells in the presence of 10 μg/ml 18B7 mAb were added in each well of BMDM culture. The culture plates were centrifuged at 1200 rpm for 1 min to settle yeast cells on the monolayer of macrophage culture. After 2 h infection, phagocytized cryptococcal cells were released by lysing the macrophages with sterilized water. The lysates were serially diluted, plated onto Sabouraud agar and incubated at 30 °C for 48 h for colony form unit (CFU) determination. This experiment was performed in triplicates for each strain.
The survival of C. neoformans in amoebae culture was performed as described previously [55]. Briefly, A. castellanii were washed twice with DPBS (Corning) and diluted in DPBS to appropriate density. A. casterllanii cells (1 × 104 cells/well) were added to 96-well plates and allowed to adhere for 1 h at 25 °C. C. neoformans cells were washed twice with DPBS and diluted in DPBS to appropriate density. Fungal cells (1 × 104) were added to wells containing amoebae or control wells containing DPBS alone, and the plates were incubated at 25 °C. At 0, 24, and 48 h, the amoebae were lysed by pulling the culture through a 27-gauge syringe needles five to seven times. The lysates were serially diluted, plated onto Sabouraud agar and incubated at 30 °C for 48 h for colony form unit (CFU) determination. Two different conditions were tested with DPBS supplemented with or without 7.5 mM urea. Two biological independent experiments were performed for each strain and condition. Viability of A. castellanii was also determined under the same conditions and time intervals by adding 1:80 dilution of Trypan Blue stain. The percentage of dead amoebae was determined by counting the number of Trypan Blue stained cells per total cell number counted. Minimal of 100 cells were counted. Control wells contain A. castellanii without C. neoformans. Two biological independent experiments were performed for each strain and condition.
After 16 h infection, macrophage-internalized cryptocooccal strains were released by lysing the host cell with sterile water. Cryptococcal cells were washed twice with PBS. The capsule was visualized by India ink negative staining by mixing cell samples with equal volume of India ink on glass slides and spreading the smear evenly with coverslips. The images with a minimum 100 randomly chosen cells was taken by using Olympus AX70 microscopy with 100x oil objective using the QCapture Suite V2.46 software (QImaging, Surrey, Canada). The areas of cell body and whole cell (cell body plus capsule) were measured using image J software. The capsule area was calculated by subtracting the area of whole cell from that of cell body. Three biological independent experiments were performed for each strain.
BMDM (1.5 × 105 cells) were seeded on 12 mm circular coverslip in 24-well plate with completed BMDM containing 0.5 μg/ml LPS and 100 U/ml IFN-γ for overnight. C. neoformans with 1.5 × 105 cells in the presence of 10 μg/ml Alexa Fluor 568 conjugated 18B7 mAb were then added into BMDM culture. The culture plates were centrifuged at 1200 rpm for 1 min to settle yeast cells on the monolayer of macrophage culture. After 10 min, 30 min, 1 h and 2 h infection, cells were fixed with 4% paraformaldehyde (Electron Microscopy Sciences, Hatfield, PA), followed by permeabilization with 0.3% Triton X-100 in PBS for 5 min and incubated with blocking solution containing 3% bovine serum albumin and 1:250 dilution of purified rat anti-mouse CD16/CD32 (Mouse BD FC Block; BD Pharmingen, San Diego, CA) for 45 min. Cells were next incubated with 1:50 dilution of Alex Fluor 488 conjugated anti-mouse Lamp-1 (rat IgG2a monoclonal antibody 1D4B; Santa Cruz Biotechnology Inc., Santa Cruz, CA) at 4 °C overnight and then washed three times with 1×PBS for 5 min each time. Coverslips were mounted using ProLong Gold Antifade Mountant (Thermo Fisher Scientific) and cured for 24 h at room temperature. The images with a minimum 100 randomly chosen cells was acquired by Zeiss Axiovert 200M inverted microscope with a 40x objective. Z-stacks were taken at 1 μm intervals through entire macrophage. Phagosome-lysosome fusion was considered to take place when there is co-localization of cryptococcal capsule (Alexa Fluor 568) and Lamp-1 (Alexa Fluor 488). Two biological independent experiments were performed for each strain.
BMDM cells (1 × 105 cells/well) were activated by LPS (0.5 μg/ml) and IFN-γ (100 U/ml) for overnight in 96-well plates. To initiate infection, C. neoformans (1 × 105 cells) with 10 μg/ml 18B7 mAb were added and settled down on macrophage monolayer culture using centrifugation at 1200 rpm for 1 min. After 24 h infection, culture supernatant in 100 μl was collected and equal volume of Griess reagent (1: 1 ratio of 0.1% naphtylethylenediamine dihydrochloride and 1% sulfanilamide in 5% H3PO4) was added. The mixture was incubated in the dark for 10 min at room temperature. The absorbance of the mixture was measured at 562 using EMax Plus microplate reader (Molecular Devices). Nitrite concentration was determined from a standard curve constructed with 0 μM–50 μM sodium nitrite. Two biological independent experiments were performed for each strain and condition.
Animal studies were performed using 6 to 8-week-old female C57BL/6 mice. Cryptococcal strains were grown for 2 days at 37 °C with shaking at 180 rpm in Sabouraud broth. Cells were washed with PBS and resuspended to 1 × 107 cells/ml in BMDM medium. Cryptococcal cells in 1 ml was added to BMDM in triplicates (1 × 107 cells per replicate) together with opsonizing 18B7 (final concentration of 10 μg/ml). After 1 h incubation to allow phagocytosis, extracellular cryptococcal cells were washed with HBSS and infected BMDM were detached with CellStripper (Corning), collected by centrifugation. The infected BMDM were then resuspended in USP grade sterile saline solution (BD, Franklin Lakes, NJ). Each mouse was injected i.v. with 200 μl of cell suspension. Mice were anesthetized with 2% isoflurane anesthesia followed by retroorbital injection of BMDM suspension, according to standard procedures [102]. Infected BMDM were lysed with sterilized water and cell lysis were plated in YPD plates to confirm inoculum CFU. After 72 h post-injection, mice were euthanized, lung and brain were isolated and homogenized by passing through a 100 μm filter. Homogenates were plated onto YPD agar for CFU enumeration. Three mice per strain per experiment were studied and two independent biological experiments were performed.
Pairwise comparisons depicted in Fig 6 are Mann-Whitney U test with both wild-type and ure1 complement strains comparing to ure1 deletion mutant. One-way ANOVA, followed by Tukey’s multiple-comparison test was used to evaluate the statistical parameters of characteristic growth values. For categorical data, Fisher’s exact test was used when sample sizes were less than 1000 or chi-square test was used when sample sizes were larger than 1000. All other continuous data were analyzed by Student’s t test.
|
10.1371/journal.ppat.1004078 | HIV-Infected Individuals with Low CD4/CD8 Ratio despite Effective Antiretroviral Therapy Exhibit Altered T Cell Subsets, Heightened CD8+ T Cell Activation, and Increased Risk of Non-AIDS Morbidity and Mortality | A low CD4/CD8 ratio in elderly HIV-uninfected adults is associated with increased morbidity and mortality. A subset of HIV-infected adults receiving effective antiretroviral therapy (ART) fails to normalize this ratio, even after they achieve normal CD4+ T cell counts. The immunologic and clinical characteristics of this clinical phenotype remain undefined. Using data from four distinct clinical cohorts and three clinical trials, we show that a low CD4/CD8 ratio in HIV-infected adults during otherwise effective ART (after CD4 count recovery above 500 cells/mm3) is associated with a number of immunological abnormalities, including a skewed T cell phenotype from naïve toward terminally differentiated CD8+ T cells, higher levels of CD8+ T cell activation (HLADR+CD38+) and senescence (CD28− and CD57+CD28−), and higher kynurenine/tryptophan ratio. Changes in the peripheral CD4/CD8 ratio are also reflective of changes in gut mucosa, but not in lymph nodes. In a longitudinal study, individuals who initiated ART within six months of infection had greater CD4/CD8 ratio increase compared to later initiators (>2 years). After controlling for age, gender, ART duration, nadir and CD4 count, the CD4/CD8 ratio predicted increased risk of morbidity and mortality. Hence, a persistently low CD4/CD8 ratio during otherwise effective ART is associated with increased innate and adaptive immune activation, an immunosenescent phenotype, and higher risk of morbidity/mortality. This ratio may prove useful in monitoring response to ART and could identify a unique subset of individuals needed of novel therapeutic interventions.
| The CD4/CD8 ratio, a hallmark of the collection of T cell defects related to aging –“immunosenescence”- and a predictor of mortality in the general population, often fails to normalize in an important proportion of HIV-infected individuals with adequate CD4+ T cell recovery after ART initiation. However, the immunological and clinical characteristics of this clinical phenotype have not been elucidated. Herein we show that during treated HIV infection, expansion of CD8+ T cells, reflected as a low CD4/CD8 ratio, identifies a subgroup of individuals with a number of immunological abnormalities and a poor prognosis. These subjects exhibit increased innate and adaptive immune activation, an immunosenescent phenotype, CD4+ and CD8+ imbalance in the gut mucosa and higher risk of morbidity and mortality. In contrast, those who normalize the CD4/CD8 ratio have traits of a healthy immune system. We observed that early ART initiation might contribute to more rapid and robust CD4/CD8 ratio normalization compared to later initiation. Hence, the CD4/CD8 ratio might help to further discriminate the risk of disease progression of successfully treated HIV-infected individuals, and a successful response to ART may require both normalization of the peripheral CD4+ T cell count and the ratio of CD4+ to CD8+ T cell counts.
| It is now anticipated that HIV-infected adults who have access to modern antiretroviral therapy (ART) should be able to suppress HIV replication indefinitely. Although treatment-mediated increases in the peripheral CD4 count are associated with reduced morbidity and mortality, compared to age-matched individuals without HIV infection, those on ART have a higher risk of morbidity and mortality. This risk is predicted in part by the on therapy CD4 count, although achieving an apparent normal CD4 count may not fully restore health [1]–[5]. Indeed, it has been shown that even those treated patients with CD4+ T cell counts above 500 cells/mm3, a further CD4+ T cell count increase is still associated with a slight benefit in terms of mortality [6]. The decreased life expectancy during ART-mediated viral suppression is largely explained by a higher than expected risk of non-AIDS-morbidity, a term that entails a group of conditions generally associated with aging, including cardiovascular, renal, liver, neurologic, and bone disease, as well as cancer [4], [7], [8].
While the mechanisms driving the increased burden of aging-associated disease in HIV-infected individuals are not fully understood, an emerging body of evidence suggests that persistent innate and adaptive immune dysfunction and/or activation are major risk factors [9]–[12]. Many of the immunologic abnormalities that persist during therapy are similar to those observed in the elderly, raising the hypothesis that age-associated decline in immune function (“immunosenescence”) contributes to disease progression and adverse outcomes [13]–[16]. Markers of innate immune activation [e.g. interleukin (IL)-6, high-sensitivity C reactive protein (hs-CRP) and soluble CD14 (sCD14)], coagulation (fibrinogen, D-dimers), bacterial translocation (lipopolysaccharide), and T cell activation (HLADR and CD38 co-expression) are elevated despite effective ART and associated with subsequent morbidity and mortality, even after adjustment for CD4+ T cell count [17]–[21]. Induction of indoleamine 2,3-dioxygenase-1 (IDO) in monocytes and dendritic cells occurs during HIV infection and has been associated with impairment of the mucosal immunity and the maintenance of a chronic inflammatory state [22]. Collectively, these observations strongly suggest that an underlying mechanism not captured by CD4+ T cell count and HIV replication might be contributing to disease progression.
The importance of CD4 counts as a strong predictor of opportunistic infections and non-AIDS events has been widely investigated, but little attention has been paid to the prognostic significance of CD8 counts. During untreated HIV infection, CD8 counts increase as CD4 counts decline [23]. During ART-mediated viral suppression, some individuals achieving CD4 counts above 500 cells/mm3 experience a simultaneous decline in CD8 counts, leading to normalization of the CD4/CD8 ratio. Others, however, maintain high levels of circulating CD8+ T cells, and hence a persistently low CD4/CD8 ratio [24]. Among elderly HIV-uninfected adults, inversion of the CD4/CD8 ratio (<1.0) predicts all-cause mortality and is considered part of the immunosenescent phenotype [25]–[29]. In the setting of untreated HIV infection, the CD4/CD8 ratio predicts time to AIDS [30] and is associated with pre-ART CD4 and CD8 counts [24], [31]. Among treated adults with spectrum of CD4+ T cell counts, the ratio appears to correlate with markers of T cell activation and senescence [32], [33] and with the risk of non-AIDS morbidity and mortality [34]–[36]. Whether this is true in those with normalized CD4+ T cell counts is unknown.
We hypothesized that among ART-treated HIV-infected individuals with CD4 counts ≥500 cells/mm3, expansion of CD8+ T cells, reflected as a low CD4/CD8 ratio, may identify individuals with persistent innate and adaptive immune activation at greater risk of serious non-AIDS events. Since early ART initiation has been shown to reduce levels of T cell activation, we hypothesized that earlier ART initiation might also accelerate the rate of CD4/CD8 ratio normalization.
Study subjects were sampled from four cohorts and two clinical trials: 1) SCOPE: a clinic-based cohort of over 1500 chronically HIV-infected participants and HIV-uninfected controls in San Francisco; 2) the Study of the Ocular Complications of AIDS (SOCA): a multicenter cohort of over 2200 HIV-infected participants who initiated ART with an AIDS diagnosis; 3) OPTIONS: a clinic-based cohort of participants diagnosed during acute/early HIV infection, previously described [37]; 4) the Madrid cohort: a clinic-based cohort of 2400 ART-treated individuals, 130 of whom developed serious non-AIDS events. 5) the raltegravir (NCT00631449) [38] and maraviroc (NCT00735072) [39], [40] ART intensification randomized, placebo-controlled, clinical trials in HIV-infected individuals. Additional information on the cohorts and the clinical trials can be found in Text S1.
These studies were approved by the UCSF Committee on Human Research or by the Ethics Committee of the University Hospital Ramón y Cajal. All participants were adults and provided written informed consent in accordance with the Declaration of Helsinki.
T cell immunophenotyping was performed on cryopreserved peripheral blood mononuclear cells (PBMC), in fresh lymph node mononuclear cells (LNMC) from inguinal biopsies obtained from HIV-infected volunteers under ART-mediated viral suppression, and in mucosal mononuclear cells (MMC) obtained from rectal biopsies in the maraviroc and raltegravir studies, as previously described [38], [39], [41]. Fresh inguinal lymph nodes were biopsied were minced and strained through a 70 micron filter to created a single cell suspension of LNMC. MMC were isolated from biopsy specimens using a protocol optimized for lymphocyte viability and yield [42]. Cells were thawed, washed, stained with LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (Invitrogen) to exclude non-viable cells and stained with fluorescently-conjugated monoclonal antibodies (recognizing CD3, CD4, CD8, HLADR, CD38, CD27, CD28, CCR5, CCR7, CD45, PD1 for PBMC and CD3, CD4 and CD8 for LNMC and MMC; see Table S1). Cells were then fixed in 0.5% formaldehyde and ≥250,000 were analyzed on a BD LSR II Flow cytometer (BD Biosciences) using FlowJo (Tree Star) to determine the proportion of CD4+ and CD8+ T cells expressing each of the T cell markers. Combinations of markers were calculated in FlowJo, using the Boolean gate function (for the gating strategy, see Figure S1). We determined in PBMC the T cell maturation subsets, defined as naïve (TN, CD45RA+CCR7+CD27+CD28+), central memory (TCM, CD45RA-CCR7+CD27+CD28+), transitional memory (TTM, CD45RA−CCR7−CD27+), effector memory (TEM, CD45RA−CCR7−CD27−CD28−), and terminally differentiated (TEMRA, CD45RA+CCR7−CD27−CD28−). For CD4+ TTM we analyzed the CD28+ subset (CD45RA−CCR7−CD27+CD28+) and for CD8+ TTM cells we analyzed the CD28− (CD45RA−CCR7−CD27+CD28−) given the strong correlation between the CD4/CD8 ratio and CD8+CD28− T cells. We also determined the phenotypes of activated/senescent CD8+ T cells (HLADR+CD38+, CD28−, CD57+CD28−, and PD-1+), and the proportion of CD28−CD8+ T cells expressing CD57, which has been recently described as a unique CD8+ T cell defect in HIV that appears to be distinct from the classical immunosenescent phenotype found with aging and that predicts mortality [43]. In LNMC and MMC we determined the % of CD4+ and CD8+ T cells. Additional information is provided in the supplemental material.
Cryopreserved plasma was assessed by immunoassay for IL-6 (R&D Systems), sCD14 (R&D Systems), hs-CRP (CardioPhase hs-CRP assay, Siemens), D-dimer (DiagnosticaStago), intestinal fatty acid binding protein (I-FABP, Cell Sciences) and zonulin-1 (ALPCO) levels. Plasma tryptophan and kynurenine levels were measured by high performance liquid chromatography tandem mass spectroscopy [22], and the activity of IDO was assessed as the plasma kynurenine to tryptophan (KT) ratio. Chronic asymptomatic CMV infection was confirmed by a positive CMV IgG titer and for a subset of HIV-infected participants without available CMV serology, >0.1% pp65/IE-specific IFN-γ+ CD8+ T cell responses by cytokine flow cytometry (ten-fold increase over limit of detection) as previously described [44].
Cross-sectional pairwise comparisons between groups were performed using Wilcoxon rank sum tests. Since a “normal” CD4/CD8 ratio remains poorly defined, for the between-group comparisons of T cell subsets and percentages of activated/senescent CD8+ T cells, we classified individuals according to the lowest quartile (≤0.4) and highest quartile (≥1.0) of SCOPE participants with ≥500 CD4+ T cells/mm3. A CD4/CD8 ratio ≤0.4 has been defined previously as the best cutoff that may predict serious non-AIDS events in well-treated HIV-infected patients [36], and 1.0 has been suggested in the general population as the cutoff for the “immune risk profile” associated with immunosenescence and mortality [26], [45]. To assess the intra-individual variability of the CD4/CD8 ratio, we used data from the control arms of ART intensification trials with raltegravir [38] and maraviroc [39], [40] to calculate the coefficient of variation (standard deviation/mean) for the CD4+ and CD8+ T cell counts and for the CD4/CD8 ratio.
To analyze the association between the CD4/CD8 ratio and the KT ratio in SOCA, we fitted a linear regression model using CD4/CD8 ratio as the dependent variable, and KT ratio as the explanatory variable, adjusting the model by age, gender, time under viral suppression and CD4 nadir. To evaluate the relative contribution of the CD4+ and CD8+ T cells to this association, we also fitted another model with both CD4+ and CD8+ T cells in which the CD4/CD8 ratio was not considered because of colinearity, adjusting for the same covariates.
We analyzed the correlations between the CD4/CD8 ratio in blood, with the ratio in lymph nodes and in GALT. For the GALT CD4/CD8 ratio measured in the MVC and RAL studies, we used only baseline measurements (before ART intensification). Since a different panel of antibodies was used for each study for flow-cytometry analysis, we fitted a linear regression analysis adjusting by the source study.
We analyzed the impact of early ART initiation on the CD4/CD8 ratio in the OPTIONS cohort among recently HIV-infected participants, focusing on those who either started ART within six months of infection (early ART) or who deferred therapy for at least two years (later ART) [37]. Longitudinal changes in CD4 and CD8 counts and in the CD4/CD8 ratio were assessed using linear mixed models with random intercepts. Age, gender, and pre-ART CD4 counts were included in multivariate analyses as fixed-effects. Interaction terms were created to assess whether these changes over time differed significantly between the early and later ART initiators. Changes in slopes before and after ART time points were assessed using linear splines.
We used data from the Madrid and SOCA cohorts to evaluate whether the CD4/CD8 ratio might be a marker of non-AIDS-related morbidity and mortality, respectively. In the nested case-control analysis in the Madrid cohort, cases who developed serious non-AIDS events and had ≥500 CD4+ T cells/mm3, were each matched to one controls by age, sex, nadir CD4, and proximal CD4 counts (N = 66). In the nested case-control study of immunological predictors of mortality in SOCA, cases with non-accidental death who had PBMC and plasma samples available within 18 months of death with confirmed plasma HIV RNA levels <400 copies/ml, were each matched to two controls by age, gender, duration of viral suppression, history of CMV retinitis, and nadir CD4 (N = 183). We used conditional logistic regression to evaluate the CD4/CD8 ratio as a predictor of non-AIDS morbidity/mortality. Continuous variables in multivariate models were log-transformed when necessary to satisfy model assumptions.
We first analyzed the correlations of the CD4/CD8 ratio with naïve T cells (TN, CD45RA+CCR7+CD27+CD28+), central memory T cells (TCM, CD45RA−CCR7+CD27+CD28+), transitional memory T cells (TTM, CD45RA−CCR7−CD27+CD28+ for CD4+ cells and CD45RA−CCR7−CD27+CD28− for CD8+ cells), effector memory T cells (TEM, CD45RA−CCR7−CD27−CD28−), terminally differentiated T cells (TEMRA, CD45RA+CCR7−CD27−CD28−) and different phenotypes of activated T cells (HLADR+CD38+, CD28−, CD57+CD28−, and PD-1+). All data were obtained from those participants in SCOPE cohort who were on effective therapy and had ≥500 CD4+ T cells/mm3 (N = 67) (for a description of the general characteristics, see Table S2). The CD4/CD8 ratio was correlated positively with the frequencies of TN (Rho = 0.35, P = 0.005), TCM (Rho = 0.272, P = 0.03), and TTM CD8+ T cells (Rho = 0.25, P = 0.05), and negatively with the frequencies of TEM (Rho = −0.37, P = 0.003) and TEMRA (Rho = −0.26, P = 0.024) CD8+ T cells. Overall, the CD4/CD8 ratio was more strongly associated with the proportions of T cell maturation subsets and proportions of activated CD8+ T cell phenotypes than were the CD4 or CD8 counts (see Table 1).
To underline the association between a low CD4/CD8 ratio and persistent T cell abnormalities during effective ART, we compared ART-suppressed HIV-infected individuals with ≥500 CD4+ T cells/mm3 (N = 67) in the lowest (≤0.4) versus highest (≥1) quartiles of CD4/CD8 ratio. We also analyzed HIV-uninfected CMV-positive adults (N = 15) (see Table S2 for the clinical characteristics of each group and Figures 1, 2 and S2 for the between-group comparisons). Median CD8 counts were markedly higher among those with low versus high CD4/CD8 ratio (1964 cells/mm3 vs. 696 cells/mm3, respectively). ART-suppressed participants with a high CD4/CD8 ratio had similar proportions of CD8+ T cell maturation subsets as those in the healthy controls (Figure 1A–B). In contrast, those participants with a low ratio had higher frequencies of TTM and TEM CD8+ T cell subsets than that observed in the healthy controls. Compared to HIV-uninfected subjects, ART-suppressed individuals with low CD4/CD8 ratio had higher proportions of activated (HLADR+CD38+) and “senescent” (CD28− and CD28−CD57+) CD8+ T cells; while those with high CD4/CD8 ratio had levels of CD8+ T cell activation and senescence close to those observed in controls. However, both ART-suppressed groups (low and high CD4/CD8 ratios) had lower proportions of CD28−CD8+ T cells expressing CD57 compared to healthy controls, consistent with prior data (Figure 2A) [43].
We sought to validate these findings among effectively treated subjects (undetectable viral load, ≥500 CD4+ T cells/mm3) within the SOCA cohort (general characteristics summarized in Table S3), and found comparable correlations between the CD4/CD8 ratio and different phenotypes of activated/senescent CD8+ T cells among ART-suppressed subjects with CD4>500 T cells/mm3. The most consistent correlates of the CD4/CD8 ratio were the %HLADR+CD38+ CD8+ T cells (Rho = −0.507, P<0.001) and %CD28− CD8+ T cells (Rho = −0.400, P = 0.009) (Table 2).
To explore the potential mechanisms driving the expansion of late-memory T cells in subjects with low CD4/CD8 ratio despite effective ART we used data from the SOCA cohort. We analyzed the correlations between CD4 and CD8 counts and the CD4/CD8 ratio and different markers of innate immune activation and epithelial integrity (Table 3). We observed across all effectively treated subjects (as defined by having an undetectable viral load) significant inverse correlations between the CD4/CD8 ratio and hs-CRP, IL-6, sCD14 and the KT ratio. However, in the subgroup of subjects with ≥500 CD4+ T cells/mm3, only the KT ratio remained significantly correlated with the CD4/CD8 ratio (Rho = −0.30, P = 0.041) (Figure 3A). This association was confirmed in a linear regression analysis adjusted for age, gender and cumulative ART exposure (Beta = −0.72, P = 0.009), where for each 10% increase in the CD4/CD8 ratio there was a 7% decrease in the KT ratio. The CD4/CD8 ratio performed better as a predictor of the KT ratio than the CD4+ or CD8+ T cell counts in a similar model (see Table 4). Since subjects with low ratio and ≥500 CD4+ T cells/mm3 were enriched for CD28−CD8+ T cells and also showed increased IDO induction, we hypothesized that a potential underlying mechanism driving expansion of CD28−CD8+ T cells might be IDO induction, and we found a positive correlation between these two variables (Rho = 0.50, P<0.001) (Figure 3B). These results indicate that the CD4/CD8 ratio predicts better the degree of IDO induction than the CD4 or CD8 counts individually, which becomes especially evident above the threshold of 500 CD4+ T cells/mm3, and that in subjects with low CD4/CD8 ratio increased IDO induction is associated with an immunosenescent phenotype (expansion of CD28−CD8+ T cells).
A persistently low CD4/CD8 count ratio in peripheral blood might conceivably be the result of differential redistribution of CD4+ and CD8+ T cells out of lymphoid tissues, but to our knowledge, no study has assessed whether the CD4/CD8 ratio in peripheral blood is reflective of the CD4/CD8 ratio in tissues. To address this question, we analyzed the correlations of the CD4/CD8 ratio in blood with the CD4/CD8 ratio in lymph node and in GALT (see Table S4). Using data from 10 individuals on ART, in whom the CD4/CD8 ratio in blood was 0.6 (0.4–1.1) and in lymph nodes 2.5 (1.5–4.1), no significant correlation between the CD4/CD8 ratio in blood and in lymph nodes was detected (Rho = −0.07, P = 0.855) (Figure 4A). For the correlation between the CD4/CD8 ratio in GALT and in blood (Figure 4B), we used data from 32 individuals on ART, in whom the CD4/CD8 ratio in blood was 0.4 (0.2–0.6) and in GALT 0.6 (0.4–0.9). We found that the CD4/CD8 ratio in blood was strongly correlated with that in rectal mucosa (Rho = 0.68, P<0.001 and Beta = 0.69, P<0.001).
To characterize the variability on CD4/CD8 ratios over time, we analyzed data from 38 HIV-infected adults who maintained undetectable viral loads on ART. Subjects had a median CD4+ count of 320 cells/mm3 (range 88 to 884) at baseline. A median of 11 determinations of CD4+ and CD8+ T cells measurements were performed during a median of 81 weeks (Figure S3). The mean coefficient of variation was significantly lower for the CD4/CD8 ratio (12%) compared to CD4+ T cell counts (16%, P = 0.017) and for CD8+ T cell counts (18%, P = 0.001), indicating that the CD4/CD8 ratio shows lower intra-individual than the CD4+ or CD8+ T cell counts over time.
We next examined the extent to which persistent abnormalities in the CD4/CD8 ratio were associated with later vs. earlier initiation of ART using the described OPTIONS cohort of recently infected adults who started therapy during the first six months of their infection (early ART) or after two years of untreated infection (later ART) (see Table S5) [37]. At the time of their diagnosis, median CD4/CD8 ratio was significantly lower in recently HIV-infected individuals compared to the HIV-uninfected group (Figure 5A, all baseline comparisons, P<0.05). The later ART group remained untreated for a median of 3 years. The CD4 count declined by 274 cells/mm3, the median CD8 count increased by 125 cells/mm3, and the CD4/CD8 ratio decreased from 0.76 to 0.38 during this untreated period.
After one year of ART, both early and late ART initiators experienced a substantial increase in CD4+ T cells (Figures 5B–C). However, while early ART subjects also showed a substantial decline in CD8+ T cells after one year of ART, the later ART group required a median follow-up of three years (Figures 5D–E). After one year of ART, early treated patients showed significantly higher median CD4/CD8 ratio (1.0 vs. 0.57, P<0.001) and had fourfold-increased odds of CD4/CD8 ratio normalization during follow-up (OR, 3.6; 95% CI, 1.2, 10.8; P = 0.022). The greater effect of early ART compared to later ART on the CD4/CD8 ratio remained statistically significant after adjustment by age, gender, and baseline CD4+ T cell counts in the mixed-effects linear model (Figures 5F–G). The mean CD4/CD8 ratio change predicted by the model was significantly higher among early ART initiators compared to later initiators after one year of ART (+0.44 vs. +0.25, respectively, P<0.001), and after a median of 3 years of ART (+0.61 vs. +0.49, respectively, P<0.001). In summary, early ART initiators experienced a faster CD4/CD8 ratio increase and reached higher CD4/CD8 ratios after a median of 3 years of ART, which was primarily driven by changes in the CD4 counts and, to a lesser degree, by changes in the CD8 counts.
Lastly, we hypothesized that the prognostic importance of the CD4/CD8 ratio might depend upon the relative predictive contribution of both CD4 and CD8 counts. In a previously reported analysis, we found that among a large cohort of treated adults in Madrid (N = 420) that the ratio was predictive disease progression [36]. Here, we performed a case-control study among the subset with ≥500 cells/mm3. A sample of 33 cases with CD4 counts ≥500 cells/mm3 was matched to 33 controls by age, gender, nadir CD4 and proximal CD4+ T cell counts (see Table S6 for the general characteristics of the study population and Table S7 for the description of non-AIDS events). We observed that both the CD4/CD8 ratio and CD8 count independently predicted the risk of non-AIDS events, with the coefficient of the CD4/CD8 ratio significantly higher (see Table 5). After controlling for age, gender, ART duration, nadir and proximal CD4 count, each 10% decrease in the CD4/CD8 ratio and each 10% increase in the CD8+ T cell counts were associated with 48% and 22% higher odds of serious non-AIDS events, respectively.
To assess the relationships with mortality, we also examined the entire SOCA cohort (median CD4+ T cell count at baseline of 340 cells/mm3, range 1 to 1498). We analyzed 62 individuals who died (cases) and 121 who did not die (controls) matched by age, gender, nadir CD4+ T cell count, and duration of viral suppression (see Tables S3 and S7). We observed that both the CD4/CD8 ratio and CD4+ T cells, but not CD8+ T cells, were independent predictors of mortality –for each 10% increase in the CD4/CD8 ratio or in CD4+ T cells and there was a 15% and 13% decrease in the risk of death, respectively (see Table 5). As this cohort enrolled individuals with advanced HIV disease, there were insufficient numbers of events in the ≥500 cells/mm3 subset to analyze.
Combining the data from four clinical cohorts and two clinical trials, we demonstrate here that a substantial subset of ART-suppressed HIV-infected adults who have achieved virologic suppression and a normalized peripheral CD4 count (≥500 cells/mm3) have persistently elevated CD8 counts and a low CD4/CD8 ratio. This ratio is correlated with markers of T cell activation and innate immune activation (IDO induction) and with the presence of a previously described immunosenescent phenotype (i.e., low naïve T cell frequencies and increased frequency of terminally differentiated). This imbalance in T cell homeostasis measured in blood is also present in GALT, and the CD4/CD8 ratio shows lower intra-individual variability than the CD4+ or CD8+ T cell counts. Although early ART (<6 month after HIV infection) is associated with more rapid normalization of the CD4/CD8 ratio, an abnormal ratio persists even in these aggressively treated individuals. Among well-treated individuals with high CD4 count, a low ratio was an independent predictor of serious non-AIDS events and mortality. Collectively, these results suggest that a persistently low CD4/CD8 ratio during ART may be a marker of persistent immune dysfunction and inflammation, and that monitoring of this ratio—which can be readily done in most clinics with current assays—may be clinically useful. A truly successful response to ART may require both normalization of the peripheral CD4 count and the CD4/CD8 ratio.
The immunologic profile of the individuals in our cohorts with a persistently low CD4/CD8 ratio despite high CD4 counts is similar to that observed in the very old. T cell “immunosenescence” is generally defined as a low naïve/memory T cell ratio, expansion of CMV-specific CD8+ T cells, enrichment for CD28− and PD-1+ T cells, increased CRP and IL-6 levels, reduced T cell telomere lengths and a low CD4/CD8 ratio [46]. Since untreated HIV infection is associated with each of these immunologic characteristics, it has been proposed that HIV might accelerate the aging of human immune system [13], [47]–[49]. The extent to which successful ART reverses these HIV-induced immune changes is currently the subject of intense investigation [31]. We found that CD8+ T cells counts often remain high even as CD4+ T cell count levels normalize, arguing against the existence of a “blinded” T cell homeostasis during long-term ART. We also found that among apparently well-treated adults (undetectable viral load, high CD4+ T cell counts) that a persistently low CD4/CD8 ratio was consistently associated with markers of inflammation (particularly CD8+ T cell activation).
We also studied a number of related markers of immunosenescence, using as a comparator group HIV-uninfected adults who were infected with CMV (as nearly all HIV-infected subjects are co-infected with this virus). We found that HIV-infected subjects who achieved CD4/CD8 ratio normalization during ART demonstrated traits of a nearly healthy immune system, with T cell maturation subsets and levels of CD8+ T cell activation/senescence comparable to those observed in healthy subjects. In contrast, a CD4/CD8 below ≤0.4 identified individuals with prominent features of immunosenescence despite CD4+ T cell recovery, including reduction of the CD8+ naïve T cell compartment, enrichment for TEM and TEMRA CD8+ cells, and increased levels of CD8+ T cell activation (HLADR+CD38+ T cells) and senescence (CD28− and CD57+CD28− T cells). Expansion of CD28−CD8+ T cells is a hallmark of replicative senescence, a term describing the phenomenon in which long-lived cells that have undergone multiple rounds of proliferation, show telomere shortening and hence, limited proliferative potential [50],[51].
When studied in all treated subjects, the CD4/CD8 ratio inversely correlated with several markers of innate immune activation (sCD14, hs-CRP, and IL-6) and with a biomarker of IDO induction (KT ratio). While the association with the KT ratio remained significant in the subgroup of individuals with ≥500 CD4+ T cells/mm3, the correlations for sCD14, hs-CRP, and IL-6 did not, perhaps because of the very high biologic variability in these assays and the loss of statistical power in the subgroup analysis. IDO is induced during HIV infection in activated dendritic cells and monocytes by interferons and toll-like receptors ligands such as LPS and sCD14, catabolizing tryptophan into kynurenine and other immunologically active catabolites; these catabolites suppress T cell proliferation and/or differentiation, resulting in impairment of the mucosal immunity. Such induction may occur even after the initiation of effective ART, as shown by increased K/T ratios in some if not all [52]. It has been argued that induction of IDO may represent a critical initiating event that results in inversion of the Treg/TH17 regulatory balance, loss of epithelial barrier integrity and thereby maintenance of a chronic inflammatory state during chronic HIV infection [22], [53]. Since the KT ratio correlated well with CD28−CD8+ T cells, our data suggest that increased IDO activity may be contributing to the replicative CD8+ T cell senescence observed in ART-treated subjects with low CD4/CD8 ratio. Alternatively, the accumulation of CD8+ T cells might be the cause of increased IDO activity, as a consequence of greater IFN-γ production.
We observed that a low CD4/CD8 ratio is a predictor of serious non-AIDS events among treated individuals with ≥500 CD4+ T cells/mm3 in the Madrid-based cohort, with much of the association driven by the CD8 counts. This observation expands upon our previous findings in this cohort, where we showed in a population of HIV-infected individuals non-restricted by a CD4 count that a low ratio is associated with increased risk of non-AIDS events and associated mortality [36]. We observed an association between the ratio and mortality in the entire SOCA cohort, but in this cohort selected based on low CD4 nadir and which had a lower range of CD4 counts, the association depended upon the CD4 counts. These two analyses suggest that while the CD4/CD8 ratio has prognostic significance in all treated adults, much of the harm associated with a low ratio is driven by the CD4 count (and presumably immunodeficiency) in those with low CD4+ T cell counts, while the harm associated with low ratio in those with higher CD4+ T cell counts is driven by CD8+ T cell count (and presumably inflammation).
As suggested by the present study, the group of ART-treated HIV-infected individuals with CD4+ T cell counts above 500 cells/mm3 represents a clinical spectrum of individuals, ranging from those with an immune system that is abnormal (e.g., with a CD4/CD8 ratio ≤0.4) to those with an apparently normal immune system (e.g., with a CD4/CD8 ratio ≥1), a finding that might serve to explain the discrepancies in cohort studies addressing morbidity and mortality among successfully treated HIV-infected individuals [4], [6]–[8], [54]–[57]. We suggest that the CD4/CD8 ratio might help to further discriminate the risk of disease progression of successfully treated HIV-infected individuals.
Mechanistically, we imagine that the maintenance of an abnormally high level of circulating CD8+ T cells could be the result of increased proliferation, decreased death, and/or changes in the rate at which these cells move between organized lymphoid structures and the peripheral blood. Although we have no evidence for increased levels of proliferation (e.g., as measured by incorporation of stable isotopes or expression of Ki67), previous studies have shown that the TEMRA subset has a longer lifespan in the setting of untreated HIV disease [58], a property that may well be found in treated individuals with a low CD4/CD8 ratio. There are also data showing that progressive as well as treated HIV disease is associated with collagen deposition and loss of the fibroblastic reticular cell network within lymphoid tissue, particularly in the context of inflammation [59], [60], and such structural changes might result in the presence of an unusually large proportion of circulating CD8+ T cells. Since each of these pathologic changes is associated with a high level of inflammation, it follows that resolution of inflammation in the effectively treated individual with a low CD4/CD8 ratio might result in normalization of the ratio over time.
There are limitations to the current study that deserve mention. First, for the analysis of the correlation between the CD4/CD8 ratio in blood and GALT we used data from two clinical trials involving individuals with suboptimal CD4+ T cell recovery; hence, further studies in individuals with CD4+ T cell recovery above 500 cells/mm3 are needed to assess whether a low CD4/CD8 ratio reflects poor GALT immune reconstitution in these subjects. Second, although we were able to find consistent associations between CD4/CD8 ratio and T cell activation/maturation in those with high CD4+ T cell counts, we did not find such associations with plasma biomarkers (e.g., IL-6) possibly because of high assay variability; larger more definitive studies assessing this question are needed. Third, in the subset of individuals above 500 cells/mm3, there were only eight instances of non-AIDS related death in the Madrid cohort, and 16 in SOCA cohort, which prevented us from performing mortality analysis in this subgroup. The prognostic significance of the ratio among well-treated adults (i.e., those with undetectable viral loads and high CD4+ T cells) will need to be confirmed in larger cohorts. It will be of interest to determine in this population if the ratio has unique prognostic capacity as compared to that observed in the general population [25], [45].
Our results have potential clinical implications for novel therapeutic strategies targeting immune dysfunction in chronically treated HIV-infected individuals, in particular those with persistent expansion of CD8+ T cells despite adequate CD4+ T cell recovery. This CD4/CD8 ratio may be useful in monitoring response to therapies aimed at reducing residual immune activation, and given that prior studies have also reported that a low CD4/CD8 ratio is associated with increased markers of HIV persistence [61], [62], subjects with a high CD4/CD8 ratio may be useful targeted candidates for HIV eradication trials. Finally, ART-suppressed HIV-infected individuals who do not have an increase in the CD4/CD8 ratio might benefit from screening programs or aggressive management of concomitant risk factors for aging-associated disease.
In summary, a low CD4/CD8 ratio among ART-treated HIV-infected individuals achieving CD4+ T cell counts above 500 cells/mm3 may define a new clinical phenotype of individuals with higher levels of CD8+ T cell activation and senescence, depletion of naïve and enrichment for TEMRA cells, increased IDO activity and higher risk morbidity and mortality. Early ART initiation may contribute to more rapid and robust CD4/CD8 ratio normalization, and the CD4/CD8 ratio may be a useful clinical endpoint to be used in evaluating novel therapies for ongoing immune dysfunction during treated infection and for HIV eradication.
|
10.1371/journal.pcbi.1003802 | The Protective Role of Symmetric Stem Cell Division on the Accumulation of Heritable Damage | Stem cell divisions are either asymmetric—in which one daughter cell remains a stem cell and one does not—or symmetric, in which both daughter cells adopt the same fate, either stem or non-stem. Recent studies show that in many tissues operating under homeostatic conditions stem cell division patterns are strongly biased toward the symmetric outcome, raising the question of whether symmetry confers some benefit. Here, we show that symmetry, via extinction of damaged stem-cell clones, reduces the lifetime risk of accumulating phenotypically silent heritable damage (mutations or aberrant epigenetic changes) in individual stem cells. This effect is greatest in rapidly cycling tissues subject to accelerating rates of damage accumulation over time, a scenario that describes the progression of many cancers. A decrease in the rate of cellular damage accumulation may be an important factor favoring symmetric patterns of stem cell division.
| Recently, highly symmetric patterns of stem cell division have been observed in a variety of adult mammalian somatic tissues. Here we identify conditions under which this behavior serves as a strategy to protect the organism against mutation accumulation. First, we find that a sufficient number of lifetime stem cell divisions must occur, potentially explaining why stem cell pools with the most symmetric divisions are rapidly cycling. Second, we find that late-occurring mutations must occur rapidly, a scenario known in cancer biology as genetic instability. These findings provide a potential explanation for the observation that cancer risks among large, long-lived organisms fail to rise as expected with lifespan and body size.
| The accumulation of heritable damage—both mutation and epigenetic change—within individual cells is thought to be a major driver of cancer [1], [2] and aging [3], [4]. Cells employ various strategies for preventing or delaying damage accumulation including DNA repair, apoptosis and senescence [5]. Unfortunately, these strategies fail when damage lacks an immediate phenotypic consequence.
One way to delay damage accumulation in the absence of phenotypic consequences is to employ a lineage hierarchy in which self-renewing stem cells produce “transit-amplifying” cells that proliferate before differentiating into cells that eventually leave the tissue [6]. The success of this strategy relies upon transit amplifying cells being short-lived (so that damage that occurs at this stage is flushed away before more damage occurs) and stem cells dividing infrequently. However recent studies of several major vertebrate tissues (e.g. epidermis, intestinal epithelium, testis) challenge both the existence of obligatorily short-lived transit amplifying cells, and the view that stem cells usually cycle slowly [7]–[11].
The “immortal strand” mechanism [6], another proposed strategy for limiting damage accumulation, is predicated on the hypothesis that stem cells divide asymmetrically (Fig. 1A), segregating parental (less damaged) DNA strands to the daughter that remains a stem cell. Not only do recent observations question whether such DNA sorting occurs, e.g. [12], but, in vertebrates at least, most adult stem cell pools—including those of the hematopoietic system, intestinal epithelium, interfollicular epidermis, testis and hippocampus—exhibit a substantial proportion of symmetric divisions [7]–[11], [13]–[18]; (Fig. 1B).
The observation that many stem cell pools undergo symmetric divisions is interesting, given that tissue homeostasis (constancy of stem and differentiated cell numbers) demands that symmetric renewal events (where one stem cell generates two stem cells) be balanced, on average, by an equal number of symmetric differentiation events (where one stem cell generates two differentiated cells; also referred to as symmetric extinction events, since such events extinguish a stem lineage). Feedback signals from differentiated cells most likely provide such a matching mechanism [14], [19]–[21].
We were intrigued by the fact that the somatic tissue with the highest degree of symmetric stem cell division observed to date (close to 100%) is the vertebrate intestinal epithelium [10], [11], [16], because its large, rapidly dividing stem cell pool [12] ought to be particularly susceptible to mutation accumulation. Indeed, measurements of microsatellite alterations in mismatch-repair deficient mice [22], and genome-wide sequencing studies of cancer genomes (summarized in Fig. 3 of [23]), both show that the mutation burden in the vertebrate intestine is significantly higher than in other tissues. This made us wonder whether a highly symmetric pattern of stem cell division might play a role in slowing the accumulation of heritable damage. Below, we show mathematically that this is indeed the case and that, in certain biologically relevant scenarios, the protection achieved can be surprisingly large.
When a stem cell undergoes an extinguishing division, it and all of its mutations (here we will use “mutation” to stand for all forms of heritable damage, genetic or otherwise) become fated to leave the body, suggesting that some of the mutation “flushing” enjoyed by short-lived transit-amplifying cells also accrues to symmetrically-dividing stem cells (Fig. 1C). Symmetric renewal divisions oppose this effect, increasing the proportion of stem cells with any given set of mutations, and elevating the risk of mutation accumulation (Fig. 1D). Since symmetric renewal and extinction must balance in homeostatic tissues (Fig. 1B), one might expect these two effects to cancel.
To test this prediction, we performed stochastic simulations of homeostatic stem cell populations of various sizes, engaging in either purely asymmetric or purely symmetric division. We allowed mutations at different loci to occur at different rates, and measured the number of stem cell divisions required for at least one stem cell to accumulate a particular number of mutations. In each simulation, the order in which specific loci mutated was fixed, allowing us to model the accumulation of mutations at K loci as a stepwise transition of cells through K stages (Fig. 1E). Later, we calculate the behavior when mutation order is not fixed (i.e. where any locus can mutate at any time).
For any cell population that chooses division outcomes stochastically, even if probabilities of renewal and extinction exactly balance, cell numbers will fluctuate around a mean value [9]–[11]; The more symmetric the division pattern, the greater the fluctuations. Such fluctuations are negligible (in relative terms) in large stem cell pools but physiologically significant in smaller ones, potentially extinguishing the entire pool. Therefore moderately sized stem cell pools that exhibit a high degree of division symmetry in vivo, yet don't display large size variation—for example intestinal crypts [10], [11]—must employ a variance-reducing process (e.g. size-dependent feedback control). Since the details of such processes are generally unknown, we cannot model them explicitly. Rather, we model the behavior of stem cell pools of <104 cells as a Moran process (a stochastic process in which constant population size is enforced at every time point), whereas for larger pools we simulated a simple branching process without imposed size constraints (Materials and Methods). Importantly, results obtained with both approaches agreed when assessed at large population sizes (Fig. S1A).
Simulation parameters consisted of N, the population size (number of stem cells); K, the number of loci in which mutations of interest may accumulate; u0,…,uK-1, the mutation rates for acquiring each of the K mutations; and L, the organism lifetime measured in stem cell cycles. For each set of parameters we determined the fraction of simulations of purely asymmetric division in which at least one stem cell had K mutations—the “asymmetric risk”—as a function of time (see Materials and Methods). The “symmetric risk” was calculated similarly, but employing a purely symmetric division pattern. One possible way to quantify the difference between the two risks (at otherwise identical parameter values) is to measure displacement, along the time axis, from one risk curve to the other, i.e. the amount of extra time a particular division strategy confers on a stem cell pool before it acquires a cell with K mutations. Though such a “mean first passage time” approach is mathematically sound, the answer one obtains is biologically irrelevant whenever the mean-first passage time is much shorter or much longer than the reproductive lifespan of the organism. We therefore measured the ratio of risks at a single time point, which we term the “Protection Factor” (PF), because a change in the probability of having a deleterious phenotype (K mutations in at least one stem cell) at a fixed time point (e.g. the end of an organism's reproductive period) is directly connected to the pressures of natural selection at the organism level.
Care must be exercised in choosing the time at which PF is evaluated since, with enough time, all risks plateau at 100%. Accordingly, PFs were typically ascertained when the asymmetric risk (always greater than or equal to the symmetric risk; see below) lay in the vicinity of 50% (Materials and Methods), i.e. at a time when a stem cell pool executing only asymmetric divisions would have a 50% chance of possessing at least one K-fold mutant stem cell. Later, we also consider cases of many small stem cell pools functioning in parallel (i.e., a compartmentalized population), for which much smaller asymmetric risks must be used.
Fig. 1F presents the results for a particular case in which the size of the stem cell pool was N∼60,000 stem cells and the number of accumulated mutations was K = 3. The symmetrically dividing stem cell population clearly displays a significantly reduced risk of arriving at the fully mutant state. Fig. 1G summarizes the results for 1000 parameter sets in which population sizes, mutation rates, and number of mutable loci (K) were chosen at random, subject to the condition that the time (in cell cycles) when the asymmetric risk was close to 50% should fall within a reasonable range of organism lifetimes (Table S1). Interestingly, 19% of simulations exhibited PF>2, i.e. symmetric division cut the risk of mutation accumulation by at least half. This suggests that significant protection against mutation accumulation occurs over a substantial fraction of parameter space.
To understand the origin of protection, we examined the dynamics of mutant stages (i.e. sub-populations with a given number of mutations per stem cell) in individual simulations. Stage size fluctuated more when divisions were symmetric versus asymmetric (Fig. 2A, B); this is expected because, in symmetrically-dividing populations, fluctuations come not only from the random entry and exit of cells as they acquire mutations, but also from transient imbalances in renewal versus extinction.
Fluctuations in one stage are expected to alter the rate of entry of cells into the next stage, depending upon the size and direction of the fluctuation. As long as upward and downward fluctuations balance—which they usually do—they should have no long-term effect on rates of mutation accumulation, and therefore offer no protection. There is one circumstance under which they will not balance, however, which is when fluctuations extinguish all the mutants in a given stage. In that case that stage must await the entry of a new mutant cell from the previous stage before upward fluctuations can resume; in principle, this effect could slow the rate of mutation accumulation. Consistent with this idea, when we visually examined simulations of symmetrically dividing populations that exhibited high PF, we always observed frequent extinctions of one or more mutant stages (Fig. 2D).
If stage extinctions are the basis for protection, then the magnitude of protection might reflect the propensity of stages to extinguish before they progress (i.e. acquire a subsequent mutation). In other words, for protected cases we expect the average time for a stage to extinguish to be much smaller than the average time to progress. For stages 1, 2 and 3 in Fig. 2B, we see that progression occurs only once a rare clone expands to a large size. The largest stage-i clone that arises during an organism lifetime (e.g. the clone indicated by an asterisk in Fig. 2D) extinguishes in a time of order , where is the (random) number of stage-i stem cells at the end of life, L, and represents an average over many realizations of the stochastic process (derived in Section 1.5 of Text S1). On the other hand, were that clone dividing asymmetrically, the mean time to acquire a new mutation would have been . Thus, clonal extinctions outcompete progression when , which can be conveniently formulated as , where , the “scaled stage size,” is defined by(1)Stages with exhibit time-dependent trajectories that are well described by the “deterministic” equations, Eq. (S5), whereas stages with are “stochastic” (Fig. 2E). Parameter sets with multiple stochastic stages are, on average, the most highly protected (Fig. 2F). In Fig. 2G we correlate the observed scaled stage size for the penultimate and antepenultimate stages (i.e. stages K-1 and K-2), color-coding the data by the observed PF; significant protection only occurs when for the penultimate stage, and is largest when for both stages. In general, protection increases with the number of consecutive stages satisfying (Fig. S1B).
What sorts of parameter values give rise to such behavior? We established a simple scaling relation (Section 1.5.2 of Text S1 and illustrated in Fig. 2H, I)(2)for the “scaled mutation rate” defined by(3)Eq. (2), implies that when a stage is protected (), the scaled mutation rate is necessarily large, . The latter condition is less stringent at later stages, where the threshold mutation rate from Eq. (3) is typically smaller, suggesting that protection is favored by high mutation rates at late stages. Indeed, numerical screens, conducted subject to the constraint that mutation rates are strictly accelerated (versus decelerated) after an arbitrarily chosen stage, show dramatic enrichment of protected parameter sets (Fig. 2 J, K).
To gain further insight, we focused on cases of mutation accumulation with just two mutant stages (Fig. 3A). With large stem-cell populations, mutation rates must be unrealistically slow to meet the requirement that asymmetric risk at biologically realistic lifetimes should be ∼50%; we therefore simulated pools of 1 to 106 stem cells (using a Moran model). Fig. 3B and C show the asymmetric and symmetric risks as a function of the secondary mutation rate and population size, with the lifetime and primary mutation rates held fixed (at 103 cell cycles, and 10−6 per cell cycle, respectively). The grey contour marks parameter combinations for which the asymmetric risk is 50% at the organism lifetime (i.e. the conditions under which risks were compared in Fig. 2). Panel D, which plots the ratio of panels B and C, i.e. PF, shows that significant protection can occur in a large region of parameter space (circumscribed by a white contour line), with protection as great as 16-fold possible. As concluded in the previous section, a high final mutation rate (i.e. u1≫u0) is necessary to achieve significant protection.
To uncover the role of the tissue renewal rate, we varied the number of stem cell cycles that occur during the organism's lifetime. Fig. 3E indicates that, though protection is impressive (PF>2) in rapidly cycling tissues (i.e. those tissues whose stem cells cycle at least 100 times during the organism's lifetime), it vanishes (PF = 1) in slowly cycling tissues (small number of cell cycles). This dependence of protection on lifetime stem cell output, which we would have missed had we simply gauged protection from mean first passage times (e.g. the times in Fig. 3E at which each risk reaches 50%), is also seen for other population sizes and secondary mutation rates (Fig. 3F–H). We derived a piecewise analytical formula for PF (Section 2.3 in Text S1; Fig. 3I) that is an excellent approximation of the simulated results, as seen by the excellent fitting of symbols in Fig. 3E, and by the resemblance of the simulation heat maps in Fig. 3F–H with their analytical approximations in Fig. 3J–L. This analysis shows that significant protection is expected when the organism lifetime is appreciably larger than the mean time it takes a symmetrically dividing, single-mutant clone to progress to the next stage. In larger populations, where the clone is unlikely to fix (Fig. S2D; “Stochastic Tunneling” regime), this progression time is(4)(Fig. 3E, I; Section 1.7 of Text S1) whereas it is in smaller populations (Fig. 3I), in which the clone first fixes before progressing (Fig. S2L; “Sequential Fixation” regime). In both regimes, protection is favored by minimizing the clonal progression time, which, these formulae tell us, occurs when the secondary mutation rate is fast, as concluded in the previous section (Fig. 2). In short, the observations that rapid stem cell divisions or fast terminal mutation rates favor protection are in fact just two sides of the same coin (Section 2.3.1 of Text S1).
So far, we have assessed the protection offered by a purely symmetric pattern of stem cell renewal, yet many tissues (e.g. the mammalian epidermis [7], [8]) employ a mixture of asymmetric and symmetric stem cell divisions (Fig. 4A). How much should “contaminating” asymmetric divisions reduce protection? Surprisingly, we find that a tissue employing symmetric divisions just 10% of the time still robustly delays mutation accumulation (Fig. 4B) via extinctions of intermediate-stage clones (Fig. 4C), just as we found in purely symmetric cases (e.g. Fig. 2B). A formula for the probability that a single-mutant clone mutates (derived in Section 1.6 of Text S1 and used to plot the theoretical curves shown in Fig. 4D)(5)(s is the fraction of divisions that are symmetric) explains why: new clones likely extinguish so long as u1≪s, i.e. provided the mutation rate is slow compared to the symmetric-division rate—not a particularly stringent condition, even with an elevated mutation rate.
We saw that symmetric stem cell divisions are protective when fast mutations occur late (Fig. 5A; Fig. 2J) but not when they occur early (Fig. 2K). If mutations are independent, however, fast mutations sometimes occur early and sometimes late (Sections 3.1 and 3.2 of Text S1). Under such circumstances, little if any protection is observed (Fig. 5B; Fig. S3D–H) because most double mutants arise via the fast-slow route (which is not protected), rather than the slow-fast one (which is protected).
One important scenario where mutation rates are not independent is the development of cancer, where alterations at “genetic stability” loci—whether by DNA sequence changes or aberrant epigenetic alterations—elevate mutation rates throughout the genome (see Discussion). Fig. 5C presents the case where A represents a genetic stability gene, so that B mutates rapidly only if A is already inactivated. In this scenario, the unprotected path (genetic stability gene mutated last) can no longer compete with the protected path (genetic stability gene mutated first). Thus protection conferred by division symmetry persists. It should be noted that, for these calculations, mutations in A were treated as neutral (i.e. not by themselves affecting fitness), which is valid as long as the increase in mutation rate is small enough that the added lifetime burden of subsequent deleterious mutations is inconsequential.
Mutation accumulation is expected to be particularly acute in the mouse small intestine because its calculated lifetime proliferative output—some 1010 stem cell divisions [24], [25]—is extraordinarily high. Under a plausible mutation progression scenario (neutral mutations in a genetic stability gene followed by rapid mutation of APC [26]–[28]; Fig. 6A) simulations show that as little as 10% symmetric divisions are significantly protective (circles in Fig. 6B). This calculation assumes, however, that mutant clones expand unimpeded whereas, in reality, stem cells are segregated into crypts, with clonal expansion beyond crypt boundaries occurring only infrequently (1–10 times per crypt per lifetime [29], [30]). We therefore accounted for this fact by computing lifetime risk in an individual crypt (Fig. 6C) and then integrating that risk over all crypts in the intestine. Fig. 6B (triangles) shows that compartmentalizing in this way does not affect the intestine-wide asymmetric risk (as expected since, in that scenario, each stem cell behaves independently), but does increase the symmetric risk, although it is still substantially lower than the asymmetric risk.
We show here that observed levels of division symmetry in vertebrate tissues [7]–[11] can lower the risk of heritable damage accumulation even if the amount of symmetry is modest, the tissue is compartmentalized, or damage is initially phenotypically silent. For this effect to be physiologically significant—substantially reducing a high cumulative incidence of mutation—conditions must favor the frequent stochastic extinction of multi-hit stages (e.g. ; see Eq. (1)), which places significant constraints on mutation rates and the order in which mutations occur.
Even when we relax the assumption that damage is phenotypically silent (e.g. by allowing mutations to be selectively advantageous), we find that protection can persist, provided that selection coefficients are (Fig. S5; see also Section 4 of Text S1). Recent analyses of cancer genome data suggest that cancers commonly evolve through multiple driver mutations [31] with selection coefficients on the order of 1% (assuming normal mutation rates) or lower (assuming genetic instability) [32]. Thus our results should be relevant not only to the accumulation of neutral mutations but also to mutations that drive cancer evolution.
Others have noted that patterns of division symmetry should have an impact on the stochastic dynamics of cancer stem cells (e.g. how division pattern influences the probability of clonal fixation [33] or the rate at which drug resistance should develop in a growing tumor [34]) whereas here we focused on the accumulation of neutral mutations that precede the development of cancer. Although some authors have dismissed the idea that division symmetry can have any effect on neutral mutation accumulation [35], a few studies have identified specific scenarios in which symmetry can be protective [36]–[38]. None of these studies systematically identified the extent of protection as a function of mutation rate, population size, and organism lifetime. By doing so here, we discovered that necessary and sufficient conditions for protection are: (1) a large enough number of stem cell divisions must occur over the organism's lifetime (Fig. 3E–H); (2) late mutations must occur rapidly (e.g. ; see Eq. (3); Fig. 2H, I).
The first condition arises because there need to be enough stem cell divisions for a substantial number of lineage extinctions to occur. This is likely the case in many mammalian tissues (e.g. epidermis, intestine, testis [7], [9], [11]), but probably not in the somatic tissues of small, short-lived animals. This difference may help explain why most observations of purely (or predominantly) asymmetric stem cell division have come from studies of invertebrates such as Drosophila [39]–[42]—in which the protective effect of symmetry is likely to be virtually nil—whereas most observations of substantial symmetry come from studies of mouse, cat, monkey and man [7], [9]–[11], [13], [15], [17], [43].
The second condition for significant protection—that late mutational steps be fast (Fig. 2J)—broadens the distribution of latencies to arrive at the final mutant state (Fig. S6). This is an example of a general principle: distributions formed by sequential stochastic processes become over-dispersed when late steps occur on a faster time scale than earlier ones. Another example is the supra-Poissonian variance in protein levels seen among cells when mRNAs are translated more rapidly than they are transcribed [44].
That protection requires late mutations to be rapid suggests that the primary physiological value of symmetric stem cell division is cancer-risk reduction. This is because genetic instability is thought to play a key role in the development and progression of many cancers, especially at late stages [45]. Mutations in genes associated with DNA polymerase proofreading [46]; DNA damage repair [47], [48]; chromosome segregation [49], [50]; chromatid cohesion [51]; DNA damage checkpoints [52]; as well as aberrant epigenetic alterations [53]; appear commonly in human cancers and/or the germline of individuals pre-disposed to cancer. The effect of these genetic or epigenetic alterations is to make other loci mutate more rapidly, creating just the scenario in which division symmetry will be protective (e.g. Fig. 5C).
Levels of genetic instability in cancer can be large. Studies of human tumors and their adjacent normal tissue find a mutational load of 0.1–100 somatic variations per megabase, a significant fraction of which exhibit mutagen signatures, e.g. [54], [55]. Colorectal tumor samples harboring alterations in more than one genetic stability gene contain 10–100 variants per megabase [47], which is 10- to 100-fold more than expected given a normal mutation rate [56]. Not only is mutation prevalence increased in cancer, so are mutation rates. When compared with normal cells, human colorectal carcinoma cell lines with DNA mismatch repair loss exhibit a 100- to 1000-fold increase in mutation rate [57], [58], and even those without this deficiency display a 10- to 100-fold increase in loss or gain of entire chromosomes [59].
Is this amount of genetic instability high enough to make protection physiologically significant? With a 1000-fold acceleration in mutation rate, fully symmetric division could lower mutation accumulation risk in the mouse intestine by 2.2-fold (Fig. 6B). Extrapolating to stem cell numbers and lifetimes representative of the human intestine, we expect protection to increase a further 2.5- to 3.6-fold at ages 30 and 60, respectively (Fig. S7A, B; see also Section 5 of Text S1); indeed even with only a 100-fold acceleration in mutation rate, calculations indicate that symmetric stem cell divisions still provide significant protection (PF = 1.5 and 1.9 at ages 30 and 60, respectively; Fig. S7C, D).
A direct test of the hypothesis that symmetric stem cell division lowers cancer risk would require experimental manipulation of division patterns. The molecular mechanisms that produce highly symmetric assignment of cell fates are unknown, although it should be noted that symmetry fractions between 50% and 100% arise spontaneously if cells simply chose their fates at random (a 50% level is achieved if fate is exclusively determined after division; a 100% level if fate is exclusively determined before). Accordingly, the acquisition of highly symmetric division patterns may be less about implementing special mechanisms than about not implementing mechanisms required to guarantee asymmetric divisions.
Although we have evaluated the effect of mutation accumulation in a single cell type, our results are easily generalized to multi-stage lineages, in which intermediate cell stages (“committed progenitors”) may be modeled as products of stem cells whose rates of differentiation exceed those of renewal (allowing for a steady-state influx of earlier-stage cells). This imbalance by itself enhances the flushing of mutants arising at such stages. Any further propensity for symmetric divisions beyond this minimum level would provide further protection against mutation accumulation, in precisely the same way as described above for stem cells.
Recently, Roche et al. argued that the documented lack of an expected inter-species correlation between cancer risk and body size/longevity (“Peto's paradox”) implies that large, long-lived species must have evolved strategies to reduce cancer risk [60]. Here, we identify symmetric stem cell division as one such strategy. Whether the protection this strategy offers is sufficient to explain natural selection for symmetry in the ancestors of mammals is difficult to know, especially since we do not know the conditions under which selection took place. It may be enlightening to determine whether a prevalence of symmetric vs. asymmetric division patterns coincides with the emergence of larger, longer-lived species.
We modeled mutation accumulation in large populations of stem cells with a discrete-time branching process where each division produces 0, 1 or 2 stem cell daughters, each of which randomly accumulates a mutation (Section 1.1 and 1.2 of Text S1). Small populations were treated using a continuous-time Moran-type model (Section 2.1 of Text S1). Populations, initially mutation-free, were simulated until either a K-fold mutant stem cell appeared or the simulation reached the organism's lifetime. For each parameter set, we generated ∼103 successful runs in which the mutant appeared within a lifetime, which was sufficient to estimate the lifetime cumulative risk of mutation accumulation to an accuracy of ∼3% (). For example, approximately 2000 total runs were required when the risk was ∼50% (e.g. Figs. 1, 2) but as many as 1011 total runs were needed for risks of the order of 10−6% (risk per human crypt in Fig. S7). For computational reasons, we updated our risk estimates after each run.
The number of accumulated mutations was sampled from a uniform distribution typically defined on the interval 1 to 10 whereas log-parameter values of population size, organism lifetime and mutation rates were sampled from a uniform distribution on the log of the ranges (Tables S1, S2, S3). We repeatedly generated parameter sets until we obtained 1,000 parameter sets in which the predicted asymmetric lifetime risk RA (Eq. (S12)) lay in a defined range (10%<RA<99.996%). For each such parameter set, we then simulated the asymmetric and symmetric risks (Fig. 1F).
|
10.1371/journal.pgen.1006516 | Reverse Pathway Genetic Approach Identifies Epistasis in Autism Spectrum Disorders | Although gene-gene interaction, or epistasis, plays a large role in complex traits in model organisms, genome-wide by genome-wide searches for two-way interaction have limited power in human studies. We thus used knowledge of a biological pathway in order to identify a contribution of epistasis to autism spectrum disorders (ASDs) in humans, a reverse-pathway genetic approach. Based on previous observation of increased ASD symptoms in Mendelian disorders of the Ras/MAPK pathway (RASopathies), we showed that common SNPs in RASopathy genes show enrichment for association signal in GWAS (P = 0.02). We then screened genome-wide for interactors with RASopathy gene SNPs and showed strong enrichment in ASD-affected individuals (P < 2.2 x 10−16), with a number of pairwise interactions meeting genome-wide criteria for significance. Finally, we utilized quantitative measures of ASD symptoms in RASopathy-affected individuals to perform modifier mapping via GWAS. One top region overlapped between these independent approaches, and we showed dysregulation of a gene in this region, GPR141, in a RASopathy neural cell line. We thus used orthogonal approaches to provide strong evidence for a contribution of epistasis to ASDs, confirm a role for the Ras/MAPK pathway in idiopathic ASDs, and to identify a convergent candidate gene that may interact with the Ras/MAPK pathway.
| The contribution of epistasis to human biology and complex trait architecture has been subject to intense debate. Despite statistical methods to detect interaction, allele frequencies and study designs have limited our power to address this question. Rather than a statistically-motivated approach, we developed a reverse pathway genetic approach to detect epistasis in in autism spectrum disorders (ASDs). Instead of traditional pathway analysis—exploiting unbiased genetic results to identify biological pathways important for pathophysiology—we hypothesized that the reverse approach of leveraging a specific biological pathway to uncover novel aspects of genetic architecture could be fruitful. We drew on knowledge from Mendelian disorders of the Ras/MAPK pathway (RASopathies) associated with ASD to inform our analyses. First, we implemented an epistasis screen in idiopathic ASD for SNPs that interact with the Ras/MAPK pathway, allowing us to detect genome-wide enrichment in epistasis signal as well as specific interacting loci. Second, we performed modifier mapping within RASopathies and identified overlapping signals with our first approach. Finally, we took the top overlapping region and experimentally demonstrated dysregulated expression in iPSC-derived neural cells from RASopathy subjects. Together, our results establish a human genetics motivated reverse-pathway strategy, which can be applied broadly to study epistasis.
| Genome-wide association studies (GWAS) of common polymorphism association with complex neuropsychiatric traits have yielded recent success mapping single nucleotide risk polymorphisms with modest additive effects, particularly in schizophrenia[1]. However, complementary approaches utilizing the same data support an even greater role for common polymorphism in complex heritable disorders like schizophrenia and autism spectrum disorders (ASDs) than explained by the identified additive effects. For example, analysis in schizophrenia and other traits suggests that heritability is not completely accounted for by common single nucleotide polymorphisms (SNPs), in models inconsistent with contribution primarily from rare SNPs, suggesting that genetic interaction could account for some of the additional contribution of common variation[2,3]. Recent studies in ASDs estimate similarly large contribution of common variants to ASD liability despite inability to identify specific highly significant SNPs[4–6]. The modest effect sizes of individual SNP associations observed to date match expectation based on severe mating and fecundity reduction in schizophrenia and ASDs, however selective pressure could allow for stronger effects of gene-gene interaction, or epistasis[7,8].
Studies of complex traits in mice and in fruit flies have revealed modest main effects and frequency distributions similar to those identified in human GWAS[9]. However, careful study design and leverage of inbreeding have proven that epistasis typically accounts for a majority of the variation in quantitative traits[10–15]. Human GWAS studies have had limited success identifying genetic interactions, although model organism and theoretical evidence suggests that such effects are likely to be important[16]. Several limitations in human studies could account for this. First, power is extremely limited for genome-wide by genome-wide exploration of interaction, due to the potential number of even two-way tests to perform, requiring ‘astronomical’ sample sizes to even begin to address. Second, epistatic variance depends on both the size of genetic effects and the allele frequencies. Alleles with strong functional effects (such as those causing highly penetrant disease) may be more likely to show epistasis, but also be rare in the population. Therefore, in order to explore epistasis in humans, we wanted to take advantage of known rare, functional variation that can contribute to symptoms of complex disease[17].
Emerging evidence suggests that Mendelian diseases (high penetrance, dominant/recessive inheritance) resembling complex disorders may affect the same genes showing common risk variation in the population. There has been longstanding skepticism that similar phenotypes with simple and complex genetic inheritance share biological etiology, however, powerful recent study designs have revealed strong overlap. In Type II Diabetes, at least four of the first 20 genes shown robustly to be associated with the common adult-onset form were previously identified as causes of Mendelian forms[18]. Moreover, comprehensive analysis of medical records suggests widespread pleiotropy such that strong associations have been identified between single-gene Mendelian conditions and complex heritable traits; these specific associations are reflected in enrichment of GWAS SNP associations at the implicated Mendelian loci and increased replication at these loci compared with other GWAS-associated SNPs[19]. Further, this study indicated that these inferred genetic variants often act in a non-additive combinatorial model for certain disorders, including ASDs[19]. Despite strong evidence for association, many single-gene disorders display variation in penetrance or expression of associated complex phenotypes, i.e. reduced penetrance for these traits compared with primary features of Mendelian disease. One theory posits that background common genetic variation could modify risk for complex disease symptoms in the presence of a Mendelian disease. A notable example is autosomal recessive cystic fibrosis (CF), in which a combination of early gene identification (CFTR), common primary mutation (delta F508), frequency (1/3500 in the United States) and familial inheritance have enabled modifier mapping studies[20]. Many variable features of CF influence morbidity and mortality, including lung, liver, intestine, and pancreatic manifestations. Interestingly, in several examples so far, modifier loci in CF overlap with common complex traits[21]. TGFB1 SNPs are associated with CF pulmonary function and with asthma and chronic obstructive pulmonary disease in the general population[22,23]. TCF7L2, CDKAL1, CDKN2A/B, and IGF2BP2 and several other susceptibility loci for type 2 diabetes in the general population are also associated with high risk for CF-related diabetes[24]. Here, we can consider an independent SNP with large effect size in the presence of a Mendelian mutation (but modest effect size in the general population) to be equivalent to gene-gene interaction. One locus is known to be present due to affection with a monogenic disease, and the other is to be identified by modifier mapping. Thus, the active biological pathways involved in complex disease can be powerfully identified in studies with ascertainment for Mendelian conditions.
ASDs are diagnosed based on core deficits in social reciprocity and communication as well as presence of restricted and repetitive behaviors, interests, or activities. These traits have also been long associated with a range of genetically simpler disorders, such as Fragile X syndrome, tuberous sclerosis, Rett syndrome, and Turner syndrome[25]. We hypothesized that Mendelian disorders associated with variable expression of ASD symptoms would be the optimal avenue for identification of gene-gene interaction. At the same time, specific study of natural variation in neurogenetic networks for behavioral traits in other organisms suggested a shift from considering single genes to pathway-based approaches[26]. Similarly, biological network knowledge has been proposed to enhance detection of epistasis[27–29]. We reasoned that a biologically informed network approach, showing promise in Crohn’s disease, bipolar disorder, hypertension and rheumatoid arthritis, may also illuminate ASD genetics[30,31]. Hence, instead of a single Mendelian disease, we chose to focus on a set of syndromes caused by mutations tightly intertwined in a single well-defined signaling pathway.
Disorders of the Ras/MAPK pathway (commonly referred to as RASopathies)[32] are ideal to study for identification of gene-gene interaction in ASD. Ras is a small GTPase with critical signaling functions in the cell, including the MAPK signaling cascade. Although best-known for its role in cancer due to acquired somatic mutations, dysregulation of genes in the Ras/MAPK pathway in human development causes disorders including neurofibromatosis type 1 (NF1: NF1[33]), Noonan syndrome and Noonan syndrome with multiple lentigines[34] (NS: CBL[35], BRAF, KRAS[36], LZTR1[37], NRAS[38,39], PTPN11[40], RAF1[41], RASA2[42], RIT1[43], SHOC2[44–46], SOS1[47,48], and SOS2[37]), Gingival fibromatosis 1 (SOS1[49,50]), Capillary malformation-arteriovenous malformation (CM-AVM) (RASA1[51,52]), Costello syndrome (CS: HRAS[53]), Cardio-facio-cutaneous syndrome (CFC[54]: BRAF,MAP2K1, MAP2K2, KRAS), and NF1-like syndrome (SPRED1[55]). Many of these syndromes share craniofacial dysmorphology, cardiac malformations and cutaneous, musculoskeletal and ocular abnormalities. We have recently studied four RASopathies (NF1, NS, CS, and CFC) and found association with both threshold measures correlated with clinical ASD diagnosis and quantitative ASD trait measures[56]. Our phenotype analyses suggested additional similarities with idiopathic ASDs, such as a male-biased sex ratio. Other groups have performed independent studies with highly consistent findings[57–62]. Studying multiple disorders in the same biological pathway could therefore both increase the power of our study and increase the likelihood of results translating to an even broader diagnostic category, idiopathic ASDs.
Instead of traditional pathway analysis—taking advantage of unbiased genetic analysis results to identify biological pathways important for disease pathophysiology—we hypothesized that the reverse approach of utilizing a specific biological pathway to uncover novel aspects of genetic architecture could be fruitful. In particular, we sought to overcome the limited power of genome-wide screens for two-way epistasis and the modest effects anticipated for common polymorphisms in complex disease. Thus, we designed two orthogonal approaches to leverage a well-defined biological pathway in order to learn more about gene-gene interaction in ASDs. 1) We performed a genetic screen for epistasis in idiopathic ASD subjects searching genome-wide for interaction partners of common polymorphisms in the Ras/MAPK pathway genes, thereby limiting one side of the pairwise test to include only polymorphisms relevant to a small number of RASopathy genes; 2) In parallel, we mapped SNPs influencing a quantitative measure of social responsiveness in RASopathy subjects ascertained for rare, major effect mutations in the Ras/MAPK pathway. In the latter case, the autosomal dominant RASopathy mutation is one locus, and the second locus involved in interaction will be identified by genome-wide SNP-based QTL mapping. We used these complementary approaches to show that ‘reverse pathway’ screening is a feasible approach to identify epistasis relevant to a complex heritable trait.
We generated an idiopathic ASD GWAS dataset by utilizing each available published dataset [Autism Genetic Resource Exchange (AGRE), Autism Genome Project (AGP), Simons Simplex Collection (SSC)] in addition to in-house generated data [University of California, San Francisco (UCSF)], all of which were comprised of family trios with one affected offspring and both parents. We performed a transmission disequilibrium test (TDT) for ASD association in our final dataset of 4,109 trios. We first performed set-based analysis for polymorphisms within 5kb of each of the known RASopathy genes in the Ras/MAPK pathway (S1 Table). Then, we compared the proportion of SNPs exceeding a false discovery rate (FDR) threshold of q = 0.2 to permutations using length-matched random gene sets, showing that common polymorphisms within Ras/MAPK genes are significantly enriched for association with idiopathic ASDs (P = 0.02) (Fig 1). This evidence for main effects of Ras/MAPK polymorphism in ASDs supports our rationale for testing epistasis involving Ras/MAPK SNPs.
We next performed ‘set-by-all’ tests for two-way genetic interactions with polymorphisms in the Ras/MAPK SNP set (S1 Table). This is performed by testing each polymorphism in the Ras/MAPK SNP set (SNP 1) for correlation with any other polymorphism (on an unlinked autosome) across the genome (SNP 2) in cases with ASDs. SNP1-SNP2 correlation across independent chromosomes in a genetically homogeneous population is considered evidence for epistasis contributing to disease risk. We calculated a genome-wide significance threshold (P = 7.6x10-10) accounting for multiple LD groups per Ras/MAPK gene and the proportion of the genome tested for each genome-wide epistasis screen (see S2 Table). At this threshold, 569 SNP pairs representing 19 independent region pairs showed epistasis in cases, which was significantly increased compared with matched pseudo-controls (OR 3.1, P < 2.2x10-16). We observed an excess of association at several additional P-value thresholds in this analysis in ASD cases compared with matched pseudo-controls (P < 2.2x10-16) (Table 1, Fig 2). We also identified individual epistasis signals surpassing a gene-based significance threshold set by dividing the GWAS significance threshold (P = 5.0x10-8) by the number of genes in the Ras/MAPK set for an approximate independent hypothesis-testing estimate (P = 2.9x10-9) considering all SNPs in a single gene part of the same hypothesis (i.e. LD-independent SNPs relevant to the same gene considered non-biologically independent due to potentially similar functional consequences) (Table 1, S1A–S1F Fig). The data underlying these results show that the primary driving genotype-combination category is double-heterozygotes, which shows dramatic increase compared with expected counts in cases, but similar to expected counts in pseudo-controls (S3 Table). This would be expected for minor allele interactions increasing risk, as most combinations including minor allele homozygotes have very low counts.
We performed a number of negative and positive control analyses in order to exclude any bias or artifact. For our first negative control, we performed similar set-by-all epistasis testing for a permuted SNP set (median result in the main effect enrichment analysis above). In testing for epistasis enrichment for a non-candidate pathway-selected set, we do not observe enrichment of epistasis results in cases compared to pseudo-controls at our top significance thresholds (OR < 1) (S2 Fig). Second, for our observed Ras/MAPK results above, we not only compared a pseudo-control correlation approach (as above; in Table 1 all 41 top SNP pairs show different OR between cases and pseudo-controls P < 0.05), but also used matched-sex parent controls, which provided similar results to pseudo-controls (no epistasis P < 10−6 for SNP pairs in Table 1; significant overrepresentation genome-wide in cases at all P-value thresholds considered from P < 10−6 to P < 7.6 10−10). As a positive control, we selected the small homogeneous set of cases and matched unaffected siblings from unique families (N = 1,136) from the SSC dataset, and similarly to pseudo-controls, we observed significant epistasis enrichment in cases and diminishing odds ratios at varying P-value thresholds down to P < 10−6 (S3 Fig). Finally, we validated the top results (P < 1.0x10-6) with the trio correlation test for epistasis. This independent method had highly correlated P-values with the case-only approach, suggesting the results are consistent with epistasis and not marginal effects (rho = 0.78, P< 2.2x10-16). We observed that all results from the PLINK[63,64] case-only epistasis analysis with P< 1.0x10-6 also had P< 4.4x10-5 in the trio correlation test, thereby confirming the results with an independent method (S4 Fig).
In order to utilize a separate approach to assess evidence for multiple genetic hits contributing to ASD symptoms, we ascertained individuals for having a RASopathy (NF1, NS, CS, and CFC). Although these individuals have a dominant germline mutation in the Ras/MAPK pathway causing some highly penetrant features, we have previously shown that ASD symptoms are variable both across and within disorders[56]. Thus, we performed a modifier screen via genome-wide association based on a quantitative social responsiveness trait, measured by the Social Responsiveness Scale (SRS) (S5 Fig). The SRS is a questionnaire measure with normally distributed and highly heritable scores in the general population strongly reflective of clinical ASDs[65]. Thus, identifying modifiers of an ASD-related trait in individuals ascertained for a RASopathy could identify interactors with the Ras/MAPK pathway relevant to ASDs, as we consider the RASopathy locus and the modifier locus interacting to influence the ASD-relevant trait. We performed quantitative trait locus (QTL) mapping within-disorder and meta-analysis across the four disorders, as well as QTL mapping in sibling controls. We did not observe any SNPs meeting criteria for genome-wide significance (P = 5.0x10-8), however, the most significant SNP in this QTL analysis (rs62621010, P = 5.6x10-7, Table 2) is 0.39 Mb from the locus with the second-most significant epistasis signal in ASDs (rs114490548, Table 1). The rs62621010 putative modifier SNP does not show evidence for association with SRS in sibling controls. The region flanking and between rs62621010 and rs114490548 (chr7:37,749,392–38,139,570) contains genes ELMO1, GPR141, NME8, SFRP4, EPDR1, and STARD3NL. The two SNPs have low LD in 1000genomes measured by r2, although there is high D’ variably across the region (Fig 3). Neither SNP is represented in both ASD and RASopathy datasets (or can be imputed) for direct comparison.
For prioritization of genes within the top region jointly identified by epistasis analysis in idiopathic ASDs and QTL-mapping in RASopathy subjects, we assessed expression level of genes in the chromosome 7 region in RASopathy neural cell lines. Our reasoning was that if a RASopathy mutation resulted in expression dysregulation of a gene, it provides strong biological plausibility for interaction with Ras/MAPK signaling. We utilized qRT-PCR to compare RNA extracted from CFC (BRAF c.770A>G, p.Q257R) patient-derived iPSC neural cell cultures (5 weeks) and compared to control-derived matched cultures[66]. All cultures were positive for mature neuronal and astrocyte markers, MAP2 and GFAP (S7 Fig). In the first experiment (CFC N = 3; control N = 3; performed in technical triplicates), we identified two genes appearing to be downregulated in CFC lines (GPR141, P = 0.02; SFRP4, P = 0.03, S5 Table, Fig 4). In order to independently replicate this result, we utilized RNA extracted from a second batch of independently-derived neural lines from a set of overlapping individuals at 5 weeks (CFC N = 3; control N = 2; performed in technical triplicates), and showed that GPR141 was again significantly downregulated (P = 0.04, S5 Table, Fig 4). GPR141 encodes a brain-expressed orphan g-protein coupled receptor in the rhodopsin family.
We find the results of this reverse-pathway genetic study compelling for a number of reasons. We limit one side of a two-way interaction test to a relevant genetic pathway, and obtain genome-wide enrichment for epistasis in a common, complex human disease. Further, specific epistasis results survive independent genome-wide or gene-by-genome multiple testing correction, and reside near compelling candidate loci. Our second approach of modifier mapping for a relevant quantitative trait in subjects with Mendelian disorders in the same pathway, RASopathies, converges on several regions overlapping the ASD epistasis results. A gene in the top overlapping region across analyses showed expression dysregulation in neural cell lines from RASopathy subjects. Together, our results from each step of this study suggest we have identified powerful approaches to unravel complex genetic mechanisms.
First, in a reverse pathway-driven approach, we obtained strong evidence for gene-gene interaction in a human complex trait. We have previously used candidate-pair approaches to find interaction effects in ASDs, which have been independently replicated and shown to have a functional basis[67–71], and other studies have performed epistasis screens limiting all discovery to a candidate set or pathway[72–76]. However, our set-by-genome strategy can not only identify novel disease-relevant biological relationships, but this study also directly shows a genome-wide excess of epistasis signal in human disease (P < 2.2 x 10−16), achieved by constraining the set of interaction partners based on biological and genetic knowledge. The approach we have used here is in principle straightforward and could be applied to other disorders or pathways.
Beyond global enrichment for epistasis in ASDs, we identified specific loci meeting criteria for gene-based or genome-wide significance considering the number of Ras/MAPK genes or LD groups for which genome-wide screening was performed (P < 2.9 x 10−9; P < 7.6 x 10−10). Loci identified to be putative Ras/MAPK interaction partners influencing ASD risk in this way include or are adjacent to some already strong ASD candidate genes. GRM7 has been identified in rare, de novo CNV deletions[77,78] and single nucleotide mutations[79] in subjects with ASD and in a candidate gene SNP study. It encodes a metabotropic glutamate receptor critical for early development[80]. KIRREL3 has likewise been implicated by rare genetic disruptions[81–84] and inclusion in expression networks of common polymorphism association[85], as well as mouse behavioral anomalies[86]. It is thought to be key for synaptogenesis. NIPA2 and CYFIP1 are located on 15q11.2 in the region affected by an interstitial microduplication syndrome associated with ASDs[87,88]. PCDH9, PITX, REEP3, NBEA, and OPRM1 are additional ASD candidate genes listed in SFARIgene[89] (see web resources).
Similarly, several genes implicated by this epistasis analysis are specifically relevant to the Ras/MAPK pathway. SPRY1 encodes a classic inhibitor of the Ras/MAPK pathway, and plays a critical role in determining the balance between proliferation and differentiation for cortical patterning and cerebellar development[90–92]. DCC encodes the receptor deleted in colorectal cancer (DCC), which recruits proteins to promote axon outgrowth and guidance during neurodevelopment, and has been shown to interact with a Ras inhibitor[93]. GATA3 is a transcription factor known for its role in T-cell development, however it also helps to control excitatory/inhibitory balance by determining GABAergic vs. glutamatergic fates during neurodevelopment[94]. Ras/MAPK signaling can regulate the stability of GATA3 post-transcriptionally by inhibiting the ubiquitin-proteasome pathway[95]. GHRHR encodes the growth hormone releasing hormone receptor which can activate Ras/MAPK signaling[96].
Second, our results are in line with the previous prediction that not only would non-additive effects be a major contributor to ASDs, but also that Mendelian syndromic genes are likely to be enriched for main effects[19]. We observe enrichment of additive association signal for common polymorphisms near the Ras/MAPK genes, supporting traditional pathway analyses that have identified this pathway as a major contributor to ASDs and indicating general overlap in biological etiology between Mendelian and complex traits [97–99]. In addition, our results in idiopathic ASD SNP datasets strongly support our findings in rare RASopathies. One of the highly significant epistasis loci is near the top result (albeit non genome-wide significant) in a QTL modifier mapping approach in RASopathy subjects. A gene in this region, GPR141, demonstrates reproducibly reduced expression in RASopathy neural cell lines. Although this is merely circumstantial with regards to the observed epistasis and QTL results, it provides biological plausibility for interaction with the Ras/MAPK pathway.
The function of GPR141, an orphan g-coupled protein receptor, is currently unknown. Sequence-structure based phylogeny suggests potential ligand association with N-arachidonylglycine (NAGly)[100]. The well-studied GPR18 in this class has been shown to mediate concentration-dependent phosphorylation of ERK 1/2 in the presence of NAGly[101]. NAGly is known to be anti-nociceptive and is thought to reversibly inhibit calcium currents in sensory neurons and have additional minor effects on sodium currents[102]. Further experimentation specific to GPR141 would be necessary to speculate about its potential interaction with the Ras/MAPK pathway or functional role in ASDs. A second gene in the region, SFRP4, showed nominal reduced expression in our first experiment and non-significant but consistent expression reduction in the second experiment. Secreted frizzled-related protein 4 (SFRP4) is an antagonist for Wnt ligands, inhibiting the canonical Wnt signaling pathway. Based on both linkage disequilibrium in the region and our experimental results, we cannot rule out that there are two separate loci or genes of interest relevant to ASD traits and Ras/MAPK signaling within this region of chromosome 7 that are proximal but do not represent overlap across analyses.
In addition, two other noteworthy loci (one between KIRREL3 and ETS1, the other between GATA3 and CELF2) overlap between top results from ASD epistasis and RASopathy modifier mapping and contain genes of particular interest. KIRREL3 is a strong ASD candidate gene, as described above, and ETS1 encodes an effector of Ras/MAPK signaling mediating cell migration and transcriptional activation expressed in astrocytes[103,104]. GATA3 encodes a known Ras/MAPK interactor, as described above, and CELF2 encodes CUGBP[105], Elav-like family member 2, a fetal and adult brain expressed regulator of alternative splicing also likely to be involved in mRNA editing and translation. Together, these loci overlapping with significant ASD epistasis results suggest that studying rare Mendelian disorders associated with symptoms of complex traits is a highly effective and relevant study design.
Further, some of the top results in the RASopathy modifier screen, although not meeting criteria for genome-wide significance, are within or adjacent to previously implicated ASD candidate genes. MACROD2 has been identified near a GWAS association, shown to be associated with ASD traits in the general population, and described as part of the gene expression network of an independent ASD GWAS locus[106–108]. In addition, the MACROD2 locus has been identified as associated with temporal lobe volume[109]. Likewise, CDH10 was initially located near a GWAS association[110]. RBFOX1 (or A2BP1) has been identified as disrupted by rare translocation or CNV in ASDs[111–113]. The encoded protein (FOX1) controls alternative splicing and transcription[114], including many other ASD candidate genes, and is thought to be a key regulator of neurodevelopment[115]. DISC1, known for identification via rare structural variant in schizophrenia, has also been associated with ASDs[78,112,116,117]. Additional genes listed in SFARIgene[89] adjacent to our top RASopathy social responsiveness QTL results include CNTN5, ESRRB, GABRB1, GNB1L, and ICA1.
Some clear Ras/MAPK related genes contain or are adjacent to top RASopathy modifier SNPs, as well. Primary RASopathy gene SPRED1 is a negative regulator of the pathway; MAP3K2 encodes MEK kinase 2; RAPGEF2 encodes a Ras activator thought to control developmental neuronal migration in the cortex and formation of the corpus callosum[118,119]. ELK3 encodes a transcription factor downstream of Ras/MAPK signaling; ETS1 and GATA3 are additional transcription factors with demonstrated relationships to Ras/MAPK signaling described above[120,121]. It is striking that the modifier-mapping approach in subjects ascertained for Mendelian disease appears to contain many plausible loci despite the small sample size compared with modern GWAS designs for case-control studies. However, it should be noted that proximity between a SNP and gene does not indicate a functional relationship, so further information would be needed to directly relate our identified loci from any analyses in this study to specific genes listed in tables or discussed here.
In summary, we have used a variety of approaches under a reverse pathway tactic of defining a relevant biological pathway and leveraging it to study a proposed genetic mechanism. In this case, we chose a pathway of interest based on human Mendelian disorders with overlapping symptoms shared by a genetically complex trait. Together, our strategies were successful in validating a role for the Ras/MAPK pathway in idiopathic ASDs, demonstrating highly significant enrichment for epistasis in ASDs, and identifying specific candidate loci interacting with the Ras/MAPK pathway pertinent to symptoms of ASDs. Our experimental data confirmed expression dysregulation of a gene within a convergently identified locus in RASopathy-specific neural cell lines. Future studies would be useful to follow up additional candidates identified by these approaches or extensions of them.
ASD genotype data sets were collected from multiple sources. We obtained previously published genotype data as study investigators (UCSF-Weiss) or by application to AGRE, SSC, and dbGAP (AGP). Genotyping of each dataset was previously performed on Illumina or Affymetrix genotyping arrays as described in S6 Table. Diagnostic criteria were previously described in the respective references (S6 Table), but in summary, the Autism Diagnostic Interview-Revised (ADI-R)[122], and/or Autism Diagnostic Observation Schedule (ADOS)[123] criteria was used for diagnosis for the AGRE, AGP and SSC datasets, and clinician diagnosis for the UCSF-Weiss dataset. The ASD affected child and both parents were included in each study. All samples were anonymized for analysis.
Data preparation, quality control, and imputation were conducted as described previously in Mitra, I. et al. [124]. First, SNPs were filtered using PLINK[63,64] (see web resources) for Hardy-Weinberg equilibrium (HWE), call rate, minor allele frequency (MAF) and Mendel errors separately in each ASD dataset (S6 Table). Next, imputation was performed separately for each dataset using IMPUTE2[125] (see web resources), following the recommended pipeline. Lastly, each ASD dataset was combined together, and the following quality control steps were performed: SNPs with severe departure (P < 1.0x10-6) from HWE in Caucasian founders were removed; SNPs were removed if they had different MAF (P < 1.0x10-6) in Caucasians between multiple datasets; SNPs were removed if they had MAF < 1% in Caucasians, or MAF < 2% in the combined dataset. We excluded chromosome X. The final dataset for the analysis included 4,471,807 autosomal SNPs.
Initial data preparation and quality control for ASD individuals was conducted for each dataset, as described previously in Mitra, I. et al. [124]. For each dataset, the following individual quality control filters were applied using PLINK[63,64]: genotyping rate, heterozygosity rate, verifying individual sex, verifying known relationships, removing individuals contributing to confounding relationship, and keeping one instance of individuals present in multiple studies. After combining the multiple datasets, sex and family structure were re-checked.
To avoid population stratification, we selected only Caucasian individuals for the analysis. Ancestry was determined using the first two principal components resulting from multidimensional scaling with PLINK[63,64] (—mds-plot option) (S8 Fig). A proband and both parents were required to fall within the Caucasian cluster for inclusion in the analysis. Only unrelated complete trios were used. The final dataset included 4,109 ASD affected cases (3,517 males and 592 females) with both parents. A QQ plot for main effect association is shown in S9 Fig. As our data were family-based, we used non-transmitted parental alleles, commonly known as pseudo-controls, generated by the—tucc option in PLINK[63,64] instead of unrelated healthy individuals. These 4,109 pseudo-controls are perfectly matched to cases for ancestry, thereby serving as a control for any population confounding.
We included the following genes in the Ras/MAPK pathway as the set of interest: NF1, BRAF, SOS1, RASA2, RASA1, SOS2, MAPK1, MAP2K1, SPRED1, CBL, SHOC2, PTPN11, RAF1, KRAS, LZTR1, RIT1, and NRAS. Genes with no SNPs represented within the ASD dataset (HRAS, MAP2K2, and MAPK3) could not be included. We extracted all SNPs within 5kb of each gene. 2,520 SNPs were included in the Ras/MAPK SNP set to be used for enrichment assessment in association and for the epistasis analysis (S1 Table).
Permutation testing procedures were implemented to establish significance of association signal enrichment in the Ras/MAPK SNP set. First, we performed a TDT (—tdt in PLINK[63,64]) in the 4,109 ASD trios to test for association. Then, at a given Benjamini and Hochberg’s FDR[126] (q = 0.2), we compared the percent of SNPs in the Ras/MAPK set (S1 Table) meeting this criterion in the TDT results to the empirical null distribution produced by permuted data. The FDR threshold was chosen to maximize power by minimizing instances of 0 or 100% of SNPs meeting the criterion, but is an arbitrary threshold not intended to indicate significance, only as a means of comparison with permuted data. At lower FDR thresholds, no SNPs are available to test, and beginning at FDR 0.2 approximately 10% of SNPs pass the threshold, which is sufficient for testing enrichment (S7 Table). To generate each of 100 permuted gene sets, a random gene was selected from 100 RefSeq (see web resources) genes with the most similar size (using the longest transcript) to each Ras/MAPK gene (S1 Table) and compiled into a SNP set using the same procedure (all SNPs within 5kb). These sets appear well matched to the Ras/MAPK set, as they have similar numbers of SNPs (Ras/MAPK 2,520; permutation median 2,528) and allele frequency (Ras/MAPK average 0.18; permutation average 0.19). The permuted gene lists were analyzed with the same protocol as the Ras/MAPK set to produce the null distribution for comparison. The empirical P-value was calculated as the proportion of results from the null distribution equal to or greater than the results from the Ras/MAPK set[127].
For the epistasis analysis, the dataset was comprised of 4,109 ASD affected cases and 4,471,807 SNPs. We used the—fast-epistasis test with the case-only and set-by-all options implemented in PLINK[63,64] to perform pairwise epistasis tests between each SNP in the defined Ras/MAPK SNP set (N = 2,520 SNPs) and SNPs across the autosomal genome. This epistasis test is performed by testing an allelic odds ratio, based on collapsing the 4N independent alleles observed at two loci in a sample of N individuals into a 2x2 table, so the allele (not the individual or haplotype) is the unit of analysis. The four cells are (a) 4*AABB+2*AABb+2*AaBB+AaBb, (b) 4*AAbb+2*AABb+2*Aabb+AaBb, (c) 4*aaBB+2*aaBb+2*AaBB+AaBb, (d) 4*aabb+2*aaBb+2*Aabb+AaBb. The odds ratio is then estimated as ad/bc with variance 1/a+1/b+1/c+1/d. This test follows a standard normal distribution under the multiplicative model of no interaction. Appropriate type I error rates have been observed in simulation and power is equivalent to a logistic regression test for epistasis. The correlation with a logistic regression analysis is high (r = 0.995) [63],[64].
Two SNPs on the same chromosome were excluded from consideration in order to conservatively eliminate linkage disequilibrium or effects of rare variants. We calculated a genome-wide significance threshold (P = 7.6x10-10) accounting for multiple LD groups per Ras/MAPK gene and the proportion of the genome tested for each genome-wide epistasis screen (calculations shown in S2 Table). We also utilized a gene-based significance threshold set by dividing the GWAS significance threshold (P = 5.0x10-8) by the number of genes in the Ras/MAPK set for an approximate independent hypothesis-testing estimate (P = 2.9x10-9), as genes were the functional unit of the set-based testing. Because we use a case-only approach with similar properties to identifying significant correlation, we have calculated power based on correlation statistics. In this instance, for our stringent P-value threshold of 7.6 x 10−10, we have >80% power to detect r2 = 0.11 with our sample size 4,109.
In order to rule out nonspecific effects and control for false-positives, we performed the same epistasis analysis on the 4,109 matched pseudo-controls as a negative control. To test if the most significant epistasis results between the ASD cases and ASD pseudo-controls were significantly different, we performed a 2x2 chi-square test using the number of SNPs meeting a given significance threshold in cases compared with pseudo-controls. We conducted a second negative control analysis to rule out results due to interactions associated with viability by using 4,109 parents that were of the same sex as the ASD cases. In addition, we performed epistasis testing in our sample for a permuted SNP set included in a non-candidate pathway selected set as a negative control and for Ras/MAPK set in a homogeneous set of cases (N = 1,136) and matched unaffected siblings (N = 1,136) from unique families from the SSC dataset as a positive control. For both, we measured odds ratios at varying P-value thresholds down to P < 10−4 (S2 and S3 Figs).
To statistically validate the results from the case-only epistasis test, we used an independent test, called the trio correlation test[128].The trio correlation test leverages information from the parental genotypes to compute the expected distribution of the offspring genotypes then used in a correlation test. This test was provided to us as an R script (see web resources). To individually test candidate SNP pairs for interaction, we tested the nominally significant interaction results (P < 1.0x10-6) from the PLINK[63,64] epistasis test. To confirm the absence of false-positive results, we performed a Spearman correlation test in R (see web resources) between the P-values of the nominally significant PLINK[63,64] epistasis test results (P < 1.0x10-6) and their corresponding P-value from the trio correlation test (S4 Fig).
We have previously described in detail the recruitment and phenotype data collection in RASopathy subjects[56]. In summary, patients with a physician (medical geneticist or neurologist) confirmed NF1, NS, CS, CFC diagnosis were included in the study. We recruited subjects from the NF/RAS Pathway Genetics Clinic UCSF, UCSF NF Symposium, RASopathy support groups (NF, Inc., Children’s Tumor Foundation, Noonan Syndrome Foundation, CFC International, Costello Syndrome Family Support Network, and Costello Kids), and three national RASopathy family meetings (Chicago, Illinois, USA, July 2011; Berkeley, California, USA, July 2009; Orlando, Florida, USA, August 2013). In addition, NF1 patients were recruited at University of California, Los Angeles (UCLA) through online postings (NF, Inc., and Children's Tumor Foundation), and the Neurofibromatosis and Neurocutaneous Disorders clinic at the Children's Hospital, Los Angeles. We enrolled the unaffected siblings of RASopathy subjects as controls.
To measure ASD symptoms, we used the SRS questionnaire[129]. The SRS questionnaire is a quantitative and continuous measure of social ability. Parents or persons well-acquainted with the study participant answered the 65-item SRS questionnaire regarding traits characteristic of ASD. Following the SRS manual, we calculated the raw score for each individual, and then calculated the sex-normalized T-scores. A total of 257 RASopathy patients and 142 RASopathy-unaffected full siblings of RASopathy patients with SRS phenotype data were recruited. Further detailed information about the SRS questionnaire data for the sample population can be found in Adviento et al. (2014)[56].
All participants provided blood or saliva samples for DNA extraction. Blood samples were collected by venipuncture using standard procedures. Saliva samples of families involved in the study were collected by mail or in person at family meetings. Participants provided saliva samples using the Oragene Discover kit (OGR-250 for children and OGR-500 for adults) by DNA Genotek (see web resources). DNA was extracted using the manufacturer’s standard protocol. All specimens were anonymized for analysis.
All DNA samples were genotyped in the Genomics Core Facility (GCF) of UCSF on the Affymetrix Axiom EUR array following standard manufacturer protocols. The Axiom EUR array contains approximately 675,000 SNPs across the genome[130]. Genotype calling was performed using Axiom GT1 algorithm as part of the Affymetrix Genotyping Console™ (GTC) Software (see web resources). For analysis, we used samples that had a dish QC (probe intensity) threshold greater than 82% and a genotype call rate greater than 97%. Additional quality control procedures were performed in PLINK[63,64]. Identification of samples was validated based on sex and familial relationships, using pairwise identity by decent (IBD) estimation (—genome). Samples failing quality control checks, including incorrect sex (—check-sex), excessive heterozygosity (—het), and other indicators of DNA contamination were removed. One sample was selected for analysis from monozygotic twin pairs or duplicate samples. To ensure that within each group all subjects were unrelated, a maximum of one person per family was selected to be in each of the control sibling, CFC, CS, NF1, and NS groups. SNPs were removed based on the following quality filters: ≥ 5% missing rate, and Hardy-Weinberg equilibrium P ≤10−4. The final dataset used for analysis included 658,746 SNPs with a 99.70% genotyping rate. The final individuals included were 209 RASopathy subjects (49 CFC, 50 CS, 60 NF1, 50 NS) and 84 control siblings.
We performed a multidimensional scaling analysis of genome-wide pairwise identity-by-state (IBS) distances in PLINK[63,64] for all individuals in the dataset. We used the first five dimensions as covariates in the analysis to correct for population stratification and batch effects. To accurately compare between the RASopathy and control sibling groups, we scaled the T-scores within each group (CFC, CS, NF1, NS, and sibling) so that the mean of the values was 0 and variance was 1, and excluded outlier values greater or less than 3 standard deviations (SD) from the mean (S5 Fig). For each group (CFC, CS, NF1, NS, sibling), we performed QTL mapping by implementing in PLINK[63,64] a linear regression analysis using the scaled SRS scores as a quantitative trait (—linear). This resulted in the multi-linear regression model Y = b0 + b1*ADD + b2*COV1 + b3*COV2… bn*COV5 + e. To analyze the RASopathy groups together (CFC, CS, NF1, and NS) with greater statistical power, we used METASOFT[131] (see web resources) to conduct a random effects meta-analysis using Han and Eskin's random effects model[131]. We also report the Cochran’s Q statistic, calculated using METASOFT[131], to analyze heterogeneity between RASopathy groups. The data underlying the top six potential modifiers are graphically represented in S6 Fig, by boxplot (MAF>0.05) or distribution (MAF≤ 0.05).
To induce neural differentiation, free floating induced pluripotent stem cell (iPSC) aggregates were formed for 24 hours in mTeSR1 (Stemcell Technologies) and then switched to a Neural medium [DMEM/F12 (Invitrogen), N2 supplement (Invitrogen), MEM-NEAA (Gibco) and 2 μg/ml Heparin (Sigma-Aldrich)] with media exchange every other day[132]. To promote neural induction, the small molecules SB431542 (5 μM, Stemgent) and LDN-193189 (0.25 μM, Stemgent) were added for 48 hours. On day 3, aggregates were attached to 6 well plates and cultured in neural media for an additional week during which rosettes appeared in the colonies. On day 11, neuroepithelial cells in the center of the colonies were mechanically removed and kept as free floating aggregates. At day 25 of neural differentiation, neurospheres were dissociated into single cells using Accutase (Stem Cell Technologies) and cultured as monolayer neural progenitor cells (NPCs). NPCs were plated into poly-ornithine/laminin -coated plates at 50,000 cells/cm2 and fed with forebrain neuronal medium [Neurobasal medium (Invitrogen), supplemented with N2 supplement (Invitrogen), and B27 supplement (Invitrogen)[132]. Cells were fed twice a week and RNA samples were extracted at 5 weeks after plating. Total RNA was isolated using RNeasy Mini kit (Qiagen) according to manufacturer’s instruction.
Complementary DNA (cDNA) was produced from 1 μg of total RNA using High Capacity RNA-to-cDNA Kit (Life technologies). The qRT-PCR assay was performed using approximately 20 ng of cDNA and Taqman gene expression master mix in a QuantStudio™ 6 Flex Real-Time PCR System (Applied Biosystems). Expression level was determined by relative quantification in comparison to the endogenous control gene GUSB. Expression of each target gene (ELMO1, GPR141, NME8, SFRP4, EPDR1, and STARD3NL) was assessed relative to a control sample (comparative Ct method). Samples were run in technical triplicates, and the threshold suggested by the instrument software was used (after visual confirmation) to calculate the Ct. Outlier replicate samples were excluded from analysis. The Taqman probes used in this study are summarized in S8 Table.
All subjects or their legal guardians gave written informed consent. This study was approved by the institutional review boards of UCSF Human Research Protection Program (CHR #10–02794) and University of California, Los Angeles (UCLA, IRB#10–000518).
Affymetrix Genotyping Console™ (GTC) Software: http://www.affymetrix.com/estore/browse/level_seven_software_products_only.jsp?productId=131535#1_1
Haploview: https://www.broadinstitute.org/scientific-community/science/programs/medical-and-population-genetics/haploview/haploview
IMPUTE2: https://mathgen.stats.ox.ac.uk/impute/impute_v2.html
LocusZoom: http://locuszoom.sph.umich.edu/locuszoom/
METASOFT: http://genetics.cs.ucla.edu/meta
Oragene Discover kit OGR-250 by DNA Genotek: http://www.dnagenotek.com/US/products/OGR250.html
Oragene Discover kit OGR-500 by DNA Genotek: http://www.dnagenotek.com/US/products/OGR500.html
PLINK: http://pngu.mgh.harvard.edu/~purcell/plink/index.shtml
R—A language and environment for statistical computing: http://www.R-project.org/
RefSeq Genes Database–UCSC: http://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/knownToRefSeq.txt.gz
SFARIgene: https://gene.sfari.org/autdb/Welcome.do
Trio Correlation Test R script: https://github.com/BrunildaBalliu/TrioEpi
The accession number for the UCSF RASopathies social responsiveness and genotype data reported in this paper is The National Database for Autism Research (NDAR) ID 1966.
|
10.1371/journal.pcbi.1000671 | Characterizing Dynamic Changes in the Human Blood Transcriptional Network | Gene expression data generated systematically in a given system over multiple time points provides a source of perturbation that can be leveraged to infer causal relationships among genes explaining network changes. Previously, we showed that food intake has a large impact on blood gene expression patterns and that these responses, either in terms of gene expression level or gene-gene connectivity, are strongly associated with metabolic diseases. In this study, we explored which genes drive the changes of gene expression patterns in response to time and food intake. We applied the Granger causality test and the dynamic Bayesian network to gene expression data generated from blood samples collected at multiple time points during the course of a day. The simulation result shows that combining many short time series together is as powerful to infer Granger causality as using a single long time series. Using the Granger causality test, we identified genes that were supported as the most likely causal candidates for the coordinated temporal changes in the network. These results show that PER1 is a key regulator of the blood transcriptional network, in which multiple biological processes are under circadian rhythm regulation. The fasted and fed dynamic Bayesian networks showed that over 72% of dynamic connections are self links. Finally, we show that different processes such as inflammation and lipid metabolism, which are disconnected in the static network, become dynamically linked in response to food intake, which would suggest that increasing nutritional load leads to coordinate regulation of these biological processes. In conclusion, our results suggest that food intake has a profound impact on the dynamic co-regulation of multiple biological processes, such as metabolism, immune response, apoptosis and circadian rhythm. The results could have broader implications for the design of studies of disease association and drug response in clinical trials.
| Peripheral blood is the most readily accessible human tissue for clinical studies and experimental research more generally. Large-scale molecular profiling technologies have enabled measurements of mRNA expression on the scale of whole genomes. Understanding the relationships between human blood gene expression profiles and clinical traits is extremely useful for inferring causal factors for human disease and for studying drug response. Biological pathways and the complex behaviors they induce are not static, but change dynamically in response to external factors such as intake/uptake of nutrients and administration of drugs. We employed a randomized, two-arm cross-over design to assess the effects of fasting and feeding on the dynamic changes of blood transcriptional network. Our work has convincingly shown that feeding or increasing nutritional load affects the human circadian rhythm system which connects to other biological processes including metabolic and immune responses. We believe this is a first step towards a more comprehensive population-based study that seeks to connect changes in the blood transcriptome to drug response, and to disease and biology more generally.
| Elucidating networks that define biological pathways underlying complex biological processes is an important goal of systems biology. Large-scale molecular profiling technologies have enabled measurements of mRNA and protein expression on the scale of whole genomes. As a result, understanding the relationships between genes and clinical traits, and inferring gene networks that better define biochemical pathways that drive biological processes, has become a major challenge to understanding large-scale data sets generated from these technologies. For the majority of published gene expression profiling experiments, they are carried out at a single pre-defined time point across all samples, where the implicit assumption is that the steady state for the corresponding biological system is well approximated at a single time point. The steady state in this context represents a baseline state of the system under study in which the system is least likely to change and has the least amount of variability due to environment.
Because biological pathways and the complex behaviors they induce are dynamic [1], transcriptional response, interactions among proteins and other such processes, take time and ultimately lead to time-dependent variations in mRNA, protein and metabolite levels. These types of temporal variation over time are difficult to study directly with measurements taken at only a single time point. Recently, studies applying time series to temporal gene expression data have been published, covering a range of experiments focusing for instance on the SOS DNA repair system in E.coli [2], the cell cycle in yeast [3], muscle development in Drosophila [4] and cell cycle processes in human cell lines [5]–[6].
Coexpression networks are based on pair-wise gene-gene correlations of expression data, revealing functional modules in the network that elucidate pathways that drive core biological processes [7]–[8] or pathways that underlie complex human disease [9]–[10]. Coexpression networks provide global views of network structures, but by themselves cannot yield causal relationship between genes or between genes and clinical traits. Using a Bayesian network approach to integrate genetic, expression, and clinical data in segregating populations, we have previously demonstrated that such causal relationships can be inferred [11]–[14]. While these network approaches have proven useful in elucidating complex traits emerging in complex systems at the population level, they have however been based on data sampled at a single time point.
A static Bayesian network (SBN) is a graphical model that encodes a joint probability distribution on a set of stochastic variables , which can be decomposed as , where represents the parent set of . Similar to a static Bayesian network, a dynamic Bayesian Network (DBN) is also a graphical model with a joint probability distribution. The main difference between them is that DBN also captures temporal relationships between variables which is the vector for variables at the time point . If there are time points, then the joint probability distribution can be decomposed as , where represents the parent set of . In general, can include variables from the same time point or the previous time points (represented as ). There are many ways to simplify the complexity of the DBN model and data required to train the model. First, we can assume first order Markov property for transitional dependence, then the parent set can be simplified as which corresponds to a general two-slice model (Figure 1A). The intra-slice links represent causal relationships inferred at static states or causal relationships happens in a shorter time than the sampling time between and . We will refer to this model as DBN in our present study. Second, we can further simplify the model and assume (the variables in current time only depend on the previous time point ), then the DBN corresponds to a simplified two-slice model without intra-slice interactions (Figure 1B). Third, if we assume that the variable is self regulated (), then the DBN can be represented as a two-slice model in Figure 1C, which is equivalent to a Granger causality test with a stationary Bivariate Auto-Regressive model (BVAR). We will refer this model as the Granger causality test in our result.
The DBN is a popular approach in computer sciences, such as Kalman filter and Hidden Markov Model (HMM) in voice recognition [15] or more recently in inferring transcriptional regulatory networks from time series data [2] and protein fragmentation process [16]. Another independent line of research of inferring causal relationship from time series is “Granger causality”. The Granger causality concept was originally developed for economic time series data [17], but has since been applied to time series data in many different domains. The Granger causality networks under some assumptions are similar to special cases of the DBN. For example, the model in Figure 1C is a DBN and a Granger causality network with a stationary BVAR model. However, while the Granger causality and the DBN have recently been applied to elucidate temporal causality networks in a number of experimental works, such as SOS DNA repair in E.coli [2], cell cycle in yeast [3], muscle development in Drosophila [4], and cell cycle in human cell lines [5]–[6], no studies to our knowledge have expanded on this concept of temporal causality to gene expression time series data collected in vivo in humans.
One of challenges of applying the Granger causality test to human samples is how to generate long time series data. We overcome the problem by combining multiple short time series. Our simulation results show that data combined from multiple short time series is as informative as a long time series. One of challenges of applying DBN to human samples is limited sample size. We tackled this problem by reconstructing the intra-slice structure from a large data set generated at static states, then reconstructing the inter-slice structure from the time series data.
In the present study we have applied methods based on Granger causality and DBN to a set of human blood gene expression profiles generated at multiple time points during the course of a day, shown in Figure 2. The blood gene expression data was generated from 40 apparently healthy males participating in a randomized, two-arm cross-over design study to assess the effects of fasting and feeding on the blood transcriptional network [18] (see Materials and Methods section for details). The fasted and fed arms of the study provided the necessary data to characterize the dynamic changes in gene expression and corresponding pathways associated with fasting and feeding states in human blood samples [18]. After removing individual scaling effects by referencing individual's time point 0, short time series were combined into virtual long time series (shown in Figure 2). Using the Granger causality test, we identified PER1 as the key regulator of the blood gene expression pattern in which multiple biological processes were under circadian rhythm regulation. Furthermore, the genes under PER1 regulation in the fed network are enriched for obesity causal genes. Finally, using the DBN, we show that over 72% of all inter-slice links are self links and when the SBN and the DBN were compared, we found that different processes such as inflammation and lipid metabolism, which are disconnected during fasting, are now dynamically linked together in response to food intake.
The two-way or three-way ANOVA analysis defining time- and state-dependent gene expression signatures provides meaningful way to characterize expression changes on a global scale [18]. However, these methods on their own do not provide any information on the causal regulators driving the time-dependent gene expression behavior. To leverage the time series data more maximally towards this end, we applied Granger causality test to gene expression traits scored systematically in the fasted/fed cohort blood samples at roughly 1 hour intervals during the course of a day (Figure 2). A gene expression trait is said to be Granger causal for gene expression trait if, at previous time points, provides significantly more information on time-dependent changes in than the historical information provides on itself. In our implementation of the Granger causality test, we test this by fitting to an autoregressive model with respect to the different time points, and then testing whether extending the autoregressive model by including improves the fit (see Materials and Methods for details). If there is a statistically significant improvement testing the model fit (assessed by comparing the models using the F test), then we declare that is Granger causal for , or simply as .
Traditionally, a long time series is required to apply Granger causality test. However, it is hard to obtain a long time series of human samples collected in vivo. We have previously shown that over 80% of transcripts have significant inter-individual variances [18], which is comparable to previously reported result [19]. Thus, we can treat time series data from 40 patients as 40 independent short time series. Assuming these 40 time series have similar dynamic behavior, but with different starting points, we can combine them together to generate a virtual long time series (shown in Figure 2, and see Materials and Methods for details). Our simulation results show that the virtual long time series are as informative as long time series with similar data points (shown as Supplementary Figures S1 and S2). We constructed causal networks for the fasted and fed states by applying the Granger causality test to all gene expression trait pairs generated in the fasting/feeding cohort described in Figure 2. For gene expression traits scored in the fasting/feeding cohort, a link was inserted into the causal network if the p-value associated with the Granger causality test was less than 0.01 after multiple testing correction. The resulting fasted and fed networks were comprised of 2010 and 967 causal links (listed in Supplementary Tables S1 and S2), respectively. The corresponding false discovery rates (FDR) [20] for the causal links in the fasted and fed networks were and , respectively. Bootstrapping test results (see Materials and Methods for details) show that 80% and 90% of links in fast and fed networks have confident values above 0.5, respectively (shown in Supplementary Figure S3). Both networks were observed to exhibit the scale-free property for out-degree distributions (shown as Supplementary Figure S4). From these data it was possible to identify all expression traits supported as Granger causal for at least one other expression trait in the network (referred to here as causal regulators), and then rank order the causal regulators according to the number of genes for which they were supported as causal, shown in Table 1.
There are more causal links inferred for fast time series than for fed time series. The fasted network consists of many small subnetworks and the fed network consists of mainly two subnetworks (shown in Figures 3A and 3B). The top causal gene in the fasted network is RNF144B, a putative ubiquitin-protein ligase that plays a role in mediating p53-dependent apoptosis. Genes under RNF144B regulation including PTEN are enriched for the GO biological process of negative regulation of cellular metabolic process (p-value = 0.008). The top causal gene in the fed network is PER1, a transcription factor regulating the circadian clock, cell growth and apoptosis. The genes under PER1 regulation are enriched for genes correlated to plasma concentration of triglyceride (p-value = 0.00045) in the Icelandic Family Blood (IFB) cohort [10]. PER1's downstream genes are involved in diverse biological processes including CREB5, in circadian rhythm, PTEN and P53INP2 in apoptosis, IL1R1, IL1RAP and TLR2 all involved in inflammation response, FASN and ACSL1 in fatty acid metabolism and MVK in cholesterol biosynthesis. These results suggest that food intake interacts with circadian rhythm and the circadian rhythm has impacts on many biological processes as has been previously shown in mouse studies [21]–[22]. Further, previous research has demonstrated circadian gene (PER1, PER2, PER3 etc.) mRNA expression rhythm in human peripheral blood cells and linked that to individual's circadian phenotype [23]–[24]. Our blood causal network where PER1 is a top causal gene illustrates a potential mechanism of how the CNS control and environmental influences (e.g. external sunlight) can affect circadian rhythm gene expression which in turn regulating a host of other biological functions. More specifically, circadian rhythm genes (PER1 in particular) play important roles in cell cycle regulation and cancer processes [25]–[26]. These reports support our observations in the fed network that several genes under PER1 control are involved in apoptosis and cell cycle regulation (e.g., PTEN and P53INP2).
380 human genes are cataloged as obesity causal genes in the human obesity map [27]. In recent years, many large genome-wide association studies (GWAS) have convincingly identified a number of genes causing human obesity. 34 genes including FTO were replicated in many populations [28]–[31] Taking consideration of these two sources, there are 409 obesity causal genes, and 246 of them were expressed in our blood data set. When the obesity causal genes were overlapped with the fasted and fed networks, 7 genes (ADA, BBS5, CBL, CCND3, FASN, FTO and SCARB1) overlapped with PER1's downstream genes in the fed network (Fisher's Exact Test p-value = 0.037) (shown in Figure 3C). It has been shown that circadian rhythm links to metabolic processes in mouse [32]–[33]. For instance, mutations in mouse genes involving circadian rhythm regulation, such as Clock, can lead to obesity [34]. Our results provide evidence that human obesity causal genes are under circadian rhythm control in a peripheral tissue like blood.
Constructing DBN using the model described in Figure 1A, requires a large amount of data and computational resources. However, when the intra-slice structure (the SBN) is known, then there is a dramatically reduced demand for large amounts of both data and computational resources. A large dataset of profiled peripheral blood samples (IFB) is already described and available [10]. The fasting feeding study group and the IFB cohort are derived from the same population both in terms of geological location and genetic background, therefore the static networks based on these two studies are assumed to be similar. The IFB data set consists of both gene expression measured in the fasting state and genotype data. Previously, we demonstrated that Bayesian networks constructed by integrating gene expression data and genotype data were of high quality [12]–[13],[35]. To match for gender, data from 455 males in the IFB cohort was used to construct a static Bayesian network which consisted of 7310 nodes (genes) and 11047 links (see Method Section for details). The static Bayesian network was fixed as the intra-slice network in the DBN model shown in Figure 1A, and then the time series data (fast or fed) were used to construct inter-slice connections.
The fasted and fed DBNs consisted of 1125 and 1290 inter-slice links (listed in Supplementary Tables S3 and S4), respectively. Among them, 846 (75%) and 936 (73%) were self links. 404 self links are common between the fasted and fed DBNs. The genes under self control (with self links in DBNs) are enriched for cis expression quantitative traits (cis eQTLs) in blood (enrichment p-values = and for the fasted and fed DBNs, respectively).
One important goal for utilizing time series data is to study the dynamic changes in molecular networks. Under static condition, many biological processes may be disconnected or loosely connected, whereas under a perturbation, these processes will change coordinately. 409 obesity causal genes mentioned above were collected from two resources, namely the human obesity map [27] and recent GWAS data [28]–[29],[36]. 138 out of the 409 genes are included in the DBNs. These 138 genes were used as seeds to construct obesity related sub-networks for fast and fed DBN and the SBN as previously described [13]. The fasted and fed subnetworks were compared with the subnetworks constructed from the SBN. The largest change was from the fed subnetwork, where three segmented subnetworks in the SBN were connected in the fed DBN by two inter-slice links (shown as red in Figure 4). CDCA7, a transcription regulator for the cell cycle, is found in the center of the connected subnetworks. It connects genes involved in lipid metabolism such as NPC1, FABP5 and APOE to the large subnetwork on the left which consists of inflammatory response genes such as STAT3, STAT5, GPR109A, TNF, NTSR1, ORM1 and IL1RN. This suggests that the expression of genes involved in either inflammatory response or lipid metabolism change coordinately in response to food intake. It is also worth noting that the circadian rhythm regulator PER1 is in the subnetwork on the left, which consists of many genes involved in inflammatory response pathways. As well, in the fed DBN, both cell cycle regulation and lipid metabolism processes are linked to the circadian rhythm.
Designing experiments to generate large-scale molecular phenotyping data and to enable inferring causal relationships among genes and between genes and clinical endpoints is now a feasible task. Genetic variants (e.g. nonsynonimous, nonsense, eSNPs etc), genetically modified animals (e.g., knockouts, transgenics, RNAi knockdown), and chemical perturbations have all been used successfully to establish a causal relationship between genes and phenotypes in mammalian systems. Here we have detailed the use of time series data in a human population to predict causal regulators using a Granger causality test and a DBN. Our Granger causality networks showed that multiple biological processes such as apoptosis, inflammation response and lipid metabolism are under circadian rhythm regulation and obesity causal genes are under circadian rhythm regulator PER1 in the fed networks. For the DBN, we showed that over 73% of inter-slice links are self links. When the SBN and the DBN were compared, we find that different processes such as inflammation and lipid metabolism are linked together during the dynamic changes in responding to food intake.
The time series data provided a path to go beyond the characterization of interesting patterns of expression and network differences associated with complex states (like fasting and feed status), by allowing for the identification of putative causal regulators driving these differences. While extensive experimental validation will be required to assess the full utility of the approach detailed in the present study, we believe these methods and the characterizations of time and state dependent changes in gene expression and network topology, will motivate a need to integrate a time domain into gene expression experiments that aim to elucidate complex system behavior.
Our data consist of many short time series from multiple individuals instead of a single long time series. Our approach for combining multiple short time series was based on the assumption that individual response slopes are similar. First, the population under study is relatively homogeneous, i.e. only males, similar age, same population, same ethnicity and each individual consumed the meal of same size and composition. Second, we reduced the individual specific variance by normalizing each individual data according to its own expression data at the first time point. This essentially reduces the number of parameters to fit in the model, at the cost of reducing the number of time points available to feed into the model. In contrast, if the population under study was genetically heterogeneous, we would treat the response slope differently for different individuals and would employ the mixed-effects model as suggested by Berhane and Thomas [37] for combining time series. In that case, we wouldn't need to normalize data for each individual, and as a result there would be an increase in the number of parameters to fit as well as an increase in the available data points. We note in passing, that the Icelandic population is relatively homogenous as regards genetic makeup and environmental parameters.
Our implementation of the Granger causality test is a special form of DBN where there is no causal structure within a single time slice. There are also many variations of the Granger causality test including stationary or non-stationary, dynamic or time-invariant Granger causality tests. Our simple implementation of Granger causality test identified the transcription factor PER1 as the main causal regulator in the fed time series.
The intra-slice network (SBN) was reconstructed from an independent data set and is fixed in our current model of DBN. Even though the SBN was reconstructed using about 455 samples, there are still many uncertainties about the network structure and edge directions. Further researches on using the SBN as flexible priors for intra-slice structure rather than fixed one are warranted.
Several simulation studies have been carried out to estimate the number of samples that are required to build SBNs or DBNs. Zhu et al. [12] showed that these numbers are related to the interaction strength between nodes. For instance, with networks consisting mainly of interactions at intermediate strength, over 80% of interactions in SBN can be recovered at 90% precision with 1000 samples. Similarly, Yu et al. [38] showed that over 85% of links in DBNs can be recovered with 2000 samples. In addition, Yu et al. showed that the sampling interval is also an important parameter. When the sampling interval is small, the difference between data at consecutive time points will be small. In other words, the independent information added is small. Our time series simulation result (Supplementary Figure 2) and the results of Yu et al., both show that network reconstruction accuracies drop when sampling intervals are large. In both our and Yu et al.'s time series simulations, all interactions have the same time lag. In reality, the time lags are different for different transcriptional regulations [39]. Zou and Conzen [3] showed that a better reconstruction accuracy of DBN could be achieved when considering time lag differences. The general DBN model shown in Figure 1a can represent mixed time lags with intra-slice interactions for zero or short time lags and inter-slice interactions for large time lags. Based on the complications discussed above, at least 1000 data points are needed to reconstruct an adequate DBN. Sachs et al. [40] suggests that even over 23,000 data points are not sufficient for reconstructing an accurate DBN. Obviously, additional priors can improve reconstruction accuracies with the same amount of data [3],[12]. To accurately estimate the amount of data required to reconstruct DBNs under different interaction strengths using different mixtures of time lags and different priors, a systematic data simulation is warranted.
The causal networks derived from either the Granger causality test or the Dynamic Bayesian network, both showed that the networks under the fasting state were fragmented (loosely connected) while the networks in the feeding state are more highly interconnected. It is well established, that the circadian rhythm interacts with metabolic [32] and immune response processes in rodents [41]. For instance TNF-alpha, which regulates immune cells and induces apoptotic cell death, is also shown to regulate key genes in the circadian rhythm, including Dbp and Per1-3 [41]. It is possible that increasing nutritional load directly affects the circadian rhythm system, possibly through ghrelin [42]. Our results in humans are consistent with the rodent data, showing that feeding is directly linked to the circadian rhythm system. Furthermore, our results suggest that the interconnections between different biological processes such as metabolic and immune responses and activated cell death are weak in the fasted state, while feeding dramatically enhances the interconnections between these different biological processes. Further experimental work is warranted to verify whether these changes still hold in the general population.
Human peripheral blood is the most readily accessible human tissue for clinical studies. Our work on peripheral blood has demonstrated that feeding or increasing nutritional load affects the human circadian rhythm system, which becomes highly connected to other biological processes including metabolic and immune responses. And these effects can be observed in peripheral blood. We believe the results of the present work have broader implications for studies of drug response and for genetic and experimental studies on blood chemistry and vascular related clinical traits. Our results suggest that how blood networks respond to feeding is an important variable that may bring us closer to dissecting the underlying causes of obesity and associated disorders. Our results also provide a guideline on how much data are required for inferring causal relationship in human blood for future experiments.
40 healthy participants from an Icelandic company were recruited to participate in a randomized, two-arm, cross-over study to examine the effects of fasting and feeding on human blood gene expression [18], shown in Figure 2. For the first period of the study the 40 participants were randomized to two treatment groups, with 20 individuals making up each group. All participants began fasting at 9pm the night before the first period of the study. The first treatment group comprised the fasted arm of the study for the first period, where participants continued to fast through the day for the duration of the study (participants were only allowed to drink water during this time). The second treatment group comprised the fed arm of the study for the first period, where participants were fed a standard meal in the morning and then fasted through the rest of the day for the duration of the study. The second period of the study was carried out one week later from the start of the first period. The protocol for the second period of the study was identical to the first period, except those in the fasted arm for the first period were switched to the fed arm, and those in the fed arm for the first period were switched to the fasted arm. Figure 2 shows the schematic for the experimental design.
A total of 560 peripheral blood samples were collected from the 40 participants at 7 time points for each period of the study. Significant inter-individual variation has been noted in human blood gene expression profiles [43]. Previous analyses carried out on this data set detailed the inter-individual variation and overall expression differences between the fasted and fed conditions [18]. In the present study we focus mainly on using temporal information to infer causal relationship by applying a Granger causality test and a dynamic Bayesian network so that possible causal drivers of dynamic changes can be identified from the causal networks. To correct for the individual differences in gene expression we referenced each individual expression profile to the corresponding individual profile at time point 0. This reduced the effective number of time points to 6 for this study.
The time series based causality test was proposed by Wiener [44] as the notion that, if the prediction of one time series could be improved by incorporating the knowledge if a second one, then the second series has a causal influence on the first. Granger was the first to formalize the idea in the context of linear regression model [17], so that time series based causality test is generally referred as Granger causality test. There is a variety of models for testing Granger causality, such as multivariate autoregressive model (MVAR) and bivariate autoregressive model (BVAR). If coefficients in the regression model do not change depending on time, the model is referred as a stationary model. Otherwise it is referred a non-stationary model. The simplest model is stationary bivector autoregressive model. Even though comparing to MVAR, BVAR tends to infer many indirected links, the causal directions of these inferred links follow causal information flows [45]. To remove potential in-direct links, for each gene, we only keep one causal link pointing to it, which has the most significant p-value in the BVAR model.
Traditionally, Granger causality test is applied to long time series. However, it is hard to collect long time course data from human samples. Our data consists of many short time series from multiple individuals. There are several theoretical studies related to combining multiple time series in a general regression frame work, including for instance that of Berhane & Thomas [37] and Guerrero & Pena [46]. Berhane & Thomas [37] proposed to use a mixed-effects model to combine time series from different locations, while Guerrero & Pena [46] outlined a weighted least squares approach. In both approaches, some constraints were applied after a number of assumptions were made.
Our approach is a simplified version of the Berhane & Thomas approach [37]. Instead of using community-specific slopes, we assumed response slopes for individuals are similar. Further, in order to reduce individual specific variation which could affect the response slope, an individual's gene expression data were normalized according to its own expression data at the first time point. Our simulation study shows that causal relationship can be accurately inferred by combining these short time series.
Under first order stationary BVAR model, a set of data was simulated for causal relationship as following:(1)There are independent time series of length , , . All coefficients and noises follow normal distributions as(2)The initial conditions are draw from an uniform distribution with mean 0. 1000 independent time series were simulated, and each series consists of 240 time point (shown as Supplementary Figure 1).
The test of Granger causality under BVAR model can be carried out by comparing the full model(3)with the autoregressive model(4)The significance of the Granger causality test (full model explains more variance than the autoregressive model) is then measured by F-test statistics(5)where and are sum of squared residuals of full model and autoregressive model, respectively; and is the length of the time series.
For the 1000 time series simulated above, the p-values of Granger causality are estimated as Eq. 5. If only partial time points are used, then the power to detect Granger causality decreases (shown in Supplementary Figure 2). It is worth to note when the same number of time points are used, it is more likely to inferred correct causality if the interval between time points is shorter.
If only 6 time points are used, no Granger causality test is significant if considering the time series independently. If assuming and are similar, then these short series can be combined together to infer Granger causality, and the Eq. 3 can be modified as(6)where and are sum of squared residuals of full model and autoregression model, respectively. For example, a virtual time series by combining the first 6 time points of randomly selected 40 time series is as informative as a long time series with the same time points.
To estimate the false positive rate, we permuted the assignment of 1000 time series generated above (for example was assigned as where ) so that the autoregressive assumption was valid. For each permutated data set, we followed the same procedure mentioned above to calculate p-values for the Granger causality test. At different p-value cutoffs, we calculated the recall (positive rate) and the false positive rate (shown in the Supplementary Figure 2).
It is of note that choosing the optimal time lag length in the autoregressive (AR) model normally requires comparing model residuals and statistics at different p-value thresholds. However, because of the small sample size (40) and limited number of time points (6), we restricted our analyses here using AR models with only first order time dependency, similar to what has been done in previous studies [5]–[6]. Similarly, we assumed the Granger causal relations were stationary from time point 1 to 6. That is, we were mainly interested in the mean α and β values in Eq.(1), which represent the averaged Granger causality between genes from time point 1 to 6.
A bootstrapping procedure of re-sampling individuals with replacement, was used. At each time, one subject (along the associated data at 6 time points) was sampled from a pool of 40 individuals. A bootstrapped data set consisted of 40 sampled individuals (40×6 data points). The same Granger causality test outlined above was applied to the re-sampled data. The bootstrapping procedure was performed 100 times. The link confident value is the percentage of a link's p-values above a multiple testing corrected threshold in the results of the 100 bootstrapping tests.
455 male samples in IFB cohort [10] was used in reconstruction of the static Bayesian network. A set of informative genes were identified as follows: (1) a gene expressed in the blood (with mean log intensity >−1.5), (2) the variation of the mean log ratio was larger than 1.23. Of the 23720 genes represented on the microarray, 7310 were selected for inclusion in the network reconstruction process as previously described [12],[35]. One thousand Bayesian networks were reconstructed using different random seeds to start the reconstruction process. From the resulting set of 1000 networks generated by this process, edges that appeared in greater than 30% of the networks were used to define a consensus network.
For a two-slice dynamic Bayesian network represented in Figure 1A, it can be decomposed as , where is the parent set of . The static Bayesian network reconstructed above was used as the intra-slice network. The intra-slice network is fixed and is not refined in the process of reconstructing dynamic Bayesian networks. Thus, only inter-slice links () are added or removed during the reconstruction process. Similar to the static Bayesian network reconstruction process, 1000 networks were reconstructed using different seeds and the Bayesian information criterion (BIC) score [47] was used for the optimization. Edges appeared in 30% of the 1000 structures are included in the final network.
|
10.1371/journal.pgen.1000680 | p63 and p73 Transcriptionally Regulate Genes Involved in DNA Repair | The p53 family activates many of the same genes in response to DNA damage. Because p63 and p73 have structural differences from p53 and play distinct biological functions in development and metastasis, it is likely that they activate a unique transcriptional network. Therefore, we performed a genome-wide analysis using cells lacking the p53 family members after treatment with DNA damage. We identified over 100 genes involved in multiple pathways that were uniquely regulated by p63 or p73, and not p53. Further validation indicated that BRCA2, Rad51, and mre11 are direct transcriptional targets of p63 and p73. Additionally, cells deficient for p63 and p73 are impaired in DNA repair and p63+/−;p73+/− mice develop mammary tumors suggesting a novel mechanism whereby p63 and p73 suppress tumorigenesis.
| p63 and p73 have been identified as important suppressors of tumorigenesis and metastasis. Although they are structurally similar to p53, they have many functions that are unique including roles in development and metastasis. Here we show, using a genome-wide analysis of cells lacking p63 and p73 individually and in combination, that p63 and p73 regulate many unique target genes involved in multiple cellular processes. Interestingly, one of these pathways is DNA repair. Further validation of differentially expressed target genes in this pathway, revealed that p63 and p73 transcriptionally regulate BRCA2, Rad51, and mre11 providing a novel mechanism for the action of p63 and p73 in tumor suppression. These findings have important therapeutic implications for cancer patients with alterations in the p63/p73 pathway.
| p53 acts as a tumor suppressor gene by transcriptionally regulating a multitude of target genes in response to DNA damage [1]. Induction of these genes results in multiple cellular fates including apoptosis and cell cycle arrest. p63 and p73 share some of the same functions as p53; however, p63 and p73 are structurally more complex containing multiple isoforms [2],[3],[4]. The TA isoforms are structurally more like p53 and contain a transactivation domain while the ΔN isoforms lack this domain and are transcribed from an internal promoter unique to these isoforms [3],[5]. Based on the fact that the TA isoforms are more similar structurally to p53, the TA isoforms were hypothesized and shown to be the major isoforms that induce transcription and are thought to have tumor suppressive functions [3],[5],[6],[7],[8]. In contrast, the ΔN isoforms have been shown to act as dominant negatives against the TA isoforms of p63 and p73 and also against p53. Because of the ability of the ΔN isoforms to act as dominant negatives and their overexpression patterns in human tumors [2],[6],[9],[10],[11], these isoforms have been hypothesized to act as oncogenes [3],[5],[6]. Interestingly, recent data have revealed that the ΔN isoforms of p73 can induce apoptosis, cell cycle arrest and transactivate target genes, such as p21, 14-3-3σ, and GADD45 [12]. Additional studies have demonstrated that expression of ΔNp73β at physiological levels can result in the suppression of cell growth in the presence or absence of p53 indicating that this isoform of p73 may act as a tumor suppressor gene [12]. Similarly, the ΔN isoforms of p63 have also been shown to have the ability to transactivate target genes [13]. In the case of p63, the ΔN isoforms are more highly expressed in epithelial tissues [14], and thus it is not be surprising that the ΔN isoforms transcriptionally regulate genes involved in the morphogenesis and differentiation of the epithelium. Given the structural complexity and expression of p63, p73, and their isoforms, the transcriptional targets of these genes are an area of growing research.
We and others have shown previously that p63 and p73 can induce apoptosis in response to DNA damage [2],[8],[15]. Many of the target genes induced by p63 and p73 are shared with p53 [2],[8]. Additionally, we have shown that the p53 family of genes is interdependent on each other in the apoptotic response and in the suppression of tumorigenesis. p53+/−;p63+/− and p53+/−;p73+/− develop some of the same tumor types as p53+/− mice, but the phenotype of the tumors in the compound mutant mice is highly aggressive and metastatic indicative of cooperativity between family members [7],[15]. Mice heterozygous for combinations of the p53 family members develop a novel tumor spectrum compared to p53+/− mice indicative of functions of p63 and p73 independent of p53 [7]. These independent functions suggest that p63 and p73 may have unique transcriptional programs.
To understand the transcriptional program of p63 and p73, we made use of MEFs deficient for each of the p53 family members individually and in combination and performed a genome wide analysis using cDNA microarray analysis to determine whether p63 and p73 transcriptionally regulate genes independently of p53 in response to DNA damage. Interestingly, we found that p63 and/or p73 transactivate sets of genes independent of p53. Among these sets of genes are those involved in homologous DNA repair, including Rad51, BRCA2, mre11 and Rad50. p63 and p73 were found to bind to these gene promoters by ChIP assay and to transactivate them as demonstrated by luciferase assay. Surprisingly, the ΔN isoforms of p63 and p73, which have been shown to be weak transactivators, transactivate the Rad51 and BRCA2 genes to high levels. In addition, p63−/−, p73−/− and p63−/−;p73−/− MEFs exhibited an impaired ability to repair their DNA and to survive in a clonogenic survival assay. Additionally, in vivo evidence from p63/p73 mutant mice supports this finding; p63+/−;p73+/− mice develop mammary adenocarcinomas at a high frequency [7]. Here, we show that these mammary tumors lose expression of p63, p73, BRCA2, and Rad51. Our findings indicate that p63 and p73 may suppress tumorigenesis by transcriptionally regulating critical genes in the DNA repair pathway.
The p53 family members, p63 and p73, have previously been shown to share many of the same target genes as p53 [2],[8]. Additionally, both p63 and p73 have the ability to bind to the p53 consensus binding site. p63 and p73 also have biological activities independent of p53. Consequently, we were interested in determining whether p63 and p73 had unique transcriptional target genes. A cDNA microarray analysis was performed using E1A expressing MEFs deficient for each p53 family member individually (p53−/−, p63−/−, p73−/−) and in combination (p63−/−;p73−/−). These cells were treated with doxorubicin, a DNA damaging agent, to induce apoptosis in wild-type E1A MEFs. p53−/− and p63−/−;p73−/− E1A MEFs have previously been shown to be resistant to this treatment while the p63−/− and p73−/− E1A MEFS are partially resistant to apoptosis [15]. Microarray analysis revealed a large number of genes differentially expressed in the MEFs deficient for each p53 family member. Because we were interested in identifying genes that are transactivated by the p53 family members in response to DNA damage, genes that are down regulated in the absence of the p53 family members were further analyzed.
After filtering and statistical analysis using SAM [16], 620 out of 15,488 genes were found to be down regulated in at least one of the single knockout E1A MEF lines compared to wild-type E1A MEFs in response to DNA damage. Eight-six of the 620 genes were down regulated in the p53−/−, p63−/−, and p73−/− E1A MEFs as illustrated by the Venn diagram (Figure 1 and Figure S1). There were also sets of genes that were uniquely regulated by each p53 family member; the p53−/−, p63−/−, and p73−/− MEFs each had 109, 148, and 131 genes down regulated respectively. Lastly, there were sets of genes that were regulated by two family members only; forty-seven were down regulated in the absence of p53 and p63, 41 in the absence of p53 and p73, and 58 in the absence of p63 and p73. The final list of differentially regulated genes was processed through multiple bioinformatic pipelines to identify biological pathways regulated by the p53 family members. Pathway analysis using the web-based KEGG, BioCarta, and GenMAPP databases indicated that the p53 family members regulate numerous pathways including: cell cycle, DNA-damage, p53 signaling, apoptosis, ribosomal proteins, metabolic pathways, and growth factor signaling (Table S1).
The putative target genes identified by microarray analysis were analyzed for the presence of a p53 or p63 consensus binding sites using a computer based genome wide search and HMMER1 software [17]. The promoter sequences (defined as 5 kb upstream and downstream of the transcription start site excluding exons) from the 724 down regulated genes were queried, and 700 of these genes were found to have p53 family member motifs. Of these, 669 genes contained p53 family member motif sites with the ideal p53 spacer of 6 nucleotides between the two half sites. Scores were then given to each identified binding site corresponding to how well they matched with previously published p53 or p63 matrices (Table S2).
Hierarchical clustering was then performed in the specified knock out MEFs relative to wild-type after DNA damage to highlight patterns between the downregulated genes in the p53−/−, p63−/− and p73−/− E1A MEFs (Figure 2 and Figure S1). Interestingly, many genes were differentially regulated in the various MEF lines (Figure 2) indicating that p63 and p73 have unique target genes. Also, many genes were found to be down-regulated in all mutant cell types supporting the hypothesis that all three transcription factors can transactivate some of the same gene targets (Figure 2).
DNA damage triggers numerous cellular responses including an extensive DNA repair pathway involving numerous genes [18]. Microarray analysis revealed that the p53 family members regulate numerous genes involved in the DNA repair pathway. Many of these genes seemed to be uniquely regulated by p63 and/or p73. After DNA damage, loss of p63 or p73 prevents induction of Brca2 (Figure 2, cluster 4), an essential co-factor in Rad51-dependent DNA repair of double-stranded breaks, and Rad51 itself (Figure 2, cluster 3) [18]. Sequence analysis also indicates that p53/p63 response elements exist in both the promoter and intronic region of Brca2 and Rad51 (Table 1 and Table S2). Clustered with Rad51 are Dbf4, a regulator of Cdc7 and a prognostic determinant for melanoma development, and Gas6, which cooperates with the tyrosine receptor kinase Axl in tumor proliferation and cell survival (Figure 2, cluster 3). We also found additional genes that were uniquely down-regulated in p63−/−, p73−/−, and p63−/−;p73−/− MEFs. These hits indicate that p63 and p73 have roles independent of p53 in the DNA-damage response pathway. For example, expression of Rad50, which forms a complex with mre11 and Nebrin, is found to be down regulated in p63−/−;p73−/− MEFs relative to wild-type MEFs treated with doxorubicin. There are also p53 family response elements upstream of the transcription start site of Rad50 (Table S2).
In addition to genes that are uniquely regulated by p63 and/or p73, genes controlled by all three p53 family members were identified. Mre11, a gene that functions in the repair of DNA double strand breaks, was found to be down-regulated in p53, p63, and p73 deficient E1A expressing MEFs (Figure 2, cluster 1). In addition, sequence analysis revealed multiple p53/p63 response elements (Table S2). Genes with similar expression profiles as mre11 include the growth factor signaling components Ghr and Sos1 as well as the apoptotic components Traf1 and Cathepsin D all of which contain p53 family member binding sites (Figure 2, cluster 1 & 6 & Table S2).
Multiple genes involved in other biological processes, including tumor progression, metastasis and development were found to be differentially regulated in the various E1A MEF cells. For example, Mmp2, a gene shown to play a role in embryonic development and tumor metastasis, is also down regulated in the absence of p73 after doxorubicin treatment. Clustered with Mmp2 are many signaling components such as Grb2, Stat1, Map3k14, and Mapk8ip3- all of which have at least one p53 family member binding motif present near its promoter (Figure 2, cluster 5 and Table S2). Interestingly, brachyury, the developmental transcription factor, was identified as a putative p63 target gene (Figure 2, cluster 2). Given the identified roles of brachyury in limb development, cancer, and hematopoetic stem cells and the development phenotype of the p63−/− mouse, this putative target has important biological significance [19],[20],[21],[22]. We found brachyury to contain multiple p53 family response elements both upstream of its transcriptional start site and within the first intron (Table S2). Other p63 dependent genes that cluster with brachyury include Abr, the GAP for the small GTPase Rac, Socs3, involved in cytokine and apoptotic signaling, and the zinc-finger transcription factor Klf9 which is implicated in control of cell proliferation, cell differentiation, and cell fate (Figure 2, cluster 2).
Strikingly, the results from the cDNA microarray indicate that genes in the DNA repair pathway are differentially regulated in MEFs lacking p63 and/or p73 after treatment with DNA damaging agents. To verify these putative transcriptional targets of p63 and p73, quantitative real time PCR was performed. The expression of mre11, BRCA2, Rad51, and Rad50 was examined in wild-type, p53−/−, p63−/−, p73−/− and p63−/−;p73−/− E1A MEFs before and after treatment with doxorubicin for 12 hours and 5 Gy of gamma radiation. Interestingly, mre11, BRCA2, Rad51, and Rad50 are all induced in wild-type E1A MEFs after these treatments (Figure 3). We measured the baseline levels of mRNA of mre11, BRCA2, Rad51, and Rad50 to determine levels of these transcripts prior to DNA damage (Figure S2). After treatment with doxorubicin or gamma radiation, levels of mre11 mRNA are not induced to wild-type levels in p63−/−and p63−/−;p73−/− E1A indicating that p63 may transcriptionally regulate this gene (Figure 3). Similarly, the levels of BRCA2 are significantly lower in p73−/− and p63−/−;p73−/− E1A MEFs than in wild-type or p53−/− E1A MEFs (Figure 3) after treatment with doxorubicin and gamma radiation. Likewise, the Rad51 gene is not induced to wild-type levels in p63−/−, p73−/−, and p63−/−;p73−/− E1A MEFs after treatment with DNA damaging agents (Figure 3), indicating again that p63 and p73 may be critical transcriptional activators of Rad51 after DNA damage. Lastly, Rad50 also showed a pattern indicative of transcriptional regulation by both p63 and p73. The mRNA levels of Rad50 are approximately 4-fold lower in p63−/−;p73−/− E1A MEFs than in wild-type E1A MEFs (Figure 3) after treatment with doxorubicin and gamma radiation. Taken together, these data indicate that mre11, BRCA2, Rad51, and Rad50 may be transcriptional targets of p63 and p73 in response to DNA damage.
As previously reported, twenty percent of mice heterozygous for p63 and p73 (p63+/−;p73+/−) develop mammary adenocarcinomas [7] (Figure 4), and ninety percent of these tumors lose the wild-type allele of p63 and p73 [7]. Given that BRCA2 plays an important role in the pathogenesis of mammary adenocarcinoma, this made it a relevant biological target for p63 and p73 in mammary tumors. The protein levels of Rad51 was first examined by Western blot analysis using wild-type and p63−/−;p73−/− MEFs. Interestingly, the basal level of Rad51 is lower in p63−/−;p73−/− MEFs compared to wild-type MEFs (Figure 4A). The levels of Rad51 in p63−/−;p73−/− MEFs are not induced in response to gamma irradiation; however, a 2-fold increase in expression of Rad51 was detected in the wild-type MEFs after DNA damage. To determine whether this change in expression pattern of Rad51 was cell-type specific, we performed immunohistochemistry on mammary adenocarcinomas from p63+/−;p73+/− mice where LOH of p63 and p73 had occurred (n = 10) (Figure 4F–4I). Indeed, Rad51 as well as BRCA expression is detected in normal mammary glands (n = 10) of p63+/−;p73+/− mice (Figure 4B and 4D) and is lost in hyperplastic mammary glands (n = 4) and mammary adenocarcinomas (n = 6) in these mice (Figure 4C and 4E).
Both the cDNA microarray and real-time RT-PCR data provide evidence that BRCA2, Rad51, and mre11 are transcriptionally regulated by p63 and p73 after DNA damage (Figure 3). Consequently, chromatin immunoprecipitation (ChIP) assay was performed to determine whether p63 and/or p73 could directly bind to the promoter region of these two genes. A subset of putative binding sites identified and summarized in Table 1 were assayed using ChIP. Sites chosen included those with the best scores for p53 and p63. Four putative binding sites were assayed for RAD51 (Table 1). RAD51-1 and 2 are located in intron 1, upstream of the start site, while RAD51-3 and 4 are found in intron 2, downstream of the start site. One putative element was assayed for BRCA2 in intron 2, 133 nucleotides downstream of the start site (Table 1). Lastly, three putative p53 family response elements were queried for mre11: MRE11-1, 2, and 3, located in intron 1, upstream of the start site (Table 1).
ChIP analysis was performed using an antibody for p53, p63 or p73 in wild-type, p53−/−, p63−/−, and p73−/− E1A MEFs treated with doxorubicin for 12 hours (Figure 5). Interestingly, p73 was the only p53 family member that binds to the RAD51 promoter after DNA damage treatment. p73 was found to bind to RAD51-2 and 3 in intron 1 and intron 2 respectively. The primers used for this PCR reaction did not distinguish between the two sites; therefore, it is possible that p73 only binds to one of these sites. p63 and p73, but not p53, were found to bind to the response element in BRCA2 after DNA damage (Figure 5). Lastly, p63 was the only family member found bound to the mre-11 promoter at site mre11-3 within intron 1, 171 nucleotides upstream of the start site. The same binding pattern in the ChIP assay was obtained with other DNA damaging agents, such as gamma radiation (data not shown).
The ChIP results clearly demonstrate that p63 and/or p73 can bind to the promoters of these genes; however to gain a clear indication of which isoforms of p63 and p73 transactivate Rad51, BRCA2, and mre11, luciferase assays were performed with TA and ΔN isoforms of p63 and p73. Regions shown to bind by ChIP assay were used to construct firefly luciferase reporters. pGL3-Rad51-1 was designed by cloning intron 1 containing RAD-51-1 and 2 (Table 1) in to the pGL3 basic vector and pGL3-Rad51-2 containing the elements, RAD51-3 and 4, was cloned in to the pGL3 basic vector. These constructs were transfected in to p63−/−;p73−/− MEFs along with a renilla luciferase gene and one of the following isoforms of p63 or p73: TAp63α, TAp63γ, TAp73α, TAp73β, ΔNp63γ, ΔNp73α, and ΔNp73β. Interestingly, both ΔNp63α and ΔNp73β are the isoforms that transactivate the Rad51 reporter gene to appreciable levels. ΔNp63α transactivates pGL3-Rad51-1 11 fold and ΔNp73β transactivates this reporter 6 fold (Figure 6A). These isoforms more modestly transactivate the pGL3-Rad51-2 reporter indicating that the p63/p73 element resides in intron 1 (Figure 6A and 6B). Surprisingly, the TA isoforms did not transactivate the reporter gene. The p63/p73 family members also transactivate this reporter gene. ΔNp63α and ΔNp73β together can transactivate the Rad51-1 reporter 19 fold (Figure 6A and 6B). Additionally, the other ΔN isoforms that modestly transactivate this reporter alone can transactivate this reporter to higher levels. For example, ΔNp63α along with ΔNp73α can transactivate this reporter gene 9.8 fold, demonstrating additive effects between these family members.
Similar to the experiments for RAD51, the BRCA2 region within intron 1 found to be bound by both p63 and p73 was cloned in to the pGL3 basic vector. Dual-luciferase reporter assay was performed in p63−/−;p73−/− MEFs as described above. Strikingly, the isoform with the highest ability to transactivate this reporter was ΔNp73β with a 4 fold induction. Additionally, ΔNp63α and ΔNp73β can transactivate the reporter 6 fold and other combinations of ΔN isoforms also show increases in transactivation of this reporter (Figure 6C).
The ability of p63 and p73 to transactivate the mre11 gene was also tested by luciferase assay. The region shown to bind to p63 by ChIP analysis was cloned in to the pGL3 basic vector to generate pGL3-Mre11. This reporter was induced 3.8 fold by ΔNp63α and ΔNp73β together (Figure 6D). pPERP-luc, which has previously been shown to be responsive to TAp63γ was used as a positive control for these experiments [23],[24].
To determine whether p53 could transactivate these reporters, p53 was transfected with each reporter and luciferase activity was measured. p53 did not induce any of the reporters assayed (Figure 6A–6D). In addition, we performed luciferase assays using the Rad51-1 and BRCA2 reporters in MEFs lacking p53, p53−/−;p73−/− (Figure 6E and 6F) and p53−/−;p63−/− (data not shown). These experiments yielded similar results as those shown in Figure 6A and 6C. Taken together, these data indicate that the trasactivation of Rad51, BRCA2, and mre11 is p53-independent.
Rad51 and BRCA2 are both involved in homologous recombination (HR) DNA repair, one of the major pathways for repair of double strand breaks (DSBs). Cells lacking genes involved in HR, like BRCA2 and Rad51, have been shown to have an impaired ability to repair their DNA [18],[25],[26],[27]. Consequently, we hypothesized that cells lacking p63 and/or p73, which have low levels of these two proteins, may have a defect in repairing DSBs in damaged DNA. To test this hypothesis, wild-type, p53−/−, p63−/−, p73−/−, and p63−/−;p73−/− primary and E1A MEFs were treated with 5 Gy gamma-radiation or doxorubicin to generate DSBs. A comet assay was then performed to determine the DSB repair capacity in these cells. Comet assay, or single cell gel electrophoresis, is a commonly applied approach for detecting DNA damage in a single cell. The unwound, relaxed DNA migrates out of the cell during electrophoresis and forms a “tail” [28]. Therefore, cells that have damaged DNA appear as comets with tails containing fragmented and nicked DNA, while normal cells do not. The degree of DNA damage is represented using the parameter known as tail moment defined as the product of the tail length and the portion of total DNA in the tail. MEFs lacking the p53 family members were treated with DNA damage and incubated for a total of 16 hours allowing the homologous recombination repair to take place. Cells were and harvested at 0 (untreated), 1, and 16 hours for the Comet assay. In all cases, p63−/−, p73−/−, and p63−/−;p73−/− MEFs were found to have the largest tail moment after DNA damage (Figure 7A–7D). The tail moment after DNA damage was significantly higher for p63−/−, p73−/−, and p63−/−;p73−/− primary and E1A MEFs (18.8) compared wild-type samples (p <0.0001). This result indicates that p63 and p73 play a critical role in DNA repair.
Because loss of p63 and p73 impair DSB repair by regulating Rad51, BRCA2, and mre11, it is likely that loss of p63 and p73 results in poor cell survival due to the inability to repair damaged chromosomal DNA. To determine whether loss of p63 and p73 results in a decrease in cell survival, a clonogenic survival assay was performed using both primary MEFs and E1A expressing MEFs after treatment with 1, 2 and 3 Gy of gamma radiation and 0.34, 0.5, and 1.0 µM doxorubicin. After 12 hours, cells were replated and assayed for the ability to form colonies. p63−/−;p73−/− E1A MEFs and primary MEFs have an impaired ability to form colonies after gamma radiation indicative of defects in DNA repair (Figure 7E and 7F). A similar result was seen after treatment with doxorubicin in these cells (Figure 7G and 7H).
p53 transactivates a vast network of genes in response to DNA damage [1]. While p63 and p73 can also transactivate known p53 target genes to varying degrees, they play roles in distinct biological functions including development and metastasis and likely have unique transcriptional targets. The advantage of the system employed here is the use of isogenic primary cells with the deletion of a single p53 family member. Here, we used early passage MEFs lacking the p53 family members individually or both p63 and p73 in combination and expressing E1A, which sensitizes them to undergo apoptosis after DNA damage to identify changes in gene expression in this process. We identified sets of genes that are regulated by individual and multiple p53 family members indicating unique and overlapping functions for this family of genes in response to DNA damage. Six hundred twenty out of 15,488 genes queried were regulated by a p53 family member. Genes identified played a role in multiple processes including apoptosis and DNA repair. In addition to engaging pathways predicted to be induced by DNA damage, genes involved in other processes like development and metastasis were also induced. These are biologically significant given the reported developmental, tumor, and metastatic phenotypes of the p63/p73 mutant mice [7],[20],[22],[29]. Lastly, the majority of the targets identified had binding sites that closely fit the p53 and p63 consensus binding site [14],[30],[31] indicating that they may be bona fide direct transcriptional targets of these family members. Indeed, we verified that Rad51, BRCA2, and mre11, genes involved in DNA repair, are direct transcriptional targets of p63 and p73.
Given the high prevalence of mammary adenocarcinoma in mice mutant for p63 and p73 (p63+/−;p73+/−), a group of genes of interest are those involved in DNA repair. These genes were induced in wild-type cells and down regulated in the absence of p63 or p73. The mechanism for the tumor suppressive activity of p63 and p73 is not completely understood [6],[7],[32]. Regulation of DNA repair genes by p63 and p73 has not been demonstrated previously and could be a pathway employed by these genes in tumor suppression. Both Rad51 and BRCA2 were found to be direct transcriptional targets of p63 and p73 indicating that these mechanisms may be triggered during tumorigenesis. Interestingly, Rad51 has been shown previously to be repressed by p53 through a site found upstream of the start site [33]. Here, we show that ΔNp63 and ΔNp73 transactivate Rad51 through a distinct element in intron 1 indicating that there is an intricate and complex regulation of this gene by the p53 family and is likely a critical target in tumor suppression by this family. We also showed that transcriptional regulation of Rad51, BRCA2, and Rad51 by p63 and p73 is p53-independent/
It was surprising that the ΔN isoforms of p63 and p73 were more potent transactivators of Rad51, BRCA2, and mre11 than the TA isoforms. The TA isoforms have an acidic N-terminal domain necessary for transactivation [2],[3], and many studies have shown previously that the TA isoforms are more potent transactivators than the ΔN isoforms [2],[8]. Furthermore, the ΔN isoforms are better known for the dominant negative activities that they impose on the TA isoforms of p63 and p73 and p53. Interestingly, a number of recent studies have shown that the ΔN isoforms are capable of transactivating target genes due to a proline-rich transactivation domain that exists in these isoforms [12],[13]. In addition, the ΔN isoforms of p63 are more highly expressed than TAp63 in certain tissues including the skin [14] making the ΔNp63 isoforms likely candidates for gene regulation in these tissues. Taken together, our results indicate that the roles of the ΔN isoforms are more complex than previously appreciated.
We have shown previously that E1A expressing MEFs deficient for p63 and p73 are resistant to apoptosis [15]. Paradoxically, we found that p63−/−;p73−/− primary and E1A MEFs are radiosensitive in long-term clonogenic assays. This finding coupled with the inability of p63/p73 deficient cells to repair DNA as shown by Comet assay indicate that p63 and p73 play a critical role in DNA repair. This new finding does not preclude that p63/p73 deficient cells are resistant to apoptosis after acute exposure to DNA damage. These data demonstrate that surviving p63−/−;p73−/− cells are unable to proliferate and establish a colony after DNA damage. This is likely due to defects in the DNA repair mechanisms.
Using a genome wide analysis, these studies have revealed novel transcriptional targets of the p53 family members. We have also identified a novel mechanism of the regulation of the DNA repair pathway by p63 and p73. Given the high incidence of mammary adenocarcinoma in p63/p73 mutant mice, these studies have unveiled a potential mechanism for p63 and p73 as tumor suppressor genes. In addition, our studies have revealed further complexity by indicating that the primary transactivators of these DNA repair genes are the ΔN isoforms of p63 and p73. These isoforms have previously been thought to act as oncogenes. More recent data have challenged this notion as these isoforms can also transactivate genes involved in apoptosis and the expression of these isoforms does not provide a growth advantage [12]. These studies provide further evidence that the ΔN isoforms may have some anti-tumor functions such as the ability to engage DNA repair pathways. Future studies using isoform specific knock out mice should yield important insights in to how each of these isoforms contributes to tumor suppression and shed light on the interactions of the complex p53 family.
The Laboratory of Genetics at The National Institute on Aging (NIA) cloned approximately 15,000 unique cDNAs into the NotI/SalI site of Ampicillin-resistant pSPORT1 vector (Life Technologies). Average insert size of the clones is 1.5 kb (0.5–3 kb). Inserts were amplified for microarray printing following a modified version of the protocol described previously [34]. In 96 well format, bacterial stocks were grown overnight in 2X YT medium (100 µg/ml ampicillin) with agitation. Ten microliters of the overnight bacterial culture was added to 90 µl ddH2O in PCR plates (MJ Research) and denatured at 95° C for 10 minutes. Following denaturation, plates were centrifuged for 10 minutes. To perform PCR, 5 µl of supernatant from each well was used as template in a 100 µl reaction with 3.5 units of AmpliTaq DNA polymerase (Applied Biosystems), forward primer (5′–CCAGTCACGACGTTGTAAAACGAC-3′) reverse primer (5′-GTGTGGAATTGTGAGCGGATAACAA-3′), and deoxynucleotide triphosphates (dNTPs). Amplification was carried out in thermocyclers with a program that contained an initial denaturation step at 95°C for 2 minutes followed by 38 cycles of 30 s at 94°C, 45 s at 65°C, and 3 minutes at 72°C, and a final extension of 5 minutes at 72°C. The amplified inserts were then purified using Montage PCR96 cleanup Filter Plates (Millipore) on a BIO-TEK Precision 2000 Automated Microplate Pipetting System to a purified volume of 100 µl. Thirty-five microliters of each purified PCR product was added to a 384-well plate, and desiccated using a large Savant Speed-vac apparatus, then reconstituted in 7 µl of 3X SSC/1.5 M betaine to a mean concentration of 600 ng/µl. The microarrays were fabricated at the MIT BioMicro Center using Corning GAPS II Gamma Amino Propyl Silane slides. cDNA clones were printed using a BioRobotics Microgrid 600 TAS Arrayer with a 32-pin print head and quill pin microfluidic liquid transfer technology.
All procedures involving mice were approved by the IACUC at U.T. M.D. Anderson Cancer Center and M.I.T. E1A-expressing mouse embryonic fibroblasts (MEFs) (wild-type, p53−/−, p63−/−, p73−/−, and p63−/−;p73−/−) were generated as described previously [15] from passage 1 primary MEFs. 3×106 E1A MEFs were plated on each of 6–15 cm dishes. Twenty-four hours after plating, the cells were treated with 0.34 µM doxorubicin. Twelve hours after treatment, total RNA (150–300 µg) was extracted from treated and untreated E1A MEFs using the RNAeasy Midi Kit (Qiagen). For each microarray hybridization, 100 µg of total RNA prepared from the reference or experimental cells were labeled by incorporating Cy3- or Cy5-labeled dUTP (NEN) using oligo d(T) (MWG) and Superscript II reverse transcriptase (Invitrogen). The resulting probes were purified using the Qiaquick PCR purification Kit (Qiagen) and recovered in a volume of 30 µl ddH20.
The printed slides were rehydrated, UV cross-linked, and blocked to reduce background using succinic anhydride (Sigma), 1-methyl 2-pyrrolidinone and sodium borate. Each slide was incubated in 60 µl total volume of hybridization solution containing Cy3- and Cy5-labeled target (one probe is the reference invariant target and the other is the experimental target), 1 µg of Mouse Cot-1 DNA (Invitrogen), 0.1 units of poly-A40–60 (Amersham Pharmacia), and 10.1 µg of Salmon Testes DNA (Sigma), 25% Formamide, 5X SSC, 0.1% SDS under a 22×40-mm lifterslip (Erie Scientific Company) at 42°C for 16 hours exactly. The slide was placed in a sealed hybridization chamber (Corning) containing two side wells with a total of 20 µl 3X SSC for humidification in a light-sealed water bath. After exactly 16 hours of hybridization, the slide was washed in 500 ml of 1X SSC, 0.03% SDS for 5 minutes after the lifterslips are gently removed in the wash solution. Then, the slides were washed for 5 minutes in 0.1X SSC, 0.01% SDS followed by 0.1X SSC. Slides were centrifuged in a speed-vac to dry. Each slide was scanned using an arrayWoRx Auto Biochip Reader that employs white light, polychromatic filter-wheel/CCD camera (Applied Precision) at wavelengths corresponding to each analog's emmision wavelength (595 and 685 nm for Cy3 and Cy5, respectively). RNA from each sample was hybridized to four independent cDNA microarrays. For 2 replicates, the invariant target was labeled with Cy3 and the experimental target was labeled with Cy5. For the other 2 replicates for each sample, the invariant target was labeled with Cy5 and the experimental target was labeled with Cy3. The invariant reference target RNA used was extracted from untreated wild type- E1A MEFs. These cells were chosen as a source of reference target RNA because this species of RNA robustly hybridized to a large percentage of genes, and it is relevant to the experimental design.
Total RNA was extracted from the E1A MEFs of the genotypes described above using the RNeasy Midi and Rnase-free Dnase kits (Qiagen). RNA was quantified and tested for quality on the Agilent 2100 Bioanalyzer (Agilent Technologies). To generate cDNA, RNA (2 µg) from each E1A MEF line treated with 0.34 µM doxorubicin was used for random hexanucleotide- primed cDNA synthesis. Each 40 µl reaction contained 1X buffer, 10 µM DTT, 1 µg random hexamer, 2 µl of Superscript II (Invitrogen), 0.5 mM each of all four dNTPs, and 80 units of RNase inhibitor (Promega). Using heating blocks, reactions were incubated at 42°C for 1 hour, 70°C for 15 minutes, 37°C for 20 minutes, and 95°C for 2 minutes. RNase H (2 units) (Invitrogen) was added to each reaction following the 70°C incubation. Afterwards, each reaction was diluted with ddH2O to a final working volume of 200 µl. cDNAs (2 µl) were added to 25-µl reaction mixtures containing 12.5 µl of 2X SYBR Green master mix (Applied Biosystems), and 40 nm of gene-specific primers. Primers were designed using Primer Express software (Applied Biosystems). Assays were performed in triplicate with an ABI Prism 7000 Sequence Detector (Applied Biosystems). All data were normalized to an internal standard (18 S ribosomal RNA; TaqMan Ribosomal RNA Control Reagents VIC Probe: Protocol: Rev C, Applied Biosystems) or GAPDH.
ChIP Assay was performed as described previously, E1A MEFs (wild-type, p53−/−, p63−/−, p73−/−, and p63−/−;p73−/−) were untreated or treated with 0.34 µM doxorubicin for 12 hours, which are the same conditions used for the array and real time PCR. Cellular proteins were crosslinked to chromatin with 1% formaldehyde. p53-DNA, p63-DNA or p73-DNA complexes were immunoprecipitated using the following antibodies: pan-p63 (4A4, Santa Cruz), pan-p73 (IMG-259a, Imgenex) or p53 (Ab-3, Oncogene Research Products). Immunprecipitated complexes were recovered by Staphylococcus A cells, treated with proteinase K, and DNA was purified. PCR was performed for putative p53 family binding elements. Putative p53 family member binding sites were identified by scanning 1000 bp of the 5′ UTR, exon 1, intron 1, exon 2 and intron 2 for the consensus p53 binding site [31]. These sites are summarized in Table 1. Sequences for primers used are available upon request.
To generate the pGL3-Rad51 luciferase reporter, DNA was amplified from a BAC clone containing the Rad51 gene (RP23-15121, CHORI BACPAC resources) using primers designed containing the p73 binding site shown by ChIP and 5′ NheI and 3′ XhoI cloning restriction enzyme sites: forward primer (5′- ACTAGCTAGCAGCAGGGCGACCAACCGAC-3′) and reverse primer (5′-CCGCTCGAGTGGCCCTCCCTATCCACAGG-3′). To construct the pGL3-BRCA2 luciferase reporter, the DNA fragment containing the p63/p73 binding site shown by ChIP was amplified from C57/B6 genomic DNA by PCR using the following primers with 5′ XhoI and 3′ BglII cloning restriction enzyme sites: forward primer (5′-CCGCTCGAGAGAGGGATCCGGCGCGTC-3′) and reverse primer (5′-GGAAGATCTGGTCTAAGCTCTGTTGCTCCTG-3′. To generate the pGL3-Mre11 luciferase reporter, DNA was amplified from a BAC clone containing the mre11a gene (RP23-149D5, CHORI BACPAC resources) using primers designed containing the p63 binding site shown by ChIP and 5′ XhoI and 3′ BglII cloning restriction enzyme sites: forward primer (5′- CCGCTCGAGACAGAGAGAACCTCACCGAGAAC -3′) and reverse primer (5′-GGAAGATCTCTGTACCAGGTTCCTCTCCAAG-3′). The resulting amplified DNA fragments were gel-purified (Wizard Prep Kit, Promega) after restriction enzyme digestion and then ligated to pGL3-basic vector (Promega) between the respective cloning sites.
6×105 wild-type and p63−/−;p73−/− MEFs were plated on 6 cm dishes. Twelve hours after plating, the MEFs were irradiated with 5 Gy of gamma-irradiation and then harvested at 10 minutes, 30 minutes, 1, 2, and 4 hours. The MEFs were lysed on ice in lysis buffer (100 mM Tris, 100 mM NaCl, 1% Nonidet P40, protease inhibitor cocktail (Roche)). Thirty micrograms of each lysate was subjected to electrophoresis on a 10% SDS PAGE for Rad51 and transferred to PVDF membrane. Rad51 was detected using the anti-Rad51 antibody (clone 51RAD01, Neomarkers), and BRCA2 was detected using the anti-BRCA2 antibody (clone H-300, Santa Cruz).
Slides were dewaxed in xylene and rehydrated in a graded series of ethanol following standard protocols [7]. Slides were incubated with primary antibodies for p63 (4A4, Santa Cruz), p73 (IMG-259A, Imgenex), Rad51 (clone 51RAD01, Neomarkers), or BRCA2 (clone H-300), Santa Cruz). at a dilution of 1∶100 for 18 hours at 4 deg C. Detection was performed using the Vectastain kit (Vector Labs) followed by the VIP kit or DAB kit (Vector Labs) and counterstained with methyl green (Vector Labs). Ten normal mammary glands and ten mammary adenocarcinomas were stained with each antibody.
p63−/−;p73−/−, p53−/−;p73−/− or p53−/−;p63−/− MEFs were plated on 6-well plates (3.5×105 cells per well). Twelve hours after plating, the MEFs were transiently transfected using Fugene HD (Roche) with 2.5 µg of the following Firefly luciferase reporter plasmids (pGL3-Rad51-1, pGL3-Rad51-2, pGL3-BRCA2) or pPERP-luc [24], 1 µg of Renilla luciferase plasmid (transfection control), and 2.5 µg of empty vector (pcDNA3) or plasmids encoding the p63/p73 isoforms (TAp63α, TAp63γ, ΔNp63γ, TAp73α, TAp73β, ΔNp73α and ΔNp73β) or p53/ In experiments where 2 isoforms of p63 and p73 were assayed simultaneously, 1.25 µg of each isoform was used. After 24 hr, cells were harvested and luciferase activity was measured using the Dual-Luciferase Reporter Assay system (Promega) and a Veritas microplate luminometer (Turner BioSystems). The relative luciferase activity was determined by dividing the Firefly luciferase value with the Renilla luciferase value and the fold increase in relative luciferase activity was determined by dividing the relative luciferase value induced by p63 and p73 isoforms with that induced by the pcDNA3 control vector. Each experiment was performed in triplicate.
E1A MEFs or primary MEFs were plated in 6-well plates (1×106 cells per well) of the following genotypes (wild-type, p53−/−, p63−/−,p73−/−, and p63−/−;p73−/−) [15]. Twelve hours later, MEFs were irradiated with 1, 2, and 3 Gy of gamma radiation or 0.34, 0.5, and 1 µM doxorubicin. After 12 hr, 1200 cells were plated on 10 cm dishes. After 12 days of incubation, the cells were stained with clonogenic reagent (0.25% of 1,9-dimethyl-methylene blue in 50% ethanol). Surviving colonies were counted, and the survival rate was calculated as the ratio of the surviving colonies after DNA damage treatment over the number of colonies for each genotype before treatment. Each experiment was performed in triplicate on three independent MEF lines for each indicated genotype.
Wild-type, p53−/−, p63−/−, p73−/−, and p63−/−;p73−/− primary and E1A MEFs were plated on 6-well dishes (1.6×105 cells per well). Twelve hours after plating, MEFs were irradiated with 5 Gy of gamma radiation. Cells were harvested 0,1, and 16 hours later for Comet Assay (Trevigen) according to the manufacturer's protocol specific for DSB detection. Briefly, cells were suspended in PBS at a density of 3×105 cell/mL. Twenty microliters of each cell suspension was mixed with 200 µL of melted low melting point agarose (LMA) and 75 µL of this mixture was placed onto the Trevigen CometSlide for electrophoresis. Subsequent to electrophoresis, samples were visualized with SYBR Green I and fluorescence microscopy. Twenty pictures were taken for each sample and at least 135 cells per experiment were examined for comet tails using CometScore software (TriTek Corporation). Three independent MEF lines for each genotype were assayed in triplicate. Student's t test was used for statistical analysis.
All experiments were performed at least in triplicate. Data are represented as the mean ± SEM. Statistics for qRT-PCR, luciferase, clonogenic, and comet assays was performed using Student's t test for comparison between two groups. A p value of 0.05 was considered significant.
|
10.1371/journal.pntd.0005528 | Transmission of Zika virus through breast milk and other breastfeeding-related bodily-fluids: A systematic review | Zika virus (ZIKV) infection is an emerging mosquito-borne disease, which is associated with an increase in central nervous system malformations and newborn microcephaly cases. This review investigated evidence of breastfeeding transmission from ZIKV-infected mothers to their children and the presence of ZIKV infection in breastfeeding-related fluids.
We conducted a systematic review of observational studies, case studies, and surveillance reports involving breastfeeding women with ZIKV infection in several international databases. Data extraction and analysis were conducted following a PROSPERO-registered protocol. From 472 non-duplicate records, two case reports met criteria for inclusion. We reviewed three cases of ZIKV infection among lactating mothers near the time of delivery. Two of the three (2/3) associated newborns had evidence of ZIKV infection. ZIKV was detected in breast milk of all three mothers. Breast milk detection results were positive in all mothers (3/3) by RT-PCR, one was positive by culture (1/3), and none was tested for ZIKV-specific antibodies. Serum samples were ZIKV positive in all mothers (3/3), and sweat was not tested for ZIKV.
We describe three cases of ZIKV-infected breastfeeding mothers who were symptomatic within three days of delivery, and two cases with ZIKV-infected newborns. While ZIKV was detected in the breast milk of all three mothers, the data are not sufficient to conclude ZIKV transmission via breastfeeding. More evidence is needed to distinguish breastfeeding transmission from other perinatal transmission routes.
| Zika virus (ZIKV) infection was considered a mild fever illness until the recent and ongoing outbreak in South America revealed that the virus can cause microcephaly and other neurological disorders. ZIKV is primarily transmitted by mosquitoes, but can also spread from person to person through sex, blood transfusion, and from mother to child during pregnancy or delivery. It is unknown if ZIKV can spread from mother to child during breastfeeding. We conducted a systematic review of the literature to summarize evidence of ZIKV transmission via breastfeeding and the presence ZIKV in breast milk. Our search resulted in 2 case reports that described 3 mothers and their newborns. ZIKV was confirmed in the blood and breast milk of all 3 mothers and in the blood of 2 newborns. More evidence is needed to confirm if ZIKV is transmitted via breastfeeding.
| Zika virus (ZIKV) infection is an emerging vector-borne disease of the Flaviviridae family, which includes dengue, yellow fever, Japanese encephalitis, and West Nile viruses [1]. ZIKV infection causes a mild, self-limiting influenza-like illness with a 10-day incubation for most cases and shares similarities with other circulating arthropod-borne viral infections like the alphavirus chikungunya [1, 2]. Many cases of ZIKV infection are asymptomatic and therefore unreported.
The World Health Organization (WHO) has developed an interim case definition to classify and report cases of ZIKV infection (Fig 1). A suspect case is a person presenting with rash and/or fever and at least one of the following: arthralgia, arthritis or conjunctivitis. A probable case is a suspected case with presence of IgM antibody against ZIKV and an epidemiological link; and a confirmed case is a person with laboratory confirmation of recent ZIKV infection: by presence of ZIKV RNA or antigen in serum or other samples or IgM antibody against ZIKV positive and plaque reduction neutralization test ≥ 90% (PRNT90) for ZIKV with titre ≥ 20 and ZIKV PRNT90 titre ratio ≥ 4 compared to other flaviviruses [3]. Due to the possible cross reactivity with other members of the Flaviviridae family, the presence of IgM is not enough to rule out ZIKV infection, and the PRNT90 will determine if the in vitro inhibition of cell growth is produced by antibodies against ZIKV [4, 5]. An enzyme linked immunoassay (ELISA) for ZIKV has been developed by the Centers for Disease Control and Prevention, but is only available upon request for emergency use [6].
The timing and the test performed could be crucial for detecting the ZIKV infection. During the first 7 days, viral RNA can often be identified by reverse transcriptase polymerase chain reaction (RT-PCR), but as viremia decreases, a negative RT-PCR does not exclude flavivirus infection, and serologic testing should be performed. On the other hand virus-specific IgM antibodies may be detectable >4 days after onset of illness, however a sample taken within 7 days of illness onset may not have detectable virus-specific IgM antibodies [7].
ZIKV transmission occurs primarily via the bite of Aedes aegypti mosquitoes, in addition to Aedes spp. Ae. africanus, Ae. albopictus, Ae. hensilli, and Ae. Luteocephalus [2, 8–12]. However, perinatal, transfusion, and sexual transmission have also been reported [13–17]. Among infected individuals, evidence of ZIKV has been detected in serum, saliva, urine, semen, and breast milk [13, 18–22]. Generally, transmission of antibodies through breast milk has been described, particularly for IgA, conferring passive immunity [23]. The presence of IgA, IgG, or IgM antibodies against similar flaviviruses such as West Nile Virus has been reported in breast milk [24]. Given recent increases of ZIKV cases in Central and South America and suggested associations with congenital microcephaly and other non-congenital neurological or autoimmune disorders, an investigation of transmission via breast milk is needed [25].
Until recently, outbreaks of ZIKV were sporadic. During the last 50 years, widespread infection throughout Africa and Southeast Asia is suspected, but the asymptomatic nature and limited diagnostics have likely hampered disease surveillance [12, 26–28]. In 2007, the disease migrated to Oceania where an outbreak in Yap State in the Federated States of Micronesia infected roughly 5,000 individuals, nearly 75% of the island population [2]. The next outbreaks occurred in French Polynesia (396 confirmed), New Caledonia (1,400 confirmed), and the Cook Islands (50 confirmed) in 2013–2014 [29–31]. The first official outbreak in the Americas arrived to Easter Island, Chile in early 2014 with 51 confirmed cases [32]. In April, 2015, Brazil reported the first confirmed autochthonous case of ZIKV infection [33]. Since then, an epidemic has rapidly expanded affecting 48 countries and territories in South and Central America [34]. The Brazilian Ministry of Health estimates the number of ZIKV cases in 2015 alone between 0.4–1.3 million [8].
During the Brazilian ZIKV epidemic, clinicians have observed a 20-fold increase in suspected cases of microcephaly in newborns [35]. Reported microcephaly and/or central nervous system malformations have affected 7,150 individuals in Brazil between 22 October 2015 and 16 April 2016 [36]. Other flaviviruses have not been known to cause microcephaly, however ZIKV has been confirmed in recent microcephaly cases, which has prompted global concern for pregnant women and a large-scale investigation [37]. A recent report from the WHO indicated that there is scientific consensus that Zika virus is a cause of microcephaly and Guillain-Barré syndrome [36, 38] based on results from a systematic review [39, 40]. Many mothers of infants with microcephaly reported no illness or symptoms associated with Zika infection [41]. Regardless of symptoms, pregnant women are at risk for infection and potential complication in any trimester [13, 42]. At this time, the WHO recommends standard breastfeeding practices for all mothers, regardless of ZIKV infection [43], unless there is an acceptable medical reason for permanent or temporary avoidance of breastfeeding [44].
The primary objective of this systematic review was to review evidence related to the transmission of ZIKV through breastfeeding. For the purposes of this review, ZIKV infection included suspected, probable, or confirmed cases as described by the WHO interim case definition. A secondary objective assessed the available literature regarding the presence of ZIKV or ZIKV-specific antibodies in breast milk and breastfeeding-related bodily fluids (i.e. blood or sweat) of lactating women. We sought to address the following questions:
Primary Outcome: Does the literature provide evidence that in ZIKV-free infants or children, breastfeeding (any or exclusive) from a ZIKV-infected lactating mother, compared to not breastfeeding, result in evidence of ZIKV infection in the infant?
Secondary Outcome: Does the literature provide evidence there are ZIKV specific antibodies present in breast milk?
Study characteristics, as well as inclusion and exclusion criteria, were defined by study designs, participants, ZIKV infection exposure, and outcomes.
A search overview is provided in the S1 Appendix.
Electronic databases: Search terms included variations and permutations of United States National Library of Medicine Medical Subject Headings (MeSH) terms and text words relating to flaviviruses (Zika, West Nile, and yellow fever), breastfeeding, transmission fluids (breast milk, blood, and sweat), and participants (mother or child) (See appendix for full search strategy). Report characteristics included a time range of all years, any language, and any publication status. The following electronic databases were searched:
Screening of search results was performed using Covidence systematic review software (Veritas Health Innovation, Melbourne, Australia). Two authors independently screened the titles and abstracts of studies based on the inclusion criteria. A third author assessed and resolved disagreements on study selection. All irrelevant titles were excluded. For studies that met eligibility criteria, full text articles were obtained and managed using EndNote (version X7·5 2016 Thomson Reuters), a reference management software.
A data extraction form was tailored for this review. One author extracted study characteristics and two authors extracted study outcome data according to the pre-designed data extraction form. For each study, information pertaining to the source, eligibility, methods, participants, exposures, outcomes, and results was entered into the data extraction form. When relevant, effect estimates including odds ratios, relative risks, mean differences, or summary effects were extracted for each outcome. All potential modifiers or confounders of study outcomes were included in the extraction form.
This review followed a pre-established protocol based on methods for systematic reviews described in the Cochrane Handbook for Systematic Reviews [45]. The protocol was registered in PROSPERO, the international prospective register of systematic reviews of the University of York and the National Institute for Health Research, under the number CRD42016036667. The authors followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines and include a checklist in the S1 Checklist.
Our search strategy (11 March 2016) identified 472 records, detailed in the S1 Table, after duplicates were removed (Fig 2). No unpublished records were identified in the search nor included in the analysis. Initial screening retained 42 records. At the time of screening, inclusion criteria included terms for West Nile and yellow fever viruses. For the purposes of this review, only ZIKV was considered for quantitative analyses, which yielded 2 records for review (Fig 2). A total of 2 studies (that included three mother child pairs) were included for analysis. The main reasons for exclusion were non-ZIKV infections or ineligible populations.
A case report from the ZIKV outbreak in French Polynesia (French territory) described two mothers, who had recently given birth, with ZIKV infection (Table 1) [13]. Mother 1 initiated breastfeeding to Newborn 1 on the day of delivery. On day 2 following delivery, mother 1 had a confirmed case of ZIKV detected by serum RT-PCR and saliva RT-PCR. On day 3, the breast milk from mother 1 was found to contain ZIKV by RT-PCR, however ZIKV breast milk culture was negative. Also on day 3, Newborn 1 had confirmed ZIKV infection by serum RT-PCR and saliva RT-PCR.
Mother 2 was confirmed with ZIKV infection on days 1 and 5 post delivery by serum RT-PCR and initiated breastfeeding on day 3. On day 8, the ZIKV RT-PCR results from mother 2 were serum negative, urine positive, and breast milk positive, however ZIKV breast milk culture was negative. Newborn 2 tested negative for ZIKV on day 0 and day 3 by serum RT-PCR, but had confirmed ZIKV infection on days 4 and 7 by serum RT-PCR and on day 8 by urine RT-PCR. On day 9, newborn 2 urine was ZIKV negative by RT-PCR. These case reports confirmed ZIKV infection in 2 breastfeeding mothers and their newborns as well as detected ZIKV in serum and breast milk of both mothers. Both mothers had clinical signs of rash within days of delivery, and the authors hypothesized that the infants were probably infected in utero or intrapartum because the infants’ sera were positive for the presence of Zika virus within one day of starting breastfeeding. The author of this study was contacted (M. Besnard, personal communication, 2016) and confirmed that no long-term complications were reported for either of the two infants at 2 years of age.
A second case report described a mother (referred to as case 3 in Table 1) from New Caledonia (French territory) who initiated breastfeeding on the day of delivery and developed fever and maculopapular rash in the following days [47]. On day 3 post delivery, mother 3 tested positive for ZIKV infection by serum RT-PCR, however the serum RT-PCR results for newborn 3 were reported as ambiguous. Breast milk was ZIKV positive by RT-PCR on day 4 and ZIKV breast milk culture was also positive. While vertical transmission was not described in this case, the presence of ZIKV in breast milk was confirmed. No long-term complications were reported for the child at 8 months of age (M. Dupont-Rouzeyrol, personal communication, 2016). The overall quality of the evidence was very low for all the proposed outcomes, as described in the GRADE Summary of Findings (Table 2).
The cases presented in these two reports confirm the presence of ZIKV RNA in breast milk from three ZIKV-infected mothers. The presence of Zika-specific antibodies was not reported in these cases. Of the three newborns delivered to ZIKV-infected mothers who were receiving breast milk with confirmed presence of ZIKV, only two were confirmed to be infected with ZIKV with no reported adverse outcomes. With regard to the presence of ZIKV in breastfeeding-related fluids, ZIKV was detected by RT-PCR in breast milk and blood of the three mothers; sweat was not measured.
Like other viral infections, mother-to-child transmission of ZIKV infection can potentially occur during antepartum, intrapartum, or postnatal periods.[48] Given the variable incubation period for ZIKV, it can be difficult to distinguish breastfeeding transmission from other perinatal routes. For the two newborns who contracted ZIKV from ZIKV-infected mothers expressing ZIKV-infected breast milk, antepartum or intrapartum transmission is suspected. Even if a newborn is ZIKV negative following delivery from a ZIKV-infected mother and contracts ZIKV infection while consuming breast milk with ZIKV, there remains a possibility for separate mosquito transmission. Identifying the time of infection and duration of an incubation period is further complicated by the asymptomatic nature of acute ZIKV infection.
There is limited evidence describing breastfeeding transmission for other flavivirus infections. West Nile virus (WNV), dengue virus, and yellow fever virus have been detected in breast milk [49, 50]. Of these infections, WNV has been associated with breastfeeding transmission in a small number of cases [24]. Like ZIKV, breastfeeding transmission for other flavivirus infections is likely underreported due to asymptomatic illness and limited access to diagnostics. We intend to review breastfeeding transmission for related flavivirus infections in the near future.
Our systematic review for ZIKV breastfeeding transmission resulted in two studies and three cases of lactating women with confirmed ZIKV infection. As new data emerges from these current outbreaks, further investigation is needed to explore ZIKV breastfeeding transmission dynamics. This includes understanding the mechanics of transmission with regards to timing of infection for mother and infant, breast milk viral load, and exposure duration as well as assessing the frequency and distribution of breastfeeding transmission among affected populations. In addition to determining viral transmission risk, research should also explore the protective properties of ZIKV-specific immunoglobulin in breast milk transferred from mothers who have experienced ZIKV infection. At this time, the data are not sufficient to conclude ZIKV transmission via breastfeeding, and the authors support the WHO breastfeeding guidelines currently in place recommending initiating breastfeeding within one hour of delivery, exclusively for 6 months and extended until 2 years or beyond [51].
|
10.1371/journal.pcbi.1000784 | Chemotactic Response and Adaptation Dynamics in Escherichia coli | Adaptation of the chemotaxis sensory pathway of the bacterium Escherichia coli is integral for detecting chemicals over a wide range of background concentrations, ultimately allowing cells to swim towards sources of attractant and away from repellents. Its biochemical mechanism based on methylation and demethylation of chemoreceptors has long been known. Despite the importance of adaptation for cell memory and behavior, the dynamics of adaptation are difficult to reconcile with current models of precise adaptation. Here, we follow time courses of signaling in response to concentration step changes of attractant using in vivo fluorescence resonance energy transfer measurements. Specifically, we use a condensed representation of adaptation time courses for efficient evaluation of different adaptation models. To quantitatively explain the data, we finally develop a dynamic model for signaling and adaptation based on the attractant flow in the experiment, signaling by cooperative receptor complexes, and multiple layers of feedback regulation for adaptation. We experimentally confirm the predicted effects of changing the enzyme-expression level and bypassing the negative feedback for demethylation. Our data analysis suggests significant imprecision in adaptation for large additions. Furthermore, our model predicts highly regulated, ultrafast adaptation in response to removal of attractant, which may be useful for fast reorientation of the cell and noise reduction in adaptation.
| Bacterial chemotaxis is a paradigm for sensory systems, and thus has attracted immense interest from biologists and modelers alike. Using this pathway, cells can sense chemical molecules in their environment, and bias their movement towards nutrients and away from toxins. To avoid over- or understimulation of the signaling pathway, receptors adapt to current external conditions by covalent receptor modification, ultimately allowing cells to chemotax over a wide range of background concentrations. While the robustness and precision in adaptation was previously explained, we quantify the dynamics of adaptation, important for cell memory and behavior, as well as noise filtering in the pathway. Specifically, we study the intracellular signaling response and subsequent adaptation to concentration step changes in attractant chemicals. We combine measurements of signaling in living cells with a dynamic model for strongly coupled receptors, even including the effects of concentration flow in the experiment. Using a novel way of summarizing time-dependent data, we derive a new adaptation model, predicting additional layers of feedback regulation. As a consequence, adaptation to sudden exposure of unfavorable conditions is very fast, which may be useful for a quick reorientation and escape of the cell.
| Cells are able to sense and respond to various external stimuli. To extend the working range of their sensory pathways, biochemical mechanisms allow for adaptation to persistent stimulation, resulting in only a transient response. The dynamics of adaptation are important as they often represent the cells' memory of previous environmental conditions, directly affecting cellular behavior [1]–[7]. Fast adaptation means that cells stop responding and that their biochemical pathways quickly “forget” the stimulus. In contrast, slow adaptation leads to long-lasting effects in the cells. The dynamics of adaptation are often difficult to understand in detail, since they emerge from multiple, simultaneously occurring processes. In higher organisms, adaptation is best documented in the insect and vertebrate visual system, where multiple mechanisms adjust the receptor sensitivity to ambient light levels. For instance, phototransduction in the vertebrate eye involves up to nine different mechanisms for adaptation [8]. However, even in the well-characterized chemotaxis sensory system in Escherichia coli [9]–[13], adaptation, in particular its molecular mechanism and dynamics, is not well understood. This constitutes a huge deficit as there has recently been immense interest in the chemotactic behavior of bacteria [14]–[18] and noise filtering [17], [19], [20]. Here, we use adaptation time-course data from in vivo fluorescence resonance energy transfer (FRET) measurements and quantitative modeling to address this problem.
The chemotaxis pathway in E. coli allows cells to sense chemicals and to swim towards more favorable environments by successive periods of straight swimming (running) and random reorientation (tumbling). Transmembrane chemoreceptors, including the highly abundant Tar and Tsr receptors, cluster at the cell poles and act as “antennas” to detect various attractants (e.g. certain amino acids and sugars) and repellents (e.g. certain metal ions) with high sensitivity [21]. Receptors activate an intracellular signaling pathway, which results in the phosphorylation of diffusible response regulator CheY (CheY-P) via kinase CheA. CheY-P binds to the flagellated rotary motors and induces tumbling. For details of the pathway see the Supplementary Text S1. The interactions between different proteins in the chemotaxis pathway during signaling have been well characterized. Specifically, FRET measurements on the response regulator CheY-P and its phosphatase CheZ have elucidated the signaling in the chemotaxis pathway [22]–[24].
Adaptation in E. coli is based on reversible methylation and demethylation of receptors at specific modification sites, catalyzed by enzymes CheR and phosphorylated CheB (CheB-P), respectively. Tar and Tsr receptors have four major modification sites. In addition, the Tsr receptor has two minor modification sites which are methylated less strongly [25]. Receptor modification regulates the receptor activity and provides a recording of experienced concentration changes [16], [26], [27]. As a consequence, the rate of tumbling was found to adapt precisely for different ligand concentrations [28], [29]. To achieve the return of the receptor activity to its pre-stimulus value, receptor activity-dependent phosphorylation of CheB provides a negative feedback on the receptor activity. In addition, the rates of methylation and demethylation depend on the receptor activity [30]–[32], representing further layers of feedback regulation. To modify receptors, CheR and CheB molecules can bind to a specific tether sequence at the carboxyl-terminus of Tar and Tsr receptors, and act on several nearby receptors, so-called assistance neighborhoods [33]. This is believed to compensate for the low numbers of CheR and CheB (hundreds of molecules) [34], although larger numbers have been reported [35].
Although a lot is known about the components of the chemotaxis pathway, several open questions remain to be answered in adaptation. (i) Despite their importance for cell behavior, memory and noise filtering, the dynamics of adaptation and the methylation level are largely unknown. This is because the methylation level is difficult to measure precisely, relying on quantification of receptor protein and radioactively-labeled methylation substrate (methionine) incorporated into receptors [25], [36]–[38]. So far, only the initial rate of adaptation was inferred from the rate of change in motor bias in response to saturating amounts of added attractant [29]. (ii) The molecular mechanism of adaptation, in particular demethylation, remains unclear. While CheR binds strongly to the tether, suggested to increase its concentration in the vicinity of methyl-accepting sites [39], the binding affinity of CheB was found to be very low [40]. Instead, binding of CheB-P to the tether induces an allosteric activation of the receptor, increasing the demethylation rate [40]. Furthermore, while the receptor activity-dependence of the methylation and demethylation rates is believed to be a requirement for robust precise adaptation (see below), it is not known if adaptation is precise at the receptor level. Time-course data from in vivo FRET experiments, monitoring receptor activity upon stimulation, is ideally suited to study the adaptation dynamics and address these questions.
Extensive mathematical modeling has described different aspects of the chemotaxis pathway. However, modeling has mainly focused on explaining the initial response to addition of attractant, as well as precise adaptation, i.e. the complete return of the signaling activity to pre-stimulus level long after the stimulus. Briefly, the Monod-Wyman-Changeux (MWC) model was used to successfully describe the signaling of two-state receptor complexes, formed by 10–20 strongly interacting receptor dimers [24], [41]–[44]. In this model, receptor-receptor coupling provides a mechanism for signal amplification and integration. Alternative receptor models are outlined in the Discussion. Furthermore, Barkai and Leibler showed that precise adaptation is robust (insensitive to parameter variations in the pathway), if the kinetics of receptor methylation depends only on the activity of receptors and not explicitly on the receptor methylation level or external chemical concentration [45]. Their idea was later extended by others, providing conditions for precision [46], [47], as well as robustness to noise by the network architecture [48] and assistance neighborhoods [42], [49]. Most recently, adaptation to exponential ramps and sinusoidal concentration changes was investigated [20]. However, none of these studies have directly compared to adaptation time-courses from FRET.
Here, we use in vivo FRET data obtained from cells adapted to ambient concentrations of attractant -methylaspartate (MeAsp; a non-metabolizable variant of amino acid aspartate) and stimulated in a flow chamber by various concentration step changes [23]. Recording the average initial response amplitudes for each added and, after adaptation, removed concentration step change results in dose-response curves (Fig. 1, symbols). We use a dynamic version of the MWC model, which, in addition to mixed complexes of Tar and Tsr receptors, includes a more detailed description of the adaptation dynamics than used in previous models of chemotaxis. Specifically, we predict multiple layers of feedback regulation during adaptation, especially for demethylation by CheB. In addition, we take into account the kinetics of attractant flow in FRET experiments. This allows us to quantitatively describe dose-response curves (Fig. 1, lines), in particular the observed reduced response amplitudes for removal of MeAsp, which previously could not be explained by the MWC model (Inset). To analyze the adaptation dynamics, we use the data collapse, a condensed representation of time courses. This data collapse enables us to evaluate the effect of ligand flow and adaptation imprecision, to infer the kinetics of the receptor methylation level, as well as to efficiently compare adaptation models from the literature to experimental data. Finally, we experimentally test two predictions to validate our adaptation model. We change the adapted receptor activity, and use a non-regulatable CheB mutant to bypass its negative feedback on the receptor activity. Our combined study of experiments and modeling shows that adaptation is relatively imprecise at the receptor level for large stimuli, and that demethylation is more tightly regulated than previously thought. This leads to very short tumbles for sudden occurrences of unfavorable conditions, allowing cells to quickly reorient their swimming direction after a short tumble.
Our dynamic MWC model, described in the following, combines the previously used MWC model for receptor signaling by strongly-coupled receptor complexes (denoted here by static MWC model), with the dynamic effects of adaptation by receptor modification, as well as ligand concentration flow. In the static MWC model, mixed receptor complexes composed of Tar (aspartate receptor) and Tsr (serine receptor, which also binds aspartate with lower affinity) are considered in their in vivo ratio. Using a two-state assumption, the activity of a receptor complex is given by its probability to be in on (active), which depends on the free-energy difference between its on and off (inactive) state [41], [43],(1)This free-energy difference, , is determined by two contributions, one from methylation (in terms of receptor methylation level ) favoring the on state, and one from attractant binding at MeAsp concentration favoring the off state. The free-energy difference also depends on several parameters such as free-energy difference per added methyl group, the number of receptor dimers in a complex, as well as the ligand dissociation constants and for Tar (Tsr) receptors in their on and off states, respectively. Most of these parameters were determined previously (see Materials and Methods). Similar free-energy based two-state models were recently used to describe clustering of ion channels [50] and small GTPases in eukaryotic cells [51]. In the new dynamic MWC model, we include the effects of variable receptor complex sizes, adaptation dynamics, and MeAsp concentration flow on the initial response to concentration changes.
The dependence of the receptor complex size on the ambient concentration and hence methylation level was determined as follows: First, the receptor complex size was obtained for each ambient concentration using a least-squares fit to addition dose-response curves (see Fig. 2A and Materials and Methods). Consistent with previous modeling results, we find that the receptor complex size increases with increasing ambient concentration [41], [52]. As the simplest assumption, we used a linear relationship between receptor complex size and ambient concentration (Fig. 2A), allowing us to interpolate the receptor complex size for removal dose-response curves. Analyzing the signaling pathway of E. coli, we also found the phosphorylation reactions are sufficiently fast to assume that concentrations of phosphorylated (and unphosphorylated) proteins are in quasi-steady state. Furthermore, the concentrations of activated proteins are approximately proportional to the receptor complex activity. Both these conditions allow us to use the receptor complex activity as a substitute for the down-stream activity measured by FRET reducing the number of model parameters for fitting to data greatly (see Supplementary Text S1). This approximation was also used in previous work, but was never explicitly tested [41]–[43].
Adaptation occurs on a similar time scale as the kinetics of the MeAsp concentration flow. In experiments, changes in MeAsp concentration are established over several seconds, due to the finite flow speed and mixing effects in the flow chamber. In our model, we assume exponentially rising and falling concentration changes upon addition and removal in line with previous measurements by Sourjik and Berg (Fig. 2B) [23]. Adaptation is mediated by methylation and demethylation enzymes CheR and CheB, respectively. The process is described by the kinetics of the average receptor methylation level in a receptor complex,(2)where the adapted receptor-complex activity is determined by the steady-state condition . According to our model, receptors are methylated when the complex is inactive, and demethylated when it is active. Furthermore, we postulate a very strong sensitivity of the demethylation rate on activity, possibly due to cooperativity of CheB-P molecules. This mechanism explains the strong asymmetry, which is observed in experimentally measured time courses (cf. Fig. 2C) where adaptation of inactive receptors (methylation) is slow compared to the rapid adaptation of active receptors (demethylation). Hence, during a concentration step change the initial response amplitude of receptor complexes is reduced by simultaneous adaptation, which is particularly important for removal of concentration (see Fig. 2B Inset). Note that the asymmetry between slow adaptation of inactive and active receptors, respectively, cannot simply be changed by adjusting the rate constants of methylation and demethylation individually, since they are constrained by the adapted activity . For details of this adaptation model see Materials and Methods, and for a potential molecular mechanism of demethylation, see Discussion.
Experimental dose-response curves (Fig. 1, symbols) describe the initial response of adapted wild-type cells to sudden changes (addition and removal) in MeAsp concentration [23]. These responses are taken from time courses measured by in vivo FRET (cf. Fig. 2). Additional, previously unpublished dose-response curves are provided in the Supplementary Text S1. For details of the experiments see Material and Methods. Our dynamic MWC model, which includes the effects of adaptation and MeAsp flow, quantitatively describes the experimental dose-response curves. Specifically, adaptation leads to a non-saturated response for large MeAsp step changes at high ambient concentrations, which is not seen in the static MWC model without adaptation dynamics (Fig. 1 Inset). Note, however, that responses eventually do saturate according to the dynamic MWC model for even larger concentration step changes due to the presence of Tsr receptors (at 0.5 mM ambient for ; not shown). The dynamic MWC model describes the dose-response data significantly better than the static MWC model, as indicated by their overall squared errors in the caption of Fig. 1, as well as residual errors detailed in the Supplementary Text S1. The receptor-complex activity, as well as FRET data were normalized by their adapted pre-stimulus values at ambient concentration to compare model and experimental data (see Materials and Methods).
The short-term behavior in the time-course data (Fig. 2C) is essential in deriving our adaptation model, used to quantitatively describe dose-response curves (Fig. 1). Can our adaptation model also describe the long-term behavior in the time-course data, and hence the complete adaptation dynamics? Our model for precise adaptation predicts that the observable rate of activity change is given by(3)where the rate of change of the methylation level is described by Eq. 2, and is the rate of change of the MeAsp concentration. After a concentration step change, the MeAsp concentration is constant with , and the rate of activity change is given by(4)where we used that (see Material and Methods). Hence, the rate of activity change is a function of the activity only, independent of ligand concentration and receptor methylation level (except for the minor dependence of the receptor complex size on the ligand concentration, see Supplementary Text S1). This predicts a data collapse of all adaptation time courses, independent of the duration, size and number of concentration step changes, onto a single curve (Fig. 3A, thick gray line). This non-monotonous function of the activity has three fixed points: the adapted activity , where methylation and demethylation rates exactly balance each other, as well as and , where the receptor complex activity is saturated in the off and on state, respectively. Figure 3A Inset shows the experimental rate of activity change as extracted from our quantitative time-course data from FRET for different concentration step changes at an ambient concentration. We observe that, in contrast to the prediction of the model, the rate of activity change depends on the magnitude of the concentration step changes. For addition of large concentration step changes (blue symbols), the rate is reduced and the activity stays below the pre-stimulus value. Furthermore, for total removal of MeAsp concentration (replacement with buffer medium, green symbols), the magnitude of the rate is reduced and the activity remains above the pre-stimulus value.
To explain the deviations from the predicted data collapse, we consider the effects of MeAsp flow and imprecise adaptation in our model. According to Eq. 3, each of the two effects contribute independently to the rate of activity change. First, we include the MeAsp flow for concentration step changes as described, and simulate time courses based on the precise adaptation model (Fig. 3A, solid lines). We find that in the demethylation regime (negative rate of activity change), the kinetics of concentration step removal gives rise to minor deviations from the curve in qualitative agreement with experiment. However, in the methylation regime (positive rate of activity change), unlike the experimental data, all time courses lie accurately on the curve. Next, we consider imprecise adaptation, i.e. the incomplete return of the activity to pre-stimulus level, which is apparent in the time courses (Fig. 2C and Supplementary Text S1 for quantification). In our model for imprecise adaptation, Eq. 7 in Materials and Methods, the kinetics of the methylation level depends explicitly on the receptor methylation level, which leads to significant deviations from the data collapse (Fig. 3A, dashed lines). Adaptation after addition of increasing concentration step changes results in a reduced adapted receptor complex activity (adapted activity after removal is always the same as the concentration is the ambient concentration). Total removal of MeAsp concentration (buffer) results in an increased adapted activity. Our imprecise adaptation model is in line with the experimental data, showing that the data collapse is an effective way to compare experimental and theoretical time courses and to quantify the effects of ligand flow and imprecise adaptation. We also studied the effect of changes in receptor-complex size on the data collapse, which we found to be minor for the concentrations considered here (see Supplementary Text S1). In addition to the adaptation dynamics, the data collapse allows us to determine the kinetics of the receptor methylation level, which is difficult to measure directly. Figure 3B shows the rate of change of the methylation level as a function of the receptor complex activity for experimental data, as well as the dynamic MWC model. The data and curves were obtained by dividing the rate of activity change following concentration step changes by . If the activity change is caused only by the adaptation dynamics, this procedure yields a function proportional to the rate of change of the methylation level, . According to our precise adaptation model Eq. 2, the rate of change of the methylation level is a monotonically decreasing function of activity with one steady state, marking the adapted receptor complex activity (Fig. 3B, thick gray line). Corresponding to the rate of activity change in Fig. 3A, the kinetics of ligand flow upon concentration step changes, as well as imprecise adaptation result in deviations from this curve. As before, we mainly find signatures of imprecise adaptation in the experimental data in Fig. 3B Inset.
The data collapse of experimental time courses enables the efficient evaluation of different adaptation models, including our model and other models from the literature (Fig. 4A). All models considered here show precise adaptation with the rates of methylation and demethylation only depending on the receptor complex activity, and the explicit activity dependencies given respectively by the first and second term in the legend of Fig. 4. For instance, the first model (solid red line) is given by Eq. 2. Two classes of models are analyzed here. The first class of models, including our model, is based on a linear activity-dependence of the methylation and demethylation rates for binding of CheR and CheB to inactive and active receptor, respectively. Feedback from the activity-dependent phosphorylation of CheB is accounted for by additional factors of the receptor complex activity. Our model includes cooperative CheB feedback (solid red line), while linear CheB feedback (dashed red line) and no CheB feedback (dotted red line) are considered as well [15], [42], [49], [53]. Another class of models has been proposed, showing ultrasensitivity with respect to CheR and CheB protein levels. In these models, the activity-dependence of the methylation and demethylation rates for enzyme binding is described by Michaelis-Menten kinetics, and linear CheB feedback (solid blue line) and no CheB feedback (dashed blue line) is considered [17]. The last model has the property that the rate of change of methylation level becomes independent of activity around the steady-state, leading to extremely long adaptation times. Details of the alternative adaptation models and the fitting procedure are given in the Supplementary Text S1. While several models are consistent with the experimental data, our model compares most favorably. The ultrasensitive Michaelis-Menten model without CheB feedback seems not to be consistent with the data. Comparing simulated time courses for the different adaptation models in Fig. 4B, our model is best to capture the experimentally observed asymmetry between adaptation to addition and removal of concentration step changes. The quality of fit between the respective models and data is indicated by their least-squares errors in the caption of Fig. 4.
To further validate our adaptation model, we experimentally tested two predictions. First, in our precise-adaptation model the data collapse depends strongly on the steady-state activity. For instance, increasing the steady-state activity from to 1/2 changes the data collapse from the solid to the dashed red line in Fig. 5A. Such an increase in the steady-state activity can be achieved by decreasing CheB expression level, corresponding to a decreasing demethylation rate, at constant CheR expression level. To validate this prediction, a different wild-type strain (WT2) was created, in which CheB expression was induced from a plasmid, while all other chemotaxis proteins were expressed as before (WT1). The steady-state activity was estimated to be (compared to 1/3 in WT1). For details of the strains, see Materials and Methods. The data collapse (Fig. 5A, orange circles) corresponds well to the predicted curve (dashed red line). Second, the activity-dependence of the demethylation rate is diminished according to Eq. 6 when considering adaptation without feedback through activity-dependent CheB phosphorylation, while keeping the steady-state activity constant (Fig. 5C, green line). To validate this prediction, a mutant strain was created, which contains non-regulatable CheB with about 10 percent of CheB-P activity. The CheB expression level was increased to produce the kinase activity of WT2 (). All other chemotaxis proteins are expressed as in WT2 cells. We find that the experimental rate of FRET-activity change from time-course data (green squares) is consistent with this prediction.
The statistical significance for each of the two predictions (Fig. 5A and C) was tested as follows: For each prediction, we randomly permuted a number of data points from the control experiment and the prediction-testing experiment. Then we calculated the distribution of squared errors between the rates of activity change from the model and FRET measurement for the permuted data sets (Fig. 5 B and D). For four permuted pairs of data points the error is always above the error for the unpermuted data sets (Fig. 5). For fewer permutations the error lies at the lower bound of the distribution (not shown). This confirms that the control and prediction-testing data sets are significantly different and match our model.
Sensing and adaptation are fundamental biological processes, enabling cells to respond and adjust to their external environment. Adaptation extends the range of stimuli a sensory pathway can respond to, while its dynamics determines how long a stimulus will affect the cell's behavior. In this work, we developed a model to quantitatively describe experimental dose-response curves, as well as time courses of chemotaxis signaling in adapting wild-type cells. Our model includes (i) the signaling activity of two-state mixed chemoreceptor complexes in response to added or removed attractant concentration step changes based on the Monod-Wyman-Changeux model, (ii) the kinetics of the ligand concentration in the flow chamber, and (iii) a detailed mechanism for adaptation, including multiple layers of feedback regulation and imprecise adaptation. In particular, we find that the finite ligand flow speed and fast, activated demethylation explains for the first time the gradually reduced amplitudes in removal dose-response curves for increasing ambient concentrations (Fig. 1). Our adaptation model introduces a strong receptor-activity dependence of the demethylation rate, and hence is able to reproduce the observed asymmetry of slow adaptation to addition of attractant and fast adaptation to removal of attractant (Fig. 2C). Such dynamics yields long runs up the gradient and short tumbles, sufficient for random reorientation of the cell and escape from potential toxins. Furthermore, this strong activity dependence has the additional benefit of reducing the fluctuations in receptor methylation level introduced by the adaptation mechanism itself. We found for the total variance of the receptor-complex methylation level compared to 2 for a previous model for precise adaptation with weaker activity dependence (details of the calculation can be found in the Supplementary Text S1). This is because a fluctuation in the receptor methylation level leads to an increased change in activity and hence increased rate to return to the adapted activity.
Our model for precise adaptation predicts the data collapse of adaptation time-courses, allowing the convenient study of the adaptation dynamics (Fig. 3A). Specifically, the data collapse allows to evaluate the effects of ligand flow and adaptation dynamics, as well as imprecise adaptation. We found that adaptation to large concentration step changes is significantly imprecise (see Supplementary Text S1 for further details). We also extracted the kinetics of the receptor methylation level from experimental time courses from the data collapse (Fig. 3B), which is difficult to measure directly when relying on the quantification of the receptor methylation level using standard biochemical methods [25], [36]. According to our model, the activity-dependence of the receptor methylation level is a monotonously decreasing function of the receptor complex activity. Ultimately, this kinetics determines the compromise between long memory of previous concentrations and quick recovery for sensing new concentration changes [14]. Furthermore, we experimentally tested two predictions to validate our adaptation model. We analyzed the effect on the adaptation dynamics when changing the adapted receptor activity, as well as introducing a non-regulatable CheB mutant to remove the negative feedback from phosphorylation of CheB by the kinase CheA. In both cases, our model is consistent with experimental measurements (Fig. 5), supporting the finding of multiple layers of feedback regulation in adaptation.
While the MWC model is relatively well established [24], [41]–[44], we also considered alternative models for receptor signaling. These include a phase-separation model with mixed complexes separating into homogeneous complexes of Tar and Tsr at high ambient concentrations, as well as a lattice model with finite coupling between neighboring receptors (see Supplementary Text S1). Lattice models were previously suggested [54], [55], including a lattice formed by coupled CheA molecules [56], but were found to be inconsistent with FRET data [57]. We found that the dynamic MWC model describes dose-response curves far better than the alternative receptor signaling models investigated, particularly the reduced response amplitudes upon removal of attractant. Furthermore, the data collapse we introduced in this paper enabled us to compare different adaptation models proposed in the literature with FRET time-course data (Fig. 4). We found that while several models are consistent with the data, our model compared most favorably with the data.
We chose a simple model for adaptation with very few fitting parameters to explain the observed asymmetry in adaptation time-courses, i.e. slow adaptation to addition and fast adaptation to removal of attractant. Compared to the static MWC model, there are minor discrepancies between our model and experimental addition dose-response curves (Fig. 1). However, these can be rectified by refitting the dynamic MWC model by adjusting adaptation rates and receptor complex size simultaneously (see Supplementary Text S1), or by choosing an adaptation model with a more complex activity dependence. It should also be noted that adaptation rates needed to accurately describe dose-response curves are larger than those found when fitting the adaptation dynamics to the data collapse. This discrepancy may in part be due to using only a single set of experimental data for the data collapse, while dose-response curves were averaged over at least three sets. In addition, more complex processes not taken into account in our simple adaptation model, e.g. limited supply of substrate (methionine) for methylation, or the binding and unbinding kinetics of ligand, may be important for describing the dynamics.
Although our adaptation model is independent of biochemical details, our predicted fast demethylation at high activities may be due to cooperativity of two CheB-P molecules. According to in vitro experiments, CheB-P binding to the carboxyl-terminus of chemoreceptors induces an allosteric activation of the receptor, increasing the demethylation rate [40]. However, in contrast to CheR, CheB-P binds only weakly to the tether [40]. Reconciling these two observations, it is conceivable that two CheB-P molecules are necessary for efficient demethylation at high activities: one CheB-P molecule may bind to a tether to allosterically activate a group of receptors (assistance neighborhood), while another CheB-P molecule demethylates the receptors in the neighborhood. As two CheB-P molecules are not required to bind to the same receptor, this mechanism is not contradicted by the small number of CheB molecules in a cell. An alternative, simpler mechanism for cooperativity is dimerization of CheB-P molecules, which, however, has not been observed experimentally [22], .
Our adaptation model likely also applies to attractants other than MeAsp, since the dynamics of adaptation only depend on the activity of receptor complexes, independent of the details of external ligand concentration. According to the MWC model, different attractants (or their mixture) are integrated at the level of the free-energy of a receptor complex, which determines its activity. However, the imprecision of adaptation we found in FRET time courses at large MeAsp concentrations is in contrast to earlier experiments, which showed that the frequency of tumbling adapts precisely to aspartate, but not serine [28], [29]. The imprecision in adaptation to serine is readily explained if the number of Tsr receptors is larger than the number of Tar receptors per complex, since the available receptor methylation sites in a complex are more quickly used up in response to serine binding to Tsr receptors [42], [49]. However, the ratio of Tar and Tsr per complex is strongly dependent on the growth conditions, making a definitive conclusion difficult [59]. Future experiments may show if the imprecision observed in adaptation to MeAsp in FRET is reflected also in the tumbling frequency, or if imprecise adaptation is compensated for in order to yield perfect adaptation at the behavioral level.
Two different wild-type strains of E. coli were used. Wild-type strain 1 (WT1) is VS104 (cheY cheZ) that expresses the FRET pair consisting of CheY-YFP (YFP; yellow fluorescent protein) and its phosphatase CheZ-CFP (CFP; cyan fluorescent protein) from a pTrc-based plasmid pVS88, which carries pBR replication origin and ampicillin resistance and is inducible by isopropyl -D-thiogalactoside (IPTG) [23]. Wild-type strain 2 (WT2) is VS124 (cheB cheY cheZ) transformed with pVS88 and pVS91, which carries pACYC replication origin and chloramphenicol resistance and encodes wild-type CheB under control of pBAD promoter inducible by L-arabinose. The CheB-mutant strain is VS124 (cheB cheY cheZ) transformed with pVS88 and pVS97, which is identical to pVS91 except it encodes the non-regulatable CheBD56E. The D56E mutation was introduced into CheB by PCR. It prevents CheB phosphorylation, but allows protein to retain basal level of activity. Cells were grown at 275 rpm in a rotary shaker to mid-exponential phase () in tryptone broth (TB) medium supplemented with ampicillin, chloramphenicol, and IPTG. WT and CheB mutant strains were induced by 0 and 0.0003% arabinose, respectively, to produce comparable kinase activity (as assessed by the change in the level of FRET upon saturating stimulation). The CheB protein level was estimated using Western blots with CheB antibodies, and was at approximately 0.5-fold (WT2) and approximately 5-fold (CheBD56E mutant) the native level of CheB.
The experimental procedures follow those detailed by Sourjik and Berg [23]. Cells were tethered to a cover slip, and placed in a flow chamber. Cells were subject to a constant fluid flow of buffer or MeAsp at indicated concentration (flow speeds for WT1, and for WT2 and CheB mutant, respectively). Concentration step changes were achieved by abruptly switching between buffer and MeAsp, or different MeAsp concentrations. Fluorescence resonance energy transfer (FRET) between excited donor, CheZ-CFP, and acceptor, phosphorylated CheY-YFP, in a population of 300–500 cells was monitored using an epifluorescence microscopy setup. Emission light from CFP and YFP was collected and their intensity ratio was used to calculate the time-dependent number of interacting FRET pairs of CheZ-CFP and phosphorylated CheY-YFP in the cell population, which reflects the intracellular kinase activity [23]. The number of FRET pairs normalized by its adapted pre-stimulus value (after adaptation to the ambient concentration, but before concentration step changes) was calculated from the ratio according to [23]. The parameters and are the ratio for a saturating dose of attractant and the pre-stimulus value, respectively, both of which are measured in each experiment. The fluorescence efficiency ratio is determined by the experimental setup [60], and was 0.43 () for WT1 (WT2 and CheB mutant) experiments. FRET measurements were taken with a time resolution of 0.2 s (1 s) for WT1 (WT2 and CheB mutant).
This model describes the response of adapted mixed receptor complexes to instantaneous MeAsp concentration step changes [24], [43], [44]. According to this model, the activity of a mixed receptor complex is given by , where(5)is the free-energy difference between the on and off states of the complex. The indexes and denote Tar and Tsr receptor, respectively. We assumed fractions of Tar and Tsr in a complex according to their wild-type ratio, . The ligand dissociation constants for MeAsp are , , , and [43]. The free-energy contribution is attributed to methylation, and was recently extracted from dose-response curves for mutants resembling fixed methylation states [41]. We used the interpolating function (for data and fit see Inset of Fig. 2A). All energies are measured in units of ( being the Boltzmann constant and the absolute temperature). Exponential rate constants for the ligand flow were obtained from fits to ligand flow profiles (cf. Fig. 2B), with and for flow speed , and and for flow speed . The receptor complex size was estimated from least-squares fits to individual addition dose-response curves corresponding to specific ambient concentrations (and therefore adapted methylation levels). Note that complex size for removal may be different for each data point as cells are adapted to the increased concentration after each step change. The complex size grows with ambient concentration [41], [52] in a roughly linear fashion, with and . Both, individually fitted values, as well as the fitting function , are shown in Fig. 2A. We assumed an adapted receptor complex activity for WT1 as assessed from the maximum and minimum values of the experimental dose-response data in Fig. 1. Steady-state activities for WT2 and CheB mutant were estimated to be . For comparison of model and data, we normalized the receptor-complex activity for WT1, WT2 and CheB mutant by their respective activities when adapted to ambient concentration.
The dynamic MWC model combines the static MWC model with a dynamical model for adaptation. In our model for precise adaptation, the rate of change of the average receptor methylation level is given by (Eq. 2)The methylation and demethylation rates do not depend directly on the concentration of MeAsp or the methylation level, only on the receptor complex activity . Such dynamics leads to precise adaptation of the receptor complex activity to adapted level for a constant MeAsp stimulus [42], [45]. This model takes into account the receptor-activity dependence of the methylation and demethylation rates, as well as additional layers of feedback regulation for demethylation by CheB, including the activation of demethylation enzyme CheB by phosphorylation and potential cooperativity between phosphorylated CheB molecules. For Fig. 1–3, we determined the demethylation rate from a least-squares fit to addition and removal dose-response curves (WT1) using the receptor complex size from the static MWC model. The methylation rate is given by the condition that at steady state () the activity equals . The fit to the data collapse in Fig. 4 resulted in (and ), used in Fig. 4 and 5 for WT1. For WT2 in Fig. 5A, we used the same methylation rate constant as for WT1, but adjusted the demethylation rate constant to account for the increased adapted activity . For the CheB mutant in Fig. 5C, the rate of change of the average receptor methylation level is predicted to be(6)where we assume that the methylation rate is the same as for wild-type cells. The demethylation rate constant includes the basal activity of non-phosphorylatable CheB. Hence, the only dependence of the demethylation rate on receptor complex activity is due to binding of CheB to active receptors.
We incorporate the effect of imprecise adaptation, as suggested by time courses (cf. Fig. 2C), by making methylation and demethylation rates for wild-type cells (WT1) depend on the methylation level [49](7)The parameter is the maximum number of methylation sites per receptor, is the lower bound for the number of sites, which need to be available for efficient methylation or demethylation. We use to only allow Tar (not Tsr) receptors to become methylated (the total number of methylation sites of a receptor homodimer being 8). Further, we use to implement reduced efficiency of methylation or demethylation at a low number of available sites. Figure 2C shows time courses for adaptation to two concentration step changes using the precise and imprecise adaptation model ( and are the same in both models). The imprecise adaptation model fits the time courses shown far better. However, there is a large variability of imprecision seen in different data sets and more experiments are needed to produce a general model of imprecise adaptation.
To calculate the rate of activity change, the time courses for adaptation to step concentration changes were smoothed by averaging every 20 subsequent data points starting approximately 10 s after the step onset. The derivative was approximated by the difference quotient.
|
10.1371/journal.ppat.1000110 | RNAi Screen of Endoplasmic Reticulum–Associated Host Factors Reveals a Role for IRE1α in Supporting Brucella Replication | Brucella species are facultative intracellular bacterial pathogens that cause brucellosis, a global zoonosis of profound importance. Although recent studies have demonstrated that Brucella spp. replicate within an intracellular compartment that contains endoplasmic reticulum (ER) resident proteins, the molecular mechanisms by which the pathogen secures this replicative niche remain obscure. Here, we address this issue by exploiting Drosophila S2 cells and RNA interference (RNAi) technology to develop a genetically tractable system that recapitulates critical aspects of mammalian cell infection. After validating this system by demonstrating a shared requirement for phosphoinositide 3-kinase (PI3K) activities in supporting Brucella infection in both host cell systems, we performed an RNAi screen of 240 genes, including 110 ER-associated genes, for molecules that mediate bacterial interactions with the ER. We uncovered 52 evolutionarily conserved host factors that, when depleted, inhibited or increased Brucella infection. Strikingly, 29 of these factors had not been previously suggested to support bacterial infection of host cells. The most intriguing of these was inositol-requiring enzyme 1 (IRE1), a transmembrane kinase that regulates the eukaryotic unfolded protein response (UPR). We employed IRE1α−/− murine embryonic fibroblasts (MEFs) to demonstrate a role for this protein in supporting Brucella infection of mammalian cells, and thereby, validated the utility of the Drosophila S2 cell system for uncovering novel Brucella host factors. Finally, we propose a model in which IRE1α, and other ER-associated genes uncovered in our screen, mediate Brucella replication by promoting autophagosome biogenesis.
| Brucella spp. are facultative intracellular pathogens that cause brucellosis in a broad range of hosts, including humans. Brucella melitensis, B. abortus, and B. suis are highly infectious and can be readily transmitted in aerosolized form, and a human vaccine against brucellosis is unavailable. Therefore, these pathogens are recognized as potential bioterror agents. Because genetic systems for studying host–Brucella interactions have been unavailable, little is known about the host factors that mediate infection. Here, we demonstrate that a Drosophila S2 cell system and RNA interference can be exploited to study the role that evolutionarily conserved Brucella host proteins play in these processes. We also show that this system provides for the identification and characterization of host factors that mediate Brucella interactions with the host cell endoplasmic reticulum. In fact, we identified 52 host factors that, when depleted, inhibited or increased Brucella infection. Among the identified Brucella host factors, 29 have not been previously shown to support bacterial infection. Finally, we demonstrate that the novel host factor inositol-requiring enzyme 1 (IRE1) and its mammalian ortholog (IRE1α) are required for Brucella infection of Drosophila S2 and mammalian cells, respectively. Therefore, this work contributes to our understanding of host factors mediating Brucella infection.
| Infectious diseases caused by intracellular bacterial pathogens are responsible for an enormous amount of worldwide pain, suffering, and mortality. Brucella spp., for example, cause brucellosis, a global zoonosis of profound importance [1],[2]. Brucella melitensis, B. abortus, and B. suis are highly infectious and can be readily transmitted in aerosolized form [3],[4]. In addition, they have eluded systematic attempts at eradication for more than a century, even in most developed countries, and a human vaccine against brucellosis is not available [3]. Therefore, Brucella spp. have been classified as potential bioterror threat agents [5], and have generated significant interest in the biosecurity and world health communities.
Understanding the molecular mechanisms of Brucella pathogenesis and host response is critical for brucellosis control, and intracellular trafficking and replication of Brucella spp. play important roles in these processes [6]–[8]. First, bacteria, internalized from the host cell plasma membrane, orchestrate the biogenesis of early Brucella-containing vacuoles (BCVs) [9],[10]. Next, BCVs acidify but nevertheless fail to accumulate mannose 6-phosphate receptors (M6PRs) and cathepsin D, markers for late endosomes and lysosomes, respectively [8],[11]. Instead, maturing BCVs fuse with membranes that contain endoplasmic reticulum (ER) resident proteins, including calreticulin and calnexin [7],[8],[11]. In addition, this trafficking involves BCV interactions with a compartment that contains the autophagosomal marker monodansylcadaverin [7],[12]. Finally, Brucella spp. replicate in an ER-like compartment, and then presumably lyse the host cell to allow the infectious cycle to begin anew [8],[13],[14].
Bacterial lipopolysaccharides (LPS) play an important role in directing the bacterium along an intracellular trafficking pathway that enables a productive infection to be established. Brucella LPS also protects the bacterium from the harsh intracellular environment, suppresses pro-inflammatory and antibacterial host responses, and interferes with antigen presentation in macrophages [15]. Unlike their smooth wild-type (WT) counterparts, B. melitensis or B. abortus mutants harboring a deletion in the phosphomannomutase gene (ΔmanBA) lack LPS O-antigens, form rough colonies on solid medium, and are rapidly internalized by macrophages via a poorly understood pathway [16],[17]. However, these mutants fail to establish an intracellular replicative niche and reportedly induce a necrotic cytopathic effect in these cells [18],[19]. The bacterial type IV secretion system (T4SS) is also important for bacterial pathogenesis, and mutant strains lacking this system fail to traffic to, or replicate in, the ER [7], [20]–[22].
To date, relatively few host factors, including Rho1, Rac1, Cdc42 [23] and Sar1 [8], have been shown to be important for Brucella infection. Phosphoinositide 3-kinase (PI3K) activities have also been implicated in supporting Brucella infection [23]. Despite these advances, factors that mediate Brucella infection of host cells remain obscure. However, Brucella intracellular trafficking from the plasma membrane to an ER-associated replicative niche involves interactions with a membrane bounded compartment that contains autophagosome markers [7],[12]. In addition, the organism replicates within a compartment that contains ER resident proteins [7],[8],[11]. These data thereby suggest that host cell autophagic pathway proteins, and ER-associated factors, may regulate the intracellular trafficking and replication of the pathogen.
Recent developments in the use of evolutionarily divergent Drosophila S2 cell model systems to study host-pathogen interactions, and RNA interference (RNAi) technology for knocking down host gene expression, have provided unprecedented opportunities for making significant progress in elucidating Brucella host factors. Drosophila S2 cells are macrophage-like cells that recapitulate conserved aspects of innate immunity [24] and that have been exploited for studying mammalian host-pathogen interactions. RNAi-based forward genetic screens in S2 cells have, for example, identified novel host factors involved in the recognition and replication of significant human bacterial pathogens, including E. coli [25], Listeria [26],[27], Mycobacterium [28], Legionella [29], and Chlamydia [30],[31]. Importantly, mammalian orthologs of hits identified in these screens have been shown to be important for bacterial infection of mammalian cells, thereby validating the utility of this Drosophila cell model for host-pathogen studies [28]–[31]. In this study, we show that the Drosophila S2 cell-Brucella interaction system recapitulates critical aspects of Brucella infection of mammalian cells. In addition, we demonstrate the power of this system by identifying novel Brucella host factors, including IRE1α, a conserved transmembrane kinase that plays a key role in regulating the host cell unfolded protein response (UPR) [32]–[34]. Finally, we demonstrate that IRE1α is required for Brucella infection of mammalian cells, and discuss a possible mechanism by which this intriguing protein may regulate bacterial infection.
If Drosophila S2 cells are to provide a model system for studying Brucella infection, then they must support bacterial entry and replication. In addition, isogenic Brucella mutants with established entry, intracellular trafficking and replication properties should behave similarly in S2 cells and mammalian macrophages. Finally, Brucella should display similar infection phenotypes in S2 and mammalian cells that have been treated with compounds that disrupt host cell functions. With these ideas in mind, we employed gentamicin protection assays [18] to examine the entry and replication of different B. melitensis and B. abortus WT and mutant strains (listed in Table S1) in S2 cells. Because S2 cells require temperatures below 30°C for growth, all infection experiments were performed at 29°C, unless otherwise indicated. Importantly, J774A.1 cells supported Brucella entry and intracellular replication at this temperature (Fig. S2).
Brucella WT (S2308 and 16M) and mutant strains displayed strikingly similar properties when infecting S2 and mammalian cells. First, B. abortus and B. melitensis strains with smooth colony morphologies (i.e., 102B2, 146D5, BA114, S2308ΔvirB2) (Fig. 1 A1 and A2) and attenuated rough mutants (i.e., CA180, S2308ΔmanBA and 16MΔmanBA) displayed corresponding entry phenotypes in Drosophila S2 and mammalian cells (Fig. 1 A2 and data not shown). Second, B. melitensis strains harboring mutations in mucR (strain 102B2) and merR (strain 146D5) failed to replicate in both J774A.1 [35] and S2 cells. B. abortus and B. melitensis strains lacking the T4SS (e.g., BA114, S2308ΔvirB2, 16MΔvirB2) behaved similarly (Fig. 1 A3, A4 and data not shown). Third, vaccine strains RB51 and S19 [36] displayed significantly decreased levels of replication in both host cell systems (Fig. 1 A4 and data not shown). Fourth, similar cytopathic effects were observed when rough strain CA180 infected S2 and J774A.1 cells [18],[19] (Fig. 1B and 1C). Finally, the number of bacteria that entered S2 cells was directly proportional to the multiplicity of infection (MOI) (Fig. S3). This feature was also observed in mammalian cell systems (data not shown).
To easily visualize the intracellular trafficking and replication of Brucella spp., we exploited a GFP-expressing 16M strain (henceforth 16M-GFP) (Fig. S4). A comparison of the intracellular trafficking of Brucella spp. in S2 and mammalian cells indicated that the pathogen follows similar pathways in both host cell systems. BCVs trafficked to and replicated within an intracellular compartment that contained ER markers (e.g., mSpitz in S2 cells) [37], and was closely associated with COPII-coatomer (Sec 23) proteins (Fig. S5A and data not shown) in both cell systems. Quantitative analysis also demonstrated that the bacterium failed to accumulate late endosome, Golgi marker (dGRASP) [38], or lysosomal markers in S2 or mammalian cells ([7],[8],[12] and Fig. S5B). In addition, heat killed, formaldehyde fixed, and ΔvirB controls did not similarly colocalize with ER markers in either system (Fig. S5A and data not shown). Therefore, the intracellular trafficking of B. abortus and B. melitensis in S2 and mammalian cells shared striking similarities.
Similar infection profiles were observed when B. abortus was used to infect mammalian or S2 cells that were treated with several compounds; these compounds disrupted host cell functions and did not impair the bacterial growth in culture, or the viability of infected S2 cells (Fig. S6). These included: cytochalasin D [23], a compound that disrupts actin polymerization; bafilomycin A1, a specific inhibitor of vacuolar H+-ATPase activity and endolysosomal acidification [39]; brefeldin A (BFA), a fungal metabolite that prevents the assembly of COPI coated vesicles and disrupts vesicular transport [7],[8] (Table S2 and Fig. S7A and S7B). Treatment of S2 and J774.A1 cells with the PI3K inhibitor wortmannin (WM) significantly reduced entry of B. abortus and B. melitensis (Fig. S7A and data not shown). However, WM treatment of S2 and J774.A1 cells had no effect on the replication efficiency of the internalized bacteria (Fig. S7C, and data not shown). These findings were similar to those previously reported in mammalian cell systems [8],[23],[39],[40]. In addition, we performed several experiments to assess the role of sphingolipids in supporting bacterial infection, and exploited myriocin (MR), a potent inhibitor of serine palmitoyltransferase (SPT), the first step in sphingosine biosynthesis [41], for these studies. B. abortus entry and survival were significantly inhibited when cells were treated with high MR concentrations (≥1 μM). Low concentrations (≤100 nM) of the compound had no effect on bacterial entry (Table S2 and Fig. S7A). However, the replication efficiency of the pathogen was decreased under these conditions (Fig. S7A).
We employed RNAi technology to examine whether host proteins that are known to support bacterial infection of mammalian cells play similar roles in S2 cells. The evolutionarily conserved host proteins Rho, Rac, Cdc42 and Sar1 have been previously shown to be required for Brucella infection of mammalian cells [8],[23], therefore, we examined whether these proteins were also required for Brucella entry and replication in S2 cells. Fluorescence microscopy image assays were employed for these studies because they offered a rapid and convenient method for assessing bacterial infection. Importantly, similar results were obtained when either fluorescence microscopy or gentamicin protection assays were performed (Table 1 and Fig. 2A). When S2 cells were depleted of Rac and Cdc42, the entry of B. abortus (S2308) or B. melitensis (16M) was impaired (Table 1 and Fig. 2A). Rho1-depleted S2 cells appeared larger than untreated controls, contained numerous enlarged intracellular vacuoles, and also displayed significantly decreased levels of Brucella entry (Table 1, Fig. 2B). Sar1-depleted S2 cells also displayed dramatically reduced levels of Brucella replication (Table 1, Table S3 and Fig. 2A) were observed in these cells. These findings were similar to results obtained when B. abortus was used to infect mammalian cells in which the activities of the corresponding human orthologous proteins had been depleted [8],[23]. Therefore, the activities of these evolutionarily conserved GTP-binding proteins were required to support bacterial infection of both S2 and mammalian cells (Table 1, Table S3, Fig. 2 and Fig. S8).
To assess whether PI3Ks played similar roles in supporting bacterial infection of mammalian and S2 cells, we performed several experiments. First, we treated S2 and J774A.1 cells with WM and found that the levels of B. abortus and B. melitensis entry decreased in a similar fashion in both host cell systems (Table S2, Fig. S7A and data not shown). Second, we employed RNAi technology to deplete S2 cells of individual PI3K proteins and then measured bacterial entry and replication. These experiments revealed that multiple classes of PI3Ks are required to support B. abortus and B. melitensis WT strain infection (Table 1, Table S3, Fig. 2A, Fig. 3A and 3B). However, rough and smooth strains exploit separate host molecular pathways for entry [42]. when B. abortus rough strain CA180 was used to infect PI3K-depleted S2 cells, bacterial entry was dramatically enhanced (Fig. 3C). These data indicated that multiple PI3Ks play differential roles in mediating the entry of smooth and rough Brucella strains into S2 cells.
If Drosophila S2 cells are to serve as a useful model host cell system, then results obtained using this system should mirror corresponding mammalian cell findings. To test this possibility, we examined whether a murine ortholog (p85) of a model Drosophila gene (Pi3K21B) that supports Brucella infection of insect cells (Table 1, Table S3, Fig. 2A, Fig. 3A and 3B) mediates bacterial infection of murine cells. We used immortalized mouse embryonic fibroblasts (MEFs) derived from knockout mice harboring deletions in class IA PI3Ks (p85α and p85β) [43] for these studies. As expected, the levels of B. abortus and B. melitensis WT strains entry into MEF cells harboring PI3K gene deletions were dramatically reduced (Table 1 and data not shown). p85α−/− p85β−/− and p85β−/− MEFs supported lower levels of B. abortus and B. melitensis WT strains entry than p85+/+ controls (Table 1 and data not shown). However, when these MEFs were infected with Brucella rough mutants (CA180 and S2308ΔmanBA), bacterial internalization significantly increased, especially in p85β−/− MEFs (Fig. 3C and data not shown). These findings were similar to results obtained in experiments in which the entry of a Brucella rough mutant into class IA PI3K-depleted S2 cells was examined (Fig. 3C). Therefore, host cell PI3K isoforms differentially mediated the infection of smooth and rough organisms in both cell systems, and supported the use of the Drosophila cell system for elucidating novel Brucella host cell factors.
We were encouraged by our findings that previously described mammalian host proteins (i.e., Rho1, Rac, Cdc42 and Sar1) played similar roles in S2 cells. In addition, we noted that the Drosophila S2 cell system enabled the first molecular dissection of host cell PI3K isoform activity during Brucella infection. We therefore examined whether the Drosophila S2 cell system and RNAi technology could be combined to identify novel Brucella host factors. To focus our experiments, we constructed and screened 240 dsRNAs, including 110 dsRNAs that targeted the knockdown all of the genes annotated to be associated with the ER in the Drosophila RNAi Library Release 1.0 (Open Biosystems, Huntsville, AL, USA). The ER was ripe for examination because Brucella is known to replicate within a poorly characterized ER-like compartment, thereby suggesting that ER-associated host factors may be involved in regulating the intracellular replication of the pathogen.
Our ER-directed RNAi screen gave several interesting results. First, our screening approach successfully identified 52 hits. A hit was defined as a sample in which the relative infection differed by more than two standard deviations from the untreated control (Table S3). Importantly, control genes (i.e., Rho1, Rac, Cdc42, Sar1 and PI3Ks) were identified as hits in the screen (Table S3). Therefore, our screening strategy was sufficiently robust to uncover known or suspected host factors. We were curious whether the hit frequency obtained in our ER-targeted screen would be the same if a set of dsRNAs that were not associated with the ER were screened. We therefore screened 130 dsRNAs that were randomly picked from 2 of the 76 96-well plates in the Drosophila RNAi library. Because the manufacturer randomly arrayed dsRNAs into the source plates, this strategy for picking dsRNAs to be screen introduced no bias in the functions of the targeted genes in the screen. Notably, this experiment uncovered only 2 hits (∼1.5% of the total) (Table S3), and therefore gave a hit frequency that was comparable to that observed in the Mycobacterium fortuitum and Listeria monocytogenes whole genome RNAi screens [26]–[28]. Interestingly, 14 out of 52 hits in our screen had been previously shown to mediate infection of S2 cells by Mycobacteria, Listeria, Legionella and Chlamydia infection [26]–[31] (Table S3 and Fig. 4). On the other hand, 29 genes were identified that had not been previously reported to be involved in supporting intracellular bacterial infection (Table S3). These novel genes were classified according to the gene ontology system of biological and molecular function, cellular component, or protein domains as reported in FlyBase (www.flybase.org). This classification revealed that the novel hits represented a variety of functional classes, including kinases, chaperones, and biosynthetic/metabolic enzymes. In addition, these 29 genes were localized to either the ER lumen (CG9429, CG30498) or ER membrane (CG6437, CG1063) (Table S3). We re-tested some of our most interesting hits in both fluorescence microscopy and gentamicin protection assays (Table S3, repeat≥3 times), and also employed quantitative reverse transcriptase polymerase chain reaction (Q-PCR) to verify that the expression of these genes in S2 cells was knocked down by dsRNA treatment. We typically obtained 60–90% knockdown of target gene expression in our screening plates (Fig. 2C and data not shown).
Although each screen hit constituted a potential entry point for investigating the mechanism by which Brucella secures a replicative niche, we were particularly intrigued with IRE1 (CG4583), a key signal transducer that plays an important role in regulating the host cell UPR [32]–[34]. RNAi mediated knockdown of IRE1 gene expression resulted in significant reductions in Brucella replication (Fig. 5A, 5B and Table S3). In addition, IRE1 had not been previously implicated as a bacterial host factor. These data raised the intriguing possibility that IRE1 may play a novel role in regulating Brucella infection. We therefore examined whether IRE1α (the mammalian ortholog of Drosophila IRE1) was important for Brucella infection of mammalian cells.
We performed several experiments to examine whether IRE1α played a critical role in supporting Brucella infection of mammalian cells. First, we infected IRE1α-null (IRE1α−/−) and WT (IRE1α+/+) control MEF cells with 16M-GFP (Fig. 5C), and also performed gentamicin protection assays to assess bacterial entry and replication (Fig. 5D). The level of bacterial entry in IRE1α−/− was not statistically different from IRE1α+/+ controls (Fig. 5D). However, bacterial replication was significantly inhibited in IRE1α-depleted S2 cells and in IRE1α-null MEF cells (Fig. 5). Trypan blue dye exclusion analysis of 16M-infected MEF cells failed to reveal differences in host cell survival (data not shown). Therefore, the differences in bacterial replication efficiencies in these cell lines were not caused by the induction of host cell pro-apoptotic programs or by differences in the survival of Brucella-infected IRE1α−/− MEFs. Instead, they appeared to reflect a specific and important bacterial requirement for host cell IRE1α activity. Finally, the levels of entry and replication of Salmonella enterica serovar typhi, and the amounts of latex bead internalization, were similar in control and IRE1α−/− cells (Fig. 6 and data not shown). These data supported the idea that IRE1α−/− cells do not possess general defects in phagocytosis, and that IRE1α activity is not required to support infection by all intracellular bacterial pathogens (Table S3 and Fig. 6).
Besides IRE1α, several other ER-associated transmembrane signaling molecules play important roles in initiating and regulating UPR in host cells, including PERK, ATF6, and BBF2H7 (mammalian BBF-2 ortholog) [32]–[34], [44]–[46]. We performed fluorescence microscopy and gentamicin protection assays to examine their role in Brucella entry and replication. B. melitensis (16M) entry and replication in PERK-, ATF6-, and BBF-2-depleted S2 cells were not significantly different from untreated controls (Fig. 5A, Fig. 7A and Table S3). Although B. melitensis replication efficiency decreased in ATF- and BBF-2-depleted S2 cells, the number of bacterial colony forming units (CFUs) in these cells at 72 hours post-infection (h.p.i.) was not significantly smaller than controls (Fig. 7A and Table S3). Importantly, bacterial replication efficiencies in PERK−/− and PERK+/+ MEF cell lines were similar (Fig. 7B and 7C). Taken together, these results demonstrated that not all UPR signaling molecules are required to support the replication of this pathogen, and that IRE1α plays a specific role in this process.
The study of host-Brucella interactions has suffered from the absence of a tractable genetic system to elucidate host factors. However, data obtained in this study indicate that Drosophila S2 cells provide a compelling model system for identifying and characterizing these important proteins. Brucella infection of Drosophila S2 cells recapitulates important aspects of mammalian cell infection. First, isogenic mutants of Brucella spp. behaved similarly in S2 and mammalian cells. In addition, these divergent host cell systems displayed similar trends in infection by smooth and rough strains with varied pathogenicity. Brucella rough mutants, such as CA180, were cytopathic to both mammalian and S2 host cells [18],[19, this study]. Therefore, these cells share conserved molecular mechanisms for recognizing and responding to Brucella LPS mutants. Second, Brucella entry and replication in S2 and mammalian cells were similarly sensitive to pharmacological perturbation by structurally diverse compounds. Of particular interest was the observation that MR, an inhibitor of STP, the rate-limiting enzyme in sphingolipid biosynthesis, dramatically reduced the amount of Brucella infection of S2 cells (this study). Previous studies have demonstrated an important role for sphingolipid enriched lipid rafts in pathogen infection [47]–[50], and our MR experiments support these observations. Third, Brucella infection of S2 cells required the activities of conserved GTP-binding proteins (Rho1, Rac, Cdc42, and Sar1), suggesting that Brucella infection of mammalian and Drosophila cells shared similar host molecular requirements. Finally, the activities of PI3Ks differentially regulate smooth and rough Brucella infection in both mammalian and Drosophila S2 cells (this study). Interestingly, the effects of PI3K knockdown in MEF cells were more dramatic than in S2 cells. In MEFs, PI3K genes are deleted, and thus the corresponding enzyme activities are absent. However, in Drosophila S2 cells, PI3K gene expression is knocked down (60–90%), and some residual activity may remain. These differences likely account for the differential infection of these cell types. Taken together, our data support the conclusion that S2 cells provide a useful model for investigating host-Brucella interactions.
Our demonstration that Drosophila S2 cells can be used to illuminate Brucella host factors is surprising because Brucella spp. do not occupy a described environmental niche outside of the mammalian host. In addition, the bacteria do not grow well in culture at temperatures below 35–37°C. However, previous reports have demonstrated B. suis multiplication within U937 cells at 30°C [51]. Therefore, Brucella growth below 37°C is not restricted to B. melitensis and B. abortus strains. Second, Brucella replication in J774A.1 and Drosophila S2 cells at 29°C share similar kinetics (Fig. S2B and S2C). Although a difference in the replication efficiency of S2308ΔvirB2 in J774 and S2 cells at 24 and 48 h.p.i was observed, no difference was detected at 72 h.p.i. Therefore, the differential growth of B. abortus and B. melitensis in these host cell systems likely results from differences in the growth temperature, and not from differential subversion of conserved host cell functions. Third, the most important criterion for judging the utility of a model non-mammalian host-pathogen interaction system is whether it can be exploited to shed new insights into the interaction in mammalian cells. In this regard, it should be noted that bacterial pathogens, such as Listeria monocytogenes [52], grow more slowly in Drosophila S2 cells than in mammalian cells; however, many host factors required for entry and survival of these intracellular pathogens have been identified using Drosophila S2 cells as a platform [25]–[31]. We expect to garner similar insights through the use of our Drosophila S2 cell-Brucella interaction system, and our demonstration that PI3Ks and IRE1α mediate Brucella infection of Drosophila S2 cells and murine embryonic fibroblasts support this view.
Our RNAi screen in S2 cells for ER-associated Brucella host factors provides new insights into how Brucella secures an intracellular replicative niche. Our screen identified 52 genes that participate in this process, 29 of which had not been previously suggested to support bacterial pathogen infection. In addition, we dissected the role of 4 PI3K isoforms. The number of identified hits (50 out of 110 pre-selected ER- associated genes) was striking, and likely reflects that sustained and multi-faceted Brucella-ER interactions are required for Brucella replication in host cells. Interestingly, 14 of the genes identified in our screen were also required for infection of S2 cells by other intracellular bacterial pathogens, including Listeria, Mycobacteria, Legionella or Chlamydia [26]–[31]. The fact that Brucella and Legionella share several ER-associated host factors is perhaps not surprising, especially given that both organisms engage in sustained interactions with the host ER as part of their virulence and replication programs [53],[54]. Finally, Brucella-specific ER-associated factors, such as IRE1 (CG4583), were uncovered in our screen. IRE1 may constitute a species-specific host factor that plays a role in mediating the unfolded protein response, thereby suggesting that the modulation of this stress-response system may be critical to bacterial intracellular survival and replication.
In eukaryotic cells, IRE1α mediated UPR induction is associated with enhanced expression of genes encoding ER chaperones and protein-folding catalysts, and proteins that participate in ER-associated degradation (ERAD) [55],[56]. IRE1α activation also induces the biosynthesis of membrane phospholipids that increase the surface area and volume of rough ER [57],[58]. In Brucella infected cells, IRE1α mediated activity may result in the biosynthesis of ER membrane that can be exploited by the pathogen to expand the size and enhance the quality of its replicative niche However, our data indicate that other UPR signal transducers, including PERK, are not required for Brucella infection in both Drosophila S2 and murine embryonic fibroblast cell systems. Therefore, not all UPR regulatory proteins are important for bacterial replication (Fig. 7 and Table S3), raising questions about the privileged status of IRE1α among these classes of molecules.
Recent reports have indicated an intriguing link between IRE1α activity and autophagic vacuole biogenesis [59],[60]. For example, IRE1α is required for the autophagy observed after cells are treated with the ER stress-inducing agents DTT, tunicamycin or thapsigargin [59],[60]. However, parallel experiments using PERK-deficient cells, and cells in which the expression of ATF6 had been knocked down, demonstrated that these UPR-associated signal transducers are not directly involved in the response to these drug treatments [60]. Therefore, IRE1α can regulate some autophagic events independently from input by these other ER associated signaling molecules.
The differential participation of IRE1α, ATF6 and PERK in regulating the autophagy observed after cells are treated with stress-inducing agents is strikingly similar to their differential roles in mediating Brucella replication. IRE1α is required for Brucella to replicate efficiently; however, Brucella replication in PERK-, ATF6-, and BBF-2-depleted S2 cells was not significantly different from untreated controls. This differential participation therefore suggests a model in which IRE1α regulates Brucella infection by modulating the host cell autophagy pathway (Fig. 8).
Based on findings from our dsRNA screen, we propose a multi-step model by which IRE1α regulates Brucella replication. First, BCVs traffic to a compartment that contains ER resident proteins. Concomitantly, BCVs trigger IRE1α activation, which in turn, stimulates the biogenesis of ER-associated autophagosomes (ERAs) [59],[60]. ERAs then fuse with BCVs to form ERA-BCVs. This process is also regulated by the activities of PI3Ks. Finally, ERA-BCVs fail to fuse with lysosomes and hence avoid degradation; instead, they fuse with the ER to form ER-derived BCVs that are permissive for Brucella replication (Fig. 8).
Several pieces of evidence support this view. First, IRE1α, but neither PERK nor ATF6, is required for the induction of autophagy in response to treatment by ER stress-inducing agents [60]. Similar requirements for host proteins are observed during Brucella replication (Fig. 5, Fig. 7 and Table S3). Second, the assembly of ERAs is dependent upon early secretory pathway molecules [61]–[63]. In yeast, the COPII mutants sec16, sec23, and sec24, are defective in autophagy. However, mutations in two other COPII genes, sec13 and sec31, do not affect ERA biogenesis and autophagy [61],[62]. In addition, PI3K activity is important for this process [64]–[66]. Our data demonstrate similar host factor requirements during Brucella infection of Drosophila cells. Specifically, depletion of Sec23, Sec24 and PI3Ks in host cells dramatically reduces Brucella replication (Table 1, Table S3, Fig. 2 and Fig. 3). However, depletion of Sec31 has no affect on this process (Table S3). Finally, Brucella trafficking to its intracellular replicative niche involves interactions with a compartment that contains the autophagosomal marker monodansylcadaverin [7],[12]. These localization data thereby establish a physical interaction between internalized Brucella and the host cell autophagy pathway. It should be noted, however, that although we cannot rule out the possibility that Brucella trafficking in MEFs differs from professional phagocytes, Brucella trafficking in HeLa cells and phagocytes share striking similarities [12]. Therefore, our observations in MEFs likely shed light on Brucella infection of phagocytes. Taken together, the data are consistent with the idea that IRE1α activity plays an important role in supporting Brucella interactions with the host cell ERA biogenesis machinery in mammalian cells (Fig. 8). Future studies will exploit the genetic power of the Drosophila S2 cell system to elucidate this intriguing possibility, and to define the precise molecular mechanisms by which Brucella secures an intracellular replicative niche.
Brucella melitensis strain 16M (WT) and B. abortus strain 2308 (WT), and their derived mutants are listed in Table S1. Bacteria were grown in tryptic soy broth (TSB) or on tryptic soy agar (TSA, Difco™) plates, supplemented with either kanamycin (Km, 50 μg/ml), or chloramphenicol (Cm, 25 μg/ml) when required. For infection, 4 ml of TSB was inoculated with a loop of bacteria taken from a single colony grown on a freshly streaked TSA plate. Cultures were then grown with shaking at 37°C overnight, or until OD600≈3.0.
Murine macrophage J774.A1 cells, MEFs and HeLa cells were routinely cultured at 37°C in a 5% CO2 atmosphere in Dulbecco's Modified Eagle's Medium (DMEM) supplemented with 10% fetal bovine serum (FBS). S2 cells were maintained at 25°C in Drosophila-SFM medium or in Schneider's Drosophila medium (Invitrogen) supplemented with 10% FBS. Cells were seeded in 24-well plates and cultured overnight before infection. For antibiotic protection assays, 2.5×105 cells were seeded in each well; for fluorescence microscopy assays (see below), 5×104 cells were seeded on 12-mm glass coverslips (Fisherbrand) placed on the bottom of 24-well microtiter plates before infection.
Host cells were infected with Brucella at an MOI of 100, unless otherwise indicated. Infected cells were then incubated at 29°C (S2 cells) or 37°C (mammalian cells) after centrifugation for 5 min at 200×g. Thirty minutes post-infection, culture media was removed, and the cells were rinsed with 1×phosphate buffered saline (PBS). Fresh media, supplemented with 40 μg/ml gentamicin, was then added for 1 hr to kill extracellular bacteria. Infected cells were continuously incubated in this antibiotic for various lengths of time at the indicated temperature. As indicated, viable bacteria in infected cells were analyzed using the antibiotic protection assay or the immunofluorence microscopy assay described below. In addition, Brucella replication efficiency ([# of CFUs at different time points post infection]/[# of CFUs of Brucella entry]) in the infected cells was also determined
At various times post-infection, viable bacteria present in infected cells were analyzed using gentamicin protection assays [18]. Briefly, infected cells were washed twice with 1×PBS buffer, lysed with 0.5% Tween 20 in sterile water, and the released bacteria were subjected to serial dilution in peptone saline [1% (wt/vol) Bacto peptone and 0.85% (wt/vol) NaCl]. Next, 10 μl of serial diluted cell lysate was plated on TSA plates. Finally, CFU were counted after three days of incubation at 37°C.
S2 cells were coincubated with or without various drugs before 1 hr of and during Brucella infection (See below). Next, the infected cells were centrifuged at 200×g for 5 min and then incubated (30 min) with assorted Brucella strains. Fresh Drosophila-SFM media, supplemented with drugs (as indicated) and 80 μg/ml gentamicin, was added to kill extracellular bacteria. The infected and gentamicin treated cells were then incubated at 29°C for various lengths of time. To quantify the viability of S2 cells, at various time points, a portion of the infected cells was removed and processed for 0.2% trypan blue vital stain analysis. At least 500 cells were counted per sample. For image analysis, infected cells were replated onto ConA (Sigma)-coated 12-mm coverslips in 24-well plates and allowed to adhere for 1 hr. Cells were stained with 0.2% trypan blue for 5 min and then fixed with 1×PBS containing 3.7% formaldehyde for 1 hr. Viability of infected cells was assessed by analyzing images obtained with an Olympus IX70 fluorescence microscope. At least 500 infected cells per sample were used for the analysis.
To visualize Brucella spp. trafficking, S2 cells were transfected with ER marker mSpitz-GFP [37], and Golgi marker dGRASP-GFP [38] before infection. Specifically, S2 cells were grown to ∼80% confluence and then transfected using Effectene Transfection Reagent (Qiagen) as per the manufacturer's instructions. 0.25 μg of each pUAS-mSpitz GFP and pAcpA-Gal4 were employed in these transfection experiments. For the Golgi visualization experiments, 0.25 μg of dGRASP-GFP was used in the transfection. Typically, 1.5×106 cells were transfected and then grown in 2.2 ml of Schneider's Drosophila medium supplemented with 10% FBS. Three days post-transfection, cells were replated onto ConA-treated 12-mm glass coverslips placed on the bottom of 24-well microtiter plates (for early time points of less than 8 hr) and immunofluorescence microscopy analysis was performed as previously described [18]. For later times points (≥8 hr), the transfected cells were reseeded directly in 24-well plates and allowed to adhere for 2 additional hours before infection with Brucella. At different post-infection time points, the infected cells were replated onto ConA-coated 12-mm coverslips and allowed to adhere for 1 hr. The cells were then washed three times with 1×PBS, fixed with 3.7% formaldehyde (pH 7.4) at room temperature for 1 hr and processed for immunofluorescence microscopy.
To elucidate Brucella spp. intracellular trafficking, S2 cells were infected with the following strains: B. melitensis (strains 16M or 16M-GFP); B. abortus (strain S2308); S2308 virB2 deletion mutants; heat killed or 3.7% formaldehyde fixed WT strains. At various post-infection time points, S2 cells were replated onto ConA-coated 12-mm coverslips and allowed to adhere for 45 min to 1 hr. Cells were then washed, fixed as described above, and processed for immunofluorescence microscopy [18]. The primary antibodies used were as follows: goat polyclonal anti-Brucella; rabbit anti-human M6PR; rabbit anti-human cathepsin D; goat-anti rabbit Sec23 (COPII marker, Affinity BioReagents, Inc., CO, USA). Samples were stained with Alexa Fluor 488-conjugated and/or Alexa Fluor 594-conjugated donkey anti-goat/rabbit (Molecular Probes, 1:1000). Cover slips were then mounted in Vectashield mounting media (Vector Laboratories, Inc., CA, USA) and visualized with an Olympus BX51 confocal microscope. For quantitative analysis, single confocal section of random fields was acquired, and colocalization of markers was scored as positive when nonsaturated signals partially overlapped. Images for all immunofluorescence assays for Brucella spp. trafficking were acquired with a Hamamatsu ORCA-ER camera mounted on the Olympus BX51 microscope and driven by Simple PCI software (Compix Imaging Systems Inc., Cranberry Township, PA.). Images were processed with Adobe Photoshop CS Software (Adobe Systems Incorporated, San Jose, CA).
Drosophila S2 cells or J774.A1 murine macrophages were coincubated in 24 well plates with assorted drugs including bafilomycin A1 (BAF), brefeldin A (BFA), cytochalasin D (CD), myriocin (MR) and wortmannin (WM) at the indicated concentrations. Cells were treated with drugs 1 hr before, and during, infection with the indicated Brucella strains. After infection, the treated cells were incubated at 29°C (S2 cells) or at 37°C with 5% CO2 (J774.A1 macrophages). To evaluate Brucella internalization, after 30 min of infection, fresh media, supplemented with the same concentration of the drugs and 80 μg/ml gentamicin was added to kill extracellular bacteria. After 45 min of incubation, the cells were lysed and the CFU per well determined by plating dilutions on TSA plates as described above. To assess Brucella intracellular replication, CFU analysis was performed at 72 h.p.i. The effect that BAF-mediated inhibition of host cell endosomal acidification exerted on Brucella replication was also examined. Briefly, BAF was added to the culture media 2 h.p.i. and continuously coincubated with infected cells for 72 hr. Cells were lysed and analyzed using the gentamicin protection assay. To investigate whether the drugs inhibit Brucella growth, the drugs were individually added to Brucella TSB cultures at 29°C or 37°C and incubated for 1 and 72 hr. CFU plating was used to assess bacterial growth in the presence of drugs, and thereby to evaluate the potential inhibitory effects.
Primers for generating RNAi that target the knockdown of Drosophila Rac1, Rac2, Rho1, Cdc42, Sar1 and PI3Ks were designed using sequence information present in flybase (http://flybase.org/). The primers were used in RT-PCR reactions to generate cDNAs. dsRNAs targeting genes to be knocked down were generated using previously described methods [26]. Briefly, gene-specific RNAi primers were used to amplify target sequences from Drosophila cDNA mixtures. The PCR products were re-amplified using the RNAi primers with T7 RNA polymerase promoter sequences in the 5′ end. The reamplified PCR products were then used as templates for the generation of dsRNAs. For generation of dsRNAs targeting ER-associated and other genes, cDNAs from commercially available Drosophila RNAi Library Release 1.0-DNA templates (Open Biosystems, Huntsville, AL, USA) were directly used as templates. One or two microliters (total ∼150 ng) of the PCR products were used to perform in vitro transcription reactions with the T7 MEGAscript kit (Ambion, Austin, TX) as per the manufacturer's instructions. Aliquots of in vitro transcription products were subjected to quality control by 1% agarose gel electrophoresis analysis and dsRNA concentrations were quantified using a NanoDrop® ND-1000 UV-Vis spectrophotometer (NanoDrop Technologies, Inc. Wilmington, DE).
1.0×106 S2 cells were seeded in 12-well plates. dsRNAs (i.e., Rho1, Rac, Cdc42, Sar1 and PI3Ks) were added to each well at a final concentration of 15 μg/ml. After 4 days of incubation with dsRNA, an aliquot of the S2 cells was removed to check the efficiency of dsRNA mediated gene knock down by quantitative RT-PCR (Q-PCR). dsRNA-treated S2 cells in the same well were also re-plated in 24-well plates and allowed to adhere for at least 2 hr before infection. At the selected time points, the dsRNA-treated and Brucella infected cells were lysed and antibiotic protection assays or fluorescence microscopy image assays were performed as described.
To evaluate the utility of the combination of S2 cells and dsRNA technology, and the consistency of the results from antibiotic protection assays, we analyzed Brucella infection using fluorescence microscopy image assays. dsRNAs that target ER-associated genes or other known or unknown genes were added to 96-well microplates at a final concentration of 15 μg/ml (dsRNAs were added in duplicate in two different plates). S2 cells were then seeded in the plates with 5.0×104 cells/well in 200 μl Drosophila-SFM medium. dsRNA-treated cells were incubated at 25°C for 4 days to allow for knockdown of target gene expression. The dsRNA-treated cells (100 μl) were replated into 96 well plates, infected with B. melitensis 16M-GFP at an MOI of 50. After 30 min of infection, the same amount of fresh media supplemented with 80 μg/ml gentamicin was added to each well and the infected cells were incubated at 29°C. At 72 h.p.i., infected cells were replated onto 96 well glass bottom plates (Greiner), that had been coated with ConA, and allowed to adhere for 1 hr. The infected S2 cells were washed 3 times with 1×PBS, fixed with 3.7% formaldehyde in 1×PBS at 4°C overnight, and stained with phalloidin-Texas red (1:1000) for 1 hr to visualize the host cell actin cytoskeleton. Brucella infected S2 cells were viewed with an Olympus IX70 inverted microscope and two 400× images from each well were acquired for image analysis. Images were analyzed using NIH Image J software (http://rsb.info.nih.gov/ij/), and the relative infection (RIF) [100×(% of infected dsRNA-treated cells)/(% of infected cells in the untreated control)] was determined. More than 1,000 S2 cells were counted to obtain the percentage of infection or infection index [(number of infected cells (at least 10 brucellae within the cell))/(number of total cells)] in a sample. The detailed process by which image analysis was performed is shown in Fig. S1. dsRNA screen was repeated once, and some of hits identified in both two round of screens were picked out to re-test in triplicate in fluorescence microcopy and gentamicin protection assay as described above.
MEFs deficient of the two regulatory isoforms of class IA PI3Ks (p85α−/− p85β−/− and p85β−/−) [43], IRE1α (IRE1α−/−) [67] and PERK (PERK−/−) [68] and their corresponding WT control p85+/+, IRE1α+/+ and PERK+/+ MEFs, were seeded in 24-well plates. After overnight culture, cells were infected with 16M-GFP and/or S2308, and their derived mutant strains. Infected cells were centrifuged for 5 min at 200×g and then incubated at 37°C for 60 min. Cells were washed with 1×PBS buffer, and fresh media supplemented with 40 μg/ml gentamicin was added. Cells were incubated for an additional 1 hr (entry) and 48 hr (replication) at 37°C. The amount of viable bacteria present in infected cells was assessed using gentamicin protection assays. For fluorescence microscopy and viability assays, 5×104 cells were seeded onto 12-mm coverslips in 24-well plates. At 48 h.p.i., infected cells were subjected to the appropriate assays as described above.
All quantitative data were derived from results obtained in triplicate wells for at least three independent experiments. The significance of the data was assessed using Student's t-test, and all the analyzed data were normalized with internal controls before analysis.
|
10.1371/journal.pntd.0003948 | Fitness and Phenotypic Characterization of Miltefosine-Resistant Leishmania major | Trypanosomatid parasites of the genus Leishmania are the causative agents of leishmaniasis, a neglected tropical disease with several clinical manifestations. Leishmania major is the causative agent of cutaneous leishmaniasis (CL), which is largely characterized by ulcerative lesions appearing on the skin. Current treatments of leishmaniasis include pentavalent antimonials and amphotericin B, however, the toxic side effects of these drugs and difficulty with distribution makes these options less than ideal. Miltefosine (MIL) is the first oral treatment available for leishmaniasis. Originally developed for cancer chemotherapy, the mechanism of action of MIL in Leishmania spp. is largely unknown. While treatment with MIL has proven effective, higher tolerance to the drug has been observed, and resistance is easily developed in an in vitro environment. Utilizing stepwise selection we generated MIL-resistant cultures of L. major and characterized the fitness of MIL-resistant L. major. Resistant parasites proliferate at a comparable rate to the wild-type (WT) and exhibit similar apoptotic responses. As expected, MIL-resistant parasites demonstrate decreased susceptibility to MIL, which reduces after the drug is withdrawn from culture. Our data demonstrate metacyclogenesis is elevated in MIL-resistant L. major, albeit these parasites display attenuated in vitro and in vivo virulence and standard survival rates in the natural sandfly vector, indicating that development of experimental resistance to miltefosine does not lead to an increased competitive fitness in L. major.
| Cutaneous Leishmaniasis (CL) is characterized by the appearance of ulcerative lesions on the skin, and results from infection with trypanosomatid parasites such as Leishmania major. Current treatments for CL are expensive and have a wide range of toxic side effects of variable severity. Miltefosine, a recently introduced treatment option, is the first oral drug for leishmaniasis treatment. Although widespread clinical resistance has not yet been established, miltefosine-resistant parasite populations are easily created in a laboratory environment. Through step-wise selection, we have created populations of L. major resistant to miltefosine. These resistant parasites grow at a similar rate to miltefosine-sensitive parasites and exhibit similar stress responses. Accordingly, miltefosine-resistant parasites display a decrease in tolerance when selective pressure of MIL is withdrawn from the population. There is no conferred resistance to treatment with other antileishmanial agents, though increased sensitivity to alternative treatments is observed in some instances. Leishmania undergoes a complex life cycle including the differentiation to highly infective forms, in a process termed metacyclogenesis. Experimental resistance to miltefosine increases metacyclogenesis in L. major, however resistant parasites display a lower fitness than their sensitive counterparts, as judged by their attenuated virulence in vitro and in vivo.
| Leishmaniasis is caused by protozoan parasites of the genus Leishmania, and presents as a variety of clinical manifestations ranging from lesions on the skin to disseminated visceral infections [1]. Cutaneous leishmaniasis (CL) often results in self-resolving lesions, whereas visceral leishmaniasis (VL) is habitually fatal when left untreated. With an annual incidence of 2 million cases and a prevalence of more than 12 million, leishmaniasis is responsible for 70,000 deaths annually [2]. 88 countries have reported infection, resulting in 350 million individuals at risk for infection and an estimated 2.4 million disability-adjusted life years (DALYs) [2]. These statistics are grossly underestimated due to misdiagnosis and insufficient disease surveillance systems.
Leishmania species have a digenetic life cycle including both extracellular promastigote and obligate intracellular amastigote forms. Extracellular flagellated promastigotes reside in the midgut of the phlebotomine sandfly vector. Following infection in the mammalian host, promastigotes are engulfed by macrophages where they differentiate into non-motile amastigotes in the phagolysosome. This differentiation is triggered by environmental cues, mainly pH and temperature [3]. Current antileishmanial drugs include pentavalent antimony, amphotericin B, paromomycin, pentamidine, and miltefosine; most are toxic and expensive. To date, no successful vaccine exists, and the few antileishmanial drugs mentioned either risk becoming ineffective due to emerging resistance, or are limited in their use due to cost and parental administration [4, 5]. Miltefosine (MIL) is an alkylphosphocholine drug with demonstrated activity against various parasite species and cancer cells, as well some pathogenic bacteria and fungi [6]. Since its registration in 2002, miltefosine remains the only oral agent used for the treatment of all types of leishmaniasis. The U.S. Food and Drug Administration (FDA) recently (March 2014) approved Impavido (miltefosine) for the treatment of cutaneous, visceral and muco-cutaneous leishmaniasis. While the mechanism of action of MIL is not understood in its entirety, several studies have pointed at alterations in phospholipid metabolism, impairment of bioenergetic metabolism, and ultimately the induction of apoptosis as potential modes of actions [7–10]. Knowledge of experimental MIL resistance in Leishmania is limited to defects in drug internalization (defective inward translocation of MIL) and increased drug efflux [11]. Previous investigations in L. donovani have revealed the presence of several key point mutations in the P-type ATPase dubbed the LdMT (L. donovani miltefosine transporter) [12]. However, subsequent studies demonstrated that the LdMT alone was not sufficient to facilitate translocation, leading to the identification of the β-subunit LdRos3 and its importance to the function of the LdMT [13]. Mutations in the LdMT and Ros3 contribute to the MIL-resistant phenotype by significantly decreasing MIL uptake. Specifically, T420N and L856P mutations in the LdMT contributed to significantly decreased MIL uptake [12]. Other mutations identified in MIL-resistant L. donovani include W210 (LdMT) and M1 (LdRos3) [14]. Sequencing of the entire miltefosine transporter was performed in both L. major and L. infantum, and all identified sequence mutations differed from those previously detailed in L. donovani (L856P, T420N, W210, and M1) [15]. In the same study, no mutations were observed in the β-subunit Ros3 in any of the MIL-resistant populations. Widespread clinical resistance has not yet been demonstrated, nonetheless two L. infantum isolates from HIV co-infected patients have been reported to exhibit MIL resistance [16, 17]. The analysis of clinical isolates from patients infected with L. donovani that had relapsed to standard MIL therapeutic regimes demonstrated that the recovered parasites were significantly more tolerant to MIL [14]. None of the resistance markers i.e. point mutations aforementioned were found in the isolates. In the absence of a definitive mechanism of miltefosine resistance, the concept of fitness or “proficiency” of drug resistant pathogens is becoming more relevant and how the acquisition of resistance may impact the life cycle of the parasite, particularly its capacity to survive both in the insect and mammalian hosts and thus its ability to compete with wild type (sensitive) parasites [18–20]. Most of these studies are focused on antimony resistance in L. donovani and more recently, drug combinations [21]. Here we present the characterization and fitness of clonal lines of L. major that have experimentally acquired resistance to miltefosine, with relevance to survival in the mammalian host and phlebotomine vector.
All studies using vertebrate animals were conducted in accordance with the U. S. Public Health Service Policy on Humane Care and Use of Laboratory Animals and followed the standards as described in the Guide for the Care and Use of Laboratory Animals. Per these standards, all vertebrate animal studies were conducted following review by the University of Notre Dame Institutional Animal Care and Use Committee under protocol #15–047 (approved October 16, 2012). The University of Notre Dame is credited through the Animal Welfare Assurance #A3093-01.
Leishmania major strain Friedlin V1 (MHOM/JL/80/Friedlin) promastigotes were cultured at 27°C in M199 medium (medium 199 (CellGro) supplemented with 10% heat-inactivated fetal bovine serum (FBS), 20 mM HEPES, 10 mM adenine, penicillin/streptomycin, hemin, biotin, L-glutamine, and 7.5% NaHCO3) and passaged every 3–4 days. Macrophages (RAW264.7 cell line) were cultured at 37°C with 5% CO2 in RPMI supplemented with 10% heat-inactivated FBS, penicillin/streptomycin, and L-glutamine, and passaged every 2–3 days.
MIL-resistant cultures of L. major were generated using step-wise selection. Cultures were passaged every 3–4 days at an initial concentration of 5x105 promastigotes/mL. Increasing concentrations of MIL (Sigma) were introduced to the cultures beginning with 2.5 μM MIL and successively to 5, 8, 10, 15, 20, 30, and 40 μM MIL. Cultures were exposed to an increased concentration of MIL when growth rates were equivalent to the growth rate of the wild-type (WT). To account for clonal variation, 2 clones of each resistant line were generating by plating in M199 plates as previously described [22]. Clones 1 and 2 were simultaneously maintained.
Growth rates were measured for each set of resistant populations and compared with the WT strain. Parasites were counted at an initial concentration of 5x105 parasites/mL and growth was measured daily using a Neubauer chamber until the population reached stationary phase.
To further assess stability and fitness, two fluorescent FACS-based apoptotic markers were used to evaluate MIL-selection. Membrane permeability was assessed using the kit YO-PRO1 (Invitrogen) according to manufacturer’s recommendations. Briefly, samples were pelleted and washed in 1X M199 complete media. Following the wash, samples were resuspended in 1X M199 complete media and YO-PRO (Invitrogen) and Propidium Iodide (Invitrogen) were added and incubated for 20 minutes. Exposure of phosphatidylserine (PS) residues was investigated with Annexin-V-FITC (Miltenyi Biotec) following manufacturer’s instructions. Analyses were performed in a Beckman Coulter FC500 Flow Cytometer.
In order to assess the MIL-resistance achieved, the half-maximal effective concentration, EC50, was performed using the resazurin-based CellTiter-Blue (Promega) method as previously described [23]. Cultures were counted using a Neubauer chamber. 1x106 parasites/mL were incubated for 48 hours at 27°C in M199 medium (CellGro) and appropriate concentrations of MIL (Sigma), pentamidine isethionate (Sigma), amphotericin B (Sigma), potassium antimony(III) tartrate hydrate (Sigma) and paromomycin sulfate salt (Sigma), were used in order to accurately evaluate the resistance. Solvent (DMSO) controls were used where appropriate. Hundred μL from each well were incubated at 37°C at 5% CO2 for 4 hours with 20 μL Cell Titer Blue (Promega). Fifty μL of 10% SDS were added to each well, and fluorescence was measured (555 nm λexc/580 nm λem) using a Typhoon FLA-9500 laser scanner (GE Healthcare) and analyzed with ImageQuant TL software (GE Healthcare). EC50 values were calculated by non-linear regression analysis using SigmaPlot (v 11.0). All experiments were done in triplicate with appropriate controls in each case.
Both WT and MIL-resistant cultures were sequenced for previously described point mutations in the L. donovani MT (T421N, L856P, W210*) and Ros3 subunit (M1) [14] and in L. major (G852D, M547del) [15]. DNA was amplified with primers outlined in S1 Table. PCR product sizes ranging from 149–277 bp were purified using the GeneJET Gel Extraction Kit (Thermo) and sent to the Genomics Core Facility at the University of Notre Dame for sequencing. Sequences were analyzed using ClustalX [24].
Total RNA was isolated from logarithmic and stationary phase promastigotes using Trizol Reagent (Invitrogen), reverse transcribed with SuperScript II Reverse Transcriptase (Invitrogen) after deoxyribonuclease I treatment with TURBO DNA-free Kit (Ambion, Invitrogen). All qRT-PCR reactions were performed in triplicate using SYBR Green (Invitrogen) fluorescence for quantification in a 7500 Fast Real-Time PCR System (Applied Biosystems). The ΔΔCƬ method was used to determine relative changes in gene expression [25] with data presented as fold change in the target gene expression in L. major MIL-resistant cultures normalized to internal control genes GAPDH and SOD, using L. major WT as a reference strain. Standard PCR conditions were: 95°C for 10 min, followed by 40 cycles of 94°C for 1 min, 60°C for 1 min, and 72°C for 2 min. Primer design was based on nucleotide sequences of L. infantum genes coding for the L. donovani MT, L. donovani Ros3, SHERP, GAPDH and SOD genes. All experiments were performed in triplicate with appropriate controls included in each case.
Two different methods were utilized to assess metacyclogenesis as described previously [26]. Briefly, a Ficoll (Sigma) gradient was set-up using 4 mL of 20% Ficoll overlaid with 4 mL 10% Ficoll in M199 medium without FBS and 4 mL of 5-day stationary-phase culture in M199 medium laid on top. The step gradients were centrifuged at room temperature for 10 min at 1300 x g without braking or acceleration to separate out the layers. The top two layers of the gradient were recovered and the percentage of metacyclic parasites was determined by counting in a Neubauer chamber before and after the enrichment procedure. For agglutination analysis, 5-day stationary-phase cultures were pelleted and resuspended in 1 mL M199 medium (CellGro) and 10 μL peanut agglutinin (50 μg/mL) (Sigma) was added. After 30 minutes of room temperature incubation, samples were centrifuged at 200g for 10 minutes. The supernatant was recovered and the percentage of metacyclic parasites was determined by counting in a Neubauer chamber before and after the enrichment procedure. All experiments were done in triplicate.
RAW264.7 murine macrophage cells were counted using Trypan Blue (Amresco) and plated at 5x105 cells/well in 12-well plates. Infections were performed with metacyclic parasites isolated as described above. Infections were carried out at a multiplicity of infection (MOI) of 10 parasites per macrophage. Free parasites were removed by one wash with RPMI without FCS 6 h post-infection and samples collected at 6, 12, 24 and 48 h post-infection by DiffQuick staining of cytospin whole-cell preparations and visualized with light microscopy. All infections were done in triplicate and at least two independent experiments were performed.
Phlebotomus papatasi (Origin: Turkey, PPTK) was reared in the Department of Biological Sciences, University of Notre Dame, according to conditions previously described [27]. For the experiment, three-to-five day old female sandflies were used. Two groups, one experimental and one control, each containing 50 female and 10 male sandflies were placed in a 500 mL plastic container (ø = 6.3 cm, height = 6.5 cm) (Thermo-Nalgene) covered with a piece of nylon mesh (0.5mm). Blood feeding was performed through a young chicken skin membrane attached to a feeding device. Prior to sandfly feeding, fresh mouse blood was heat inactivated for 30 min at 56°C. Infection of sandflies with L. major FVI strain promastigotes was done by addition of 1×107 logarithmic parasites/mL into the blood meal. Sixteen to twenty four hours after blood feeding, the presence or absence of blood in the sandfly digestive tract was verified by anesthetizing flies with CO2 and observing the midgut distension under a stereomicroscope (Carl Zeiss). One week post-blood meal, midguts of blood-fed sandflies were individually dissected and thoroughly homogenized in 30 μl PBS buffer (pH 7.4) using a hand held tissue homogenizer and pestle. Parasites were counted in a Neubauer chamber.
5x105 metacyclic parasites isolated by peanut agglutinin (see above) from stationary cultures of L. major FVI were injected subcutaneously in the left hind footpad of Balb/c mice, as previously described [26]. Lesion development was monitored by measuring weekly the thickness of the footpad using a Vernier caliper. Number of parasites at lesion site were enumerated by limiting dilution assay [28]. Cell lines were passaged at least once through mice before performing in vivo virulence studies to minimize the loss of virulence after prolonged in vitro culture.
Significance was determined by p-values calculated from a two-tailed student’s T-test in GraphPad Prism 6.0 unless otherwise stated.
L. donovani MT: GenBank accession number AY321397.1; L. donovani Ros3 GenBank accession number DQ205096.1; SHERP: GenBank accession number XM_001683391; GAPDH: GenBank accession number XP_001684904, and SOD: GenBank accession number XP_001685502.
L. major FVI MIL-resistant parasites were generated using step-wise selection up to 40 μM MIL. Parasites were unable to proliferate in higher MIL concentrations, likely due to reaching the critical micellar concentration of MIL leading to degradation of the membrane due to the detergent effects of MIL [29]. FVI WT promastigotes were plated in solid M199 media and two random clones were used for MIL selection in flasks. In order to assess the degree of MIL-resistance in our lab populations of L. major we measured EC50 values using the resazurin-based CellTiter-Blue (Promega) assay. MIL-resistant cultures exposed to the highest concentrations of MIL (30 μM, 40 μM), and labeled R30 and R40 herein, have accordingly higher EC50 values than R10 and R20 (Fig 1). MIL-resistant cultures growing in the absence of MIL exhibited lower EC50 values than their counterparts under constant MIL-selection. However, it is important to note that this decreased EC50 value of MIL-resistant L. major is still higher than the EC50 of WT L. major cultures (Fig 1, dotted line) after at least 95 passages (2 passages per week, ca 11 months). This suggests that once any degree of resistance is accrued MIL-resistant cultures do not revert back to WT phenotype, despite the removal of MIL selective pressure (Fig 1). It is worth noting that a different resistant phenotype may be obtained if drug selection is performed in axenic promastigotes or intracellular amastigotes, as shown for paromomycin selection in antimony-resistant L. donovani [17, 30].
We next determined any difference in growth patterns between the sensitive (WT), resistant (R30) and resistant grown in the absence of MIL (R30no) L. major populations. Growth curves showed that MIL-resistant L. major proliferation is similar to L. major WT and cured lines (Fig 2), indicating that increased MIL exposure has no effect on proliferation in L. major. We used a FACS-based approach to detect two different apoptotic markers i) membrane permeability and ii) PS exposure to determine the response of parasite to stress after MIL selection. L. major R30 cell lines exhibit minimal stress and are comparable to WT populations judging the histogram levels corresponding to Annexin V and YO-PRO as analyzed by flow cytometry (S1 Fig).
Experimental MIL-resistance in L. donovani has previously been attributed to identified point mutations in the MT and Ros3 subunit (T421N, L856P, W210, and M1) [31]. We sequenced the regions of the transporter and subunit in two independent clones of the R40 line (highest concentration; R40.1 and R40.2) that had been under drug selection for at least 75 passages. As shown in Table 1, these mutations were not found in our lab populations. These results are in accordance with previous characterization of MT in MIL-resistant L. major [15]. Two genuine mutations identified in the L. major MT were pinpointed for this study: a three-nucleotide deletion (M547del) and a transition mutation (G852D) [15]. As seen in Table 1, our lab populations displayed identical sequences to WT. Although our data do not eliminate the possibility of other unidentified genetic mutations having a role in MIL-resistance in L. major, it is interesting to observe that even at higher concentrations (R40) and after long-term exposure to MIL (at least 75 passages) none of the reported mutations were found.
We investigated the possibility of any conferred resistance to alternative antileishmanial treatments by measuring EC50 values as described in Material and Methods. No cross-resistance was found in any of the R30 clones or cured lines to amphotericin B, antimony (III) and paromomycin (Table 2). Interestingly, miltefosine resistance significantly increases the sensitivity of the parasite to treatment with pentamidine 3-fold lower than WT (Table 2). When MIL has been withdrawn, the sensitivity of the parasite to this particular treatment is restored to levels comparable with the wild-type (Table 2), suggesting a potential synergistic mechanism. A similar synergy has been reported for sitamaquine/pentamidine combinations in L. donovani [32], although the use of a combined therapy of miltefosine and pentamidine is hindered by the high toxicity of pentamidine [33]. Lastly, treatment of R30 MIL-resistant cultures with paromomycin had a significant effect on the sensitivity (ranging from 2–4 fold lower than WT) of one of the clones (R30.2), indicative of potential clonal variability.
Procyclic L. major promastigotes differentiate into highly virulent metacyclic promastigotes during metacyclogenesis [34]. This process occurs in the midgut of sandflies and can be mimicked in vitro when acidification occurs in the medium. Due to the lack of phenotypic differences in our clonal lines we performed the following in vitro and in vivo experiments with the R40.2 line. We enriched metacyclic promastigotes by Ficoll 400 step gradient and peanut agglutination, as described in Material and Methods. Analyses of metacyclogenesis showed that L. major R40 had higher percentages (2-fold) of metacyclics than L. major WT (Fig 3, right panel). qRT-PCR was used to amplify SHERP gene, which is almost exclusively and highly expressed in infective and non-replicative stages of the parasite [35]. SHERP expression was significantly elevated in R40 parasites (Fig 3, left panel), confirming our metacyclic enrichment approaches. Increased metacyclogenesis has been reported in antimony-resistant L. donovani clinical isolates [36], and metacyclogenesis is regarded as a major contributor to the fitness of the parasite. In New World cutaneous species, L. mexicana resistant to Glibenclamide, an ATP-binding-cassette (ABC)-transporter blocker exhibited a reduced expression of the Meta-1 protein [37].
The stationary phase-specific differences of R40 primed us to study their capacity to infect RAW264.7 murine macrophage cells. We routinely passage our L. major cell lines through Balb/c mice to compensate for the loss of virulence due to in vitro culture. 5-day stationary cultures were subjected to peanut agglutination, and R40 and WT lines were incubated with RAW264.7 cells at a multiplicity of infection of 10 metacyclics per host cell. Intracellular parasite burden was determined by nuclear staining and microscopy at 6, 12, 24, and 48 h postinfection. Initial levels of R40 infections are comparable to the control (Fig 4A). A significant difference in R40 infectivity was apparent 48 hours post infection. This was further corroborated by decreased intracellular proliferation of R40 cells 48 hours post infection by over 20% (Fig 4B). Pentamidine-resistant L. mexicana showed no differences in the in vitro infectivity in resident mouse macrophages when compared with the wild-type clone [38].
In contrast, higher metacyclogenesis levels in clinical isolates of L. donovani resistant to antimony translated into higher in vitro infection levels [36].
We next investigated the virulence of WT and R40 using an established experimental mouse infection [39]. Control and R40 were normalized for virulence through one passage in Balb/c mice [40]. 105 WT and R40 metacyclic parasites were inoculated into the hind footpad of groups of five-six female Balb/c mice. A Vernier caliper was used to monitor lesion formation by measuring the increase in footpad size weekly. Control parasites attained a lesion size of ca. 4 mm, 5 weeks after inoculation and resulted in necrotic lesions (Fig 5). Interestingly, R40 were highly attenuated and lesions were only apparent 4 weeks after infection. Our observations in vitro with R40 cells showing a decreased infectivity and intracellular proliferation seem to have extended well to an in vivo mouse model. Amphotericin-resistant L. mexicana parasites were able to infect Balb/c mice, but the resulting lesion growth was slower than that after infection with susceptible parasites [41]. In contrast, several clinical isolates of L. donovani resistant to pentavalent antimonials showed a greater virulence in a mouse model of visceral leishmaniasis [42]. Importantly, our data suggest that metacyclogenesis alone is not a reliable marker of fitness, at least in MIL-resistant L. major, and in vitro and in vivo studies are necessary to further assess its competitive fitness. In this scenario, the L. major / MIL combination resembles the reduction in fitness widely observed in Plasmodium falciparum populations resistant to chloroquine [43].
Fitness of Leishmania parasites is linked to transmission success in the natural insect vector, therefore we tested whether MIL resistance would impact the capacity of Leishmania to survive in the natural sandfly vector. Three-to-five day old female Phlebotomus papatasi (Origin: Turkey, PPTK) sandflies were infected with 1×107 logarithmic parasites/mL as described in Material and Methods. 24h post-blood meal, the presence or absence of blood in the sandfly digestive tract was verified and one week post-blood meal, 9 midguts of blood-fed sandflies infected with WT and 14 midguts from the R40 group were individually dissected. Parasite load per individual midgut was assessed. No significant differences were observed between the two groups (Fig 6) suggesting that MIL resistance does not affect the survival capacity of L. major in the natural vector.
In summary, as shown for L. donovani [44], the generation of experimental resistance to MIL is easily achieved by step-wise selection in L. major. Axenic resistant promastigotes proliferate as control cells, and the phenotype is stable. As suggested by our data, metacyclogenesis is an important process in the life cycle of the parasite, but should be carefully interpreted as a fitness marker. A combination of in vitro, in vivo and vector studies are necessary to fully assess the competitive fitness of MIL-resistant L. major, and studies would be further strengthened with the use of recent clinical isolates of both MIL-sensitive and MIL-resistant L. major parasites. Further studies will attempt to understand the impaired ability of MIL-resistant L. major to survive in the mammalian host at the molecular level. Overall, our findings are relevant for current and future antileishmanial chemotherapy strategies.
|
10.1371/journal.pcbi.1003863 | The SH2 Domain Regulates c-Abl Kinase Activation by a Cyclin-Like Mechanism and Remodulation of the Hinge Motion | Regulation of the c-Abl (ABL1) tyrosine kinase is important because of its role in cellular signaling, and its relevance in the leukemiogenic counterpart (BCR-ABL). Both auto-inhibition and full activation of c-Abl are regulated by the interaction of the catalytic domain with the Src Homology 2 (SH2) domain. The mechanism by which this interaction enhances catalysis is not known. We combined computational simulations with mutagenesis and functional analysis to find that the SH2 domain conveys both local and global effects on the dynamics of the catalytic domain. Locally, it regulates the flexibility of the αC helix in a fashion reminiscent of cyclins in cyclin-dependent kinases, reorienting catalytically important motifs. At a more global level, SH2 binding redirects the hinge motion of the N and C lobes and changes the conformational equilibrium of the activation loop. The complex network of subtle structural shifts that link the SH2 domain with the activation loop and the active site may be partially conserved with other SH2-domain containing kinases and therefore offer additional parameters for the design of conformation-specific inhibitors.
| The Abl kinase is a key player in many crucial cellular processes. It is also an important anti-cancer drug target, because a mutation leading to the fusion protein Bcr-Abl is the main cause for chronic myeloid leukemia (CML). Abl inhibitors are currently the only pharmaceutical treatment for CML. There are two main difficulties associated with the development of kinase inhibitors: the high similarity between active sites of different kinases, which makes selectivity a challenge, and mutations leading to resistance, which make it mandatory to search for alternative drugs. One important factor controlling Abl is the interplay between the catalytic domain and an SH2 domain. We used computer simulations to understand how the interactions between the domains modify the dynamic of the kinase and detected both local and global effects. Based on our computer model, we suggested mutations that should alter the domain-domain interplay. Consequently, we tested the mutants experimentally and found that they support our hypothesis. We propose that our findings can be of help for the development of new classes of Abl inhibitors, which would modify the domain-domain interplay instead of interfering directly with the active site.
| The expression of the constitutively active BCR-ABL fusion tyrosine kinase is sufficient for the initiation and maintenance of chronic myelogenous leukemia (CML) in humans [1]. BCR-ABL is the result of the t(9;22) chromosomal translocation that leads to the fusion of the Abelson tyrosine kinase (ABL1) and the breakpoint cluster region (BCR) gene [2], [3]. The dysregulated fusion protein activates a number of signaling pathways associated with inhibition of apoptosis and uncontrolled proliferation.
In the light of the above it is not surprising that the mechanisms regulating the activation and deactivation of both the wild type c-Abl and BCR-ABL tyrosine kinases have attracted a considerable interest [4]–[9].
In physiological conditions the catalytic activity of tyrosine kinases is tightly regulated through the interplay between various protein domains, phosphorylation events and associated conformational states of the catalytic domain (CD) [10]. During the catalytic cycle, its high intrinsic flexibility allows the CD to react to the regulatory elements by switching reversibly between a number of distinct active and inactive states.
In most non receptor-type tyrosine kinases, the catalytic domain is preceded by a Src homology 2 (SH2) domain [11] (Figure 1A). The importance of the SH2 domain in the auto-inhibition and/or activation of the catalytic domain has been shown in c-Src [6], [8], [12]–[14], Hck [15]–[17], Fes [18], [19] and c-Abl, among others. The role of the SH2 domain in c-Abl is of special interest, because it is involved both in auto-inhibition and activation of the CD [18], [20], [21], and mutations in the SH2 domain have been related to imatinib-resistance in CML patients [18], [19], [22]. In the auto-inhibited state, the SH3 and SH2 domains and the SH2-kinase linker form a rigid clamp around the CD, which is locked in place by an N-terminal myristoyl modification of the N-terminal cap region inserted deeply into the CD [7], [23] (Figure 1B). This grip reduces the flexibility of the CD and, in particular, dampens the opening and closing of its N- and C-termini around the active site [16], . This so-called hinge or breathing motion of the CD is required for catalysis, and its impairment is associated with low catalytic output [26]–[28].
While the molecular basis for the role of the SH2 and SH3 domains in Abl autoinhibition is well understood, the mechanism of their activating effect is less straightforward. Recent crystal structures and small angle X-ray scattering studies have revealed that the transition from the auto-inhibited to the fully activated form of Abl requires a complete reassembly of the complex formed by the CD, the SH2 and the SH3 domains, leading to migration of the SH2 domain from the C-lobe to the N-lobe of the CD (“top-hat” conformation) (Figure 1C) [24]. Importantly, the effect of this rearrangement is not reduced to merely revoking the autoinhibition of Abl by removing the SH2-SH3 domain clamp, but the SH2 domain, when bound to the N-lobe of the CD, enhances the activity of the kinase, although it bears no direct contact to the catalytic site. Recently, it was shown that the I164E mutation in the SH2 domain, which interrupts the hydrophobic interactions at the interface between SH2 and CD in the “top-hat” conformation, leads to deactivation of Abl [5], [18]. A similar domain arrangement and activating effect has been observed in other kinases [19], [29], such as Fes [18] and Btk [30]. Hence, in multiple tyrosine kinases, the SH2 domain acts as an allosteric effector. Comparison of the crystal structures of the auto-inhibited and the activated forms of Abl does not reveal any marked conformational changes, particularly at the active site, that could explain the mechanism of activation by the SH2 domain, a finding that points towards a dynamic rather than static nature of the allosteric effect. The essential features of this allosteric effect and the molecular mechanism by which it is transferred from the N-lobe to the catalytic site still remain elusive.
The SH3/linker region has also been shown to be involved in the regulation of Abl activation [31]. However, here we focus on the SH2 domain that, even in the absence of SH3 and the linker, has been shown to have strong activating effect.
We used a multi-disciplinary approach combining elastic network models, extensive molecular dynamics simulations, free energy calculations and functional assays following mutagenesis to characterize the allosteric coupling of the CD of Abl with the SH2 domain as well as the modulation of the dynamic properties of the assembly by interactions in the “top-hat” conformation. Both the long atomistic simulations and the elastic network models indicate a significant change in the dynamics of key regulatory elements, providing a simple explanation of the mechanism of allosteric activation. Based on the computational results we designed a number of point mutants to validate the proposed model. These mutants were expressed in human cells and tested for kinase activity. We identified mutations, all distant from the active site, that were either activating the catalytic output of the kinase or were disrupting. Interestingly, we also identified a residue that when mutated lead to a decoupling of the activity of the CD from interaction with the SH2 domain. Collectively, the data suggested an effect of SH2 binding that results in changes of the properties of the αC helix, reminiscent of the effect that cyclins exert on cyclin-dependent kinases.
To elucidate the mechanism by which the SH2 domain stimulates the catalytic activity of the Abl kinase we first characterized the dynamics of the CD alone and with the SH2 domain bound in the activating conformation using elastic network models. In the free CD, the two predominant modes emerging from the normal mode analysis (NMA) corresponded to the well-described hinge motion [26], [28], [32] and to a twist of N- and C-lobe against each other (Supplemental Figure S1A). When the SH2 domain was included in the elastic network model, the hinge motion continued to be the principal motion, but the corresponding normal mode included a sliding of the SH2 domain along the binding interface, while the amplitude of the movement of the N-lobe of the CD was reduced (Supplemental Figure S1B). This finding suggests that one role of the SH2 domain may be to regulate the hinge motion while restricting lobe twists during catalysis.
We analyzed the allosteric coupling of local conformational fluctuations along these normal modes [33] (see Methods and Text S1). Figure 2A and Supplemental Figure S1C show that local distortions in the SH2 domain are associated with conformational changes in both lobes of the CD. Important couplings are detected between the SH2 domain and specific, spatially separate motifs in the C-lobe, including the catalytically important activation loop (A-loop), the αD-αE loop, the αF-αG loop and the αG helix. The αD-αE loop forms part of the myristoyl binding pocket. The αG helix is known to have an important role in the catalytic mechanisms of many kinases. In some of them, c-Abl among them, it forms part of a platform for substrate binding [34]. Furthermore, it has been proposed that in Abl and EGFR the αF helix and the αF-αG loop are allosterically coupled to the αC helix and are involved in the dynamically enhanced stabilization of active conformations [20], [29]. Virtually all other motifs coupled to the SH2 domain in turn also couple to the crucial A-loop.
The allosteric coupling analysis thus suggests that the SH2 domain acts primarily on the N-lobe loops and the A-loop. All these motifs are coupled among them through a dense network of allosteric interactions, which affect also the P-loop and the αC helix, two structural motifs that, together with the A-loop, participate actively in the catalytic process.
We next carried out 1 µs long all-atom molecular dynamics (MD) simulations of the free CD and the activated SH2-CD complex in solution. The structure of the CD does not, in general, deviate much from the crystal structure (Supplemental Figure S2A, blue line). However, some motifs have a much higher degree of intrinsic flexibility than others, as can be seen from the root mean square fluctuations (RMSF) of the backbone (Figure 2B and Supplemental Figure S2B, blue lines). The largest fluctuations were observed in the N-lobe loops, the αC helix, and the A-loop. Enhanced flexibility of the N-lobe together with partial unfolding of the αC helix has been observed in other protein kinases, such as FES [18] and EGFR [35]–[37]. The flexibility of these motifs are thought to be crucial for the regulation of activity and sensitivity towards specific kinase inhibitors [38]–[40]. The partial unfolding of the αC helix results in the rupture of a conserved salt bridge (Glu305 to Lys290), which needs to be formed and maintained stable for Abl activation [41] (Figure 3C, blue line).
When the SH2 domain was included in the simulation, the CD domain is more rigid (Figure 2C and Supplemental Figure S2A, orange line), while the relative orientation of the two domains and the position of the SH2 domain change noticeably (Figure 3A and Supplemental Figure S2A, red line). Compared to the crystal structure, conformations in which the SH2 domain is shifted towards the active site predominate.
A further hint at the mechanism by which the SH2 domain is connected to the dynamics of the catalytic domain is given by the comparison of the corresponding flexibility patterns (Figure 2B, C and Supplemental Figure S2B). By far the strongest impact is on the P-loop, which is known to be vital for ATP binding and kinase selectivity [36].
Figure 3B compares the most representative structures of the CD in absence (blue) and presence (red) of the SH2 domain, obtained by a cluster analysis of snapshots from the trajectories. The SH2 domain directs the β3-αC loop towards the active site and the A-loop (Supplemental Figure S3A, B). In this conformation, the β3-αC loop is stacked over the P-loop, fixing it in a conformation in which it points towards the DFG motif in the active site and the αC helix. This interplay between β3-αC loop, P-loop and αC helix does not take place in the free CD. Furthermore, the rearrangement and stiffening of the N-lobe motifs results in a stabilization of the crucial salt bridge between Glu305 and Lys290 (Figure 3C, red lines).
While the SH2 domain strongly stabilizes the N-lobe in general and in particular the P-loop, it enhances the flexibility of the A-loop and the αF-αG loop. The A-loop plays a fundamental role in the inactive-to-active conformational switch and together with the other elements, is involved in substrate/product binding (Supplemental Figures S2C, D). It is thus conceivable that the interaction with SH2 might play a role both in the inactive-to-active equilibrium and in the release of the products, which has been proposed to be the rate determining step of c-Abl-dependent phosphorylation.
Next, we addressed the issue of how the SH2 domain affects global motions of the CD by using principal component analysis (PCA), which provides a description of the dominant motions of the CD during the simulations.
Figure 3D shows the projection of the trajectories in absence and presence of the SH2 domain on the eigenvectors (EVs) of the two most predominant modes of motion of the backbone of the CD. These principal eigenvectors match the first normal modes of the simplified elastic network model quite well, representing again the hinge motion and the lobe-twist, The only significant difference is found in the second PCA eigenvector, that also includes a distortion of the αC helix (Supplemental Figure 1D). The agreement of these two completely independent methods in describing global changes in the CD dynamics is remarkable and indicates that the conformational changes observed during the PCA are facilitated by the low-frequency, global motions that are intrinsic to the structure.
The SH2 domain significantly restricts the movement along both EVs and leads to changes in the conformational equilibrium. The amplitude of the hinge motion along EV1 is shifted towards conformations in which the lobes close down over the active site (Figure 3D, left side of the graphic, and Supplemental Figure S1D, blue lines), which is also reflected directly by the distance between N- and C-lobe (Supplemental Figure S3C). The effect of the SH2 domain on the motion along EV2 is also strong. The twist between the N- and C-lobes, which leads to conformational changes at the active site of the kinase, is constrained and the distortion of the αC helix is strongly reduced (Figure 3D). A PCA analysis of the trajectory of the SH2-CD construct revealed that, unsurprisingly, the first eigenvector still represents the pure hinge motion. However, eigenvectors 2 to 4 all include a certain degree of hinge closing but virtually no distortion of the active site motifs (Supplementary Figure S1E). The corresponding eigenvalue spectra furthermore confirm that in presence of the SH2 domain the hinge motion (EV1) gains importance relative to all other modes (Figure 3E).
The MD simulations indicate that the SH2 domain bound in the “top-hat” conformation exerts a double effect. On the one hand, it locally reduces the flexibility of a number of structural motifs of the CD, which participate in the catalytic process, such as the P-loop, the β3-αC loop and the αC helix, locking them in their active conformations. On the other hand, it modifies the collective motions of the CD, channeling energy away from unproductive twists and distortions into the catalytic hinge motion, and shifting the hinge motion towards closed conformations and the A-loop towards more “active-like” conformations.
To investigate the effects of the SH2 domain on the A-loop conformation propensities we computed the inactive-to-active free energy landscape. To that aim we used a multiple-replica free energy algorithm (parallel-tempering Metadynamics) and a structure based hybrid force field. A similar computational strategy has been successfully used in the case of the CD of the highly homologous Src kinase to study the A-loop opening, where it was able to provide an accurate reconstruction of the free energy surface underlying the transition [42]. This force field reproduces fairly well the flexibility patterns observed with long all-atom explicit solvent simulations, apart from a small discrepancy in the region corresponding to the αG-helix in the CD. This region appears to be somewhat more rigid that it should be in the CD, but recovers the correct flexibility in the complex (Supplemental Figure S4).
In absence of SH2, the A-loop of the CD is mostly closed, as expected from its in-vitro low catalytic activity (Figure 4, left). The A-loop active-like, or “open”, conformation is still a local minimum of the free energy but at a much higher value (ΔGO-C≈6 kcal/mol). Moreover, the large free energy barrier separating the closed to the open A-loop state (ΔG‡-O≈14 kcal/mol) disfavors the transition.
In contrast, the effect of the SH2 regulatory domain is to shift the equilibrium towards an active-like conformation of the A-loop (Figure 4, right) rendering it as stable as the closed conformation. The free energy barrier between the two states is greatly reduced (∼6 kcal/mol lower than without SH2) and the open basin is widened increasing the A-loop flexibility.
Based on these computational results, we proposed a number of point mutations, both in the SH2 and in the catalytic domains, that we expect to modify the dynamic interaction between the domains by interrupting essential interactions or altering the flexibility of important motifs (Table 1 and Text S1). We mainly focused on mutations in the β3-αC loop and the hinge region, as in the simulations these motifs were most strongly affected by SH2 binding. Both motifs are known to be relevant for catalysis, so mutations in these regions can be expected to affect kinase activation in general. However, based on the simulations, we predicted that for some of them the effect should be different in the free CD and in the SH2-CD construct and therefore strengthen our proposal of the how the SH2 domain interferes with c-Abl dynamics.
The effect of the mutations was characterized by assays of c-Abl kinase activity in vivo and in vitro. None of the candidate mutations have been described previously as clinically relevant and could in principle have either a moderate activating, neutral, or disruptive effect. The effect of the mutations on the catalytic activity of c-Abl was assessed both in the context of the SH2-CD module and in the isolated CD, in order to discern whether the mutation generally affects the fold or activation state of the kinase domain, or if it specifically targets the SH2-mediated regulation mechanism. The mutations were introduced in HA-tagged Abl SH2-CD or CD-only constructs which were transiently expressed in human embryonic kidney 293 (HEK293) cells, and the level of their in vivo activity was assessed by phospho-Y412-Abl and total phosphotyrosine immunoblotting of the crude lysates. The mutated proteins mostly accumulated to levels comparable to the wild-type constructs, indicating that they did not affect protein stability. The impact on c-Abl enzymatic activity using an in vitro assay with an optimal peptide substrate was assessed for selected mutations (Figure 5). As observed previously [5], under the conditions of this assay, the wild-type SH2-CD module exhibits at least 2-fold higher activity than the wt CD construct. Since the peptide substrate contains a single phosphorylation site, the observed difference in activity could be ascribed to the SH2 domain-mediated activation that is independent of phosphotyrosine binding and processive phosphorylation events that may ensue. The comparison of SH2-CD wt and SH2-CD S173N (a FLVRES motif mutant which abolishes pTyr binding by SH2) using this assay shows no difference in kinase activity, hence pTyr binding contributions can be excluded [5]. Thus, this system should be well suited to study the effects resulting from SH2-kinase domain interactions.
Some of the tested mutations were neutral, such as N165A and E187K in the SH2 domain (Supplemental Figure S5), or F330A in the β5 sheet. Others abolished kinase activity and could not be rescued by the SH2 domain, like K266E in the β1-β2 loop, P328G P329G in the β4-β5 loop, F330A in the β5 sheet, or G340P in the hinge. In these cases, the effect of the mutation goes beyond interruption of the CD-SH2 interaction and interferes directly with the catalytic mechanism.
All of the tested mutations at the T291 position in the β3-αC loop exerted a moderately disrupting effect both in the context of the SH2-CD module, as well as the CD alone, suggesting that the effect of the mutation is not likely to be explainable solely by the abolishment of the interaction with the SH2 domain (Supplemental Figure S6). It is possible that T291 plays an important role due to its localization in the key β3-αC loop, where it might be required to confer or sustain an active conformation of the kinase domain.
A number of mutants (Figure 1D), however, were more revealing in terms of understanding the SH2 activation mechanism. M297 in the β3-αC loop turned out to be very sensitive to mutations. Based on the MD simulations, we hypothesized that the β3-αC loop may act as a lever, transmitting the signal from the SH2 domain to the catalytic site and positioning the αC helix correctly. The relatively conservative M297L mutation led to a several-fold decrease in kinase activity (Figure 5A). Surprisingly, the more drastic change of the M297G mutation did not impair kinase activity, but had a slightly activating effect on the isolated CD, while it was neutral in the context of the SH2-CD construct (Figure 5A, B). The slight increase in CD activity had not been anticipated from the simulations of wt c-Abl, and suggests a de novo effect, which underlines the importance of this region for modulating c-Abl activity. In the presence of SH2, the activating effect of M297G was either suppressed, masked by similar changes, or compensated for by a corresponding drop in c-Abl activity. This suggests that SH2 indeed uses the β3-αC region as a key lever to efficiently redirect the conformational changes of CD. Changing the side chain and therefore the highly sensitive interaction network (M297L) severely decreases c-Abl activity, while introduction of additional flexibility (M297G) has a slightly activating effect and makes c-Abl activation less dependent on SH2 domain binding.
The pivotal role of the β3-αC loop is further confirmed by the effect of mutating E294 and V299 to prolines, which are the equivalent residues in c-Src, a protein closely related to c-Abl but not known to be activated by SH2. We would expect the E294P V299P double mutant to stiffen the β3-αC loop lever and, consequently, the αC helix. In turn, this should activate the CD and enhance the effect of the SH2 domain. The double mutant was indeed found to be markedly activating, both in the context of the catalytic domain alone as well as within the SH2-CD construct (Figure 5C). The introduction of E294P and V299P resulted in a substantial increase in enzyme velocity compared to wt SH2-CD (Figure 5D). The single mutation E294P also exerted an activating effect on Abl activity, however the effect was more pronounced in the presence of V299P, suggesting that the two mutations might act synergistically (Figure 5C).
Lastly, we have investigated the effect of changes in the flexibility of the hinge region. In the MD simulations, Y339 in the hinge region showed the highest fluctuations. Mutation of the tyrosine to glycine, which should make the hinge even more flexible, has no effect on c-Abl kinase activity (Figure 5E). In contrast, the Y339P mutation, which should rigidify the hinge region, was indeed found to be disruptive to c-Abl activity (Figure 5E). Mutation of another flexible hinge residue, G340, to proline also abrogated c-Abl activity, as observed by a decrease in phosphorylation of cellular proteins on tyrosine.
Finally, another striking evidence of the stabilizing effect of the SH2 domain on the kinase domain comes from the changes of in vitro measured c-Abl activity upon temperature increase (Figure 5F, G). At higher temperature, the activity of the c-Abl CD constructs decreased substantially, most likely due to an increase in mobility that is undirected and unproductive. However, under the same conditions, the activity of c-Abl SH2-CD increased, suggesting that the SH2 domain is able to direct the enhanced movements of the kinase towards more productive states, as had been indicated by the changes in the PCA eigenvalue spectrum due to SH2 binding. Interestingly, the Y339P mutation which should rigidify the hinge region, could not be rescued by an increase in temperature, and the activity of the mutant SH2-CD was only slightly raised at 35°C as compared to CD Y339P (Figure 5F). Contrary to the Y339P mutant in the hinge region, the M297G and E294P V299P mutants in the β3-αC loop preserve the effect of the temperature increase (Figure 5G), suggesting that it is, in fact, the hinge motion that is responsible for the effect. We suggest that the hinge region always has to maintain an important degree of flexibility in order for the kinase to be functional. In the free CD, conformations with a large opening of the hinge dominate, and additional twists and distortions render the hinge motion largely ineffective. Upon SH2 binding, the non-catalytic motions are strongly restricted and the hinge motion is directed to the optimal amplitude and opening needed for catalysis.
In the experiments, the M297G single mutant and the E294P V299P double mutant modified the activity of the CD as well as the effect of the SH2 domain in the activating “top-hat” conformation. Based on the simulations of the wt kinase, we had predicted that changes in the flexibility of the β3-αC loop should affect the directing effect the SH2 domain exerts on the CD. We carried out unbiased 1 µs simulations of the free CD and the CD-SH2 construct of both mutants to gain insight at atomic level into the mechanism underlying the observed effects and test our hypothesis.
The M297G mutant, which was chosen to introduce additional flexibility in the β3-αC loop and weaken the coupling of the SH2 domain with the αC helix, was slightly activating in the free CD and neutral in the CD-SH2 construct. In the simulation of the mutant CD, the expected enhanced flexibility of the β3-αC loop was confirmed (Supplemental Figure S7A). The simulation also provided us with an explanation for the rather unexpected increase in activity of the CD. The expansion of accessible conformations lead to the loss of interactions between the β3-αC and the P-loop, and enabled the system to assume a conformation similar to that observed in the wt in the presence of SH2 without requireing the directing effect of the SH2 domain (Supplemental Figure S8A, C). Contrary to the wild type, the flexibility of the CD in the M297G complex was enhanced, not reduced upon SH2 binding (Supplemental Figures S7B, S8B), and no additional strengthening of the salt bridge between the αC helix and the β3 sheet was observed (Supplemental Figure S7C). The picture emerging from the simulations is thus that the M297G mutant in the CD partly mimics SH2 binding. At the same time, due to the enhanced flexibility of the N-lobe loops, the effect SH2 domain on the arrangement of the catalytic important motifs is much smaller than in the wt, leading to a situation, where the difference in activity between CD and SH2-CD is markedly reduced.
To summarize, the simulations indicate that introduction of a glycine residue in the β3-αC loop enhances the flexibility of the N-lobe, allowing the P-loop to adopt an activated conformation, while it partly uncouples kinase activation from SH2 domain binding.
In the E294P V299P double mutant, the introduction of two proline residues in the β3-αC loop was expected to stiffen it, limiting the distortions of the αC helix. In our experiments, the E294P V299P double mutant was, indeed, found to be significantly activating, while, in contrast to the M297G mutant, the enhancing effect of the SH2 domain was preserved.
In the MD simulation, the mutant CD exhibited a flexibility pattern very similar to the wt with a general reduction in flexibility, which was strongest in the P-loop, the β3-αC loop and the αC helix (Supplemental Figure S7D and Supplemental Figure S8D). Surprisingly, the SH2 domain, however, increased the conformational fluctuations (Supplemental Figure S7E). Closer inspection of the corresponding structures, however, showed that this is due to concerted movements of P-loop, β3-αC loop and αC helix, which follow the movement of the SH2 domain and enhance the hinge motion (Supplemental Figure S8E). The essential salt bridge between E305 and K290 is strongly stabilized even in absence of SH2 (Supplemental Figure S7F). In principle, the substitution of Glu294, in close vicinity of the SH2 domain, could alter the domain-domain interplay, above and beyond changes in flexibility, due to the loss of possible electrostatic interactions between the glutamate side chain and residues of the SH2 domain. We compared the essential dynamics of wt and mutant, and found that in the free CD the hinge movement and the lobe twist are strongly affected by the mutation, while they remain practically unchanged in SH2-CD (Supplemental Figure S8F, G). This finding supports the view that it is the altered dynamical properties rather than lost interactions between E294 and the SH2 domain that leads to the changes in kinase activation.
The combination of structure-based normal mode analysis, unbiased MD simulations and free energy calculations with experiments in vitro and in vivo has shown how the SH2 domain in the “top-hat” conformation changes the dynamics of the catalytic domain of the Abl kinase. Based on the MD simulations, we could identify three residues (M297, E294 and V299) that are key to the interplay between catalytic and SH2 domain in different ways.
Contrary to what could have been expected, the residues forming the hydrophobic spine are not directly affected by the formation of the activating complex. The picture that emerges from our studies is that the SH2 domain stabilizes and repositions the loops that form the binding site, which, in turn, interact with the surrounding loops, stabilizing the P-loop and the αC helix, redirecting the movement of the N-lobe and changing the inactive to active conformational equilibrium of the A-loop. The stabilization and repositioning of the αC helix also leads to an enhancement of the salt bridge between E305 and K290, which has been identified as one of the structural elements essential for kinase activation. The β3-αC loop preceding the αC helix is the key player in this complex mechanism, acting as a lever that transmits the effect from the SH2 domain-binding site towards the substrate binding residues and the catalytically important motifs. The effects of the M297G and E294P V299P mutants demonstrate the crucial role of the β3-αC loop. The observation that these mutations have a different impact on the CD and on the SH2-CD constructs shows that the β3-αC loop is involved in Abl activation by the SH2 domain. While the introduction of a glycine breaks the chain of transmission from the SH2 domain towards the active site and uncouples SH2 domain binding and kinase activation, the stiffening of the β3-αC lever by the introduction of two proline residues enhances the activating effect of the SH2 domain.
The finding that at higher temperatures the unbound catalytic domain becomes less effective while the activity of the CD-SH2 construct increases indicates that the main role of the SH2 domain is to channel kinetic energy into directed, catalytically relevant motions. We propose that the SH2 domain modifies the so-called hinge motion, an opening and closing of the N-and C-lobes upon the active site. This is underlined by the fact that stiffening the hinge by the Y339P mutation decreases substantially the activity of the free CD, which cannot be rescued in the CD-SH2 construct. The hinge motion and the other changes in dynamics upon SH2 domain binding are involved in both the inactive-to-active conformational changes of the A-loop and the release of the products, the phosphorylated substrate and ADP. The latter has been proposed to be the rate-determining step of the phospho-transfer reaction in some kinases [38], [43]. Analysis of the flexibility pattern of the M297G and E294P V299P mutants reveals that the SH2 domain does not simply stabilize catalytically important motifs but that it regulates the Abl kinase through a complex interplay between stabilization, subtle local conformational changes and direction of the concerted hinge motion.
It is interesting to note that the effect on the β3-αC loop are reminiscent of the binding of cyclin to cyclin-dependent kinases [44], [45] and the dimerization of the EGF Receptor, in which the catalytic domain of the activator kinase engages the helix αC of the receiver kinase [46]. The effect of the SH2 domain in Abl, however, occurs at a distance, exploiting a network of interactions and motions that effectively may be equivalent to interaction with the so-called αC “hydrophobic patch” by different structural elements in other kinases, such as the mentioned CDKs, EGFR but also Aurora, Rho kinase and other (reviewed in [41]). Interestingly, Fes, the prototypic SH domain containing tyrosine kinase, engages the SH2 domain also in a “top hat” position relative to the kinase domain, but here the SH2 domain reaches down sufficiently to interact directly with the αC helix [18], [19]. The discovery of the allosteric coupling of catalytically active residues in Abl with SH2 binding may instruct strategies aiming at the development of small molecules interfering with the SH2-kinase interface, but also conformation-depending kinase inhibitors [5], [47], [48]
Normal mode analyses (NMA) were applied to elastic network models of the CD alone and with the SH2 domain bound at the N-lobe in the activating conformation. The models were built from the crystal structure of the CD-SH2 complex (pdb-id: 1opl, chain B).
Allosteric couplings were calculated following the approach of Ref. [33]. They measure correlations in structural distortions along protein normal modes, quantifying interactions between distant residues. See also Text S1.
The CD of c-Abl with the SH2 domain bound in the top-hat conformation (PDBid: 1OPL, chain B) was solvated in tip3p water molecules [49], and the system charges were neutralized, adding the proper number of positive (Na+) or negative (Cl−) ions. For the free CD, the SH2 domain was removed from crystal structure and the protein prepared accordingly. Simulations were performed with the AMBER99SB**-ILDN force field [50], using GROMACS 4.5 [51].
The free energy calculations were performed with parallel tempering metadynamics (PT-metaD) [52], and a structure based hybrid force field.
For details on the set-up of the simulations and the coarse-grained model see Text S1.
Root mean square deviations (RMSD), root mean square fluctuations (RMSF), clustering, distances and principal component analysis were performed with GROMACS tools (g_rms, g_rmsf, g_cluster, g_dist, g_covar and g_anaeig).
All structural figures were created using PyMol [53], except Figure 1 where ICM Browser was used (Molsoft L.L.C), and Supplemental Figures S1A, B where VMD was used [54].
Point mutations were obtained using the Quikchange Site-directed Mutagenesis kit (Stratagene) and pcDNA3.1-2×HA-TEV c-Abl CD and pcDNA3.1-2×HA-TEV c-Abl SH2-CD as templates [5].
Transfection of HEK 293 cells and immunoprecipitation of HA-tagged Abl protein was carried out as described previously [5], [7] using the monoclonal Anti-HA agarose conjugate (Clone HA-7, Sigma-Aldrich). Immunoblotting was done using the following antibodies: anti-phosphotyrosine (4G10, Millipore), anti-phospho-Abl (Tyr412, Cell Signaling Technology), and anti-HA labeled with AlexaFluor 680 (Molecular Probes). Secondary antibody was goat anti-mouse IgG labeled with AlexaFluor 680 (Molecular Probes). Alternatively, peroxidase-labeled mouse anti-HA antibody (HA-7, Sigma) was used. Quantification of the relative amounts of immunoprecipitated Abl protein was done using the Odyssey system (Li-Cor) or ImageJ program.
The in vitro kinase assays were carried out as described previously [5], [7]. Immunoprecipitated HA-tagged c-Abl constructs were resuspended in kinase assay buffer (20 mM Tris-Cl pH 7, 10 mM MgCl2, 1 mM DTT). A peptide with the preferred c-Abl substrate sequence carrying an N-terminal biotin (biotin-GGEAIYAAPFKK-amide) was used as substrate. The reaction was done at 24°C unless otherwise indicated. The terminated reaction was spotted onto a SAM2 Biotin Capture membrane (Promega). The activity was normalized for the relative amount of immunoprecipitated protein and for the activity of c-Abl CD wild-type. For the measurements of KM, each experiment was normalized for a reaction performed in the absence of the peptide, and the amount of kinase.
|
10.1371/journal.ppat.1006534 | Bacterial effector NleL promotes enterohemorrhagic E. coli-induced attaching and effacing lesions by ubiquitylating and inactivating JNK | As a major diarrheagenic human pathogen, enterohemorrhagic Escherichia coli (EHEC) produce attaching and effacing (A/E) lesions, characterized by the formation of actin pedestals, on mammalian cells. A bacterial T3SS effector NleL from EHEC O157:H7 was recently shown to be a HECT-like E3 ligase in vitro, but its biological functions and host targets remain elusive. Here, we report that NleL is required to effectively promote EHEC-induced A/E lesions and bacterial infection. Furthermore, human c-Jun NH2-terminal kinases (JNKs) were identified as primary substrates of NleL. NleL-induced JNK ubiquitylation, particularly mono-ubiquitylation at the Lys 68 residue of JNK, impairs JNK’s interaction with an upstream kinase MKK7, thus disrupting JNK phosphorylation and activation. This subsequently suppresses the transcriptional activity of activator protein-1 (AP-1), which modulates the formation of the EHEC-induced actin pedestals. Moreover, JNK knockdown or inhibition in host cells complements NleL deficiency in EHEC infection. Thus, we demonstrate that the effector protein NleL enhances the ability of EHEC to infect host cells by targeting host JNK, and elucidate an inhibitory role of ubiquitylation in regulating JNK phosphorylation.
| Enterohemorrhagic Escherichia coli (EHEC) can cause attaching and effacing (A/E) lesions to form in the colons of animals and humans, contributing to severe bacterial infection. NleL, an E3 ubiquitin ligase from EHEC O157:H7 is one of the bacterial type III secretion effectors that may be involved in the regulation of A/E lesions. However, NleL’s exact host targets and the detailed mechanistic actions are still unclear. Here, we report that the effector protein NleL effectively promotes EHEC-induced A/E lesions and bacterial infection by targeting the host JNK protein. Specifically, we find that NleL-mediated JNK ubiquitylation abolishes phosphorylation and activation of host JNK, subsequently suppressing the host JNK/AP-1 signaling pathway to favor the formation of EHEC-mediated actin pedestals on the surface of mammalian cells. Collectively, our work has not only discovered the A/E lesion-promoting function of NleL during EHEC infection, but also revealed a novel regulatory mechanism of host JNK protein.
| EHEC is a globally spread, pathogenic Escherichia coli that infects animals and humans [1,2]. Particularly, O157:H7, as the most prominent serotype in the EHEC group, is a leading cause of diarrhea or hemorrhagic colitis in humans [3]. These pathogens belong to a distinct family of enteric bacteria that cause marked cytoskeletal changes and form unique attaching and effacing (A/E) lesions on intestinal epithelium [4,5]. A/E lesions are characterized by effacement of microvilli, intimate adherence between the bacterium and the host cell membrane, and the generation of actin pedestals, polymerized actin structures beneath the adherent bacteria [2]. Although the specific functions of actin pedestals are currently unclear [6], many studies suggest that the capability of A/E pathogens to form actin pedestals correlates with their ability to cause disease in hosts [7–9]. The type III protein secretion system (T3SS), as well as additional EHEC effector proteins, were reported to be involved in actin pedestal formation [10–12], but it remains incompletely understood how pedestal formation can be modulated.
Although the ubiquitin (Ub) system is exclusive to eukaryotes, prokaryotic bacteria have produced many E3 ligase-like effectors [13,14]. Recently, a bacterial T3SS effector NleL (Non-Lee-encoded effector ligase; also named EspX7) from EHEC O157:H7 was shown to be a HECT-like E3 ligase in vitro, with Cys753 as the catalytic site (Fig 1A) [15]. Later biochemical work revealed that NleL interacts with human E2 UbcH7 and is capable of assembling heterotypic Ub chains in vitro [16–18]. While NleL has been proposed to modulate EHEC-induced actin-pedestal formation [19], NleL’s specific host targets and functions in EHEC infection remain elusive.
In this study, we have identified human JNK as the first substrate of the bacterial E3 ligase NleL. The JNK (also known as stress-activated protein kinase, SAPK) family includes three highly homologous isoforms: ubiquitously expressed JNK1 and JNK2, and tissue-specific JNK3 [20]. JNKs are phosphorylated and activated by upstream kinases and regulate a wide range of cellular functions [21]. However, little is known about post-translational modifications other than phosphorylation regulating JNK functions. Here, we report that JNK proteins are ubiquitylated and inactivated by a bacterial effector NleL in EHEC infection, which promotes EHEC-induced A/E lesion formation and infection.
To evaluate the effect of NleL on EHEC infection, a nleL-deletion mutant (ΔnleL) was first constructed by chromosomally inactivating the nleL gene in the parental EHEC O157:H7 Sakai strain (RIMD 0509952) as described previously [22]. Deletion of nleL has no effect on the growth of EHEC in culture medium (S1A Fig). We then assessed the ability of EHEC strains to infect mammalian cells. As a hallmark of EHEC infection, bacteria closely attach to cultured mammalian cells [23]. As shown, deletion of nleL from EHEC significantly reduced bacterial attachment to mammalian cells. More importantly, complementation of the nleL-deletion strain with wild-type NleL (ΔnleL + pNleL), but not the enzymatically-dead NleL mutant C753A (where the active site Cys at position 753 is replaced with Ala) (ΔnleL + pC753A), effectively restored the strong adherence of EHEC to host cells (Fig 1B and 1C). These data indicate that the bacterial effector NleL enhances the ability of EHEC to attach to mammalian cells in a manner dependent on its E3 ligase activity.
To identify the host targets of NleL, a human ORFs library was screened by a yeast two-hybrid (Y2H) system with full-length NleL as the bait. A cDNA encoding human JNK1, Mapk8, was identified in the Y2H screen (S1B Fig). A series of assays were then carried out to confirm the interaction of NleL with JNK1. Both wild-type NleL and the mutant C753A (also NleL-CA) readily co-immunoprecipitated with JNK1, suggesting NleL could form a complex with JNK1 independent of its E3 activity (Fig 1D and 1E and S1C Fig). A GST pull-down assay with the recombinant proteins confirmed that JNK1 can directly interact with NleL or its C753A mutant in vitro (Fig 1F). Compared to NleL170–782, a truncation mutant of NleL frequently used in structural or in vitro biochemical analyses [15,17], full-length NleL was shown to interact with JNK1 with significantly higher affinity (Fig 1G and S1D Fig), suggesting that the N-terminal unordered region of NleL might be involved in the interaction with JNK1. Moreover, we further demonstrated that endogenous JNKs interact with secreted NleL from EHEC in the infected mammalian cells (Fig 1H). Altogether, NleL interacts with host protein JNK1, providing a physical basis for their potential functional interplay.
We next asked whether the interaction between NleL and host JNK might cause JNK ubiquitylation. As shown in Fig 2A, infection with wild-type EHEC O157:H7 increased the ubiquitylation of JNK, but infection with the ΔnleL strain had little or no impact on JNK ubiquitylation. Moreover, complementation of ΔnleL with wild-type NleL (but not the C753A mutant) effectively promoted JNK ubiquitylation in the infected cells. Thus, NleL could induce JNK ubiquitylation in the EHEC-infected host cells.
Although NleL170-782 was sufficient to mediate the assembly of poly-ubiquitin (poly-Ub) chains in vitro [15,16], the full-length form of NleL was used for all subsequent ubiquitylation assays and functional assays because of its stronger interaction with JNK1 and its intact E3 activity in vitro and in vivo (S2A and S2B Fig). Wild-type NleL, but not C753A, was found to efficiently promote mono- and poly-ubiquitylation of JNK1 in vivo and in vitro, depending on its E3 activity (Fig 2B–2D). Thus, these results established human JNK1 as the substrate for NleL.
NleL was previously shown to assemble Lys 6 and/or Lys 48 linked poly-Ub chains in vitro, while auto-ubiquitylation of NleL occurred preferentially via other Ub linkages [15,16]. Here, we found that NleL-catalyzed ubiquitin chains on JNK1 were primarily linked via Lys 27, Lys 29 and Lys 33 (K27, K29 and K33) linkages, especially the K29 linkage (S2C and S2D Fig). As shown in Fig 2E, NleL readily modified JNK1 with mono-Ub and K29-linked Ub chains, which can be completely removed by Usp2cc, the catalytic core of human ubiquitin-specific protease 2 (USP2). Our data also indicated that several E2s, particularly UbcH7, could support NleL-mediated JNK1 ubiquitylation in cells (S3A Fig).
We next mapped potential ubiquitylation sites in JNK1. JNK1α1, as the canonical isoform of human JNK1, contains 29 lysine residues. Trypsinolysis of ubiquitin conjugation yields signature “diGly remnants”, which could be enriched with anti-diGly monoclonal antibody for mass spectrometry (MS) analysis [24,25]. MS analysis of JNK1 purified from NleL-expressing 293T cells identified 6 diGly-containing peptides, which corresponded to ubiquitylation at Lys residues 68, 166, 250, 251, 265, and 308 (Fig 2F). The other five ubiquitylation sites in JNK1 (K140, K153, K203, K222, and K236) were identified by protein-protein docking analysis. Therefore, 11 Lys residues of JNK1 were the putative ubiquitylation sites for NleL (Fig 2G).
To further pinpoint the major Lys residues of JNK1 for NleL-induced ubiquitylation, in vivo ubiquitylation assays were performed with JNK1 mutants bearing Lys-to-Arg substitutions at each potential ubiquitylation site. Single Lys-to-Arg substitution on each of three sites (K153R, K222R or K265R) markedly attenuated NleL-induced JNK1 poly-ubiquitylation (Fig 2H and S3B and S3C Fig). Moreover, combined mutation of these three sites (K153/222/265R, 3KR) significantly abolished NleL-induced poly-ubiquitylation of JNK1 (Fig 2I). On the other hand, the K68R substitution alone completely abolished mono-ubiquitylation of JNK1 by NleL in vivo and in vitro, while the 3KR mutant was still clearly mono-ubiquitylated (Fig 2J–2L). Thus, four Lys residues (K68, K153, K222 and K265) of JNK1 were established as the major NleL-associated ubiquitylation sites, with the K68 residue predominantly responsible for mono-ubiquitylation.
It has been established that activation of JNK signaling constitutes an early cellular response to bacterial infection [26,27]. We next investigated whether NleL functionally regulates this JNK role. As shown in Fig 3A and 3B, while wild-type EHEC induced slight phosphorylation of endogenous JNK in mammalian cells, infection by ΔnleL strain elicited much stronger JNK phosphorylation. Complementation of the ΔnleL strain with wild-type NleL restored the EHEC inhibitory effect on JNK phosphorylation, but the C753A-mutant–complemented ΔnleL strain did not (Fig 3A). Additionally, TNFα stimulation did not induce JNK phosphorylation when the mammalian cells were infected by the EHEC overexpressing wild-type NleL (but not C753A) (Fig 3B). These results prompted us to further investigate whether NleL alone is sufficient to suppress JNK phosphorylation.
Indeed, overexpression of wild-type NleL, but not NleL C753A, efficiently reduced the basal phosphorylation level of JNK (Fig 3C and 3E). Even when the cells were stimulated by TNFα at either low (1.0 ng/ml) or high (10.0 ng/ml) concentration, wild-type NleL significantly suppressed JNK phosphorylation (Fig 3C and 3D). Thus, the E3 ligase activity of NleL is sufficient to suppress JNK phosphorylation in host cells.
We further investigated the relevance of NleL-induced ubiquitylation to JNK1 phosphorylation. In the presence of wild-type NleL (but not C753A), JNK1 was ubiquitylated efficiently but poorly phosphorylated, suggesting that NleL-mediated JNK1 ubiquitylation might adversely impact JNK1 phosphorylation (Fig 3F). Moreover, a JNK kinase assay with recombinant c-Jun as the substrate revealed that the E3 activity of NleL markedly impaired the total kinase activity of JNK1 (Fig 3G). We then tried to explore the roles of JNK1 mono-ubiquitylation at the K68 residue and poly-ubiquitylation at other sites (K153, K222 and K265) in suppressing JNK1 phosphorylation. As shown in Fig 3H, NleL almost completely abolished TNFα-induced phosphorylation of wild-type JNK1 as well as the 3KR mutant; however, NleL had only negligible effects on the phosphorylation of the JNK1 K68R mutant. In addition, NleL had no effect on the phosphorylation of p38 or Erk, two other members of the MAPK superfamily (Fig 3I). Therefore, NleL-induced JNK1 ubiquitylation, particularly mono-ubiquitylation of K68, specifically inhibited JNK1 phosphorylation.
JNK family proteins share over 90% sequence homology to each other (Fig 4A and S4 Fig), including a conserved Lys 68 residue. We next asked whether NleL might also target JNK2 or JNK3, the other two members of the JNK family. Binding assays confirmed that NleL indeed interacted with JNK2 and JNK3 (Fig 4B and 4C). NleL, but not C753A, effectively catalyzed ubiquitylation of JNK2 and JNK3 (Fig 4D and S5A–S5C Fig), and inhibited their phosphorylation as well (Fig 4E and S5D Fig). Furthermore, similar to JNK1, ubiquitylation of JNK2/3 by NleL was negatively correlated with JNK2/3 phosphorylation and activation (Fig 4F and 4G). Thus, NleL appeared to target all the members of JNK family.
Currently, two MAP2Ks, MKK4 and MMK7, are known to phosphorylate and activate JNK proteins, with MKK7 being more specific to JNK [28]. It is natural to speculate that NleL-induced ubiquitylation might suppress JNK phosphorylation by either inhibiting the kinase activity of MKK7 or disrupting the interaction between MKK7 and JNK. Since NleL had little or no effect on the phosphorylation of MKK7 (Fig 5A), we proceeded to explore the latter possibility. Usually, JNK1 or JNK2 interacts with MKK7 in host cells. NleL readily reduced MKK7 association with JNKs, but the C753A mutant had little effect (Fig 5B and 5C and S6A and S6B Fig), suggesting that NleL impairs the MKK7-JNK interaction independent of direct competition against MKK7. We also ruled out the possibility that NleL might obstruct MKK4 phosphorylation or MKK4-JNK association (Fig 5D and S6C Fig). Moreover, NleL disrupted the recruitment of wild-type JNK1 to MKK7, but had no effect on the interaction of K68R mutant with MKK7, although wild-type JNK1 and its K68R mutant had the same ability to interact with MKK7 (Fig 5E and 5F). Thus, the K68 residue of JNK1 is required for NleL to suppress the MKK7-JNK interaction.
Since E. coli do not have the ubiquitin system, NleL should not conjugate Ub to JNK1 in bacterial cells. If NleL could mediate modifications in addition to ubiquitylation, it could still potentially modify and inactivate JNK1 when co-expressed with JNK1 in E.coli. However, we found that MKK7 readily phosphorylated purified JNK1 that was co-expressed with either GST-tagged NleL or GST alone in the E. coli BL21 (DE3) strain (Fig 5G). These data suggested that NleL did not inactivate JNK1 through other post-translational modifications.
Altogether, we conclude that NleL-induced ubiquitylation at the K68 residue of JNK1 suppresses the phosphorylation of JNK1, through disrupting the JNK1-MKK7 interaction.
We found that NleL-mediated JNK inactivation was independent of NF-κB signaling (S7A Fig). Next, we examined the possible effects of NleL on downstream of JNK. As expected, wild-type NleL suppressed the basal-level phosphorylation of endogenous c-Jun, a bona fide physiological substrate of JNK, while the C753A mutant did not (Fig 6A). Immunofluorescence microscopy analysis also revealed that EGFP-tagged NleL, but not EGFP, markedly reduced c-Jun phosphorylation (Fig 6B and S7B Fig). Consistently, overexpression of NleL (but not C753A) impaired TNFα-stimulated phosphorylation of c-Jun (Fig 6C). However, depletion of Jnk1/2 using short hairpin RNAs in mammalian cells almost completely abolished the inhibitory effect of NleL on c-Jun phosphorylation (Fig 6C). Therefore, NleL inhibits c-Jun phosphorylation by targeting JNK proteins.
As c-Jun is a major component of the AP-1 transcription factor [29], we next investigated the regulation of AP-1 activity by NleL. An AP-1 luciferase reporter assay showed that wild-type NleL, but not C753A, suppressed AP-1 activity in cells (Fig 6D and S7C and S7D Fig). AP-1 is known to regulate the expression of a large number of genes, e.g. cyclin D1 (CCND1) [30]. As expected, the basal expression of CCND1 in mammalian cells was down-regulated by NleL (but not C753A) (Fig 6E). Consistently, NleL also diminished JNK phosphorylation and CCND1 expression induced by different stimulators (Fig 6F and 6G and S7E and S7F Fig). These results suggest that the E3 activity of NleL is required to suppress AP-1 activity and the expression of AP-1 target genes.
Based on previous work by multiple groups [4,5,31], some AP-1 targets (e.g. CD44, TPM1, ARPC1B and EZR) are known to be important for the formation of EHEC-induced pedestals. A protein-protein interaction (PPI) network analysis was performed to characterize the potential interplay among the actin-associated proteins targeted by AP-1 and the host proteins identified in EHEC actin pedestals (Fig 6H). The data strongly suggested an emerging role of AP-1 signaling in regulating the formation of EHEC actin pedestals. Furthermore, we found that ectopically expressed NleL suppressed the phosphorylation of VSAP (Fig 6I), one of the critical components in actin pedestals that was recently reported to be modulated by JNK/AP-1 signaling [32]. Meanwhile, CCND1, another AP-1 target protein shown above to be down-regulated by NleL, also interacts with AP-1 actin-associated targets and EHEC pedestal proteins (Fig 6H). Thus, NleL might promote the formation of EHEC actin pedestals by modulating the host JNK/AP-1 pathway.
We next explored the role of NleL on the formation of EHEC actin pedestals. Compared to the wild-type EHEC strain, deletion of nleL from EHEC reduced actin-pedestal formation on mammalian cells (Fig 7A). Complementation of ΔnleL strain with NleL, but not C753A, significantly restored the EHEC pedestal-forming abilities (Fig 7A and 7B). On the other hand, Jnk1/2 depletion in HeLa cells promoted actin-pedestal formation by each of the EHEC strains. Additionally, ΔnleL and wild-type EHEC had similar pedestal-forming abilities on Jnk1/2-silenced host cells (Fig 7A and 7B). Thus, NleL promoted the formation of EHEC actin pedestals through targeting host JNKs.
C. rodentium has been used as an alternative approach to study EHEC. However, differing from our findings on the role of NleL in EHEC infection, NleL-deficiency was found to have no effect on the ability of C. rodentium to form actin pedestals and attach to HeLa cells or mice colon (S8 and S9 Fig). Instead, overexpression of EHEC NleL in the ΔnleL C. rodentium suppressed the ability of C. rodentium to attach to HeLa cells and form actin pedestals. Thus, C. rodentium was not a suitable model system to study NleL.
As an A/E pathogen, the pedestal-forming ability of EHEC is considered to correlate with its capability to colonize host [4]. Caco-2 is a human colorectal epithelial cell line that can form an epithelial cell monolayer when cultured for 6 days; continued Caco-2 monolayers growth for 21 days can become differentiated with a brush border that more closely resembles human intestinal epithelium [33–35]. We next performed EHEC infection assays on 6-day-old and 21-day-old Caco-2 monolayers. NleL increased the ability of EHEC to colonize Caco-2 monolayers in a manner dependent on NleL’s E3 ligase activity. Moreover, treatment of Caco-2 with the JNK inhibitor SP600125 effectively rescued the capability of the ΔnleL strain to attach to the Caco-2 monolayer (Fig 7C and S10A and S10B Fig). Similar results were also observed in HeLa cells (S11 Fig). We additionally observed EHEC-induced A/E lesions on 21-day-old Caco-2 monolayers (Fig 7D). Wild-type EHEC (but not ΔnleL) caused A/E lesions on 21-day-old Caco-2 monolayers, marked by microvillus damage. Complementation of the ΔnleL strain with wild-type NleL, but not the C753A mutant, restored the EHEC A/E lesion-forming ability; SP600125 treatment also rescued the ability of ΔnleL to form A/E lesions on the Caco-2 monolayer. These data suggested that the suppression of JNK1/2 functions in host cells could compensate for nleL-deletion-caused loss of the EHEC ability to colonize the host and form A/E lesions. JNK proteins are thus identified as the critical host targets for NleL to promote EHEC O157:H7 infection (Fig 7E).
In this work, we have demonstrated that NleL, a bacterial effector and HECT-like E3 ubiquitin ligase from EHEC O157:H7, is critically involved in promoting actin-pedestal formation during EHEC infection. Notably, this finding is different from a previous report by Piscatelli et al., in which NleL was shown to down-regulate the EHEC-induced formation of actin pedestals [19]. If NleL down-regulates actin-pedestal formation in infection as suggested by Piscatelli et al., the presence of NleL should have disrupted bacterial attachment to the host cells. However, results from the same study by Piscatelli et al. indicated that EHEC NleL, reintroduced into the nleL-deleted C. rodentium, was actually required for efficient infection in vivo. Further work is needed to clearly understand the discrepancy between our findings and the findings of Piscatelli et al.
As described above, C. rodentium cannot be a suitable surrogate for EHEC O157:H7 to study NleL. Multiple studies have indicated that EHEC can trigger actin-pedestal formation in host cells in ways differing from C. rodentium [36–38]. While EHEC O157 uses the bacterial effector Tir and TccP adaptor protein to trigger actin polymerization, C. rodentium relies on the phosphorylation of Tir Y471 and the host protein Nck. This may partially account for why the NleLs from these two pathogens have very different effects on actin-pedestal formation, despite their high homology. Given that EHEC NleL targets and suppresses the JNK/AP-1 pathway, it is likely that NleL plays roles in the later stages, rather than the triggering stage, of actin polymerization. Further work is warranted to elucidate the mechanisms underlying why NleL has different roles in these two bacteria. On the other hand, we found that the differentiated Caco-2 monolayers (grown for 21 days) could be used as an in vitro infection model to study A/E lesions. Caco-2 monolayers can mimic human colonic epithelium for EHEC infection, providing an alternative approach to in vivo infection.
Different Ub chain linkages have different impacts on targeted proteins (e.g. K48- or K63-linked Ub chains usually cause protein degradation) [39–41]. Here, NleL primarily assembled mono-Ub and K29-linked Ub chains on JNK, suggesting a non-proteolytic function of NleL-induced JNK ubiquitylation. Instead, NleL-induced mono-ubiquitylation at K68 of JNK, but not the poly-ubiquitylation at other residues, disrupted the interaction between JNK and MKK7. Although a recent report showed that JNK1 binds MKK7 using multiple binding sites [42], K68 of JNK1 is not in the MKK7-JNK1 binding interface. Other mechanisms, such as those involving allostery, may underlie how JNK1 mono-ubiquitylation disrupts the MKK7-JNK1 interaction.
NleL-mediated JNK mono-ubiquitination appeared to drastically suppress JNK activity, although the mono-ubiquitylated sub-population of JNK1 seemed limited (5% ~ 10% of all cellular JNK1). Currently there are two accepted explanations: 1) cellular proteins can exist in different subcellular populations, so targeted ubiquitylation of a particular subpopulation can be sufficient to generate significant impacts on a specific pathway; 2) NleL-induced JNK1 ubiquitylation should be a dynamic process balanced by removal of ubiquitin by deubiquitylating enzymes (DUBs), as there are plenty of DUBs in mammalian cells [43–45]. In other words, although only part of JNK were observed to be modified by NleL, it is highly possible that most of the JNK molecules undergo NleL-mediated ubiquitylation and then deubiquitylation by DUBs.
The JNKs are master regulators in mammalian cells [46]. It is well established that JNKs are phosphorylated by upstream kinases and then activate downstream targets. However, little was known about the posttranslational modifications other than phosphorylation that might occur on JNKs until recently, when several ubiquitylation and acetylation sites of endogenous JNKs were uncovered by proteomic analyses [24,47,48]. Whether these potential modifications might impact the functions of JNKs remains poorly understood. We thus demonstrate for the first time that ubiquitylation of JNKs by the bacterial effector NleL negatively regulates the function of JNKs. It will be intriguing to investigate whether an unknown endogenous E3 ligase might exist to catalyze the ubiquitylation of JNKs and regulate JNK signaling in mammalian cells.
Worldwide, outbreaks of EHEC O157:H7 infection constitute constant and serious threats to human population and live stocks, without effective treatments [49]. Antibiotic therapy is generally contraindicated as it may promote expression of Stx toxin protein and increase the risk of the hemolytic–uremic syndrome (HUS) [50–52]. Only supportive care can be provided for the infected patients who have developed HUS. Thus, a diversity of treatment and prevention strategies should be developed to protect against EHEC. As NleL promotes EHEC infection by suppressing host JNK, disrupting the NleL-JNK interaction may represent a novel strategy against EHEC O157:H7 infections.
All animal use procedures were in strict accordance with the Guide for the Care and Use of Laboratory Animals (8th edition, National Research Council, 2011), approved by the Institutional Animal Care and Use Committee (Protocol number SIBCB-S330-1512) of Shanghai Institute of Biochemistry and Cell Biology, Chinese Academy of Sciences.
DNAs for NleL amplified from the genomic DNA of E. coli O157:H7 Sakai strain was inserted into pCDNA3.0 and pEGFP-C1 for mammalian expression, and pGEX-4T-1 for recombinant expression in E. coli. NleL DNA was also ligated into the pTRC99A vector for complementation in EHEC (under the trc promoter; pTRC99A is kindly provided by Dr. Xueli Zhang from Tianjin institute of industrial biotechnology, Chinese academy of sciences). Genes for encoding JNK1α1, JNK2α2 (kindly provided by Dr. Jinzhang Zeng from Xiamen University, China), JNK3α2 (Addgene #13759) and MKK7 (Addgene #14623) were cloned to pCDNA3.0 vector with a C-terminal Flag tag (or HA tag), and pET28a vector with 6× His tag. AP-1 luciferase reporter plasmid was a gift from Dr. Jine Yang (Sun Yat-sen University, Guangzhou, China). Plasmids expressing HA-tagged Ub and its mutants were described previously [53]. All the point mutations were generated by using the QuickChange Site-Directed Mutagenesis Kit (Stratagene) according to manufacturer’s protocol. All constructs were verified by DNA sequencing.
Antibodies for JNK1 (2C6) (#3708), phospho-SAPK/JNK (81E11) (#4668), phospho-SAPK/JNK (G9) (#9255), phospho-c-Jun (54B3) (#2361), phospho-c-Jun (#9261), caspase-3 (#9662), VASP (9A2) (#3132) and phospho-VASP (Ser157) were obtained from Cell Signaling Technology. Anti-JNK2 antibody (EP1595Y) (ab76125), anti-MEK7 antibody (EP1455Y) (ab52618), anti-JNK1/2/3 antibody (ab179461) and anti-MEK7 (phospho S271 + T275) (ab4762) were purchased from Abcam. Rabbit anti-HA antibody (H6908), mouse monoclonal anti-Flag antibody (F1804), and Anti-Flag M2 affinity gel (A2220) were from Sigma. Other antibodies were purchased from BD pharmingen for anti-JNK1 (551197), Santa Cruz Biotechnology for anti-ubiquitin (P4D1), Bioword for anti-c-Jun (G237), Absci for anti-MKK4 (#AB21132) and anti-phospho-MKK4 (Ser80) (#AB11177), HangZhou HuaAn Biotechnology for anti-His6 tag (M0812-3), Proteintech for mouse anti-Cyclin D1 (60186-1-Ig), rabbit anti-Flag tag (20543-1-AP) and mouse anti-GAPDH (60004-1-Ig). Peroxidase-conjugated goat anti-rabbit and goat anti-mouse IgG secondary antibodies were purchased from Jackson ImmunoResearch Laboratories, Inc. Chemicals were purchased from Sigma if not otherwise indicated: PS-1145 (Santa Cruz Biotechnology), ATP (Thermo Scientific Fermentas). JNK inhibitor SP600125 (S1460) was from Selleck. HeLa, Caco-2 and HEK293T cells were obtained from the American Type Culture Collection (ATCC). All cell culture products were from Corning.
HEK293T, HeLa, Caco-2 cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) (Hyclone) supplemented with 10% fetal bovine serum (FBS), 2.0 mM L-glutamine, 100 units/ml penicillin and 100 mg/ml streptomycin. Cells were maintained in 5.0% CO2 at 37°C. Transfections were carried out with Lipofectamine 2000 (Invitrogen) according to the manufacturer’s instruction.
Enterohemorrhagic Escherichia coli (EHEC) O157:H7 Sakai strain (RIMD 0509952) and C. rodentium (CR) strain DBS100 (ATCC 51459) were used as wild-type strains. These bacteria were commonly cultured at 37°C LB broth. Deletion of nleL from EHEC O157:H7 genome was achieved through standard homologous recombination, as reported previously [22]. For CR, the nleL-deleted mutant and escR-deleted mutant had been constructed through a homologous recombination method “Gene doctoring” [54,55]. The mutants were verified by PCR and DNA sequencing. For rescue assay, the ΔnleL strain was transformed with plasmid encoding wild-type Flag-tagged NleL or its C753A mutant.
The infection was performed as described before with slight modifications [56,57]. Briefly, EHEC or CR strains were cultured overnight in 2 × YT (16.0 g/L tryptone, 10.0 g/L yeast extract, 5.0 g/L NaCl) medium without shaking at 37°C. Bacterial cultures were then diluted by 1:40 with serum-free DMEM medium, and cultured for an additional 3 ~ 4 h at 37°C in the presence of 5% CO2 to induce the expression of type III secretion system before infection. For the complementation assay in EHEC or CR, the medium was added with 1.0 mM Isopropyl-B-D-thiogalactopyranoside (IPTG). Bacterial cells were then collected and suspended in PBS. After measuring the O.D. of cultured bacteria, infections were performed with mammalian cells at a multiplicity of infection (MOI) of 100:1 or 20:1, if not indicated otherwise, with a centrifugation at 800 g for 10 min, and then proceeded with incubation at 37°C in 5% CO2 for 2 ~ 3 h. After that, cells were washed three times and further cultured in fresh DMEM medium for another 2.5 ~ 5 h. Then the infected cells were subjected to immunofluorescence assay or IB analyses with indicated antibodies.
E. coli BL21 (DE3) strains harboring the corresponding recombinant plasmids were grown in LB medium supplemented with appropriate antibiotics. Protein expression was induced overnight at 16°C with 0.3 mM IPTG when OD600 reached 0.6 ~ 0.8. To purify His6-tagged proteins, bacteria were harvested and lysed in lysis buffer containing 50mM Tris-HCl (pH 7.6), 300 mM NaCl, 40 mM imidazole and 5.0 mM beta-mercaptoethanol, and then proteins were purified with affinity chromatography using Ni-NTA beads (Qiagen) according to the manufacturer’s instruction. For GST-fusion proteins, purifications were performed by affinity chromatography using Glutathione Sepharose Fast Flow beads (GE Healthcare). Eluted proteins were further dialyzed overnight at 4°C. Recombinant proteins were concentrated and then frozen-stored in a buffer containing 50 mM Tris-HCl (pH 7.4), 300 mM NaCl and 15% Glycerol (V/V). Protein concentrations were determined using Bradford colorimetric assays, with their protein purities examined with SDS-PAGE followed by Coomassie Blue staining.
Nucleotide sequences for the human Jnk1/2-specific shRNAs used were described before [58]. For stable knockdown of Jnk1/2, lentiviral particles harboring specific shRNA expression vector (pLKO.1; Sigma-Aldrich) were produced by transfection of HEK293FT cells with the shRNA expression plasmid and lentiviral packaging mix. Target cells (HEK293T and HeLa) were incubated with the viral supernatant in the presence of 8 μg/ml polybrene (Sigma) and selected with 2 μg/ml puromycin (Clontech).
As previously described with minor modifications [59], in vitro ubiquitylation assays were carried out in a 30 μl reaction system containing E1 (100 ng), His6-tagged UbcH7 (200 ng), GST-NleL or C753A (500 ng), Flag-JNK1 (500 ng) and ubiquitin (1.0 μg) in ubiquitylation buffer (50 mM Tris–HCl, pH 7.5, 5.0 mM MgCl2, 2.0 mM ATP, 1.0 mM DTT) at 37°C for 60 min. After the reaction was terminated by adding one-tenth volume of 10 × SDS sample buffer, the resulted mixtures were boiled and then subjected to SDS-PAGE, followed by IB analysis with indicated antibodies.
In vivo ubiquitylation were also carried out as described previously [60]. Cells were collected and lysed in denaturing RIPA buffer containing 50 mM Tris-HCl (pH 7.6), 150 mM NaCl, 1.0% Triton X-100, 1.0% sodium deoxycholate and 0.1% SDS supplemented with 1.0% protease inhibitor cocktail (Roche). Cell lysates were then centrifuged at 15,000 g for 10 min, 4°C. Following overnight incubation of anti-Flag beads with cell lysate at 4°C, the beads was washed five times with RIPA buffer. Finally the beads were boiled in SDS sample buffer and subjected to IB analysis with indicated antibodies.
HEK293T cells in 6 cm dish were transfected with 5.0 μg pCDNA3-Flag-JNK1 and cultured for 24 h. Cells were then lysed in IP buffer (containing 50 mM Tris-HCl, pH 7.6, 150 mM NaCl, 1.0% Triton X-100), and centrifuged (15,000g) for 15 min at 4°C. The supernatants were then subjected to IP with anti-Flag M2 beads to enrich Flag-JNK1. Then Flag-JNK1 was eluted by using 3 × Flag peptides (150 μg/ml in PBS) at 4°C for 30min. The eluted protein was verified by Coomassie Blue staining of SDS-PAGE gels.
For Co-IP assay, 24 h after transfection with indicated plasmids, HEK293T cells were lysed in Triton X-100 buffer (50 mM Tris-HCl at pH 7.4, 150 mM NaCl, 1% Triton X-100) plus 1% protease inhibitor cocktail at 4°C for 1 h. The corresponding antibody-conjugated Sepharose beads were added into the lysates supernatant. Following incubation overnight, the beads were washed five times (10 min each). For GST pull-down assay, purified GST-NleL, His-JNKs or cell lysate (appropriate volume) were mixed in total 400 μL reaction system (50 mM Tris-Cl at pH 7.5, 150 mM NaCl, 1% (v/v) Triton X-100, 1 mM EDTA) at 4°C for 6 h. The beads were then pelleted and washed for 5 times (10 min incubation at 4°C for each washing). Then the recovered beads were boiled in 1× SDS-PAGE loading buffer and subjected to SDS-PAGE, followed by Coomassie Blue staining or IB with indicated antibodies.
For immunofluorescence staining, cells were fixed with 4% paraformaldehyde for 20 min at room temperature, permeabilized for 10 min with 0.2% Triton X-100 in PBS, and then blocked for 60 min with 1.0% BSA, followed by incubation with indicated antibodies. All the cell nuclei were counterstained with DAPI before imaging.
In vitro kinase assays were performed as described before [61]. Cells were transfected with plasmids encoding NleL and Flag-JNK1 or JNK2. 24 h after transfection, cells were treated with or without TNFα (10 ng/m1) for 10 min. Cell were then lysed and subjected to IP with anti-Flag M2 beads in M2 buffer (containing 20 mM Tris-HCl, pH 7.6, 250 mM NaCl, 0.5% NP-40, 3.0 mM EDTA, 3.0 mM EGTA). Then the beads with the precipitated JNK proteins were washed and subjected to in vitro kinase reaction (30 μl total) containing 1.0 μg of GST-c-Jun (1-79aa) in the kinase buffer (30 mM HEPES, pH 7.4, 3.0 mM DTT, 30 mM PNPP, 0.2 mM NaVO3, 30 mM MgCl2) supplemented with 2.0 mM ATP. The in vitro kinase assays involving human JNKs were carried out at 30°C for 60 min. Reactions were stopped by adding 1 × SDS-PAGE loading buffer and subjected to IB analysis with anti-p-c-Jun antibody.
Luciferase reporter assays were performed as described previously [62]. HEK293T cells seeded in 24-well plates were transiently co-transfected with 50 ng of pAP-1-Luc and 10 ng of pRL-tk-Luc together with or without 500 ng of pCDNA3-NleL, using Lipofectamine 2000 reagent. The total amounts of DNA were kept constant by supplementing empty vector (pCNDA3.0). 24 h after transfection, cells were subjected to TNFα (R&D Systems Inc., 10 ng/ml) or PMA (Sigma, 20 nM) for indicated time. Cells were then lysed in 1 × Passive Lysis Buffer (5 × concentrate diluted in ddH2O, Promega) for 15 min at room temperature with vigorous shaking. AP-1 activities were finally determined using the dual luciferase assay kit (Promega). Results were independently replicated for at least three experiments.
Total RNAs from indicated cells were extracted using TRIzol (Invitrogen). RNA concentrations were determined on Nanodrop ND-1000 sepectrophotometer. 0.1 μg total RNAs were used to cDNA synthesis with the ReverTra Ace qPCR RT kit (FSQ-201, TOYOBO), according to the manufacturer’s instructions. The relative levels of genes expression (normalized to those of Gapdh) were assessed in triplicate wells of a 96-well reaction plate by subjecting 10 ng cDNAs per well to a Bio-Rad CFX96 Touch Detection System with SYBR Green chemistry using the following primers:
human Ccnd1-forward: 5’-CCGTCCATGCGGAAGATC-3’;
human Ccnd1-reverse: 5’ -GAAGACCTCCTCCTCGCACT-3’ [63];
human Gapdh-forward: 5’-TGCCCTCAACGACCACTTTG-3’;
human Gapdh-reverse: 5’-TTCCTCTTGTGCTCTTGCTGGG-3’ [64].
qRT-PCR data were analyzed on Bio-Rad CFX Manager 3.0.
Protein sequences were retrieved from the RefSeq database, aligned with ClustalW2, and further processed on GeneDoc.
Caco-2 cells were fixed in 2.5% glutaraldehyde and then processed for scanning electron microscopy (SEM) analysis as previously described [65]. SEM samples were examined at 25 kV using a FEI Tecnai G2 Spirit TEM (SIBCB, China).
Y2H screening with full-length NleL as the bait was performed as described previously [53].
Statistical significance of the data was determined using the Student’s t-test. In all experiments, only P value of < 0.05 was considered to be statistically significant.
|
10.1371/journal.pntd.0002188 | Sulfated Polysaccharide, Curdlan Sulfate, Efficiently Prevents Entry/Fusion and Restricts Antibody-Dependent Enhancement of Dengue Virus Infection In Vitro: A Possible Candidate for Clinical Application | Curdlan sulfate (CRDS), a sulfated 1→3-β-D glucan, previously shown to be a potent HIV entry inhibitor, is characterized in this study as a potent inhibitor of the Dengue virus (DENV). CRDS was identified by in silico blind docking studies to exhibit binding potential to the envelope (E) protein of the DENV. CRDS was shown to inhibit the DENV replication very efficiently in different cells in vitro. Minimal effective concentration of CRDS was as low as 0.1 µg/mL in LLC-MK2 cells, and toxicity was observed only at concentrations over 10 mg/mL. CRDS can also inhibit DENV-1, 3, and 4 efficiently. CRDS did not inhibit the replication of DENV subgenomic replicon. Time of addition experiments demonstrated that the compound not only inhibited viral infection at the host cell binding step, but also at an early post-attachment step of entry (membrane fusion). The direct binding of CRDS to DENV was suggested by an evident reduction in the viral titers after interaction of the virus with CRDS following an ultrafiltration device separation, as well as after virus adsorption to an alkyl CRDS-coated membrane filter. The electron microscopic features also showed that CRDS interacted directly with the viral envelope, and caused changes to the viral surface. CRDS also potently inhibited DENV infection in DC-SIGN expressing cells as well as the antibody-dependent enhancement of DENV-2 infection. Based on these data, a probable binding model of CRDS to DENV E protein was constructed by a flexible receptor and ligand docking study. The binding site of CRDS was predicted to be at the interface between domains II and III of E protein dimer, which is unique to this compound, and is apparently different from the β-OG binding site. Since CRDS has already been tested in humans without serious side effects, its clinical application can be considered.
| There is no specific approved antiviral and vaccine for treatment or prevention of dengue, an acute mosquito-transmitted viral disease that affects more than 50 million people each year. Dengue virus (DENV) entry is a critical step that establishes the infection and enables virus replication. Curdlan sulfate (CRDS) is known to inhibit the entry and propagation of HIV-1 in the laboratory. Here we applied a computational binding site identification strategy, which suggested that CRDS could be a probable entry inhibitor of the viral surface E protein. CRDS potently blocked DENV infection at an early stage of the virus lifecycle in vitro. In addition, CRDS prevented antibody dependent enhancement, which is considered to be one of the most important clinical observations in DENV-infected patients. CRDS shows a favorable selectivity index against all serotypes of DENV. Further computational docking indicates that the compound binds to a pocket on the DENV E protein. Since CRDS has already been tested in humans without serious side effects, it can be a good candidate for clinical application.
| Globally, an estimated 50–100 million people are infected with DENV each year [1]. Dengue is the most important mosquito-transmitted viral disease in the world, particularly in the tropical and subtropical countries [2], [3]. The virus is transmitted to humans via the mosquito vectors, Aedes aegypti and Aedes albopictus. There are four antigenically distinct serotypes of DENV, DENV-1 to DENV-4 [1]. All serotypes cause a range of human diseases, ranging from mild dengue fever (DF) to dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS), which can be fatal [4]. Primary infection by the virus induces immunity against the infecting serotype. However, a secondary infection with a different serotype has been shown to enhance the risk of developing DHF/DSS, a phenomenon termed antibody-dependent enhancement (ADE) [5].
Dengue virus binds to, and enters, a permissive host cell via uncharacterized receptors, undergoing receptor–mediated endocytosis. Upon acidification of the endocytic vesicle, viral and vesicular membranes fuse, allowing entry of the nucleocapsid into the host cytoplasm. The viral RNA is uncoated and released in the cytoplasm and it is directly accessible for translation. Virus assembly occurs in the endoplasmic reticulum, and upon maturation of the virions transported in secretory vesicles, using the Golgi network, mature viruses are released from the cell [6]. This multi-step DENV infection cycle presents potential drug targets during entry, viral membrane fusion, translation, assembly, and maturation. Traditionally, antiviral agents designed against Flaviviruses have focused on the non-structural proteins required for viral RNA replication, specifically, inhibition of essential enzymes like NS3 protease and NS5 RNA-dependent RNA polymerase [7]–[10]. However, recent advances in the X-ray crystallographic and electron microscopic analysis of the DENV structure and its component proteins have led to the identification of other potential targets for drug development [11]–[13]. The E protein has emerged, due to such studies, as a promising target for the inhibition of virus entry into cells [14]–[20]. The DENV-2 E protein is a class II viral fusion protein, consisting of three domains: domain I is the central structure, domain II is the dimerization domain containing the fusion peptide, and domain III has the putative receptor-binding site. Crystal structures of the DENV E protein at different stages of virus life cycle reveal that the protein exists in different conformations and states of polymerization, brought about by domain rearrangements, in the mature, immature and the post-fusion stages [21]. Structural studies of the E protein by Modis et al. [13] revealed a hydrophobic pocket near the hinge region between domains I and II which was found to accommodate a molecule of the detergent N-octyl-beta-d-glucoside (βOG). These studies have provided impetus for antiviral drug development targeting the viral entry process.
The anti-HIV effect of CRDS was first discovered based on our observation that several sulfated polysaccharides, dextran, xylofuranan, and ribofuranan, but not their non-sulfated counterparts, completely prevented HIV-induced cytopathic effects with very high selectivity indices [22]–[23]. CRDS, with its branched β-d-(1→3) glucan backbone with piperidine-N-sulfonic acid, consists of molecules with various molecular masses and degrees of sulfation. We showed that the anti-HIV activity of CRDS was positively correlated with its molecular weight [23]. The anti-coagulant activity of CRDS was significantly lower than that of standard dextran sulfate and heparin [24]. CRDS was thus expected to have potential as an AIDS drug through inhibition of virus entry in the early stage of infection. CRDS has also been shown to be effective against Plasmodium falciparum in vitro [25]. Accordingly, Phase I trial of CRDS against HIV infection and Phase II trial against severe/cerebral Plasmodium falciparum malaria had been performed in the US and in Thailand and South Africa, respectively. The results showed that the treatment was well tolerated by the patients and it showed some clinical benefits [26].
In the present study, based on a preliminary in silico blind docking study which indicated that CRDS could be a probable inhibitor of the DENV E protein, we have characterized its inhibitory activity through a cell-based anti-DENV screening effort and identified that this polysaccharide can block DENV at both the binding and fusion steps very efficiently. Our in silico docking model indicates that the compound binds to a pocket on the DENV E protein. CRDS shows a favorable selectivity index against all serotypes of DENV. Since the compound has already been tested in humans without serious side effects, it provides a possibility for clinical application.
The coordinates of the DENV E protein were obtained from PDB from the crystal structure 1OKE [13]. The crystal structure details the E protein in its dimeric pre-fusion conformation. For the purpose of the study, the crystal structure was modified by the Protein Preparation Wizard module of Schrodinger Suite 2012 (Schrodinger).
The binding site identification of the CRDS in the E protein was performed by the blind docking method using the Molegro Virtual Docker (MVD) program (Molegro). Overlapping grids of 30 Å radius were used to define the search space on the E protein. The grid based MolDock scoring function was used to define the energy terms to rank the potential binding sites [27]. The MolDock Simplex evolution algorithm was chosen for the prediction. A population size of 50, with 1500 maximum iterations was used over ten runs per grid. The simplex minimization procedure was performed with 300 iterations, and the neighbor distance factor set to 1.00. For pose generation, the energy threshold was set to 100 [27].
The Induced Fit module of Schrodinger Suite 2012 (Schrodinger) was used to predict the best binding pose of the E protein, taking into account the conformational changes induced by the binding of the CRDS molecule [28]. The blind docking procedure employed to predict the binding pocket predicted that the CRDS might form strong H-bonding interactions with Arg2. Based on these results, the search grid was constructed around Arg2 of one of the two chains of the E protein, the size set to 26 Å around each side of the residue. The binding pocket predicted by the blind docking procedure was used as the search space. Docking was done by allowing receptor flexibility to account for the conformational changes induced upon ligand binding. The ligand too was docked flexibly to sample all possible conformations of the ligand due to torsions about the rotatable bonds. The van der Waals radii for both the receptor and ligand were scaled by a factor of 0.50. Residues within 5 Å of the docked ligand were subjected to refinement [29]–[31].
The root mean square deviation of the best docked pose of the E protein, as determined by Induced Fit with CRDS, from its crystal structure conformation (pre-fusion) was determined using the rmsd.py script (Schrodinger Suite 2012 script collection). The RMSD per individual residue was determined using the rmsd_by_residue script of Schrodinger Suite 2012 script collection.
LLC-MK2 cells were grown in Eagle's minimum essential medium (EMEM)(SIGMA, St. Louis, MO) supplemented with 10% fetal bovine serum (FBS) (Invitrogen, Carlsbad, CA) with penicillin and streptomycin. HL-60, THP-1 and Raji-DC-SIGN cells were maintained in RPMI 1640 (Invitrogen, Carlsbad, CA) containing 10% FBS, and incubated at 37°C in 5% CO2. Aedes albopictus C6/36 cells were maintained in RPMI-1640 medium with 25 mM HEPES supplemented with 8% FBS and incubated at 28°C. A549 replicon cells (containing DENV-2 NS genes) were grown as described elsewhere [32].
Clinical samples of DENV-1 and DENV-4 and laboratory-adapted New Guinea C (NGC) strain of DENV-2 were kindly provided by Dr. Justin JH Chu (National University of Singapore, Singapore). DENV-3 strain EDEN8630K1 was isolated by EEO [33]. These viruses were propagated in C3/36 cells grown in RPMI-1640 medium supplemented with 25 mM HEPES, 8% FBS and antibiotics. The supernatant from infected cells was centrifuged to remove cell debris, then aliquoted and stored at −80°C.
For virus concentration, supernatant from DENV-infected cell cultures was loaded onto a 30% (wt/wt) sucrose cushion and centrifuged for 16 h at 100,000× g at 4°C.
The virus fraction obtained was re-suspended in PBS.
CRDS was obtained through one of the authors (TY) from the Ajinomoto Co Inc., Tokyo, Japan [average molecular mass, 41 kDa; sulfur content, 11.5% (wt/wt)] (Fig. 1A) and was dissolved in sterile phosphate-buffered saline (PBS) [22]–[24], [34]. To synthesize CRDS analogues, commercial Curdlan (M, = 8.9×104, Wako Pure Chemical Industries, Tokyo), pyridine-SO3 complex (SO3-pyridine) (Tokyo Chemical Industry, Tokyo), and dry Me2SO (Aldrich Chemical, Milwaukee, WI) were used without further purification. Piperidine-N-sulfonic acid was prepared from piperidine and chlorosulfonic acid according to the method of Nagasawa et al. [34], [35]
Tenfold serial dilutions of virus were added to LLC-MK2 cells in 6-well plates, followed by 1 h incubation at 37°C with gentle shaking every 15 min. The medium was aspirated and replaced with 0.8% methylcellulose (CALBIOCHEM, San Diego, CA) in maintenance medium (EMEM, 10% FBS, penicillin and streptomycin). At 5 days post-inoculation, cells were fixed with 4% paraformaldehyde in PBS at room temperature for 20 min. Next, they were washed with water before the addition of 1 mL 1% crystal violet at room temperature for 20 min. The plates were then washed and dried, and the plaque forming units per milliliter (pfu/mL) were calculated.
Anti-dengue virus activity and cytotoxicity in LLC-MK2 cells were monitored by infecting the cells with DENV. Infected and non-infected cells were exposed to a range of concentrations of test compounds, after which, the LLC-MK2 cells were allowed to proliferate for 5 days, and the number of viable cells was quantified by the 3-[4,5-dimethylthiazol-2-yl]-2,5-diphenyltetrazolium bromide (MTT) (SIGMA, St. Louis, MO) method in 96-well plates [36], [37].
A549 cells containing a luciferase-reporting replicon of DENV-2 were seeded at a cell density of 25×104 cells/well in 6-well plates [31]. The cells were treated with 100 µg/mL of heparin, 100 µg/mL of CRDS, or medium with DMSO. After incubation for 48 h at 37°C with 5% CO2, luciferase activity was measured using the Renilla luciferase assay system (Promega, Madison, WI). Results were normalized by protein quantity.
Cells (LLC-MK2, HL-60, and THP-1) were incubated with DENV-2 at 37°C for 1.5 h to allow viral adsorption in the presence or absence of CRDS. The cells were then thoroughly washed and cultured for 4 more days and viral antigen expression was analyzed by FACS analysis as described below.
The viral samples (5.0×106 pfu/mL) were directly mixed with CRDS (500 µg/mL) at 4°C for 1 h followed by filtration through a Vivaspin 500, 100 kDa molecular weight cut off (GE Healthcare, Buckinghamshire, UK) to separate the DENV. The DENV titers were monitored by a plaque assay as described previously.
The alkyl CRDS-coated membrane filter was prepared as described by Muschin et al. [38]. Briefly, before preparation, the weight of the nitrocellulose membrane filter was measured precisely by a Mettler microbalance instrument. The average weight of 10 filters was used. A 1% aqueous solution (2 mL) of alkyl CRDS with a degree of alkylation (DOA) of 0.03 was passed slowly through the nitrocellulose membrane filter in the holder by a syringe and dried overnight under vacuum below 40°C to give an alkyl CRDS-coated membrane. The amount of fixed CRDS on the membrane filter was determined by weighing to be around 2.0 mg. Concentrated virus was incubated with the membrane with or without alkyl CRDS for 5 min at 4°C, and then viral titers were monitored by a plaque assay as described previously.
Dengue virus with a titer of 1×107 pfu was diluted in 10 mM HEPES (SIGMA, St. Louis, MO) buffer to a final concentration of 7×105 pfu which was used for analyzing the virus-compound interaction. These solutions were incubated separately on 0.01% poly-l-lysine coated 300 carbon nickel grid for 1 min. After fixing the samples with 1% Glutaraldehyde (SIGMA, St. Louis, MO) for 1 min, the samples were washed 2 times with ddH2O followed by a staining with 2% aqueous phoshotungstic acid (SIGMA, St. Louis, MO), pH 7.4 for 1 min. After that, the samples were visualized with the electron microscope JEM1010 from JEOL and analyzed with Digital Micrograph version 1.18.78 from Gatan Inc. USA.
LLC-MK2 cells were seeded at 30×104 cells/well in 6-well plates. The cells were incubated with DENV-2 (500 pfu) at 4°C for 1.5 h on a rocking platform. The cells were then washed with cold PBS(−) 3 times and the plates were shifted to a 37°C incubator and cultured. During the time for viral adsorption and infection, plaque assay medium containing 100 µg/mL of either heparin (MP Biomedicals, Solon, OH) or CRDS was added to the cells at appropriate time points (−1.5, 0, 1, 2, 3, 4 and 5 h) for 5 days. Viral amounts were monitored by a plaque assay. Experiments were conducted in triplicates and mean percentage inhibition was calculated relative to control, which consists of the same set-up but without inhibitors [39].
This assay was used to detect the inhibition of cell fusion by test compounds at low pH. C6/36 cells were seeded with a cell density of 1.0×106 cells/well in 6-well plates one day prior to the assay. Dengue virus was inoculated at multiplicity of infection (MOI) 0.03 onto seeded C6/36 cells along with either 100 µg/mL of heparin, CRDS, or medium with DMSO. Each set-up was then incubated for 2 days at 28°C. Thereafter, the medium was acidified to induce fusion by addition of 50 µl of 0.5 M 2-(N-morpholino)ethanesulfonic acid (MES) (pH 5.0) (SIGMA, St. Louis, MO), followed by incubation at 28°C for 2 days. Fusion cells were then stained with Giemsa stain, modified solution (SIGMA, St. Louis, MO) according to manufacturer's protocol. The stained plates were analyzed under the microscope [40]–[41].
Cells were fixed with 4% paraformaldehyde in PBS for 5 min and treated with 0.5% saponin and 2% Fc-receptor blocking solution (Human Trustain FcX, BioLegend, San Diego, CA) in PBS(−) 4°C for 30 min. Immunostaining of the cells was performed with the first antibody (1/10 dilution HB114) [42] at 4°C for 30 min followed by staining with the second antibody (1/100 dilution goat anti-mouse IgG PE conjugated, SC-3738 Santa Cruz Biotechnology, Santa Cruz, CA) at 4°C for 30 min. Cells were analyzed with CyAn Flow Cytometer (Beckman Coulter, Brea, CA).
3H5 chimeric human/mouse IgG1 antibodies (h3H5) were constructed and used as previously described [43]. Briefly, 0.391 µg/mL of h3H5 (ADE) or media (No antibody control) was incubated with DENV-2 for 1 h at 37°C. CRDS was either incubated with the immune complexes (pre-incubation) or after immune complex formation (post-incubation), following which, the mixture was added to THP-1 at MOI 10. After 24 h post-infection, the cells were washed thrice in PBS, followed by RNA extraction using RNAeasy kit (QIAGEN, Germantown, MD), cDNA synthesis (Bio-Rad, Hercules, CA) and real-time qPCR (Roche, Indianapolis, IN) according to the manufacturer's protocol. DENV primers used were targeted at the 3′ untranslated region [44]:
DEN-F (5′-TTGAGTAAACYRTGCTGCCTGTAGCTC);
DEN-R(5′-GAGACAGCAGGATCTCTGGTCTYTC). GAPDH primers were obtained from Origene and infection in cells is expressed as units relative to GAPDH [44].
Raji/DC-SIGN+ (0.5×106 cells/well) (a gift from Dr Timothy Burgess) were infected with DENV (1000 pfu/mL) in the absence or presence of the compound for 4 h at 37°C. The cells were washed twice with medium to remove excessive virus and incubated at 37°C in fresh culture medium. Raji/DC-SIGN+ cells were analyzed for DENV infection on day 4 after infection [45].
The Protein Data Bank (PDB) accession code for the DENV-2 Envelope protein used for the computational modeling in this paper is 1OKE [13].
The blind docking protocol (MVD) indicated that CRDS might exhibit the potential to bind the DENV E protein. The binding site of the CRDS molecule on the E protein was predicted to lie near the fusion loop of the E protein monomer, at the interface of the Domain II and Domain III of the E protein dimer (Supplementary Fig. S1).
The screening for anti-DENV activity was performed by the conventional MTT assay using LLC-MK2 cells infected with DENV-2 (Fig. 1B). The compound inhibited DENV-2 replication at concentrations starting as low as 0.1 µg/mL in LLC-MK2 cells in a dose-dependent manner (Fig. 1B). CRDS showed an EC50 of 7 µg/mL and a CC50 of more than 10 mg/mL with LLC-MK2 cells. Thus, the selectivity index (ratio of CC50/EC50) of CRDS was >1428, indicating that this compound is both potent and selective. Control compound, heparin, also showed a very potent anti-DENV activity with EC50 of 0.5 µg/mL and a CC50 of more than 10 mg/mL in the same experimental system.
The DENV can be propagated in broad range of host cells in culture, which include cell lines of mammalian and insect origin. The anti-DENV-2 activity of CRDS was further studied in different cell lines. Non-adherent cells such as HL-60 and THP-1, and adherent cells such as BHK-21 and C6/36 cells were analyzed by MTT and FACS methods, respectively. The efficient inhibitory effect of 100 µg/mL of CRDS on DENV replication in HL-60 cells 4 days after infection is shown in Fig. 2. CRDS was found to inhibit DENV-2 replication in all cell lines studied at similar concentration range (data not shown).
Some compounds analogous to CRDS were synthesized to assess the structure-activity relationship against DENV-2 (Table 1). Two CRDS analogs, CS0125 and CS0202, amongst 4 samples, exhibited weak anti-DENV activity with an EC50 in the range of 0.2 mg/mL to 0.3 mg/mL, with no detectable cytotoxicity up to 5 mg/mL in LLC-MK2 cells.
We investigated the activity of CRDS against three other serotypes of DENV (DENV-1, 3 and 4). CRDS was found to be active against all serotypes of DENV, with EC50 values of 0.262, 0.01 and 0.069 mg/mL against DENV-1, 3 and 4, respectively (Table 2). Of all the DENV serotypes, DENV-1 appeared to be less susceptible to CRDS. These data clearly showed that the CRDS is effective against all four DENV serotypes, though the extent of anti-DENV effect of CRDS is dependent on the serotypes of the virus [46].
Based on our previous data showing that CRDS efficiently inhibits the entry process of HIV, we presumed that this compound might also act at the early stages of DENV replication. To confirm that CRDS exerts its antiviral effect at the early step of DENV life cycle and not at a later stage, the effect of the molecule on the replication of DENV subgenomic replicon, encoding only non-structural viral proteins, was studied in replicon-transfected cells. Neither CRDS nor heparin inhibited replication of the DENV subgenomic replicon in the luciferase activity assay, whereas the replication inhibitor NITD008 [47] used as a control showed significant inhibition of replication (Fig. 3A). To confirm that CRDS indeed inhibits an early step of the fusion process, LLC-MK2 cells were incubated with DENV-2 at 1,000 pfu with or without various concentrations of CRDS at 37°C for 1.5 h. The cells were then thoroughly washed and cultured without the compound further for 4 days. The cells were then subjected to FACS analysis. The results indicated that CRDS might suppress the binding of DENV to host cells at all the three CRDS concentrations used (Fig. 3B). The control compound, heparin, inhibited this step similarly.
To identify the exact step of infection at which CRDS exerts its inhibitory effect, we characterized the kinetics of compound activity using time of addition assays. We examined whether CRDS blocked the initial attachment step in the LLC-MK2 cells, or a downstream event in the viral entry process. When CRDS or heparin was added together with DENV to cells during the 4°C attachment step (−1.5 h), and then removed prior to shifting to 37°C, both sulfated polysaccharides (100 µg/mL) could inhibit DENV infection completely. Interestingly, CRDS blocked viral infection almost completely even when it was added at the time of temperature shift to 37°C (0 h) allowing viral entry and membrane fusion to proceed. However, heparin inhibited DENV only by about 50% under the same experimental conditions. CRDS was then added at various time intervals up to 5 h post infection to address the kinetics of CRDS activity. The results demonstrated that inhibition was only seen by 0 h but not after infection, confirming its point of action during DENV entry/membrane fusion. Hence, heparin was only effective in preventing entry when added during the viral attachment step while the CRDS was also effective even during the post-attachment stage (Fig. 4A, B). Taken together, the data suggest that CRDS can block an event in DENV entry that lies temporally downstream of attachment to cells, either before or during fusion.
Ten-fold dilutions of virus solutions were prepared in 24-well plates. The alkyl CRDS-coated membrane filter (one or three sheets) was then placed in the wells. After 5 min, each solution (500 µL) was removed from the plate and transferred to another 24-well plate, and viral titers were monitored as described in Materials and Methods. The DENV titers were reduced by more than 70% when viruses were treated with membrane filters coated with alkyl CRDS (2.0 mg) as compared to those with membrane filters without alkyl CRDS, showing that the viruses were adsorbed more efficiently to CRDS-coated membrane (Fig. 5A).
DENV were mixed with CRDS for 1 h at 4°C and then, viral particles were separated with a Vivaspin 500 to monitor viral titer by plaque assay. The results showed that DENV titer was reduced to about 60% of control level which was treated with PBS. Hence, direct binding of DENV to CRDS was suggested (Fig. 5B).
Electron microscopy (EM) was used to visualize the effect of the CRDS on DENV-2 viral particles. We noticed that the treated DENV virions were visible as aggregates while control dengue virions were found mainly as solitary particles. Also, while control virions exhibited the normal, nearly smooth, outer surface typical of mature flaviviruses, the surface of the CRDS-treated virus particles seemed to be rougher, implying a possible structural change of the viral envelope (Figure 5C).
Cell-to-cell transmission, in addition to cell-free virus infection, is considered to be the key mechanism of spread of DENV infection, although syncytia formation does not necessarily indicate virus spread. Flavivirus induces cell fusion very efficiently upon infection in Aedes albopictus C6/36 cells at low pH. When C6/36 cells were incubated with DENV at 4°C for 1.5 h during viral adsorption in the presence of 100 µg/mL of either CRDS or heparin followed by culture for 4 days at 28°C, the appearance of fused cells was completely blocked by both compounds (Fig. 6A). Consistent with the time of addition studies, CRDS inhibited syncytia formation even when it was added at the time of temperature shift to 28°C, whereas significant syncytia formation was observed in the cells treated with Heparin (Fig. 6B).
We also carried out the co-culture experiments of naïve C6/36 cells with 2 day-old DENV infected C6/36 cells in the presence or absence of CRDS for 2 days to observe its effect on syncytium formation. The compound completely blocked the appearance of DENV-induced fused cells (data not shown).
The observation that CRDS prevents virus binding to the host cell surface prompted us to evaluate whether this compound can also exert an antiviral effect towards DENV particles pre-opsonized with antibodies, in THP-1 cells. For this purpose, the infectious properties of DENV particles pre-opsonized with increasing concentrations of antibody was determined in the presence of 1 mg/mL of CRDS. The antibody enhanced DENV infectivity by 4 times in THP-1 cells as compared to control without the antibody. Under this condition, CRDS prevented more than 70% ADE when the simultaneous treatment of DENV with CRDS was done, before the addition of antibody. Also, nearly 60% inhibition was seen even when treatment was performed after mixing of virus and antibody concurrently (Fig. 7, ADE with CRDS (late)).
We also evaluated the effect of CRDS on DENV infection in DC-SIGN expressing Raji-DC-SIGN cells. Cells were infected with a large concentration of DENV (1,000 pfu) in the presence of CRDS. The cultures were analyzed for DENV infection by FACS analysis 4 days later. CRDS inhibited DENV infection in these cells rather weakly but significantly (by only about 10–15%) at a concentration of 100 µg/mL (Fig. 8).
To further specify the possible protein-ligand interactions and to model the conformational changes of the protein brought about by binding of the inhibitor, a flexible receptor docking study was performed. The receptor and ligand were both docked flexibly, in order to obtain the docking pose with the least entropy possible, as would be the case in vitro. The Chemscore based docking algorithm of Glide module (Schrodinger Suite 2012) imposes constraints based on rewarding of hydrogen bonding and hydrophobic interactions, and penalization of steric clashes, followed by minimization of energy due to non-bonded interactions based on OPLS-aa force field.
The induced fit docking procedure produced five minimized binding poses of the CRDS to the E Protein. Rank-ordering of poses was done based on Glide Score, and the best scoring pose is presented here as the most probable and optimal binding model of CRDS to the DENV E protein. This model predicted that CRDS binds to the E protein at the DII and DIII-DI interface of the two monomers. The proposed model also indicated that the CRDS fits at one end, in a pocket below the flexible ‘kl’ loop lining the hydrophobic cavity, described as the BOG binding pocket by previous research (Fig. 9). The tail end of the CRDS interacts with residues that neighbor the fusion loop (DII – residues 100–108).
The H-bonding interaction map of the CRDS at the binding site revealed that the ligand forms hydrogen bonds with the two conserved residues His 244 and Lys 310 (both 100% conserved across all DENV serotypes) and Asn 153 and Lys 247 (80% conserved), apart from other residues (Table 3, Fig. 10A, B) [48].
The LigPlot map [49] of the CRDS-E protein docking model indicated hydrophobic interactions (Fig. 11) with Trp 101, Asn 103, His 244 (all 100% conserved residues), and Gly 28 (80% conserved) [48].
The RMSD deviation of the docked pose of the protein obtained by induced fit, when compared to the conformation of the crystal structure, was found to be 0.14 Å, following pairwise alignment. The per residue RMSD revealed the highest deviation to be that of Lys 157(B), at 5.69 Å, followed by Arg 2 (B), which was 2.91 Å.
As do the other viral envelope proteins, the DENV E protein plays a crucial role in both binding of the virus particle to host cell receptors and fusion of the viral membrane with the target membrane. Hence, the E protein is one of the more attractive targets to inhibit DENV entry, to develop novel anti-viral drugs, as well as to discover effective vaccines. In this paper, we report that a sulfated 1→3- β -D glucan, CRDS, inhibits entry and fusion steps of the DENV life cycle very efficiently in several different cells of mammalian and mosquito origin. CRDS is very potent and selective since its minimal effective concentration is as low as 0.1 µg/mL in LLC-MK2 cells while toxicity is only seen at the concentrations more than 10 mg/mL. Although the anti-viral activity was comparatively weaker, it is important to note that CRDS could also inhibit DENV-1, 3, and 4.
DENV-2 and DENV-3 infections are very efficiently inhibited, whereas DENV-1 and DENV-4 infections require much higher concentration of the compound to achieve inhibition. The results of the experiments on anti-DENV effect of polysaccharides with different serotypes of the DENV obtained by Talarico et al. and Pujol et al. correlate with our results [46], [50]. This phenomenon is most probably not related to the difference in amino acid sequence of the E protein. Wang et al. derived a phylogenetic tree from the comparison of the DENV E gene nucleotide sequence, where they put forth that DENV-1 and DENV-3 are sister groups [51]. We believe that this difference could be due to the differences in the internalization/endocytosis pathway of the DENV particle of different serotypes. DENV utilizes different pathways for entry in different cells, and for that matter, even in the same cell type. In fact, distinct receptors have been proposed for different serotypes in the same host cells [52], [53].
Our results do indicate that CRDS inhibits the virus binding to the host cells, and also pH dependent cell fusion. The mechanism by which CRDS prevents virus binding is most probably by inhibition of virus internalization into the host cells via endocytosis. However, endocytosis of the virus particle may occur by any of the mechanisms, viz., phagocytosis, macropicnocytosis, clathrin-mediated endocytosis or caveolin-mediated endocytosis. We don't have evidence that CRDS can inhibit endocytosis by any or all of these routes. However, it could be that the DENV-CRDS complex undergoes internalization into the host cell. But the presence of CRDS prevents the viral fusion with the host vesicular membranes and thus inhibits the release of viral genome into the host cytoplasm. The vesicle is probably then degraded.
From the anti-viral therapeutic point of view, sulfated polysaccharides are compounds of particular interest because they have been shown to exhibit potent entry inhibitory activity against diverse viruses such as HIV, CMV, VSV, and HSV [54], [55], [56]. As expected, while CRDS did inhibit the early step of DENV infection, it failed to inhibit the replication of DENV subgenomic replicon. Time of addition experiments demonstrated that the compound inhibited viral infection at an early step of DENV infection. Although heparin showed similar activity, there are apparent differences in the activities of CRDS and heparin; CRDS inhibits both viral binding to cells and an early post-attachment step of entry (membrane fusion), while heparin acts mainly at the virus binding step. Our finding that CRDS inhibits cell-to-cell infection/spread of DENV more efficiently than heparin, is also consistent with our proposal that CRDS also inhibits membrane fusion between the host cell and the viral envelope.
DENV attachment to host cells has been proposed to be mediated by the binding of receptor glycosaminoglycans (GAGs) to the DIII of the E protein [57]. Two putative receptor GAG binding motifs have been mapped to the DIII [57], [58]. Abd-Jamil et al. proposed that while the GAG binding motif of loop I could mediate the E protein interaction with the host cell GAG, the residues on loop II could mediate a later, more receptor-specific interaction to facilitate virus attachment [58].
We investigated the possibility of direct binding and interaction of CRDS to DENV particles by 3 different approaches; virus titration after either mixing of DENV with CRDS followed by Vivaspin 500 or adsorption of DENV-2 to alkyl CRDS-coated membrane filter, and EM studies. Both procedures showed that the DENV titer was reduced to less than 30% of control level which was treated with PBS, suggesting direct binding of CRDS to DENV (Fig. 5A, B). Recently, two of the co-authors of this paper (TM and TY) showed that the alkyl CRDS-coated membrane filter was found to have a specific adsorptive functionality for influenza A, but not B, virus in vitro [38]. However, the membrane filter without the compound did not effectively remove Influenza viruses, and thus a membrane filter without alkyl CRDS was not effective against Influenza viruses. These results, taken together with ours, strongly suggest that the alkyl CRDS-coated membrane filter removed DENV and influenza A viruses by interactions between the negatively charged sulfate groups and the positively charged envelope proteins of both viruses. Therefore, it is likely that CRDS might recognize DENV through specific interaction with the surface glycoproteins of the DENV. The detailed adsorptive mechanism requires further investigation.
The above interpretation was further supported by EM, which was used to visualize the effect of the CRDS on DENV-2 viral particles. CRDS apparently induced aggregation of DENV virions by possible alteration of viral surface. As compared to control virions that exhibited the normal, nearly smooth, outer surface which are typical for mature Flaviviruses, the surfaces of the CRDS-treated virus particles seemed to be rougher, implying a possible manipulation of the viral envelope (Figure 5C). This observation is reminiscent of the recent report by Costin et al. [59] wherein it was demonstrated, using biolayer interferometry and cryo-electron microscopy respectively, that the anti DENV-2 peptides, which were newly developed, interfere with viral binding to cells through direct interaction with the E proteins, eventually leading to changes of the viral surface. Thus, it is possible that, like these peptides, CRDS can trap the viral E proteins in some conformational arrest which is not suitable for viral binding to cells and entry.
According to the binding model proposed by us here, CRDS forms strong H bonds with Lys 310, a residue that is 100% conserved among the flaviviruses [48], which is also a part of the GAG binding motif of loop I. The proximity of the loop I residues to the fusion peptide, near which the CRDS molecule is proposed to bind, in the dimer conformation of the E protein, could also mean that the ability of the loop I residues to form substantial interactions with the host cell surface GAG moieties will be hindered. This looks particularly plausible in view of the fact that the actual CRDS molecule is lengthier and spans more of the dimer surface than can be analyzed by Induced Fit, due to the limitation on the number of atoms of the ligand molecule specified by the software. The Structure-activity-relationship (SAR) of CRDS was studied based on its anti-DENV2 activity by synthesizing 4 analogues with various degrees of molecular weights as well as different chemical compositions (Table 1). The anti-HIV activity of CRDS is dependent on not only the degree of sulfation but also on the molecular weight of the molecules [24]. In our experiment, CRDS showed much higher anti-DENV activity as compared to the two CRDS analogues, CS0125 and CS0202, both with higher extent of sulfation, but lower molecular weight than CRDS. Hence, it is possible that molecular weight is important for the anti-DENV activity of CRDS. However, since we did not test sulfated polysaccharides with molecular mass greater than that of CRDS, more extensive studies are essential to draw a conclusion to that effect. The possible mechanism of inhibition of the virus – host membrane fusion process by CRDS could be explained by the steric hindrances and non-covalent ligand-protein interactions, as predicted by the flexible receptor-ligand docking. CRDS is predicted by our Induced fit docking protocol to bind in the proximity of the previously described BOG binding pocket beneath the kl hairpin. The kl hairpin, comprising residues 268–280, has been described to play a key role in the conformational changes driving the transition of the E protein from the dimer form to the fusion competent trimer form. Modis et al. proposed that a shift in the position of the kl loop towards the interface between DI and DII of the dimer partners causes the DII to swing away from its dimer contact and thus project the fusion peptide at its distal tip towards the host membrane for fusion [21].
In our proposed model for CRDS binding, it is seen that the ligand forms hydrogen bonds with Gly 275, and hydrophobic interactions with Asn 276, both residues of the kl hairpin. These interactions coupled by the steric hindrance offered by the physical presence of the CRDS molecule at the DI-DII interface of the dimeric partners of the E protein could very well prevent the shifting motion of the kl hairpin said to be the initiating action for the dimer-trimer transitions.
Modis et al. [21] also proposed a post-fusion conformation model of the DENV E protein, describing the conformational changes required to effect the trimerization. Here, they proposed that the kl hairpin shift causes the DII to undergo a 30° rotation with respect to DI of the E monomer, followed by a 70° rotation of DIII, corresponding to a 36 Å shift towards DII. Our model of CRDS binding indicates that the occupation of the space at the dimer interface by CRDS molecule will offer steric hindrance to the movement of DIII and prevent its rotation, thus preventing the trimerization process.
One of the more significant observations from our experiments is the ability of CRDS to inhibit ADE-mediated DENV infection in THP-1 cells expressing Fc receptor. As is well appreciated, ADE is thought to be the major cause of DHF/DSS, and fear of inducing ADE has hampered the development of a DENV vaccine [60]. Based on our observations, it is important to address whether CRDS indeed can block ADE in mice before moving on to human subjects to explore the possibility for it to serve as a candidate for clinical trials for the treatment of DENV infections.
At early stages of infection, the virus is said to replicate in DC-SIGN-positive cells [61]. DC-SIGN is considered to be a critical receptor for DENV, because it renders non-permissive cells susceptible for DENV infection and DC-SIGN is highly expressed in immature DC [62]–[64]. Though immature DCs are still not functional, they are equipped with receptors that mediate attachment, such as DC-SIGN, to capture diverse pathogens. DC-SIGN preferentially recognizes high-mannose sugars. Our study details the inhibitory effect of CRDS on DENV infection of DC-SIGN expressing cells. This points to the possibility that therapeutic application of CRDS can be effected in the early stages of DENV infection.
In the possible use of sulfated polysaccharides clinically, one of the concerns is about their anti-coagulant activity. The application of heparin, especially in its systemic use, has been associated with apparent anti-coagulant activity. Though CRDS is also a polyanionic substance like heparin, it is largely devoid of anticoagulant potential [23]. CRDS has been shown to be effective against HIV-1, Plasmodium falciparum in vitro [65], and Babesia infections [66]. The anticoagulant activity of CRDS in humans is also well documented in clinical trials on HIV patients [67], [68] and in studies on asymptomatic malaria patients [26]. The dosage of CRDS given to patients could be higher than that of heparin. Though results obtained in trials on HIV-infected patients were disappointing in terms of the efficacy, this could be due to the chronic nature of the disease as well as the HIV, which requires that CRDS be administered continuously to suppress viral activity. However, in case of cerebral malaria, upon CRDS treatment, disease symptoms such as fever, coma and organ involvement were delayed relative to parasite clearance and resulted in the favourable immune response [26]. Any effect on coagulation with prolonged usage of CRDS can be monitored easily, with subsequent adjustment of dosage. However, since DENV infection is notoriously associated with hemorrhage, DENV-infected patients should be monitored with additional care to note the effect of CRDS on coagulation time.
In conclusion, CRDS acts by interfering with the viral binding and membrane fusion steps, early and critical events of virus replication in DENV infection. We propose that the CRDS molecule binds to the DENV E protein at the DI-DIII and DII interface of the dimer partners. Since CRDS is well tolerated in clinical trials for HIV and cerebral malaria, further in vivo and clinical studies are warranted.
|
10.1371/journal.pntd.0003570 | Why Latrines Are Not Used: Communities’ Perceptions and Practices Regarding Latrines in a Taenia solium Endemic Rural Area in Eastern Zambia | Taenia solium cysticercosis is a neglected parasitic zoonosis occurring in many developing countries. Socio-cultural determinants related to its control remain unclear. Studies in Africa have shown that the underuse of sanitary facilities and the widespread occurrence of free-roaming pigs are the major risk factors for porcine cysticercosis. The study objective was to assess the communities’ perceptions, practices and knowledge regarding latrines in a T. solium endemic rural area in Eastern Zambia inhabited by the Nsenga ethno-linguistic group, and to identify possible barriers to their construction and use. A total of 21 focus group discussions on latrine use were organized separately with men, women and children, in seven villages of the Petauke district. The themes covered were related to perceived latrine availability (absence-presence, building obstacles) and perceived latrine use (defecation practices, latrine management, socio-cultural constraints).The findings reveal that latrines were not constructed in every household because of the convenient use of existing latrines in the neighborhood. Latrines were perceived to contribute to good hygiene mainly because they prevent pigs from eating human feces. Men expressed reluctance to abandon the open-air defecation practice mainly because of toilet-associated taboos with in-laws and grown-up children of the opposite gender. When reviewing conceptual frameworks of people’s approach to sanitation, we found that seeking privacy and taboos hindering latrine use and construction were mainly explained in our study area by the fact that the Nsenga observe a traditionally matrilineal descent. These findings indicate that in this local context latrine promotion messages should not only focus on health benefits in general. Since only men were responsible for building latrines and mostly men preferred open defecation, sanitation programs should also be directed to men and address related sanitary taboos in order to be effective.
| Livestock owners from small scale farms are most vulnerable for Neglected Zoonotic Diseases (NZD) in developing countries and their risk behavior leads to more intense and complex transmission patterns. Studies in Africa have shown that the underuse of sanitary facilities and the widespread occurrence of free-roaming pigs are the major risk factors for porcine cysticercosis. However the socio-cultural determinants regarding its control remain unclear. We hypothesize that via a bottom-up culture-sensitive approach, innovative control strategies can be developed that are more adapted to the local reality and more sustainable than current interventions. By assessing the communities’ perceptions, practices and knowledge regarding latrines in a T. solium endemic rural area in Eastern Zambia, we found that more than health, seeking privacy underlies motivation to use latrines or not. The identified taboos related to sanitation practices are in fact explained by the matri- or patrilineal descent and because men are responsible for building latrines, sanitation programs should focus more often on men’s knowledge and beliefs. In order to contribute to breaking the vicious cycle between poverty and poor health among livestock owners in developing countries, disease control strategies should always consider the socio-cultural context.
| Taenia solium taeniosis/cysticercosis is an important neglected parasitic zoonosis prevailing in many developing countries. The adult tapeworm lives in the intestines of humans, causing taeniosis, while the metacestode larval stage (cysticercus) usually develops in pigs following the ingestion of eggs excreted with the stool of tapeworm carriers, causing cysticercosis. Cysticercosis may also occur in humans upon accidental ingestion of eggs via faeco-oral contamination and may cause severe neurological disorders when cysticerci lodge in the central nervous system (neurocysticercosis, NCC) [1]. NCC is the most important parasitic neurological infection, to which almost 30% of acquired epilepsy cases are attributed in endemic areas [2].
Many surveys carried out in Africa have identified the general lack of and use of sanitary facilities as the major risk factors for cysticercosis [3–6]. Studies have demonstrated the positive effects of health education on the incidence of porcine cysticercosis in Tanzania [7] and on the prevention of epilepsy in Kenya [8]. However, an increased use of latrines could not be demonstrated. Many sanitation projects, implemented by governments or NGOs, which led to the construction of latrines in rural areas, faced refusal of the communities to use them and adopt safe hygienic practices [9–11] because the drives to motivate latrine adoption were often not identified and interpreted in messages and strategies to promote sanitation grounded in a given cultural context [12]. Unfortunately, in many African rural communities, open defecation practices were not adequately analyzed or taken into account before project formulation and implementation. Practicing open air defecation is linked not only to the presence or absence of water or latrines, but also to social and cultural determinants [13].
Improved latrine use as a control measure potentially has implications for many other sanitation-related pathogens [1,14,15], such as soil-transmitted helminths [16] and diarrhoeal agents [17]. According to the World Bank, 2.5 billion people worldwide live today without access to improved sanitation and 1 billion of these people practice open defecation. In sub-Saharan Africa, 70% of the population still lack access to improved sanitation, thereby indicating the urgent need for improvement [18].
Currently, T. solium control program managers need to understand why latrines are not used in endemic areas of Africa. Even though the significance of social and behavioral influences on the spread of human cysticercosis is known [7], culturally adapted control measures have not yet been implemented in endemic areas such as Zambia where the prevalence of T. solium cysticercosis in rural areas (in both human and pigs) is very high [19–22].
The objective of this research was therefore to assess the communities’ perceptions, practices and knowledge regarding latrines in a T. solium endemic rural area in Eastern Zambia, in order to identify possible barriers to their construction and use and to propose, eventually, adaptations of strategies to overcome cysticercosis, and other sanitation related diseases locally.
Focus group research was conducted in a rural area (Kakwiya) in Petauke district in the Eastern province of Zambia. The Kakwiya Rural Health Centre (RHC) has a catchment population of 11,344 (Clinic headcount records). People practice subsistence farming, growing mostly maize and groundnuts primarily for home consumption. Pig production is common; most households have owned pigs at least once to resolve financial issues.
The main ethno-linguistic group in this area is the Nsenga, which have a matrilineal descent. The district was selected based on reports indicating high porcine [20] and human cysticercosis prevalence, presence of a high number of free-roaming pigs, and reports of cysts observed in pigs slaughtered in backyards [21].
The Kakwiya catchment counts approximately 261 households and 138 individual toilets which is equivalent to an overall toilet coverage of 52.9%. The number of toilets varied quite markedly between villages (Table 1). There are no communal toilets as such. The sanitation facilities found in the study area were built following the simple pit latrine model. Completed, partially completed or abandoned, they generally consist of a pit dug into the ground, sometimes covered by a hygienic slab made from crushed stones and cement with a hole. Latrines were covered with a shelter (with or without a roof) and fitted simply with a sack or sometimes with a door.
Twenty-one focus group discussions (FGDs) were conducted totaling 172 participants including 56 men, 58 women and 58 children (below the age of 18) from seven villages (Table 2). The seven villages were randomly selected from villages around the health center because of its central position. They were not included in recent biomedical surveys to avoid information and sensitization biases. Separate FGDs were held with men, women and children in each village since these groups have different perceptions and behaviors regarding sanitation (gender dependent) [11]. In addition, working with heterogeneous groups is likely to hamper the quality of the data [23,24]. For children, the FGDs were gender-mixed because, unlike adults, they were able to speak freely regardless of age and gender.
To ensure the validity of the data collected, FGDs have been conducted until reaching data saturation of the information from the seven different villages and from the three different subgroups.
The data collection took place from July to August 2010. Each FGD consisted of approximately 8 participants. Participants were selected from the villages based on their availability and willingness to participate. The FG discussion guide was pre-tested and fine-tuned in one FGD performed with male participants from a village outside the study area.
Three facilitators (a female nurse, a male environmental health technician and a male community health volunteer), all familiar with the Nsenga language, were identified and trained to moderate, observe and record the FGDs. The training consisted of a two-day course during which they were briefed on the study objectives and on FGD moderation skills. Facilitators switched roles for each discussion. All the FGDs took place at the Kakwiya RHC because of its central geographical location and practical aspects. To avoid biases related to the fact that the venue was not neutral in terms of health, the first set of questions was about general pig management.
The average duration of the discussions was about an hour. The following topics were covered: the perception of pig breeding in the communities, knowledge and perceptions of taeniosis/cysticercosis infection and related risk behaviors such as people’s latrine perception and reasons for not using latrines (defecation practices, latrine management, building responsibility, socio-cultural obstacles); and opinions on control measures.
All discussions were recorded on a video camera to facilitate the transcription of a discussion involving several individuals at the same time. Encouraged by our key informants, the use of a video camera was pre-tested and did not seem to be intrusive or affecting the discussions. The facilitator was always assisted by a reporter. To ensure the good implementation and follow up of the study, the main researchers (Séverine Thys & Kabemba E. Mwape) attended every discussion.
In this paper, only results pertaining to people’s latrine perception, reported practices and factors that lead to lack of use of these sanitary facilities are presented and discussed.
The FGDs were transcribed and translated into English by two research assistants and two researchers who took turns in both tasks. To improve interpretation reliability, the written transcripts were reviewed independently by the two same researchers before accepting them for analysis. The analysis of the transcriptions and the notes taken during the FGDs was supported by the NVivo 8 software (QSR International Pty. Ltd., Melbourne, Australia, 2008), which allows to classify and sort data; examine relationships and trends in the data. The major themes were separately identified through coding by the same two main researchers of the study following an inductive approach. Any differences were discussed until consensus was reached.
Ethical clearance was obtained from the University of Zambia Biomedical Research Ethics Committee (003–02–10) and from the Ethical Committee of the Antwerp University Hospital in Belgium (10 03 3 704). Further approval was sought from the local authorities and community leaders before commencement of the study. Finally, before the start of each FGD, permission was sought from the individual subjects to enter the research and to video record the discussion. Written informed consent was obtained from each participant and from parents (or guardians) for children under 18 years old. Participation in the discussion was voluntary and no names nor pictures were recorded in the transcripts. Questions were appropriately phrased to avoid embarrassing people and also to tackle sensitive issues or taboos. FGDs with children took place after school hours.
The results highlight the different themes that emerged in the analysis. To reflect as much as possible what was expressed in the discussions, the order used to present the themes in each sub-sections reflects the level of importance given by the participants to these topics (going from a strong to a weaker consensus). No significant differences were observed between villages (very homogeneous), we indicate when the main ideas were mentioned across all the FGDs and where consensus or differences arose the most among the three different categories of FGDs conducted (men, women and children).
Results are illustrated with anonymous quotes, selected on the basis of their representativeness, appropriateness and revealing quality.
Topic 1: Perceived presence and absence of latrines. In this section, we describe how people perceived the presence and absence of latrines in their village in order to identify factors that explain latrine availability.
People generally referred as much to situations with as without the presence of latrines. On the overall, participants agreed on: 1) the general absence of latrines at home (no latrines at home, no latrines for visitors, not yet completed), especially women; 2) the presence of latrines in some homes (latrines at home, shared and not shared with neighbors) (acknowledged by all categories); 3) that latrines are public among neighbors, a perception mostly shared among men and women groups (Table 3). The distinction between having a latrine at home and the presence of latrines in the village revealed a distinction between private and communal uses of sanitation facilities.
Participants further stated that a household with a latrine had dignity and respect as visitors, passersby or guests unaccustomed to using the bush, could easily be allowed to use the facility. A latrine therefore was a necessary feature of hospitality. This was especially highlighted by people whose household was situated in close proximity to the roads:
Conversely, at village level, the presence of latrines (latrines shared with neighbors, few and many latrines in the village) was more mentioned than latrine absence (no latrines in the village, not in the field and not shared with neighbors), except among children FGDs who pointed out that if you need to defecate while you are working in the field, you do not have other options that doing it in the open.
Even if latrines were mentioned to be available in the community, men and women stated that there were few of them, and that sharing these facilities was a common practice.
Finally, even if very few participants mention that some of the latrines were incomplete (Table 3), They expressed a certain willingness to build latrines, although it was not often considered a priority.
Topic 2: Obstacles to build latrines. From all the 21 FGDs, eight obstacles (Table 3) were identified as contributing to the lack of latrines.
Because the Nsenga observe traditionally a matrilineal descent, a newly married couple lives in the wife’s relative’s household and the custom implies that when a man gets married, he ought to build his own latrine because of the taboos of a man sharing a latrine with his parents-in-law (see paragraph on taboos). In this cultural context, the responsibility of latrine construction clearly belongs to men but a constraint mentioned by the participants was that men did not consider the construction of a latrine for themselves as a priority. This lack of motivation was mostly explained (mainly by women) by the fact that some men were lazy, or preferred to spend time drinking alcohol.
The existence of other latrines in the village was the second consensual argument raised by the participants to explain the non-prioritization of latrine construction. Indeed no taboos were observed for sharing latrines with people from another household or with non-relatives. Participants commonly stated that it was also not well accepted that the toilet owner refused access to other community members. Refusal could create conflicts or have negative consequences on social relations. In addition, some male and female participants recognized that refusing access to neighbors would reduce all the benefits of having a latrine (prevent diseases, prevent pigs from eating human feces, prevent contamination of kitchen utensils) by forcing people to defecate in the bush or near their homes. The hesitations expressed about the placement of locks on latrine doors and its implications for the sanitation of the village reflected the tension between private use (leading to eventual envy, jealousy, no more benefits for the community health) and communal use (leading to increased latrine cleaning and maintenance, rapid filling of the pit, sharing the cost and responsibility). In both kinds of use, the risk of disrupting interpersonal relations was a potential obstacle to start constructing latrines.
As expressed by one man participant:
A third obstacle mentioned was the lack of means to construct a latrine. This was an important constraint and commonly identified in the three groups. It was stated that some people could not afford to build a latrine of good quality materials, according to the local standard criteria of a latrine (roof, proper door, walls high enough,…) pursuing the gain of visual privacy. The physical appearance of the latrine had a bearing on whether it was used or not regarding the level of privacy offered (see latrine perceived advantages). However, local materials were often not of sufficient quality.
A fourth and fifth constraint highlighted in our study were the lack of knowledge on how to build latrines and the lack of awareness on their advantages for some participants. They pointed out that educating people about the benefits of a latrine would eradicate all the misunderstandings or erroneous conceptions. Men insisted more on the need of more “persistent” and “sustained” sanitation education campaigns, women made more reference to the hygienic benefits that campaigns would result in.
The 6th reason described by some women is that unmarried women were facing great difficulties to have latrines built since the construction of latrine is a man’s responsibility.
Finally, the last obstacle raised to latrine construction was that it was simply not yet considered as a habit to defecate in a toilet.
Topic 3: Latrine use versus open defecation. Even when latrines are present, going to the bush in order to defecate in the open is a common practice, and a culturally accepted norm in the area. It appeared that men were the ones enjoying more to defecate in the open (see obstacles to use latrines) than others.
As a general finding about pros and cons of latrine use, more comments against than in favor were mentioned during the FGDs.
Topic 4: Arguments in favor of the use of latrines. Participants manifested a strong consensus that latrine use contributed to a better hygiene and prevents diseases. Additionally, greater comfort, dignity and increased privacy were mentioned. When exploring the benefits of sanitation within communities and households, the last common argument shared, especially among women was that “there is simply nothing good about open defecation” (Table 3).
Latrine use contributes to good hygiene. All groups and especially women considered that the presence of a latrine ensured hygiene in a household mainly because it prevented pigs from eating human feces, and avoided them contaminating kitchen utensils left on the ground with dirt and feces that could bring diseases (see next section).
Men and women also pointed out that, as long as all households had no latrine, no benefits would be realized, as many would still be openly defecating.
Another common perceived advantage was the prevention of food contamination by flies, as by using latrines, all human feces would be gathered in one pit instead of being everywhere in the open.
According to some comments from children, latrines were the place where you could “discard all the bad things from the intestines”.
Latrine use prevents diseases. It appeared that participants connected the use of latrines with their own improved health but not always straight line.
They alluded to the fact that latrines prevented diseases in general and some specific diseases as cholera or dysentery by preventing pigs, flies and unwashed hands to contaminate food with human feces.
There were linkages with the risk of diarrhea and HIV transmission only when participants referred to the pigs’ habit of eating feces of sick persons.
Latrine use is more comfortable, provides more privacy and increases dignity. In the FGDs, there was a strong consensus, especially among women, that “one advantage of using latrines was not being disturbed by pigs pushing you before finishing”.
Adults stated that latrines were more often used when someone was suffering from diarrhea. Afraid of not reaching the bush on time and be embarrassed in front of fellows, they would rather use latrines (indicating a matter of comfort and convenience rather than family or personal health protection).
Mainly for men, using latrine offered a greater comfort when they were situated closer than the bush and when it rained.
It was also very important for many to avoid being seen defecating in the open, especially men, by the opposite sex or by their in laws:
Another major factor in favor of latrines, mentioned by participants in the context of privacy, was the seasonal availability of good defecation sites around the village. In the dry season, the bush was usually burnt for agricultural purposes, making them not dense and high enough anymore to hide villagers who wanted to defecate in the open.
Some participants, mostly males, revealed that it was quite more convenient to use latrines at night. As latrines did not always have a proper door, using it at night will avoid others to see you.
At night, latrines presented the additional advantage of reducing the risk of being exposed to hazards in the bush.
Along with the seeking of more privacy, participants further stated that the use of latrine gave more dignity in the sense that you could hide from the others when defecating:
Topic 5: Obstacles to use latrines. Thirteen reasons were identified for not using latrines (Table 3) in order to facilitate the flow of the reading, some themes are grouped together. The greatest consensual reasons among all FGDs that arise were: 1) the taboos related to sanitation practices, 2) the fact that not all households had a latrine and 3) the fact that latrines did not offer enough privacy resulting in a loss of dignity for the user.
For women and children, the main factor that leaded to not using latrines was the unavailability of the facilities, while for men traditional taboos seemed to be the central issue.
Latrine use entails cultural taboos. In general when the different socio-cultural obstacles for the use of latrines were addressed, most of the comments were made by men. This showed that men were much more concerned about the respect of taboos than the other two groups.
As such, other people, especially children, are not allowed to see their parents or adults go to the latrine.
In the study area, traditional taboos meant that the head of household (father) could not share the same latrine with his mother-in-law, his children-in-law, older children (adults) of his own household, his grown-up daughters and his younger children when the risk to be seen was too high or when young children will use the latrine just after their father.
Often, men went to the bush pretending to go to the field, gather firewood or hunt mice (a common delicacy in the region) not to be seen entering a latrine by children.
It seemed that these taboos were strongest between in-laws and in particular between mothers and sons-in-law.
Bypassing these prohibitions was considered as a lack of respect and decency similar to being seen undressed. For this reason, some of the participants suggested having two latrines at the same household. When asked if they had no problems being seen going to the latrine in full view of their daughter in-law, a male participant stated:
On the other hand no taboos seemed to be observed between parents and very young children, between wife and husband, between women and neighbor’s children, in town and with neighbors as they often did share latrines in the community. Sons were freer to share the same latrine with the head of the family than daughters. Fewer taboos were observed between people of the same gender.
Although the origin of those taboos and reasons to observe them were not very explicitly explained in the discussions, in one way or another all the further arguments against the use of latrines developed in this section were linked to the importance of respecting those sanitation taboos.
Compounded by the existing taboos, women considered that if not every home owns at least one latrine, the practice of open defecation will not end and no benefits will be realized. Despite the men also stating this and acknowledging the benefits that would arise from the use of latrines, they were still the major obstacle towards the construction of more latrines (see above).
Latrine use causes a lack of privacy and is less convenient and comfortable. At the first sight, this sub-section can look contradictory with the advantages foreseen earlier by the participants regarding the use of latrine. However when participants were asked why people did not use latrines, some responded that most of the available latrines were not in a very good state. The walls were too low, they lacked a roof and a lockable door (many only had a cloth or a sack as a door) thus compromising privacy. This lack of privacy mainly mentioned by men, included the fear of leaving dirt after the latrine use as well as the risk to see nakedness. They also mentioned that latrines were often built in the center of the village, which prevented people from using them because they would be seen entering or leaving them.
The convenience perception about the use of latrines was not unanimous and presented also different opinions. It appeared that for some women and men the use of latrines was not necessarily more convenient than open defecation to fulfill its benefit of improving hygiene. First, because it was not easy to wash hands after the use of the latrine (no water supply nearby) and secondly, because it was not convenient to carry material to clean the latrine.
For some men especially, the few latrines available created quite rapidly a queue, which was not convenient in case of an urgent need (e.g. diarrhea). In addition, the queue led people know that you need to defecate (risk to be seen). Complete cleanness of such public latrines, needed to fulfill the required social norms of privacy, convenience and comfort, was also impossible to achieve due to maintenance difficulties. The most important reason was to be exposed to dirt, bad smell and flies. It was also mentioned to be a scary place for children (dark, big hole wherein they could fall,…).
According to some children, it was simply easier to go to the bush to relieve oneself also because it was more difficult to find a latrine when people work in the fields or when they were travelling.
One woman mentioned that the way the latrines were built (pit latrine) did not allow checking for worms and could delay the identification of a parasitic infection. Another inconvenience to use latrine according to some men and children, was that it did not allow pigs to feed on their feces. Open defecation was a common and affordable solution for the pig’s owner to face feed shortage.
Finally, more children than women admitted that using latrines was simply not a habit and that men from the older generation manifested a strong reluctance to build latrines.
Limited knowledge on latrines. If preventing diseases was an argument in favor of the use of latrines, it was not always evident for the participants that adequate maintenance was one of the most important determinants to ensure the health benefit of a latrine. According to the men it was difficult to teach children how to use a latrine properly. Also, equipment sometimes freely distributed by sanitation programs was not always properly used.
Mainly men considered that latrine promotion failed, as it did not convince them to construct and use latrines properly.
Considering the results presented in this manuscript, it is clear that the transmission of T. solium can easily continue in this very suitable fecal contaminated environment where few infrastructure for safe excreta disposal are available and correctly built or used and therefore allowing free roaming pigs to maintain the lifecycle in such endemic area.
While poverty may be a contributing reason for the lack of latrines in many communities, it does not explain why some people continue to practice open defecation long after their community has been provided with water points and learned about latrines and hygiene practices [25].
Choosing latrines means changing defecation practices and because sanitation behaviors tend to be strongly culturally conditioned, we chose to discuss our results mainly through user lenses [12] and socio-cultural lenses [26]
Like Jenkins and Curtis (2005) in Benin, we found among the arguments in favor of using latrines, drives related to prestige (gaining more respect and dignity from visitors), to well-being (better hygiene by protecting pigs from eating feces, more comfort by not being disturbed by pigs pushing you before finishing, more visual privacy) and situational drives (when it’s raining, when bush has been burnt, when you have diarrhea).
But from a consumer perspective, these authors also demonstrated through their model of motivation for latrine adoption in rural Benin that at least one drive is needed (among the 11 that they found) to motivate real changes in sanitation behavior as long as barriers do not suppress the expression of this specific drive [12].
Out of our FGDs, several obstacles for building and using latrines were identified such as sanitation taboos (avoiding sons and mothers-in-law to share the same latrines), the lack of privacy (latrines not well built, the queue) and the lack of comfort (too scary for children, flies, smell). It means that in our studied context the different obstacles identified would need to be addressed first in order to arise sufficient intensity for positive drives to be translated in concrete action taken by the target population.
Complementary of Jenkins and Curtis’ model, Avvannavar and Mani’s conceptual model of people’s approach to sanitation [11] can let us better understand the interplay of socio-cultural factors that determine how people take care of their primal urge.
In the category “culture” and “fear & superstition”, their model offers us the opportunity to discuss the taboos we identified related to sanitation practices. For the authors, attitudes and beliefs about revulsion to feces vary between cultures. Examples are numerous in Africa. In the Akan culture (Ghana) for example, the word “shit”, is as taboo as the thing itself and people when going to the bush to defecate, need to wear a blinder pretending that they will not be seen if they see nobody [27]. In Uganda, sharing latrines with in-laws is a taboo and the use of latrines could affect women’s fertility and also cause miscarriage [28]. In the Eastern Cape Province of the Republic of South-Africa, human feces were found in the bush because people were afraid to share latrines to avoid being bewitched [9].
However our results demonstrate that the type of descent, matrilineal or patrilineal, is also an important factor that has a significant influence on the sanitation practices of a community.
In our study, although no taboos seem to be observed between wife and husband, between women and neighbor’s children, in town or with neighbors, a man could not share the same latrine with his mother-in-law, his children-in-law, older children (adults) of his own household, his grown-up daughters, his younger children.
Similar taboos have been reported in a number of other African communities, reflecting important social norms. In Eastern Cape Province (RSA), for instance, stakeholders stated that people prefer to defecate in the bush because sharing a latrine as a father-in-law with his daughter-in-law is perceived as a disgrace [9]. In the case of a sanitation program (Community-Led Total Sanitation program, CLTS) introduced in a Kenyan district, the taboo for a father-in-law’s feces to mix with those of his daughter(s)-in-law was also described resulting to gender-segregated open defecation sites in the forests [29].
The origin of the taboos we identified and reasons to respect them are not very explicitly explained by the participants. It seems however logical in a matrilineal society to observe very strict proscribed behaviors towards the maternal in-law’s in order to limit contact and ensure respect. Mary Douglas in “Purity and danger: An analysis of the concepts of pollution and taboo” explained that the father in the matrilineal Trobrianders and Ashanti is credited with being an involuntary source of danger; he is an intruder [30]. If we look at the previous examples from patrilineal societies, the reported latrine practices taboos (Father and daughter-in-law) are in fact simply reversed in our setting (Mother and son-in-law) because we are in a matrilineal society.
Furthermore, being in a matrilineal or patrilineal system allows us to better understand firstly, the gender division of tasks about latrine construction and secondly, the social norms in such societies in terms of privacy, both having a strong influence on why latrines are not built or not used.
Seeking privacy, in general and from in-laws in particular, seems to be the main underlying motivation for people to use or not latrines. Using latrines when natural vegetation does not suffice to hide in the bush contextualizes and highlights the importance given to privacy and the fear to be seen going or openly defecating, especially by relatives. Embarrassment and shame that occur if being seen is the expression of a transgressed norm that underlies a number of taboos related to the matrilineal descent of this community. The perceived benefits of a latrine depends on the way it has been built and its construction is in turn strongly linked to the respect of socio-cultural sanitation practices. In our setting the benefit would be not to be seen (privacy) in order to feel free, less shy and respectful towards cultural taboos. However, the actual sanitation situation entails no or badly constructed latrines that do not offer enough privacy (e.g. no proper door, too small walls, no roof, rickety superstructures) and therefore also contributes to why latrines are not used. Avvannavar & Mani (2008) consider that the human tendency to seek privacy is the modified animal behavior attributed to the deep primal territorial tendencies. In our research, we can apply this theory as such: “By not willing to be seen or to see you when you are defecating” could be another way for men in a matrilineal system to mean that “by not defecating in the same latrine of my mother-in-law, I show respect about her territory as I am a stranger in her family”.
In general, the topic of sanitation was mostly developed by female participants. Women were more spontaneous and free to speak about latrine issues and related sanitary behaviors. The particular social, economic and political structures in most African contexts make women more concerned about sanitation and domestic duties than men [31] who carry out construction and maintenance of facilities according to the gender division of tasks. That explains why the common practice in the water and sanitation sector is to involve women, not only as a target group, but in the organization of activities at the local level [32]. In our study however, resistance to abandoning open defecation practice was mainly expressed by men. As the role of latrine construction belongs to male participants, addressing men’s knowledge and beliefs could benefit sanitation programs in many ways.
If men do not see latrine construction/use as a priority or if they do not know how to build it, latrine coverage and use will not raise. In this specific situation, not getting married can be a handicap for women to have their own pit latrine built.
In regard to sanitation taboos, in certain cases, taboos themselves can be used as arguments in favor of the use of latrines in order to facilitate even more their respect. In the CLTS study in Kenya for instance, where the targeted communities are observing quite similar cultural norms and defecation practices as in our study, the facilitators were able to break one of the defecation taboos by showing that a pit latrine located within the homestead will complicate the task for an intruder who sought to bewitch others by accessing their intended victim’s feces (less discrete, difficult to dig up the feces) [29].
The existing challenges of cysticercosis control in endemic regions require a “people-centered” preventive approach that addresses both the perception of the disease and its management. Control strategies should also be directed to the patterns of people’s behavior associated with the phases of transmission of the disease [33]. In this specific study we focused on people’s perceptions, knowledge and reported behaviors regarding the use and the construction of latrines.
Out of our findings, several entry points for promoting the use of latrines were identified and discussed.
Seeking privacy and taboos were both identified as the key factors influencing the possession and use of sanitation facilities. These findings reinforce why latrine promotion messages should not only focus on health benefits.
Some taboos can be explained by the type of descent (matri- or patrilineal). By acknowledging that the descent is also a factor that influences sanitation behaviors and regulates a number of norms and practices, we can more easily anticipate the type of taboos that could entail the adoption of hygienic practice related to sanitation.
A concrete proposition that could be made is to start building per homestead gender specific latrines instead of household specific latrines, each of them located in two different places to respect privacy.
But unless program planners are not totally convinced of the necessity to direct interventions not only at women but at men as well and focus also on men issues (practices, beliefs and knowledge), latrine building and use will not be efficiently promoted. Our results also stress the importance of anthropological studies for an in-depth understanding of sanitation practices within particular contexts in order to enhance the design of adapted interventions.
|
10.1371/journal.pbio.1001719 | ANXUR Receptor-Like Kinases Coordinate Cell Wall Integrity with Growth at the Pollen Tube Tip Via NADPH Oxidases | It has become increasingly apparent that the extracellular matrix (ECM), which in plants corresponds to the cell wall, can influence intracellular activities in ways that go far beyond their supposedly passive mechanical support. In plants, growing cells use mechanisms sensing cell wall integrity to coordinate cell wall performance with the internal growth machinery to avoid growth cessation or loss of integrity. How this coordination precisely works is unknown. Previously, we reported that in the tip-growing pollen tube the ANXUR receptor-like kinases (RLKs) of the CrRLK1L subfamily are essential to sustain growth without loss of cell wall integrity in Arabidopsis. Here, we show that over-expression of the ANXUR RLKs inhibits growth by over-activating exocytosis and the over-accumulation of secreted cell wall material. Moreover, the characterization of mutations in two partially redundant pollen-expressed NADPH oxidases coupled with genetic interaction studies demonstrate that the ANXUR RLKs function upstream of these NADPH oxidases. Using the H2O2-sensitive HyPer and the Ca2+-sensitive YC3.60 sensors in NADPH oxidase-deficient mutants, we reveal that NADPH oxidases generate tip-localized, pulsating H2O2 production that functions, possibly through Ca2+ channel activation, to maintain a steady tip-focused Ca2+ gradient during growth. Our findings support a model where ECM-sensing receptors regulate reactive oxygen species production, Ca2+ homeostasis, and exocytosis to coordinate ECM-performance with the internal growth machinery.
| Tip-growing cells, such as plant root hairs and pollen tubes or fungal hyphae, are characterized by a tip-focused Ca2+ gradient. These tip-growing cells tightly coordinate the loosening and pressure-driven deformation of their extracellular matrix (ECM)—the cell wall in plant cells—by locally adding new membrane and cell wall materials. In pollen tubes, which grow at amazing speeds to effect fertilization in plants, a class of kinases called the ANXUR receptor-like kinases (RLKs) sense perturbations in cell wall integrity, and their loss leads to pollen tube rupture. Here, we gain new insights into the mechanism of cell wall surveillance by these RLKs in the model plant Arabidopsis. We show that over-expressing ANXUR RLKs over-activates exocytosis, causing an over-accumulation of secreted cell wall material that eventually leads to growth arrest. Moreover, we find that the ANXUR RLKs function upstream of NADPH oxidases, which are membrane-anchored enzymes that produce reactive oxygen species (ROS). Using H2O2- and Ca2+-sensitive reporters, we show that NADPH oxidases generate tip-localized H2O2 production, which is required to maintain a steady, tip-focused Ca2+ gradient that is essential for pollen tube growth. We postulate that ECM-sensing receptors, such as the ANXUR RLKs, regulate ROS production, Ca2+ homeostasis, and exocytosis to coordinate the status of the ECM with the cell's internal growth machinery.
| It is well established that growing animal cells control the biogenesis, deposition, and remodeling of their extracellular matrix (ECM). In vivo the ECM contributes to the bulk, shape, and strength of many tissues and, therefore, plays a central role in development [1]. However, it is often underappreciated that the ECM also controls intracellular activities far beyond providing mechanical stability. For example, the ECM is under continuous cellular surveillance in order to monitor the loss of adhesion to the surrounding matrix, which leads to apoptosis. Consequently, disruption of signaling between the ECM and the cell is associated with tumorigenicity [2]. Similarly, growing plant cells direct the deposition of the primary cell wall (CW): the plants rigid, carbohydrate-rich ECM that resists turgor pressure, yet is flexible enough to allow cell expansion. Growing plant cells tightly coordinate the loosening and pressure-driven deformation of the CW with the addition of new membrane and CW materials through exocytosis. Thus, the cell must be kept informed about any environmental changes modifying the CW properties in order to avoid growth arrest or rupture. To circumvent these catastrophic scenarios, it has become increasingly evident that plant cells have developed mechanisms to sense CW integrity, which relay information about CW performance to the internal growth machinery. The molecular nature of this relay mechanism, however, remains largely unknown [3].
Since the first reports on THESEUS1 (THE1 [4]) and FERONIA (FER [5]), these Arabidopsis receptor-like kinases (RLKs) of the Catharanthus roseus RLK1-like subfamily (CrRLK1L) have received increasing attention as putative sensors that coordinate cellular growth and CW integrity (reviewed in [6]–[8]). How this coordination precisely works and which molecular players of the growth machinery are involved remained elusive, although Rho GTPases of plants (ROPs) and the production of NADPH oxidase-dependent reactive oxygen species (ROS) have emerged as putative downstream components. The role of NADPH oxidases, the ROS-producing enzymes that, based on their homology to the catalytic glycoprotein subunit of the mammalian phagocyte oxidase (gp91phox), are also called “respiratory burst oxidase homologues (Rboh)”, has been firmly established in various fundamental processes. These include localized lignin deposition [9], stomatal closure [10], pathogen responses [11], and root hair growth [12]. NADPH oxidases are plasma membrane (PM)-bound enzymes with six trans-membrane domains, an N-terminal region that contains EF-hands, and a C-terminal oxidase domain responsible for oxidizing O2 to produce superoxide radicals in the apoplast (reviewed in [13],[14]). The latter can quickly be dismutated, enzymatically or otherwise, into H2O2 that can freely diffuse back from the apoplast into the cytosol.
Connections between members of the CrRLK1L and NADPH oxidase families have been proposed or established for THE1 and FER, respectively. For example, THE1 has been reported to be a positive regulator of CW damage-induced ROS production in seedlings, possibly through RbohD [15], while FER is both a negative regulator of H2O2 production in unchallenged leaves [16] and of ROS in guard cells [17]. Furthermore, in root hairs that elongate by tip-growth, FER is a positive regulator of ROS production through the ROP2-RbohC pathway [18]. Similar to the rbohC loss-of-function mutant (also called root hair defective2 [rhd2]), disruption of FER leads to an impairment of ROS production and defective root hairs that burst [12],[18]. Disruption of the redundant CrRLK1Ls ANXUR1 (ANX1) and ANX2, the two closest homologues of FER, triggers the rupture of pollen tubes (PTs), the tip-growing male gametophytes of flowering plants, resulting in male sterility [19],[20]. Similar to fer root hairs, anx1 anx2 double mutant pollen form bulges and burst, failing to maintain their integrity during growth. This indicates that the FER and ANX RLKs could be cell-surface receptors that control CW integrity in tip-growing cells. In PTs, genetic evidence for the involvement of NADPH oxidases is lacking, but several studies have revealed a role for ROS during PT growth that remains to be precisely characterized. For example, it has been shown that either the use of ROS scavengers or the NADPH oxidase inhibitor diphenylene iodonium (DPI), or the down-regulation of a NADPH oxidase, reduces PT growth in tobacco [21]. In addition, the application of DPI at higher concentrations has also been reported to induce PT rupture in lily [22].
Because of the difficulty to image the dynamics of ROS production with good spatial and temporal resolution, and because of its multi-faceted impacts on CW properties and the activation of intracellular signaling, it is unknown how NADPH oxidase-dependent ROS control polar growth [23],[24]. It was first reported that RbohC/RHD2 is required for calcium influx via the stimulation of Ca2+ channels and for the generation of a tip-focused gradient of cytosolic free calcium [Ca2+]cyt, which is essential for polar growth [12],[25]. Later, Monshausen and colleagues reported that, under certain conditions, rbohC root hairs still display a tip-focused Ca2+ gradient, showing that RbohC was not essential for its establishment [26]. Moreover, they showed that artificially increasing or decreasing apoplastic ROS leads to growth cessation and root hair bursting, respectively, consistent with a role for ROS in regulating CW properties [27]. Finally, oscillations in apoplastic ROS levels just behind the tip were reported during root hair growth and correlated with growth rate, leading the authors to propose a model in which ROS rigidify the CW behind the tip, such that growth would be restricted to the tip [26]. However, due to the irreversible nature of the ROS-sensitive oxidation of the dye they used, the observed oscillations are unlikely to reflect the true nature of ROS dynamics [28]. Nonetheless, both models—namely the growth-promoting effect at the tip related to intracellular Ca2+ signaling and the growth-inhibiting effect behind the tip by rigidifying the CW—are not mutually exclusive as they could recruit different forms of ROS at different times and in different locations.
In this study we show that over-expression of the ANX RLKs inhibits PT growth by the over-activation of exocytosis and the over-accumulation of secreted membrane and CW materials. Genetic interaction studies coupled with a phenotypic characterization of loss-of-function mutants of two partially redundant, pollen-expressed NADPH oxidases, RbohH and RbohJ, demonstrate that the ANX RLKs function upstream of these NADPH oxidases. Furthermore, analyses of the genetically encoded H2O2-sensitive HyPer and Ca2+-sensitive YC3.60 sensors in NADPH oxidase-deficient pollen revealed that NADPH oxidases generate tip-localized, pulsating ROS that are responsible—possibly through activation of Ca2+ channels—for maintaining a steady, tip-focused Ca2+ gradient.
We have previously shown that ANX1-yellow fluorescent protein (YFP) and ANX2-YFP protein fusions are polarly localized in the PM at the tip of growing PTs in independent T1 transgenic Arabidopsis lines [19]. Although in T1 lines, which contain a mixture of untransformed and transformed pollen grains, no obvious fertilization-related phenotypes could be detected, in vitro pollen germination and growth assays of homozygous lines that carry a single insertion of the constructs in the T3 generation revealed that ANX1-YFP and, to a lesser extent, ANX2-YFP inhibit pollen germination and PT growth compared to the wild type (WT) (Figures 1A, 1B, and S1). To investigate whether these phenotypes are due to over-expression or non-functionality of the fusion proteins, we transformed anx1-2/anx1-2 anx2-2/ANX2 and anx1-1/anx1-1 anx2-1/ANX2 plants with ANX1-YFP and ANX2-YFP fusions, respectively. In all T1 anx1-2/anx1-2 anx2-2/ANX2 lines expressing ANX1-YFP and anx1-1/anx1-1 anx2-1/ANX2 expressing ANX2-YFP PT rupture was reduced compared to the corresponding untransformed genotype (). Moreover, in all T3 homozygous lines with good ANX1/2-YFP expression in the anx1 anx2 double mutant background, pollen germination, PT rupture, and PT length was indistinguishable from the WT (Figures 1A, 1B, S1, and S2). Thus, both ANX1-YFP and ANX2-YFP fusion proteins are functional, and the phenotypes observed in WT pollen expressing these fusion proteins are due to over-expression. Hereafter, independent homozygous lines expressing the ANX-YFP fusion proteins in the anx1 anx2 background will be called either complemented lines or ANX-YFP in anx1 anx2, while homozygous lines expressing the same fusion proteins in a WT background will be referred to as ANX-OX or ANX-YFP in WT.
Pollen of ANX-OX lines germinated poorly and produced shorter and wider PTs than pollen of either WT or complemented lines (Figure 1A, 1B, and S1). To check whether these in vitro phenotypes impact the fitness of PTs in vivo, the male transmission efficiency (TEM) was assayed for each of the ANX-OX lines. Male transmission of ANX-YFP fusions was significantly decreased for all but one ANX2-OX line, showing that PTs over-expressing ANX-YFP fusion proteins are not as competitive as untransformed WT PTs (Figure 1C). Interestingly, the difference in the severity of these phenotypes between ANX1-OX lines or between ANX2-OX lines nicely correlated with the difference in the level of YFP fluorescence imaged at the PM of growing ANX-OX PTs (Figure 1D). The two strongest ANX1-OX lines (#1 and #4) and one ANX1-YFP complemented line were selected for further investigations.
Time-lapse imaging of YFP fluorescence in growing PTs 6 hours after incubation showed that all PTs of the complemented line display the previously reported asymmetric distribution of YFP in the PM at the PT tip [19] and were growing normally (n>100, Figure 2A, left panels). In contrast, only 43% to 47% of ANX1 over-expressing PTs grew and exhibited the same YFP distribution (n>100 PTs, ANX1-OX #4 and #1, respectively). The remaining ANX1-OX PTs (53% to 57% of all PTs) had ceased to elongate and displayed PM invaginations at the PT tip as observed with both YFP and the lipophilic dye FM4-64 (Figure 2A, right panels). Intriguingly, in the ANX1 over-expressing PTs that had ceased to elongate, the PM at the tip kept growing inwards, creating tunnel-like structures, instead of outwards as normally observed for tip-growing cells (Figures 2B and S3A; Video S1). Invaginations can start early as they were observed even in pollen grains that did not yet produce a tube (Figure S3B). The PM invagination phenotype was also observed in ANX2 over-expressing lines, while we never saw it in any of the ANX1-YFP or ANX2-YFP complemented lines (n>100 PTs, two independent lines for each fusion protein). PM invaginations were accompanied by thick extracellular deposits of CW material (Figure 2A, asterisk), which were pectinaceous as revealed by Ruthenium red staining (Figure 2C). This finding indicates that secretion of CW material still occurred at the site of PM invagination. In addition, detailed observations of ANX-OX PTs showed that apical CW thickening occurred before the invagination of the apical membrane.
Since the surface of both the PM and the secreted CW material increase at the tip, we hypothesize that the balance between endocytosis and exocytosis rates might be tilted towards exocytosis at the tip of ANX-OX PTs. This could be achieved by a decrease or increase in the rate of endocytosis or exocytosis, respectively, or by a combination of both. For example, CW accumulation and PM invaginations have also been reported for tobacco PTs that over-express the phosphatidylinositol-4-phosphate 5-kinases PIP5K4, PIP5K5, and PIP5K6 [29]–[31]. The PIP5K-OX phenotypes originate from an over-initiation of aborted endocytosis in PIP5K-OX PTs, which show a dramatic inhibition of FM4-64 uptake [30],[31]. To investigate this phenotype further, we conducted two types of experiments on growing PTs of complemented and ANX1-OX lines before they start to show PM invaginations and apical CW accumulation. First, we labeled PTs for 5 min with FM4-64, a styryl dye that quickly labels the PM and is internalized via endocytosis. In growing PTs of complemented lines (as in WT), FM4-64 is observed at the PM and in the apical cytoplasm as an inverted cone that presumably contains both endocytotic and secretory vesicles (Figure 3A and Video S2, upper panels). In growing ANX1-OX PTs, the same distribution was observed (Figure 3A and Video S2, lower panels), indicating that, in contrast to PIP5K-OX PTs, FM4-64 uptake and thus endocytosis was not impaired [30],[31]. However, FM4-64 fluorescence intensity at the PM versus the apical cytoplasm was significantly lower than in complemented PTs, suggesting that there were globally more endocytotic and secretory vesicles in ANX1 over-expressing PTs (Figure 3A and 3B; n>25 each, p<0.01).
As evidenced by Brefeldin A (BFA) treatment, a well-known inhibitor of exocytosis, ANX1-YFP is inserted at the apical PM via exocytosis (Figure S4A). Thus, we performed fluorescence recovery after photobleaching (FRAP) experiments for ANX1-OX and complemented PTs to analyze exocytosis dynamics in growing PTs as described previously [31],[32]. Photobleaching was applied to the tip of growing PTs and measurements of the recovery of YFP fluorescence in the apical PM of the PT tips were carried out every 4 seconds. For ANX1 complemented PTs, the relative fluorescence recovery in the PM 10 seconds after photobleaching (I10sec) reached on average 47%±9% of the maximum relative fluorescence with a PT growth rate of 4.02±1.41 µm min−1 (n = 18; Figure 3C and 3D, left panels; Table S1; Video S3). No correlation was observed between I10sec and the fluorescence intensity pre-bleaching (R2 = 0.0092; Figure S4B), suggesting that the secretion rate of new ANX1-YFP fusion protein in the PM is independent of the amount of fusion protein originally present in the PM. Furthermore, no correlation was observed between I10sec and PT growth rate (R2 = 0.0001; Figure S4C), indicating that exocytosis and PT growth rate do not share a direct linear relationship.
Interestingly, for PTs of both ANX1 over-expressing lines, I10sec was significantly higher than in the complemented line (57%±10%, p<0.01 for line #1; 57%±13%, p<0.05 for line #4), while their PT growth rate was significantly decreased to 1.34±0.68 µm min−1 and 1.78±0.53 µm min−1, respectively (n = 17 and n = 20 for ANX1-OX lines #1 and #4, respectively, p<0.01; Figure 3C and 3D right panels; Table S1; Video S3). The faster fluorescence recovery is unlikely to be due to a secondary effect of slow PT growth, because all the mutant PTs tested so far in FRAP experiments, namely DN-ROP1-OX (dominant negative ROP1), CA-ROP1-OX (constitutively active ROP1), RIC3-OX, RIC4-OX, and PIP5K6-OX grow slower than controls and show an inhibition of fluorescence recovery [31],[32]. Thus, increased rates of fluorescence recovery at the apical PM indicate that the rate of exocytosis is increased at the apical PM of growing ANX1-OX PTs as compared to controls.
Altogether, our results support the hypothesis that ANX over-expression tilts the balance of exo- to endocytosis towards more exocytosis, which progressively leads to CW accumulation. PT growth slows down as the apical CW thickness increases. When the latter reaches a certain threshold where the CW is not deformable anymore, expansion ceases and apical PM grows inwards due to continuing exocytosis.
A better understanding of how the ANX RLKs regulate exocytosis requires the identification of downstream components of the ANX-dependent pathway. Recently, FER, which is the closest homologue of the ANX RLKs in Arabidopsis, has been shown to function as an upstream regulator of the ROP2/NADPH oxidase RbohC signaling pathway that controls ROS-dependent root hair growth [18]. Moreover, down-regulation of a pollen-expressed NADPH oxidase and application of ROS scavengers inhibit PT growth in tobacco [21]. Thus, we hypothesized that pollen-expressed NADPH oxidases could be downstream components of the ANX RLK pathway that coordinates CW integrity and PT growth. In Arabidopsis, NADPH oxidases belong to a family with ten members, two of which, RbohH (At5g60010) and RbohJ (At3g45810), sharing 81% amino acid identity, define a subgroup that is preferentially expressed in pollen (Figure S5C) [13],[14]. We isolated two independent single T-DNA insertional mutants for each of these NADPH oxidases, namely rbohH-1 (GABI_028G04), rbohH-3 (SALK_136917), rbohJ-2 (SAIL_31_D07), and rbohJ-3 (SALK_050665), which show little or no expression of the corresponding gene (Figure S5B). Pollen germination assays showed that PTs of single rbohJ-2 and rbohJ-3 mutant plants behaved like WT (∼8.5% bursting), while around 57% of PTs of single rbohH-1 and rbohH-3 mutant plants ruptured in vitro (Figure 4A). However, this mild PT rupture phenotype did not significantly reduce seed set or TEM in vivo (Figure 4A; Table S2). To investigate whether RbohH and RbohJ are redundant, single mutants were crossed to generate independent double mutant rbohH-1 rbohJ-2 and rbohH-3 rbohJ-3 plants. First, double homozygous mutant plants were only rarely found in the progeny of rbohH-1/RbohH rbohJ-2/rbohJ-2 and rbohH-3/RbohH rbohJ-3/rbohJ-3 (Table S3). Secondly, PTs of both independent rbohH-1 rbohJ-2 and rbohH-3 rbohJ-3 double mutants ruptured up to 80% in vitro (Figure 4A and 4B). The remaining germinating grains produced longer PTs, but they eventually also burst (Figure 4B). As a consequence, double homozygotes for rbohH-1 rbohJ-2 and rbohH-3 rbohJ-3 were partially sterile, producing only five to seven seeds per silique as compared to ∼60 seeds in WT or single mutant plants (Figure 4A and 4C). As evidenced by aniline blue staining after reciprocal crosses of rbohH rbohJ with WT, this sterility was due to the double mutant pollen being unable to grow sufficiently in vivo to reach and fertilize the female gametophytes (Figure 4D). This was further supported by analyses of male and female transmission efficiencies (TEs) of the rbohH-1 rbohJ-2 and rbohH-3 rbohJ-3 mutations, which showed a greatly reduced TEM while TEF was not significantly affected (Table S2).
Taken together, these results provide compelling evidence that disruption of both RbohH and RbohJ leads to spontaneous PT rupture, preventing PTs to reach and fertilize the female gametophytes in vivo. Interestingly, all the above mentioned phenotypes are reminiscent of the anx1 anx2 double mutant phenotype [19],[20]. Moreover, our results show partial functional redundancy between pollen-expressed NADPH oxidases, with RbohH being able to perfectly substitute for the loss of RbohJ, while the latter can only partially substitute for the loss of RbohH.
To test whether RbohH and RbohJ are indeed downstream effectors of the ANX RLK-dependent pathway, the strong ANX1-OX line (#4) was crossed to rbohH-1 rbohJ-2 double mutant plants. Partially male sterile plants homozygous for rbohH-1 rbohJ-2 and homozygous for ANX1-YFP were retrieved in the F2 generation. Intriguingly, rbohH-1 rbohJ-2 pollen strongly expressing ANX1-YFP behaved exactly like rbohH-1 rbohJ-2 pollen with germination and PT rupture rates of ∼70% and 82%, respectively (Figure 5A). Furthermore, none of the growing PTs (n>100) of rbohH-1 rbohJ-2 plants homozygous for ANX1-YFP displayed CW accumulation or PM invagination, phenotypes observed in ANX1-OX PTs. To independently confirm these results, we directly transformed rbohH-1 rbohJ-2 mutant with ANX1-YFP fusion. Four independent, partially male sterile rbohH-1 rbohJ-2 transgenic lines homozygous for ANX1-YFP were recovered in the T2 generation. Again, neither CW accumulation nor PM invagination was observed in growing PTs (n>100 PTs for each), which eventually ruptured similar to the rbohH-1 rbohJ-2 PTs without the ANX1-OX construct. Furthermore, FRAP analyses also showed that the fast recovery rate at the apical PM observed in ANX1-OX PTs was suppressed in the rbohH rbohJ background as I10sec for rbohH rbohJ PTs over-expressing ANX1-YFP was similar to the complemented line (45%±9%, n = 23; Figure 5B and 5C; Table S1). Interestingly, a few rbohH rbohJ PTs over-expressing ANX1-YFP did not recover 80% of the initial fluorescence, a phenomenon that was never observed in controls or ANX1-YFP over-expressing PTs (Figure S6), indicating that exocytosis may become defective in these rbohH rbohJ ANX1-OX PTs.
In summary, these results demonstrate that ANX1-OX phenotypes are dependent on functional RbohH and RbohJ and, consequently, that these pollen-expressed NADPH oxidases are positive downstream effectors of the ANX RLK-dependent pathway.
To check whether disruption of RbohH and RbohJ impairs the production of ROS, we made use of the fluorescent ROS-sensitive dye 5-(and 6-)chloromethyl-2′,7′-dichlorodihydrofluorescein diacetate (CM-H2DCFDA) to stain PTs of WT and rbohH rbohJ double mutants. Fluorescence quantification of the apical cytoplasm in growing PTs treated for 5 minutes with 2 µM CM-H2DCFDA showed that PTs of the rbohH-1 rbohJ-2 and rbohH-3 rbohJ-3 double mutants displayed only 25% of the CM-H2DCFDA-derived fluorescence signal observed in WT PTs (p<0.01; Figure S7A and S7C). These low levels of CM-H2DCFDA-derived fluorescence were not due to a defect in dye uptake, as mutant and WT PTs exhibited the same level of fluorescence derived from the ROS-insensitive dye fluorescein diacetate (FDA) (Figure S7B and S7D). These results show that ROS production is indeed impaired in rbohH rbohJ PTs as expected for NADPH oxidase mutants [12]. However, because CM-H2DCFDA oxidation is sensitive to different reactive oxygen and nitrogen species, sensitive to light, and irreversible, this dye cannot be used to monitor ROS production over time in growing PTs. Thus, we generated stably transformed Arabidopsis lines with PT expression of the genetically encoded YFP-based ratiometric sensor HyPer, which has been shown to faithfully report H2O2 production in bacteria, animal, and plant cells [33],[34]. Curiously, in growing WT PTs expressing cytosolic HyPer (n = 27), the HyPer activity measured as the ratio of F488/F405 was stronger in the shank of PTs than at the tip (Figure 6A and 6B). We hypothesized that this strong shank activity could either be due to the presence of H2O2-producing organelles, such as mitochondria and/or peroxisomes in this region, an artifact of HyPer due to its pH sensitivity, or a combination of both. Indeed, it was shown that HyPer's activity artificially increases when the pH increases [33] and that PTs display a pH gradient with an acidic tip and a alkaline shank [35]. Interestingly, at the tip of growing PTs, HyPer activity displayed irregular oscillations originating from the tip periphery (Figures 6B and S8A; Video S4). However, oscillations of HyPer activity did not seem to correlate with growth rates (Figure S8B). In growing rbohH-1 rbohJ-2 PTs, HyPer activity was 16 and 18 times lower at the tip and in the shank, respectively, as compared to the WT (n = 22, p<0.001; Figure 6A and 6C). This indicates that membrane-bound RbohH and RbohJ are responsible for most of the H2O2 production revealed by the HyPer sensor. Moreover, since HyPer activity in the shank was also strongly reduced in rbohH-1 rbohJ-2 double mutant PTs (Figure 6A and 6C), the strong activity in the shank of WT PTs is likely due to propagation of the tip-derived H2O2 in the alkaline shanks, which artificially increases HyPer activity.
To investigate if Rboh localization is consistent with the Rboh-dependent H2O2 production observed at the tip-periphery, we transformed partially sterile rbohH-3 rbohJ-3 plants with a green fluorescent protein (GFP)-RbohH fusion. Forty-four independent T1 transgenic lines out of 50 displayed rescue of sterility with WT-like elongated siliques (e.g., for three independent T1 lines with good GFP expression, the average of seeds per silique was 40.8±3.8, 40.9±11.1, and 43.4±6.9 as opposed to 6.2±3.5 in untransformed rbohH-3 rbohJ-3, n = 12 siliques per plant). In vitro pollen growth assays confirmed that the rbohH rbohJ bursting phenotypes were rescued by GFP-RbohH (Figure S8C) and that GFP-RbohH localized polarly to the plasma membrane at the tip of growing PTs (Figure S8D, left panels). These results show that GFP-RbohH protein fusion is functional and that its localization is consistent with both ANX1-YFP localization (Figure 2A, left panels) and Rboh-dependent H2O2 production at the tip periphery (Figure S8A). Furthermore, unlike the rbohH rbohJ complemented plants, in WT plants expressing the GFP-RbohH fusion, PM invagination and over-accumulation of CW material were also observed (Figure S8D, right panels), although these phenotypes appeared milder and less frequent than in ANX1-OX PTs.
ROS and H2O2 have been shown to regulate calcium-permeable channels, e.g., in protoplast of root hairs [12] and pear pollen [36], and a tip-focused Ca2+ gradient is essential for polar growth [37]. Therefore, we crossed WT plants expressing the genetically encoded FRET-based Ca2+-cameleon YC3.60 in PTs [38] with the anx1-2 anx2-2 and rbohH-1 rbohJ-2 double mutants, and partially male sterile anx1-2 anx2-2 and rbohH-1 rbohJ-2 plants homozygous for YC3.60 were recovered in subsequent generations. Cytosolic Ca2+ concentrations ([Ca2+]cyt, measured as FCFP/FVenus) were monitored over time at the PT tip and behind the tip when possible, and compared to YC3.60-expressing WT PTs grown and imaged under the same conditions. First, we attempted to study Ca2+ dynamics in anx1 anx2 bulges before bursting, young growing WT PTs, and arrested WT bulges. Bulges of anx1 anx2 never produced a growing tube, and only two out 17 burst during imaging. Interestingly, for both anx1 anx2 bulges that eventually burst, a sudden increase of [Ca2+]cyt was observed (Figure S9A, white arrow) before the first visible sign of rupture (Figure S9A, black arrow; Video S5). However, before bursting, [Ca2+]cyt in non-growing anx1 anx2 bulges was on average lower than at the tip of growing WT PTs but similar to the arrested WT bulges (Figure S9B). Because one cannot conclude if the decreased levels of [Ca2+]cyt in anx1 anx2 are due to the lack of ANX RLKs or rather to an indirect effect of arrested growth, we focused on studying Ca2+ dynamics in growing WT PTs and rbohH rbohJ pollen grains, which produce a few growing PTs that eventually burst.
In steadily growing WT PTs, the tip-focused Ca2+ gradient (i.e., higher [Ca2+]cyt at the PT tip compared to behind the tip) was always observed and quite stable (n = 46; Figures 7A, 7C, S10A, and S10D; Video S6). Furthermore, as reported previously for in vitro grown Arabidopsis PTs [38],[39], but unlike lily PTs [40], we did not observe regular oscillations for either the PT growth rate or [Ca2+]cyt (Figure 7A). In growing rbohH rbohJ PTs (n = 30), [Ca2+]cyt was significantly lower than in the WT (p<0.001 for both tip and behind the tip regions; Figure 7B and 7C; Video S6). However, the tip-focused Ca2+ gradient and the PT growth rate were on average similar to that of WT PTs (Figure 7B and 7C; p = 0.054 for gradient, p = 0.84 for growth rate). But both the tip-focused Ca2+ gradient and the PT growth rate were significantly less steady over time in the rbohH rbohJ double mutant than in the WT, as evidenced by a significantly higher variance (p = 4.1280•10−13 and p = 0.008737 for [Ca2+]tip/[Ca2+]behind and growth rate, respectively; Figure S10A–S10F). The steady and jerky growth rate of WT and rbohH rbohJ PTs, respectively, were quite obvious during live-imaging of growing FM4-64 stained PTs (Video S7).
These results indicate that disruption of the pollen-expressed NADPH oxidases RbohH and RbohJ does not abolish the tip-focused Ca2+ gradient but results in PTs that display (i) overall lower [Ca2+]cyt levels, and (ii) unstable tip-focused Ca2+ gradients and growth rates. Finally, increasing the external [Ca2+] in the germination medium from 5 mM to 15 mM or 30 mM, significantly decreased the rupture of rbohH rbohJ PTs, while lowering the external [Ca2+] had the opposite effect (p<0.05; Figure S11). These findings indicate that supplementing Ca2+ externally can partially stabilize the growth of rbohH rbohJ PTs. Conversely, decreasing external [Ca2+] strongly, increases the frequency of PTs that rupture in the WT (Figure S11), consistent with a pioneering report from the 1980s [41].
In tip-growing root hairs, FER and RbohC/RHD2 have been proposed to function in the same pathway based on the facts that: (i) fer and rbohC/rhd2 display similar phenotypes with stunted, collapsed, and bursting root hairs, and (ii) that roots and root hairs of fer and one FER-OX line accumulate less and more ROS than WT, respectively [12],[18]. Similarly, we show here that two independent rbohH rbohJ double mutants display anx-like phenotypes, i.e., PTs that burst, preventing them from growing to fertilize the female gametophytes (Figure 4). Consequently, both anx1 anx2 and rbohH rbohJ mutant plants are nearly male sterile. In addition, over-expression of both ANX1-YFP and GFP-RbohH triggers over-accumulation of membrane and CW materials (Figures 2 and S8D, right panels). Furthermore, we provide strong genetic evidence for the NADPH oxidases to function downstream of the ANX RLKs, by showing that the phenotypes observed in ANX-OX lines are abolished in the rbohH rbohJ mutant background (Figure 5). Therefore, the CrRLK1L-NADPH oxidase signaling module appears to be conserved in tip-growing cells. However, it is unlikely that the CW integrity pathway in pollen is a carbon copy of the root hair pathway, as the biological functions, growth habits and patterns, CW compositions, and growth environments are quite different between these tip-growing cells [42]. For example, in root hairs FER has been shown to positively regulate RbohC-dependent ROS production through ROP2-signaling [18]. In PTs, however, it remains unclear whether ANX RLKs also activate the NADPH oxidases RbohH and RbohJ through ROP-signaling, because over-activation of ROP-signaling leads to growth depolarization but does not trigger CW accumulation, PM invagination, or increased apical exocytosis [32],[43],[44], as we observed it in ANX1-OX lines (Figures 2 and 3).
Our understanding of the role of NADPH oxidase-derived ROS signaling in plant development and in responses to abiotic and biotic stresses has improved tremendously over the past few years [13],[14],[27]. Production of different ROS species has been imaged in different plant tissue and cell types, but because of the irreversible oxidation of the different dyes used (e.g., diaminobenzidine tetrahydrochloride, nitro blue tetrazolium [NBT], H2DCFDA and derivatives) meaningful information about the dynamics of ROS production is still scarce [45]. GFP-based, genetically encoded sensors such as roGFPs and HyPer, which display reversible changes in fluorescence to alterations in redox/ROS levels, have been successfully developed and tested in plant cells [45]. However, none of them have been assayed in a mutant background affecting ROS-producing enzymes. Here, we used the cytoplasmic H2O2-selective HyPer sensor expressed in PTs in a rbohH rbohJ NADPH oxidase-deficient mutant background to gain more insights into H2O2 production in tip-growing cells. HyPer activity displayed irregular oscillations at the tip of growing WT PTs (Figures 6B and S8). HyPer oscillations are unlikely due to pH oscillations reported for the tip of growing PTs, because (i) pH at the tip varies [35] in a range where HyPer is not really pH-sensitive [33], and (ii) HyPer activity is completely abolished in growing rbohH rbohJ mutant PTs (Figure 6A and 6C). Moreover, HyPer activity originated from the periphery of the growing tip (Figure S8A), which is consistent with (i) the tip-preferential PM RbohH localization (Figure S8D, left panels) and the reported PM localization of other NADPH oxidases [9],[46], (ii) the NADPH oxidase activity reported at the PM [47],[48], and (iii) the extracellular, tip-localized O2.− distribution revealed by NBT staining of PTs [21],[48].
The exact mechanism by which NADPH oxidase-dependent ROS regulate polar growth is still not fully understood. One reason for this is that quantitative information with good temporal and spatial resolution is difficult to obtain from growing CrRLK1L and/or NADPH-oxidase mutant cells (root hairs or PTs), owing to their rapid loss of cellular integrity. On one hand, the NADPH oxidase RbohC has been proposed to generate ROS that activate Ca2+-permeable channels at the PM to establish the tip-focused Ca2+ gradient and to promote expansion at the tip of root hairs [12],[25]. On the other hand, a tip-focused Ca2+ gradient was observed in rbohC root hairs under certain conditions, indicating that RbohC was not essential to generate the Ca2+ gradient, but rather plays a role in restricting growth to the tip by rigidifying the CW behind the tip [26]. On the basis of our results we propose a third alternative. Unlike anx1 anx2, a small but appreciable number (∼20%) of germinating rbohH rbohJ pollen grains are able to produce longer tubes in vitro that, however, will eventually burst, too. We took advantage of this opportunity to study [Ca2+]cyt dynamics with a good spatial and temporal resolution on growing NADPH oxidase-deficient PTs. First, the tip-focused Ca2+ gradient, visualized by the ratio between [Ca2+]tip/[Ca2+]behind, was clearly visible in growing rbohH rbohJ PTs, confirming that NADPH oxidases are not required to generate the Ca2+ gradient. However, unlike steadily growing WT PTs, which maintain a constant Ca2+ gradient over time (Figure S10A, S10D, and S10E), rbohH rbohJ PTs displayed a very unstable gradient, which could sometimes be steep but was abolished a few seconds later (Figure S10B, S10D, and S10E). This was correlated with more variable growth rates in rbohH rbohJ mutant compared to steadily growing WT PTs. Moreover, the global cytosolic Ca2+ levels were significantly lower in the growing rbohH rbohJ mutant PTs compared to WT PTs (Figure 7C). An increase in external [Ca2+] partially rescued the rupture of rbohH rbohJ PTs, while lowering the external [Ca2+] increased PT rupture in both the mutant and WT (Figure S11). This is in agreement with previous studies, which showed that lowering external [Ca2+] or limiting/blocking Ca2+ influx causes PTs and root hairs to burst [41],[49]. The data reported here are consistent with NADPH oxidase-dependent ROS activating Ca2+-permeable channels for Ca2+ influx [12],[36]. However, we propose that these yet unidentified channels do not generate the tip-focused Ca2+ gradient on their own; rather, they fine tune the Ca2+ gradient by stabilizing it to sustain steady growth of Arabidopsis PTs. It is noteworthy that different types of PM-localized Ca2+ channels have been characterized recently in tip-growing cells [37],[50]. Among these, the Cyclic-Nucleotide-Gated Channel (CNGC) family is of particular interest, because single cngc18 or double cngc7 cngc8 mutant PTs spontaneously rupture after germination or produce kinky PTs that often burst as well [51],[52]. Thus, CNGCs constitute good candidates for Ca2+ channels that are regulated by the CrRLK1L-NADPH oxidase signaling module at the PM. Annexins are also possible candidates as ANN1 has recently been shown to function as a ROS-activating Ca2+ transporter in root cells [53].
One of the many roles proposed for the tip-focused Ca2+ gradient is to facilitate and stimulate exocytosis at the site of growth [54],[55], where the exocyst complex has been shown to function [56]. Increasing external [Ca2+] leads to root hair and PT growth inhibition and CW thickening [41],[49]. However, in this case it is not clear whether the accumulation of secreted CW material is due to an increase of the exocytosis rate or to uncoupling of exocytosis (that otherwise remains the same) from growth. Interestingly, we found that ANX over-expressing PTs grow slower than controls and also display CW accumulation (Figure 2). By performing FRAP analyses in the apical membrane of growing PTs of WT, anx1 anx2, and rbohH rbohJ plants expressing the ANX1-YFP fusion protein, we show that the rate of exocytosis is significantly increased in ANX1-OX PTs compared to controls. In contrast, for some of the rbohH rbohJ mutant PTs that have low calcium levels and an unsteady Ca2+ gradient, the recovery was impaired during our analysis. In agreement, the Ca2+channel blocker LaCl3, which has been shown to trigger the rupture of root hairs [49], appears to inhibit FRAP at the PT tip [32].
Altogether our data are consistent with the following sequence of events: ANX RLKs positively regulate the NADPH oxidases RbohH and RbohJ, possibly through ROP signaling, to periodically produce ROS. Subsequently, ROS activate Ca2+-permeable channels for calcium influx to fine tune the tip-focused Ca2+ gradient, which in turn sustains secretion at the apical tip enabling PTs to elongate steadily without a loss of CW integrity. Perturbations of the pathway by over-expressing ANX RLKs would lead to a NADPH oxidase-dependent over-production of ROS and Ca2+ influx at the PT tip, which in turn would increase the secretion rate of membrane and CW materials, progressively leading to growth cessation and membrane invagination. Conversely, disrupting the ANX RLKs or NADPH oxidase would abolish NADPH oxidases-dependent ROS production and impair the opening of ROS-activated Ca2+-permeable channels, thus limiting the cell's ability to buffer [Ca2+]cyt variation that is required to maintain a steady tip-focused Ca2+ gradient. Consequently, the Ca2+ gradient and exocytosis at the PT tip would become erratic and, if not stabilized by compensatory mechanisms, the CW thickness would decrease until turgor pressure would lead to PT rupture. Finally, our model does not exclude that, in parallel to the signaling events described above, NADPH oxidase-dependent ROS and/or Ca2+ could directly alter CW properties, thereby affecting PT tip-growth. To investigate this possibility, direct measurements of the impact of ROS on CW properties during tip-growth would need to be established. We are confident that combining the continuously improving polar growth models and techniques to measure mechanical properties of growing cells [57],[58] with genetic approaches, will soon uncover some of the remaining mysteries of the fascinating coordination between CW integrity and polar growth.
After 3 h to 5 h of in vitro incubation on solid germination medium, 100 µl of liquid germination medium containing 0.01% Ruthenium red (Sigma, R-2751) or 2 µM of either FM4-64 (Molecular Probes, Invitrogen, T3166), CM-H2DCFDA (Molecular Probes, Invitrogen, C6827), or FDA (Sigma, F7378) were applied for 5 min to PTs, then washed away with fresh dye-free medium before imaging. PTs stained with Ruthenium red were imaged with a Leica DM6000 (Leica Microsystems). PTs stained with either FM4-64, CM-H2DCFDA, or FDA were imaged with a Leica SP2 or SP5 confocal microscope. For FM4-64 stained PTs, the apical PM region was defined as the first 2.5 µm along PM at the apex. A circle (2.5 µm in diameter) 3 µm away from tip was chosen for measurement of apical cytosol intensity. Relative localization of the FM4-64 dye on the PM versus the apical cytosol was calculated to illustrate the degree of FM4-64 internalization. For CM-H2DCFDA and FDA stained PTs, a circle (4 µm in diameter) 3 µm away from tip was chosen to measure apical cytosol intensity. All dye-derived fluorescence intensities were measured using the ImageJ 1.47d software after background subtraction. PTs of different genotypes were all imaged and quantified under the same conditions.
Growing PTs expressing ANX1-YFP in an anx1-2 anx2-2 (complemented line), WT (ANX-OX, lines #1 and #4), and rbohH-1 rbohJ-2 backgrounds were used for FRAP analyses with the same imaging and quantification parameters. The apical region of PTs was photobleached using 100% power of a 514-nm laser (Leica SP5), and the recovery of fluorescence was monitored every 4 s in the following 2 min. The apical PM region was defined on the bright-field pictures at every time frame as the first 2.5 µm along PM at the apex, and fluorescence intensities were measured with ImageJ 1.47d software after background subtraction. Relative intensity of PM-localized ANX1-YFP compared with fluorescence before photobleaching was used to quantify the speed of fluorescence recovery. See Table S1 for curve fitting.
Fluorescence in growing PTs of WT and rbohH-1 rbohJ-2 expressing either HyPer or YC3.60 were acquired (Leica SP2 confocal microscope) and quantified (ImageJ 1.47d) in the exact same conditions. For HyPer, fluorescence was acquired with the sequential mode and excitation at 488 nm and emission between 500–540 nm for F488 and excitation at 405 nm and emission between 500–540 nm for F405. Two circular regions of interest (ROIs; 4 µm in diameter), one 0.5 µm, the other 20 µm away from the apex were drawn for measurement of apical cytosol and far behind the tip intensities, respectively, for each single time point of each PT. For YC3.60, excitation was 458 nm then emission 469–501 nm for FCFP and 522–554 nm for FVenus. Two circular ROIs (4 µm in diameter), one 0.5 µm, the other 10 µm away from the apex, were drawn for measurement of apical cytosol and behind the tip intensities, respectively, for each single time point of each PT. All ratiometric measurements, i.e., F488/F405 and FCFP/FVenus, were determined with ImageJ 1.47d and its Ratio ROI Manager plugin after background subtraction. Ratiometric pictures were generated with the plugin Ratio Stack after median filtering. The YC3.60 titration curve (Figures 7D) was obtained as described before [38].
All primers used in this study are listed in Table S4. Additional protocols are described in Text S1.
|
10.1371/journal.pntd.0005174 | Emerging Infectious Disease Implications of Invasive Mammalian Species: The Greater White-Toothed Shrew (Crocidura russula) Is Associated With a Novel Serovar of Pathogenic Leptospira in Ireland | The greater white-toothed shrew (Crocidura russula) is an invasive mammalian species that was first recorded in Ireland in 2007. It currently occupies an area of approximately 7,600 km2 on the island. C. russula is normally distributed in Northern Africa and Western Europe, and was previously absent from the British Isles. Whilst invasive species can have dramatic and rapid impacts on faunal and floral communities, they may also be carriers of pathogens facilitating disease transmission in potentially naive populations. Pathogenic leptospires are endemic in Ireland and a significant cause of human and animal disease. From 18 trapped C. russula, 3 isolates of Leptospira were cultured. However, typing of these isolates by standard serological reference methods was negative, and suggested an, as yet, unidentified serovar. Sequence analysis of 16S ribosomal RNA and secY indicated that these novel isolates belong to Leptospira alstonii, a unique pathogenic species of which only 7 isolates have been described to date. Earlier isolations were limited geographically to China, Japan and Malaysia, and this leptospiral species had not previously been cultured from mammals. Restriction enzyme analysis (REA) further confirms the novelty of these strains since no similar patterns were observed with a reference database of leptospires. As with other pathogenic Leptospira species, these isolates contain lipL32 and do not grow in the presence of 8-azagunaine; however no evidence of disease was apparent after experimental infection of hamsters. These isolates are genetically related to L. alstonii but have a novel REA pattern; they represent a new serovar which we designate as serovar Room22. This study demonstrates that invasive mammalian species act as bridge vectors of novel zoonotic pathogens such as Leptospira.
| Leptospirosis is a global zoonotic disease. Pathogenic species of Leptospira are excreted in urine from asymptomatic carrier hosts which facilitates disease transmission to new hosts. To date, there are 10 species of pathogenic leptospires which comprise more than 200 serovars. Disease transmission of these strains is maintained by a wide range of domestic and wild animal species. In this work, we discovered that an invasive mammalian species, the greater white toothed shrew, which was first identified in Ireland in 2007, acts as a carrier for a species of leptospires never before identified in Ireland. Results demonstrate that invasive mammalian species act as bridge vectors of novel zoonotic pathogens such as Leptospira.
| The greater white-toothed shrew (Crocidura russula) is an exotic species to Ireland first recorded in 2007[1], and now classified as an invasive mammalian species[2]. According to recent studies, this species is rapidly spreading with radial expansion estimates of approximately 5.5 km/yr[2]. The source of this invasive population is from Europe as opposed to North Africa[3], and evidence suggests that the greater white-toothed shrew is associated with the local extinction of indigenous populations of the pygmy shrew (Sorex minutus)[2]. However, a comprehensive investigation on the One Health implications of this invasive species has yet to be performed.
Pathogenic species of Leptospira cause leptospirosis, a bacterial zoonotic disease with a global distribution affecting over one million people annually[4, 5]. Leptospires colonize the renal tubules of reservoir hosts, from where they are excreted via urine into the environment and survive in suitable moist conditions. Contact with infected urine, or contaminated water sources can result in disease since pathogenic leptospires can penetrate breaches of the skin, or mucosal surfaces, and disseminate haematogenously to cause a range of clinical symptoms from mild fever, to icteric Weil’s disease and pulmonary hemorrhage syndrome. In developed countries, leptospirosis is primarily a recreational disease, or occupational disease of farm workers, veterinarians, and slaughter plant workers. In developing countries, it is a socioeconomic disease perpetuated by rapid urbanization, rodent infestation and transmission via contaminated water sources associated with limited infrastructures and severe weather events. Both rodents and domestic farm animal species can serve as reservoir hosts of infection and sources of disease transmission to humans.
Leptospirosis is endemic in Ireland[6–12]. The mean annual incidence for 2009 was 5.6 per million inhabitants per annum, compared to that of 1.4 per million across the EU[13]. The predominant serovars associated with human infection were serovars Icterohaemorrhagiae and Hardjo, indicative of rodent/recreational and occupational exposure respectively. Rats are reservoir hosts for serovar Icterohaemorrhagiae whilst cattle act as reservoir hosts for serovar Hardjo[14]. Over 80% of Irish beef suckler herds, and more than 40% of individual beef producing animals, show evidence of exposure to leptospires[15]. Similarly, 79% of unvaccinated dairy herds were positive for antibodies to Leptospira by bulk tank milk testing[16]. Leptospirosis continues to be a leading cause of bovine abortion[17]. Other domestic animals species that show evidence of exposure to pathogenic leptospires in Ireland include pigs, sheep, horses and dogs[18–26].
There is clear evidence that invasive species act as vectors for pathogens and parasites that can have environmental conservation, and human health, implications. Globalization has facilitated the movement of exotic and invasive species, and a range of associated pathogens e.g. mosquitoes and West Nile Virus[27]. The combination of invasive species and degradation of ecosystems presents a substantial threat in relation to emerging infectious diseases[27, 28]. Novel pathogens can have devastating effects on naive communities; examples include the invasive grey squirrel (Sciurus carolinensis) which carries squirrelpox virus that severely adversely affected native red squirrels (Sciurus vulgaris) in Britain and Ireland[29, 30]; the introduced raccoon dog (Nyctereutes procyonoides) in Europe, which has an expanding range, and which can facilitate the spread of infectious diseases including echinococcosis, trichinellosis and rabies[31]. In this study, we identified that a recently introduced mammalian species (C. russula) in Ireland is a reservoir host for a novel strain of pathogenic Leptospira.
Greater white-toothed shrews (GWTS) were live-trapped and euthanized by cervical dislocation. All animal experimental procedures were performed in accordance with relevant guidelines and regulations, and as approved by the National Parks and Wildlife Service (NPWS) in Ireland and the Animal Research Ethics Committee in University College Dublin (AREC-13-24).
Kidneys were removed from GWTS at time of euthanasia and immediately processed for the culture of leptospires[32]. In brief, a single kidney was aseptically removed using a disposable forceps and scalpel and placed in 5 ml 1% Bovine Serum Albumin (BSA). The kidney was subsequently macerated with scalpels and the resulting mixture homogenized by passing it through a 10ml syringe (without needle attachment). Each tissue homogenate was serially diluted 10-fold (to a final dilution of 10−3) into 1% BSA and 500μl of this mixture was used to inoculate the surface of 10ml EMJH medium containing 200μg 5-Fluoruracil and 0.2% noble agar. Cultures were transported back to the laboratory and maintained at 29°C. Cultures were examined at weekly intervals by dark-field microscopy.
L. alstonii Serogroup Ranarum Serovar Pingchang Strain 80–412 and L. alstonii Serogroup Undesignated Serovar Sichuan Strain 79601 were sourced from the WHO/OIE Leptospirosis Reference Laboratory at the Royal Tropical Institute, The Netherlands. L. alstonii strains MS267, MS311 and MS316 were kindly provided by Department of Bacteriology, Faculty of Medical Sciences, Kyushu University, Japan.
Growth assessment in the presence of 8-azaguanine was performed as previously described[33]; in brief, leptospires were cultured in EMJH medium with 1% rabbit serum and 225 μg/ml 8-Azaguanine (A5284 8-Azaguanine, Sigma, St. Louis, MO). Duplicate tubes were inoculated with the shrew isolates while Leptospira biflexa (ATCC® 23582™) was used as a positive control. Cultures were incubated at 30°C for 14 days. The cultures were counted by dark-field microscopy at days 1, 3, 5, 7 and 14 using a Cellometer® disposable cell counting chamber (Nexcelom Bioscience).
Serological strain identification was initially attempted by cross-agglutination. In this procedure, the Microscopic Agglutination Test (MAT) was carried out using a panel of 19 reference antisera against the 17 major pathogenic Leptospira serogroups[34–36]. The Leptospira serogroups tested included Australis (serovars Australis and Bratislava), Autumnalis, Ballum, Canicola, Celledoni, Cynopteri, Grippotyphosa, Hebdomadis, Icterohaemorrhagiae, Javanica, Louisiana, Mini, Pomona (serovar Pomona and Altodouro), Pyrogenes, Sejroe, Semaranga and Tarassovi. In addition, rabbit sera generated against each of the three shrew isolates were then tested against the panel of Leptospira antigens from the 17 serogroups mentioned above, and additionally against a panel of 9 antigens from serogroups comprised of: Andamana, Semaranga, Hursbridge, Sarmin, Lyme, Louisiana, Shermani (serovar Shermani and Aquaruna), Bataviae, Ranarum, and against one undesignated serogroup (serovar Sichuan).
Four hundred ml culture grown from each shrew isolate of Leptospira was harvested and whole cell leptospiral DNA purified as previously described[18]. DNA concentration was estimated after spectrophotometric measurement using a Nanophotometer Pearl (Implen). Restriction endonuclease digestion with EcoRI, electrophoresis and gel analysis were carried out as previously described[18].
Rabbit sera were prepared as previously described with slight modification[34] and as licensed under the Animals (Scientific Procedures) Act (1986). In brief, rabbits were injected intraperitoneally at weekly intervals with live leptospires at a density of 2 x 108 per ml. The weekly injected doses were 5, 10, 15, and 20 ml respectively. Rabbits were bled by cardiac puncture one week after the last injection.
Genome sequencing was performed by the Centre for Genomic Research at the University of Liverpool. Genomic DNA material was purified with 1x cleaned Ampure beads (Agencourt) and the quantity and quality was assessed by Nanodrop and the Qubit assay. In addition, the Fragment Analyser (using a high sensitivity genomic kit) was used to determine the average size of the DNA and the extent of degradation. This procedure was also used at the steps indicated below to determine average fragment size of the DNA. DNA was sheared using Covaris G tubes by centrifugation at 7,000 rpm in an Eppendorf 5415R centrifuge. The fragment size was checked as before. DNA was purified with 0.5x ampure beads and treated with Exonuclease VII at 37°C for 15 minutes. The ends of the DNA were repaired as described by Pacific Biosciences protocol. Each sample was incubated for 20 minutes at 37°C with DNA Damage Repair Mix supplied in the SMRTbell library kit (Pac Bio). This was followed by 5 minutes incubation at 25°C with End Repair Mix. DNA was cleaned using 0.5x ampure and 70% ethanol washes. DNA was ligated to adapter sequences overnight at 25°C. Ligation was terminated by incubation at 65°C for 10 minutes followed by exonuclease treatment for 1 hour at 37°C. The SMRTbell library was purified with 0.5x ampure beads. The quantity of library and therefore the recovery was determined by Qubit assay and the average fragment size determined by Fragment Analyser. SMRTbell library was annealed to sequencing primer at values predetermined by the Binding Calculator (Pac Bio) and a complex made with the DNA Polymerase (P6/C4 chemistry). The complex was bound to Magbeads and this was used to set up 3 SMRT cells for sequencing. Sequencing was done using 240 minute movie times.
The 16S rRNA gene sequence identified within the newly sequenced organism described herein was used to retrieve 108 similar sequences from the Ribosomal Database Project (RDP) via the SeqMatch tool[37]. Sequences were aligned with MUSCLE[38], and divergent and ambiguously aligned alignment blocks were removed with Gblocks[39]. The modelTest feature of Phangorn[40] was used to calculate the Bayesian Information Criterion (BIC) for a variety of models, and guided the selection of the HKY model. The model parameters for computing the maximum likelihood of phylogeny were optimized using optim.pml, and bootstrap.pml was used to perform a bootstrap analysis[40]. The phylogenetic reconstruction with bootstrapped values assigned to the edges was graphically rendered with TreeDyn[41].
The secY gene sequence identified within the newly sequenced organism described herein was compared with other sequences of secY from the genus Leptospira, as retrieved from GenBank[42]. Sequences of secY were aligned with CLUSTAL W[43]. Phylogenic analysis was conducted with MEGA4[44] and the maximum likelihoods method was used for estimation of distance of aligned sequences[45].
Golden Syrian hamsters were inoculated by intraperitoneal (IP) injection as previously described[46]. Groups of three hamsters each received 107 of GWTS isolate #1, #2 or #3 IP respectively. Three hamsters acted as negative controls and received media alone. All animal experimental procedures were performed in accordance with relevant guidelines and regulations, and as approved by USDA Institutional guidelines.
The microscopic agglutination test was performed as previously described according to OIE guidelines[47].
The fluorescent antibody test was performed as previously described[32].
The annotated assembly for L. alstonii serovar Room22 strain GWTS#1 is available in GenBank under the accession numbers CP015217 (Chromosome I) and CP015218 (Chromosome II).
Culture of leptospires was attempted from a single kidney in each of 18 trapped GWTS. Kidneys from three of the GWTS were culture positive as confirmed by dark-field microscopy and the isolates were named GWTS Isolate #1, #2 and #3 respectively.
Each GWTS isolate of Leptospira was tested against a standard panel of reference antisera, representing 19 serovars from 17 serogroups and representative of the geographical locale, for typing purposes, Table 1. No significant reactivity was detected between any GWTS isolate and any reference sera. In a further attempt to type each GWTS isolate, rabbit antisera specific for each GWTS isolate was then prepared and tested against an additional panel of reference strains of Leptospira, representing 9 serogroups, one undesignated serogroup, and 13 serovars, Table 2. Slight reactivity was detected by antisera specific for GWTS isolate #1 & #2 against serovar Shermani, which belongs to Leptospira santarosai. However, the lack of a consistently high MAT titre detected between GWTS isolate-specific antisera and reference antigen indicated an inconclusive serological typing classification of any of the GWTS isolates, and suggesting that they were of an as yet unidentified serovar.
The inability to serologically type the GWTS Leptospira isolates using reference antisera and reference antigens indicates that the GWTS Leptospira isolates are atypical compared to those previously identified in Western Europe. Therefore, whole genome sequencing was performed on a single strain, GWTS isolate #1. The gene sequence for 16S rDNA was extracted from the complete genome and compared to 108 16S rDNA sequences available for Leptospira from the Ribosomal Database project (https://rdp.cme.msu.edu/). Phylogenetic analysis indicated that GWTS isolate #1 clustered among 4 strains of Leptospira recently isolated from soil samples in Fukuoka, Japan (designated as MS267, MS306, MS311, and MS316 respectively[48]), Fig 1 and S1 Fig. These, in turn, cluster most closely with Leptospira genomospecies 1, which has recently been renamed L. alstonii, and is comprised of two serovars of Leptospira that were originally isolated from frogs in China[49]; serogroup Ranarum serovar Pingchang and serogroup Undesignated serovar Sichuan. Similarly, the sequence for secY was extracted from the genome and phylogenetic analysis performed; the secY sequence of GWTS isolate #1 aligned most closely with that of L. alstonii serovar Pingchang and L. alstonii serovar Sichuan, Fig 2. However, rabbit antiserum specific for GWTS isolate #1, 2 or 3, failed to agglutinate with either of these two serovars representative of L. alstonii, Table 2. Nucleotide sequence for 16S rDNA and secY of GWTS #1 is provided (S2 Fig).
Restriction enzyme analysis was performed on DNA purified from each GWTS isolate #1, 2 & 3 for comparison with 5 of the 6 available isolates of L. alstonii that have been cultured to date, Fig 3. Results indicate that GWTS isolate #1 and #3 have an identical REA pattern that differed slightly from that of GWTS isolate #2. Results also indicate that the REA patterns are significantly different to that of any of the L. alstonii isolates. Analysis of REA patterns compared with a reference database of Leptospira strains held in the OIE Reference Laboratory (AFBI Stormont, Northern Ireland) did not identify any similar REA patterns.
Collectively, these results provide evidence of the unique and novel molecular attributes of each of the GWTS isolates, which we designate as L. alstonii serogroup Undesignated serovar Room22.
Leptospira alstonii is considered to be a member of the pathogenic complex of Leptospira, as defined by DNA-DNA relatedness, 16S rDNA and secY sequence. In addition to these criteria, the genome sequence of GWTS#1 contains lipL32, which to date has only been identified in pathogenic leptospires (S2 Fig). Each of the GWTS isolates was also tested for growth in the presence of 8-azagunaine; as with all pathogenic leptospires, none of the shrew isolates were able to grow in the presence of 8-azaguanine.
To further assess virulence properties of GWTS isolates, 3 groups of three hamsters were experimentally inoculated with 107 leptospires of GWTS isolate #1, #2 and #3 respectively. No hamster showed any sign of acute disease as determined by weight gain which remained comparable to non-infected controls at all times. All experimentally infected hamsters seroconverted, Table 3, as determined by a positive MAT titre on sera collected at 3 weeks post-inoculation. Sera from experimentally infected hamsters were only reactive with the challenge isolate; no cross-reacting MAT titres were detected when tested against an MAT panel representative for Ireland, and which included serogroup Bratislava, Canicola, Grippotyphosa, Hardjo, Icterohaemorrhagiae or Pomona. Kidneys from experimentally infected hamsters were culture negative for leptospires.
Serological evidence indicates that each of the GWTS isolates have uncharacterized antigens that fail to mediate agglutination, the basis of current standard typing and diagnostic methodologies. Since FAT is routinely used on infected host tissue to detect leptospires in situ by specialist laboratories, an FAT test was performed to determine reactivity with GWTS isolate #1, Fig 4. The positive result indicates that antibody prepared for the detection of leptospires by FAT is able to detect conserved antigens expressed by GWTS isolates.
This study demonstrates that an invasive mammalian species identified in Ireland is infected with a novel bacterial pathogen, designated L. alstonii serogroup Undesignated serovar Room22. This pathogen has not previously been identified in Ireland, or Europe, and never before been cultured from a mammalian host. Whilst there have been numerous accidental or deliberate introductions of mammalian and avian species into Europe[50], the GWTS population established in Ireland is most likely sourced from within Europe[3]. Regardless, invasive species have unique attributes to facilitate the dissemination of emerging infectious diseases[51]: firstly, invasive species may be more efficient at transmitting pathogens and, as in the case of our study, novel and as yet undescribed, pathogens. Secondly, invasive species tend to thrive in heavily anthropogenic habitats thus increasing the risk of transmission to humans. Thirdly, invasive species tend to have high dispersal rates as exemplified by the GWTS in Ireland with estimates of radial expansion rates of 5.5 km/yr[2]. Finally, invasive species facilitate the establishment of new emerging infectious diseases which are potentially zoonotic.
Leptospirosis is one of the most geographically widespread zoonotic diseases in the world[52]. Historically, all pathogenic leptospires were classified as Leptospira interrogans (sensu lato) which were subdivided into serovars, a division based on shared agglutinating lipopolysaccharide antigens and for which more than 200 serovars have been described[53, 54]. With the advent of genomics, pathogenic species of leptospires are now divided into 10 species, based on in silico hybridization of whole genome sequences, and include Leptospira alexanderi, L. alstonii, L. borgpetersenii, L. interrogans (sensu stricto), L. kirschneri, L. kmetyi, L. mayottensis, L. noguchii, L. santarosai and L. weilii [55–57]. However, the serologic and genomic based typing mechanisms are not mutually exclusive, as exemplified by serovar Hardjo, a significant pathogen in bovine populations throughout the world[58], which may belong to either L. interrogans or L. borgpetersenii. Nevertheless, the serologic classification of leptospires continues to play an important role in the epidemiology of leptospirosis and is the basis for the current “gold standard” serologic diagnostic assay, the microscopic agglutination test (MAT). In the MAT, serum from a patient (human or animal) is incubated with a panel of serovars of leptospires to test for a positive agglutination reaction, with the selected panel being representative of a geographical region; one of the obvious limitations of this assay is the composition of the diagnostic panel which will remain negative if tested with serum from a patient that is infected with a serovar not represented in the panel. Such is the case in our studies; when L. alstonii serovar Room22 was used to inoculate hamsters, all hamsters seroconverted and were MAT positive when tested against serovar Room22; but all were negative, with no cross-reactivity, when tested against six common pathogenic serovars, as typically found in Ireland. Nor was specific antiserum for L. alstonii serovar Room22 reactive with a range of pathogenic leptospires (Tables 1 and 2). Thus, prior to this study, no mammalian isolate of L. alstonii was ever available for serological diagnostics by MAT.
L. alstonii has been cultured from a mammalian host for the first time. Prior isolates of L. alstonii are derived from the amphibians Bombina orientalis and Rana nigromaculata, which belong to Neobatrachia species in China, or are derived from soil samples in Japan or Malaysia [48, 55, 59]. Whether L. alstonii serovar Room22 is pathogenic for domestic or wild animal species in Ireland or other parts of Europe and Northern Africa in which the GWTS exists, remains to be determined; such studies can now be facilitated, either by a comprehensive seroprevalence study by MAT, or culture, from other animal species. Alternatively, specialist Leptospira laboratories use fluorescent antibody testing (FAT) to detect leptospires in host infected tissue using polyclonal antibodies which cross reacts with L. alstonii serovar Room22 (Fig 4).
Our results suggest that the GWTS acts as a reservoir host for L. alstonii. Three isolates of Leptospira were identified, none of which had could be typed according to standard serological typing assays for Leptospira. Genome sequencing identified GWTS#1 as belonging to L. alstonii; restriction enzyme analysis (REA) confirmed that GWTS#3 has an identical pattern to that of GWTS#1, which differed slightly to that of GWTS#2. All REA patterns were different to that of other strains of L. alstonii cultured to date (Fig 3). Similarly, GWTS isolates have no agglutinating titres when tested against the reference strains of L. alstonii or conversely, when antisera specific for each of the GWTS isolates was test against more recently acquired strains of L. alstonii. In contrast to incidental hosts which typically suffer an acute limited disease that may include symptoms that range from a mild fever to more severe icteric disease with limited urinary excretion, reservoir hosts are asymptomatic, and may be MAT negative despite persistent renal colonization and excretion of leptospires via urine into the environment[60, 61]. Unique associations between specific host species and certain serovars of leptospires have been recognized; for example, Rattus norvegicus acts as a reservoir host for serovar Copenhageni and cattle are reservoir hosts for serovar Hardjo. Both serovar Copenhageni and serovar Hardjo can cause lethal infections in non-reservoir hosts. Whilst the GWTS likely acts as a reservoir host for L. alstonii serovar Room22, no evidence for acute or chronic disease was detected when serovar Room22 was used to experimentally infect hamsters. These results are similar to those previously described for soil isolates of L. alstonii in Japan and in which the authors concluded that such results likely reflect attenuation of strains due to continued maintenance under in vitro laboratory conditions[48]. Alternatively, a more appropriate animal model is required; in any case, culture of L. alstonii from the kidneys of the multiple GWTS confirms its pathogenicity. More recently, an in silico analysis of 102 isolates of Leptospira included the genomes of 3 strains of L. alstonii as originally isolated from amphibians in China[55]; results not only confirm that L. alstonii is a pathogen, but that the independent lineages of L. alstonii gained 504 genes (including three virulence genes) during its evolution, whilst no gene loss was observed. Such observations are interpreted to facilitate the adaptation by Leptospira to different hosts and an expanding range of environments.
The GWTS was originally identified in Ireland from skeletal remains in the pellets of barn owls (Tyto alba) and kestrels (Falco tinnunculus). Barn owls are susceptible to leptospirosis[62]. However it remains to be determined if birds of prey in Ireland are also infected with L. alstonii serovar Room22, or indeed if the decline of the native pygmy shrew in those areas inhabited by the GWTS is due in part to incidental infection with serovar Room22. There is little information available to assess the implications of the GWTS and associated pathogens on domestic animals and wildlife.
Our results raise additional questions yet to be answered; did the GWTS bring serovar Room22 to Ireland or did it acquire it in Ireland? There is no evidence of serovar Room22 in Ireland prior to capture of GWTS, but nor is there evidence of it in Western Europe or in Africa. Does serovar Room22 infect other domestic or other wild animal species? Up until now, this question could not be addressed by conventional serological surveys. The availability of an isolate of L. alstonii serovar Room22 from the current studies provides for an isolate to be included in conventional MAT panels, and for the preparation of specific antiserum that can be used in immunohistochemistry or FAT. Molecular assays are still applicable e.g. for the detection of lipL32, but such assays do not routinely type positive samples and still rely on a cultured isolate. This was the case in two recent surveys of the greater white-toothed shrew in Germany[63, 64]; in one study, 5 of 24 kidneys were PCR positive for lipL32[64]. Additional molecular typing suggested that kidneys were positive for L. kirschneri but results are not conclusive since the serovar was not identified. Culture was not attempted in either study.
The findings of the current study highlight the importance of screening wildlife for diseases. The current focus on wildlife health surveillance is primarily on human and livestock diseases that are outside the domestic and domiciled environments[65]. This emphasizes a lack of appreciation for the role that sylvatic ecosystems have in the development of zoonotic diseases[28, 66]. To carry out effective wildlife surveillance of emerging infectious diseases that are zoonotic or otherwise, there is a requirement to apply a systematic collaborative approach with veterinarians, ecologists, medical doctors, wildlife biologists, microbiologists and molecular biologists[67]. To date the surveillance of emerging diseases in wildlife is inherently passive[67]. There are clear conservation biology implications of this finding in conjunction with domestic animal health, and potentially human health. Globalization means there are likely to be more introductions of invasive species and therefore societies need to be in position to respond to the effect that these species and their associated pathogens and parasites have on ecosystems[51]. The current study demonstrates precisely what unwanted gifts an invasive species can bear but, to date, the exact consequences of such gifts have yet to be determined.
|
10.1371/journal.pcbi.1003375 | Detecting Genetic Association of Common Human Facial Morphological Variation Using High Density 3D Image Registration | Human facial morphology is a combination of many complex traits. Little is known about the genetic basis of common facial morphological variation. Existing association studies have largely used simple landmark-distances as surrogates for the complex morphological phenotypes of the face. However, this can result in decreased statistical power and unclear inference of shape changes. In this study, we applied a new image registration approach that automatically identified the salient landmarks and aligned the sample faces using high density pixel points. Based on this high density registration, three different phenotype data schemes were used to test the association between the common facial morphological variation and 10 candidate SNPs, and their performances were compared. The first scheme used traditional landmark-distances; the second relied on the geometric analysis of 15 landmarks and the third used geometric analysis of a dense registration of ∼30,000 3D points. We found that the two geometric approaches were highly consistent in their detection of morphological changes. The geometric method using dense registration further demonstrated superiority in the fine inference of shape changes and 3D face modeling. Several candidate SNPs showed potential associations with different facial features. In particular, one SNP, a known risk factor of non-syndromic cleft lips/palates, rs642961 in the IRF6 gene, was validated to strongly predict normal lip shape variation in female Han Chinese. This study further demonstrated that dense face registration may substantially improve the detection and characterization of genetic association in common facial variation.
| Heritability of human facial appearance is an intriguing question to the general public and researchers. Although it is known that some facial features are highly heritable, the exact genetic basis is unknown. Previous studies used simple linear measurements such as landmark distances, to evaluate the facial shape variation. Such approaches, although easy to carry out, may lack statistical power and miss complex morphological changes. In this study, we utilized a new 3D face registration method that enables subtle differences to be detected at high resolution 3D images. Based on this, we tried to test and characterize the associations of 10 candidate genetic variants to common facial morphological variations. Different types of phenotype data were extracted and compared in the association tests. Our results show that geometry based data performed better than simple distance based data. Furthermore, high density geometric data outstood the others in capturing small shape changes and modeling the 3D face visualization. Interestingly, a genetic variant from IRF6 gene, which is also a well-known risk factor of non-syndrome cleft lip, was found to strongly predispose the mouth shape in Han Chinese females.
| The human face plays an essential role in everyday life. It hosts the most important sensory organs and acts as the central interface for expression, appearance, communication and mutual identification. Inheritance of facial appearance from parents to their offspring is a constantly intriguing question to the public and scientific community. Indeed, human facial morphology is highly heritable. Twin studies have shown that heritability of facial features is as high as 80% [1], [2]. On the other hand, non-genetic factors also play important roles in shaping the human face, such as age and climate [2]–[6]. High heritability suggests that one's facial characters might be predicted to a certain extent, as long as the genetic determinants are identified and their effects fully understood. Face prediction based on genetic profiling, if feasible, may revolutionize forensics [7] and strongly benefit medical diagnosis [8]. However, the influences of common genetic variants on facial morphogenesis are largely unknown. The current understanding of facial morphogenesis has mainly arisen from developmental biology studies in model organisms. Facial morphogenesis involves a coordinated growth of facial prominences in a precise temporal and spatial sequence, which is tightly regulated by many signaling pathways, including the BMP, SHH, FGF, GHR and Wnt/β-catenin pathways [9]–[16].
In humans, knowledge of the effects of genetic variation on facial morphology has mainly arisen from studies of congenital craniofacial abnormalities. Non-syndromic cleft lip with or without cleft palate (NSCL/P) is the most common congenital craniofacial defect [3], [16], [17]. Great efforts have been made towards identifying the genetic factors that predispose carriers to NSCL/P, and a large number of candidate risk genes have been proposed [17]–[19]. Among these, the IRF6 gene has shown the most convincing and consistent signals for association across many studies [17], [20]–[24]. Many other craniofacial abnormalities can also result from rare genetic disorders, such as Down syndrome, Rubinstein-Taybi syndrome, Sotos syndrome, Bardet-Biedl syndrome and Noonan syndrome [25]–[29]. Nevertheless, these studies have mainly focused on pathological facial morphological changes.
Relatively few studies have attempted to associate genetic polymorphisms to common facial morphological variations. Several non-synonymous changes in the growth hormone receptor (GHR) were suggested to affect mandible shape in Japanese and Chinese populations [30]–[32]. Ermakov et al. found that a SNP in ENPP1, a gene essential in bone physiology, was significantly associated with upper facial height in Chuvashians [33]. In the FGFR1 gene, a genetic marker was found to be associated with the cephalic index in multiple populations [34]. Interestingly, a recent study examined several high frequency SNPs associated with differential risks of NSCL/P in a few healthy cohorts, and found that two were associated with normal facial shape variation [6]. This suggests that disease risk alleles may also modulate the phenotypes of unaffected carriers, although within a range of normal variation. Subtle shape alteration patterns induced by disease risk alleles, if properly defined, may help to screen carriers of disease alleles, and therefore facilitate disease prevention. In addition to these candidate gene studies, two genome wide association studies (GWAS) have also recently been carried out in Europeans, to search for genetic loci that influence common facial shape variation, and five loci were found to significantly modulate several nose related features [2], [35].
Anthropometric phenotypes, especially facial features, are highly complex and diverse. Traditional phenotype collection involves the manual measurement of specific distances and angles directly on the specimen or subjects, which is infamously tedious and error prone. In recent years, new imaging technologies, have been developed to allow fast and accurate acquisition of three dimensional facial landscapes without direct physical contact with the subject. Such imaging technologies have greatly facilitated human evolutionary analyses of craniofacial phenotypes [4], [5], [35], [36], as well as genetic association studies of human facial morphological variations [2], [6], [35]. However, the analysis post image acquisition still generally involves manual annotation of landmarks on digital images [4], [5], [35], [36]. More importantly, these inter-landmark distances were the most widely used phenotype measurements in the recent genetic studies of human facial morphology [2], [6], [33]–[35]. Inter-landmark based approaches have several problems. First, when pairwise distances are used as phenotypes, the number of phenotypes increases exponentially with that of landmarks, which often results in over conservative p values after multiple-testing correction. Second, the information on shape changes that is conveyed by inter-landmark distances is usually obscure. For example, an extended distance between the nasion and nose tip could signal either more pointed or overall bigger nose. Third, the facial shape cannot be fully reconstructed based on pairwise distances and it is, therefore, hard to perceive the biological meaning of the variation in distances. Thus, methods that directly examine the geometrical configuration of shapes are more desirable for general shape analyses. Such methods involve superimposing sample shapes according to their landmarks, followed by multivariate analyses/tests based on landmark coordinates [37]. More recently, new methods have also emerged to better use high resolution geometrical information. Instead of using only the limited number of traditional landmarks, these methods establish high density correspondence for thousands of mathematical landmarks [8][38][39]. Based on such methods, rare genetic diseases could be precisely identified and the syndrome effects could be extracted, predicted and visualized in great detail [40]–[42].
In this study, we first applied the method of high resolution 3D image registration to test the potential genetic associations of the complex normal facial variations, and to infer the detailed effects of genetic variants on face. In brief, we applied high density face registration (HDFR) to capture the comprehensive facial variation information of ∼30,000 3D points (referred to as marker points hereafter) [39]. Based on HDFR, three different schemes of phenotype representation were systematically compared for the detection of genetic associations with 10 candidate SNPs. The first scheme used traditional inter-landmark distances; the second represented the face geometrical shapes based on 15 major landmarks; the third is the high density geometric approach that we first proposed in such kind of studies. It uses the complete geometric data of over 30,000 marker points. The high density geometric data was then further used to examine the detailed phenotype changes associated with candidate SNPs.
We reviewed the literature for candidate SNPs that may be involved in the morphogenesis of the human face.10 SNPs from 4 genes, ENPP1, GHR, FGFR1 and IRF6 were identified and their functional relevance was listed (Table 1). The ectonucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1) gene is a key regulator of bone mineralization. Ermakov et al found that the upstream promoter and 3′ un-translated regions in this gene harbor genetic variants associated with the upper facial height and other phenotypes [33]. Four SNPs, rs7773292, rs6925433, rs6569759, rs7754561 that carry the strongest association signals were added to our candidate list. GHR is the growth hormone receptor, which plays essential role in the development. Mutations in this gene induce idiopathic short stature and Laron syndrome, marked by a characteristic facial appearance [31]. Several non-synonymous SNPs, including Pro561Thr (rs6184), I526L (rs6180) and C422F(rs6182) were suggested to contribute to mandibular measures in East Asian populations [15], [31], [32]. In this study, we included rs6180 and rs6184 in our study, as they were validated in Han Chinese [32]. FGFR1, the fibroblast growth factor receptor 1 plays an important role in facial morphogenesis, and mutations in this gene lead to syndomes associated with facial abnormality, such as the type 1 Pfeiffer syndrome (MIM 101600) and Kallmann syndrome 2 (KAL2) (MIM 147950) [34]. A tagging SNP of this gene, rs4647905 showed moderate signals of association with cephalic index in multiple ethnic groups [43]. We added another tagging SNP rs3213849 to span the full length of FGFR1. The Interferon regulatory factor 6 (IRF6) plays a critical role in keratinocyte development. Genetic variants of IRF6, especially SNP rs642961, were found consistently associated to NSCL/P throughout many candidate gene and GWAS studies [17], [23], [44], [45] As the genetic risk factors of NSCL/P may also contribute to normal facial variation in healthy carriers [16], we enrolled rs642961 into our study. We further included the SNP rs2236907 of IRF6, which seems to have a particularly strong signal in Han Chinese [23], [46].
The genetic effects of these SNPs were examined in a Han Chinese population from Taizhou, Jiangsu province on the east coast of China. The complete work flow is summarized in figure 1. In total 1001 self-reported Han Chinese individuals were enrolled in the analyses (604 females and 397 males), with an age range of 17∼25 years. Saliva was collected to obtain DNA. For the phenotype data, we collected high resolution 3D facial images for each individual. Facial images were automatically annotated with 15 salient landmarks (see Fig. 2 for the full list of the landmarks), using a novel landmark recognition method (see Methods) [39]. This was followed by HDFR that resulted in 32,251 mathematically derived marker points, which were corresponded one to one across all individuals (see Methods) [39]. The facial shape phenotypes were represented with three different schemes. In the first scheme, the Euclidean distances between pairs of the landmarks were taken as phenotypes, and hereafter collectively referred to as the landmark-distance (LMD) data. In the second scheme, the 15 landmarks of different individuals were first superimposed into a common coordinate system, by partial general procrustes analysis (PGPA) (see methods) [37]. PGPA removes the differences in location and rotation, while keeping the size and shape information. The coordinates of the aligned landmarks were thus used as the second type of phenotypes, hereafter referred to as landmark-geometric data (LMG). In the third scheme, all the 32,251 marker points were used to describe the phenotypes. The marker points were similarly superimposed onto a common 3D space using PGPA, and the coordinate vectors specified a phenotype data space of 32,251×3 = 96,753 dimensions. This data is hereafter referred to as dense-geometric (DG) data.
As sampling was carried out during winter time, many 3D images were affected by the participants' collared clothing, especially around the upper neck and lower jaw area. Furthermore, heavy facial hair in males caused defects and holes in some surface meshes. During quality control, the images with obvious caveats were removed from further analysis (105 individuals). 40 individuals were further removed due to the poor DNA quality (uv light absorption A260/280<1.6 or total DNA quantity lower than 300 ng). In the end, 856 individuals were successfully processed for their 3D images and have corresponding DNA. We carried out the genetic association study in two stages. The individuals of the original cohort were randomly assigned to 2 panels: panel I included 376 individuals (168 males and 208 females), and panel II included 480 individuals (174 males, 306 females). Tests were carried out separately for different genders. In stage I, all 10 candidate SNPs were genotyped for panel I. Then in stage II, the markers that showed preliminary evidence of correlation were validated using panel II. For stage I analysis, individuals were assigned into 3 possible genotype groups for each SNP. None of these SNPs deviated significantly from the Hardy-Weinberg equilibrium. For the LMD data, the landmark-distances were tested for association with SNP genotypes using the Tukey's honestly significant difference test (Tukey's HSD test). Of the total 105 possible pairwise distances, 6 departure from normal distribution according to Shapiro-Wilk normality test. As normality is required in Tukey's HSD test, these phenotypes were removed from further analysis. For the remaining 99 phenotypes, the raw p values were calculated and corrected for multiple-testing with 10,000 permutations (see Methods). Table 2 shows the summary of the preliminary association signals. Several SNPs demonstrated some preliminary association signals in terms of nominal Tukey test p value (p value<0.01) (Table 2). In particular SNPs rs642961and rs6184, showed enriched signals (Table 2). For SNP rs642961, many signals appeared in females between the mutant (TT) and the other two groups CC and CT. Interestingly, the strongest signals seemed to all point to the area around mouth and lower nose area. The distances between the right/left lip corners and the right/left alare (RLipCn – RAla and LLipCn – LAla) had nominal Tukey test p values between 0.002∼0.004 in both the CC/TT and CT/TT comparisons (Table 2). The distance between the upper lip point and lower lip point (ULipP-LLipP) also suggested potential shape difference between the CC and TT groups (nominal p value = 0.0023, Table 2). The suggestive involvement of this SNP with mouth shape is consistent with the known role the host gene IRF6 plays in NSCL/P [17], [23], [44], [45]. SNP rs6180 and rs6184 both showed some signals in males, which seemed to mainly involve the two lip corners and their relative positions to the middle line landmarks such as Pronasale, Nasion, Subnasale, lower lip point and chin (Table 2). These phenotypes may suggest size differences in the lower face among different genotype groups, but the overall trend is not clear. However, after the permutation correction of the multiple testing, none of these phenotypes stood significant at the individual SNP level, before accounting for multiple SNPs and different genders (Table 2).
For the LMG and DG data, we did the test for the whole geometric shapes, in a similar way to that previously described [37]. Briefly, the mean shapes were computed for each genotype group (see Methods), and the mutual distances between any two mean shapes were calculated. The mutual distance was calculated as the point-wise Procrustes distances (PPD), which is the Procrustes distance normalized by the number of landmarks/marker-points (see Methods). PPD distance can be directly compared between the LMG and DG data. The observed PPD distances were compared to 5000 random permutations to calculate empirical p values (see Methods). This procedure resulted in a single empirical p value for each comparison. The geometric permutation tests indicated that several SNPs had a nominal significance of association in stage I, and these signals were highly consistent between the LMG and DG data (Table 3, Table S1). To facilitate the visualization of the detailed differences, we also calculated the point-wise Euclidean distances between the mean shapes of the DG data, plotted as color gradients across the whole face (see Methods, Fig. S1). In gene IRF6, two SNPs rs2236907 and rs642961 exhibited moderate evidence of association. rs2236907 showed preliminary signals in both males and females (Table 3). However, a stronger association was found for rs642961 in females, where the CC and CT types both differ substantially from the TT type. The LMG tests had lower p values (nominal p = 0.005 and 0.006 for the CC/TT and CT/TT comparisons) than the DG tests (nominal p = 0.04 and 0.02 for the CC/TT and CT/TT comparisons) in both comparisons. Color gradient plots reveal that the major changes occur around the lips (Fig. S1A). The GHR SNP rs6184 showed some potential association between CC and AA in males (Table 3, Fig. S1 J). Two SNPs in the ENPP1 gene, rs6925433 and rs7773292 that were previously found to be associated with vertical upper face measurements in the Chuvashian population [33], also showed potential association signals in our data (Table 3). The preliminary signals were in males, although rs7773292 may be involved in forehead shape (Fig. S1B), whereas SNP rs6925433 may be related to the chin area (Fig. S1D). SNP rs7773292 had the second strongest association signal among all the 10 markers, with the corresponding nominal p values scoring 0.015 and 0.034 in LMG and DG data respectively (Table 3). The highly consistent pattern of p values between LMG and DG suggests that the 15 landmarks for the LMG data captured the total facial shape variation well. It is also worth noting that signals based on LMD data (rs642961, rs2236907 and rs6184) overlapped substantially with those from LMG and DG data, suggesting a general compatibility among the three different schemes. The signals from geometric tests (LMG, DG) were stronger than those of LMD, as their p values stood nominally significant at individual SNP level, whereas none of the LMD tests passed the single SNP significance level after permutation correction. Globally, none of the LMG/DG proved significant after Bonfferoni correction assuming 60 independent tests (3 genotypes and 2 genders ×10 SNPs).
Since the geometric tests gave obviously stronger association signals than the LMD tests, we chose the candidate SNPs based on the LMG/DG results for further re-validation. The two SNPs rs642961 and rs7773292, from genes IRF6 and ENPP1 respectively exhibited the most prominent signals in stage I tests, and were selected to be revalidated in sample panel II. The same tests as in stage I were carried out either solely with panel II or with the combined panel of I and II together. The LMD data showed strong associations between rs642961 and several distances involving mouth landmarks, e.g. LLipP, ULipP and Stm (Table S2). In particular in panel II, six pairwise distances, RAla-Stm, RAla- LLipP, LAla-LLipP, Stm-Sbn, ULipP-Sbn and LLipP-Sbn remained significant or marginally significant for the CC/TT and TT/CT comparisons (corrected significance level 0.01, Table S2). For the combined panel, the distance between LAla and LLipP gave corrected p values of 2.0×10−4 and 3.0×10−4 respectively for the CC/TT and CT/TT comparisons (Table S2). Association signals in rs642961 were much more significant when the tests were carried out using the geometric data (Table 4). In panel II alone, the females remained significant in the CC/TT comparison (corrected p values 0.022 and 0.011), and marginally significant in the CT/TT comparison (corrected p values 0.089 and 0.054) after correcting for 12 tests (2 SNPs×6 comparisons). The same 4 comparisons were more significant in the combined panel (corrected p values 0.001∼0.065) after correcting for all 60 possible tests with 10 SNPs (Table 4). The color gradient plots based on the dense geometric data in combined panel revealed substantial facial morphological differences between rs642961 TT and the other two genotypes (Fig. 3 A,C,E), which were also highly consistent with the patterns revealed in panel I (Fig. S1 A). These plots clearly show that the strongest changes occur around the mouth region. The comparison of the face profile lines revealed that the TT carriers on average had a slightly elevated forehead, as well as thicker and more protrusive (2–3 mm outwards) lips, than the other two genotypes (Fig. 3 B, D, F). However, the signals from rs7773292 completely disappeared in all stage II tests (Table S3), suggesting a possible false positive signal.
To investigate the mouth shape changes associated with SNP rs642961 in more details, we extracted the mouth DG data from the whole face by retrieving a defined set of marker points for the mouth. The 5 mouth landmarks (LLipCn, RLipCn, ULipP, Stm, LLipP) were also extracted to compose the mouth LMG data. The landmark-distance analyses were not repeated as they remained the same despite the extraction of the mouth data. Geometric permutation tests were conducted as before for the mouth LMG and DG data. In general, the results seemed to be much more significant than the corresponding whole face comparisons (Table 4). In panel II, the extreme nominal p value of 7×10−5 (corrected p = 0.00084) occurred between CC and TT in females in the LMG data. In the combined panel, the CC/TT comparison in females had the minimum p value of 1×10−5 (corrected p = 0.00012) in both the LMG and DG data. It should be noted that these p values for mouth region do not indicate any formal statistical significance as they were conditional on the prior information of the genetic association in mouth shape. Nonetheless, the extreme p values suggested there are substantial impacts of genetic variants on normal mouth shape variation. One potential problem that may affect the mouth shape analysis is the stomion point. Stomion is the central point between the upper and lower lips. None-neutral expressions or open mouth may induce altered distances between stomion and other mouth landmarks, therefore confound the association signals. Our image dataset has been carefully screened for such cases. In order to formally test the impacts, we removed stomion from the landmark set, and re-ran the image registration procedure and the LMG/DG analyses for SNP rs642961. As can be seen in Table S4, the results remained largely unchanged, indicating that our results were not confounded by stomion variation. Another potential confounding factor is age, as facial appearance changes during the time course of aging. We carried out formal tests to examine whether there were non-negligible age effects in our sample. As age 18 and 21 seemed to define the tails of the sample age distributions (Fig. S2), we grouped the individuals of 18 years or younger, and of 21 years or older, from the combined panel. The average shape difference was tested on the DG data using permutation (see methods). Neither test was significant (p value = 0.267 for female; and 0.576 for male). The same test was performed between other age groups, and also did not reveal any significant age/face interactions. This suggests that age has little impact to the overall analyses in this study.
The mouth shape changes among different genotypes seem to involve complex shape changes, thus we performed further high-dimensional data analyses to describe such changes. In the following analyses, we used the combined female panel unless otherwise specified. We first carried out principle component analysis (PCA, see Methods) on both the LMG and DG data. In the DG data, the first PC mode best distinguished the TT and CC genotypes (t-test nominal p = 1.3×10−6), and the TT/CT comparison was also highly significant (t-test nominal p = 1.8×10−6) on this PC. The first PC from the LMG data revealed similarly strong differences in the TT/CC (t-test nominal p = 2.7×10−6) and TT/CT (t-test nominal p = 2.2×10−6) comparison. The large differences between TT and the other two genotypes and the little difference between CC and CT suggested that this locus may follow a dominant model. To formally test this, we constructed an additive model and a dominant model based on the standard linear model (see Methods). The additive model did not suggest any statistical significance, whereas the dominant model was highly significant both with the LMG (nominal p = 1×10−6) and the DG data (nominal p = 6.8×10−6). Based on the dominant model, the genotypes of rs642961 explained a substantial proportion of the total variance (5.24% in the LMG data; 4.46% in the DG data) in PC1. Interestingly, when we tested these two models in a combined panel that included both males and females, the additive model remained insignificant, and the dominant model also became only marginally significant (nominal p = 0.003 in the LMG data; nominal p = 0.0159 in the DG data). This suggests that the effect of TT is female specific. To extract the facial pattern that best distinguishes TT from the other genotypes, we further carried out a simple linear discriminant analysis (LDA). As a hyperline that transects the mean points of TT and CC groups would best separate these two groups, this line was defined as a new data axis onto which individual data points were projected to generate hyperline (HL) scores. The HL scores were plotted against the PC1 scores to visualize data distribution (Fig. 4). As can be seen from Fig. 4, the distributions on PC and HL are highly correlated (r2 = 0.97). The TT distribution differed substantially from that of CC and CT. Specifically, the average PC1 score of 0 found 18 of the 19 TT individuals at the minus side; similarly, the average HL score of 0.444 had 18 out of 19 TT individuals at the minus side. To visualize the mouth shape changes, we transformed the mean shape (Fig. 4B) by adding or subtracting 3 standard deviations along either dimension as: st = sm±3σvv, where st was the transformed shape, sm the average shape, v the Eigen vector of the dimension and σv the standard deviation. The resulting shapes were defined as PC1+, PC1−, HL+ and HL− respectively in Fig. 4. The PC1− shape (Fig. 4A), which represents the trend for TT, has more protrusive and thicker lips compared to the finer and thinner lips in the PC1+ shape (Fig. 4E). The whole mouth region of PC1− is also more prominent and bigger than that of PC1+. Consistent with the high correlation between HL and PC1, the face models along the HL dimension reveal similar shape changes. (Fig. 4).
To the authors' knowledge, this study is the first to use high resolution face image registration to test the genetic association for common facial variation. Human face is a highly complex geometric surface. The simple inter-landmark distances used in previous studies may have over-simplified the common variation of human faces. As the high throughput acquisition of high content 3D image data becomes easier, methods based on shape geometric information, especially of high definition, become increasingly necessary to enable comprehensive and fully quantitative analyses of the complex facial traits. Based on high density 3D face registration, we compared three different schemes of phenotype during tests of genetic association, including LMD, LMG and the high resolution geometric data DG. We found that, in general, the three schemes produced consistent signals for the candidate SNPs. In the stage I test, the LMD method had only moderate association signals, mainly due to the large number of tests. The 15 landmarks gave rise to 105 possible tests in each genotype comparison (Table 2). One strategy to reduce the number of tests is to use only the essential distances, e.g. the conventional craniometrical measures that correspond to obvious anatomical structures. However, this risks missing the strongest signals. The other major problem with distance data is the difficulty in perceiving the underlying shape changes. For example in stage I, SNP rs642961 did not show a clear involvement with mouth shape changes in the LMD tests (Table 2). However, such an involvement was already quite clear on the DG comparison in stage I (Fig. S1). The LMD method seemed to improve both in the test power as well as the inference of shape changes (most significant landmark-distances involved the mouth landmarks) when larger sample sizes were used in stage II tests.
The two geometric schemes were generally found to give stronger association signals, implying better statistical power for the geometric methods. This may be due partially to the fact that the geometric tests were carried out in one step, which avoided a complex test structure. Interestingly, the LMG data of only 15 landmarks showed highly consistent test signals with that based on DG data. This suggests that these 15 landmarks capture the majority of the normal facial morphological variation. When only shape difference is to be tested, the LMG method seems to provide better efficiency (given the smaller data involved) and potentially higher test power. However, the strong consistency between LMG and DG in the association signals attributed to rs642961 may be partially accounted for by the high landmark density around the mouth area (5 out of 15 chosen landmarks). Features with fewer landmarks would confer lower power in the LMG tests. On the other hand, the DG data has other unique advantages for shape change inference and modeling. We also show here that the point-wise distance distribution between the mean faces can highlight the areas of shape changes in high definition (Fig. 3), which can guide future in depth exploration. Furthermore, the effects of potential genetic factors may also be modeled visually as realistic 3D face images (Fig. 4). This may have hugely beneficial applications to forensic studies.
Variants in the IRF6 gene have been found to predispose to the risk of NSCL/P [21]–[23], [47]. Nevertheless, a link between the IRF6 gene and common facial variation has not been established. This is the first study that provides strong evidence that rs642961 also affects normal facial shape variation. In particular, TT individuals may have more protrusive and thicker lips (Fig. 4). Interestingly, such an effect is very likely female specific as the tests in males did not yield significant signals. Combination of both sexes in the dominant model test also suggested that males did not contribute to the association signals. This is not uncommon. For example, various types of NSCL/P have been found to have sex specific spectra, suggesting sex is an important epistatic factor in mouth morphogenesis [16], [48]. In females, the TT individuals showed a highly specific distribution on the plane defined by PC1 and hyperline (Fig. 4). This could be used during diagnosis to pre-screen the risk allele carriers by interpreting 3D pictures, therefore facilitating early prevention of NSCL/P.
We have also detected preliminary associations for other SNPs. Failure to validate these association signals does not exclude them from the candidate list of loci related to normal facial shape variation. Extended sample sizes as well as inclusion of samples from other populations will be needed to further increase our understanding of the genetics of human facial morphology.
Sample collection in this study was carried out with the approval of the ethics committee of the Shanghai Institutes for Biological Science and in accordance with the standards of the Declaration of Helsinki. Written informed consent was obtained from every participant.
In total 1001 combined individuals (604 females, 397 males) from self-reported Han Chinese were sampled from Taizhou, Jiangsu province. Age ranges were between 17 and 25 years. 2∼3 ml of saliva was collected from each participant for DNA extraction. Individuals with obvious health problems or any history of facial surgery were excluded from the study.
Genomic DNA was extracted from saliva following a modified Phenol–chloroform protocol [49], then suspended in Tris-ethylenediaminetetraacetic acid (TE) buffer (0.01 m TrisHCl, 0.001 m EDTA, pH 8.0) and stored at −20°C. SNP genotyping was performed with the SNaPshot multiplex system on an ABI3130xl genetic analyzer by Genesky Biotech, Shanghai, China.
The 3dMDface system (www.3dmd.com/3dMDface) was used to collect high-resolution 3D facial images from participants. This system captures 3D facial images at a speed of ∼1.5 milliseconds and a geometry accuracy of 0.2 mm RMS.
We applied a new approach to achieve high density point-wise registration across all 3D facial images [39]. In brief, 17 salient facial landmarks were first automatically annotated, based on the PCA projection of both texture and shape information. In this study, 15 out of the 17 landmarks were used in analysis (Fig. 2). Two earlobe points were excluded as many 3D images, mainly of female participants were missing parts of the ears due to the unbound long hair. Afterwards, a facial image of high quality and smooth shape surface was chosen as the reference, and its surface mesh was re-sampled for an even point distribution at a density of 1pixel per mm2. The reference face was then warped to register every sample face by matching all the 15 landmarks, via a non-rigid thin-plate spline (TPS) transformation. The mesh points of the reference face were then projected to the sample surface to find their one-to-one correspondents. The projection points were then used to re-define the mesh of the sample facial surface [39]. As the same reference face was always used, the re-defined 3D point sets in the sample faces were also point-wisely corresponded across all samples. The non-rigid registration guided by the 15 landmarks also ensured that the point-wise correspondence was approximately anatomically homologous. Each sample face was represented by a set of 32,251 3D points, with their coordinate values stored in a 3×32,251 matrix. Generalized Procrustes analysis (GPA) was used to align the sample facial shapes into a common coordinate system. The details of the dense correspondence registration approach are described elsewhere [39].
Assuming each shape is represented as a vector: s = [x1, y1, z1, x2, y2, z2 … xn, yn, zn], where xi, yi, zi stand for the x, y, z coordinates of the ith point. There are n points in total. For two shapes s and s′, the squared Euclidean distance for the ith point is,and the PPD is defined as:
For the LMD data, in order to correct for the large number of sub-tests within each SNP, we performed a permutation procedure. For each of the 99 traits, raw p values were first calculated with Tukey's HSD test. A permutation procedure was used to correct the raw p values for multiple-testing. Briefly, the genotypes were reshuffled among the participants for 10,000 times and the Tukey's test was similarly carried out. The lowest p value from each permutation was combined to derive a null distribution. The empirical raw p values were then ranked against the null distribution to give the corrected permutation p values [50].
For the LMG and DG data, genotypes were randomly reshuffled among the individuals, and the PPD distances were calculated for the permutated genotype groups. 5000 permutations were carried out in stage I analyses and 100000 permutations in stage II analyses due to the much more significant P values. The PPD distribution under permutation was compared to the observed PPD value. The proportion of the permutation sets that had PPD values smaller than or equal to the observed PPD was taken as the nominal one-sided p value.
The prcomp function in the R statistics package was used for PCA analysis. An un-scaled PCA analysis was carried out, assuming equal variance for all points.
We established both the dominant model and the additive model based on the standard linear model. The additive model was implemented by encoding genotypes as 0, 1 and 2. The dominant model was built by assuming CC and CT as 0 and TT as 1. Model test and analyses were conducted with the R statistics package.
|
10.1371/journal.pntd.0006987 | Antibody responses to Plasmodium vivax Duffy binding and Erythrocyte binding proteins predict risk of infection and are associated with protection from clinical Malaria | The Plasmodium vivax Duffy Binding Protein (PvDBP) is a key target of naturally acquired immunity. However, region II of PvDBP, which contains the receptor-binding site, is highly polymorphic. The natural acquisition of antibodies to different variants of PvDBP region II (PvDBPII), including the AH, O, P and Sal1 alleles, the central region III-V (PvDBPIII-V), and P. vivax Erythrocyte Binding Protein region II (PvEBPII) and their associations with risk of clinical P. vivax malaria are not well understood.
Total IgG and IgG subclasses 1, 2, and 3 that recognize four alleles of PvDBPII (AH, O, P, and Sal1), PvDBPIII-V and PvEBPII were measured in samples collected from a cohort of 1 to 3 year old Papua New Guinean (PNG) children living in a highly endemic area of PNG. The levels of binding inhibitory antibodies (BIAbs) to PvDBPII (AH, O, and Sal1) were also tested in a subset of children. The association of presence of IgG with age, cumulative exposure (measured as the product of age and malaria infections during follow-up) and prospective risk of clinical malaria were evaluated.
The increase in antigen-specific total IgG, IgG1, and IgG3 with age and cumulative exposure was only observed for PvDBPII AH and PvEBPII. High levels of total IgG and predominant subclass IgG3 specific for PvDBPII AH were associated with decreased incidence of clinical P. vivax episodes (aIRR = 0.56–0.68, P≤0.001–0.021). High levels of total IgG and IgG1 to PvEBPII correlated strongly with protection against clinical vivax malaria compared with IgGs against all PvDBPII variants (aIRR = 0.38, P<0.001). Antibodies to PvDBPII AH and PvEBPII showed evidence of an additive effect, with a joint protective association of 70%.
Antibodies to the key parasite invasion ligands PvDBPII and PvEBPII are good correlates of protection against P. vivax malaria in PNG. This further strengthens the rationale for inclusion of PvDBPII in a recombinant subunit vaccine for P. vivax malaria and highlights the need for further functional studies to determine the potential of PvEBPII as a component of a subunit vaccine for P. vivax malaria.
| Plasmodium vivax is responsible for most malaria infections outside Africa, with 13.8 million vivax malaria cases reported annually worldwide. Antibodies are a key component of the host response to P. vivax infection, and their study can assist in identifying suitable vaccine candidates and serological biomarkers for malaria surveillance. The binding of P. vivax Duffy binding protein region II (PvDBPII) to the Duffy Antigen Receptor for Chemokines (DARC) is critical for P. vivax invasion of reticulocytes. Although the binding residues for DARC are highly conserved across PvDBPII, the parasite displays high sequence diversity in non-binding residues of PvDBPII. Other regions such as PvDBPIII-V are relatively conserved. Recently, sequencing of P. vivax field isolates, identified a homologous erythrocyte-binding protein (PvEBP), which harbors a domain, region II (PvEBPII), that is homologous to PvDBPII. To date, there has been limited investigation into the naturally acquired immunity to both PvDBPIII-V and PvEBPII in human populations. Using a longitudinal cohort study, we have characterized the serological response to PvDBPII, PvDBPIII-V, and PvEBPII among 1–3 years old PNG children and investigated associations with protection against clinical malaria. This study shows that both total IgG and IgG3 to the predominant PvDBPII AH allele in PNG, and total IgG and IgG1 to PvEBPII were associated with protection from P. vivax malaria.
| Plasmodium vivax, which is the most widely distributed plasmodium species that infects humans [1], is considered the key challenge to malaria elimination efforts outside Africa. This is largely due to the ability of P. vivax to relapse from dormant stages in the liver [2]. These liver hypnozoites are undetectable with current diagnostic tools and treatment, which is currently limited to 8-aminoquinolines and cannot be safely prescribed to G6PD-deficient individuals [3]. Thus, additional tools to target P. vivax are urgently needed [4]. Vaccines could play an important role in the elimination of this parasite. However, as primary infections are likely to cause most of the clinical episodes and to contribute proportionately more to onward transmission than a single infection [5], it will be essential to incorporate blood-stage antigens in a candidate vaccine to reduce blood stage parasitemia and gametocytemia in breakthrough infections.
P. vivax preferentially invades young red blood cells called reticulocytes. Invasion into these cells relies on the interaction between parasite proteins and reticulocyte receptors. A well-characterized ligand-receptor pair involved in invasion is the interaction of P. vivax Duffy binding protein (PvDBP) with the Duffy Antigen Receptor for Chemokines (DARC) [6]. The virtual absence of P. vivax malaria in West Africa, where populations are generally DARC negative, highlights the central role of this pathway in P. vivax infection [7]. PvDBP is a 140 kDa type 1 integral membrane protein that consists of seven regions: a leader peptide sequence and N-terminal region (region I), conserved cysteine-rich regions (regions II and VI), central uncharacterized regions (regions III to V), and a transmembrane region followed by a cytoplasmic domain. Region II (PvDBPII) contains three-subdomain (SD) protein, with SD2 contributing key residues for binding to DARC on red blood cells (RBCs) [8, 9]. It has been proposed that binding of PvDBPII with its receptor may lead to dimerization of PvDBPII [10]. No functional role has yet been identified for central regions III to V of PvDBP. However, as they are conserved among isolates from various geographical regions [11], the potential of antibodies against these regions as correlates of protective immunity against vivax malaria deserves investigation.
In highly endemic areas of PNG, naturally acquired immunity against P. vivax controls parasite densities leading to reduced risk of clinical disease in the second and third year of life [12]. Immune responses to PvDBPII increase with age, suggesting they may play an important role in acquired immunity [13]. Strong naturally acquired humoral immunity to PvDBPII has been associated with reduced risk of high-density parasitemia in PNG children [14]. In addition, anti-PvDBPII antibodies purified from plasma from PNG individuals with the ability of blocking P. vivax invasion of reticulocytes provide the rationale of PvDBP as promising vaccine candidate [15]. However, one of concerns related to the development of PvDBP as a vaccine candidate is that sequence diversity of PvDBPII may allow the evasion of human immune responses [16, 17]. In regions of PNG with high P. vivax endemicity, the most predominant allele of PvDBPII is AH with a proportion of 26% in circulating strains [14]. In parallel to PvDBPII polymorphisms, antibody responses to PvDBPII showed strain-specific immunity to the P. vivax strains circulating in PNG [14]. When antisera were tested for the presence of functional antibodies that block PvDBPII-DARC interaction, it was found that a small proportion of individuals (<10%) were able to make high levels of binding inhibitory antibodies (BIAbs) that blocked binding of diverse strains [18]. The presence of such high levels of BIAbs was associated with protection against P. vivax infection and reduced parasite densities [18, 19]. Although PvDBPII has significant polymorphisms, the binding residues for DARC are highly conserved, which makes the development of strain-transcending BIAbs possible [18]. The reason why the development of such BIAbs is not common remains to be understood. Here, we used a functional binding inhibition assay to investigate the presence and association of anti-PvDBPII BIAbs and protection against clinical P. vivax malaria in young children in PNG of 1–3 years age.
Whole genome sequencing of Cambodian field isolates identified a second putative erythrocyte binding protein (PvEBP) with all the features of a Plasmodium erythrocyte-binding protein, including a N-terminal signal peptide, a Duffy-binding like domain (Region II, PvEBPII), a C-terminal cysteine-rich domain, and a transmembrane domain [20]. Although harboring all the characteristics typical of DBP superfamily member, PvEBPII seems to be distant from PvDBP in phylogeny, and no inhibition of PvEBPII-reticulocyte binding was observed by using mouse anti-PvDBPII IgG [21]. The genetic distance of PvEBPII and PvDBPII indicates that PvEBPII is not a recent gene duplication, and its apparently lower proportion of single nucleotide polymorphisms suggests that it is unlikely to be under the same level of immune selection as PvDBP [21]. In a recent screen of 38 P. vivax antigens in plasma from naturally exposed children in PNG, antibodies to both PvDBPII and PvEBPII were frequently identified among five-antigen combinations with the strongest protective effects against clinical malaria [22]. A more in-depth evaluation of the functional importance of antibody responses to variants of PvDBPII, PvDBPIII-V and PvEBPII is thus warranted.
The IgG isotype determines antibody function, and in humans, cytophilic IgG1 and IgG3 are important mediators of pathogen clearance. Numerous studies have reported that IgG subclass profiles differ among antibodies targeting different P. falciparum antigens [23–27]. The properties of the antigen appear to be one of the main determinants of the type of IgG subclass generated [25]. In addition, for some Plasmodium antigens, a switch from a predominant IgG1 response in young children to an increase or even predominance of IgG3 response in older individuals is a characteristic feature of natural acquisition of clinical immunity to malaria [28–30]. It remains to be confirmed if this switch is due to a history of increased exposure and/or the maturing of the immune system. Elucidating the subclasses of IgG against different PvDBPII variants and PvEBPII, and their association with clinical diseases may help better understand the importance of development of IgG subclass immunity for protection against malaria.
Plasmodium infection is considered to be one of the key driving forces of the evolution of the human genome. Polymorphisms in RBC proteins are particularly common in malaria endemic regions [31–33]. Gerbich deficiency is associated with the deletion of exon 3 in the glycophorin C gene (GYPCΔex3) [34]. Gerbich-negative erythrocytes were first identified in 1960 but are of low prevalence globally [35]. However, in some Melanesian populations from PNG, 50% of them have inherited the Gerbich phenotype [36]. Few studies have shown consistent associations between the Gerbich phenotype and Plasmodium infection [37–39], with one study observing a lower prevalence of P. falciparum infection among the population with this phenotype [37]. Its potential association with protection against P. vivax and its relationship with the acquisition of immunity remains unknown.
In this study, parameters of naturally acquired immunity to four variants of PvDBP (several alleles of PvDBPII and PvDBPIII-V), as well as PvEBPII were characterized in PNG children of 1–3 years of age from a 16-month longitudinal cohort study. In addition, levels of total IgG to PvDBPII, IgG subclass, and presence of anti-PvDBPII BIAbs were measured and their association with P. vivax infection, clinical episodes, and Gerbich negativity was explored.
Plasma samples used in the current study were collected as part of a longitudinal cohort study of young PNG children (1 to 3 years old) previously described [12]. In brief, participants were followed for up to 16 months, with visits twice/month for symptomatic illness and infection status as detected by microscopy and PCR. All P. vivax infections were genotyped, allowing for the calculation of the incidence of genetically distinct blood-stage infections acquired during follow-up (i.e. the molecular force of blood-stage infections, molFOB) [40]. Host genotyping for the presence of the GYPCΔex3 deletion associated with the Gerbich blood group was done by PCR, as previously described [41]. Children who were homozygous for the GYPCΔex3 deletion were considered to be Gerbich negative. Plasma samples collected at the start and at the end of the study from 224 children were used in the present study.
The four PvDBP variants used in this study were binding domain II from strains AH, O, P and Sal1 [14]. The recombinant PvDBPII variants were expressed in E. coli. Proteins were solubilized from inclusion bodies, purified by affinity chromatography, followed by refolding and ion exchange chromatography as per methods described earlier [42]. An 1176 bp fragment corresponding to the PvDBPIII-V region (aa 508–899) from Sal1 reference sequence was codon optimized for expression in E. coli. The protein was purified by metal affinity chromatography. Recombinant PvEBPII (aa 161–641) with a C-terminal 6-His tag was expressed as a soluble protein in E. coli SHuffle cells. Following cell lysis, the recombinant PvEBPII was purified from cell lysate as a soluble protein by metal affinity chromatography using standard procedures.
Recombinant PvDBPII, PvDBPIII-V, and PvEBPII fragments were conjugated to Luminex Microplex microspheres as previously described [43]. To conjugate proteins to 2.5x106 beads, we used 0.300 μg/mL of PvDBPII AH, 0.125 μg/mL of PvDBPII O, 0.094 μg/ mL of PvDBPII P, 0.225 μg/mL of PvDBPII Sal, 0.031 μg/mL of PvDBPIII-V and 0.250 μg/mL of PvEBPII. The Luminex multiplex bead-based antibody detection assay was performed as described elsewhere with the following modifications [29, 43]. Plasma samples were diluted 1:100 in PBS with 1% BSA and 0.05% Tween (PBT). Diluted samples were incubated with a mix of antigen-conjugated beads (0.1 uL of each bead position) (1:2) for 30 minutes under constant agitation. PE-conjugated donkey anti-human IgG Fc (0.1 mg/mL, Jackson ImmunoResearch) was used as a secondary antibody. IgG subclasses were detected using the following antibodies: mouse anti-human IgG1 hinge-PE (0.1 mg/mL, clone 4E3, Southern Biotech); mouse anti-human IgG2 Fc-PE (0.1mg/mL, clone HP6002, Southern Biotech); mouse anti-human IgG3 hinge-PE (0.1 mg/mL clone HP6050, Southern Biotech); or mouse anti-human IgG4 Fc-PE (0.1 mg/mL, clone HP6025, Southern Biotech). All these antibodies were diluted 1:100 in PBS to detect total IgG, IgG1, IgG2, IgG3, and IgG4 respectively. Beads were read on a Bio-Plex 200 reader set for 75 beads per analyte. Results were reported as median fluorescence intensity (MFI). One blank well without plasma was used for determination of the true fluorescence background. Positive controls consisted of pooled serum from immune PNG adults (>18 years) from the Madang (n = 10) and East Sepic Provinces (n = 10) who were highly exposed to malaria. Such positive controls were included in ten two-fold serial dilutions (1:50–1:25600) as pervious described [29]. Negative control sera in all assays were from the same individual, which was from the Australia Red Cross donor. The donor was anonymous resident of Melbourne, Australia with no known previous exposure to malaria.
An ELISA plate-based semi-quantitative binding assay was used to test binding of PvDBPII with DARC and estimate the binding inhibitory activity of serum as described earlier [44]. Briefly, the N-terminal 60 amino acid extracellular region of DARC was expressed as a fusion with Fc region of human IgG (nDARC-Fc), purified using protein A column and used to coat ELISA plates. Recombinant PvDBPII was incubated with nDARC-Fc coated plates in the presence of different concentrations of anti-PvDBPII serum or purified anti-PvDBPII IgG. Bound PvDBPII was detected with anti-PvDBPII rabbit sera followed by anti-rabbit IgG horse radish peroxidase (HRP)-conjugated goat antibodies. Percent binding inhibition was determined at different serum or IgG concentrations using a standard curve as previously described [44].
Standard curves from each Luminex assay plate were used for transformation of MFIs into relative antibody units (expressed as dilution factors that range from 1.95 x 10−5 or 1/51200 to 0.02 or 1/50) using a five parametric logistic regression model as described previously [29]. Statistical analyses were performed using STATA version 12 (StataCorp) and R version 3.2.1 (htpp://cran.r-project.org). Spearman's rank correlation was used to assess the associations between antibody levels and age, and correlations among antibody responses against different antigens. Differences in antibody reactivity between categorical variables were assessed using Wilcoxon signed-rank sum test (for two groups) and Kruskal Wallis test (for multiple groups). Differences in proportions were evaluated by chi-square test. Antibody responses were used to predict molecular force of blood stage infection (molFOB) using a general linear model (GLM) stratified by concurrent infection status at the last visit of the study. Concurrent infection was defined as positive if PCR test was positive at the time of antibody measurement (i.e. enrollment). Antibody levels were stratified into tertiles to analyze the relationship with prospective risk of clinical P. vivax episodes (defined as axillary temperature > 37.5°C or history of fever in the preceding 48 hours with a concurrent parasitemia > 500 P. vivax /μl) and prevalence of infection diagnosed by PCR and light microscopy over the 16 months of follow-up [12]. Generalized estimating equation (GEE) with exchangeable correlation structure and semi-robust variance sandwich estimator were used and analyses were done by comparing the incidence rate ratio (IRR) of clinical malaria between the highest and lowest tertiles, and medium and low antibody levels groups. Differences in geometric mean parasitemia and incidence of clinical episodes among GYPCΔex3 genotypes were analyzed using GEE.
To examine the effect of combining antibody responses to different antigens on the risk of clinical disease, we examined all possible combinations of 2 and 3 antigens. For this, IgG responses for each antigen were assigned a score starting from 0 to 3 for low, medium or high antibody levels (i.e. quartiles). These scores were then added up for each different combination. The scores of any combination were equally divided into three groups and used in our GEE model.
All datasets were available in the Dryad repository: https://doi.org/10.5061/dryad.n14p52b [45].
Ethics clearance was obtained from the PNG Medical Research and Advisory Committee (MRAC 05.19) and the Walter and Eliza Hall Institute (HREC 07/07) for the use of field samples. All parents/guardians of the participants signed a consent form prior to enrollment. The Melbourne control sample was obtained under ethics approval HREC 13/07.
The pooled serum from immune PNG adults was assumed to represent the equilibrium antibody levels to all antigens achieved following repeated natural exposure. Here, we determined at enrollment the number of children who had already acquired IgG levels equivalent to >50%, >25%, >10%, >5%, or >1% of the IgG levels in adults (Table 1). Plasma from the PNG children were reactive to all four PvDBPII alleles and PvEBPII. However, total IgG levels were relatively low for the most common PvDBPII PNG variant AH, with only 25.9% and 3.6% of children achieving >5% and >25% of hyper-immune adult levels respectively. Immunogenicity of other PvDBPII alleles and PvEBPII were similar (range: 14.3–20.1%, >5% of hyper-immune adult levels) (Table 1).
Total IgG levels were strongly correlated between all proteins measured at the beginning of the cohort study (rho = 0.67–0.98, P<0.001), with the strongest correlation found between PvDBPII AH and PvDBPII O variants. Total IgG to PvEBPII shows weak to moderate correlation with different PvDBP antigens (Spearman’s rho = 0.22–0.63) (Fig 1). Similar correlation patterns were observed in the plasma samples collected at the last time point of the longitudinal study (rho = 0.22–0.99, P≤0.001) (S1 Fig).
The prevalence of P. vivax infection was 55.4% (124/224) among young PNG children at enrolment, as determined by PCR (Fig 2A). Individuals with a concurrent P. vivax infection by PCR had significantly higher IgG levels (P<0.015) to all variants of PvDBPII, PvDBPIII-V and PvEBPII (Fig 2A and S1 Table), indicating that even asymptomatic P. vivax infections may boost immune responses to PvDBP and PvEBPII in settings of high P. vivax endemicity.
We examined the relationship between total IgG antibody levels with age and cumulative exposure, which was defined as the product of age and the corresponding individual molFOB [40]. Collectively, both categorical and continuous measures of antibody levels were positively correlated with age and cumulative exposure for PvDBPII AH and PvEBPII (P<0.001–0.047), and PvDBPIII-V with a borderline significance (P = 0.051), but only in children free of P. vivax infection at enrolment (Fig 2B and S1 Table). This might reflect that the acquisition of clinical immunity in this cohort of young children was mainly driven by individual exposure heterogeneity [40]. To further understand the associations of antibody responses with prevalence of infection in the longitudinal cohort study, antibody responses to PvDBPII AH and PvEBPII were found associated with increased risk of infection detected by PCR for the study period (S2 Table).
In addition, higher antibody levels were observed in the last visit for PvDBPII AH, PvDBPIII-V, and PvEBPII compared to their levels at enrolment (S2 Fig). IgG antibodies to PvDBPIII-V and PvEBPII were indeed significantly correlated with molFOB at the last visit (rho = 0.15, P≤0.020), with a borderline significance for PvDBPII AH (rho = 0.12, P = 0.070) (S2 Fig).
In PNG adults, IgG1 was the predominant antibody subclass for all PvDBPII proteins (S3 Fig). IgG2 and IgG3 were the subdominant antibody subclasses, with substantially higher amounts of IgG3 detected against the most common PvDBPII AH variant than the other strains. Balanced responses with detectable amounts of IgG1 and IgG3 antibodies were found for PvDBPIII-V and PvEBPII. No detectable levels of IgG4 were observed for any of the antigens tested (S3 Fig).
Between 27.7–75.0% of the children had levels of IgG1 against PvDBPII and PvEBPII that were >5% of the IgG1 levels observed in adults to all antigens tested. However, only a small subset of children (range: 3.6–11.2%) had IgG1 levels exceeding 25% of adult levels (Table 1). In contrast to what was observed in adults, children showed a strong IgG1 predominance among antibodies to PvEBPII. Detectable levels of IgG3 were observed for PvDBPII AH and PvEBPII, both in much lower levels than IgG1, suggesting that acquisition of IgG1 was faster than IgG3 for both PvEBPII and PvDBP antigens. Polarization from IgG1 towards IgG3 was only identified for PvDBPII AH, as suggested by the decreasing ratio of IgG1/IgG3 with increase in age (rho = -0.24, P<0.001). No detectable levels of IgG2 and IgG4 were observed for any of the antigens tested among these children.
PCR-positive children had increased IgG1 for PvDBPII AH, PvDBPII O, and PvEBPII (S1 Table). Consistent with the patterns observed for total IgG, significant increase in IgG1 with age and cumulative exposure were identified for PvDBPII AH and PvEBPII in PCR negative children (rho = 0.25–0.42, P≤0.013) but not in those with concurrent infections (S1 Table). Increases in IgG3 against PvDBPII AH and PvEBPII were significantly associated with age and cumulative exposure (rho = 0.19–0.39, P≤0.044) (S1 Table).
Children with medium and high levels of IgG to PvDBPII AH allele had 31% and 44% reduction in the risk of a P. vivax clinical episodes, respectively, compared to children with low antibody levels (adjusted incidence risk ratio medium versus low antibody levels (aIRRM 0.69, 95% CI: 0.49–0.98, P = 0.037); high versus low antibody levels (aIRRH 0.56, 95% CI, 0.40–0.77, P <0.001) (Fig 3 and S3 Table). IgG to PvDBPII O showed similar, but slightly lower significant protective association (Fig 3 and S3 Table). Therefore, total IgG responses to PvDBPII AH and PvDBPII O may be biomarkers of protective immunity. Only children with high IgG antibody levels to PvDBPII P, PvDBPII Sal1 and PvDBPIII-V had a significant reduction in the risk of P. vivax malaria. Antibodies to PvEBPII correlated with stronger protection than all variants of PvDBPII (aIRRM = 0.73, P = 0.022; aIRRH = 0.26, P<0.001, Fig 3 and S3 Table) and in a multivariate model, only antibodies to PvEBPII remained independently associated with protection against clinical malaria (S3 Table).
We further examined the possible effect of combining antibody levels against PvDBPs and PvEBPII on the risk of clinical disease. Combinations of PvDBPII AH and PvEBPII showed evidence of an increased protective effect (aIRR = 0.30, 95% CI, 0.19–0.46, P<0.001, S4 Table). Combinations of 3 antigens did not show an additional increase in protection (S4 Table).
For all of the antigens tested, with the exception of PvDBPII O, high levels of IgG1 were associated with decreased risk of clinical malaria in the adjusted models (aIRRH = 0.38–0.70, P ≤ 0.027) (Fig 3 and S3 Table). For IgG3 responses, the analysis was restricted to PvDBPII AH and PvEBPII as detectable antibody levels were only observed for them. Both PvDBPII AH and PvEBPII responses also showed a protective effect (aIRRH = 0.55–0.68, P≤0.021). In a multivariate model incorporating IgG1 and IgG3 for all antigens, only IgG3 to PvDBPII AH and IgG1 to PvEBPII remained associated with clinical protection (aIRR = 0.38–0.63, P ≤ 0.006) (S3 Table).
Plasma obtained at first (n = 8) and last visit (n = 8) exhibited substantial binding inhibitory antibodies against diverse PvDBPII alleles (Table 2). Binding inhibitory antibodies against the three PvDBPII variants was also significantly correlated (P<0.001) with the highest correlation observed between the two most prevalent alleles PvDBPII AH and PvDBPII O (rho = 0.66, P<0.001).
Twelve children (7.14% of 168 tested) had ≥60% blocking activity for at least one variant of PvDBPII. Eight children (4.76%) had high levels (≥80% blocking activity) of inhibition to one variant, six of which showed high blocking activity against all three variants. Children with concurrent P. vivax infections showed moderate-high blocking activity (>60% blocking against all three alleles, PCR positive: 13.3% vs. PCR negative: 3.5%, P = 0.027). Although median blocking activity did not vary with age (P>0.19), five of the six children with high levels of inhibitory, strain-transcending antibodies were older than 21 months of age (P = 0.115).
Blocking activity in plasma samples collected at the end of follow-up was very similar to the start of the study (Table 2). After 16 months of additional exposure, concurrent P. vivax infections were no longer associated with an increase in blocking activity. Only three children (1.9%) had high blocking antibodies (≥80%) against all three variants at both time points. Among the three children with constant high blocking activity, two (66.7%) were homozygous for the Gerbich blood group (i.e. GYPCΔex3) compared to 14 (10.5%) in those with lower or no blocking activity (P = 0.036).
When assessing the association between the ability of antibodies to block PvDBPII binding to red blood cells and prospective risk of P. vivax malaria, children with high blocking ability against AH (IRR = 0.44, P = 0.059), O (IRR = 0.52, P = 0.119), Sal1 (IRR = 0.52, P = 0.081), or all three alleles combined (IRR = 0.45, P = 0.083) at enrolment showed a tendency for a reduced incidence of P. vivax episodes of any density (Table 3). These effects were almost entirely due to the three children with high strain-transcending blocking activity at both enrolment and end of follow-up (IRR = 0.16, 95% CI, 0.03–1.04, P = 0.055).
Neither blocking nor total IgG antibodies to any of the P. vivax proteins showed any protective association with the risk of P. falciparum clinical episodes, but all children with high levels of total IgG were associated with increased episodes of clinical P. falciparum malaria, suggesting antibodies against PvDBP and PvEBP were correlates of increased risk of P. falciparum exposure (S5 Table).
In this study, 29 children harbored Gerbich phenotype caused by double deletion of exon 3 of GYPC gene, 111 of them with single deletion of the same region named as heterozygote and other 84 were without any deletion called wild-type. Children with Gerbich phenotype had a reduced risk of malaria episodes in comparison to those with wild-type, and the strength of this relationship increased with increasing parasite densities (aIRR = 0.69, 95%CI = 0.41–1.01, P = 0.040, for P. vivax >500 parasites /μL; aIRR = 0.53, 95%CI = 0.28–1.00, P = 0.050 for >2,000 parasites /μL; aIRR = 0.40, 95%CI = 0.17–0.94, P = 0.036 for >10,000 parasites /μL; Table 4). Similarly, the geometric mean parasitemia was significantly lower in children with homozygous Gerbich phenotype than those with wild-type (P = 0.003).
Children with wild-type phenotype had the lowest levels of antibodies against all PvDBP variants, while homozygotes had the highest levels to almost all antigens (P<0.011–0.046), except for PvEBPII (P = 0.501) (Fig 4). These results suggested that GYPCΔex3 may contribute to the acquisition of antibodies to PvDBP but not to PvEBPII in PNG children.
This study confirmed that total IgG and infrequently detected BIAbs against PvDBPII were associated with an overall lower incidence rate of clinical vivax malaria in young children who were developing clinical immunity to P. vivax. Antibody responses against PvDBPII were higher in those Gerbich-negative, a common red blood cell polymorphism within the East Sepik region of PNG. We also observed a strong association with protection for total IgG and IgG1 antibodies to PvEBPII, but no difference by Gerbich phenotype.
In this study, only 4.8% (8/168) of the young children aged 1 to 3 years had acquired high levels of BIAbs (> 80% binding inhibition) against at least one PvDBPII allele, with six children exhibiting binding-inhibitory antibodies against diverse strains, while none of them had obtained levels higher than of 90% of BIAbs against any PvDBPII allele. An earlier study among school-age children (5–9 years) in PNG identified 9% of children with BIAbs >90% to PvDBPII [18] and in the Brazilian Amazon, 26.6% and 20.5% of the residents of all age groups presented >80% and >90% BIAbs activity to PvDBPII [19]. In these two studies [18, 19], high anti-PvDBPII BIAbs activity blocked diverse strains. The overall lower detection of high-levels PvDBPII BIAbs in our study indicate that acquisition of BIAbs is at least partially related to the increase in life-time malaria exposure. However, the observation that BIAbs are not common even in adult population with high levels of P. vivax exposure indicates that these functional blocking antibodies are difficult to acquire under conditions of natural exposure. Nevertheless, since the target epitope of BIAbs to PvDBPII was conserved [46], once PvDBPII BIAbs are acquired, they may provide strain transcending binding inhibitory activity, even in young children with limited and developing clinical immunity.
In contrast to the DARC conserved binding residues, polymorphisms thought to be associated with immune evasion [10] are common throughout nonfunctional regions of PvDBPII distal to the binding site of DARC. Our study and previous reports show that these regions of PvDBPII are exposed to the immune system resulting in observable immune responses to PvDBPII generated at a young age, even in our cohort of young children with limited immunity [14, 18]. However, antibody responses to these polymorphic PvDBPII regions are likely to be strain-specific [14] and potentially not functionally associated with clinical protection [18, 19]. In this young cohort of PNG children, IgG to the dominant variants PvDBPII AH and O were both more prevalent and strongly associated with protection against clinical vivax-malaria in the adjusted models. As such, total IgG responses to these PvDBPII variants may serve as biomarkers for protective immunity.
Regions II and VI of PvDBP are cysteine-rich domains and under immune selection, while regions III to V do not have a known function [8]. Nevertheless, one study in P. falciparum showed that antibodies to regions III-V of PfEBA175 and PfEBA140 could inhibit P. falciparum invasion [47]. Our study showed that antibodies to region III-V predicted protection from vivax clinical malaria. However, antibody titers to region III-V were significantly correlated to antibodies against some PvDBPII variants and did not retain a significant association with protection in the multivariate models. Future functional studies will be required to investigate the potential functional role of antibodies to region III-V of PvDBP.
Among all antigens tested, the strongest association of protection from vivax malaria was identified for PvEBPII. In one previous study, PvEBPII was characterized as a functionally and antigenically distinct P. vivax ligand with a stronger binding preference for Duffy-positive than Duffy-negative reticulocytes in vitro [21], suggesting that although antigenically distinct from PvDBP, it may function as a redundant invasion pathway when immune activity blocks the principal PvDBPII-DARC pathway. The relatively low correlation between antibody responses to PvDBP and PvEBPII indicates co-acquisition of antibodies to both antigens, rather than cross-reactivity between them. Combination of PvDBPII AH and PvEBPII immune responses offered a further increase in protection against clinical vivax-malaria and thus supports the inclusion of these two antigenically distinct ligands in a combination vaccine. PvEBPII should be further investigated as a potential vaccine candidate, and the efficacy of anti-PvEBP antibodies will need to be confirmed in functional assays.
Some polarization of IgG1 towards IgG3 was only identified for PvDBPII AH, as similar switch towards a balanced IgG1/IgG3 response was observed in the narrow age range of our study population. The possible reason for this might be related to the higher circulation of AH strain in this setting, thus young children might be exposed to the AH strain enabling them to acquire higher IgG3 to AH. Apart from the possible higher exposure, by comparing with other alleles included in this manuscript, AH was shown to have two special mutations K371E and K386Q, both of which were mapped to SD2. As most of the PvDBPII polymorphism are in this subdomain, our results suggest that these two mutations were important for immune reactivity to PvDBPII AH. Whereas, after a longer period of immune exposure to other strains, including O, P and Sal1, significantly higher levels of IgG3 to them were also observed in the adults. Consistent with previous reports [25, 28], these results indicated that IgG3 switching may be driven by the nature of the antigen and influenced by exposure and maturation of the immune system. In multivariate models of isotype-specific responses, for IgG3 against PvDBPII AH and high IgG1 to PvEBPII were significantly associated with a reduced risk of clinical vivax-malaria. It remains to be confirmed whether IgG1 and IgG3 antibodies target different epitopes and/or differ in their functionality or if they simply differ in their utility as correlates of risk of future infection or protection against vivax malaria.
In addition to antibody-mediated protection against vivax malaria, it is believed that specific red blood cell polymorphisms can induce resistance to clinical malaria [33]. There is limited evidence that individuals with Gerbich negativity may have a lower risk of P. falciparum and/or P. vivax infection [37]. Our longitudinal study now provides the first indication for a protective role of the Gerbich phenotype with reduction of P. vivax malaria among young children aged 1 to 3 years with limited clinical immunity. However, Gerbich negativity was not associated with significant protection against blood-stage infection in school-aged children (5–14 years) from another PNG longitudinal cohort study [39]. It is assumed that the acquired, clinical immunity among older semi-immune children and immune adults may mask the protective effects of specific genotypes against uncomplicated malaria infection [39]. In this study, a higher proportion of Gerbich homozygotes was found among children with high titers of PvDBPII-specific BIAbs. This was consistent with a previous observation that children with the South-East Asian ovalocytosis (SAO, caused the SLC4A1Δ27 deletion in the human Band 3) were 3.3 times more likely than non-SAO children to have high levels of PvDBPII-specific BIAbs [48]. The mechanism by which Gerbich affects anti-PvDBP antibody responses is unknown and will require further investigation.
In summary, our study highlights the association of total antibody and BIAbs to PvDBPII variants with lower risk of clinical P. vivax malaria episodes in PNG children and further strengthens the rationale for PvDBPII as a potential vaccine antigen. Interestingly, both naturally acquired immunity to PvDBPII and Gerbich homozygosity showed significant protection against P. vivax. Antibodies to PvEBPII were more strongly association with clinical protection against P. vivax malaria in these young children. Further studies will be needed to clarify the mechanism of protection afforded by Gerbich and the importance of PvEBPII in P. vivax reticulocyte invasion and immune protection.
|
10.1371/journal.pntd.0004999 | Development and Assessment of a Geographic Knowledge-Based Model for Mapping Suitable Areas for Rift Valley Fever Transmission in Eastern Africa | Rift Valley fever (RVF), a mosquito-borne disease affecting ruminants and humans, is one of the most important viral zoonoses in Africa. The objective of the present study was to develop a geographic knowledge-based method to map the areas suitable for RVF amplification and RVF spread in four East African countries, namely, Kenya, Tanzania, Uganda and Ethiopia, and to assess the predictive accuracy of the model using livestock outbreak data from Kenya and Tanzania. Risk factors and their relative importance regarding RVF amplification and spread were identified from a literature review. A numerical weight was calculated for each risk factor using an analytical hierarchy process. The corresponding geographic data were collected, standardized and combined based on a weighted linear combination to produce maps of the suitability for RVF transmission. The accuracy of the resulting maps was assessed using RVF outbreak locations in livestock reported in Kenya and Tanzania between 1998 and 2012 and the ROC curve analysis. Our results confirmed the capacity of the geographic information system-based multi-criteria evaluation method to synthesize available scientific knowledge and to accurately map (AUC = 0.786; 95% CI [0.730–0.842]) the spatial heterogeneity of RVF suitability in East Africa. This approach provides users with a straightforward and easy update of the maps according to data availability or the further development of scientific knowledge.
| Rift Valley fever (RVF) is a zoonotic disease affecting ruminants and humans. It occurs mostly in Africa, causing human deaths and important economic losses in the livestock sector. The RVF virus (RVFV) is transmitted from ruminant to ruminant by mosquitoes. Different climatic, environmental, and socio-economic factors may impact the transmission of the virus. Our work uses all current available knowledge on the epidemiology of the disease and geographic data to map areas suitable for RVFV. The study area includes four East African countries: Kenya, Tanzania, Uganda, three countries which have been historically affected by RVF, and Ethiopia, where the disease has never been reported but which shares borders with infected countries. The resulting maps are compared with the locations of outbreaks reported in livestock. Our results demonstrate the capacity of the spatial multi-criteria evaluation method to map with accuracy the areas suitable for RVF occurrence. Thus, the method we developed provides users with risk maps that could be used for early warning detection and implementation of control measures.
| Caused by a Phlebovirus (Bunyaviridae) that affects both humans and livestock, Rift Valley fever (RVF) is considered to be one of the most important viral zoonoses in Africa. The RVF virus (RVFV) is transmitted from ruminant to ruminant by mosquitoes [1]. Although never demonstrated, there is field, serological and virological evidence of transmission without any use of vectors [2], suggesting an alternative transmission of the RVFV between ruminants through direct contact. Humans become infected mainly through direct contact with ruminant viremic fluids, such as blood or abortion products, but also through mosquito bites.
Although in the majority of human cases RVF infection is asymptomatic or causes mild illness, severe forms are characterized by retinitis, encephalitis or hemorrhagic fever. In ruminants, RVF infection causes abortion storms in groups or flocks of pregnant females and acute deaths in newborns [3]. Both health and economic impacts can be greatly reduced when control measures, such as vaccination, insecticide spraying and dissemination of information, are quickly implemented. The delay between case detection and control measure implementation depends on, among other factors, the efficiency of surveillance networks; therefore, an accurate definition of at-risk areas needs to be monitored along with other factors.
RVF virus circulation has been reported in several eco-climatic areas: arid in western Africa and the Arabic Peninsula [4, 5]; sub-humid in East Africa [6, 7]; wet forests in central Africa [8]; dam and irrigated agricultural land under hot climatic conditions in Egypt, Mauritania and Sudan [9–11]; and humid highlands in Madagascar [2, 12].
Depending on the areas of concern, different risk factors have been identified, either for transmission, spread or human and/or livestock occurrence. Potential mosquito vectors of the RVFV belong to the genera Aedes, Anopheles, Culex, Eretmapodites and Mansonia. The majority of the factors driving mosquito vector presence and abundance, thus driving the risk of RVF transmission, are related to climate, water and landscape. The Aedes genus is mostly associated with temporary water bodies such as flooded area, temporary pond, puddles, and rice fields. Culex and Anopheles mosquito breeding areas are diverse and could be temporary (rice fields, swamps) or permanent (lakes, ponds) bodies of water. Stagnant and permanent water bodies are the habitats of Eretmapodites and Mansonia, respectively [13].
In fact, the presence of temporary water bodies and floodplains are outbreak risk factors for RVF in semi-arid areas in eastern Africa, the Arabian Peninsula and western Africa [4]. In eastern and southern Africa, the risk of RVFV infection has been shown to vary as a function of rainfall, temperature, and a remotely sensed vegetation index (NDVI: normalized difference vegetation index) [14, 15]. Artificial water bodies, such as dams and irrigated rice fields, are also known to be associated with the high abundance of RVFV vectors in western Africa [4]. In addition to eco-climatic factors, cattle density has been identified as a risk factor for transmission of the RVFV [6, 16]. Habitat, gender, profession, and contact with ruminants and ruminant products have also been identified as risk factors of RVF occurrence in humans [17].
The Horn of Africa has been historically affected by RVF [18]. However, the occurrence of RVF has never been reported in Ethiopia, which shares borders with infected countries, namely, Kenya [6], northern Somalia [19], and Sudan [20]. In Uganda, although no outbreak in humans or animals were reported until 2016, a recent serological survey revealed that RVFV was endemic in goats in four districts [21].
In Kenya and Tanzania, where RVF is endemic, historical knowledge indicates that ‘dambos’ are areas at risk of RVF [22]. Moreover, recent eco-epidemiological studies identified the main environmental risk factors for RVF, which has allowed for health targeted surveillance by health authorities [6, 14, 23]. However, the application of these models to regions outside of the study area of interest may lead to incorrect inferences. Moreover, information related to the suitability of both Ethiopia and Uganda ecosystems for the transmission of the RVFV are scarce. Given this lack of information, pragmatic approaches must be developed to provide risk maps that could be used for early warning detection and implementation of control measures.
Spatial multi-criteria evaluation (MCE) is a rapid and pragmatic knowledge-based method adapted for mapping disease suitability in the absence of large epidemiological datasets. Defined as ‘a process that transforms and combines geographical data and value judgments (expert and bibliographic knowledge, including uncertainties, subjective and qualitative information) to obtain appropriate and useful information for decision making’ [24], this method has been used to map suitable areas for RVF transmission in Africa, on a continental scale [25] and in Senegal [26], and in the European countries of Spain [27] and Italy [28]. However, the predicted maps could not be validated for European countries which are disease-free regions, while in Senegal the validation could only be performed in a qualitative way [26]. Moreover, in these studies, only the ‘amplification step’, defined as the local transmission of the RVFV to its hosts by mosquito vectors, was considered and did not account for the possible transportation of the virus from a primary outbreak to secondary foci in a ‘spread step’. This process may involve different risk factors than those of the amplification step, such as animal trade [12, 29].
The goal of the present study was to adapt a geographic knowledge-based method [25] to identify suitable areas for RVF amplification and spread in four Eastern African countries, namely, Kenya, Tanzania, Uganda and Ethiopia, and to assess the predictive accuracy of the model using livestock outbreak data from Kenya and Tanzania.
Under suitable conditions and after the introduction or low-level transmission of a given pathogen, the pathogen can be locally transmitted to a ‘primary’ host through direct or vectorial transmission and then transferred from the primary infectious host to several secondary hosts; this is the “amplification” process [30]. Therefore, ‘spread’ is defined as the transportation of the pathogen from the primary outbreak to secondary foci. In this study, ‘suitability’ is defined as the ability of a habitat to support either the amplification or the spread of RVF. Amplification is a necessary condition for primary RVF occurrence. Both amplification and spread are needed for secondary outbreaks.
Following the spatial multi-criteria evaluation (MCE) methodology that has been detailed elsewhere [25, 31], we first identified the amplification and spread risk factors of RVF through a literature review. PubMed and ISI Web of Knowledge were searched for articles published from 1980 to 2014 using the search terms ‘‘rift valley fever” AND (separately) ‘‘model” or ‘‘spatial”, or “risk factors” or “analysis” using the ‘‘all fields” option to allow for the retrieval of articles in which the search terms appeared in the titles, abstracts, or keywords. Inclusion criteria were reviews and/or articles using expert knowledge, and/or statistical and mathematical modelling approaches to model RVF risk to ruminants or humans. A total of 62 references were thus included.
In Table 1, we listed the factors associated with the risks of amplification and spread of RVF according to the published literature review. Only risk factors that could be mapped were selected for the risk mapping process. The following risk factors were thus included:
A search was conducted to obtain digital geographical data for each identified risk factor (Table 1). The different sources of the data that were used and their main characteristics are provided in S1 File and S1 Table. The data were imported into a geographic information system (GIS) and processed to produce standardized spatial risk factor layers, namely a mosquito index (reflecting the suitability for RVF mosquito vectors), sheep density, goat density, cattle density, proximity to markets, road density, railways density, proximity to water bodies, proximity to wildlife national parks (software: ESRI ArcMap and ArcMap Spatial Analyst Extension, Redlands, CA, USA). At the end of the process, each image layer was raster-based, with pixel dimensions of 300 m x 300 m. The scale of all spatial risk factor layers was continuous, ranging from 0 (completely unsuitable) to 1 (completely suitable). The different sources of the data used and the calculation method of the standardized geographical layers are provided in S1 File. The resulting maps for the risk factor layers are presented in S1 Fig.
We assumed that the weight of each risk factor in the amplification and spread processes were not equivalent. For example, small ruminants are known to be more susceptible than cattle for the transmission of the RVFV [42] and, therefore, more prone to play a more important role during the amplification phase than the latter. Markets are aggregation points where ruminant herds meet and have potential contact with each other before returning back to their living area; markets are, therefore, more important for spread than roads that may be used by herders for travelling. Based on the literature review and our own expertise, we ranked RVF risk factors for virus amplification and spread according to their putative relative importance in both processes [78]. Factors were compared two at a time: 1) We first specified whether risk factor A was more or less important than risk factor B; and 2) We specified the degree of importance of factor A regarding factor B on a nine-point scale using the Saaty scale (factor A can be extremely more important, very strongly more important, strongly more important, moderately more important, equally important, moderately less important, strongly less important, very strongly less important or extremely less important than factor B), resulting in a pair-wise comparison matrix. A numerical weight was then derived for each risk factor from the pair-wise comparison matrix [79]. We calculated pair-wise comparison matrices separately for amplification and spread, considering the vector distribution being much more important in the amplification process than in the virus spread phase.
Then, three different maps were generated, considering two distinct groups of risk factors and their associated weights for the amplification and spread steps.
Assuming that vector presence is a necessary condition for RVF amplification, the suitability of areas for RVF amplification was calculated for each raster cell as the product of the mosquito index map (computed as described in S1 File) and a weighted linear combination (WLC) of each of the standardized geographical risk factor layers associated with RVF amplification using its corresponding weight. Regarding the spread process, the raster maps for RVF risk factors associated with RVF spread were combined by computing a WLC with their corresponding weights.
The suitability maps for amplification and spread were then combined to create two different suitability maps for RVF occurrence. First, the values of suitability indices for amplification and spread were recoded in three classes (low/medium/high suitability) using a quantile discretization. These two recoded maps were then merged into a primary synthetic RVF suitability map with nine classes corresponding to all possible combination of amplification suitability (low/medium/high) and spread suitability (low/medium/high). Second, the areas with the highest risk (suitability estimates greater than 0.1, i.e., in the 90th percentile) of RVF amplification were selected. The Euclidean distance between these areas was calculated and transformed into a ‘proximity to RVF amplification areas’ index, which assumed a sigmoid-shaped decreasing relationship between 0 and 100 km and negligible risk thereafter. Finally, suitability estimates for RVF occurrence, combining the spread and amplification processes, were computed as the product of the suitability estimates for RVF spread and the proximity to RVF amplification areas index, resulting in a second synthetic RVF suitability map expressed as a continuous suitability index.
A sensitivity analysis (SA) was conducted to assess the sensitivity of the method to the expert choices. To determine the effect of a change in the weights applied to each risk factor, a range of weight values to explore was defined by adding and subtracting 25% from the original weights. Ten weight values within this range were tested (+/-5%, +/- 10%, +/-15%, +/-20% and +/-25%). Each of the newly calculated weights was incorporated into the modelling process, while other factor weights were proportionally decreased or increased such that the sum of the weights equaled 1. For each combination of weights obtained, maps of suitability indices for RVF amplification, spread and occurrence were calculated, and a total of 169 suitability maps for RVF occurrence were generated. Based on these different realizations, the contribution of the different risk factor weights to model output variance was calculated for each country (see S2 File for details).
Finally, an uncertainty surface was produced. It represented the standard deviation of the different suitability maps resulting from the change in weights [80].
RVF outbreaks in livestock reported in Kenya and Tanzania between 1998 and 2012 were collected to assess the consistency of RVF suitability map (Source: FAO EMPRES-i database: http://empres-i.fao.org). A total of 145 outbreaks were located using geographic coordinates (Fig 1).
Then, 150 locations of disease ‘pseudoabsence’ data were randomly generated in these two countries, under the condition of being 25 km from another ‘pseudoabsence’ or outbreak location. The value of the quantitative suitability estimates for RVF occurrence was extracted for each ‘presence’ or ‘pseudoabsence’ location and the AUC (area under curve) of the ROC curve [81], and the sensitivity and specificity were calculated to evaluate the quality of the suitability map.
The resulting weights of risk factors for RVFV amplification and spread are presented in Table 2 (see S1 and S2 Tables for the details of the pair-wise comparison matrices).
Regarding amplification, we assumed that the mosquito index was the most important factor, and a necessary condition for RVFV amplification. Then, small ruminant densities were identified as important factors, followed by (in descending order) cattle density, proximity to markets that are aggregation points for animals, proximity to roads, water bodies and railways, and proximity to wildlife parks.
Regarding spread, we considered that viremic hosts were the most important means of virus dissemination and that markets were, again, an aggregation point for people and their herds. The proximity to roads and water bodies were also important factors because they allow for trade movements. Finally, proximity to wildlife national parks and the mosquito index were factors of low influence in the spread process.
Fig 2 presents the different maps produced from the MCE process: a map of suitability for RVF virus amplification (Fig 2a), a map of suitability for RVF spread (Fig 2b) and a final map of suitability for RVF occurrence in domestic ruminants (Fig 2c) (maps of all standardized risk factors are provided in S1 Fig).
According to the results of the MCE, areas suitable for RVF amplification were located in the low elevation areas of Kenya (the eastern coast, the northeastern portion that borders Ethiopia and Somaliland, and, to a lesser extent, the northwest region bordering Uganda), Tanzania (northeastern portion) and Ethiopia (in the Northwest and Southwest). Uganda presented a very low suitability for RVF amplification in domestic ruminants.
In addition, the suitability map for RVF spread in domestic ruminants (Fig 2b) showed a different pattern, identifying the highlands of Ethiopia, Kenya, and Uganda as areas favorable to RVF spread. In Tanzania, areas suitable for RVF spread were located in the northern part of the country.
The combination of the ‘amplification’ and ‘spread’ maps resulted in two final synthetic maps of the areas suitable for RVF occurrence in domestic ruminants that were complementary (Fig 2c and 2d). According to the map that highlights the different categories of amplification/spread combinations (Fig 2c), the majority of eastern Kenya was identified as highly suitable for RVF occurrence, with a medium-to-high suitability for RVF amplification combined with medium-to-high suitability for RVF spread. This pattern was also observed in northeastern Tanzania and southwestern Ethiopia. Areas with medium suitability for both RVF amplification and spread, such as northwestern Tanzania, or areas with low amplification suitability but high spread suitability, such as western Kenya, the majority of Uganda and the Ethiopian highlands, were identified as suitable for RVF occurrence. This first map also highlighted areas with low suitability for RVF occurrence (low suitability for amplification and spread), such as eastern Ethiopia and central Tanzania. Taking into account the proximity of the areas with the highest suitability for RVF amplification, the second synthetic RVF suitability map (Fig 2d) highlighted the different patterns of the areas suitable for RVF occurrence: the majority of Kenya was identified as suitable; however, in the three other countries, the areas suitable for RVF occurrence were smaller than those shown in Fig 2c.
The uncertainty of the surface-based model produced by the data for the four countries showed that the predictions of the location of suitable areas for RVF occurrence in livestock were robust, meaning that they remained stable when varying the risk factor weights in the ‘amplification’ and ‘spread’ steps. Indeed, the maximum standard deviation (STD) of the suitability maps for RVF occurrence was less than 0.1. The results highlighted a spatial heterogeneity in uncertainty, with higher uncertainty in the western parts of Ethiopia and Kenya (S2 Fig).
The sensitivity analysis showed that the variation in the suitability index was explained by four factors for the amplification step and by seven factors for spread (Fig 3). The most sensitive parameters for the amplification step were the sheep density and the proximity to markets, wildlife national parks, and water bodies. Regarding the spread step, the most sensitive parameters were the cattle and goat densities, the density of roads and railways, and the proximity to national parks and water bodies. The importance of these sensitive parameters varied among the four countries, particularly the importance of the livestock densities (cattle, sheep and goats) (Fig 3).
The ROC AUC associated with the suitability map for RVF occurrence in Kenya and Tanzania showed a good fitting (AUC = 0.786; 95% CI [0.730–0.842]) (Fig 4a), demonstrating the capacity of the model to distinguish ‘presence’ from ‘absence’ locations with good predictive accuracy (Fig 4b). With a cut-off point of 0.3 maximizing both sensitivity and specificity, the sensitivity was 0.74, and the specificity was 0.75. A total of 74% (107 out of 145) of the RVF outbreak locations were mapped in at-risk areas, which were defined as the areas with a suitability index for RVF occurrence greater than 0.3, the cut-off point value maximizing sensitivity and specificity (Fig 5b).
Many regions from Kenya and Tanzania were previously and heavily affected by RVF outbreaks [6, 82]. However, some areas may be at-risk without having experienced outbreaks in past years. The identification of these areas is essential for implementing risk-based surveillance and reducing the impact of RVF human and animal outbreaks in the coming years. Until 2016, Uganda and Ethiopia remained free from outbreaks, but their geographical locations as well as the livestock exchanges they have with their neighbors make these two countries highly vulnerable to the disease.
In this context, the implementation of the GIS-based MCE method for RVF risk mapping appeared to be a very efficient method to map suitability areas for the amplification and spread of the virus based on freely available geographic data. To our knowledge, this is the first study aiming to produce regional suitability maps for RVF using MCE methodology combined with outbreak dataset validation.
Validation of the suitability map using disease presence and background data randomly generated produced good results according to the ROC AUC method (AUC = 0.786). However, the use of randomly generated ‘pseudoabsence locations’ may be controversial; indeed, the absence of reported outbreaks is not an evidence of absence of pathogen transmission. The results of regional serological surveys may give a more precise evaluation of the RVF suitability map.
Nevertheless, 74% of the reported RVF outbreaks in livestock were located in areas with the highest predicted suitability for RVF occurrence (Fig 5b). Interestingly our MCE-based model performed better than other predictive models based purely on climatic anomalies and previously validated with human outbreaks [14]. These models, which showed the highest accuracy in the Eastern African region, included 65% of the human case locations in predicted at-risk areas. Two human cases of RVF have been reported in early March 2016 in the Kabale District, southwestern Uganda [83]. The outbreak occurred in an area that was identified by our model as poorly suitable for RVF amplification but highly suitable for RVF spread (Fig 5a). This result is highly consistent with the socio-economic and ecological environment of Kabale district. Indeed, Kabale is an important commercial center with six animal markets, a situation associated with a higher risk for RVF spread according to our assumptions. Being outside of the ‘potential epizootic area mask’ [14], this area is not predicted by the climate-based model [83]. Despite the strong 2015–2016 El Niño phenomenon and the associated abnormal rainy season in East Africa, no substantial climatic anomalies were observed in the Kabale area during the 2016 epidemics. Differently from the southeastern and central districts in Uganda and neighboring countries, such as Kenya and Tanzania, both cumulative precipitation as well as NDVI values were lower than or equal to average in Kabale area during the period September 2015 to February 2016, except for short periods in October and December 2015. We therefore hypothesize a little role of vector-borne transmission in the Kabale outbreak.
Thus, our results highlighted the importance of taking into account livestock data and the factor of animal trade in addition to environmental factors to develop predictive maps of RVF occurrence. Moreover, these maps increase the confidence level for the approach applied to RVF free-areas [27, 28].
Indeed, the MCE approach we applied to four countries of eastern Africa was very similar to previous modelling studies that used the same approach in different geographic contexts [25–28]. All of these studies considered two main categories of risk factors: on the one hand, those related to domestic ruminant densities, and on the other hand, those related to vector presence (i.e., vector distributions or proxies of vector distributions, such as temperature, elevation, rainfall, and proximity to aquatic areas). One of the distinctive features of our study was the ability to distinguish RVF amplification and spread steps in the modelling process, thus considering risk factors related to animal trade and movements (markets, roads and railways). Moreover, the hypothesized role of wildlife reservoirs in the amplification and spread of RVF was considered.
Identifying areas of low amplification with high spread suitability and vice versa (Fig 5a) was expected because these two epidemiological phenomena imply different mechanisms: vector and host densities favor local amplification, whereas animal movements favor the long-range spread of the disease. From a control perspective, surveillance strategies should be adapted; active surveillance in sentinel herds would be relevant in amplification areas that act as virus sources for areas that are not at risk of amplification, and analysis of the trade network and the existing links between amplification areas and other regions could be used as an early warning tool to protect spread areas from viremic ruminant introduction in case of primary foci.
However, the limits of our method must be noted. First, in the absence of homogeneous information on RVFV vector abundance and distribution in our study area, we used environmental variables to map a vector index reflecting the suitability of locations for the presence of RVF vectors. These variables were identified through a study performed in Kenya; this study may not be perfectly relevant for the three other countries because this vector suitability map needs to be validated by landscape-targeted mosquito trappings in each country. Among mosquito species vectors recorded in the countries of concern, several were demonstrated to be competent in the lab [42]. However, even if competency measures were to provide elements to infer the potential role of a given mosquito species in RVF outbreaks, these measures are not sufficient to definitely incriminate the species. Indeed, mosquito abundance and foraging behavior are major elements that also shape the epidemiology of arboviruses. Better knowledge of these entomological characteristics should be considered to improve the vector index map.
Second, the spatial scale chosen for mapping the suitable areas for disease transmission has a great impact on the produced maps due to the spatial resolution of the data used to calculate the risk factors and the choice of the risk factors included to map suitability areas for pathogen transmission. Moreover, the weight attributed to each risk factor may differ between regional and national scales. In this study, we provided regional maps of suitability for RVF; however, maps produced on a national scale with higher spatial resolution, derived from risk factors and weights discussed with experts of each country (with particular attention paid to the most sensitive weights identified by the SA), would be more accurate and useful for surveillance and control purposes. Moreover, the threshold used to define the areas highly suitable for virus amplification (90th percentile threshold considering the whole study area) may introduce a bias in suitability predictions. This threshold value should be adapted for each country to provide better predictions at the national level.
Lastly, limitations of the produced map are related to the availability of data used as risk factors and their quality. For instance, due to a lack of geographic data on the locations of veterinary services and surveillance networks within each country, this information was not taken into account in our model, although both are key factors for the control of animal diseases—the spread intensity and magnitude from primary foci depending on the detection delay and, thus, on the efficiencies of veterinary and surveillance services. Moreover, the immunological status of ruminant populations induced by previous virus circulation episodes or vaccination campaigns were not taken into account although they are important factors, as demonstrated in South Africa in 2010 [84]. Additional information on how landscape features and socio-economic factors impact domestic and wild ruminant movements may also be important to refine the cost distance calculations of markets, water and wildlife park proximity indices.
In this study, we focused on mapping the areas suitable for RVF amplification and spread; thus, we focused on the spatial dimension of RVF risk. Although transmitted by mosquitoes and probably by direct contact, RVF is a seasonal disease, occurring during or at the end of the rainy season when mosquito abundance is at its highest. Future work should take into account this temporality to provide seasonal suitability maps for RVF transmission in livestock. Coupling climate-based models [14] with the RVF suitability map, which includes livestock and commercial variables, would allow for the development of seasonal suitability maps for RVF transmission in livestock. However, it must be stressed that this requires a good understanding of the drivers of RVF emergence. Indeed, with the exception of Kenya, where a strong association was demonstrated between heavy rainfall events and outbreak occurrence [22], rainfall may not be the only key factor for RVF emergence. Host density, associated with suitable climatic conditions and the introduction of the virus by ruminant trade, probably led to the 2000 outbreak in Yemen [32]. In Madagascar in 2008, no abnormal rainfall was noticed before the outbreak [85]. In Senegal in 2003, an intense transmission was described without any abnormal rainfall [44]. Soti et al. (2012) observed that in this region, the rainfall pattern rather than rainfall abundance could be responsible for triggering outbreaks [86]. Therefore, the seasonality of outbreaks should be incorporated in models with caution, depending on the area considered.
The present study confirmed the capacity of GIS-based MCE method to synthesize available scientific knowledge and map with accuracy the spatial heterogeneity of RVF suitability in four countries of East Africa. Moreover, such an approach enables users a straightforward and easy updating of the maps according to data availability or scientific knowledge development to include more precise geographic data or additional risk factors and to modify the weights of each factor.
|
10.1371/journal.pbio.0050078 | Glucocorticoids Play a Key Role in Circadian Cell Cycle Rhythms | Clock output pathways play a pivotal role by relaying timing information from the circadian clock to a diversity of physiological systems. Both cell-autonomous and systemic mechanisms have been implicated as clock outputs; however, the relative importance and interplay between these mechanisms are poorly understood. The cell cycle represents a highly conserved regulatory target of the circadian timing system. Previously, we have demonstrated that in zebrafish, the circadian clock has the capacity to generate daily rhythms of S phase by a cell-autonomous mechanism in vitro. Here, by studying a panel of zebrafish mutants, we reveal that the pituitary–adrenal axis also plays an essential role in establishing these rhythms in the whole animal. Mutants with a reduction or a complete absence of corticotrope pituitary cells show attenuated cell-proliferation rhythms, whereas expression of circadian clock genes is not affected. We show that the corticotrope deficiency is associated with reduced cortisol levels, implicating glucocorticoids as a component of a systemic signaling pathway required for circadian cell cycle rhythmicity. Strikingly, high-amplitude rhythms can be rescued by exposing mutant larvae to a tonic concentration of a glucocorticoid agonist. Our work suggests that cell-autonomous clock mechanisms are not sufficient to establish circadian cell cycle rhythms at the whole-animal level. Instead, they act in concert with a systemic signaling environment of which glucocorticoids are an essential part.
| To guarantee normal growth and to avoid tumor formation, the timing of cell division must be under strict control. Remarkably, cells, from bacteria to man, often divide only at certain times of day, suggesting the influence of internal biological clocks. A central pacemaker structure in the brain controls diurnal rhythms of behavior and hormone release. However, biological clocks are also encountered in almost every cell type (so-called “peripheral” clocks), in which they regulate daily changes in cell biology, including cell division. Very little is known to date about how the two clock systems interact. Here, by examining zebrafish strains with defects in hormone production, we find that peripheral clocks require the steroid hormone cortisol to generate daily rhythms of cell proliferation. Interestingly, the daily changes in cortisol levels observed in normal zebrafish are not required to achieve this control; treating the cortisol-deficient strains with constant levels of a drug that mimics the effects of cortisol restores normal cell-division rhythms. Thus, it appears that internal cell timers cooperate with hormonal signals to regulate the timing of cell division.
| The physiology of most plants and animals changes significantly between day and night. These daily rhythms are generated by endogenous clocks or pacemakers, and persist under constant conditions with a period length of approximately 24 h (hence, they are termed circadian). In vertebrates, cell-autonomous circadian clocks are present in most cell types, and are termed peripheral clocks. In addition, a limited number of specialized central pacemakers such as the suprachiasmatic nucleus (SCN) of the hypothalamus [1,2] appear to play a key role in coordinating the function of peripheral clocks. Although it is known that ocular photoreception synchronizes the SCN pacemaker with the environment, the identity of the pathways that subsequently transmit timing information to the peripheral clocks remains elusive. Current models implicate multiple humoral signals that result indirectly from the SCN circadian control of systemic function, such as feeding behavior [3].
Both cell-autonomous and systemic regulatory mechanisms have been implicated in clock output pathways that relay timing information from the clock to physiological systems. Circadian E box enhancers represent key regulatory elements within the core transcription–translation feedback loop of the vertebrate clock. These promoter elements direct circadian rhythms of transcription of clock genes by acting as binding sites for the clock components CLOCK and BMAL1. Circadian E boxes are also encountered in the promoters of many non-clock genes (so-called clock-controlled genes, e.g., see [4]). Via such target genes and their downstream effectors, peripheral circadian clock components directly regulate many aspects of cell physiology, such as membrane trafficking, detoxification, nutrient metabolism, and the cell cycle [4]. The central SCN pacemaker, in contrast, has been documented to influence systemic functions ranging from locomotor activity rhythms and the sleep–wake cycle to endocrine activity. Thus, the circulating levels of many hormones are under circadian control and so exert their effects only during specific times of day. A major unexplored issue is the relative contribution of cell-autonomous and systemic factors in directing circadian clock outputs. Do certain clock outputs rely solely upon direct peripheral clock regulation, or do they require input from systemic factors, acting either upstream or downstream of the peripheral clocks? Are other outputs driven solely by circadian oscillations of systemic signals?
A particularly interesting clock output is the timing of cell proliferation. Daily rhythms of cell division are conserved across huge evolutionary distances, from cyanobacteria to humans [5,6]. This property has been proposed as a strategy for minimizing the ultraviolet damaging effects of sunlight during critical steps of the cell proliferation. In vertebrates, circadian gating of certain cell cycle steps also occurs in cell lines [7,8]. Furthermore, clock components have been implicated in controlling the transcription of cell cycle regulatory genes [9–12]. These observations imply that the circadian clock may regulate cell cycle progression via cell-autonomous mechanisms. However, given that systemic factors such as hormones are well-known regulators of cell proliferation [13,14], one important question is whether cell-autonomous regulatory mechanisms are sufficient to direct circadian cell cycle rhythms at the whole-animal level.
The zebrafish represents a valuable model for exploring the vertebrate circadian clock and its regulation of cell cycle timing. Robust daily S-phase rhythms are observed in larvae raised under light–dark (LD) cycles [7]. The persistence of these rhythms following transfer of the larvae to constant darkness (DD) conditions demonstrates that they are under control of the circadian clock. Furthermore, consistent with other clock outputs [15], exposure to a LD cycle is essential for the establishment of these rhythms because they are absent in larvae raised in DD. Circadian rhythms of S phase, albeit with lower amplitude, are also observed in zebrafish primary cell lines, implicating cell-autonomous regulation by peripheral clock mechanisms [7]. Interestingly, peripheral clocks in this species can be entrained by direct exposure to LD cycles [16]. However, zebrafish also possess central pacemakers: a structural counterpart of the SCN and a photosensitive pineal complex where nighttime synthesis of the hormone melatonin is directed by an endogenous clock [15,17–20].
Extensive panels of zebrafish mutants that show specific developmental defects in a range of organ systems have been assembled, thanks to large-scale screening efforts [21]. These animals represent potentially powerful tools to dissect the functional contribution of specific organs and tissues to the generation of clock outputs at the whole-animal level. Here, by studying a set of blind mutants, we have demonstrated that ocular photoreception is not required to establish circadian cell cycle rhythms during early larval development. In contrast, a severe attenuation of cell cycle rhythms is observed in mutants that exhibit a reduction or absence of the corticotrope cell lineage in the pituitary gland. Importantly, high-amplitude circadian cell cycle rhythms can be rescued by exposing corticotrope-deficient larvae to tonic concentrations of the glucocorticoid receptor (GR) agonist dexamethasone. Our work reveals the contribution of systemic factors to establishing circadian cell cycle rhythms at the whole-animal level.
We have previously demonstrated that exposure to a LD cycle is a prerequisite for circadian cell cycle rhythms to be established during early larval development. Whereas zebrafish peripheral clocks are directly light entrainable in vitro, light input through the eyes also plays an important role in entraining the circadian timing system in most vertebrates. We therefore asked whether ocular photoreception might contribute to establishing circadian cell cycle regulation in the zebrafish. We examined cell cycle rhythms in a set of functionally blind mutants using bromodeoxyuridine (BrdU) incorporation as a marker for the S phase of the cell cycle. lakritz/ath5 (lak) mutants, which carry a null mutation in a basic Helix-Loop-Helix transcription factor gene, the atonal homologue 5, lack the retinal ganglion cell layer [22]. These cells relay light information from the inner retina to the brain. Furthermore, in mammals, the retinal ganglion cells themselves function as a circadian photoreceptor [23]. These mutants are thus particularly well suited for examining the role of ocular photoreception in clock outputs. Mutant and wild-type siblings were raised under a LD regime, and BrdU incorporation was tested at four time points on the sixth day post-fertilization (dpf), before feeding starts. BrdU incorporation rhythms in the lak mutant larvae are indistinguishable from their wild-type siblings (Figure 1A) and thus not affected by the absence of ocular light input. To confirm these results, we examined chokh/rx3 mutant fish (carrying mutations in the retinal homeobox gene 3), which show a severe impairment of eye and retinal development [24–26]. We analyzed two alleles (“weak,” chkt25181 = chkw, and “strong,” chkt25327 = chks) with differing severity of the morphological phenotype [26]. As for the lak mutants, in the weak rx3 larvae, BrdU incorporation rhythms are indistinguishable from the wild-type siblings (Figure 1B). Surprisingly, however, the strong rx3 mutant larvae show a severe attenuation of the circadian S-phase rhythm compared with their wild-type siblings (amplitude reduced from 4.0-fold to 2.3-fold, Figure 1C and Figure S1). Since both lak and weak rx3 mutants are blind, we conclude that the attenuated S-phase rhythm of the strong rx3 mutant is functionally unrelated to its blindness. This suggests the presence of additional distinct defects in this mutant.
What is the cause of the attenuated cell cycle rhythms in the strong rx3 mutants? Given the proposed direct link between the cell cycle and the circadian clock [7–10], we first tested whether the cell cycle phenotype was due to a deregulation of the circadian clock itself. We examined mRNA expression of clock genes in larval RNA extracts from wild-type and rx3 mutant siblings raised in the same conditions as those of the BrdU experiments. We assayed expression of clock as a representative of the positive limb [27] and of per4 for the negative limb [28] of the circadian feedback loop [29]. As shown in Figure 2A, rhythmic expression of clock and per4 in both rx3 alleles is equivalent to that of their wild-type siblings. Thus, the cell cycle defects in the strong allele cannot be explained by a global deregulation of the circadian clock.
Could the attenuated cell cycle rhythms of the strong rx3 mutants be attributed to the disruption of a systemic pathway conveying circadian timing information to the cell cycle? Nocturnal production of the hormone melatonin by the pineal gland is a key central clock output [30]. Recently, melatonin has also been implicated in regulating cell proliferation in larval zebrafish [31]. Pineal mRNA expression of the rate-limiting enzyme of the melatonin synthesis pathway, arylalkylamine N-acetyltransferase (AANAT), is under circadian regulation, with expression high during the night and low during the day ([32] and Figure 2B, upper row). This expression pattern has been used widely as a reliable indicator of the levels of melatonin synthesis in the zebrafish [32–35]. The whole-mount in situ hybridization for aanat2 in strong rx3 mutant larvae shown in Figure 2B (lower row) reveals a circadian expression rhythm indistinguishable from wild-type siblings. Thus, the attenuated cell cycle rhythms in strong rx3 larvae seem unlikely to be explained by defects in this endocrine pathway. As an additional test of the contribution of melatonin, we treated wild-type larvae with luzindole, a specific antagonist of the MT1 and MT2 high-affinity melatonin receptors [31,36]. Cell cycle rhythms were not significantly affected by this treatment, confirming the hypothesis that melatonin production is not required for establishing cell cycle rhythms (Figure 2C). In lower vertebrates, the pineal is a directly photosensitive structure, and light directly affects the production of melatonin [2]. Therefore, our results also indicate that direct pineal photoreception is unlikely to contribute to circadian cell cycle rhythms.
The hypothalamic–pituitary axis is another endocrine pathway with a crucial role in the control of cell proliferation that shows circadian variations of activity [37]. We examined expression of a set of specific pituitary cell-lineage markers in the rx3 mutants: The transcription factor pit1 [38], growth hormone, gh [39], prolactin, prl [39], and glycoprotein hormone alpha subunit, α-gsu [38]. Expression of these markers is equivalent in rx3 mutants of both alleles when compared with their wild-type siblings (Figure 3A). Thus, the somatotrope (gh), lactotrope (prl), and gonadotrope/thyrotrope (α-gsu) lineages appear to be normally formed in the strong rx3 mutants. However, for the corticotrope/melanotrope lineage marker proopiomelanocortin (pomc, [39]), two expression domains show a marked reduction in strong allele rx3 mutants. The anterior pituitary domain is strongly reduced (arrowhead), and the expression corresponding to the β-endorphin/MSHα synthesizing cells of the arcuate nucleus ([40], arrow) is essentially absent, whereas the posterior pituitary expression domain (asterisk) appears normal. All these domains have a wild-type–like appearance in the weak allele mutant larvae (Figure 3A).
To test whether these differences reflect a general disorganization of the diencephalon, we examined the expression of a number of hypothalamic markers (somatostatin3, isotocin, and corticotropin releasing factor; for details, see Figure S2). The structures labeled by these markers are present in both mutant alleles and show no major disruption, despite the lack of normal eyecups. Thus, the strong rx3 mutation specifically seems to affect the pomc-expressing cells in the anterior pituitary and the arcuate nucleus. Previous studies have established that the anterior pituitary expression domain of pomc consists mainly of cells of the corticotrope lineage, whereas the posterior domain also contains melanotrope cells [41]. Furthermore, the arcuate nucleus has been implicated in regulation of the corticotrope axis in mammals [42]. Thus, the reduced number of corticotrope cells in the strong rx3 mutants might additionally lack normal hypothalamic control. These findings implicate the corticotrope lineage in circadian cell cycle regulation.
Given the pituitary defect in the strong rx3 mutant, we asked whether disruption of the hypothalamic–pituitary axis would cause similar circadian cell cycle defects to those seen in the strong rx3 mutants. To address this issue, we examined rhythms of BrdU incorporation in a series of zebrafish mutants that lack either the entire pituitary or specific subsets of pituitary lineages (Figure 3B–3E) [38,40,43–46]. The fibroblast growth factor 3 mutant lia/fgf3 (two alleles, [43]) and the proneural basic Helix-Loop-Helix transcription factor achaete scute-complex like 1a mutant pia/ascl1a [46], which lack the entire pituitary, show severely attenuated rhythms (Figure 3B and unpublished data). Thus, genetic ablation of the pituitary creates a circadian cell cycle phenotype highly similar to that observed for the strong rx3 mutant. Since lia and pia mutants show normal pomc expression in the arcuate nucleus, we can also exclude a non-pituitary–mediated contribution of β-endorphin/MSHα–expressing arcuate nucleus neurones to cell cycle rhythm generation.
To pinpoint the precise pituitary cell type responsible for the establishment of normal circadian cell cycle rhythms, we examined BrdU incorporation rhythms in two other pituitary mutants that lack subsets of the pituitary lineages (Figure 3E): The protein tyrosine phosphatase eyes absent 1 mutant aal/eya1 [44,45], which possesses only the lactotropes, and the POU-domain transcription factor pit1 mutant [38], in which only the corticotropes/melanotropes and the gonadotropes are present. The aal mutants show a similar phenotype to the lia, pia, and the strong rx3 mutants (Figure 3C), demonstrating that the lactotrope lineage alone is not sufficient for establishing circadian cell cycle rhythms. In contrast, the pit1 mutants are indistinguishable from their wild-type siblings (Figure 3D). Thus, the presence of only the corticotrope/melanotrope and gonadotrope lineages is sufficient to establish wild-type circadian cell cycle rhythmicity. Together with the reduced number of corticotropes observed in the rx3 mutant embryos, this result strongly suggests that the corticotropes are required for the establishment of the circadian cell cycle rhythms.
The principal target organ of signaling by the corticotrope axis is the medulla of the adrenal gland (interrenal gland in fish [47]), where it regulates production of glucocorticoids such as cortisol. To explore the mechanism of the cell cycle defect in the strong rx3 mutants, we measured cortisol levels in 6-d-old mutant and wild-type sibling larvae of both alleles raised under a LD cycle (Figure 4A). All larvae tested show higher cortisol levels at zeitgeber time (ZT)17 than at ZT1. However, mutant larvae of the strong allele have significantly lower levels (p < 0.0001) than all other larvae at both time points. Thus, the reduction of corticotrope cells in the strong allele mutant pituitary seems to strongly reduce cortisol levels, pointing to cortisol as a candidate systemic signal required for circadian cell cycle rhythmicity.
If cortisol is indeed the systemic signal, it should be possible to rescue circadian cell cycle rhythms by artificially stimulating glucocorticoid signaling in the strong rx3 mutants. Mutant larvae and wild-type siblings were raised in the presence of the potent glucocorticoid agonist dexamethasone. We then measured BrdU incorporation at four time points on day 6 of development. Strikingly, dexamethasone treatment fully restores high-amplitude circadian cell cycle rhythms in the mutants (Figure 4B). Non-treated control mutant larvae show the typical severely attenuated rhythms (Figure 4C). Similarly, aal mutant larvae treated with dexamethasone are indistinguishable from their wild-type siblings (Figure 4D and 4E). In conclusion, tonically activating glucocorticoid signaling during the early days of development can rescue the circadian cell cycle rhythms in cortisol deficient larvae.
Our previous work has shown that the diurnal cell cycle rhythms of zebrafish larvae are under control of the circadian clock, because these rhythms persist upon transfer into constant darkness [7]. We therefore asked whether the rescue effect of dexamethasone treatment could also operate without direct light input. Tonic dexamethasone application could indeed rescue BrdU incorporation rhythms in mutant larvae that were transferred to DD after 5 d of entrainment under a LD cycle (Figure S3), clearly showing that the rescue is due to interaction of glucocorticoids with the circadian clock and not dependent on direct light input.
We wished to explore in more detail how cortisol affects cell cycle rhythmicity. We first tested the hypothesis that the circadian clock might regulate expression levels of the glucocorticoid receptor gene (GR) and thereby confer a circadian rhythm of sensitivity to the receptor ligand. Such a mechanism would enable even tonic levels of the ligand to activate GR signaling pathways with a circadian rhythm. Furthermore, recent reports have highlighted that many members of the nuclear receptor superfamily show circadian cycling of transcript levels [48]. We thus prepared a time course of total RNA and protein extracts from wild-type sibling larvae during one LD cycle. Quantitative real-time PCR analysis failed to detect any significant change in GR mRNA levels during the course of the LD cycle (Figure 5A). Consistently, levels of an 82-kDa immunoreactive protein corresponding to the zebrafish GR also did not cycle as determined by Western blotting analysis (Figure 5B). These results indicate that the circadian clock does not simply affect global expression levels of the GR.
Many studies have documented the functional complexity of the glucocorticoid signaling pathway in vivo [49]. Lack of cortisol during development and larval growth might generally alter larval physiology and thereby also indirectly affect cell cycle rhythms. Rescue of high-amplitude cell cycle rhythms in rx3 mutant larvae by continuous exposure to dexamethasone from early development onwards could act via rescuing developmental defects as well as by affecting cells more directly. In order to address this point, we systematically tested the effect of reducing the duration of exposure to dexamethasone on cell cycle rhythms. We supplemented the medium with dexamethasone at progressively later stages before harvesting on day 6 of larval development (Figure 5C and unpublished data). Addition of dexamethasone as late as the night before sampling (day 5) still resulted in a significant increase in cell cycle amplitude. Dexamethasone delivered at later time points failed to rescue the rhythm. Given that all the major organ systems have developed and are functional at this freely feeding larval stage [50], this would exclude a major role for indirect developmental mechanisms.
The zebrafish represents an attractive model to explore how circadian clock outputs are regulated at the whole-animal level. We have previously implicated a contribution of directly light-entrainable peripheral clocks in the cell-autonomous control of circadian rhythms of S phase. Here, by studying panels of zebrafish mutants affecting development of the eye and hypothalamic–pituitary axis, we have been able to define the regulatory contribution of these structures to establishing circadian cell cycle rhythms. This study illustrates the power of using complementary sets of zebrafish mutants in the genetic dissection of physiological pathways in vivo.
We have shown that ocular photoreception is dispensable for establishing this clock output function, potentially reinforcing the notion that cell-autonomous light sensing plays a key role in cell cycle entrainment in the zebrafish. However, in common with most lower vertebrates, zebrafish possess additional extraocular specialized photoreceptor tissues: the pineal complex [2] and also the so-called deep brain photoreceptors that line the third ventricle of the diencephalon [51,52]. Since no zebrafish mutants are available to date that specifically lack these photoreceptors, it is problematic to assess their contribution. We used an alternative pharmacological approach to interfere with the melatonin signal, the major output of the pineal gland, and thereby tested whether photoreception through the pineal complex might affect cell cycle rhythms. Because treatment of larvae with the melatonin receptor antagonist luzindole did not change circadian cell cycle rhythms, pineal light reception is not strictly required for the timing of circadian cell cycle progression. However, it is still conceivable that light input from the pineal conveyed via neuronal pathways may contribute to the timing of the cell cycle [53–55]. Also, one type of dedicated photoreceptor might be able to substitute for lack of input from the other types, leading to functional redundancy of inputs from, e.g., the eye and the pineal complex. Finally, the direct peripheral light reception alone might also be sufficient to time this clock output in the context of the whole animal. Ultimately, mutant zebrafish that lack all specialized photoreceptor cells will be required to assess the relative contribution of directly light-sensing peripheral clocks to entraining the circadian cell cycle.
By studying pituitary mutants with overlapping cell-lineage defects, and by our demonstration that pituitary corticotropes are severely reduced in strong allele rx3 mutant larvae, we have implicated this pituitary cell lineage in the establishment of circadian cell cycle rhythms. Furthermore, levels of cortisol in strong rx3 mutants are reduced and, importantly, a GR agonist rescues the circadian cell cycle defects in both strong rx3 and pituitary mutants. These findings point to glucocorticoids as a requirement for high-amplitude cell cycle rhythms.
Previously, glucocorticoids have been implicated in the entrainment of peripheral circadian clocks in mammals. Injection of dexamethasone into mice can reset the phase of peripheral oscillators [56]. However, mice lacking the GR show normal clock gene expression rhythms in the liver [56]. Furthermore, here we observe normal circadian clock gene cycling in the strong rx3 mutant larvae (Figure 2A), despite their low levels of cortisol. Consistently, also in the aal mutants, per4 exhibits normal circadian rhythms of expression (Figure S4). Finally, zebrafish cells grown in cortisol-depleted medium still show normal clock expression under a LD cycle (Figure S5B, see below). Thus, glucocorticoid signaling is not an absolute requirement for the regulation of the peripheral pacemakers themselves. Our data directly implicate glucocorticoids in the regulation of clock output.
Glucocorticoids are well known to influence cell proliferation, both in vitro and in vivo. For example, dexamethasone can stimulate myoblast proliferation [57] and enhance the mitogenic response of fibroblasts to epidermal growth factor (EGF) [58], whereas it inhibits cell division in a lymphosarcoma cell line [59]. In addition, at the whole-animal level, corticosteroids have been reported to affect the capacity of the liver to regenerate after hepatectomy in rats [60]. Our results indicate that part of the effect of corticosteroids on cell proliferation might be brought about by cooperation with circadian clock output pathways.
At which level of organization might this interaction occur? Given the wealth of physiological targets of glucocorticoids, they might function through indirect, systemic pathways. For example, in zebrafish larvae, they might be involved in the maturation of organ systems and their physiological functions during development. However, our finding that dexamethasone treatment starting 10 h before sampling still rescues cell cycle rhythms in rx3 mutants makes a long-term effect on development unlikely. Nevertheless, other indirect systemic pathways might act within this time frame. For example, glucocorticoids could stimulate the release of mitogens from neighboring tissues or influence the release of other hormones (e.g., see [61–63]). Alternatively, they might act directly at the level of the proliferating cells themselves. We tested a potential cell-autonomous role for glucocorticoids in circadian cell cycle rhythms by examining zebrafish cell cultures grown in charcoal-treated (and thereby steroid-depleted) medium (Figure S5). Whereas circadian clock gene expression is normal in these cultures (Figure S5B), circadian cell cycle rhythms are severely attenuated (Figure S5A), thus mimicking the situation in cortisol-deficient larvae. However, in this cell culture system, treatment with dexamethasone does not rescue the attenuation (unpublished data). One explanation for this result could be that glucocorticoids might need to act synergistically with other substances that are also depleted from the medium by charcoal treatment. Taken together, although there are some hints for a direct cell-autonomous action of glucocorticoids on circadian cell cycle rhythms, more indirect systemic effects certainly cannot be excluded. Indeed, given the multifaceted actions of glucocorticoids, confinement of their effects on cell cycle progression to a single level would be rather surprising. In the animal, glucocorticoids might help to create a systemic signaling environment, within which they could also exert more direct effects.
What is the relative importance of systemic and peripheral circadian clock mechanisms in glucocorticoid-mediated circadian cell cycle regulation? Circadian rhythms of circulating glucocorticoids have long been recognized as a clock output in various vertebrates [37]. However, here we show that the role played by glucocorticoids in circadian cell cycle rhythms does not necessarily involve conveying timing information via changes in circulating levels. Rather, a constant glucocorticoid signal can rescue rhythms of cell cycle in cortisol-deficient mutant larvae. In the animal, the timing information might stem from another cycling systemic signal, or it might be provided by peripheral circadian clocks. Circadian changes in glucocorticoid levels might then reinforce the peripheral timing information in cell cycle control, or they might be required for other physiological functions.
How can a constant glucocorticoid signal lead to a rhythmic output? We propose a working model in which a certain level of glucocorticoid signaling may be permissive (Figure 6A, green arrow) for peripheral circadian clock regulation of genes involved in cell proliferation (Figure 6A, red arrow). Alternatively, the clock could gate responsiveness to the glucocorticoid signal, which would then act to regulate cell proliferation (Figure 6A, blue arrow). In either scenario, in corticotrope-deficient mutants, GR signaling is attenuated, and then peripheral clock input alone would be insufficient to generate full cell cycle rhythms (Figure 6B). Furthermore, in the rescue experiments, tonic dexamethasone treatment activates GR signaling and would restore cell cycle rhythms in the mutants with timing information being provided by the peripheral circadian clock (Figure 6C). In this model, cycling cortisol levels are not required for, but they might reinforce cell cycle rhythmicity.
An attractive mechanism for gating responsiveness to glucocorticoids would involve circadian clock regulation of expression of the GR itself. Expression of many nuclear receptor transcripts is under circadian clock control in different peripheral tissues [48]. However, here we show that the levels of the GR transcript as well as the protein are not subject to significant day–night variation. Thus, a simple scenario in which transcription of the GR gene is a regulatory target of peripheral clock components, e.g., via E box elements, appears unlikely. One can speculate that if indeed the receptor is under clock control, then this may occur at the post-translational level. Alternatively, circadian clock control might operate at various levels on other elements of the glucocorticoid signaling pathway, perhaps downstream of the receptor itself [64,65].
Glucocorticoids can increase the amplitude of cell cycle rhythms in strong rx3 mutants even when they are delivered the night before assaying BrdU incorporation. Specifically, they have to be present at least before ZT17, or 16 h before the peak of BrdU incorporation at ZT9 is reached. Adding dexamethasone at ZT21 (12 h before the expected peak) is not sufficient. This might reflect either a minimum time needed to exert downstream effects on the cell cycle, or a requirement for the presence of glucocorticoids at a certain time of day. Interestingly, the last time point with rescue capability coincides with the natural peak in cortisol levels in wild-type larvae (Figure 4A). Experiments involving a precisely controlled temporal activation and inactivation of the glucocorticoid signaling pathway will be needed to decide between these possibilities.
In summary, our results call for a re-evaluation of existing models that account for control of circadian cell cycle timing purely via the direct regulation of gene expression by peripheral clock components. We reveal a requirement for endocrine regulation involving glucocorticoids that operates downstream of the clock mechanism itself. It is tempting to speculate that many other clock outputs may involve similar contributions from cell-autonomous and systemic control elements.
Fish were raised and bred according to standard procedures [50]. RNase protection analysis was carried out as described previously [27], and the per4, clock1, and β-actin probes have been described [27,28,66]. The raising of larvae under controlled lighting and temperature conditions, and the BrdU labeling procedure have been described in [7]. Briefly, ten mutant and ten wild-type sibling larvae each were sorted into cell culture flasks on day 2 post-fertilization and raised at 25 °C under a 12-h light:12-h dark cycle until day 6 of development. Three hours after “lights on” (ZT = 3), the larvae of one flask were incubated for 20 min with BrdU before fixation in 4% paraformaldehyde/PBS, and this procedure was repeated at three additional time points at 6-h intervals. The cell culture BrdU incorporation experiments (AB.9 cells, ATCC) were carried out as described in [7]. Cells were raised in L15 medium supplemented with gentamycin, streptomycin, penicillin [16], and either 15% fetal bovine serum (FBS) or 15% charcoal-treated FBS (both Biochrom, Berlin, Germany).
The pituitary and lak mutants are not morphologically distinguishable from wild-type siblings early in development. Thus, two flasks with 25 larvae each from a cross of heterozygous carriers were sorted and processed as described above to ensure the presence of a sufficient number of mutants per time point. lak mutant larvae can be recognized by their expanded melanophore phenotype later in development. To identify the pituitary mutants, larvae were stained for growth hormone expression by whole-mount in situ hybridization (see below) and then stained for BrdU incorporation.
Dexamethasone treatment was carried out essentially as described by [41]. Dexamethasone (Sigma, St. Louis, Missouri, United States) was dissolved in distilled water at 1 mM as a stock solution, then diluted further in E3 medium to a final concentration of 25 μM. Ten strong allele rx3 mutants and wild-type siblings were sorted into each cell culture flask on day 2 of development and raised in the presence or absence of dexamethasone until BrdU labeling on day 6 as described above. Luzindole treatments were performed similarly, with a stock solution of 0.01 M luzindole (Sigma) in ethanol diluted further in E3 to a final concentration of 0.00001 M [31]. Twenty larvae were raised in 25 ml of E3 at this luzindole concentration from day 2 of development, and on day 4, an additional 30 μl of luzindole stock solution were added to compensate for potential degradation of the compound. Control larvae were treated with equivalent solvent (ethanol) concentrations only.
In situ hybridization was carried out essentially as described [50], with the following modifications: The 4% paraformaldehyde fixation step after rehydration was omitted. Larvae were washed twice in PBS+0.1% Tween-20 (PBST), then rinsed for 2 min in distilled water, incubated for 7 min in pre-cooled acetone at −20 °C, passed through distilled water for 2 min at room temperature, and washed 3×5 min in PBST. Then, larvae were digested with 1 mg/ml of collagenase P (Roche, Basel, Switzerland) in PBS with 1% BSA and 1% DMSO at room temperature for 45 min, before fixation for 20 min at 4 °C in 4% paraformaldehyde. After five washes with PBST, larvae were prehybridized with HYB buffer for 4–6 h, then incubated with the probes overnight. After antibody incubation, larvae were washed 6 × 15 min in PBST, then staining was developed for several hours at room temperature and up to overnight at 4 °C. To remove pigmentation, larvae were incubated in 3 ml of 10% H2O2/methanol overnight, followed by addition of 12 ml of PBST, incubation overnight, and several washes with PBST. Probes used were aanat2 [32], pit1 [38], gh [39], prl [39], α-gsu [38], pomc [38], isotocin [67], ppss3 [68], and per4 [28]. corticotrophin relasing factor (crf ) was cloned in our laboratory.
Total RNA was extracted from triplicate samples using Trizol RNA isolation reagent (GIBCO-BRL, San Diego, California, United States) according to the manufacturer's instructions. The RNA (3 μg) was reverse-transcribed using Oligo(dT) primer (Amersham Biosciences, Little Chalfont, United Kingdom) and SuperScript III reverse transcriptase (Invitrogen, Carlsbad, California, United States). per4 mRNA levels were determined by real-time PCR using the DNA Engine Opticon thermocycler (Bio-Rad, Hercules, California, United States) following the manufacturer's instructions. First-strand cDNA aliquots from each sample were diluted 20× and served as templates in a PCR consisting of master mix, SYBR Green I fluorescent dye (Bio-Rad), and 400 nM gene-specific primers. Copy numbers were normalized using β-actin controls. Primer sequences were per4: 5′-CCGTCAGTTTCGCTTTTCTC-3′ and 5′-ATGTGCAGGCTGTAGATCCC-3′; glucocorticoid receptor: 5′-CGGACAGAGCTTCCTCTTTG-3′ and 5′-CTGCTGCATTCCACTGACAT-3′; and β-actin: 5′-TCCTGCTTGCTAATCCAC-3′ and 5′-ACCACCTTCAACTCCATC-3′.
Protein extracts were prepared by homogenizing 20 larvae per time point in 100 μl of Laemmli buffer. A total of 10 μl of the homogenate was loaded on a 6% SDS polyacrylamide gel, then Western blotting (BioRad) was carried out using an anti-human glucocorticoid receptor α (sc-1002; Santa Cruz Biotechnology, Santa Cruz, California, United States) and anti-mouse CREB antibodies (Upstate Biotechnology, Billerica, Massachusetts, United States), and visualized with the ECL detection system (Amersham Biosciences).
Twenty-five larvae each were raised in cell culture flasks with 20 ml of E3 medium under a LD cycle and a constant temperature of 25.3 °C. At each time point, larvae were rapidly transferred to 2-ml Eppendorf tubes, the medium was removed, and the larvae were snap frozen in liquid nitrogen and stored at −80 °C until further processing. For each time point and condition, three identical flasks were used per experiment. Larvae were homogenized on ice with a microgrinder (Eppendorf, Hamburg, Germany), then extracted with 500 μl of cold ethanol. After centrifugation for 10 min at 3,000 rpm at 4 °C, the supernatant was recovered and evaporated in a SpeedVac. The resulting pellet was resuspended in 20 μl of standard A buffer of the IBL cortisol LIA kit (IBL, Hamburg, Germany) and measured following the instructions provided by the supplier. The LIA plate was read by a VICTOR light 1420 luminometer (Wallac/Perkin Elmer, Wellesley, Massachusetts, United States), and the raw data were analyzed using the MikroWin2000 program, version 4.23 (Mikrotek Laborsysteme, Overath, Germany).
Statistical analysis was performed using the GraphPad Prism version 4.00 for Windows (Graph Pad Software, http://www.graphpad.com).
The GenBank accession numbers (http://www.ncbi.nlm.nih.gov/Genbank) for the genes discussed in this paper are α-gsu (NM_205687), aal/eya1 (NM_131193), aanat2 (NM_131411), atonal homologue 5 (AB049457), clock (NM_130957), corticotropin releasing factor (DQ250674), glucocorticoid receptor (NM_001020711), lia/fgf3 (NM_131291), isotocin (AY069956), per4 (NM_212439), pia/ascl1a (NM_131219), pit1 (AY421970), pomc (NM_181438), prl (NM_181437), retinal homeobox gene 3 (NM_131227), and somatostatin3 (BI472739 and BI473045).
|
10.1371/journal.pbio.0050067 | Arkadia Enhances Nodal/TGF-β Signaling by Coupling Phospho-Smad2/3 Activity and Turnover | Regulation of transforming growth factor-β (TGF-β) signaling is critical in vertebrate development, as several members of the TGF-β family have been shown to act as morphogens, controlling a variety of cell fate decisions depending on concentration. Little is known about the role of intracellular regulation of the TGF-β pathway in development. E3 ubiquitin ligases target specific protein substrates for proteasome-mediated degradation, and several are implicated in signaling. We have shown that Arkadia, a nuclear RING-domain E3 ubiquitin ligase, is essential for a subset of Nodal functions in the embryo, but the molecular mechanism of its action in embryonic cells had not been addressed. Here, we find that Arkadia facilitates Nodal signaling broadly in the embryo, and that it is indispensable for cell fates that depend on maximum signaling. Loss of Arkadia in embryonic cells causes nuclear accumulation of phospho-Smad2/3 (P-Smad2/3), the effectors of Nodal signaling; however, these must be repressed or hypoactive as the expression of their direct target genes is reduced or lost. Molecular and functional analysis shows that Arkadia interacts with and ubiquitinates P-Smad2/3 causing their degradation, and that this is via the same domains required for enhancing their activity. Consistent with this dual function, introduction of Arkadia in homozygous null (−/−) embryonic stem cells activates the accumulated and hypoactive P-Smad2/3 at the expense of their abundance. Arkadia−/− cells, like Smad2−/− cells, cannot form foregut and prechordal plate in chimeras, confirming this functional interaction in vivo. As Arkadia overexpression never represses, and in some cells enhances signaling, the degradation of P-Smad2/3 by Arkadia cannot occur prior to their activation in the nucleus. Therefore, Arkadia provides a mechanism for signaling termination at the end of the cascade by coupling degradation of P-Smad2/3 with the activation of target gene transcription. This mechanism can account for achieving efficient and maximum Nodal signaling during embryogenesis and for rapid resetting of target gene promoters allowing cells to respond to dynamic changes in extracellular signals.
| In development, cells respond to secreted signals (called morphogens) by turning on or off sets of target genes. How does gene activity adjust quickly in response to rapidly changing extracellular signals? This should require efficient removal of old/used signaling effectors (signal-activated transcription factors) from the promoters of target genes to allow new ones to assume control. We previously discovered Arkadia, an E3 ubiquitin ligase, and showed that it is an essential factor for normal development. (Ubiquitin ligases trigger the addition of ubiquitin residues to proteins, typically marking them for degradation.) Here, we show that Arkadia is required for high activity of the major signaling pathway, TGF-β/Nodal. Arkadia has a dual role to degrade Smads, the TGF-β signaling effectors, and enhance their transcriptional activity. This coupling of degradation with activation provides a mechanism to ensure that only effectors “in use” are degraded, allowing the new ones to proceed. It is possible that very similar mechanisms operate in other pathways to establish dynamic regulation and efficient signaling, while their failure may be associated with developmental abnormalities and disease, including cancer.
| Transforming growth factor-β (TGF-β) signaling controls a diverse set of cellular processes, including cell proliferation, differentiation, apoptosis, and specification of fate in vertebrate and invertebrate species. Disruption of signaling leads to developmental abnormalities and disease, including cancer. Activin and Nodal TGF-β ligands have been shown to act as morphogens in vertebrate development [1–4]. For example, in the mouse, Nodal is required for gastrulation, including development of the anterior primitive streak and the formation of the germ layers, endoderm and mesoderm [5,6]; for maintenance of pluripotency in the epiblast [7,8]; and for the specification of the anterior-posterior [9,10] and left-right axes [11]. Loss-of-function mutations in the Nodal gene, including enhancer deletions, lead to a reduction of Nodal RNA [12] and reveal that the highest level of Nodal signaling is required during gastrulation for the induction of the anterior primitive streak. This contains the precursors of the mammalian equivalent of the amphibian Spemann's organizer, and it gives rise to the anterior endoderm, the node, and the mesendoderm (notochord and prechordal plate), all of which are required for subsequent patterning of the vertebrate embryo [6]. Complementary experiments in Xenopus embryos, where increasing amounts of Nodal RNA are injected, show that it functions as a dose-dependent inducer and that the highest level induces Spemann's organizer [13]. The dynamic changes in the concentration of ligands, to elicit different cellular responses, demand that the responding cells have rapid turnover of the signaling-effectors and frequent refreshing of target gene promoters. Therefore, how TGF-β is regulated, and particularly, how signaling is terminated in the nucleus after gene transcription, is key in understanding cell fate decisions and patterning in vertebrate development.
TGF-β signals bind to cognate serine/threonine kinase receptors leading to phosphorylation and activation of the Smad family of signal transducers. Two different Smad signaling branches have been described. Ligands, like Activin, Nodal, Gdf1, Vg1, and TGFβ1 are transduced by the receptor-activated Smad2 and Smad3 (Smad2/3) [14,15]. The phosphorylated form of Smads (phospho-Smads [P-Smads]) complex with Smad4 and together translocate into the nucleus, where they function as transcription factors in association with DNA-binding partners such as FoxH1, Mixer, Jun/Fos, Runx, ATF3, and E2F4/5, etc., which provide target gene specificity [15]. In the mouse, loss-of-function mutations affecting core components of the Nodal signal transduction pathway give patterning and cell fate defects similar to that of Nodal itself [16–20].
Extracellular cofactors [21], antagonists [22,23], and proteases [24,25] have been shown to regulate Nodal activity during mouse development. However, little is known about the role of intracellular regulation in cells receiving Nodal. Intracellular regulators of the pathway include negative regulators such as inhibitory Smads (Smad6/7) that block TGF-β signaling by competing with Smads for association with the receptors or by targeting receptors for ubiquitin-mediated degradation [26–28]; in the nucleus, Ski and SnoN incorporate in the Smad DNA-binding complex to prevent them from binding to the transcriptional coactivator p300/CBP and repress transcription by recruiting histone deacetylase [29,30]. More recently, a phosphatase, PPM1A/PP2Ca, has been identified and is shown to de-phosphorylate P-Smad2/3 [31] and abrogate their signaling activity. Furthermore, proteasome-mediated degradation of ubiquitin-modified core components of the TGF-β signaling cascade has been shown to play a major role in controlling signaling output [32]. Poly-ubiquitination and proteasome-dependent degradation of proteins is one of the most prominent turnover mechanisms in the cell. Ubiquitination of protein substrates involves a cascade of enzymatic reactions. E3 ubiquitin ligases are the critical components responsible for the recognition of specific substrates for ubiquitination [33]. They are generally classified into the HECT- and RING-domain classes and exhibit substrate specificity [34,35]. Several ubiquitin ligases are known to reduce signaling by mediating the degradation of individual components of the pathway [28,36]. However, ligases that terminate signaling by degrading activated Smads (P-Smads) have not been identified.
One of the important unanswered questions is how long the activated Smads transcribe target genes and how the promoters are refreshed to allow rapid intracellular responses to dynamic changes in concentration of ligands. Both de-phosphorylation of P-Smads followed by cytoplasmic recycling [37] and proteasome-mediated degradation of P-Smads have been proposed to terminate signaling in the nucleus [38]. However, there was no explanation for how “used” versus “unused” activated effectors could be distinguished by these mechanisms. An obvious mechanism to limit the time that an effector works is to link its turnover with its ability to drive transcription.
We have shown previously that Arkadia, a nuclear RING-domain ubiquitin ligase, enhances Nodal signaling and is essential for the induction of the organizer/node [39,40]. In somatic tumor cell lines, Arkadia has been shown to enhance TGF-β signaling by ubiquitin-mediated degradation of Smad6/7 [41,42]. However, we show here that Arkadia functions by a different mechanism in embryonic cells. Specifically, we find that Arkadia directly ubiquitinates and degrades P-Smad2/3 and that this is coupled with their high activity. The link between activity and degradation provides a mechanism to ensure that only “used” P-Smad2/3 effectors are degraded and that signaling is terminated at the end of the cascade and not before. Therefore, Arkadia can account for rapid resetting of actively transcribed promoters, forcing transcription to rely only on fresh P-Smads and allowing cells to respond to dynamic changes in signaling. Similar mechanisms may operate in other signaling pathways in development to achieve peak efficiency and dynamic responses of cells during development.
The phenotype of Arkadia−/− embryos consists of loss of anterior primitive streak derivatives (node, notochord, prechordal plate, and anterior definitive endoderm [ADE]/foregut) leading to anterior patterning defects including head truncations [39,40]. We have shown previously that while Arkadia or Nodal heterozygous (+/−) mice are normal, a small number of double heterozygotes for Arkadia and Nodal recapitulate the Arkadia−/− phenotype [39,40]. This suggested a functional interaction between Arkadia and Nodal. To investigate the extent of Arkadia's role in Nodal signaling, and in additional Nodal-dependent developmental events, we generated Arkadia−/− embryos with only one wild-type copy of the Nodal gene (Akd−/−, Nodal+/−). The majority of Akd−/−, Nodal+/− embryos that were analyzed (n = 19/33) exhibited phenotypes never observed in Arkadia−/− embryos. Using whole mount in situ hybridization, we performed marker analysis to define whether these phenotypes are Nodal-dependent (Figure 1).
Before gastrulation, Nodal signaling is responsible for anterior-posterior axis specification via the induction of the anterior visceral endoderm (AVE) domain at the distal tip of the mouse embryo and its migration to the prospective anterior [2,43]. Expression analysis of the AVE markers Hex [44] and Cerl [45] demonstrates that the Arkadia−/− embryos are always able to induce an AVE, which correctly migrates to the prospective anterior of the embryo (compare Figure 1A and 1D to 1B and 1E), while the primitive streak marker Brachyury is normally expressed posteriorly (Figure 1A and 1B). However, Akd−/−, Nodal+/− embryos have incomplete AVE-specific gene expression, as they are rarely able to induce an AVE (one out of seven) that expresses Hex (none out of two; Figure 1C), Cerl (one out of three; Figure 1F), or Lefty1 (none out of two; unpublished data). Furthermore, in the Akd−/−, Nodal+/− embryos, the Cerl-expressing embryo AVE domain remains distal (Figure 1F) and the Brachyury domain remains proximal (Figure 1C), indicating that when the AVE is induced it cannot migrate to the anterior. In addition, more than 50% of the Akd−/−, Nodal+/− embryos examined at mid-streak stage have a constriction between the embryonic and extraembryonic compartments. This phenotype, which is not observed in Arkadia−/− embryos, is thought to be due to a failure of the AVE to migrate and define the anterior-posterior axis, leading to failure of primitive streak elongation and mesoderm formation along the embryonic-extraembryonic boundary [46]. These new phenotypes in Arkadia−/− embryos carrying only one Nodal wild-type allele indicates that Arkadia is also involved in AVE formation and suggests that it enhances Nodal signaling at pre-gastrulation stages.
During gastrulation, Nodal is responsible for the formation and patterning of endoderm and mesoderm. Cardiac mesoderm is considered an anterior mesodermal tissue [47], and Arkadia−/− embryos form heart [39]. In contrast, all Akd−/−, Nodal+/− embryos fail to form morphologically visible heart (n = 4; Figure 1L), as it is also shown by loss of Nkx2.5 expression; one of the earliest markers of myocardial differentiation [48] (Figure 1G−1I), Shh expression (Figure 1J) marks the mesendoderm, the midline of the neural tube (floor plate), and the gut endoderm [49]. In Arkadia−/− embryos, Shh expression is reduced [39] due to the absence of mesendoderm and foregut, but it is present in the midgut and hindgut (Figure 1K). Akd−/−, Nodal+/− embryos, however, have a severe reduction in Shh expression indicating not only loss of mesendoderm but also of all endoderm (n = 2; Figure 1L), as confirmed by histological analysis (unpublished data). These new phenotypes seen in Arkadia−/− embryos carrying only one Nodal wild-type allele indicate that during gastrulation, Arkadia is involved in the formation of the entire endoderm and anterior mesoderm and suggest that this is mediated by its ability to enhance Nodal. As Arkadia facilitates Nodal signaling broadly, before and during gastrulation, it is likely to be a regular partner factor of the Nodal signal transduction pathway.
To find at what position within the Nodal signaling cascade Arkadia functions, we compared the level of the receptor-activated (phosphorylated) signaling effector P-Smad2 in embryos and embryonic stem (ES) cells by Western blotting (Figure 2). We examined 20 wild-type and 20 Arkadia−/− embryos at 8.5 days post-coitum (dpc) (Figure 2A and unpublished data) and three wild-type and three Arkadia−/− blastocyst-derived ES cell lines (Figure 2G). We found that while the total levels of Smad2 protein remain the same, P-Smad2 was always at least two times higher in all Arkadia−/− embryos and ES cells compared to the wild-type samples. Similarly, the other Nodal signaling effector P-Smad3 was found to be elevated in Arkadia−/−ES cell lines (Figure S1). Therefore, in embryonic cells, P-Smad2/3 are more abundant in the absence of Arkadia than in its presence. As the phosphorylation of Smad2/3 depends on the kinase activity of the ligand-activated receptors, the data suggest that in the absence of Arkadia the receptors are more active or that P-Smad2/3 are more stable after their phosphorylation.
We examined the activity of the receptors by stimulating Arkadia−/− and wild-type ES cells with Activin A ligand (Activin) and comparing Smad2 phosphorylation over time (Figure 2B–2D). Although Arkadia−/−cells start with higher basal levels of P-Smad2 compared to wild-type, Smad2 phosphorylation peaks 1 h after Activin addition in all ES cell lines (Figure 2B–2D), indicating normal receptor kinase activity in the presence or absence of Arkadia. Interestingly, after peak stimulation in wild-type ES cells, P-Smad2 decreases to basal levels within 2 h (Figure 2B–2D), but in Arkadia−/− ES cells P-Smad2 is maintained at peak levels (>90%) for at least 6 more h (Figure 2B). The data suggest that after receptor phosphorylation, Smad2 is more stable or maintains the phosphorylation longer in the absence of Arkadia. The total Smad2 levels do not change during the course of the experiment, indicating that the increased stability is associated with only the phosphorylated fraction (Figure 2C and 2D). To examine the possibility that other factors, such as a receptor inhibitor, is induced specifically in wild-type cells causing reduction of Smad2 phosphorulation, we repeated the above experiment in the presence of a protein synthesis inhibitor (cycloheximide). The decay of P-Smad2 was found, as before, to be slower in Arkadia−/− cells (Figure S2), suggesting that differences in protein stability rather than synthesis account for the increase in P-Smad2 levels.
To exclude the possibility that the receptors generate the differences in P-Smad2 levels, we blocked all the TGF-β receptors with the serine/threonine kinase inhibitor, H7 (Figure 2E and 2F) or used SB431542 selective inhibitor of Alk receptors (SB), which blocks the receptors that specifically phosphorylate Smad2/3 (Figure S1A). We found that in wild-type cells treated with two different concentrations (5 or 25 μM of H7), P-Smad2 levels declined 40% and 60%, respectively, within 30 min and they diminished to 30% after 90 min. In Arkadia−/− ES cells, however, even after 90 min, with the highest amount of inhibitor, P-Smad2 levels were not significantly changed (Figure 2F). The data indicate that the receptors are not responsible for generating the increase of P-Smad2 in Arkadia−/− cells and suggest that this is caused by P-Smad2 stabilization. The total Smad2 protein levels do not change during the course of the experiment (Figure 2E), but as the fraction of P-Smad2 is most likely small, differences within this fraction may not be visible when the total levels are examined. All of the above experiments were reproducible in three different Arkadia−/− ES cell lines (unpublished data) and the same results were obtained for the other Nodal/Activin effector, P-Smad3 (Figure S1B). Collectively, the data suggest that Arkadia acts downstream of the receptors and destabilizes the phosphorylated forms of Smad2/3.
P-Smad2/3 complex with Smad4 and translocate to the nucleus where they activate target genes and are subjected to different mechanisms of turnover and signaling termination such as ubiquitination/proteasome-mediated degradation [38] or de-phosphorylation [31] and nuclear export [50,51]. Cytoplasmic retention of P-Smad2/3 can result in both inability to activate target genes and increased stability. As in the absence of Arkadia, P-Smad2/3 are more stable and less transcriptionally efficient; it is possible that they are cytoplasmic and Arkadia regulates their nuclear localization. We therefore examined in three different wild-type and three Arkadia−/− ES cell lines the localization of P-Smads using Western blots of nuclear and cytoplasmic fractions (Figure 2G) or by immunofluorescence (Figure 2H) with antibodies against P-Smad2 or Smad2/3. We did not find evidence that P-Smads are cytoplasmic in the absence of Arkadia. On the contrary, the Western blots revealed that P-Smad2 accumulates in the nucleus of Arkadia−/−cells. Furthermore, as P-Smad2/3 complex with Smad4 to translocate to the nucleus [15], an increase of Smad4 in the nucleus of Arkadia−/− ES cells was observed (Figure 2G). Therefore, we conclude that Arkadia destabilizes P-Smad2/3 without affecting Smad4 complex formation and nuclear localization. Furthermore, as Arkadia is nuclear, the data suggest that Arkadia regulates P-Smad2/3 stability in the nucleus.
Arkadia is an E3 ubiquitin ligase and could be destabilizing P-Smad2/3 directly by poly-ubiquitinating them, leading to their proteasome-dependent degradation. To test this hypothesis, we examined whether Arkadia interacts specifically with phosphorylated Smads. We used HEK293T (293T) cells stably expressing moderate levels of full-length Arkadia, tagged either with Flag on the N-terminus and Myc on the C-terminus, or with green fluorescent protein (GFP) fused to the N-terminus (GAkd). We performed immunoprecipitation (IP) with anti-Flag (Figure 3A and 3B), -Myc, or -GFP (unpublished data) antibodies and Western blotted with anti-P-Smad2, -Smad2 (Figure 3A), or -P-Smad3 (Figure 3B). The results show that with Activin stimulation, Arkadia coIPs with the phosphorylated endogenous Smad2 (Figure 3A) and Smad3 (Figure 3B). However, when the cells are treated with the SB receptor inhibitor, which eliminates Smad2/3 phosphorylation, Arkadia does not coIP unphosphorylated Smad2/3 (Figure 3A and unpublished data). Furthermore, we examined the interaction of Arkadia with other phosphorylated Smads (Smad1/5/8; Figure S3A) or with Smad4 (unpublished data) and found no evidence of interaction. We therefore conclude that Arkadia interacts specifically with the phosphorylated forms of Smad2/3.
To test how direct this interaction is, we performed the IP in vitro (Figure 3C) using in vitro transcribed/translated (recombinant) Arkadia protein labeled with S35 and phosphorylated Flag-tagged Smad2 isolated by IP (Figure S3B) from 293T cells stimulated with constitutive active Alk4 (Alk4*). The data show that phosphorylated flag-Smad2 protein can IP recombinant full-length Arkadia but not the N-terminal portion (1–510 amino acids [aa]), or the luciferase control (Figure 3C). The data suggest that Arkadia interacts directly with P-Smad2/3 via a C-terminal domain.
Arkadia is a 989-aa protein and its C-terminal half (516–989 aa) contains a highly conserved domain of 100 aa (889–989 aa), which includes the RING-ubiquitin ligase activity-domain at the C-terminal region (947–965 aa), a nuclear localization signal (NLS) (903–909 aa), and a conserved domain (889–903 aa), termed here NRG, of unknown function (Figure S3C). To understand the interaction of Arkadia with P-Smad2/3, we mapped further the responsible domain. We used transient transfections of 293T cells to test the ability of various deletions of Arkadia (all GFP-tagged; Figure 3D) to IP P-Smad2/3. The data show that the last 100 aa of Arkadia containing the NRG, the NLS, and the RING (G-NRG-RING, 889–989 aa; Figure 3D) are sufficient for the interaction (Figure 3D). Deletion of the C-terminal end of Arkadia (GAkdR*) that eliminates one of the Zinc-binding fingers of the RING domain (965–989 aa) does not affect the interaction with P-Smad2/3 (Figure 3F). However, deletion of the NRG completely abrogates the interaction with P-Smad2 (Figure 3E). To confirm that the NRG domain is necessary for the interaction with P-Smad2/3 within the context of the full-length Arkadia protein, we generated an internal partial deletion of only the first eight residues of the NRG (GAkdNRG*) and showed that it diminishes the interaction with P-Smad2/3 (Figure 3F). As judged by fluorescence from the GFP tag, all of the above mutant Arkadia proteins are localized in the nucleus and are expressed at comparable levels to full-length Arkadia (Figure S4), indicating that loss of the interaction of the various Arkadia deletion constructs is not due to instability or differential localization. Collectively, the above data indicate that Arkadia interacts directly with P-Smad2/3 via its 100-aa C-terminal portion and that within this domain, a 14-aa NRG motif is essential for this interaction. As the 293T cells that we used for the IPs do not express FoxH1 (unpublished data), one of the major P-Smad2/3 transcription partners in early embryogenesis, we conclude that the interaction of Arkadia with P-Smad2/3 may not depend on a particular partner.
As P-Smad2/3 interact with Arkadia, it is possible that they are substrates of Arkadia ubiquitination. To address this, we examined the ubiquitination status of P-Smad2 in the presence or absence of Arkadia expression. We used Arkadia−/− mouse embryonic fibroblasts lines (MEFs) to exclude any endogenous Arkadia activity and introduced Flag-Smad2 and Alk4* to obtain phosphorylated Flag-Smad2, in the presence of full-length or mutant forms of Arkadia. Western blot analysis of the IPs with anti-P-Smad2 antibodies showed the existence of higher molecular weight forms of P-Smad2 associated specifically with the presence of full-length Arkadia (Figures 4A and 4B and S4) suggesting poly-ubiquitination. Probing with ubiquitin antibodies confirmed that these modifications contain ubiquitin chains (Figures 4A and 4B and S4). Furthermore, these blots show that P-Smad2 is not ubiquitinated in the presence of mutant Arkadia proteins lacking either ubiquitin ligase activity (GAkdR*) or the P-Smad2 interaction domain (GAkdNRG*; Figure 4A and 4B). Therefore, Arkadia ubiquitinates P-Smad2 in vivo, and this depends on both its ubiquitin ligase activity and the P-Smad2/3 interaction domain, suggesting that Arkadia ubiquitinates them directly.
To verify that P-Smad2/3 ubiquitination is directly dependent on Arkadia, we performed the assay in vitro by adding all the components of ubiquitination separately along with recombinant full-length Arkadia, the C-terminal part of Arkadia containing the RING and the NRG, or N-terminal Arkadia. In this reaction we added the phosphorylated Flag-Smad2 obtained by IP as shown before (Figure S3B), and examined by Western blot, with anti-P-Smad2 antibody, whether or not it becomes modified by the ubiquitination reaction. We found that the full-length and C-terminal Arkadia are capable of poly-ubiquitinating in vitro Flag-P-Smad2, only when all the ubiquitination components were present, while the N-terminal Arkadia does not (Figures 4C and S5B). In addition, the Flag-P-Smad2 substrate does not become ubiquitinated without the addition of recombinant Arkadia, indicating that the IP is not contaminated with ubiquitin ligases from the cells. Collectively, the data show that P-Smad2/3 are ubiquitinated directly by Arkadia in vivo and in vitro, and therefore, they are qualified substrates of Arkadia ubiquitination.
Poly-ubiquitination of proteins usually leads to degradation via the proteasome [34,35]. Consistent with this, our data show that P-Smad2/3 are unstable in the presence of Arkadia (Figure 2), but it was unknown whether this instability is mediated by the proteasome. To address this, we transfected Arkadia−/− MEFs with full-length Arkadia or Arkadia lacking ubiquitin ligase activity under ligand stimulation and examined with an anti-P-Smad2 antibody the stability of transfected P-Flag-Smad2 (Figure 4D) or endogenous P-Smad2 (Figure 4E) in the presence or absence of MG132 proteasome inhibitor. We found that P-Smad2 levels are reduced specifically in the presence of full-length Arkadia and that MG132 can inhibit this. Therefore, Arkadia, via its ubiquitin ligase activity, is sufficient to induce P-Smad2 proteasome-dependent degradation.
In the above experiments, the degradation of P-Smad2 was achieved by the transfection of exogenous Arkadia in −/− MEFs. To examine whether endogenous Arkadia is necessary for proteasome-dependent degradation of P-Smad2, we compared its decay in Arkadia−/− and wild-type ES cells in the presence or absence of MG132. For this we first stimulated the ES cells with Activin (1 h), then added SB inhibitor to prevent further phosphorylation of Smad2 and examined its decay at different time points (Figure 4F). We found that MG132 protects P-Smad2 in wild-type ES cells but has very little effect in Arkadia−/− (Figure 4F). Therefore, Arkadia, by direct poly-ubiquitination, most likely mediates degradation of P-Smad2 by the proteasome.
We showed above that loss of Arkadia leads to the stabilization and nuclear accumulation of P-Smad2/3. But do these higher levels correspond to an increase in target gene transcription? Analysis of Arkadia−/− and compound Arkadia−/−, Nodal+/− embryos showed that Nodal signaling is defective (Figure 1 and [39,40]) suggesting that Smad2/3 target gene transcription is compromised. We examined the transcriptional activity of P-Smad2/3 in three Arkadia−/− and three wild-type ES cell lines. To estimate the relative levels of P-Smad2/3 transcriptional activity, we used two different target gene luciferase reporters, 0.9-P1, (hereafter termed Pitx2-luc) regulated by P-Smad2/3 and its partner factor FoxH1, and 9xCAGA–luc, a Smad3 specific reporter [52,53]. Although the Arkadia−/− ES cell lines always have a higher amount of P-Smad2 protein compared to wild-type, they have on average 30% (Figure 5A) lower luciferase from the wild-type cell line with the lowest activity (WT 3 designated as reference = zero). Stimulation with Activin did not change significantly the luciferase reporter expression (unpublished data); indicating that under standard culture conditions ES cells exhibit ligand-saturated signaling (autocrine signaling). In addition, real-time PCR showed that the expression of the endogenous Nodal gene, which like the Pitx2-luc luciferase reporter is regulated by FoxH1/Smad2/3 binding sites (known as ASE) [12], is reduced by about 70% in Arkadia−/− ES cells (Figure 5B). Together, these observations suggest that Arkadia is necessary for efficient target gene expression and suggest that in its absence, the stable and high levels of P-Smad2/3 are hypoactive or prevented from activating their target genes. Therefore, although Arkadia degrades P-Smad2/3, it is necessary for efficient P-Smad2/3 transcriptional activity.
To address whether Arkadia is sufficient to activate P-Smad2/3, we performed gain-of-function experiments in ES cells. We introduced full-length Arkadia (GFP-tagged; GAkd) in three −/− ES cell lines and showed that it enhances the expression of the luciferase reporters on average 100% (Figure 5C) and 230% (Figure 5D) above the level of a GFP-expressing plasmid. Interestingly, in wild-type ES cells, Arkadia does not significantly change the reporter activity (Figure 5C and 5D) even after Activin stimulation (unpublished data). This indicates that endogenous Arkadia is adequate to activate all P-Smad2/3 generated by the receptors. Together, these results suggest that Arkadia is necessary and sufficient to enhance P-Smad2/3 target gene transcription.
According to the above data, Arkadia has two functions: to destabilize P-Smad2/3 and enhance their activity. As it is possible that different domains mediate the two opposing functions, we used the −/− ES cell functional assay to identify domains that are essential for Arkadia to enhance reporter activity. We found that Arkadia constructs with ubiquitin ligase domain mutations (GAkdR* and GAkdR2*) or Arkadia without the P-Smad2/3 interaction domain (NRG deletion, GAkdNRG) fail to enhance (Figure 5E); suggesting that like the degradation of P-Smad2/3, activation also requires a direct interaction with Arkadia and its ubiquitin ligase activity. Examination of several N-terminal deletions (Figure 3D) showed that the C-terminal half (516–989 aa) of Arkadia is the minimum region sufficient to enhance the reporter efficiently (unpublished data). Furthermore, loss of the P-Smad2/3 activation, but not the degradation properties of Arkadia, is expected to convert it to a repressor of signaling. However, none of the deletions and mutations of Arkadia separated the two functions. Together, the above data suggest that Arkadia activates and degrades P-Smad2/3 via the same domains and that the two functions are most likely coupled.
To test whether the activation of the hypoactive P-Smad2/3 by Arkadia occurs at the expense of their levels, we examined the relationship between levels of endogenous P-Smad2 and the degree of enhancement after expression of Arkadia in −/− ES cells. To visualize levels of endogenous P-Smad2, we isolated by fluorescence-activated cell sorting (FACS) pure populations of cells transfected with GAkd or enzymatically (ubiquitin ligase) inactive Arkadia (GAkdR*) or control GFP constructs. A portion of the cells was used for luciferase assays and the rest was used to examine endogenous P-Smad2 and total Smad2 levels in Western blots. As before, Arkadia enhances signaling only in −/− ES cells in a ubiquitin-ligase-dependent manner (Figure 6A); and this phenomenon was accompanied by an 80% reduction in the level of P-Smad2 (Figure 6B), confirming that activation of the hypoactive P-Smad2 is followed by its degradation.
We performed the same experiment in wild-type ES cells and found that overexpression of Arkadia does not change the levels of endogenous P-Smad2 or the reporter activity (Figure 6A and 6B). We conclude that in wild-type cells endogenous Arkadia must be in excess and sufficient to activate all available P-Smad2/3 in the nucleus. The fact that P-Smad2 is not eliminated and signaling is never repressed by the overexpression of Arkadia suggests that only a fraction of P-Smad2 interacts with and gets degraded by Arkadia, i.e., nuclear P-Smads and perhaps those engaged in transcription. The above data confirm the dual role and coexisting functions of Arkadia, suggesting that Arkadia enhances P-Smad2/3 activity at the expense of their levels.
An interesting observation is that in −/− ES cells, Arkadia expression not only restores the transcriptional deficit but it enhances reporter activity on average 50% (line in Figure 5C) or 100% (Figure 5D) above the maximum level that can be achieved in wild-type ES cells. The simplest explanation for this phenomenon is that the extra activity most likely reflects the accumulated levels of the hypoactive P-Smad2/3 that is being simultaneously activated by the expression of Arkadia. As this enhancement exceeds the maximum that can be achieved in wild-type ES cells even under ligand stimulation conditions, we termed it super-activation. According to this hypothesis, Arkadia expression in −/− ES cells will eventually “consume” (activate and degrade) the accumulated P-Smad2/3, releasing their activity to produce a transient super-activation of target genes. Subsequently, target gene transcription will be reduced to basal levels similar to that of wild-type cells.
We tested this prediction by comparing the percentage of enhancement by GAkd over that of the GFP control at different time points after transfection in Arkadia−/− ES cells. The results indicate that maximum super-activation occurs as early as 9 h post-transfection and coincides with the appearance of GFP fluorescence, declines after 15–18 h, and disappears after 30 h (Figure 5F). GFP fluorescence remained high throughout the experiment and past 48 h (unpublished data) and does not account for the loss of super-activation at 30 h. The above data suggest that in −/− ES cells Arkadia releases the activity of the hypoactive and stable P-Smad2/3 pool causing target gene transcription above wild-type levels (super-activation), and as this is a transient phenomenon, it occurs at the expense of P-Smad2/3 abundance. Therefore, Arkadia functions by a mechanism that consumes P-Smad2/3 as it activates them.
All the above analysis shows that in ES cells Arkadia functions as a coactivator of P-Smad2/3 transcription. To address whether this also occurs in other embryonic cells and if this is the underlying cause of the Arkadia−/− phenotype in the embryo, we examined the expression of known Smad2 target genes in Arkadia−/− embryos. The FoxH1/P-Smad2 complex directly upregulates the Nodal gene and is responsible for its tissue-specific expression in the visceral endoderm (VE) at pre-gastrulation stages [10,12,54]. Whole mount in situ hybridization, as expected, revealed that in Arkadia−/− embryos (n = 10) Nodal expression is dramatically reduced in the epiblast and almost lost in the VE (Figure 7A and 7B).
Later in development, FoxH1/Smad2/3 regulates directly Nodal [12], Pitx2, and Lefty2 [55,56] expression, in the left lateral plate mesoderm (LPM). However, this expression cannot be assessed in Arkadia−/− embryos because they lack a node and mesendoderm, which are essential for establishing left-right asymmetry. Using tetraploid chimeras (TC), we have previously shown that restoration of Arkadia expression in the extraembryonic lineages is sufficient to rescue the node and notochord formation in an embryo that consists entirely of Arkadia−/− cells [39]. We therefore generated TC embryos by injecting Arkadia−/− ES cells in tetraploid wild-type blastocysts. The rescue of node formation in the Arkadia−/− TC embryos is shown with the appearance of Nodal expression around the node (Figure 7D), which is absent in Arkadia−/− embryos (Figure 7E). We used these rescued Arkadia−/− embryos to examine target gene expression in the left LPM.
In the left LPM, expression of Nodal, Lefty2, and Pitx2 is present in wild-type (Figure 7G, 7I, and 7K) and absent in Arkadia−/− embryos (Figure 7F), while in the TC embryos (n = 9), Nodal and Pitx2 expression is severely reduced (Figure 7H and 7L) and Lefty2 is undetectable (Figure 7J), indicating that Arkadia is required for Nodal target gene expression in the LPM. Furthermore, the expression of Nodal in the node of the TC embryos is not reduced (compare Figure 7C and 7D) and is consistent with previous findings that this expression does not depend on Smad2/3 [12,57]. This confirms that Arkadia is required specifically for the expression of Smad2/3 target genes in vivo. Therefore, in the absence of Arkadia, Nodal target gene expression is reduced broadly in many cells and tissues throughout early embryogenesis, before and during gastrulation, and can account for the developmental defects observed in the −/− embryos. In addition, the effect of Arkadia on Nodal gene expression can explain its non-cell autonomous functions in the Arkadia−/− TC embryos where wild-type VE (expressing Nodal) can rescue node formation, and those reported in the literature for Xenopus assays [39,40].
According to the biochemical and functional data in ES cells, Arkadia interacts with P-Smad2, and this interaction is essential for the full expression of target genes. It is therefore expected that Arkadia−/− embryos and ES cells will have similar defects and exhibit the same phenotypes with those of Smad2−/− cells. The phenotype of mice with conditional FoxH1 or Smad2 deletion exclusively in the epiblast shows that they cannot form ADE/future foregut or prechordal plate (the most anterior mesendoderm). Furthermore, injection of FoxH1−/− or Smad2−/− ES cells in wild-type blastocysts shows that the −/− cells cannot colonize tissues such as the gut and the prechordal plate, indicating a cell autonomous requirement for these factors in these chimeras (mosaic embryos) [58,59]. It is therefore expected that Arkadia−/− embryos and ES cells will have the same defect and exhibit the same phenotypes with those of Smad2−/− cells. Expression of Hex, an ADE/foregut marker [44] (Figure 7M and 7N) and Shh, a prechordal plate marker (Figure S6), show that although Arkadia−/− TC embryos develop node and mesendoderm, they exhibit a deficit in these tissues. Consistent with the role of the ADE and the prechordal plate in maintaining and patterning the head folds, we observed that the Arkadia−/− TC exhibit reduction of the head folds (Figure 7J, 7L, and 7N) that can also account for the reduction of Pitx2 expression in the forebrain (Figure 7L).
The requirement for Arkadia expression within the cells that form the ADE/foregut and the prechordal plate was addressed in mosaic chimeras generated either by injection of wild-type ES cells into Arkadia−/− blastocysts (Figure 7O) or by Arkadia−/− ES cells into wild-type blastocysts (Figure 7P) [39]. In both types of chimeras, the embryo consists of a mixture of wild-type (unstained) and Arkadia−/− cells (β-galactosidase stained) [39]. We found that mosaic chimeras (n > 100) exhibit normal morphology, as long as wild-type cells colonize the foregut and the prechordal plate (Figure 7O and 7P). Therefore, like Smad2, Arkadia is required cell autonomously for ADE/foregut and prechordal plate formation. The above data show that Arkadia loss-of-function phenocopies that of Smad2 in embryonic cells, supporting a functional interaction between the two factors.
An important unanswered question is how the promoters of target genes are cleared and refreshed from transcription factors that are activated by signal transduction pathways to allow rapid intracellular responses to dynamic changes in the concentration of extracellular ligands. We show here that Arkadia terminates signaling at the level of transcription, by linking the ubiquitination/degradation of P-Smad2/3 to their transcriptional activity. Therefore, this mechanism enhances transcription but limits the time that the effector works by inducing its turnover. This represents the first example of coupled activation and degradation of signaling effectors, as well as a critical role for this mechanism in development. It is possible that very similar mechanisms operate in other pathways to establish dynamic regulation and efficient signaling, while their failure may be associated with disease.
Experiments using somatic tissue culture cell lines showed that Arkadia might enhance TGF-β signaling by degrading the Smad6/7, which are known to inhibit Smad2/3 phosphorylation mainly by mediating the degradation of the receptors [41,42]. If this is the case, then embryos and ES cells should have reduced levels of P-Smad2/3 in the absence of Arkadia. Contrary to this prediction, all Arkadia−/− embryos and ES cells have at least 2-fold higher P-Smad2/3 levels compared to wild-type. Therefore, Arkadia must enhance signaling via a different mechanism during early development.
In vivo, in ES cells, MEFs, and embryos, loss of Arkadia causes P-Smad2/3 stabilization and accumulation in the nucleus and makes them resistant to proteasome degradation (Figure 4), suggesting that Arkadia directly or indirectly is associated with their stability and degradation. Destruction of P-Smad2 had been shown previously to occur in the nucleus [38], but a nuclear ubiquitin ligase that interacts specifically with the phosphorylated form of Smad2/3, as well as the role of proteasome-mediated turnover of P-Smads during development, remained unknown. We present here a number of experiments indicating that Arkadia, a nuclear E3 ubiquitin ligase, interacts with P-Smad2/3 and directly ubiquitinates them leading to their degradation by the proteasome. These include IP experiments showing that Arkadia interacts specifically with the phosphorylated forms of Smad2/3 via the conserved NRG domain (Figure 3A and 3B), and in vitro ubiquitination assays showing that Arkadia directly poly-ubiquitinates P-Smad2 (Figures 4A–4C and S5). Furthermore, introduction of Arkadia in −/− cells causes P-Smad2 poly-ubiquitination and decreases its abundance in a proteasome-dependent manner (Figure 4D–4E). However, as Arkadia is nuclear, it most likely degrades nuclear P-Smads.
One of the major questions is how effector activity is terminated after target genes have been activated and not before. This should involve a mechanism to distinguish between effectors actively engaged in transcription and those that are fresh and unused. A modification or a change of conformation of the effectors when they interact with the target-gene promoter complex may allow recognition by a critical component of the termination mechanism (referred to subsequently as the terminator). In this case, absence of the terminator should lead to a prolonged, persistent response to signaling, as the unhindered effectors will continue to transcribe. On the other hand, excess terminator could cause repression of signaling if it degrades and deactivates the effectors prematurely, i.e., as soon as they bind to the promoter. A more precise and efficient mechanism would involve degradation of the effectors after they activate at least one round of transcription. This requires a terminator directly involved in transcription by the effectors or an even more stringent mechanism, where initiation of transcription is linked with the destruction of the transcription factor: a “suicide” model. In this case, absence of the terminator will pause transcription, while its overexpression will not degrade/deactivate the effectors prematurely to repress target gene transcription. We show here that Arkadia fulfils the criteria of such a termination mechanism as it degrades P-Smad2/3 and enhances their target gene transcription via the same domains.
Does Arkadia directly activate P-Smad2/3? In the absence of Arkadia, P-Smad2/3 are stable and accumulate in the nucleus, but they are repressed or hypoactive as target gene transcription is reduced and in some cases lost (Figure 7). Introduction of Arkadia in −/− ES cells not only restores signaling but super-activates P-Smads, meaning that the expression of the reporter reaches levels higher than those that can be achieved in wild-type ES cells under maximum ligand stimulation conditions. The above findings suggest that the extra activity is not generated by receptors or ligands, but by the release of activity from the pools of the accumulated hypoactive P-Smad2/3. Consistent with this hypothesis, super-activation by Arkadia in −/− ES cells is a transient phenomenon, as it depends on the excess P-Smad2 that is used up. Therefore, Arkadia degrades and spends P-Smad2/3 to activate their transcriptional activity. These observations suggest that P-Smad2/3 is reminiscent of fuel, which in the absence of Arkadia is stable and can be stored, while in its presence it is “ignited” and “burns,” releasing activity. These observations suggest that phosphorylation of Smad2/3 by the receptors is not sufficient for their transcriptional activity and that they require an additional step mediated by the function of Arkadia as a coactivator in the nucleus.
We have shown previously that in Xenopus animal cap assay, injection of Arkadia RNA enhances the ability of Nodal but not that of Smad2 RNA to induce mesendoderm [40]. This appears inconsistent with our current data, however; as in the animal cap there is no ligand to phosphorylate Smad2, it remains a mystery how Smad2 RNA injection alone induces mesoderm and mesendoderm. Furthermore, the amount of Smad2 RNA required for this is very high and most likely does not represent physiological signaling conditions. More importantly, we show here that Arkadia recognizes P-Smad2 and not Smad2, and therefore it is not surprising that in the Smad2-injected animal caps Arkadia does not induce mesendoderm.
While the exact mechanism of how P-Smad2/3 are activated by Arkadia remains unknown, it is important to point out that Arkadia does not degrade P-Smads directly, it only modifies them by ubiquitination and is the proteasome that degrades them. Although the role of ubiquitin modification in mediating proteasome degradation is well established, it has become apparent that addition of ubiquitin onto proteins may also affect their properties. Recent studies have suggested a direct role of ubiquitin modifications and of the proteasome in transcriptional activation (reviewed in [60–63]). It is possible, however, that Arkadia interacts with, and degrades simultaneously, P-Smads and a closely linked repressor. Future studies will determine the mechanism of P-Smad2/3 activation and how this is coupled with immediate degradation. In conclusion, by linking P-Smad2/3 activation with degradation, Arkadia provides a way to achieve rapid resetting of P-Smad2/3 target gene transcription in development and allows efficient and dynamic responses to Nodal/TGF-β signaling events.
We had shown previously that Arkadia−/− embryos do not develop anterior primitive streak and its derivative tissues: ADE/future foregut, mesendoderm (prechordal plate and notochord), and node [39]. However, the molecular mechanisms that underlie these abnormalities were not understood. Specifically, we had shown that the formation of the node is restored in chimeric embryos consisting of tetraploid wild-type extraembryonic lineages (such as VE) and exclusively Arkadia−/− embryonic tissues derived from −/− ES cells [39]. This indicates that the development of the node in the embryonic lineages requires Arkadia expression and function within a different lineage (extraembryonic). As Arkadia is nuclear and unlikely to be secreted, this effect was presumed to be mediated by a downstream-secreted factor. Here we show that Arkadia increases directly the transcription of P-Smad2 target genes, which include the Nodal gene itself, particularly in the VE and early epiblast when node precursors are induced (Figure 7). This suggests that in the Arkadia−/− TC, the wild-type VE provides sufficient Nodal to partially restore the overall level of Nodal in the Arkadia−/− epiblast, thus bringing it up to the threshold required for node formation (Figure 7D and unpublished data). Therefore, organizer/node induction requires high levels of Nodal, which can be achieved with a considerable contribution from the VE.
While the node may be rescued, to what extent expression of Arkadia in the extraembryonic lineages is sufficient to restore normal development in the Arkadia−/− embryo of the TC? We show here that these embryos cannot form the most anterior derivatives of the primitive streak, such as ADE/foregut and prechordal plate (Figure 7), and exhibit laterality defects, including delayed turning and heart looping ([39] and unpublished data). This is consistent with our findings that Arkadia is required for efficient P-Smad2/3 target gene transcription. Specifically, the reduction of Nodal expression in the left LPM along with its known target genes (Lefty2 and Pitx2) can account for the left-right axis defects. The Nodal target genes that are responsible for foregut and prechordal plate development remain unknown. However, the development of these tissues must depend on the expression of Arkadia within these cells or their precursors. This was shown by the analysis of chimeric embryos consisting of a mixture of wild-type and Arkadia−/− cells, which revealed that the −/− cells could not contribute to these tissues (Figure 7O and 7P). Furthermore, as Smad2−/− cells [59] and FoxH1−/− cells [58] behave similarly in chimeras, our data provide in vivo evidence for the functional interaction between Arkadia and Smad2, and reveal that the three factors together regulate target genes essential for foregut endoderm and prechordal plate formation.
Our data show that Arkadia is essential for the development of tissues that depend on very high Nodal signaling such as ADE/foregut, and that its introduction in −/− ES cells can boost signaling above levels obtained by just high concentrations of ligand “super-activation.” This latter phenomenon is presumably generated by the simultaneous activation of the accumulated P-Smad2/3, which has been stabilized and reached higher than normal levels in the absence of Arkadia. It is possible that such a mechanism may be responsible for maximizing Nodal signaling in the embryo. However, it would require the transient absence of Arkadia in the precursors of the ADE/foregut. Although Arkadia RNA is present broadly in the embryo, protein analysis is needed to reveal whether super-activation occurs in the embryo.
In conclusion, we reveal a novel ubiquitin-mediated mechanism of TGF-β signaling regulation that depends on Arkadia, involves the activation of P-Smad2/3 signaling effectors downstream of receptor-phosphorylation, and couples their high activity with turnover and signaling termination. As the TGF-β/Nodal signaling pathway has been linked to cancers and genetic diseases, the identification of key molecular players such as Arkadia would be useful for the development of new drug targets and therapeutic intervention.
Chimeras were generated as described before [39]. Both adult mice and embryos were genotyped using allele-specific PCR amplification of genomic DNA. The oligonucleotide primer pairs TGAGGTAGGCATACCTAGAG and TGACTTAAGCCCTGCAATCC; TGAGGTAGGCATACCTAGAG and CTGAGTGATTGACTACCCGT; and TCTGGATTCATCGACTGTGG and CTGGATGTAGGCATGGTTGGTAG were used to give diagnostic amplification products of 313 bp for the wild-type Arkadia allele, 293 bp for the disrupted Arkadia allele, and 925 bp for the disrupted Nodal allele, respectively. Histology and in situ hybridizations and marker plasmids used are as previously described [39].
Total RNA was extracted from ES cells using Trizol (Invitrogen, http://www.invitrogen.com) followed by digestion with RQ1 RNase-Free DNase (Promega, http://www.promega.com) to remove DNA contamination. Synthesis of cDNA from total RNA was performed with SuperScript II Reverse Transcriptase (Invitrogen). Experiments were performed in quadruplicates using the DNA Engine Opticon Real-Time PCR Detection System (MJ Research, http://www.bio-rad.com) and SYBRGreen PCR Master Mix (Applied Biosystems, http://appliedbiosystems.com). The Nodal 375-bp amplicon was produced by the forward/reverse primer pair AAGACCAAGCCACTGAGCAT and GCCTTTGCACACAATTTCAA. Nodal expression was quantified by normalizing against endogenous controls GAPDH (primer pair TGCACCACCAACTGCTTAGC and GGCATGGACTGTGGTCATGAG) or YWHAZ (primer pair CGTTGTAGGAGCCCGTAGGTCAT and TCTGGTTGCGAAGCATTGGG) using the delta Ct method.
Tagged Arkadia: Mus musculus ring finger 111 (Rnf111) constructs were generated by eliminating the first 9 aa of Arkadia and fusing in frame to GFP (pEGFP-cI; Clontech, http://www.clontech.com) or Flag-tag (synthetic oligonucleotide). A single Myc-tag (synthetic oligonucleotide) was fused to the carboxy-terminus of Arkadia. The various tagged Arkadia sequences and the GFP control gene were subcloned into SmaI/XhoI of the pTriEx2-hygro (Novagen, http://www.novagen.com) vector. The Arkadia RING domain mutations were constructed by either deleting the last 24 aa that included the second Zinc finger (GAkdR*) or by point mutation leading to amino acid substitutions (C2–A2) that disrupt both Zinc fingers (GAkdR2*). GAkdNRG* consists of an internal deletion of the first 7 aa of the NRG domain by PCR. The truncated Arkadia containing only the C-terminal region of Arkadia was constructed by fusing GFP in frame with the N-terminus of the truncated Arkadias. G-NRG-RING includes aa 889–989; G-NLS-RING, 903–989; GRING 947–989. Pitx2-luc and 123-luc (gifts of H. Hamada, Osaka, Japan) and 9xCAGA-luc; 6 Myc-Smad3; constitutively active forms of Alk4 and Alk6 (gifts of K. Miyazono, Tokyo, Japan); HA-Ub (gift of Y. Fujita, London, United Kingdom); and the Flag-Smad2 construct (gift of J. Smith, Cambridge, United Kingdom).
The ES cell lines used in this study were derived directly from blastocysts as described previously [9]. Subsequently, they were maintained feeder-free in 20% FCS in DMEM and LIF (Invitrogen). Cells were transiently transfected with Lipofectamine 2000 (Invitrogen), Lipofectamine Plus (Invitrogen), or TransIT-293 (Mirus, http://www.mirusbio.com). Arkadia null primary mouse fibroblasts of mixed 129Sv/MF1 genetic backgrounds were isolated from embryos at 9.5 dpc and immortalized with pBabe-puro-SV40-TA (gift from Parmjit Jat, Ludwig Institute, London, United Kingdom), selected with 0.3 μg/ml puromycin and maintained in F12 (Invitrogen) medium with 20% FCS. Signaling stimulation was achieved using 50 μng/ml BMP2 or 10 μng/ml Activin A (Sigma, http://sigmaaldrich.com); signaling inhibition using serine/threonine kinase inhibitor H7 (Calbiochem, http://www.calbiochem.com) or SB431542 (Sigma and gift from GSK); and translation inhibition using 100 μg/ml cycloheximide (Sigma).
Cells were cotransfected with the Firefly reporter plasmids, a Renilla luciferase transfection control pRL-SV40 (Promega), and various GFP-tagged Arkadia constructs in a 5:1:5 ratio. Dual luciferase assays were performed according to manufacturer's protocols (Promega) 18 h after transfection, unless stated otherwise. After normalizing the Firefly values to Renilla, the data was presented as relative luciferase values or as percent increase.
For IP studies, cells in 10-cm culture dishes were treated for 5 h with 50 μM of the proteasome inhibitor MG-132 (Calbiochem) before lysis. The cells were lysed by adding 1 ml per 10-cm dish of RIPA (150 mM NaCl, 50 mM Tris [pH 8.0], 0.5% DOC, 0.1% SDS, and 1% NP-40) or NP-40 buffer (20 mM Tris [pH 7.5], 150 mM NaCl, 0.5% NP-40), which was specifically used for the IPs shown in Figure S5A. Both IP solutions contained 10 mM N-ethylmaleimide (cysteine protease inhibitor, Sigma), 100 μM MG-132 (Calbiochem), 100 μM Epoxomicin (Calbiochem), 100 μM clasto-Lactacystin β-Lactone (Calbiochem), and protease/phosphatase inhibitor cocktails (Sigma) at a final concentration of 1% each. After centrifugation at 100,000 g for 30 min, the supernatants were immunoprecipitated for 1 h with the appropriate antibody and 50 μl of protein agarose beads (Sigma). 2 μg of an anti-GFP mouse monoclonal antibody (clone 7.1 and 13.1; Roche, http://www.roche.com) or 2 μg of the anti-FLAG M2 antibody (Stratagene, http://www.stratagene.com) with 50 μl of Protein G Sepharose beads (Amersham Biosciences, http://www.gehealthcare.com) was used for the IP. Samples were eluted off the beads by boiling in 2× Laemelli buffer followed by standard SDS-ployacrylamide gel electrophoresis (SDS-PAGE) and Western blot analysis.
For nuclear and cytoplasmic fractionation, cells were lysed in hypotonic buffer (20 mM Hepes [pH7.5], 10 mM NaCl, 0.2 mM EDTA, 20% glycerol, 1.5 mM MgCl2, 0.1% Triton X-100) containing protease and phosphatase inhibitors as above. Nuclei were pelleted by centrifugation at 1,000 rpm for 10 min, and the cytoplasmic fraction obtained by retaining the supernatant. Nuclear extracts were obtained by rocking the nuclear pellet in five times the volume of hypertonic solution (hypotonic buffer + 500 mM NaCl) for 1 h at 4° C and subsequent centrifugation at 13,000 rpm for 5 min. Nuclear fraction was obtained in the supernatant, and 30 μg of each protein sample was loaded on each lane. When required to inhibit protein degradation, MG-132 (50 μM) was added to both the hypotonic and hypertonic buffers.
For Western blotting, cells and embryos were lysed with the RIPA buffer as for IPs. 30 μg of each protein sample was loaded on each lane for SDS-PAGE. The following antibodies were obtained and used according to the manufacturer's instructions: rabbit anti-P-Smad2 (Calbiochem), rabbit anti-P-Smad3 (Cell Signaling Technologies, http://www.cellsignal.com and also gift from E. Leof, Mayo Clinic, United States), rabbit anti-P-Smad1/5/8 (Cell Signaling Technologies), rabbit anti- Smad2 (Zymed Laboratories, http://www.invitrogen.com), rabbit anti-GFP (Abcam, http://www.abcam.com), rabbit anti- cyclin D2 (H-295; Santa Cruz Biotechnology, http://www.scbt.com), as well as mouse monoclonal antibodies against Smad4 (B-8; Santa Cruz); actin (Santa Cruz); PCNA (PC-10; Santa Cruz); tubulin (B-5-1-2; Sigma); FLAG (M2; Stratagene); histone H3 (Upstate, http://www.upstate.com); ubiquitin (Covance and Bethyl Laboratories, http://www.bethyl.com) and HA (Roche).
[35S]-methionine-labeled or unlabeled recombinant proteins full-length Arkadia, N-Akd (1–516 aa), C-Akd (510–989), and luciferase were generated by the TNT Quick Coupled in vitro transcription/translation kit (Promega). Flag-P-Smad2 was obtained with IP from 293T cells. Beads bound to Flag-P-Smad2 protein were washed with RIPA buffer and incubated for 1 h by constant rotation at 4 °C with radiolabeled recombinant proteins in 1 ml of RIPA buffer. Protein-bead complexes were then washed four times with RIPA buffer (200 × bed volume) and re-suspended in 2 × Laemelli buffer. The presence of radiolabeled Arkadia protein in pull-downs was detected by SDS-PAGE and autoradiography. Assays were carried out in 20 μl of ubiquitination assay buffer (20 mM Tris-HCl [pH7.7], 100 mM KCl, 0.1 mM CaCl2, 1 mM MgCl2, 1 mM DTT) containing 15 μg of GST-ubiquitin (Boston Biochem, http://www.bostonbiochem.com), 1 μg of E1 (Boston Biochem), 1.5 μg of E2 (GSTUbcH5b and GST-UbcH5c; Boston Biochem), 0.6 μl of Energy Regeneration Solution (Boston Biochem), and Flag-P-Smad2 immunoprecipitated from 293T cells. In vitro translated Arkadia and Arkadia mutants (N-terminal corresponding to aa 1–516 and C-terminal truncation corresponding to aa 510–989) were generated using the TNT Quick Coupled in vitro transcription/translation kit (Promega). Reactions were incubated for 1 h under constant rotation at 37 °C. The reaction was terminated by the addition of 2 × Laemelli buffer, and the presence of poly-ubiquitinated substrates was detected by Western blotting.
Immunofluorescence was performed using the following primary antibodies: anti-P-Smad2 antibody (1:50; Calbiochem); anti-Smad2/3 (1:100; BD Biosciences, http://www.bdbiosciences.com); Alexa Fluor 568 anti-rabbit secondary (1:400; Molecular Probes, http://www.invitrogen.com); and mounting medium with DAPI (Vectashield, http://www.vectorlabs.com). Arkadia−/− and wild-type ES cells were grown on cover slips and rinsed in PBS for 5 min prior to fixation with 4% (w/v) paraformaldehyde in PBS for 10 min. Following fixation, cells were permeabilized with 0.5% Triton-X 100 in PBS for 2 min on ice and rinsed with PBS for 5 min and incubated in 10% FCS/PBS for 30 min to block nonspecific binding of antibodies. Subsequently, the cells were incubated with the appropriate primary antibody diluted in 10% FCS/PBS for 1 h at room temperature. A no primary control was included to verify antibody specificity. Cover slips were then washed several times with 10% FCS/PBS and incubated with secondary antibody for a further 1 h. This was followed by washes in PBS and cover slips mounted for fluorescence in medium containing DAPI (Vectashield). The cells were visualized on a Leica TCS SP2 (http://www.leica.com) confocal microscope at 100× magnification.
The GenBank (http://www.ncbi.nlm.nih.gov/genbank) accession numbers for the entities from the discussed in this paper are Arkadia (NM_033604), GAPDH (NM_001001303), Nodal (NM_013611), and YWHAZ (NM_011740).
|
10.1371/journal.pmed.1002686 | Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists | Chest radiograph interpretation is critical for the detection of thoracic diseases, including tuberculosis and lung cancer, which affect millions of people worldwide each year. This time-consuming task typically requires expert radiologists to read the images, leading to fatigue-based diagnostic error and lack of diagnostic expertise in areas of the world where radiologists are not available. Recently, deep learning approaches have been able to achieve expert-level performance in medical image interpretation tasks, powered by large network architectures and fueled by the emergence of large labeled datasets. The purpose of this study is to investigate the performance of a deep learning algorithm on the detection of pathologies in chest radiographs compared with practicing radiologists.
We developed CheXNeXt, a convolutional neural network to concurrently detect the presence of 14 different pathologies, including pneumonia, pleural effusion, pulmonary masses, and nodules in frontal-view chest radiographs. CheXNeXt was trained and internally validated on the ChestX-ray8 dataset, with a held-out validation set consisting of 420 images, sampled to contain at least 50 cases of each of the original pathology labels. On this validation set, the majority vote of a panel of 3 board-certified cardiothoracic specialist radiologists served as reference standard. We compared CheXNeXt’s discriminative performance on the validation set to the performance of 9 radiologists using the area under the receiver operating characteristic curve (AUC). The radiologists included 6 board-certified radiologists (average experience 12 years, range 4–28 years) and 3 senior radiology residents, from 3 academic institutions. We found that CheXNeXt achieved radiologist-level performance on 11 pathologies and did not achieve radiologist-level performance on 3 pathologies. The radiologists achieved statistically significantly higher AUC performance on cardiomegaly, emphysema, and hiatal hernia, with AUCs of 0.888 (95% confidence interval [CI] 0.863–0.910), 0.911 (95% CI 0.866–0.947), and 0.985 (95% CI 0.974–0.991), respectively, whereas CheXNeXt’s AUCs were 0.831 (95% CI 0.790–0.870), 0.704 (95% CI 0.567–0.833), and 0.851 (95% CI 0.785–0.909), respectively. CheXNeXt performed better than radiologists in detecting atelectasis, with an AUC of 0.862 (95% CI 0.825–0.895), statistically significantly higher than radiologists' AUC of 0.808 (95% CI 0.777–0.838); there were no statistically significant differences in AUCs for the other 10 pathologies. The average time to interpret the 420 images in the validation set was substantially longer for the radiologists (240 minutes) than for CheXNeXt (1.5 minutes). The main limitations of our study are that neither CheXNeXt nor the radiologists were permitted to use patient history or review prior examinations and that evaluation was limited to a dataset from a single institution.
In this study, we developed and validated a deep learning algorithm that classified clinically important abnormalities in chest radiographs at a performance level comparable to practicing radiologists. Once tested prospectively in clinical settings, the algorithm could have the potential to expand patient access to chest radiograph diagnostics.
| Chest radiographs are the most common medical imaging test in the world and critical for diagnosing common thoracic diseases.
Radiograph interpretation is a time-consuming task, and there is shortage of qualified trained radiologists in many healthcare systems.
Deep learning algorithms that have been developed to provide diagnostic chest radiograph interpretation have not been compared to expert human radiologist performance.
We developed a deep learning algorithm to concurrently detect 14 clinically important pathologies in chest radiographs.
The algorithm can also localize parts of the image most indicative of each pathology.
We evaluated the algorithm against 9 practicing radiologists on a validation set of 420 images for which the majority vote of 3 cardiothoracic specialty radiologists served as ground truth.
The algorithm achieved performance equivalent to the practicing radiologists on 10 pathologies, better on 1 pathology, and worse on 3 pathologies.
Radiologists labeled the 420 images in 240 minutes on average, and the algorithm labeled them in 1.5 minutes.
Deep learning algorithms can diagnose certain pathologies in chest radiographs at a level comparable to practicing radiologists on a single institution dataset.
After clinical validation, algorithms such as the one presented in this work could be used to increase access to rapid, high-quality chest radiograph interpretation.
| Chest radiography is the most common type of imaging examination in the world, with over 2 billion procedures performed each year [1]. This technique is critical for screening, diagnosis, and management of thoracic diseases, many of which are among the leading causes of mortality worldwide [2]. A computer system to interpret chest radiographs as effectively as practicing radiologists could thus provide substantial benefit in many clinical settings, from improved workflow prioritization and clinical decision support to large-scale screening and global population health initiatives.
Recent advancements in deep learning and large datasets have enabled algorithms to match the performance of medical professionals in a wide variety of other medical imaging tasks, including diabetic retinopathy detection [3], skin cancer classification [4], and lymph node metastases detection [5]. Automated diagnosis from chest imaging has received increasing attention [6,7], with specialized algorithms developed for pulmonary tuberculosis classification [8,9] and lung nodule detection [10], but the use of chest radiographs to discover other pathologies such as pneumonia and pneumothorax motivates an approach that can detect multiple pathologies simultaneously. Only recently have the computational power and availability of large datasets enabled the development of such an approach. The National Institutes of Health’s release of ChestX-ray14 led to many more studies that use deep learning for chest radiograph diagnosis [11–13]. However, the performance of these algorithms has not been compared to that of practicing radiologists.
In this work, we aimed to assess the performance of a deep learning algorithm to automatically interpret chest radiographs. We developed a deep learning algorithm to concurrently detect the presence of 14 different disease classes in chest radiographs and evaluated its performance against practicing radiologists.
The ChestX-ray14 dataset [14] was used to develop the deep learning algorithm. The dataset is currently the largest public repository of radiographs, containing 112,120 frontal-view (both posteroanterior and anteroposterior) chest radiographs of 30,805 unique patients. Each image in ChestX-ray14 was annotated with up to 14 different thoracic pathology labels that were chosen based on frequency of observation and diagnosis in clinical practice. The labels for each image were obtained using automatic extraction methods on radiology reports, resulting in 14 binary values per image, where 0 indicates the absence of that pathology and 1 denotes the presence (multiple pathologies can be present in each image). We partitioned the dataset into training, tuning, and validation (see S1 Table for statistics of dataset splits used in this study).
The training set was used to optimize network parameters, the tuning set was used to compare and choose networks, and the validation set was used to evaluate CheXNeXt and radiologists. There is no patient overlap among the partitions.
A validation set of 420 frontal-view chest radiographs was selected from ChestX-ray14 for radiologist annotation. The set was curated to contain at least 50 cases of each pathology according to the original labels provided in the dataset by randomly sampling examples and iteratively updating the selected examples by sampling from the examples labeled with the underrepresented pathologies. The radiographs in the validation set were annotated by 3 independent board-certified cardiothoracic specialist radiologists (average experience 15 years, range 5–28 years) for the presence of each of the 14 pathologies. The majority vote of their annotations was taken as a consensus reference standard on each image. To compare to the algorithm, 6 board-certified radiologists from 3 academic institutions (average experience 12 years, range 4–28 years) and 3 senior radiology residents also annotated the validation set of 420 radiographs for all 14 labels. All radiologists individually reviewed and labeled each of the images using a freely available image viewer with capabilities for picture archiving and communication system features such as zoom, window leveling, and contrast adjustment. Radiologists did not have access to any patient information or knowledge of disease prevalence in the data. Labels were entered into a standardized data entry program, and the total time to complete the review was recorded. The Stanford International Review Board (IRB) approved this study, and all radiologists consented to participate in the labeling process.
The deep learning algorithm, called CheXNeXt, is a neural network trained to concurrently detect the 14 pathologies in frontal-view chest radiographs. Neural networks are functions with many parameters that are structured as a hierarchy of layers to model different levels of abstraction. In this study, the selected architecture was a convolutional neural network, a particular type of neural network that is specially designed to handle image data. By exploiting a parameter sharing receptive field, convolutional neural networks scan over an image to learn features from local structure and aggregate the local features to make a prediction on the full image. The neural network used in this study is a 121-layer DenseNet architecture [15] in which each layer is directly connected to every other layer within a block. For each layer, the feature maps of all preceding layers are used as inputs, and its own feature maps are passed on to all following layers as inputs.
Once specifying the neural network architecture, the parameters are automatically learned from a large amount of data labeled with the presence or absence of each pathology. The learning process consists of iteratively updating the parameters to decrease the prediction error, which is computed by comparing the network’s prediction to the known annotations on each image. By performing this procedure using a representative set of images, the resulting network can make predictions on previously unseen frontal-view chest radiographs.
The training process consisted of 2 consecutive stages to account for the partially incorrect labels in the ChestX-ray14 dataset. First, multiple networks were trained on the training set to predict the probability that each of the 14 pathologies is present in the image. Then, a subset of those networks, each chosen based on the average error on the tuning set, constituted an ensemble that produced predictions by computing the mean over the predictions of each individual network. The ensemble was used to relabel the training and tuning sets as follows: first, the ensemble probabilities were converted to binary values by computing the threshold that led to the highest average F1 score on the tuning set across all pathologies. Then, the new label was taken to be positive if and only if either the original label was positive or the ensemble prediction was positive. Finally, new networks were trained on the relabeled training set, and a subset of the new networks was selected based on the average error on the relabeled tuning set. The final network was an ensemble of 10 networks trained on the relabeled data, where again the predictions of the ensemble were computed as the mean over the predictions of each individual network.
Before both stages of training, the parameters of each network were initialized with parameters from a network pretrained on ImageNet [16]. The final fully connected layer of the pretrained network was replaced with a new fully connected layer producing a 14-dimensional output, after which the sigmoid was applied to each of the outputs to obtain the predicted probabilities of the presence of each of the 14 pathology classes. Before inputting the images into the network, the images were resized to 512 pixels by 512 pixels and normalized based on the mean and standard deviation (SD) of images in the ImageNet training set. For each image in the training set, a random lateral inversion was applied with 50% probability before being fed into the network. The networks were updated to minimize the sum of per-class weighted binary cross entropy losses, where the per-class weights were computed based on the prevalence of that class in the training set. All parameters of the networks were trained jointly using Adam with standard parameters [17]. Adam is an effective variant of an optimization algorithm called stochastic gradient descent, which iteratively applies updates to parameters in order to minimize the loss during training. We trained the networks with minibatches of size 8 and used an initial learning rate of 0.0001 that was decayed by a factor of 10 each time the loss on the tuning set plateaued after an epoch (a full pass over the training set). In order to prevent the networks from overfitting, early stopping was performed by saving the network after every epoch and choosing the saved network with the lowest loss on the tuning set. No other forms of regularization, such as weight decay or dropout, were used. Each stage of training completed after around 20 hours on a single NVIDIA GeForce GTX TITAN Black. Each network had 6,968,206 learnable parameters, and the final ensemble had 69,682,060 parameters.
The open-source deep learning framework PyTorch (http://pytorch.org/) was used to train and evaluate the algorithms.
In order to interpret predictions, CheXNeXt produced heat maps that identified locations in the chest radiograph that contributed most to the network’s classification through the use of class activation mappings (CAMs) [18]. To generate the CAMs, images were fed into the fully trained network, and the feature maps from the final convolutional layer were extracted. A map of the most salient features used in classifying the image as having a specified pathology was computed by taking the weighted sum of the feature maps using their associated weights in the fully connected layer. The most important features used by CheXNeXt in its prediction of the pathology were identified in the image by upscaling the map to the dimensions of the image and overlaying the image.
We provide a comprehensive comparison of the CheXNeXt algorithm to practicing radiologists across 7 performance metrics, namely, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, F1 metric, positive and negative predictive value (PPV and NPV), and Cohen’s kappa [19]. To convert the probabilities produced by CheXNeXt to binary predictions, we chose pathology-specific thresholds through maximization of the F1 score on the tuning set (more details presented in S1 Appendix).
To compare the CheXNeXt algorithm to radiologists using a single diagnostic performance measure, we used the AUC metric. Because the radiologists only provided yes/no responses for each image and not a continuous score, the receiver operating characteristic (ROC) was estimated for the radiologists as a group using partial least-squares regression with constrained splines to fit an increasing concave curve to the specificities and sensitivities of 9 radiologists. We specify knots at each 1/20th and assume symmetry. An example with R code is provided in S1 Appendix.
Because we estimate the ROCs for the radiologists, we cannot use standard confidence intervals (CIs) for the radiologists' AUCs, and so to ensure a fair comparison, we calculated and compared the respective AUCs in the same manner, as follows. We first estimate the ROC for the radiologists using constrained splines—as described above—and the ROC for the algorithm and then estimate the AUCs for both the algorithm and the radiologists using linear interpolation and the composite trapezoidal rule. Finally, we use the robust bootstrap method, described below, to construct CIs around the AUCs.
In addition to individual-level and pathology-specific performance measures, the CheXNeXt algorithm was evaluated over all pathologies and against radiologists as a group. To evaluate CheXNeXt against resident radiologists as a group and board-certified radiologists as a group, the micro-averages of the performance measures were computed across all resident radiologists as well as across all board-certified radiologists. Micro-averages for groups of radiologists were calculated by concatenating the predictions of group members and then calculating the performance measures. For example, to calculate the sensitivity for board-certified radiologists in predicting hernia (420 images), we concatenated each of 6 board-certified radiologists' predictions into a single array of length 420 × 6 = 2,520, repeated the reference standard for hernia 6 times to create an array of the same length, and then calculated sensitivity. To provide an overall estimate of accuracy, the proportion correct was calculated for each image across all 14 pathologies, and the mean and SD of these proportions are reported.
The nonparametric bootstrap was used to estimate the variability around each of the performance measures; 10,000 bootstrap replicates from the validation set were drawn, and each performance measure was calculated for CheXNeXt and the radiologists on these same 10,000 bootstrap replicates. This produced a distribution for each estimate, and the 95% bootstrap percentile intervals (2.5th and 97.5th percentiles) are reported [20].
Because AUC is a single measure on which to compare the CheXNeXt algorithm to the radiologists as a group, the difference between the AUCs on these same bootstrap replicates was also computed. To control the familywise error rate when testing for significant differences in AUCs, the stringent Bonferroni-corrected [21] CIs of 1 − 0.05/14 are reported. If the interval does not include 0, there is evidence that either CheXNeXt or the radiologists are superior in that task.
All statistical analyses were completed in the R environment for statistical computing [22]. The irr package [23] was used to calculate the exact Fleiss’ kappa and Cohen’s kappa. The boot package [24] was used to perform the bootstrap and construct the bootstrap percentile intervals (95% and 99.6%). The ConSpline package [25] was used to estimate the ROC for the radiologists using partial least-squares regression with constrained splines, the pROC package [26] was used to estimate the ROC for the algorithm, and the MESS package [27] was used to calculate the AUC for both the radiologists and CheXNeXt. Figures were created using the ggplot2 [28] and gridExtra [29] packages.
The ROC curves for each of the pathologies on the validation set are illustrated in Fig 1, and AUCs with CIs are reported in Table 1; statistically significant differences in AUCs were assessed with the Bonferroni-corrected CI (1 − 0.05/14). The CheXNeXt algorithm performed as well as the radiologists for 10 pathologies and performed better than the radiologists on 1 pathology. It achieved an AUC of 0.862 (95% CI 0.825–0.895) for atelectasis, statistically significantly higher than radiologists' AUC of 0.808 (95% CI 0.777–0.838). The radiologists achieved statistically significantly higher AUC performance on cardiomegaly, emphysema, and hiatal hernia, with AUCs of 0.888 (95% CI 0.863–0.910), 0.911 (95% CI 0.866–0.947), and 0.985 (95% CI 0.974–0.991), respectively, whereas CheXNeXt’s AUCs were 0.831 (95% CI 0.790–0.870), 0.704 (95% CI 0.567–0.833), and 0.851 (95% CI, 0.785–0.909), respectively. There were no statistically significant differences in the AUCs for the other 10 pathologies.
Performance measure results for mass, nodule, consolidation, and effusion are illustrated in Fig 2 (panels a–d), and numerical values for those pathologies are reported in S1 File. The CheXNeXt algorithm detected masses and nodules with sensitivities of 0.754 (95% CI 0.644–0.860) and 0.690 (95% CI 0.581–0.797), respectively, which was higher than the micro-average sensitivities of board-certified radiologists at 0.495 (95% CI 0.443–0.546) and 0.573 (95% CI 0.525–0.619), respectively (Fig 2). CheXNeXt maintained high specificity in both tasks, achieving 0.911 (95% CI 0.880–0.939) in mass detection and 0.900 (95% CI 0.867–0.931) in nodule detection compared with radiologist scores of 0.933 (95% CI 0.922–0.944) and 0.937 (95% CI 0.927–0.947) for mass and nodule, respectively. In identifying consolidation, algorithm specificity was 0.927 (95% CI 0.897–0.954) and sensitivity was 0.594 (95% CI 0.500–0.688), compared with micro-average board-certified radiologist specificity 0.935 (95% CI 0.924–0.946) and sensitivity 0.456 (95% CI 0.418–0.495). The CheXNeXt algorithm detected effusion with a specificity of 0.921 (95% CI 0.889–0.951), higher than micro-average board-certified radiologist specificity of 0.883 (95% CI 0.868–0.898) while achieving a sensitivity of 0.674 (95% CI 0.592–0.754), comparable to micro-average board-certified radiologist sensitivity of 0.761 (95% CI 0.731–0.790). The results for the other 10 pathologies are shown in S1 Fig, and numerical values are provided in S1 File.
The effects of training set prevalence and the relabeling procedure on algorithm performance are illustrated in S2 Table. The algorithm performed significantly worse than radiologists on cardiomegaly, emphysema, and hernia, all of which had low prevalence in the original training set. On pneumonia, fibrosis, and edema, however, the algorithm performed as well as radiologists even though the prevalence of labels in the original training set was low. Our relabeling procedure resulted in an increase in the number of positive labels for every pathology. Using the new labels, the algorithm’s performance improved on 11 pathologies and worsened for 3 pathologies.
The mean proportion correct values with SDs of the algorithm and the radiologists are shown in S3 Table. The algorithm had a mean proportion correct for all pathologies of 0.828 (SD 0.12) compared with 0.675 (SD 0.15) and 0.654 (SD 0.16) for board-certified radiologists and residents, respectively. This indicates that over all 14 pathologies, the algorithm predictions agreed with the cardiothoracic specialist radiologists' findings more often than board-certified general radiologists (on average, 15.3% more often). S4 Table and S5 Table display additional performance and radiologist agreement results.
The average time for radiologists to complete labeling of 420 chest radiographs was 240 minutes (range 180–300 minutes). The deep learning algorithm labeled the same 420 chest radiographs in 1.5 minutes and produced heat maps highlighting areas of the image that are indicative of a particular pathology in 40 additional seconds. Fig 3 panels a and b show examples of heat maps for different pathologies, and more examples can be found in S2 Fig.
The results presented in this study demonstrate that deep learning can be used to develop algorithms that can automatically detect and localize many pathologies in chest radiographs at a level comparable to practicing radiologists. Clinical integration of this system could allow for a transformation of patient care by decreasing time to diagnosis and increasing access to chest radiograph interpretation.
The potential value of this tool is highlighted by the World Health Organization, which estimates that more than 4 billion people lack access to medical imaging expertise [30]. Even in developed countries with advanced healthcare systems, an automated system to interpret chest radiographs could provide immense utility [31,32]. This algorithm could be used for worklist prioritization, allowing the sickest patients to receive quicker diagnoses and treatment even in hospital settings in which radiologists are not immediately available. Furthermore, experienced radiologists are still subject to human limitations, including fatigue, perceptual biases, and cognitive biases, all of which lead to errors [33–37]. Prior studies suggest that perceptual errors and biases can be reduced by providing feedback on the presence and locations of abnormalities on radiographs to interpreting radiologists [38], a scenario that is well suited for our proposed algorithm.
An additional application for CheXNeXt is screening of tuberculosis and lung cancer, both of which use chest radiography for screening, diagnosis, and management [39–43]. The CheXNeXt algorithm detected both consolidation and pleural effusion, the most common findings for primary tuberculosis, at the level of practicing radiologists. Similarly, CheXNeXt achieved radiologist-level accuracy for both pulmonary nodule and mass detection, a critical task for lung cancer diagnosis, with much higher specificity than previously reported computer-aided detection systems and comparable sensitivity [44–47]. Although chest radiography is not the primary method used to perform lung cancer screening, it is the most common thoracic imaging study in which incidental lung cancers (nodules or masses) are discovered. For example, in a large study of incidentally discovered lung cancers in 593 patients, 71.8% were diagnosed incidentally on chest X-ray and the remaining on computed tomography (CT) scan [48]. This would suggest that, despite the recommendation and widespread use in modernized healthcare environments for the use of screening CT, chest radiographs remain the primary modality by which lung cancer is imaged. Additionally, lung cancers are sometimes diagnosed on chest CT and then identified in retrospect as “missed” on previous chest radiographs. This scenario is not rare and has a considerable medicolegal impact on the field of radiology. Furthermore, the vast majority of the world’s population does not have access to chest CT for lung cancer screening or diagnosis and therefore must rely on the versatile and less resource-intensive chest radiograph for the detection of thoracic pathologies, including lung cancer and tuberculosis. Once clinically validated, an algorithm such as CheXNeXt could have impactful clinical applications in healthcare systems.
While CheXNeXt performed extremely well in comparison to board-certified radiologists on acute diagnoses, it performed poorest in the detection of emphysema and hiatal hernia. The symmetric "global" radiographic appearance in emphysema (symmetric pulmonary overexpansion) may have been more challenging to recognize as opposed to asymmetric "localized" findings such as pulmonary nodule, effusion, or pneumothorax. In addition, hiatal hernia was the least prevalent of all the 14 labels in the training data. These shortcomings could be addressed in the future by obtaining more labeled training data for these pathologies.
Additionally, the sensitivity of board-certified radiologists in the detection of mass was low. To investigate this, we evaluated the sensitivity of the board-certified radiologists and algorithm after grouping the mass and nodule pathology classes as lung lesion (if the label was positive for either nodule or mass, the new label was positive for lung lesion; otherwise, it was negative). Before collapsing these classes, the board-certified radiologists achieved a sensitivity of 0.573 in detecting nodules and 0.495 in detecting masses. After collapsing, the board-certified radiologists achieved a sensitivity of 0.667 in the detection of lung lesions. This indicates that the board-certified radiologists frequently selected the nodule label when the ground truth was mass but did accurately detect a pulmonary lesion. CheXNeXt had higher sensitivities for mass and nodule than board-certified radiologists (0.754 and 0.690, respectively) and maintained a higher sensitivity (0.723) after grouping.
This study has limitations that likely led to a conservative estimate of both radiologist and algorithm performance. First, the radiologists and algorithm only had access to frontal radiographs during reading, and it has been shown that up to 15% of accurate diagnoses require the lateral view [1]. The lack of lateral views in the dataset may limit detection of certain clinical findings such as vertebral body fractures or subtle pleural effusions not detected on frontal views alone; future work may consider utilizing the lateral views when applicable for diagnosis and algorithm development. Second, neither CheXNeXt nor the radiologists were permitted to use patient history or review prior examinations, which has been shown to improve radiologist diagnostic performance in interpreting chest radiographs [49,50]. Third, the images were presented to the radiologists and the CheXNeXt algorithm at a resolution of 1,024 pixels and 512 pixels, respectively, and chest radiographs are usually presented at a resolution of over 2,000 pixels. Fourth, the reference standard was decided by a consensus of cardiothoracic radiologists, and no access to cross-sectional imaging, laboratory, or pathology data was available to determine the reference standard. The comparison to gold standard cases for all pathologies is outside the scope and purpose of this study. Instead, the goal is to evaluate the performance of a deep learning algorithm in diagnostic tasks on radiographs using a retrospective approach based on the interpretations of an expert panel compared with the interpretations of individual nonspecialist radiologists. Finally, consolidation, infiltration, and pneumonia are all manifestations of airspace opacities on chest radiographs yet were provided as distinct labels. While any given radiograph can be marked with one or more of these 3 labels, certain radiographic patterns of airspace opacities are characteristic of pneumonia and, when combined with clinical information, can determine the pneumonia diagnosis specifically. Even in the absence of clinical data, identifying airspace opacity patterns characteristic of pneumonia is useful, particularly in parts of the world where access to expert diagnostics is limited.
This work has additional limitations that should be considered when interpreting the results. This study is limited to evaluation on a dataset from a single institution, so future work will be necessary to address generalizability of these algorithms to datasets from other institutions. Additionally, the experimental design used to assess radiologists in this work does not replicate the clinical environment, so the radiologist performance scores presented in this study may not exactly reflect true performance in a more realistic setting. Specifically, disagreement in chest radiograph interpretation between clinical radiologists has been well described and would not always be interpreted as error in clinical practice, e.g., atelectasis is not always a clinically important observation, particularly if other findings are present. In that way, the labeling task performed by the radiologist readers in this study differs from routine clinical interpretation because in this work, any/all relevant findings in each image were labeled as present no matter the potential clinical significance. Finally, the primary performance metric comparison in this study required estimating the ROC for radiologists. While we assumed symmetry in the specificities and sensitivities, allowing for a better fit, we acknowledge that this is not a perfect comparison, and for this reason, we also provided a comprehensive view of how the algorithm compares to radiologists on 6 other performance metrics (Fig 2 and S1 Fig). All performance metrics and estimates of uncertainty should be taken together to better understand the performance of this algorithm in relation to these practicing radiologists.
We present CheXNeXt, a deep learning algorithm that performs comparably to practicing board-certified radiologists in the detection of multiple thoracic pathologies in frontal-view chest radiographs. This technology may have the potential to improve healthcare delivery and increase access to chest radiograph expertise for the detection of a variety of acute diseases. Further studies are necessary to determine the feasibility of these outcomes in a prospective clinical setting.
|
10.1371/journal.ppat.1007745 | STAT2 dependent Type I Interferon response promotes dysbiosis and luminal expansion of the enteric pathogen Salmonella Typhimurium | The mechanisms by which the gut luminal environment is disturbed by the immune system to foster pathogenic bacterial growth and survival remain incompletely understood. Here, we show that STAT2 dependent type I IFN signaling contributes to the inflammatory environment by disrupting hypoxia enabling the pathogenic S. Typhimurium to outgrow the microbiota. Stat2-/- mice infected with S. Typhimurium exhibited impaired type I IFN induced transcriptional responses in cecal tissue and reduced bacterial burden in the intestinal lumen compared to infected wild-type mice. Although inflammatory pathology was similar between wild-type and Stat2-/- mice, we observed decreased hypoxia in the gut tissue of Stat2-/- mice. Neutrophil numbers were similar in wild-type and Stat2-/- mice, yet Stat2-/- mice showed reduced levels of myeloperoxidase activity. In vitro, the neutrophils from Stat2-/- mice produced lower levels of superoxide anion upon stimulation with the bacterial ligand N-formylmethionyl-leucyl-phenylalanine (fMLP) in the presence of IFNα compared to neutrophils from wild-type mice, indicating that the neutrophils were less functional in Stat2-/- mice. Cytochrome bd-II oxidase-mediated respiration enhances S. Typhimurium fitness in wild-type mice, while in Stat2-/- deficiency, this respiratory pathway did not provide a fitness advantage. Furthermore, luminal expansion of S. Typhimurium in wild-type mice was blunted in Stat2-/- mice. Compared to wild-type mice which exhibited a significant perturbation in Bacteroidetes abundance, Stat2-/- mice exhibited significantly less perturbation and higher levels of Bacteroidetes upon S. Typhimurium infection. Our results highlight STAT2 dependent type I IFN mediated inflammation in the gut as a novel mechanism promoting luminal expansion of S. Typhimurium.
| The spread of invading microbes is frequently contained by an inflammatory response. Yet, some pathogenic microbes have evolved to utilize inflammation for niche generation and to support their metabolism. Here, we demonstrate that S. Typhimurium exploits type I IFN signaling, a prototypical anti-viral response, to foster its own growth in the inflamed gut. In the absence of STAT2-dependent type I IFN, production of neutrophil reactive oxygen species was impaired, and epithelial metabolism returned to homeostatic hypoxia. Consequently, S. Typhimurium was unable to respire in the absence of type I IFN, and failed to expand in the gut lumen. Furthermore, perturbation of the gut microbiota was dependent on type I IFN signaling. Taken together, our work suggests that S. Typhimurium utilizes STAT2-dependent type I IFN signaling to generate a niche in the inflamed intestinal tract and outcompete the gut microbiota.
| A healthy gastrointestinal microbiota is characterized by the dominance of obligate anaerobic members of the phyla Bacteroidetes and Firmicutes. The expansion of facultative anaerobic Enterobacteriaceae (phylum Proteobacteria) is considered a microbial signature for gut inflammation and dysbiosis [1, 2]. This signature is observed in severe human intestinal diseases including inflammatory bowel disease (IBD), [3–5] colorectal cancer [6] and necrotizing enterocolitis [7]. Several mechanisms by which the enteric pathogen, Salmonella enterica serovar Typhimurium, capitalizes on multiple processes induced by inflammation and outcompete the commensal have been described. Infection with S. Typhimurium starts with the invasion of intestinal epithelial cells using its type III secretion system (T3SS-1) [8]. After crossing the intestinal barrier, the bacterium is rapidly recognized by Pattern Recognition Receptors (PRRs), such as Toll-like receptors (TLRs) and Nod-like receptors (NLRs), and is internalized by macrophages or dendritic cells. In macrophages, S. Typhimurium survives using its T3SS-2 [9]. Epithelial invasion, recognition of Pathogen-Associated Molecular Patterns (PAMPs) and macrophage survival leads to the production of chemokines and cytokines triggering an inflammatory environment and acute colitis [10–12]. In the lumen, S. Typhimurium employs mechanisms to utilize unique respiratory electron acceptors (e.g. tetrathionate and nitrate) which are generated as byproducts of the inflammatory burst. Most commensal members of the microbiota are unable to metabolize nitrate and tetrathionate [13, 14]. As a result, S. Typhimurium outcompetes the healthy microbiota enabling its luminal expansion and eventually facilitating the transmission to subsequent hosts [13–16].
Although S. Typhimurium succeeds in expanding its luminal population during inflammation leading to a decline in the commensal microbiota, the coordinated actions of multiple immune cell defense pathways mediate the clearance of the pathogen. Activation of the Interferon (IFN) signaling pathway is critical for successful host defense against many infections. Type II IFN (IFN γ) plays a central role in generating inflammatory responses to clear S. Typhimurium [17–20]. However, the role of a closely related pathway involving the actions of type I IFNs (IFNs α and β) is less clear. Type I IFN signaling is well-documented as essential for mounting antiviral responses. Pre-exposure of cells to type I IFNs induces an antiviral state by blocking viral replication [21, 22]. It has recently become evident that activation of this pathway also plays a pivotal role during bacterial infections by acting directly or indirectly on many immune cell types including NK cells, T cells, B cells, Dendritic Cells (DCs), neutrophils and phagocytic cells. Depending on the bacterial agent, the role of type I IFNs exert seemingly opposing roles. For instance, while type I IFNs restrict the growth of Legionella pneumophila or Streptococcal species [23–27], activation of the same pathway impairs the clearance of intracellular Mycobacterium tuberculosis leading to tuberculosis [28, 29]. Recent studies highlighted the role of type I IFN signaling during systemic infection with S. Typhimurium. Mice deficient in type I IFN receptor (IFNAR), or IFN β exhibit greater resistance to S. Typhimurium [30]. Furthermore, type I IFNs are critical for inflammasome formation, caspase activation, and inflammatory cell death following infection with S. Typhimurium [31–34]. The role of this pathway during intestinal bacterial induced inflammation and the subsequent impact on the luminal bacterial population remains unclear.
IFNAR activation by type I IFNs (IFNs α and β) not only leads to the transcription of type I IFN stimulated genes (ISGs) induced by ISGF3, the heterotrimeric transcriptional complex composed of STAT1/STAT2/IRF9, but also by inflammatory gene activation via the formation of STAT1 homodimers. As STAT1 homodimers can also be activated by IFN γ, earlier studies that used Ifnar-/- or Stat1-/- mice did not clearly differentiate the contribution of each IFN pathway to driving inflammation (Fig 1). Here we used Stat2-/- mice, which causes the genetic ablation of type I IFN signaling, in combination with the streptomycin pretreated mouse model to pinpoint the role of type I IFNs in host response to Salmonella infection. Overall, we conclude that STAT2-driven type I IFN response leads to the transmigration of functional neutrophils into the lumen creating a microaerophilic environment, which enables the pathogen to outgrow the microbiota.
Type I IFNs released during bacterial infections may affect many arms of the immune response including inhibition of bacterial invasion, amplification of the immune response and production of antimicrobial genes. To investigate the possible role of a STAT2-dependent type I IFN signaling pathway during S. Typhimurium induced intestinal infection, wild-type C57BL/6, Stat1-/- (deficient in both IFN-α/β and IFN-γ signaling), and Stat2-/- (deficient only in IFN-α/β signaling) mice were orally infected with 109 CFU of S. Typhimurium following streptomycin pretreatment. All strains eventually succumbed to S. Typhimurium infection, but Stat2-/- mice survived significantly longer than wild-type and Stat1-/- mice (p = 0.0026) (Fig 2). This finding is notable because our previous study [35] reported increased mortality with Stat2-/- mice during LPS-induced sepsis suggesting that type I IFNs play different roles when compared between mucosal and systemic sites during infection.
To determine the role of STAT2 during S. Typhimurium infection, we first evaluated intestinal immune responses by analyzing gene expression by qPCR in the cecum of mice at 48 hours post infection, a time point where no animal death was observed and found to be optimal for investigating inflammatory responses [36]. When we examined the expression of genes that have previously been identified to be dependent on STAT2 [37], we found that there were significantly lower transcript levels of Irf7, Isg15, Oas1b, Rsad1, and IrgM1 in the cecum of infected Stat2-/- mice when compared to cecum of infected wild-type mice (Fig 3). We found no significant differences in genes known to be regulated by IFN γ and the IFNGR such as Cxcl10 (Fig 3). Furthermore, no differences in the transcription levels of genes previously shown to be important for S. Typhimurium infections including Tnfα, Il6, Ifnγ, and Mcp1 were observed between wild-type and Stat2-/- mice (Fig 4). This result was further confirmed when we analyzed the systemic cytokine responses in the serum using a cytometric bead assay. We did not observe a significant difference in the serum levels of TNFα, IFNγ, MCP1 (also known as CCL2), IL-12, IL-6 or IL-10 between wild-type and Stat2-/- infected mice (S1 Fig). Overall, these results show that type I IFN signaling is distinctively blocked in Stat2-/- mice as classical inflammatory gene expression was unaffected by this deficiency.
When we investigated the bacterial burdens at 48 hours post infection, the time point where we observed changes in immune responses, we found that there were significantly fewer bacteria in the cecum and colon contents of Stat2-/- mice compared to wild-type mice (Fig 5). No differences were observed in bacterial numbers in mesenteric lymph nodes (MLN). Although there was a trend towards lower numbers in spleen and liver at this time point (Fig 5), it was not statistically significant. The fact that a deficiency in STAT2 signaling leads to decreased bacterial burden specifically in the lumen suggests a role for a STAT2-induced inflammatory environment in S. Typhimurium expansion.
In response to infection with S. Typhimurium, neutrophils migrate into the tissue as well as the lumen [38, 39]. Studies using different pathogens have suggested that type I IFNs not only mediate the migration of neutrophils into the infection site but also enhance their function [40, 41]. To determine whether there was a defect in neutrophil migration as well as pathology, cecal tissue samples were fixed and stained with H&E. No differences were observed in overall histopathology between wild-type and Stat2-/- mice infected with wild-type S. Typhimurium at 48 hours (S2 Fig). Neutrophil numbers (PMN/field) were similar between wild-type and Stat2-/- mice infected with wild-type S. Typhimurium (Fig 6A). No differences in neutrophil abundance were noted when comparing uninfected wild-type and Stat2-/- (S3A Fig). However, when we quantified levels of myeloperoxidase (MPO), a neutrophil marker, we surprisingly found that there was less MPO in the cecal tissue of S. Typhimurium-infected Stat2-/- mice than infected wild-type mice (Fig 6B). This result was confirmed using immunohistochemistry with an MPO-specific antibody (Fig 6C). These results suggested that although the presence of type I IFNs does not affect the transmigration of neutrophils into the infection site, they somehow alter the function of these immune cells. This finding is somewhat surprising because neutrophils play a major role in clearing S. Typhimurium. The fact that there were more bacteria in the presence of neutrophils indicated to us that a novel mechanism allows this pathogen to thrive in the presence of neutrophils.
To determine if intestinal oxygenation caused the difference in bacterial burdens in the intestines of wild-type versus the Stat2-/- mice, we infected mice with the S. Typhimurium cyxA mutant. The cyxAB operon encodes a cytochrome bd-II oxidase enzyme that facilitates growth of S. Typhimurium under oxygen-limiting conditions [42–45]. CyxA is essential for S. Typhimurium survival in the post-antibiotic treatment model [45]. Mice were streptomycin pretreated and then orally administered a 1:1 mixture of wild-type S. Typhimurium and cyxA mutant. Four days post infection colon contents were collected for bacterial enumeration by determining colony forming units (CFU), and the competitive index (CI) was determined by dividing the output ratio (wild-type CFU/cyxA CFU) in the colonic contents of mice by the input ratio (wild-type CFU/cyxA CFU). Wild-type S. Typhimurium exhibited a fitness advantage over the cyxA mutant in wild-type mice (higher numbers of wild-type bacteria recovered), consistent with previous findings [45]; however, the cyxA gene provided no advantage in the Stat2-/- mice (both strains were recovered at same numbers) (Fig 6D). There was no observable phenotype in systemic sites such as the liver where there was no fitness advantage conferred by the cyxA mutant in either the wild-type or STAT2-/- mice (Fig 6E). These data suggested that oxygenation in the intestine of wild-type mice is different from that of Stat2-/- mice. This was confirmed using pimonidazole (PMDZ), a marker of hypoxia. Mice were injected intraperitoneally with PMDZ (Chemicon; 2.0 mg/20 g body weight in 100 μl PBS) 1 hour prior to euthanasia, and PMDZ was detected in tissue sections by immunohistochemistry. It was previously reported that hypoxia decreases in the intestine during S. Typhimurium infection [45]. We determined that the intestinal environment in Stat2-/- mice was more hypoxic than in wild-type mice as shown by higher levels of pimonidazole staining (red) (Fig 6F). Both wild-type and Stat2-/- mice showed comparable hypoxia staining without infection (S3B Fig).
Obligate anaerobes of the healthy gut microbiota were previously reported to become depleted from the microbiota at later stages of S. Typhimurium infection in streptomycin-treated mice through a neutrophil-dependent mechanism [46]. To determine if STAT2 signaling led to changes in the microbiota, we analyzed the phylogenetic composition of the intestinal microbial communities using 16S rRNA profiling (S4A Fig). We observed a drastic reduction in the relative abundance of Bacteroidetes phylum with approximately 10% remaining in wild-type mice infected with S. Typhimurium as opposed to 60% in uninfected wild type mice. No significant shifts were detected in the abundance of Bacteroidetes in Stat2-/- infected mice when compared to that of uninfected wild-type and Stat2-/- control mice (S4B Fig). Conversely to the CFU recovered from feces of wild-type and Stat2-/- mice upon S. Typhimurium infection (Fig 5), there was a significantly higher relative abundance of Proteobacteria observed in the wild-type infected mice with an average of 70% than in the Stat2-/- infected mice (30%; S4C Fig). To verify that the experimental changes we observed between wild-type and Stat2-/- mice was not due to differences in the overall microbiome content of the two strains of mice that arose because they were housed separately, we co-housed the mice starting at the day of weaning for 5 weeks. The co-housed wild-type and Stat2-/- mice were infected with S. Typhimurium following streptomycin pre-treatment. Forty-eight hours post infection, fecal and cecal contents were collected and the phylogenetic composition of the microbial communities at the phylum level was determined using 16S rRNA profiling (Fig 7A and 7B). Uninfected mice were also included as controls. Infection of wild-type mice with S. Typhimurium led to a reduction in the relative abundance of the Bacteroidetes phylum. The relative abundance of Bacteroidetes in Stat2-/- infected mice remained comparable to that of uninfected control mice (Fig 7C). It’s important to emphasize that the striking reduction in Bacteroidetes abundance we detected earlier in non-cohoused infected and non-infected wild type mice were no longer observed under co-housing conditions. Nevertheless, differences between infected wild type and Stat2-/- mice remained unchanged. Moreover, relative abundance of Proteobacteria was increased in the wild-type infected mice and this expansion was at a lower level in Stat2-/- infected mice (Fig 7D). Detailed microbial analysis revealed that the members of Proteobacteria that expanded in their relative abundance in infected wild-type mice belonged to the Salmonella genus (S5 Fig). This result was validated by directly enumerating S. Typhimurium in the colon contents of co-housed infected wild-type and Stat2-/- mice (Fig 7E). Overall, the results obtained in both experiments using co-housed and non-cohoused mice demonstrated that STAT2 enabled the expansion of S. Typhimurium. These data also strongly indicate that post-infection, Stat2-/- mice retained a protective microbiome against pathogenic bacteria.
To also confirm the previous findings on neutrophils (Fig 6), we performed a semi-quantitative analysis of overall pathology and quantified neutrophil abundance. There were no differences in the overall pathology (Fig 8A) and the neutrophil numbers between the S. Typhimurium infected wild-type and Stat2-/- mice (Fig 8B). The cecal neutrophils were also quantified using flow cytometry. Following the identification of live cells, neutrophils were identified as CD45+, CD3-, NK1.1-, B220-, Ly6G+ cells using the gating strategy described in S6 Fig. While there was an increase in percentage of neutrophils in wild-type mice infected S. Typhimurium compared to uninfected wild-type mice, there were no significant difference observed in the percentage of neutrophils when comparing S. Typhimurium infected wild-type and with S. Typhimurium infected Stat2-/- mice (Fig 8C). As we did not observe any differences between numbers of neutrophils transmigrating into the infection site between wild-type and Stat2-/- mice but there was a difference in MPO levels in the colon contents of these mice (Fig 6C), we next determined whether Stat2-/- neutrophils were functional. Bone marrow neutrophils from wild-type and Stat2-/- mice were stimulated with the Gram-negative bacterial ligand N-formylmethionyl-leucyl-phenylalanine (fMLP) in the absence or presence of IFNα. Superoxide anion generation was then measured. There were no differences in superoxide anion generation between the neutrophils of wild-type and Stat2-/- mice upon stimulation with fMLP. However, when the cells were pre-treated with a type I IFN, IFNα, there were reduced levels of superoxide anion generation in the neutrophils isolated from Stat2-/- mice compared to wild-type mice (Fig 8E)
The immune system deploys multiple mechanisms to eradicate invading microbes and infections. Induction of type I IFNs is a critical mechanism that the immune system exploits to fight viral infections. Type I IFNs (IFNs α and β) induce antiviral responses by binding to their cognate receptor IFNAR ubiquitously expressed on many cell types. The transcription factor STAT2 takes center stage in the type I IFN response as it is essential to mediate an antiviral state that helps the host clear a viral infection [47]. Research over the past few years has suggested that type I IFNs are intricate players during bacterial infections. Although type I IFN responses mounted against viral infections provide a common anti-viral state among a broad range of viral pathogens, the type I IFN response generated against bacteria varies based on the specific bacterial pathogen. Recently, it was reported that Ifnβ−/− mice exhibit greater resistance to oral S. Typhimurium infection and a slower spread of S. Typhimurium to distal sterile sites [30]. These results are consistent with our findings using Stat2-/- mice (Fig 2). Nevertheless, the previous study did not use streptomycin pre-treatment to induce colitis during infection, which models more accurately the course of S. Typhimurium infection. Hence the role of type I IFNs during gut inflammation and dysbiosis has remained unclear. Several studies have emerged showing that not all type I IFN responses involve the classical ISGF3 complex. STAT2 homodimers have been shown to bind IRF9 and activate ISG expression of antiviral genes in the absence of STAT1 [37, 48]. The expression of a subset of ISGs stimulated by STAT2 homodimers/IRF9 exhibits a delayed kinetics compared to the classical ISGF3, however, this is sufficiently robust to evoke an innate response [49]. These observations together with our own findings indicate that S. Typhimurium exploits the type I IFN pathway by relying on STAT2 and potentially in the absence of STAT1.
S. Typhimurium successfully establishes an infection with the coordinated actions of its two distinct populations; the first invades the tissue and increases inflammation while the second luminal population counter intuitively benefits from the generation of host derived nitrate and oxygen [14, 15, 50]. The regulatory host signaling pathways that control the availability of these electron acceptors are not known. Here, we determined that type I IFN pathway is activated during S. Typhimurium infection and leads to oxygenation of the gut mucosa allowing the pathogen to respire and expand its luminal population. As we observed blunted expression of type I IFN stimulated genes (Fig 3), similar numbers of neutrophils in the cecal mucosa but lower levels of MPO (Fig 6A and 6B), a neutrophil activation marker, in the cecum of Stat2-/- mice infected with S. Typhimurium, these results suggest that type I IFNs do not effect the migration of neutrophils to the site of infection but may effect the antimicrobial activity of these cells. It was previously established that upon activation, neutrophils release reactive oxygen species as antimicrobial measures. The release of reactive oxygen species also contributes to oxygenation of the lumen, and superoxide dismutases encoded by Salmonella allow the bacteria to detoxify the oxygen radicals promoting bacterial survival [51]. The competition experiments between wild-type S. Typhimurium and cyxA mutant as well as hypoxia staining confirmed that oxygenation of the gut lumen of Stat2-/- mice was lower compared to that of wild-type mice. Furthermore, our in vitro experiments confirmed that neutrophils from the Stat2-/- mice were blunted in their ability to generate superoxide anion. Overall, these results suggest that in response to S. Typhimurium, neutrophils invade the gut lumen and contribute to the oxygenation of the gut via a type I IFN mediated mechanism. In return, the professional pathogen S. Typhimurium takes advantage of this mechanism and expands its population.
One of the many benefits of the gut microbiota to the host is to limit the expansion of enteric pathogens. Gut microbiota provides metabolites such as butyrate that fuels colonocyte metabolism resulting in the consumption of oxygen, thereby rendering the lumen hypoxic. It was recently shown that the epithelial PPAR-γ-signaling pathway limits oxygenation of the gut epithelium in the presence of butyrate, which in turn limits the expansion of S. Typhimurium [50]. In addition to the neutrophils, we do not know whether STAT2 signaling may also have a direct effect on colonocyte metabolism (Fig 6D), which impacts the bioavailability of oxygen in the gut lumen. Previous studies have shown that the abundance of dominant microbial phyla, Bacteroidetes and Clostridia can directly be affected by drastic changes in the luminal environment during enteric infections [52–54]. The depletion of Clostridia from the microbiota at later stages of S. Typhimurium infection in streptomycin-treated mice through a neutrophil-dependent mechanism was reported [46]. Our results demonstrate that a STAT2 mediated type I IFN response triggered during S. Typhimurium infection directly affects Bacteroidetes phyla in the gut.
Our study highlights the importance of STAT2 signaling in neutrophils during Salmonella infection. To date, most of what has been described for STAT2 signaling in pathogenic infections was centered on immune cells, such as macrophages and dendritic cells. In models of viral infection, STAT2 signaling is exploited by measles virus and choriomeningitis virus to interfere with dendritic cell (DC) development and expansion [55]. Furthermore, STAT2 signaling in macrophages is critical to activate a transcriptional response and control early dengue virus replication [37, 56]. We speculate that STAT2 signaling in colonocytes during Salmonella infection is equally important as in neutrophils for crosstalk and the release of chemokines for the recruitment and activation of neutrophils and macrophages. Future studies involving a more detailed analysis are warranted to delineate the far-reaching effects of type I IFN signaling on both the microbiota and oxygenation by colonocytes and neutrophils.
Salmonella enterica serovar Typhimurium strain IR715, a fully virulent, spontaneous nalidixic acid resistant derivative of strain ATCC 14028, was grown in Luria-Bertani broth (LB) supplemented with 50 μg/ml nalidixic acid at 37°C [57]. Salmonella Typhimurium IR715 cyxA [45], generously provided by Andreas Baumler, was supplemented with 100 μg/ml carbenicillin LB broth and was grown at 37°C.
Eight- to ten-week-old female C57BL/6 (wild-type) mice were age and sex matched to mice deficient in STAT1 (kindly provided by Dr. David Levy, NYU) or STAT2 (generously provided by Dr. Christian Schindler on the SvJ background that we backcrossed 10 generations onto the B6 genetic background). All mice were bred at the animal facility of the Lewis Katz School of Medicine at Temple University. All mice were streptomycin treated prior to bacterial infection. Mice were monitored twice daily after infection. Humane terminal endpoints included inability to ambulate and/or labored breathing. Briefly, mice were inoculated intragastrically with 20 mg of streptomycin (0.1 ml of a 200 mg/ml solution in water) 24 hours prior to bacterial infection. Bacteria were grown with shaking in LB broth containing nalidixic acid (50 μg/ml) at 37°C overnight. For infection, groups of 3 to 5 mice were inoculated intragastrically with either 0.1 ml of sterile LB broth (mock infection) or 109 CFU of S. Typhimurium. Mice were sacrificed at indicated time points after infection. To determine the number of viable S. Typhimurium, samples of cecum (proximal section), liver, spleen, mesenteric lymph nodes, and colon contents were collected from each mouse and homogenized in 5 ml PBS. 10-fold serial dilutions were plated on LB agar plates containing nalidixic acid (50 μg/ml). The tip of the cecum was collected for histopathological analysis. The center section of the cecum was immediately snap-frozen in liquid nitrogen and stored at −80°C for RNA isolation. All animal experiments were repeated at least three times with identical results.
To determine to role of luminal oxygenation in bacterial survival, 24 hours prior to inoculation 6 to 8 week-old age matched wild-type C57BL/6 and Stat2-/- mice were orally gavaged 0.1 ml of a 200mg/ml streptomycin solution. Mice were orally infected with 108 bacteria in a 1:1 ratio of S. Typhimurium IR715:cyxA. Four days after infection mice were euthanized and colon contents, cecum and liver were collected to determine the CFU of IR715 and cyxA mutant. The cecum was snap frozen in liquid nitrogen and stored at -80°C for MPO ELISA, and sections of the colon were collected for histopathological analysis. Organs for bacterial enumeration were homogenized as mentioned above and plated on selective media using 10-fold serial dilutions. The competitive index (CI) was calculated as the ratio of recovered bacterial strains (output ratio) divided by the ratio present in the inoculum (input ratio). All animal experiments were at least repeated three times with identical results.
RNA was extracted from snap-frozen tissues or tissue culture cells using 1 ml TriReagent (Molecular Research Center, TR118) according to the manufacturer's protocol. RNA was then treated with DNase according to the manufacturer’s protocol (Ambion, AM1906). Reverse transcription of total RNA (1 μg) was performed in 25 μl volume according to manufacturer's instructions using the TaqMan Reverse Transcription Kit (Invitrogen, N8080234). Real-time PCR was performed using the SYBR green (Applied Biosystems, 4309155) or TaqMan (Applied Biosystems) according to the manufacturer's instructions. Real-time PCR was performed for each cDNA sample (5 μl per reaction) in duplicate using the Step One Plus real-time PCR system (Applied Biosystems). The primers sequences are listed in Table 1. Results were analyzed using the comparative ΔCT method. Data was normalized to Gapdh or β-actin for SybrGreen or TaqMan reagents, respectively. Fold increases in gene expression in infected or mock-infected Stat2-/- mice were calculated relative to the average level of the respective cytokine in the mock-infected wild-type mice.
Hypoxia studies were performed as described by the manufacturer’s instructions (Hypoxyprobe-1 Plus Kit, Chemicon, Temecula, CA, USA) [45]. One hour prior to euthanasia, wild-type and Stat2-/- infected mice were injected with 100 mg/kg of PMDZ diluted in DMSO. After euthanasia, colon samples were collected and fixed with 10% formalin. Unstained paraffin embedded tissue samples were probed with 1:50 FITC-conjugated IgG1 mouse monoclonal antibody clone 4.3.11.3 (Hypoxyprobe, Inc.), and stained with 1:150 Cy3 conjugated AffniPure Goat Anti Mouse IgG (H+L) (Jackson ImmunoResearch, 115-165-0003). Briefly, tissue sections were incubated at 50°C for 10 minutes and then deparaffinized by washing for 10 minutes with xylene 2x, 3 minutes with 95% ethanol 2x, 3 minutes with 80% ethanol 1x, and then rehydrated by washing with 70% ethanol 1x. The antigens were retrieved by incubating sections with 20μg/ml Proteinase K (Fisher, BP1700-100) in TE buffer (10mM Tris, 1mM EDTA, pH 8.0) for 15 min at 37°C in a humidified chamber. The slides were washed with PBS for 10 minutes and then blocked for 45 minutes with blocking buffer. Samples were incubated with the primary antibody 1: 50 FITC-conjugated IgG1 mouse monoclonal antibody clone 4.3.11.3 over night at 4°C in a humidified chamber. After PBS washing (5 minutes, trice), each slide was incubated with the secondary antibody 1:150 Cy3 conjugated AffniPure Goat Anti Mouse IgG (H+L) at room temperature for 90 minutes in a humidified chamber. DAPI (Invitrogen, P21490) was used as a counter-stain (1μg/ml, incubated at room temperature for 5 minutes in the dark). Slides mounted with Vectashield (Vector Labs, H-1000) and were visualized using an Olympus BX60 Fluorescent Microscope with Spot Insight2 camera at 10x magnification.
MPO activity in the cecal tissue was determined as previously described [58]. Snap frozen cecum samples were lysed by homogenizing in 0.5% HETAB (hexadecyltrimethyl ammonium bromide) in 50mM KPi (phosphate buffer) at a ratio of 0.1g sample per 1ml buffer. Master mixes were prepared by mixing 10 μL homogenized sample with 3 μL o-dianisidine hydrochloride (20mg/ml stock), 3 μL 20 mM hydrogen peroxide and 284μL 50mM KPi. Ten-fold serial dilutions of the MPO standard 100UG (Millipore, 475911) were prepared in the same fashion as the samples, with the top standard being 125μg/ml. Samples were plated in clear 96 well plate and incubated at 37°C for 10 minutes, taking absorbance measurements at 460 nm every 2 minutes for 10 minutes using a Flex Station, Molecular Devices plate reader. The reaction was halted by adding 3 μL of 30% NaN3 to each well and a final absorbance reading at 460nm was taken. The concentration of MPO was calculated using the absorbance values obtained from the standard curve.
To visualize the presence of MPO within the cecum of wild-type C57BL/6 and Stat2-/- mice, unstained paraffin embedded tissue sections were heated at 55–60°C for 30 minutes. The tissue was deparaffinized by washing in xylene 2x for 5 minutes, absolute ethanol 3x for 3 minutes, 95% ethanol 1x for 3 minutes, 90% ethanol 1x for 3 minutes, 70% ethanol 1x for 3 minutes. Antigens were retrieved by boiling slides in sodium citrate buffer (10mM Sodium Citrate, 0.05% Tween 20, pH 6.0), for 10 minutes, and allowing to cool to room temperature for 20 minutes. The samples were washed in Tris Buffered Saline, TBS (0.5M Tris Base, 9% NaCl) 1x for 5 minutes. The tissue was blocked in TBS supplemented with 3% BSA for 30 minutes and then incubated with 1:200 MPO Heavy Chain (L-20), (Santa Cruz sc-16129) in TBS supplemented with 3% BSA for two hours at room temperature. The samples were washed 3x for 5 minutes each in TBS supplemented with 3% BSA. The samples were then incubated for 40 minutes at room temperature with 1:1000 Rabbit anti-Goat (H+L) Super Clonal Secondary antibody, Alexa Fluor 488 conjugated (Fisher A27012). The slides were washed with 3x for 5 minutes each in TBS supplemented with 3% BSA. Samples were counter stained with 1 μg/ml DAPI (Invitrogen, P21490) and then washed in TBS supplemented with 3% BSA [59]. Slides were mounted with Vectashield (Vector Labs, H-1000) and visualized using Olympus BX60 Fluorescent Microscope with Spot Insight2 camera at 10x magnification.
DNA from fecal contents of wild-type and Stat2-/- infected mice was extracted using the PowerSoil DNA Isolation Kit (MoBio, 12888–50) according to manufacturer’s protocol. High quality isolated DNA was then submitted to SeqMatic for 16S rRNA V4 sequencing using the Illumina MiSeq platform. Data was then analyzed using Qiime pipeline as described [2]. For the co-housing, wild-type (two or three) and Stat2-/- mice (two or three) were placed in the same cages at the time of weaning and housed together for 5 weeks.
The immune cells from the cecal tissue were isolated using the mouse lamina propria dissociation kit (Miltenyi Biotech, 130-097-410) according to manufacturer’s protocol. Briefly, 1x106 cells were resuspended in PBS and stained with live/dead cell discriminator (Invitrogen, L34597) according to manufacturer’s protocol. Cells were then rinsed with PBS, spun down at 400g for 10 minutes and resuspended in 20ml of mouse Fc Block (Miltenyi,130-092-575). The cells were then incubated at room temperature for 15 minutes. The mouse Fc Block was left on the cells and the cells were then stained with CD45 PE-Cy7 Rat (Clone30-F11; Biolegend, 103114), CD3 APC Rat (Clone 17A2; Biolegend, 100236), B220 APC rat (Clone RA3-6B2, Biolegend, 103212), NK1.1 APC mouse (Clone PK136, Biolegend, 108710) and Ly6G Alexa 488 rat (Clone 1A8; Biolegend, 127626) diluted in fluorescence activated cell sorting (FACS) buffer according to the manufacturer’s instructions for 30 minutes at 4°C in the dark. The FACS buffer was composed of PBS, 0.5% BSA and 2% FBS. Cells were then rinsed with FACS buffer, spun down at 400g for 10 minutes and then resuspended in 100 ml of 4% paraformaldehyde (BD, 554655), which fixed the cells. Following a 20 min incubation in the dark at room temperature, the cells were washed with FACS buffer. Finally, the cells were spun down and resuspended in 400 ul of FACS buffer. Cells were analyzed on a BD FACS Canto flow cytometer (BD Biosciences) and analyzed using FlowJo software (TreeStar, Inc., Ashland, OR).
Mouse bone marrow neutrophils were isolated according to the method of Mocsai et al. [60]. Wild-type and Stat2-/- mice were euthanized and the femur and tibias from the hind legs harvested. Neutrophils were isolated by Percoll density gradient sedimentation, followed by hypotonic lysis to remove erythrocytes.
Superoxide anion (O2-) generation was measured spectrophotometrically as superoxide-dismutase (SOD)-inhibitable cytochrome c reduction. In 96 well plates, bone marrow neutrophils (1.5 X 106) from wild-type and Stat2-/- mice were activated with fMLP (10-8M) in the presence of 5 μg/ml cytochalasin B. For experiments examining the effect of type I IFN on fMLP-stimulated O2- generation, the neutrophils were pre-treated with IFNα (1000 units/ml) prior to the addition of fMLP. The generation of O2- was monitored over a 10 min time-period [61, 62].
For analysis of bacterial numbers, competitive indices, relative abundance of bacterial populations and fold changes in mRNA levels, values were converted logarithmically to calculate geometric means. Parametric test (Student t test) or one-way ANOVA test was used to calculate whether differences were statistically significant (P < 0.05) using GraphPad Prism software.
All animal experiments were performed in BSL2 facilities with protocols that are approved by AALAC-accredited Temple University Lewis Katz School of Medicine Institutional Animal Care and Use Committee (IACUC# 4561) in accordance with guidelines set forth by the USDA and PHS Policy on Humane Care and Use of Laboratory Animals under the guidance of the Office of Laboratory Animal Welfare (OLAW). The institution has an Animal Welfare Assurance on file with the NIH Office for the Protection of Research Risks (OPRR), Number A3594-01.
|
10.1371/journal.ppat.1000123 | Human Cytomegalovirus UL18 Utilizes US6 for Evading the NK and T-Cell Responses | Human cytomegalovirus (HCMV) US6 glycoprotein inhibits TAP function, resulting in down-regulation of MHC class I molecules at the cell surface. Cells lacking MHC class I molecules are susceptible to NK cell lysis. HCMV expresses UL18, a MHC class I homolog that functions as a surrogate to prevent host cell lysis. Despite a high level of sequence and structural homology between UL18 and MHC class I molecules, surface expression of MHC class I, but not UL18, is down regulated by US6. Here, we describe a mechanism of action by which HCMV UL18 avoids attack by the self-derived TAP inhibitor US6. UL18 abrogates US6 inhibition of ATP binding by TAP and, thereby, restores TAP-mediated peptide translocation. In addition, UL18 together with US6 interferes with the physical association between MHC class I molecules and TAP that is required for optimal peptide loading. Thus, regardless of the recovery of TAP function, surface expression of MHC class I molecules remains decreased. UL18 represents a unique immune evasion protein that has evolved to evade both the NK and the T cell immune responses.
| HCMV establishes a lifelong latent infection and causes serious disease in immunocompromised individuals. Cytotoxic T lymphocytes (CTL) and natural killer (NK) cells are the primary effectors for the immune defense against HCMV. However, HCMV has evolved to evade both the innate and adaptive cellular immunity to viral infection. HCMV US6 glycoprotein inhibits TAP function, resulting in down-regulation of MHC class I, while HCMV UL18 is an MHC class I homolog that functions as a surrogate to prevent host cell lysis. Despite significant sequence and structural homology between UL18 and MHC class I molecules, US6 down regulates surface expression of MHC class I, but not UL18. Here, we describe a mechanism by which UL18 circumvents the self-derived TAP inhibitor, US6. UL18 abrogates US6 inhibition of TAP-ATP binding and restores TAP-mediated peptide translocation, thereby making peptides available for the assembly and subsequent surface expression of UL18. Together UL18 and US6 inhibit binding of MHC class I to TAP, thus down regulating surface expression of MHC class I molecules. UL18 represents a unique immune evasion protein resistant to both the NK and T cell immune responses. Our data provide a molecular basis for persistent HCMV infection and will aid in the development of a therapeutic vaccine.
| Human cytomegalovirus (HCMV), a β-herpesvirus, is prevalent in human populations worldwide [1]. Commonly, HCMV infection remains in a chronic, latent state. However, HCMV reactivation can result in morbidity or death in immunocompromised individuals, such as organ transplant recipients or AIDS patients [2]. MHC class I-restricted CD8+ cytotoxic T lymphocytes (CTL) play a major role in controlling viral infections. CTLs recognize and lyse virus-infected cells through engagement of the T cell receptor with MHC class I molecules presenting viral antigens at the surface of infected cells [3].
Newly-synthesized MHC class I heavy chain associates with β2-microglobulin (β2m) to form a heterodimer. This heterodimer is recruited into the MHC class I peptide-loading complex, which consists of MHC class I, β2m, TAP, calreticulin, ERp57, tapasin, and protein disulfide isomerase [4],[5], that controls the optimal peptide loading into the peptide-binding groove of MHC class I molecules. Only peptide-loaded class I molecules exit from the endoplasmic reticulum (ER) for transport to the cell surface. Peptides are generated primarily in the cytosol by proteasomes and then translocated to the ER by TAP for binding to MHC class I molecules [6],[7]. The TAP1/TAP2 heterodimer is a member of the ATP-binding cassette transporter superfamily and utilizes energy from ATP hydrolysis to translocate peptides into the lumen of the ER [8].
With selective pressure from the host immune response, HCMV has evolved several gene products which interfere with antigen presentation and eventual cell surface expression of MHC class I molecules. HCMV encodes four individual gene products of the unique short region protein (US): US2, US3, US6, and US11. Each protein is independently able to reduce class I surface expression [9]. For example, US6 is an ER-resident glycoprotein that binds directly to TAP in the ER lumen [10],[11]. US6 blocks the binding of ATP by TAP1 through a conformational change and subsequently inhibits TAP-mediated peptide translocation to the ER [12].
Although interference in antigen presentation and consequent MHC class I down-regulation on the cell surface might allow infected cells to evade virus-specific CTL, down-regulation of MHC class I molecules makes these cells susceptible to lysis by NK cells, which target cells lacking MHC class I molecules [13]. HCMV encodes multiple genes that function to evade NK-mediated cell lysis of infected cells by distinct mechanisms [14]–[16], suggesting that NK cells are crucial components of the innate defense against HCMV. UL18 is a MHC class I homolog that acts as a decoy to block NK cell cytotoxicity [17]. UL18 binds LIR-1, an NK cell inhibitory receptor, with an affinity exceeding that of MHC class I to LIR-1 by greater than 1000-fold [18]. UL18 is a type I transmembrane glycoprotein that shares a high level of amino acid sequence identity with MHC class I [19],[20]. Like MHC class I molecules, UL18 can associate with β2m [21] and bind endogenous peptides that are similar to peptides loaded on an MHC class I molecule [22]. Surprisingly, despite the sequence and structural similarities between MHC class I molecules and UL18 [20], US6 down regulates only MHC class I and not UL18 [23].
Here, we describe a novel mechanism of action for UL18 that may account for the specific down-regulation of one homolog and not the other. UL18 restores the peptide transport activity of TAP by inactivating US6. In addition, UL18 impairs optimal peptide binding by MHC class I molecules by interfering with the assembly of the peptide-loading complex. Hence, even though TAP function is recovered, the surface level of MHC class I molecules remains down regulated. Our data provide insight into how the viral MHC class I homolog UL18 has effectively evolved to evade both NK and CTL immune responses.
We previously reported that unlike MHC class I expression, cell surface expression of HCMV UL18 is resistant to the self-derived TAP inhibitor US6 [23]. Considering both the temporal coexistence of the UL18 and US6 gene products during infection and sequence similarities between MHC class I and UL18, we raised the intriguing question of how UL18 avoids ‘self attack’ by US6. We first confirmed that US6 differentially regulates surface expression of MHC class I and UL18. We infected HeLa cells stably expressing US6 (HeLa-US6) or parental HeLa cells with either a recombinant UL18-expressing vaccinia virus (vvUL18) or a wild-type vaccinia virus (vvWT). To measure cell surface levels of MHC class I and UL18, we used the monoclonal antibodies (mAbs) W6/32 and 10C7 and flow cytometry analysis. Consistent with our previous observation [23], US6 down regulated the surface expression of MHC class I molecules (Fig. 1A, left column), but did not influence surface levels of UL18 (Fig. 1A, right column).
To date, the possible dependence of UL18 on TAP for surface expression has not been confirmed. To address this question, we assessed the effect of ICP47, the herpes simplex virus-derived TAP inhibitor [24],[25], on the surface expression of UL18. In contrast to US6, ICP47 binds to the peptide-binding site on the cytosolic side of TAP, thereby, blocking transport of peptide ligands into the ER [26],[27]. As expected, the surface level of MHC class I molecules was down regulated in cells stably expressing ICP47 (HeLa-ICP47) (Fig. 1B, left column). The surface level of UL18 was also reduced in HeLa-ICP47 cells (Fig. 1B, right column) unlike in HeLa-US6 cells, indicating the TAP-dependency of UL18 for surface expression. Surface levels of UL18 were significantly reduced in TAP-deficient mutant T2 cells compared to normal T1 cells (Fig. 1C), further supporting the dependence of UL18 on TAP for surface expression.
The above data suggest that UL18 modulates the inhibitory action of US6 on TAP activity. To test this hypothesis, we directly measured TAP-dependent peptide transport in cells expressing either UL18 alone, both UL18 and US6, or both UL18 and ICP47. A significant reduction in peptide translocation was observed in cells expressing the US6 protein alone (HeLa-US6) and cells expressing ICP47 alone (HeLa-ICP47). Expression of only the UL18 protein did not affect peptide transport. Interestingly, upon ectopic expression of UL18 in HeLa-US6, peptide translocation into the ER lumen was markedly restored (Fig. 2), suggesting that UL18 attenuates the inhibitory effect of US6 on TAP. UL18-induced recovery of TAP function was not observed in cells expressing ICP47 (Fig. 2). The peptide transport results correlate with the cell surface expression of UL18 molecules (Fig. 1A).
Peptide translocation by TAP occurs in two steps: peptide binding to TAP and peptide translocation involving ATP binding and hydrolysis [28],[29]. US6 inhibits TAP by preventing ATP from binding to TAP1 [12]. To examine whether UL18 affects the ATP binding of TAP1 inhibited by US6, we compared the ATP-binding capacity of TAP in HeLa-US6 with that in HeLa-US6 expressing UL18. Cells were lysed and incubated with ATP-agarose. Proteins bound to the ATP-agarose were eluted, separated by SDS-PAGE, and probed with anti-TAP1 antibody. Consistent with the results of the previous study [12], US6 inhibited ATP binding by TAP1 (Fig. 3A, lane 3). UL18 alone did not interfere with ATP binding of TAP1 (Fig. 3A, lane 2). Interestingly, inhibition of ATP binding to TAP1 by US6 was abolished upon expression of UL18 (Fig. 3A, lane 4), demonstrating that UL18 restores the ability of TAP to bind ATP. To further define the mechanism underlying the reversal of US6-mediated inhibition of ATP binding to TAP, we investigated the effect of UL18 on the physical association between US6 and TAP that is required for inhibition of TAP by US6 [11]. HeLa-US6 cells were infected with vvWT or vvUL18 and lysed in 1% digitonin, and cell lysates were coimmunoprecipitated with anti-US6 antibody. Anti-US6 immunoprecipitated proteins were resolved by SDS-PAGE, followed by immunoblotting with antibodies. UL18 disrupted the interaction between US6 and TAP1 while not affecting the association of US6 with TAP2 (Fig. 3B, first and second panels). Therefore, we conclude that UL18 induces the dissociation of US6 from TAP1, restoring TAP-mediated peptide transport.
How does UL18 specifically block the binding of US6 to TAP1 but not to TAP2? To address this question, we examined the interactions among MHC class I, TAP, and UL18 in UL18-expressing HeLa and HeLa-US6 cells. Coimmunoprecipitation and western blot analysis indicated that MHC class I molecules bind to both TAP1 and TAP2 irrespective of US6 (Fig. 3C, lane 3 and Fig. 3D, lane 3). In the absence of US6, UL18 also interacted with both TAP1 and TAP2 (Fig. 3C, land 4). Upon US6 expression, however, the interaction of UL18 with TAP2 was virtually undetectable, while its interaction with TAP1 remained unaffected (Fig. 3D, upper and lower panels, lane 4). Taken together, these results demonstrate that US6 binds more competitively to TAP2 than does UL18, while UL18 more competitively binds to TAP1.
Because the peptide supply was recovered upon coexpression of US6 and UL18, we expected that the surface level of MHC class I molecules would increase to the steady-state level in normal cells. However, surface levels of MHC class I molecules were not recovered (Fig. 4A). Expressing UL18 in HeLa-US6 cells further down regulated surface levels of MHC class I molecules when compared to those of parental HeLa-US6 cells (Fig. 4A, right), suggesting that down-regulation of MHC class I by UL18 and US6 might be additive. Next, we tested whether UL18 alone inhibits MHC class I surface expression. In the presence of UL18, we observed consistent down-regulation of the surface expression of MHC class I (Fig. 4B, upper right panel), albeit at significantly reduced efficiency relative to the extent of class I down-regulation observed in cells expressing US6 alone (Fig. 4A, left panel). As a control, the cell surface level of CD59 was not affected in the same cells (Fig. 4B, lower right panel), confirming that the specific effect of UL18 on the surface expression of MHC class I. Similar results were obtained using a transient transfection system. In agreement with a recent report that low levels of UL18 are detectable on the surface of UL18-transfected HeLa cells [30], we also detected transiently expressed-UL18 at the surface of HeLa cells (Fig. 4C, upper left panel). More importantly, surface levels of MHC class I molecules were reduced slightly upon transient expression of UL18 (upper right panel), similar to the vaccinia virus system. This result is consistent with the observation that surface levels of MHC class I are slightly increased in cells infected with HCMV AD169 mutant with the UL18 gene deleted [31]. Surface levels of MHC class I molecules were not recovered in HeLa-US6 cells transiently expressing UL18, as was the case with the vaccinia virus system (lower right panel). These results further eliminate the possibility that results obtained using vvUL18 are an over expression artifact in the vaccinia virus infection system.
To identify the molecular mechanisms for down-regulation of MHC class I molecules by UL18, we investigated the effect of UL18 on optimal peptide binding by MHC class I molecules. We assessed the extent of optimal peptide binding by an established assay that correlates thermostability with the affinity of MHC class I peptide cargo [32]. Radiolabeled detergent lysates of HeLa and HeLa-US6 cells infected with either vvWT or vvUL18 were incubated at 4°C, 37°C, or 50°C prior to immunoprecipitation with mAb W6/32. In vvUL18-infected HeLa cells relative to vvWT-infected cells, MHC class I molecules were less thermostable at all temperatures, with the most prominent difference observed at 37°C (Fig. 5A). In HeLa-US6 cells, we did not detect differences in thermostability in the absence or presence of UL18 (Fig. 5B). The lack of significant UL18 influence on thermostability in HeLa-US6 cells is presumably due to the overwhelming effect of US6. Collectively, our data suggest that UL18 inhibits optimal peptide loading of MHC class I molecules and subsequently down regulates surface expression of MHC class I molecules.
Proper formation of the MHC peptide-loading complex is required for effective peptide loading in the ER and subsequent expression of MHC class I molecules at the cell surface [33]–[35]. Thus, regardless of the recovery of the peptide supply in the ER with coexpression of UL18 and US6, down-regulation of MHC class I at the cell surface might represent a defect in the formation of the MHC peptide-loading complex. We initially examined whether the expression of components of the peptide-loading complex is affected by UL18 or UL18/US6. Expression of UL18 or US6, individually or in combination, did not affect the expression level of MHC class I, TAP1, TAP2, or tapasin as evidenced by western blotting of whole cell lysates (Fig. 6A). Since assembly of the peptide-loading complex involves physical associations among components of the peptide-loading complex, we tested whether UL18 interfered with associations among components of the peptide-loading complex. HeLa cells infected with vvWT or vvUL18 were lysed in 1% digitonin. Cell lysates were analyzed by immunoprecipitation and western blot. Coimmunoprecipitation experiments revealed a reduced interaction between MHC class I and TAP1 or tapasin in the presence of UL18 alone (Fig. 6B). Interestingly, in the presence of US6, the ability of UL18 to perturb the association between MHC class I and TAP1 or tapasin was further enhanced (Fig. 6C). UL18 alone, or together with US6, did not affect the association of tapasin and TAP1 (Fig. 6D). These results indicate that UL18 disrupts an interaction specifically at the MHC-TAP1 and MHC-tapasin interfaces, but not at the tapasin-TAP1 interface. Given the importance of the MHC class I/tapasin/TAP interaction in optimal peptide loading of MHC class I molecules [36],[37], these results raise the possibility that UL18 inhibits optimal peptide loading of MHC class I molecules by competitively binding TAP.
HCMV encodes multiple proteins that down regulate the surface expression of MHC class I molecules [38]. This down-regulation allows infected cells to evade recognition by CTL but renders them susceptible to NK cells. HCMV encodes the UL18 protein, an MHC class I homolog that may protect infected cells from NK-mediated attack [19]. Despite the sequence and structural similarities between MHC class I molecules and UL18, the reason why UL18 molecules are not affected by US6 remains unclear [23]. In this study, UL18 induced dissociation of US6 from TAP1 to restore TAP-mediated peptide transport. UL18 also impaired optimal peptide loading onto MHC class I molecules by blocking the assembly of the MHC class I peptide-loading complex. Therefore, in the presence of UL18 and US6, surface levels of MHC class I molecules remained down regulated in spite of TAP function recovery.
The folding, assembly, and regulation of UL18 remain largely uncharacterized. UL18 loads peptides from endogenous protein precursors [22]. However, this precursor-peptide relationship does not distinguish between TAP-dependent or -independent peptide presentation. Our data demonstrate that normal surface expression of UL18 requires TAP function, as evidenced by the reduced surface expression of UL18 in TAP-deficient T2 cells (Fig. 1C). This TAP dependency was further confirmed by the down-regulation of UL18 surface expression by HSV-derived cytosolic TAP-inhibitor ICP47 [24],[25] (Fig. 1B).
Our data provide an example of how the immune evasion gene products of HCMV have developed delicate mechanisms to avoid ‘self-attack’ and specifically target host proteins. Despite the TAP-dependency of UL18 for surface expression, surface expression of UL18 was insensitive to the self-derived TAP inhibitor US6 (Fig. 1). In contrast, the nonself-derived TAP inhibitor ICP47 down regulated surface expression of UL18. UL18 attenuated the inhibitory effect of US6 on TAP, resulting in the recovery of TAP function. UL18 then utilized the TAP-transported peptides for binding and subsequent surface expression. Given that UL18 inhibited the interaction between US6 and TAP1, but not between US6 and TAP2 (Fig. 3B), and that UL18 interacted with TAP1 (Fig. 3, C and D), an explanation for reversal of US6-mediated inhibition of TAP by UL18 might be that binding of UL18 to TAP1 causes conformational changes in TAP1 inducing dissociation from US6. The luminal domain of US6 associated with TAP [10]–[12], and under our experimental conditions, we did not detect the association of UL18 with US6. Thus, dissociation of US6 from TAP1 seems to occur indirectly as a consequence of UL18 binding to TAP1. Since US6 inhibits ATP binding to TAP1 but does not affect ATP binding to TAP2 [12], dissociation of US6 from TAP1 by UL18 would sufficiently restore TAP function. Our data show that TAP1 exhibits a preference for binding UL18 over US6 (Fig. 4, C and D). In contrast, TAP2 preferentially binds US6 over UL18 (Fig. 3, B and D). This differential affinity might be the molecular basis by which UL18 specifically modulates only the US6-TAP1 but not US6-TAP2 interaction (Fig. 7). The functional importance of physical association between TAP2 and US6 is not yet clear.
Interestingly, surface expression of MHC class I molecules was not recovered although the peptide supply was recovered upon coexpression of US6 and UL18 (Fig. 4A). Down-regulation of MHC class I in cells expressing UL18 alone was consistently observed in repeated independent experiments (Fig. 4B upper panel), although to a much lesser extent compared with the reduction levels induced by the US2, US3, US6, and US11 gene products [11],[23]. This finding is consistent with the observation that surface MHC class I expression is slightly increased in cells infected with UL18-deleted HCMV AD169. Depletion of the β2m pool by UL18 is not likely a mechanism for reducing the surface expression of MHC class I molecules, because high levels of free β2m are observed in HCMV-infected cells [39]. Instead, our findings suggest that UL18 and US6 interfere with the assembly of the MHC peptide-loading complex, despite the recovery of peptide supply upon coexpression of UL18 and US6. UL18 inhibited the physical interaction of MHC class I with both TAP1 and tapasin (Fig. 6, B and C) that is required for peptide loading and subsequent expression of MHC class I molecules on the cell surface, while the interaction between tapasin and TAP1 was not affected by UL18 (Fig. 6D). Given that tapasin is a known TAP stabilizer [40], preserving the tapasin-TAP interaction may allow UL18 to acquire peptide ligands. Due to the ability of UL18 to modulate US6, UL18 may still acquire peptides and serve as an inhibitory ligand for NK cells on the infected cell surface. In concert, UL18 might collaborate with US6 to disrupt assembly of the peptide-loading complex for CTL evasion.
The current findings may be of physiological relevance in the context of HCMV infection, if the temporal expressions of UL18 and US6 coincide during infection. US6 mRNA is expressed 8 to 120 h postinfection, and UL18 mRNA is detected later, from 54 to at least 120 h postinfection [23], indicating that indeed UL18 coexists with US6. Recently, surface-expressed UL18 on HCMV-infected fibroblasts was detected during this time frame [30]. As not all MHC class I alleles require the TAP-scaffolded peptide-loading complex for peptide loading [41], dissociation of MHC class I from TAP by UL18 might not be sufficient for evading CTL recognition. HCMV encodes a series of US gene products that function to down regulate the cell surface expression of MHC class I [42],[43]. Compared to the degree of MHC class I down-regulation by the US gene products, the effect of UL18 alone on MHC class I down-regulation was negligible (Fig. 4, B and C). Therefore, evading CTL immunity appears not to be the natural function of UL18. Instead, the physiological role of UL18 in modulating CTL recognition might be important, particularly in the context of US6 expression. Our study shows that UL18 utilizes US6 for its surface expression without compromising the ability of US6 to inhibit the cell surface expression of MHC class I. Thus, UL18 is a unique protein that has evolved to evade both the NK and the T cell immune response.
HeLa cells were obtained from the American Type Culture Collection (ATCC, Manassas, VA) and cultured in DMEM (Life Technologies, Gaithersburg, MD) supplemented with 10% FBS (HyClone Laboratories, Logan, UT), 2 mM L-glutamine, 50 U/ml penicillin, and 50 µg/ml streptomycin. Human foreskin fibroblast (HFF) cells (passage 8–10) were grown in DMEM supplemented with 10% FBS under 5% CO2 at 37°C. Stable transfectants expressing tetracycline-inducible US6 (HeLa-US6) were described previously [11]. Transient transfection was performed using Lipofectamine (Invitrogen) according to the manufacturer's instructions. Recombinant vaccinia virus expressing UL18 was produced by cloning cDNA encoding the respective gene behind the early/late vaccinia virus p7.5 promoter into a modified pSC11 plasmid as previously described [44], and each plaque was purified three times in thymidine kinase-deficient 143B cells under bromodeoxyuridine selection (50 µg/ml). Cells were infected with recombinant vaccinia viruses at a multiplicity of infection of 10 for 1 h at 37°C in 500 µl of PBS supplemented with 10% BSA (Sigma-Aldrich, St. Louis, MO).
mAb 10C7 recognizing UL18 was purchased from the ATTC. K455 recognizes MHC class I heavy chain and β2m in both assembled and nonassembled forms [45]. Anti-MHC class I polyclonal Ab H-300 (sc-25619) was purchased from Santa Cruz Biotechnology (Santa Cruz, CA). Normal mouse IgG was purchased from Sigma-Aldrich. Polyclonal antisera specific for US6, human TAP1, and TAP2 raised against synthetic peptides have been described [11],[46]. Anti-CD59 antibodies were obtained from BD Biosciences (Mountain View, CA).
The surface expression of UL18 and MHC class I molecules was determined by flow cytometry (FACSCalibur; BD Biosciences, Mountain View, CA). Cells were washed twice with cold PBS containing 1% BSA and then incubated for 1 h at 4°C with either 10C7 for UL18 or W6/32 for MHC class I molecules. Normal mouse IgG was used as a negative control. The cells were washed twice with cold PBS containing 1% BSA and then stained with FITC-conjugated goat anti-mouse IgG for 40 min. A total of 10,000 gated events were collected by the FACSCalibur cytometer and analyzed with CellQuest software (BD Biosciences).
For coimmunoprecipitation, cells were lysed in 1% digitonin in PBS supplemented with protease inhibitors. After preclearing, samples were incubated with the appropriate antibodies for 2 h at 4°C, before Protein G-Sepharose beads were added. Beads were washed four times with 0.1% digitonin, and bound proteins were eluted by boiling in SDS sample buffer. Proteins were separated by SDS-PAGE, transferred onto a nitrocellulose membrane, blocked with 5% skim milk in PBS with 0.1% Tween 20 for 2 h, and probed with the appropriate antibodies for 4 h. Membranes were washed three times in PBS with 0.1% Tween 20 and incubated with horseradish peroxidase-conjugated streptavidin (Pierce) for 1 h. The immunoblots were visualized with ECL detection reagent (Pierce).
The peptide transport assay was performed as described [42]. HeLa and HeLa-US6 cells infected with vvWT or vvUL18 were permeabilized with Streptolysin-O for 20 min at 37°C and incubated for 10 min at 37°C in 5 µM of an FITC-conjugated peptide (RYNATGRL) in the presence of 1 mM DTT and 10 mM ATP. Following lysis in 1% NP-40, the glycosylated peptide fraction was isolated using ConA-Sepharose beads, and fluorescence of the peptides eluted from ConA was measured using a fluorescence reader (HTS 7000 Bio Assay Reader; Perkin-Elmer, Norwalk, CT). The ATP-agarose binding assay was performed as described [12]. Cells infected with either vvWT or vvUL18 were used for analysis. Thermostability assays were performed essentially as described [32].
|
10.1371/journal.pgen.1004153 | A Cohesin-Independent Role for NIPBL at Promoters Provides Insights in CdLS | The cohesin complex is crucial for chromosome segregation during mitosis and has recently also been implicated in transcriptional regulation and chromatin architecture. The NIPBL protein is required for the loading of cohesin onto chromatin, but how and where cohesin is loaded in vertebrate cells is unclear. Heterozygous mutations of NIPBL were found in 50% of the cases of Cornelia de Lange Syndrome (CdLS), a human developmental syndrome with a complex phenotype. However, no defects in the mitotic function of cohesin have been observed so far and the links between NIPBL mutations and the observed developmental defects are unclear. We show that NIPBL binds to chromatin in somatic cells with a different timing than cohesin. Further, we observe that high-affinity NIPBL binding sites localize to different regions than cohesin and almost exclusively to the promoters of active genes. NIPBL or cohesin knockdown reduce transcription of these genes differently, suggesting a cohesin-independent role of NIPBL for transcription. Motif analysis and comparison to published data show that NIPBL co-localizes with a specific set of other transcription factors. In cells derived from CdLS patients NIPBL binding levels are reduced and several of the NIPBL-bound genes have previously been observed to be mis-expressed in CdLS. In summary, our observations indicate that NIPBL mutations might cause developmental defects in different ways. First, defects of NIPBL might lead to cohesin-loading defects and thereby alter gene expression and second, NIPBL deficiency might affect genes directly via its role at the respective promoters.
| The cohesin complex is crucial for chromosome segregation during cell divisions but was recently also implicated in transcriptional regulation and chromatin architecture. Cohesin's binding to chromatin depends on NIPBL, a factor that was found to be mutated in 50% of the cases of the human developmental disorder Cornelia de Lange Syndrome (CdLS). To understand the role of NIPBL for cohesin, we need to know when and where the cohesin is loaded onto DNA. Our experiments have identified high-affinity NIPBL binding sites in different cells lines which do not overlap with cohesin-binding, but colocalize with specific transcription factors at active promoters. The activity of the respective genes depends on NIPBL but not cohesin. This is in contrast with other published data showing colocalization of NIPBL and cohesin, and we reveal the existence of different types of NIPBL binding sites that are detected differently by the antibodies used in the different studies. Our observations reveal a dual role for NIPBL in cohesin loading and as potential transcription co-factor, which yields novel insights into how NIPBL defects could cause Cornelia de Lange Syndrome since NIPBL mutations might directly influence developmentally important genes.
| Genomes need to be stably inherited over numerous cell generations. For each cell division the genetic information has to be replicated, the copies identified and then equally distributed between daughter cells. This process crucially depends on the cohesin complex, consisting of the core subunits SMC3, SMC1A, RAD21, SA1/STAG1 or SA2/STAG2 and several transiently associated regulatory proteins (reviewed in [1]). Cohesin tethers two sister chromatids together from S-phase on, allowing for their proper segregation in mitosis. Furthermore, cohesin is important for DNA damage repair (for review see [2]), for chromatin insulation in cooperation with the chromatin insulator protein CCCTC-binding factor (CTCF) [3]–[5], for chromosomal long-range interactions [6]–[8], and for development [9]–[12]. The latter functions implicate cohesin in regulating gene expression; indeed, a large number of genes are misregulated after cohesin depletion [3], [13].
How exactly cohesin associates with DNA is not understood, since none of the subunits binds directly to DNA. Rather, cohesin is hypothesized to bind to DNA by embracing the DNA strands with a “protein ring” formed by the core subunits [14], [15].
Cohesin's binding to chromatin is tightly regulated throughout the cell cycle. To enable chromosome segregation it is removed from chromosomes during mitosis. A prophase pathway depending on WAPL and specific phosphorylation of cohesin subunits dissociates cohesin from chromosome arms. The remaining cohesin is removed by proteolytic cleavage of the RAD21 subunit at anaphase onset (reviewed in [1]). Cohesin re-associates with chromatin at the G1-S-phase transition in yeast but in vertebrates already earlier during G1 phase.
The chromosomal localization of cohesin is determined by several factors. First, the cohesin loading factors NIPBL (also known as IDN3 or Delangin; Nipped-B, Drosophila melanogaster; Scc2, Saccharomyces cerevisiae) and MAU2 (also KIAA0892; Scc4 in Saccharomyces cerevisiae) are crucial for the re-loading of cohesin in G1-phase after its complete dissociation from chromatin during mitosis (reviewed in [1]). In yeast, it has been shown that cohesin associates first with Scc2 binding sites and then relocalizes to different positions [16], [17]. In Drosophila melanogaster cohesin colocalizes with NIPBL to actively transcribed genes [18] and in mouse ES cells a subset of cohesin binding sites was described to colocalize with NIPBL and the mediator complex [13]. Second, factors co-localizing with cohesin on chromatin such as CTCF [3] and Estrogen receptor [19] determine where cohesin is positioned.
Mutations in NIPBL and cohesin subunits, have been linked to the “Cohesinopathy” Cornelia de Lange syndrome (CdLS, OMIM #122470, #300590 and #610759). This dominant, genetically heterogeneous developmental disorder has a high degree of variability in its clinical presentation with multiple organ systems affected. It is estimated to occur in 1∶60000 to 1∶45000 live births. Characteristic features include craniofacial anomalies, growth retardation, intellectual disability, upper limb defects, hirsutism, and involvement of the gastrointestinal and other visceral organ systems [16]. Clinically, CdLS phenotypes can range from very mildly affected (no structural abnormalities, minor intellectual disability) to severely affected (upper limb defects, severe intellectual disability). Heterozygous mutations of NIPBL, ranging from nonsense and frameshift mutations to truncation mutations, have been found in 50% of CdLS patients and mutations of the cohesin subunits SMC1A, and SMC3 were found in another 5% (reviewed in [17]). Observations in patients and mouse models show that in cells with heterozygous NIPBL mutations the NIPBL transcript levels are only reduced by ∼30% due to an increased expression from the intact allele [18], [19]. A clinical phenotype is observed with a modest 15% reduction in expression [20]. This indicates that NIPBL expression levels are tightly regulated and are critical for cells. Defects in cohesin-dependent chromosome cohesion were not observed at this level of NIPBL reduction in CdLS patients or any model systems [19], [21]. However, a reduction of cohesin binding sites was observed in cells derived from CdLS patients, which was most obvious in close proximity to genes [18]. This suggested that the clinical features of CdLS are the collective outcomes of changes in the expression level of multiple genes during development.
NIPBL has already been linked to gene regulation. In Drosophila, NIPBL was found to facilitate the activation of the cut and Ultrabithorax genes by remote enhancers. In the case of the cut gene, NIPBL facilitates its long-range activation while cohesin has an inhibitory effect on cut expression [22]. Further, human NIPBL was already shown to bind histone deacetylases (HDAC1, HDAC3) [23] and heterochromatin protein 1 (HP1) [24].
These observations implied a “dual role” for NIPBL, in loading cohesin and in gene regulation. It is not known whether these two functions are independent of each other, or if NIPBL mediates gene regulation via loading of cohesin onto DNA.
In this study we have aimed to determine when and where NIPBL binds to chromatin to determine where cohesin is initially loaded. Furthermore we wanted to elucidate whether the position of NIPBL binding in the genome accounts for the altered gene expression patterns observed in CdLS patients carrying NIPBL mutations [18].
To gain insight into the cohesin loading mechanism it is crucial to understand when cohesin interacts with these factors during the loading process. We have therefore compared the timing of the chromatin-localization of cohesin with that of NIPBL and CTCF. Mitotic HeLa cells were fixed with paraformaldehyde (PFA) and immunostained with antibodies specific for CTCF, NIPBL and the cohesin subunits RAD21 and SA2/STAG2 (Fig. 1; Suppl. Fig. S1B, C). Specificity of the antibodies was demonstrated by immunostaining of siRNA-depleted cells (Suppl. Fig. S1A). It was then determined at which stage the signals of these proteins appeared on chromatin during the exit from mitosis (Fig. 1). These results were also correlated with the reassembly of the nuclear envelope in HeLa cells expressing Lamin B-EGFP. Similar to cohesin we find the signals of NIPBL and CTCF to be largely excluded from metaphase chromosomes. However to our surprise both NIPBL and CTCF signals appear on chromatin at an earlier stage of the mitotic exit than cohesin (Fig. 1), actually before the nuclear envelope is reassembled as shown by comparison to Lamin B signals (Suppl. Fig. S1B). Therefore NIPBL and CTCF are already present on chromatin, before the cohesin complex begins to re-associate with chromatin. This suggests that NIPBL binds first to chromatin and subsequently recruits cohesin. The fact that CTCF associates with chromatin before cohesin enforces our earlier observation that cohesin is dispensable for CTCF localization on chromatin [3].
To analyze the genomic localization of NIPBL binding sites relative to cohesin and CTCF, we selected the NIPBL antibody (referred to as NIPBL#1) that performs best in human cells (Suppl. Fig. S2) and performed ChIP-sequencing for NIPBL, cohesin and CTCF using HB2 cells (1-7HB2) [25] enriched in G1 phase (Suppl. Fig. S3A) and for NIPBL in lymphoblastoid cells (LCL; B-cell population immortalised by EBV-transformation) derived from a normal control (N5) and CdLS patients (PT1, PT9).
Furthermore, we have determined the transcriptional activity by RNA-sequencing, and identified active transcription start sites in HB2 cells by ChIP-sequencing of RNA Polymerase II (RNA Pol II). ChIP for NIPBL, SMC3, CTCF and RNA Pol II was performed as described [3], but for SMC1A ChIP a SDS-free protocol was used to maximize the ChIP-efficiency [26].
To prove the specificity of the identified peaks for NIPBL we have depleted NIPBL by RNAi and observed greatly reduced ChIP-qPCR signals for the analysed sites (Suppl. Fig. S4 A–C).
Using the criteria described in the Materials and Methods section, we identified 1138 NIPBL sites, 35668 CTCF sites, 22572 SMC3 sites and 29441 SMC1A sites in HB2 cells and between 1600 and 2000 NIPBL sites in lymphoblastoid cells (LCL). The data from the different LCL's and the conclusions for CdLS are discussed in detail in a later section.
Surprisingly, in HB2 cells the NIPBL binding sites do not overlap with cohesin or CTCF binding sites (Fig. 2A). Heatmaps centred on NIPBL (Fig. 2B), cohesin or CTCF binding sites (Fig. 2C, D), show no overlap of cohesin or CTCF signals with NIPBL sites. As expected, there was a high correlation between cohesin and CTCF signals. The absence of overlapping NIPBL and cohesin sites was confirmed by qPCR analysis of several NIPBL and cohesin binding sites in SMC3 and NIPBL ChIP experiments, where we observe only background levels of NIPBL binding on cohesin sites and vice versa (Fig. 2E).
Cohesin binding was previously observed on centromeric repeats and Alu elements [27]–[29], therefore we also analysed sequencing reads mapping uniquely to repeat sequences (Table S8). NIPBL ChIP highly enriches rRNA repeats (13 fold), in particular the large (LSU, 15 fold enriched) and small subunit (SSU, 14 fold enriched) repeat families but not at the repeat classes described for cohesin. rRNA repeats are pseudogenes of unknown function distributed all over the human genome [30]. In total we observe NIPBL at 20 out of 467 known LSU/SSU regions (Hg19 assembly of the human genome) and by ChIP-qPCR with primers specific for LSU and SSU repeats we confirmed NIPBL-binding to four of five LSU repeat regions and one of three SSU regions (Suppl. Fig. S5A).
The missing colocalization between NIPBL and cohesin is in contrast with observations in mouse embryonic stem cells (mESC) [13]. To address this we critically reviewed the ChIP-sequencing data analysis from Kagey et al., the ChIP protocols used and the different antibodies, NIPBL#1 from our study and NIPBL#6 used by Kagey et al.. Our review of the ChIP-seq data analysis from Kagey et al. confirmed their general finding that cohesin and NIPBL ChIP signals overlap, although we did not find such a colocalization of NIPBL and cohesin in our study. Further, we compared the different ChIP protocols by performing ChIP from mESC using both protocols and both antibodies (Suppl. Fig. S5B, C). We observe a better ChIP/IgG-control ratio using our protocol, which includes a more stringent washing of the beads (Suppl. Fig. S5C). For three NIPBL sites at promoters (Nanog, Lefty, Oct4), identified by Kagey et al. in mESC [13], both antibodies perform weakly but equally well, independent of the ChIP protocol (Suppl. Fig. S5B, C). To demonstrate once more the specificity of both antibodies for NIPBL, we have performed ChIP with both antibodies from control mESC and mESC derived from a Nipbl+/− mouse embryo (Suppl. Fig. S5D) (S. Goldberg, F. Grosveld unpublished data) and observe with both antibodies a 20–40% decreased Nipbl binding at all sites (Suppl. Fig. S5E). This is consistent with previous reports on Nipbl+/− mESC that heterozygous knockout cells still have 70% of wild-type Nipbl mRNA levels [19].
However, on the NIPBL binding sites that we find to be conserved between human HB2 cells and mES (Tiam1, Ankhd1, Sp1), the ChIP is strikingly better enriched for NIPBL#1 than NIPBL#6 in both cell types (Suppl. Fig. S5B, C, F). Therefore, different chromatin morphologies between pluripotent and differentiated cells do not account for the different binding patterns.
We conclude from these results that there are two different types of NIPBL binding sites. The NIPBL#1 antibodies highly enrich for a set of “major sites” that localize at promoters and do not overlap with cohesin. The NIPBL#6 and NIPBL#1 antibodies both detect a set of low-enriched sites (“minor sites”, low ChIP/seq signals) which overlap with cohesin binding sites.
NIPBL “major binding sites” are distributed over the entire genome (repetitive sequences were omitted during the mapping of the reads to the genome) but localize very specifically to the promoter area (+/−1000 bp from transcription start sites) (Fig. 3A). We observe such localization for 912 of 1138 (80%) NIPBL sites in HB2 cells, while only ∼10% of the cohesin and CTCF sites localize to promoters. About 89% of NIPBL-bound promoters are CpG island promoters (Table S4). Analysis of RNA-sequencing data from HB2 cells revealed that >98% of these NIPBL-bound genes are actively transcribed (Fig. 3A and Table S3), indicating a preferential binding of NIPBL to active promoters. Comparison with RNA Pol II binding sites showed that NIPBL preferentially binds 100–200 nucleotides upstream of RNA Pol II (Fig. 3B). This correlation is also visible as bimodal distribution of the RNA Pol II signal since orientation of transcription was not considered in this plot (Fig. 2B).
To analyse the properties of NIPBL binding sites further, we used the NIPBL binding sites observed in the control LCL's (N5), since a large number of data for histone modifications and transcription factors is available for lymphoblastoid cells like GM12878 from earlier publications [31] and ENCODE [32].
Comparing the pattern of different histone modifications around NIPBL sites, we observed that the sites are flanked by histone marks linked to active promoters and enhancers (H3K4me3, H3K27ac and H3K9ac) (Fig. 3C). However, the H3K4me1 mark, characteristic for enhancers, does not show enrichment (Fig. 3C). NIPBL itself apparently resides in nucleosome-free areas.
The missing enhancer-specific histone mark is in contrast with observations in mouse ES cells showing a colocalization of NIPBL with enhancers and cohesin [13]. Therefore we also compared the NIPBL binding with the enhancer marker p300 (Fig. 3D) and the cohesin subunit RAD21 (Fig. 3E) and again observed no correlation.
Motif analysis of NIPBL binding sites in HB2 cells and LCL's using MEME [33] reveals that the motifs for the transcription factor NFYA (subunit of the NF-Y complex) are present at 80% of NIPBL sites and for SP1 at 50% of the sites (Fig. 3F). NF-Y binds the CCAAT box, which correlates well with the presence of CpG islands at promoters; also, a connection between NF-Y and SP1 has often been reported with presence of both motifs at the same promoter. To test whether the presence of the NFYA motif is correlated to the CpG-island promoter or a genuine property of the NIPBL-bound promoters we analyzed NIPBL-bound CpG island promoters versus randomly selected CpG island promoters and observe a statistical significant preference (Fisher test, p<0.001) of NFYA for NIPBL-bound CpG island promoters. ChIP with anti NFYB antibodies from HeLa cells confirms binding of the NF-Y complex to NIPBL binding sites determined above (Fig. 3G).
To investigate whether other transcription factors colocalize specifically with NIPBL we compared the NIPBL sites in LCL's with available ChIP-sequencing data for transcription factors for GM12878 cells collected by ENCODE [32]. Specifically, we analyzed in total 66 binding profiles and generated heat maps covering +/−500 bp around NIPBL binding sites conserved in lymphoblastoid cells. By visual inspection of the maps we identified five transcription factors present on NIPBL sites: NFYA/NFYB and SP1, which is consistent with the presence of the motif, as well as PBX3, C-FOS and IRF3 (Fig. 3H). The heatmaps displaying the signals of the other transcription factors on NIPBL binding sites show a very good correlation between all five factors. When the signals are plotted respective to NFYB sites sorted according to peak intensity, it shows that NIPBL and several other factors overlap only with the strongest NFY peaks (Fig. 3I).
NIPBL-bound genes in HB2 cells were analyzed using IPA (Ingenuity Systems, www.ingenuity.com) and found to be linked to diverse cellular functions, such as cell cycle control, gene expression, cell death, RNA post-translational modification and control of cellular growth and proliferation (Table S5). Out of 1118 NIPBL-bound protein-coding genes, 122 (11%) were classified as transcription factors by Vaquerizas et al. 2009 [34], which is not a statistically significant enrichment compared to the number of transcription factors in lists with randomly selected genes, but indicates that important developmental genes might depend on NIPBL. Examples are SP1, SP2, SP3, BBX and STAT3, all important transcription factors for development and NIPBL binding at their promoters could be important for their appropriate expression.
To address whether NIPBL is important for the active transcription of the associated genes, we selected functionally different genes with conserved NIPBL binding at the promoter, but no cohesin binding site close to or on the gene, and asked whether their transcription changes in HB2 cells after knockdown of NIPBL, MAU2 or SMC3. To avoid problems in cell division due to impaired sister chromatid cohesion, we synchronized cells in G2 phase during the siRNA treatment (Suppl. Fig. S3B). Out of the seven initially selected genes, five showed statistically significant changes after NIPBL RNAi depletion: GLCCI1, a glucocorticoid inducible transcript; TSPAN31, encoding a transmembrane protein involved in signal transduction and growth-regulation; BBX, encoding a HMG-BOX transcription factor; ZNF695, an uncharacterized zinc-finger protein and ARTS-1/ERAP1, an endoplasmic reticulum aminopeptidase. Transcript levels were analyzed by RT-PCR and qPCR and normalized against the housekeeping gene NAD. Depletion of NIPBL and also of MAU2 leads to a statistically significant (t-test, P-values<0.05) decrease of gene expression levels of the candidate genes (Fig. 4), indicating that NIPBL and MAU2 dosage are important for maintaining expression levels. The depletion of SMC3 did not significantly reduce the expression of these transcripts, although the expression of the known cohesin-regulated MYC gene [35] was reduced. This indicates that the changes in expression as a result of NIPBL depletion are not the indirect result of reduced cohesin binding and cohesin's role for transcription.
Mutations in the NIPBL gene have been identified in approximately 50% of CdLS patients. Our discovery that NIPBL binds to active promoters prompted us to identify the major NIPBL binding sites in lymphoblastoid cells (LCL's) derived from blood samples of severely affected CdLS patients with NIPBL truncation mutations and normal controls (Fig. 5A, B).
Using the NIPBL#1 antibody we detected 1612 major NIPBL sites in the control (N5) and 2061/2009 sites in the patient-derived lines (PT1/PT9), with 1295 sites overlapping between N5/PT1 and 1273 sites between N5/PT9. In summary 80% of the sites in the control N5 are also found in PT1 and PT9 (Fig. 5C). The majority (74%) of the sites observed in HB2 cells is consistent with these conserved sites, indicating conservation between different tissues. Consistent with our observations in HB2 cells, most NIPBL binding sites in the LCL's localize to the 5′ ends of genes and are enriched for the motifs of the transcription factors NF-Y and/or SP1. Gene ontology analysis of the LCL NIPBL-bound genes showed similar classes of genes as for HB2 cell, but no cell type-specific functions such as immune response.
Although expected from patient-derived cell lines with NIPBL haploinsufficiency, we did not observe significant differences in peak number or peak intensity between controls and patient-derived LCL's. This is explained by the rather small differences of NIPBL protein levels between CdLS patients and controls [22] due to increased transcription from the intact allele. The ChIP-sequencing method is not quantitative and therefore small changes of NIPBL levels might not be reflected by peak intensity. To address this we performed NIPBL ChIP-qPCR from four control cell lines and four CdLS patient cell lines with primers for seven NIPBL binding sites and one cohesin binding site (negative control). QPCR revealed a reduction of the NIPBL signal between the control and patient-derived cell lines (Fig. 5D; Suppl. Fig. S6), but also variations among individual control- and patient-derived cell lines. In general, strong NIPBL binding sites (OSBP, GPR108) seem to be more reduced than weaker binding sites.
The position of NIPBL at promoters could be important for the emergence of the developmental defects seen in CdLS cases. Therefore we compared NIPBL binding sites with a list of genes found to be differentially expressed between LCL's from CdLS patients and controls [22]. We compared the list of 1501 unique genes (FDR<0.05) found to be differentially expressed between controls and CdLS patients [22] with our list of 1671 genes neighbouring a NIPBL site (+/−2 kb) in the patient-derived LCL's (PT1) and found that 155 (10%) of these genes are differentially expressed (Table S7), a statistically significant number when compared to a random list of genes (Fisher test, p<0.001).
In its best-studied function NIPBL promotes the initial deposition of the cohesin complex onto chromatin, but is dispensable for maintaining the subsequent association of cohesin and chromatin. Rules that regulate the place and time of cohesin loading and targeting to its various functions (sister chromatid cohesion, transcriptional regulation, mediating long-range chromatin interactions and DNA damage repair) are only partly understood. Factors interacting with chromatin-bound cohesin such as the chromatin insulator CTCF [3], [4], [36] and to a smaller extend estrogen receptor alpha (ERa) [37] determine the localization of cohesin, but not its general chromatin binding [3]. They might either direct NIPBL-dependent cohesin loading to their binding sites or capture cohesin complexes that slide along the DNA fibre.
First, we have addressed when cohesin, CTCF and NIPBL associate with chromatin. So far, only very weak and probably transient interactions have been reported between cohesin and NIPBL in the non-chromatin-bound pool of nuclear proteins [38]. If these transient interactions are sufficient for NIPBL to bind cohesin and recruit it onto chromatin, we would expect the proteins to appear on chromatin at the same time after mitosis. The same is true for CTCF. Analysis of cells exiting mitosis by immunofluorescence staining showed that NIPBL, CTCF and cohesin are largely excluded from metaphase chromosomes, as seen before [3]. The signals of NIPBL and CTCF reappear on DNA before the nuclear envelope reassembles; however, cohesin overlaps with chromatin only during or after the nuclear envelope reformation, reinforcing what was previously described by Gerlich et al. [39]. NIPBL and CTCF are therefore already present when cohesin starts to associate with chromatin. This is consistent with cohesin being dispensable for CTCF localization [3]. NIPBL very likely associates first with chromatin and then recruits' cohesin which is subsequently localized by CTCF to the co-occupied binding sites.
Second, we determined the genomic localization of NIPBL by ChIP using a NIPBL-specific antibody (NIPBL#1) form HB2 cells enriched in G1 phase. We observed about 1100 highly enriched NIPBL sites, mostly at active CpG-island promoters but also at several LSU and SSU rRNA repeat regions. However, we do not observe colocalization with cohesin or CTCF. Missing overlap between NIPBL and cohesin was observed before. In yeast, non-overlapping foci were observed for Scc2 (NIPBL ortholog in S. cerevisiae) and Scc1 (RAD21 ortholog in S. cerevisiae) by immunofluorescence microscopy on spread chromatin [40]. Further, a ChIP-microarray study in budding and fission yeast observed a transient overlap between cohesin and Scc2 in G1 phase cells and a subsequent relocalization of cohesin to more permanent positions between convergently transcribed genes [41]. Another study in yeast confirmed this property of cohesin [42] while a third study observed that colocalization of Scc2 with cohesin persists also after cohesin loading [43]. In D. melanogaster the NIPBL ortholog, Nipped-b, was found to colocalize with cohesin and often overlap with RNA polymerase II, decorating entire active transcriptional units [44]. Remarkably, cohesin does not colocalize with CTCF in the fruit fly.
However, a study in mouse embryonic stem cells (mESC) used a different NIPBL antibody (NIPBL#6) and reported that NIPBL occupies enhancers and core promoter regions of transcriptionally active genes which are also bound by cohesin and Mediator, a huge transcriptional co-activator complex [13] (for review see [45]).
Although we observe a similar localization of NIPBL, we did not detect cohesin binding at NIPBL sites, even with relaxed parameters for peak calling and using different ChIP protocols. We have considered that the apparent discrepancies in the binding patterns might arise due to the different ChIP protocols or differences between pluripotent and differentiated cells, but have disproved these hypotheses by ChIP-qPCR experiments using both antibodies (Suppl. Fig. S5). Importantly, we do observe significant differences between the performances of both antibodies. Immunoprecipitation experiments showed that the NIPBL#1 antibodies recognize more bands originating from NIPBL than NIPBL#6 antibodies (Suppl. Fig. S2). The NIPBL#1 antibodies we use show a similar weak enrichment in ChIP-qPCR experiments as observed for the NIPBL#6 antibodies in mESC (Suppl. Fig. S5C). However, the NIPBL sites identified by our study are highly enriched only by the NIPBL#1 antibodies, not by NIPBL#6. We therefore conclude that the NIPBL#1 antibodies very specifically recognize a number of “major” NIPBL binding sites at active promoters where NIPBL localizes independently from cohesin. The striking localization of NIPBL to promoter of active genes suggested that NIPBL may have a direct role for the transcription of the associated genes. We observe that the transcript levels of several NIPBL-bound genes decrease after RNAi depletion of NIPBL and MAU2. An effect on the transcripts by impaired cohesin loading cannot be entirely excluded but we observe that depletion of SMC3 does not have the same effect on the transcripts. Therefore we hypothesize that NIPBL could have a role as transcription factor, independent from its function for cohesin. A differential effect of NIPBL and cohesin has already been observed in the fly. Nipped-b facilitates activation of the cut gene, but stromalin/Scc3, the fly orthologs of the SA1/SA2 cohesin subunit, inhibits its activation. A recent study in zebrafish using morpholino knockdown observed a reduced transcription of several genes, including the transcription factors sox17, foxa2 and sox32, after NIPBL knockdown but not in smc3 and rad21 morphants [46].
We found that 11% of NIPBL-bound genes are transcription factors according to Vaquerizas et al. 2009 [34]. A number of them are very important during development and can also be found on the list of genes differentially expressed in CdLS, for example STAT3 and YBX1 (Table S7). Studies using mouse models show that the absence of some of these factors (STAT3, YBX1) leads to severe developmental defects and embryonic lethality [47]–[49]. NIPBL deficiency could therefore interfere with the proper timing and expression of transcription factors during development.
The observation that NIPBL might be important for gene expression lead us to ask whether NIPBL haploinsufficiency in CdLS can be linked to transcriptional changes observed in these patients. We have determined NIPBL sites in unsynchronized LCL's derived from CdLS patients with NIPBL haploinsufficiency and normal controls. These binding sites are again mostly located at CpG island promoters, not overlapping cohesin or CTCF. The sites are in part conserved between different tissues, indicating that there are constitutive and cell-type specific sites. The positions of the NIPBL binding sites are conserved between the LCL's from patients and controls, but the actual levels of NIPBL binding are reduced in patients with a hypomorphic NIPBL truncation. To link NIPBL sites to differential gene expression we compared NIPBL-bound genes identified in a patient cell line (PT1) with candidate CdLS target genes identified by Liu et al. [18] and observed that a significant percentage (11%, Fischer test p<0.001) of these genes have a NIPBL binding site. When we asked whether NIPBL RNAi affects gene expression (Fig. 4) a subset of these genes was tested and found to be sensitive for NIPBL knockdown. This lead us to the conclusion that a part of the differentially expressed genes in CdLS could be direct targets of NIPBL, and the observed CdLS phenotype could be a cumulative effect of small changes in the transcriptional program of a larger number of genes.
Comparison of NIPBL sites in LCL's with published binding profiles of transcription factors in the lymphoblastoid cell line GM12878 revealed that NIPBL colocalizes with several transcription factors (SP1, NFY, PBX3, c-FOS, IRF3). Pbx3 belongs to the Pbx family of TALE (three amino acid loop extension) class of homeodomain transcription factors, which are implicated in developmental and transcriptional gene regulation in numerous cell types. Pbx3-deficient mice die after birth due to neuronal malfunctions [50]. The factor is important for facial development in mice [51] together with Pbx1 and Pbx2, and a human Pbx3 mutation was linked to heart defects [52]. IRF3 (interferon regulatory factor 3) is an IRF family transcription factor which translocates from the cytoplasm to the nucleus upon activation, where it acts together with CBP/p300 to activate transcription of interferons alpha and beta, as well as other interferon-induced genes (for review see [53]). C-FOS is part of the AP-1 (activator protein 1) transcription factor complex, which also contains the JUN, ATF and MAF proteins. The complex regulates genes involved in cell proliferation, differentiation, apoptosis, angiogenesis and tumour invasion and can have oncogenic but also anti-oncogenic properties depending on cell type or differentiation state [54]. How these factors functionally interact with NIPBL remains to be investigated.
In summary, in this study we have addressed when and where NIPBL binds to the human genome. We have discovered that a subset of very strong “major” NIPBL binding sites preferentially localizes to active promoters, together with a specific set of other transcription factors. NIPBL is important for the activity of the bound genes, suggesting that NIPBL influences transcription in two ways; directly due to its binding to the promoters and indirectly by loading of cohesin complexes which then regulate genes by chromatin insulation and chromosomal long-range interactions. The possibility that NIPBL directly affects expression suggests that NIPBL-deficiency also directly contributes to the complex CdLS phenotype by altering the transcriptional program of developmentally important genes.
If different antibodies for the same protein were used the antibodies were numbered to clearly identify them in the different experiments.
NIPBL#1 - polyclonal rabbit anti-NIPBL antibody raised against residues 2598–2825 of the X. laevis Scc2-1B, purified using the epitope used for immunization (133M).
NIPBL#2 – polyclonal rabbit anti-NIPBL antibody raised against residues 787–1164 of X. laevis Scc-1B, purified using the epitope used for immunization (114M). Generation and characterisation of the NIPBL #1 and NIPBL #2 antibodies have been published already [38].
NIPPBL#3 – monoclonal rat anti-NIPBL, isoform A (long isoform) NP_597677 (Absea, China, 010702F01 clone KT54)
NIPPBL#4 – monoclonal rat anti-NIPBL, isoform B (short isoform) NP_056199 (Absea, China, 010516H10 clone KT55)
NIPPBL#5 - polyclonal rabbit anti-NIPBL antibody raised against a region between amino acid residues 550 and 600 of human NIPBL (Bethyl Laboratories A301-778A)
NIPPBL#6 - polyclonal rabbit anti-NIPBL antibody raised against a region between amino acid residues 1025 and 1075 of human NIPBL (Bethyl Laboratories A301-779A)
CTCF#1 –monoclonal mouse anti-CTCF (BD 612149)
CTCF#2 – polyclonal rabbit anti-CTCF antiserum (Millipore 07-729)
SA2 – monoclonal rat anti-SA2(STAG2) antibody (Frank Sleutels and Niels Galjart)
SMC1A#1 - polyclonal rabbit anti-SMC3 antibodies (Bethyl Laboratories A300-055A)
SMC3 – polyclonal rabbit anti-SMC3 antibodies obtained from Jan-Michael Peters, described for immunoprecipitation and ChIP in [55] and [3].
MAU2 – polyclonal rabbit anti-MAU2(Scc4), described in [38].
RNA Pol II – polyclonal rabbit antibody (N-20) (Santa Cruz sc-899)
Tubulin – mouse anti-tubulin (Sigma)
Control IgG – rabbit whole serum
Rad21 – polyclonal rabbit anti-RAD21 (Jan-Michael Peters), described in [29]
HeLa cells were cultured in DMEM supplemented with 0.2 mM L-glutamine, 100 units/ml penicillin, 100 mg/ml streptomycin and 10% FCS.
HB2 cells (1-7HB2, a clonal derivative of the human mammary luminal epithelial cell line MTSV1-7, [25]) were cultured in DMEM supplemented with 0.2 mM L-glutamine, 100 units/ml penicillin, 100 mg/ml streptomycin, 10% FCS, 5 µg/ml hydroxycortisone and 10 µg/ml human insulin.
Lymphoblastoid cell lines derived from controls and Cornelia de Lange syndrome patients (Fig. 5B) were obtained from Ian Krantz (The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, United States of America) and cultured in RPMI medium supplemented with 0.2 mM L-glutamine, 100 units per ml penicillin, 100 mg per ml streptomycin, 20% FCS.
SMC-LAP and Lamin-LAP Hela cells were were cultured in DMEM supplemented with 0.2 mM L-glutamine, 100 units/ml penicillin, 100 mg/ml streptomycin and 10% FCS and 0.2 mg/ml G418.
The following siRNA oligos purchased form AMBION were used to deplete the respective proteins for ChIP-analysis and analysis of transcription
GL2
sense CGUACGCGGAAUACUUCGAtt
antisense UCGAAGUAUUCCGCGUACGtt
NIPBL
sense GCAUCGGUAUCAAGUCCCAtt
antisense UGGGACUUGAUACCGAUGCtt
MAU2
sense GCAUCGGUAUCAAGUCCCAtt
antisense UGGGACUUGAUACCGAUGCtt
SMC3
sense AUCGAUAAAGAGGAAGUUUtt
antisense AAACUUCCUCUUUAUCGAUtg
The following hairpin siRNA constructs in the pLKO.1–puro vector were obtained from the TRC Mission human library (Sigma) and were used to deplete NIPBL demonstrate the specificity of the NIPBL antibodies:
Control (clone SHC002) non-targeting sequence
NIPBL (clone TRCN0000129033) targeting sequence GCAGAGACAGAAGATGATGAA
The transfection of the siRNA oligos was performed with Lipofectamine RNAiMAX (Invitrogen) according to the manufacturer's instructions. The transfection of the hairpin siRNA constructs was performed with Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions. Cells were harvested 48 hours after transfection.
HeLa cells were grown on 18-mm cover slips and fixed with 4% PFA. After permeabelization with TX100 and blocking with 3% BSA the slides were stained with the respective antibodies.
Images were taken on a Leica DMRBE microscope equipped with a Hamatsu CCD (C4880) camera with a 100× objective. Images were processed with Image J, the colouring; overlay of the images was done with Adobe Photoshop.
Cells were fixed with methanol and after RNAse treatment the DNA was stained with propidium iodine. The cells were analyzed with a BD FACS Aria Cell sorter and FlowJo software.
See Table S6.
To prepare nuclear extracts from HeLa cells the cells were first lysed by gentle resuspension in hypotonic buffer (20 mM Hepes-KOH pH 8.0, 5 mM KCl, 1.5 mM MgCl2, 0.1 mM DTT). Nuclei were collected by centrifugation and extracted for 30 min on ice with extraction buffer (15 mM Tris-HCl pH 7.5, 1 mM EDTA, 0.4 M NaCl, 10% sucrose, 0.01%TX-100, 1 mM DTT and 1 tablet Complete (Roche) per 50 ml buffer). Debris were removed by centrifugation (14000 rpm, 30 min).
The nuclear extract was diluted to 50% with IP buffer (20 mM Tris-HCl pH 7.5, 100 mM NaCl, 5 mM MgCl2, 0.2% NP40, 1 mM NaF, 0.5 mM DTT) and incubated for 1 h at 4°C with the respective antibodies. Affi-Prep Protein A support beads (BioRad) were added and incubated 1 h at 4°C. The beads were washed 3 times with IP buffer and eluted by boiling with SDS-page loading buffer. Western blots were analyzed with ECL+ reagent and Alliance imaging system.
Chromatin immunoprecipitation was performed as described before [3]. In brief, cells at 70–80% confluence were crosslinked with 1% formaldehyde for 10 min and quenched with 125 mM glycine. After washing with PBS cells were resuspended in lysis buffer (50 mM Tris-HCl pH 8.0, 1% SDS, 10 mM EDTA, 1 mM PMSF and Complete protease inhibitor (Roche)) and chromatin was sonicated (Diagenode Bioruptor) to around 500 bp DNA fragments. Debris were removed by centrifugation, the lysate diluted 1∶4 with IP dilution buffer (20 mM Tris-HCl pH 8.0, 0.15 M NaCl, 2 mM EDTA, 1% TX-100, protease inhibitors) and precleared with Affi-Prep Protein A support beads (BioRad).
The respective antibodies were incubated with the lysate for 14 h at 4°C, followed by 2 h incubation at 4°C with blocked protein A Affiprep beads (Bio-Rad) (blocking solution: 0.1 mg/ml BSA or 0.1 mg/ml fish skin gelatine). The beads were washed with washing buffer I (20 mM Tris-HCl pH 8.0, 0.15 M NaCl, 2 mM EDTA, 1% TX-100, 0.1% SDS, 1 mM PMSF), washing buffer II (20 mM Tris-HCl pH 8.0, 0.5 M NaCl, 2 mM EDTA, 1% TX-100, 0.1% SDS, 1 mM PMSF), washing buffer III (10 mM Tris-HCl pH 8.0, 0.25 M LiCl, 1 mM EDTA, 0.5% NP-40, 0.5% sodium desoxycholate) and TE-buffer (10 mM Tris-HCl pH 8.0, 1 mM EDTA). The beads were eluted twice (25 mM Tris-HCl pH 7.5, 5 mM EDTA, 0.5% SDS) for 20 min at 65°C. The eluates were treated with proteinase K and RNase for 1 h at 37°C and decrosslinked 65°C over night. The samples were further purified by phenol-chloroform extraction and ethanol-precipitated. The pellet was dissolved in 50 µl TE buffer.
This protocol was used to perform ChIP-qPCR or ChIP-sequencing for CTCF, SMC3, NIPBL and RNA polymerase II. For SMC1A a milder ChIP protocol from Duncan Odom's group was used [26].
For NIPBL ChIP sequencing HB2 cells were synchronized in G1 phase by double thymidine block as described [8] (Suppl. Fig. S3). All other preparations were done from unsynchronized cells.
For NIPBL ChIP after depletion of NIPBL or control by RNAi the cells were synchronized in G1 phase by double thymidine block, starting 6 hours after transfection of the siRNA oligos. Details of the thymidine block to obtain HeLa cells in G1 phase are described [3].
Samples were either submitted for genomic sequencing or analyzed by qPCR using Platinium taq (Invitrogen) according to the manufacturer's instructions as described [3]. ChIP-qPCR experiments at least three times and one representative example is shown (SD was determined from qPCR replicates).
The ChIP DNA library was prepared according to the Illumina protocol (www.illumina.com). Briefly, 10 ng of ChIPped DNA was end-repaired, ligated to adapters, size selected on gel (200±25 bp range) and PCR amplified using Phusion polymerase as follow: 30 sec at 98°C, 18 cycles of (10 sec at 98°C, 30 sec at 65°C, 30 sec at 72°C), 5 min at 72°C final extension. Cluster generation was performed using the Illumina Cluster Reagents preparation. The libraries for NIPBL, CTCF, SMC3, RNA PolII and the respective controls generated from HB2 cells were sequenced on the Illumina Genome Analyzer II, the SMC1A ChIP samples from HB2 cells, the NIPBL ChIP samples from LCLs and the respective controls were sequenced with the Illumina HiSeq2000 system. Read lengths of 36 bases were obtained. Images were recorded and analyzed by the Illumina Genome Analyzer Pipeline (GAP 1.6.0. and 1.7.0.). The resulting sequences were mapped against Human_UCSChg18 using the Bowtie [56] alignment software, with the following parameters: bowtie -m 1 -S -k 1 –n 1. Unique reads were selected for further analysis.
Peak calling for the ChIP sequencing data was performed with SWEMBL (URL: http://www.ebi.ac.uk/~swilder/SWEMBL/) as described [37] with the respective parameters described in Table S1.
Co-localization read density profiles were done by extending a region around each peak summit by +/−200 bp. Regions from each data set were chosen in succession as viewpoints, and the position of 5′ends of the reads in corresponding regions in all data sets was plotted. The profiles were ordered by the significance score determined by the Swembl peak caller.
To investigate the repeat enrichment pattern, we used both uniquely- and multiply-aligned reads. Multiply-aligned reads were divided equally amongst all locations (N-times matched reads were weighted as 1/N reads). The reads which were aligned to reference genome more than 10 times were discarded. We applied RPKM measure (reads per kilobase per million reads) which was utilized for RNA-seq analyses [56], but we used “per 10 million reads” instead of “per million reads”. We counted the reads which were aligned to each repeat class and normalized the counts against the total number of aligned reads (whole-genome) and the total length of each repeat class.
HB2 cells were enriched in G1 phase by double thymidine block as described [8]. The RNA was isolated using TRI reagent (Sigma) according to the supplier's protocol. Two microgram of total RNA was converted into a library of template molecules suitable for sequencing according to the Illumina mRNA Sequencing sample prep protocol. Briefly, polyA containing mRNA molecules were purified using poly-T oligo attached magnetic beads. Following purification, the mRNA is fragmented into ∼200 bp fragments using divalent cations under elevated temperature. The cleaved RNA fragments are copied into first strand cDNA using reverse transcriptase and random primers. This is followed by second strand synthesis using DNA polymerase I and RNaseH treatment. These cDNA fragments are end repaired, a single A base is added and Illumina adaptors are ligated. The products are purified and size selected on gel and enriched by PCR. The PCR products are purified by Qiaquick PCR purification and used for cluster generation according to the Illumina cluster generation protocols (www.illumina.com). The sample was sequenced for 36 bp and raw data was processed using Narwhal [57].
RNA Seq reads were mapped to the Human UCSChg18 genome with Bowtie using the same parameters as for the ChIP seq analysis. The coverage vector was calculated from unique reads and the expression value was determined for each gene by taking the RPKM [58] of the most highly expressed isoform (the sum of coverage over exons was used as the numerator of the equation). All genes with RPKM>0.6 were designated as expressed.
Motif analysis was performed with the tools MEME and MEME-ChIP [33]. Residues +/−50 bp of NIPBL binding site peaks were retrieved and submitted to MEME-ChIP using standard parameters.
To analyse whether the presence of the NFYA motif at NIPBL sites is due to the presence of CpG islands or is a genuine property of NIPBL binding we selected NIPBL binding sites close to only one CpG island promoter (692 sites) and selected the same number of CpG island promoters at random. The presence of the NFYA motif was detected and the counts statistically analyzed using a Fischer-test.
We obtained from ENCODE [32] ChIP-sequencing data tracks for transcription factors generated from GM12878 cells and deposited by the Myers lab (HudsonAlpha Institute for Biotechnology) and the Snyder lab (Yale University). When called peaks were available they were used, else replicates were pooled and peak calling performed with MACS [58]. Peaks were sorted for intensity and for the 10000, 5000 and 1000 (in case of NIPBL) strongest peaks heatmaps were generated centred on NIPBL binding sites conserved in the different lymphoblastoid cell lines and also centred on the peaks of the respective transcription factors. Overlapping patterns were selected by visual inspection of the maps.
HB2 cells were transfected with the respective siRNA oligos using Lipofectamine 2000 and were harvested after 48 hours. The RNA was prepared using TRI reagent (Sigma). Remaining DNA was removed by DNAse treatment and cDNA synthesis was performed with Superscript reverse Transcriptase (Invitrogen) using oligo-dT primers. The qPCR analysis was performed as described [3].
This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the Institutional Review Board of The Children's Hospital of Philadelphia. All patients provided written informed consent for the collection of samples and subsequent analysis.
|
10.1371/journal.pcbi.1002329 | A Dynamical Model of Oocyte Maturation Unveils Precisely Orchestrated Meiotic Decisions | Maturation of vertebrate oocytes into haploid gametes relies on two consecutive meioses without intervening DNA replication. The temporal sequence of cellular transitions driving eggs from G2 arrest to meiosis I (MI) and then to meiosis II (MII) is controlled by the interplay between cyclin-dependent and mitogen-activated protein kinases. In this paper, we propose a dynamical model of the molecular network that orchestrates maturation of Xenopus laevis oocytes. Our model reproduces the core features of maturation progression, including the characteristic non-monotonous time course of cyclin-Cdks, and unveils the network design principles underlying a precise sequence of meiotic decisions, as captured by bifurcation and sensitivity analyses. Firstly, a coherent and sharp meiotic resumption is triggered by the concerted action of positive feedback loops post-translationally activating cyclin-Cdks. Secondly, meiotic transition is driven by the dynamic antagonism between positive and negative feedback loops controlling cyclin turnover. Our findings reveal a highly modular network in which the coordination of distinct regulatory schemes ensures both reliable and flexible cell-cycle decisions.
| In the life cycle of sexual organisms, a specialized cell division -meiosis- reduces the number of chromosomes in gametes or spores while fertilization or mating restores the original number. The essential feature that distinguishes meiosis from mitosis (the usual division) is the succession of two rounds of division following a single DNA replication, as well as the arrest at the second division in the case of oocyte maturation. The fact that meiosis and mitosis are similar but different raises several interesting questions: What is the meiosis-specific dynamics of cell-cycle regulators? Are there mechanisms which guarantee the occurence of two and only two rounds of division despite the presence of intrinsic and extrinsic noises ? The study of a model of the molecular network that underlies the meiotic maturation process in Xenopus oocytes provides unexpected answers to these questions. On the one hand, the modular organization of this network ensures separate controls of the first and second divisions. On the other hand, regulatory synergies ensure that these two stages are precisely and reliably sequenced during meiosis. We conclude that cells have evolved a sophisticated regulatory network to achieve a robust, albeit flexible, meiotic dynamics.
| The mitotic division cycle is the sequence of events by which a growing cell replicates all its components, including DNA, and divides them, after mitosis, into two nearly identical daughter cells [1]. Meiosis is an alternative mode of cell division in which a diploid cell undergoes two successive divisions without intervening DNA synthesis, to create haploid cells called gametes or spores [2]. In vertebrate species, for instance, meiosis occurs during oocyte maturation, which is initiated in response to an hormonal signal with the specificity that oocytes are thereafter arrested, usually at the metaphase stage of MII, awaiting fertilization [3]. Meiotic maturation shares with mitosis many morphological events, such as metaphase and anaphase, as well as regulators such as the cyclin B-Cdk1, known as the M-phase promoting factor (MPF). However, it also involves a unique sequence of decision steps - meiotic resumption, transition and arrest - which clearly diverges from the mitotic one (Fig. 1A). Investigating the regulation of meiotic maturation is therefore an opportune strategy to understand the remarkable plasticity of the cell cycle, which unfolds a diversity of decision patterns at different stages of multicellular development.
The specific decision pattern of the oocyte meiotic maturation is intimately linked to the tightly controlled temporal dynamics of MPF (Fig. 1B). The rise and the first peak of MPF activity triggers germinal vesicle break down (GVBD) and entry into MI. The transition from MI to MII is typified by an unusual partial decrease of MPF activity followed by an increase and stabilization at a plateau level associated with metaphase II arrest in Xenopus oocytes. The time course of MPF is shaped by a complex web of interaction with other cell-cycle regulators. At the first arrest of Xenopus oocyte in a G2-like state, MPF kinase is stored in an inactive state called pre-MPF in which, among the five isoforms of cyclin B described in this animal model, only cyclin B2 and B5 are found associated to Cdk1 [4]. As during mitosis, MPF activity is primarily regulated by its interaction with a dual protein-phosphatase (Cdc25), a cyclin-dependent kinase inhibitor (Myt1) and the anaphase promoting complex (APC). During meiotic maturation, this module is supplemented with a layer of control which involves the MAPK (Mitogen Activated Protein Kinase)/ERK(Extracellular Regulated Kinase) pathway, whose main upstream and output components in the context of meiotic maturation are proteins Mos and Rsk, respectively. These components of the MAPK pathway are involved not only in meiotic spindle morphogenesis during oocyte maturation [5] but also at several decision points of the oocyte maturation process including meiotic resumption (G2/MI), meiotic transition (MI/MII) and maintenance of metaphase II arrest [6]–[8]. A key advance was to identify Rsk-mediated phosphorylation of APC inhibitor Emi2 as leading to MPF reaccumulation at the MI/MII transition [9]–[11]. In turn, MPF tightly controls phosphorylated levels of Mos [12] or Emi2 [13].
Two decades of experimental studies have thus documented manifold levels of interaction between the MPF and Mos/MAPK pathways, whose respective roles in various decision stages of maturation remain difficult to disentangle. In an attempt to clarify the interactions between both pathways, we use a modeling approach which has already been harnessed to gain insight into cell cycle control during animal development, as with the syncytial mitotic cycles in Drosophila embryos [14], fertilization process in mammals [15], the oocyte maturation initiation switch [16] but not yet for the whole oocyte meiotic maturation process. This approach has been remarkably successful, not only to check whether known molecular interactions can explain observed contextual and functional cell-cycle behaviors, but also to uncover the design principles of the molecular network in terms of feedback and feedforward topology [17]–[19]. Our modeling effort will thus be devoted to address two complementary issues: Are documented interactions between MPF and MAPK pathways necessary and sufficient to account for the observed properties of meiotic maturation? What are the network design principles that robustly enforce the progression of cells through a specific sequence of meiotic decisions?
To answer these questions, we first build a computational model that incorporates the major signaling pathways involved in the meiotic maturation of Xenopus laevis oocytes. Appropriate parameterization of the model allows us to reproduce the temporal dynamics of MPF and of other key regulators when the oocyte progresses from meiotic entry to metaphase II arrest. The dynamical mechanisms underlying these transitions are further analyzed using bifurcation analysis, which unmasks the existence of two main positive-feedback systems in addition to the core negative-feedback loop along which MPF represses itself by upregulating its own inhibitor, the anaphase promoting complex (APC). Remarkably, the architecture of these two subcircuits is unambiguously identified using a parameter sensitivity analysis, which reveals that they independently regulate meiotic resumption on the one hand, and meiotic transition on the other hand. The significance of the model is further assessed by simulating how alteration of the underlying molecular network using chemical manipulations or antisense strategies may induce maturation defects including initiation delays [8] or failures to transit from first to second meiosis [4], [6], [7], [10], [20]. Revisiting the relation between topology and dynamics in the maturation regulatory network leads us to identify and discuss the design principles that underlie the complex and reliable decision sequence studied here, and which could apply in various other cellular contexts.
The interaction graph shown in Fig. 1C incorporates all molecular actors and interactions known to be involved during the cell-cycle progression from prophase I to metaphase II arrest. Following the basic principles of biochemical kinetics, we translate this graph into a set of ordinary differential equations (see Methods). The unknown rate constants of the model are estimated by fitting the qualitative model behavior to the available data, including the well-characterized temporal profile of MPF activity during the meiotic maturation (Fig. 1B) as well as the bistable behavior of the MAPK modules and the oscillatory dynamics of the MPF-APC module (Methods). It was found that the behaviors observed depend little on the specific parameter set chosen under these constraints. This global maturation regulatory network of Fig. 1C connects together functional modules that so far have been studied only separately: (i) the MPF autoamplification loop, which triggers G2/M transition; (ii) the MAPK phosphorylation cascade, which is characterized by specific upstream and downstream components during meiotic maturation and is tightly bound to the MPF autoamplification loop; (iii) the underlying CPEB-dependent translational network, which controls temporal expression during the maturation process. We describe below how these subcircuits function and how they interact.
A remarkable feature of oocyte meiotic maturation is that a basic hormonal signal (a pulse-like or constant exposure of progesterone) induces a complex non-monotonous MPF activity profile (see Fig. 1B). The term non-monotonous refers to the fact that MPF activity does not continously increase or decrease during maturation but falls rapidly after rising to a first peak, before increasing again toward a plateau. In this section, we investigate the network dynamical properties underlying this sophisticated temporal profile of MPF activity, which drives the sequence of meiotic decisions from resumption to transition and arrest.
Fig. 2A shows how the mathematical model responds to a constant exposure of progesterone. The numerical simulation reproduces the typical MPF temporal profile observed during meiotic maturation of Xenopus oocytes (compare Fig. 2A and Fig. 1B). In contrast with the non-monotonous time course of MPF, components of the MAPK pathway (Mos, MEK, ERK, Rsk) or of the autoamplification loop (Plx1, Cdc25, Myt1) exhibit a sharp activation (or inactivation for Myt1) followed by a plateau. Besides CPEB1, which is inactivated and degraded at the meiotic transition, APC is the only actor which exhibits a transient activation, following the MPF peak associated with anaphase events. Another key feature of the simulation is Emi2 activity rising only at the very end of meiosis I before reaching a plateau. As we shall see later, proper timing of Emi2 activation is crucial to allow MPF to reaccumulate after a full activation of APC. The activity profiles obtained in this simulation are fully consistent with experimental data collected for Mos, MAPK, Plx1, APC/Cdc27 or Emi2 during oocyte maturation of Xenopus laevis [4], [6], [7], [13], [34].
The one-parameter bifurcation diagram in Fig. 2B shows how the steady state value of MPF activity varies as a function of progesterone level. At least two stable solution branches coexist for some range of progesterone concentration, including the case of no progesterone. The coexistence of two (or more) stable solutions for the same parameter value is a phenomenon known as bistability (or multistability). Coexisting stable solution branches (nodes) are generally connected by an unstable branch solution (saddle), which acts as a separatrix between them. Stable and unstable branches connect at saddle-node points, where they annihilate together (the stable and unstable solution can be found on one side only of the saddle-node point). The lower branch (low MPF activity) corresponds to the prophase I-arrest state whereas the upper branch (high MPF activity) corresponds to the Metaphase II-arrest state. A sufficiently strong progesterone input, even transient, is therefore able to switch the cellular state from G2-arrest to metaphase-II arrest in an irreversible manner. An important feature of this bifurcation diagram is also the existence of four saddle-node bifurcation points (I, II, III, IV) instead of the two saddle-node points associated with the classic bistability scheme. The bifurcation point I controls the progesterone level required to destabilize the G2-arrested state and to trigger sharp MPF activation. The bifurcation point IV determines the stability of the metaphase II arrest characterized by high MPF activity. As long as this bifurcation point is associated with a negative value of progesterone signal, oocytes cannot leave the metaphase II state. If it is shifted to positive values of the progesterone signal by parameter changes, however, high MPF levels cannot be maintained, and the arrest state is unstable. The presence of the two additional saddle-node points II and III allows the slopes of the unstable solution branches originating from points IV and I to be largely independent, a feature which may persist even if the saddle-node points collide upon variation of another a parameter. This configuration effectively decouples the bifurcation points I and IV and allows them to be controlled separately, which will prove crucial in the following. The global structure of the bifurcation diagram of Figs. 2B, with its double bistability cycle, reflects in fact the coordinated actions of two bistable positive-feedback systems [18], [35], [36], which are relatively independent although one tends to activate the other and which will be identified in the next section. The non-monotonous behavior of MPF activity during the transition from the low activity state to the high activity state is not directly related to the structure of the bifurcation diagram. The fact that MPF activity rises, then decreases before increasing again is due to a negative feedback control based on the interaction between MPF and APC.
This feedback-based bifurcation structure underlying the maturation dynamics is expected to provide robustness against environmental or intrinsic noises that could bias the trajectory toward inappropriate cellular states (e.g. G2-like, interphase-like, oscillations). Fig. 3 shows that such major disruption is very improbable. Indeed, the dynamical trajectory of MPF in state space starting from a G2-arrest state to a metaphase-II arrest state (bottom panel of Fig. 3A) is remarkably insensitive to changes in the progesterone input profile (top panel of Fig. 3A), with fluctuations mostly affecting the timing of maturation (middle panel of Fig. 3A). In particular, whether progesterone input is constant or transient has almost no effect on the dynamical trajectory of MPF, provided the input is sufficiently strong, which is consistent with experiments in which oocytes are either treated with transient or continuous exposure of progesterone [37]–[39]. It also confirms the result anticipated by the bifurcation diagram of Fig. 2B that oocyte meiotic maturation is indeed a bistable process in which the transition can be triggered by a transient perturbation. Similarly, the model also displays a robust behavior with respect to variability in kinetic rates since the qualitative structure of the trajectory remains unaffected when all kinetic parameter values are randomly changed with a coefficient of variation (CV) of (Fig. 3B). For a CV of , only a few cases display abnormal MPF profiles. However, when the CV of parameter changes is increased beyond the large value of , maturation failures occur in more than half of the trials.
The type of robustness oberved here emphasizes that achieving the appropriate sequence of decisions depends on the sequence of biochemical states traversed, not on the exact times at which they are reached. Thus, the state space trajectory is more relevant than time profiles. This robust dynamical behavior stems from the existence of an attracting slow manifold that canalizes the trajectory in state space.
In the previous section, the bifurcation analysis revealed the existence of two positive-feedback control systems operating independently to sequentially drive the G2/M and meiotic transitions, in addition to the core MPF-APC negative feedback loop. To identify these two systems, numerical simulations and bifurcation analysis are here supplemented with a systematic parameter sensitivity analysis, which allows us to characterize the effect of each parameter on the maturation process.
We first focus on quantitative indicators of the MPF activity profile, which are the time of occurence of the first MPF peak (signaling the G2/MI transition), as well as MPF levels and associated with the trough and the plateau of the MPF time course (signaling the MI/MII transition and the further metaphase II arrest). We measure their sensitivities to parameter variation as:(1)where is the perturbed parameter and the weigh the different indicators. As discussed above, G2 arrest and MII arrest are directly controlled by the saddle-node bifurcation points I and IV, respectively. Denoting by and the progesterone thresholds associated with bifurcation points I and IV, respectively, the sensitivity of and to parameter variation should also be a relevant parameter sensitivity measure of maturation dynamics. These sensitivities can be written as:(2)For both types of sensitivities, it is useful to define normalized sensitivities, defined by and with and . For example, a value of close to (resp., ) indicates that parameter affects the progesterone threshold (resp., ) much more than (resp., ).
These two complementary sensitivity measures are expected to indicate whether a given parameter tends to affect early or late stages of maturation as illustrated in the right panels of Figs. 4A and 4B. Fig. 4C provides a synthetic view of sensitivity values for all kinetic parameters that control interactions between two molecular actors and are therefore associated to a link in the network diagram of Fig. 1C. Interestingly, the values of sensitivities and (and therefore and ) are seen to be highly correlated. The two sensitivities thus essentially provide the same information, confirming that dynamic response to progesterone signals is very much controlled by the bifurcation diagram. Noting that most values of the normalized sensitivities are either close to 0 or 1, a natural partition of kinetic parameters into two classes emerges, according to whether they preferentially control transition G2/MI () or the MI/MII transition (). The classification so obtained allows us to disentangle the complex regulatory network shown in Fig. 1C by isolating two separate subnetwork module, such that all links in a module control the same transition. It is quite remarkable that most molecular actors appear in one module or the other but not in both, with the notable exception of Mos and MPF. Note that links corresponding to kinetic parameters with very small sensitivities have been neglected. Presumably, these molecular interactions have biological roles not directly related to maturation control or have a specific impact that could not be identified given the chosen model parameters.
The first circuit drawn in Fig. 4D displays a coherent feedback organization where only positive feedback loops are present. The post-translational interactions between Plx1, Cdc25, MPF and Myt1 constitute the core set of positive-feedback loops that contributes to the MPF autoamplification loop. Additional feedforward loops mediated by the activation of the translation machinery (i.e., CPEB) and positive-feedback loops mediated by the phosphorylation of Mos by MPF and of Myt1 by Mos are also involved in G2/MI transition. The architecture of the circuit controlling MI/MII transition (Fig. 4E) markedly differs from that of the meiotic resumption module. First, MPF is now regulated through CPEB-dependent synthesis and APC-dependent degradation, which control only its turnover. Second, this circuit relies on an antagonism between two negative feedback loops, where the direct interactions of MPF with APC and Emi2 promote its own inactivation through the degradation of its cyclin subunits, and a positive-feedback loop where MPF-dependent activations of MAPK and CPEB4 cooperate towards the accumulation and activation of Emi2, which itself opposes the APC-dependent degradation of MPF. This feedback antagonism results in an incoherent feedforward loop, which is key to the precise temporal gap between the G2/MI and MI/MII transitions.
Importantly, our model accounts not only for the main features of meiotic maturation in wild-type eggs, but also of phenotypes of eggs treated by antisense oligonucleotides-based strategies or by chemical inhibitors. Our simulations of these phenotypes are summarized in Fig. 5 and Table 1.
To identify the role of protein synthesis in the initiation of Xenopus oocytes maturation, experiments have been performed to inhibit cyclin B or/and Mos synthesis using antisense oligonucleotides [8]. They showed that ablation of either Mos or cyclin B alone does not prevent maturation initiation yet induces significant delays, whereas combined ablation impairs initiation. Our model is able to reproduce such delays in the absence of cyclin or Mos synthesis (Fig. 5A and B), reflecting the existence of cooperative mechanisms between translational and post-translational controls during meiotic initiation (e.g., Mos synthesis and Mos-dependent inactivation of Myt1). In addition, oocytes where cyclin B is disabled by antisense strategies fail to reaccumulate MPF at MI/MII transition [4]. This is also observed in simulations (Fig. 5A) where, after the post-translational activation of preMPF by the Plx1 pathway, depletion of preMPF and degradation of active MPF by APC are not counterbalanced by the synthesis of new cyclins, thereby precluding MPF reaccumulation. Meiotic transition also fails in Mos-ablated oocytes, due to the absence of MAPK activation [7], with the possibility however to form a transitory interphase nucleus after completion of meiosis I and to reactivate MPF so as to mimic the mitotic cell cycle of early embryos [7]. In numerical simulations, oocytes lacking Mos are indeed unable to transit appropriately to MII. However, an oscillatory pattern of MPF activity may be also observed although it is highly sensitive to model parameters (Fig. 5B). Besides the defects for maturation initiation associated with inhibition of protein synthesis, disruption of the progesterone-dependent Plx1 activation also significantly delays meiotic resumption in progesterone-treated oocyte [22], [40], which is reported as well in numerical simulations (Fig. 5C). Note that any combination of the disruption of cyclin synthesis, Mos synthesis and Plx1 activation leads in model simulation to maturation initiation failures (result not shown), emphasizing the synergistic role of multiple translational and post-translational mechanisms.
Inhibition of MAPK activation in oocytes can also be achieved using MEK inhibitor U0126 [5], [6], [41]. In U0126-treated oocytes, MAPK inactivation prevented cyclin B reaccumulation after MI, by allowing APC-mediated degradation similarly as in the case of Emi2 ablation [10]. In simulations (Fig. 5D), MPF concentration does not vanish as in the case of inhibition of cyclin synthesis but remains at an intermediate level as is observed in experiments [5], [6]. Simulations do not reproduce the delay observed in these experiments, which can be due to our model not taking into account the regulation of cyclin B synthesis by MAPK as has been reported by Abrieu et al [42]. In addition, chemical inhibitor U0126 might target other translational regulators besides MEK1, and such non-specific effects may account for discrepancies in the observations made when Mos is ablated.
Experiments inducing deletion or overexpression of Emi2 demonstrate the crucial role of this protein in meiotic transition. Ectopic expression of Emi2 at physiological MII levels can arrest maturing oocytes at metaphase I [11], which is easily explained by the fact that Emi2 counteracts APC activity and subsequently cyclin degradation, maintaining a sustained MPF activity (Fig. 5E). Conversely, our simulations also reproduce the effect of inhibiting Emi2 synthesis (Fig. 5F), which leads to complete and rapid degradation of cyclin B at MI exit, causing an inappropriate exit into interphase and a failure to reaccumulate cyclin B [10], [11], [13]. These experiments and simulations showing maturation failure for overexpression or deletion of Emi2 strongly support that a strict temporal control over Emi2 levels is critical for a reliable MI/MII transition.
In this work we have designed and analyzed a detailed mathematical model describing the meiotic maturation process that begins when an oocyte is released from G2-arrest and terminates when it is arrested in metaphase of meiosis II. We have unveiled how maturation is driven by a highly dynamic coordination between the core mitotic oscillator, based essentially on MPF, Cdc25 and APC, and the MAPK signaling pathway, which are both stimulated by the same extracellular signal. Although the model does not incorporate several regulatory schemes discovered recently [34], [43], it appears to be sufficiently detailed to gain insight into the essential features of maturation and to discriminate the roles of different regulatory motifs. It can therefore serve as a solid basis for further explorations.
Resumption of meiosis requires MPF activation, which is potentially mediated by a multiplicity of pathways, including Plx1-dependent changes of Myt1-Cdc25 balance as well as Mos-dependent inhibition of Myt1 and Cyclin synthesis. Inhibiting one of these pathways in the model delays or compromises maturation initiation, as in experiments [8], [40]. This suggests redundant and cooperative roles between these various translational and post-translational MPF activation schemes. However, model parameters can be adjusted so that cyclin B synthesis or Mos activation occur after GVBD and would therefore be fully dispensable for MPF activation as it is observed in various organisms [42]. The role of MAPK is more pronounced at the meiotic transition where it leads to the Rsk-dependent activation of Emi2 required for MII entry and metaphase II arrest [10], [20]. Simultaneous activation of MAPK and MPF raised nevertheless intriguing questions (Wu and Kornbluth, 2008): why do eggs arrest at MII but not at MI? What causes the delayed and sharp activation of Emi2? Our model reconciles different views on these questions by showing that both the late translational control by CPEB4 (Igea and Mendez, 2010) and the temporally-controlled antagonistic roles of MPF and Rsk in stabilizing Emi2 [13] contribute to this delay.
It is worth mentioning that our modeling analysis does not capture the role of the few links that couple downstream effectors of MAPK - essentially ERK and Rsk - with components Cdc25 and Myt1 of the autoamplification loop [30], [31], [33], [44]. It remains unclear whether and how these MPF-activating pathways contribute to G2/M transition and to meiotic transition. A controversial hypothesis is the existence of a transient activation of MAPK or/and Rsk shortly after progesterone injection, stimulating MPF activation [33], [45], [46], which could be easily incorporated into models if needed. Another possibility is that these regulations also play a role in consolidating Myt1 inactivation and Cdc25 activation at meiotic transition to compensate the transient MPF activity decrease, an hypothesis that needs to be further tested in both models and experiments.
The availability of a regulatory network model that qualitatively reproduces a broad spectrum of experimental data allows us to investigate the design principles that underlie a reliable maturation process. Bifurcation and sensitivity analyses of the model unveiled the existence of two independent subcircuits where feedback loops are subtly interlocked so as to achieve two coordinated but separable transitions (see Fig. 6).
The first transition, meiotic resumption, relies on a circuit that involves several signaling pathways and positive-feedback loops. This module is organized around the core autoamplification loop which includes MPF, Myt1, Cdc25 and Plx1 and drives the sharp post-translational activation of MPF associated with G2/MI transition (Fig. 6A). This loop is supplemented with other positive-feedback loops and coherent feedforward loops featuring CPEB1 and Mos, which ensures a simultaneous activation of MPF, the Mos/MAPK pathway and the translational machinery (Fig. 6B). The role of these combined positive-feedback motifs in the MPF and MAPK modules is not related to robustness against noise [47], activation threshold tuning [48] or multistability [18], [35], but is rather aimed to induce and sustain high MAPK activity throughout maturation, independently of the MPF activity level which decreases due to the APC-dependent degradation of cyclin subunits at the end of the first meiosis (Fig. 6C).
The second transition from meiosis I to meiosis II indeed requires high MAPK levels to promote MPF stabilization. Late reactivation of MPF is driven by a delayed positive-feedback loop involving Emi2 that counteracts the negative feedback mediated by APC. Delayed activation of Emi2 is itself the result of the incoherent feedforward loop in which MPF both activates and inactivates Emi2 (Fig. 6D). This sophisticated regulatory scheme provides an interesting example of how the combination between positive and negative feedback loops gives rise to complex dynamics such as non-monotonous bistable behaviors, besides those that have already been studied in the context of oscillatory, excitable and bistable dynamics [36], [49], [50].
Meiotic maturation poses a difficult challenge to oocyte cells. A single transient signal must be followed by a coordinated sequence of two crucial and distinct decisions, MI entry and MI/MII transition, which both require a sharp MPF activation. Our findings reveal the sophisticated molecular network mechanisms that provide an original solution to this problem. Firstly, like in other biological decision-making processes, the two main meiotic decisions rely on two distinct positive-feedback-based circuits, each of which combining multiple loops so as to create sharp and robust transitions. Secondly, interference and retroactivity between the two decision circuits are minimized by using separate and partly independent regulatory schemes based on post-translational modifications and protein turnover control, respectively. Lastly, the coordination of the decision systems is mediated by the existence of a negative feedback loop and an incoherent feedfoward loops, which are known to be efficient for scheduling temporal gaps between successive decisions [51], [52]. Thanks to this specific regulatory and feedback architecture, a transient signal can trigger complex dynamical and phenotypical trajectories which are attracted by a one-dimensional slow manifold and follow it throughout maturation. This dynamical process is reminiscent of the phenomenon of canalization during multicellular development [53]. Overall, this encourages further efforts to decipher the dynamical behavior of molecular networks with complex feedback and feedforward topology, especially when they combine oscillatory and irreversible behaviors, as occuring during meiotic maturation.
The mathematical model for the maturation regulation network is based on the molecular interactions reported in Xenopus laevis between the 12 proteins CPEB1 (abbreviation C1), CPEB4 (abbr. C4), MPF, Cdc25 (abbr. C25), Myt1 (abbr. Myt), APC, Mos, MEK, ERK, Rsk, Plx1 (abbr. Plx) and Emi2 (abbr. Emi). We assume that the activity of each protein in the list above can be post-translationally regulated, typically through phosphorylation, such that they can be either in activated or inactivated forms. The 12 molecular actors can be distinguished according to whether their total concentration is also regulated (class I: Mos, MPF, APC, Emi, C1, C4) or can be considered as constant on the time scale of maturation (class II: Plx, C25, Myt, MEK, ERK, Rsk).
For class-I proteins, the concentrations of the active and inactive proteins evolve in time according to a set of differential equations (Table 2). where and denote the concentration of the active or inactive forms of protein , whereas , , and denote, respectively, the synthesis, degradation, activation and inactivation rates of these proteins. We assume Michaelis-Menten kinetics for activating and inactivating reactions where and are the maximum rate of the reactions (Table 2). Only Emi2 is assumed to have more than two states: (i) partially activated when unphosphorylated with a dephosphorylation reaction rate ; (ii) inactivated through phosphorylation by Rsk with a reaction rate ; (iii) fully-activated through phosphorylation by Rsk of unphosphorylated or MPF-phosphorylated forms, with a reaction rate . The concentrations of these three forms are denoted by , and , respectively. The assumption that the total concentrations of class-II proteins remain relatively stable throughout the maturation process allows us to use the quasi-steady state approximation. Moreover, if phosphorylation and dephosphorylation reactions operate in the linear regime, the steady-state concentrations can be obtained as a function of the total concentration (normalized here to 1) and of their maximum activating and inactivating reaction rates and . These expressions depend on whether activation is achieved through a one-step phosphorylation (C25, Myt, Plx) or a two-step phosphorylation (MEK, ERK, Rsk) (Table 2).
The kinetic rates , , and that appear in both differential equations and steady-state concentrations can be either considered as constant parameters on the time scale of maturation process or as time-dependent variable as they may depend on the concentration of other dynamic regulators of the maturation process. Such dependence is depicted as links in the network representation of Fig. 7 is given in Table 2.
The model contains a large number of kinetic parameters (73), which, for the most, have not been estimated experimentally so far. A preliminary step toward the adjustment of parameters is to reduce their number. To account for the dynamical features of the maturation process, the mathematical model only needs to describe the evolution of protein concentrations relative to each other. We can therefore normalize protein concentrations. First, the total concentration of class-II proteins is normalized to 1. Second, only the relative value between activation rates and inactivation rates are relevant for class-II proteins, such that we can introduce a free parameter which determines their absolute value. Third, the Michaelis constants for all activation and inactivation processes of class-I proteins are set to . Actual values of the concentrations can always be recovered by scaling the variables appropriately, keeping in mind that the present modeling study focuses on the temporal profile of protein activity rather than quantitative predictions. The normalization procedure can reduce the number of parameter to 54. The other parameters used in this study (Table 3) have been selected in a semi-arbitrary manner constrained by qualitative fitting of the time course of several components in various contexts. The kinetic parameters for the MAPK pathway have been adjusted to display the classic bistable behavior of this cascade (Fig. 7A). The kinetic parameters for Cdc25, Myt1, MPF, APC and their respective interactions have been adjusted to produce an excitable or oscillatory behavior commonly associated with a specific underlying bifurcation structure of the dynamics: a saddle node bifurcation on an invariant circle (Fig. 7B). Finally, the kinetic parameters coupling these two modules between themselves and to the input signal have been adjusted to match the temporal profile of MPF activity that is typically observed in various experimental prototocols (Fig. 7C, see also Fig. 1B and 5).
|
10.1371/journal.pgen.1004436 | Silencing Is Noisy: Population and Cell Level Noise in Telomere-Adjacent Genes Is Dependent on Telomere Position and Sir2 | Cell-to-cell gene expression noise is thought to be an important mechanism for generating phenotypic diversity. Furthermore, telomeric regions are major sites for gene amplification, which is thought to drive genetic diversity. Here we found that individual subtelomeric TLO genes exhibit increased variation in transcript and protein levels at both the cell-to-cell level as well as at the population-level. The cell-to-cell variation, termed Telomere-Adjacent Gene Expression Noise (TAGEN) was largely intrinsic noise and was dependent upon genome position: noise was reduced when a TLO gene was expressed at an ectopic internal locus and noise was elevated when a non-telomeric gene was expressed at a telomere-adjacent locus. This position-dependent TAGEN also was dependent on Sir2p, an NAD+-dependent histone deacetylase. Finally, we found that telomere silencing and TAGEN are tightly linked and regulated in cis: selection for either silencing or activation of a TLO-adjacent URA3 gene resulted in reduced noise at the neighboring TLO but not at other TLO genes. This provides experimental support to computational predictions that the ability to shift between silent and active chromatin states has a major effect on cell-to-cell noise. Furthermore, it demonstrates that these shifts affect the degree of expression variation at each telomere individually.
| Genetic diversity is often high at telomeres, the chromosome ends where genes are readily amplified and modified. Phenotypic diversity, e.g., growth properties under a given condition, is affected by stochastic variations in gene expression exhibited among cells in a homogenous environment. Our studies found that individual subtelomeric genes show high variability of gene expression both between cells within a single population and also between separate sub-populations. Cell-to-cell variation, termed Telomere-Adjacent Gene Expression Noise (TAGEN), affected single telomeric genes. We found that classical telomeric silencing and TAGEN are tightly linked, with both being dependent upon proximity to telomeres and the Sir2 chromatin modifying enzyme. In addition, both are coordinately regulated locally—at the DNA level: at a telomere with transcription that is continually silenced or activated, the level of expression variability is reduced. This work provides experimental support for computational work that predicted this relationship between stochastic chromatin silencing and expression plasticity at each telomere individually. Furthermore, it demonstrates that these shifts affect the degree of cell-to cell noise of telomere-adjacent loci.
| Responsiveness to minor changes in the environment requires exquisitely sensitive phenotypic plasticity. This can be executed via many different mechanisms, operating on different time scales, with different types of condition-specific responses, but usually includes changes in transcriptional and translational profiles. Variation between independent populations of cells that are presumed to be isogenic can be due to altered epigenetic properties, such as chromatin status of specific genes or chromosomal regions [1], [2], to cell-to-cell variations in gene expression [3], [4]. Such population and cellular variations are likely to operate continuously in natural environments. Microbes living within a mammalian host encounter a variety of host niches. For example, organisms that reside throughout the GI tract must be able to survive conditions in the oral cavity (pH 6.5–6.9, 33–35°C), the stomach (pH 2, 37°C), the small intestine (pH 7.4, 37–40°C), and anaerobic niches in the colon. Accordingly, the ability to acclimate rapidly to changing environments is thought to provide a selective advantage and is supported by studies in yeast and bacteria [5]–[9].
Gene expression noise, defined as cell-to-cell variation in levels of transcription and/or translation, provides phenotypic diversity within an isogenic population, enabling sister cells to respond differently to environmental challenges. Noise can be extrinsic, generally assumed to be due to differences in an environment or to natural variations in cell components such as transcription or translation factors that affect multiple alleles similarly [2], [3], [10]. By contrast, intrinsic noise is allele-specific and is often due to changes in the frequency with which transcription initiates from a given promoter [11], [12]. Intrinsic noise can provide a larger range of responses to environmental conditions, because the relative amounts of one gene product to another can shift more dramatically [13]. The quantitative contributions of extrinsic and intrinsic noise can be distinguished using different fluorescent protein fusions driven from otherwise identical alleles; extrinsic noise will result in correlated relative expression of both alleles, while intrinsic noise will result in independent relative expression of each allele [13]. The degree to which these types of noise contribute to different aspects of organismal survival by producing phenotypic diversity remains to be determined.
C. albicans is an organism that survives and flourishes in a wide range of niches within its human host. It engages in a benign commensal lifestyle, residing in the oral cavity and colonizing the GI tract [14]. In some hosts, especially following antibiotic treatment or immune suppression, it switches to a pathogenic state and becomes blood-borne, colonizing internal organs including the kidney, heart, or brain. C. albicans is generally found in the diploid state and it is known to tolerate high levels of genotypic and protein variation including aneuploidy and codon ambiguity [15]–[17]. Under stress conditions, e.g. during drug exposure, certain aneuploidies can provide improved fitness, largely due to increased expression of genes specifically found in extra copies on the aneuploid chromosomes [18]–[21]. Furthermore, while aneuploidy in general often incurs a high fitness cost, some aneuploidies have very little cost, even under non-selective conditions [22]–[24]. C. albicans also has a highly variable proteome because of the ambiguous CUG codon, which encodes serine most of the time. The CUG codons also encode leucine at low frequency in cells under non-stress conditions and at higher frequencies if cells are stressed [15].
C. albicans is the most virulent of the CUG clade organisms and this is thought to be due, at least in part, to amplification of several gene families thought to be important for virulence. These include the SAP [25], LIP [26], and ALS [27] gene families that encode proteases, lipases and cell wall adhesins, respectively. The most amplified of all the gene families in C. albicans are the TLO genes, present in 1 copy in most CUG family members, in 2 copies in C. dubliniensis [28] and in 14 copies in C. albicans [29]. All but one of the TLO genes is telomere-adjacent, usually found as the most telomere-proximal, or the penultimate, gene on the chromosome [30]. The TLO gene family encodes a set of related proteins with a Med2 domain, all of which are thought to function as exchangeable Med2 subunits for the Mediator transcription regulation complex [31]. However, how TLO gene expression is regulated and whether Tlo proteins contribute to the phenotypic plasticity of C. albicans has not been explored.
In many organisms, genes at telomeres are subject to telomere position effect (TPE), a transient transcriptional silencing due to specific chromatin complexes that are thought to nucleate at the telomeres and to spread inward along the chromosome arm [32], [33]. Studies of TPE generally detect two expression states (“ON” or “OFF”) using phenotypic read outs interpreted as indicating a biphasic open or closed chromatin state at a given telomere [34]. TPE is dependent upon the Silent Information Regulator proteins Sir2p, Sir3p and Sir4p in S. cerevisiae [35], [36]. Sir2p, an NAD+-dependent histone deacetylase (HDAC), is highly conserved in prokaryotes as well as eukaryotes [37] and contributes to silencing at the telomeres of organisms ranging from S. pombe to mice [38].
In S. cerevisiae, gene expression noise has been reported to be position-dependent. In one study, noise of two unrelated genes was shown to be influenced by their positions at internal loci on two different chromosome arms [12]. Bioinformatic meta-analysis of gene expression along all chromosome arms showed increased gene noise correlated with increased:1) proximity to the telomere; 2) prevalence of genes with promoters containing TATA box motifs; 3) intermediate levels of expression and 4) transitions between silencing-specific histone modifications [39]. The latter is not surprising, given that a number of histone modifiers affect gene expression noise through effects on transcription burst size as well as burst frequency [40]. This likely occurs through the regulation of nucleosome occupancy, which is different between promoters with TATA motifs and those without TATA motifs [41] and likely involves interactions with transcription factors as well [42].
Many of the chromatin modifier genes that affect noise encode HDACs. These include RPD3 and HDA1 [40]. In C. albicans, HDACs have been characterized to some degree, with Sir2 being reported to affect phenotypic switching under at least some conditions [43] and Hst3, Hda1, and seven other chromatin modifiers have been shown to alter white-opaque switching [44], [45]. Additionally, the Set3C complex, Set3 and Hos2, inhibit the yeast-to-filamentous transition by modulating transcriptional kinetics of key morphogenic regulators [46].
The association of noise with telomere proximity has only been explored experimentally in one study using C. glabrata, a pathogenic yeast most closely related to S. cerevisiae. EPA1, a subtelomeric gene that encodes a virulence-related adhesin [47], is subject to TPE and silencing contributed to high levels of EPA1 gene expression noise [48]. This study detected effects at one telomeric locus but did not address the question of whether the effect was due to the telomere-adjacent position of the gene. Nonetheless, this work suggests that telomeric silencing by Sir2p may be associated with the highly variable expression of telomere-adjacent genes.
Here we investigated the expression of telomere-adjacent genes in C. albicans, with a focus on the TLO gene family. We detected high levels of variability between isogenic isolates at the population level, and, on average, genes that are most telomere-proximal on each chromosome have higher than average expression plasticity. Furthermore, telomere-adjacent genes exhibited high levels of noise (cell-to-cell variation in expression levels) that was largely due to intrinsic noise. Importantly, this telomere-adjacent gene expression noise (TAGEN) was dependent on genome position; TLO genes had lower noise levels when moved to an internal locus and a non-telomeric gene had higher noise when moved to a sub-telomeric locus. Similar to telomeric silencing, TAGEN was dependent upon NAD-dependent HDAC activity and, to a large degree, upon Sir2p. Finally, selection for either constitutive expression or constitutive silencing of a TLO-adjacent URA3 gene specifically reduced the expression plasticity of the neighboring TLO, in cis, but had no effect on expression plasticity at other TLO genes in trans. Thus, TAGEN generates expression variability as a consequence of dynamic, local chromatin-mediated position-dependent silencing.
In the course of measuring TLO gene expression under a range of growth conditions, we found that expression levels for many individual TLO genes was strikingly variable (up to several orders of magnitude) between isogenic biological populations grown from single colonies under identical conditions (Fig. 1A–C). Furthermore, the level of TLO gene expression variation, measured as the coefficient of variation (CV; standard deviation divided by the mean; at least five replicates per gene-condition) [49], was far greater than that seen for two control genes, SOD2 and HGT20, that were expressed at similar average levels, irrespective of the growth conditions (Table S4). Transcript abundance measurements were reproducible for individual populations (average standard deviation among technical replicates = 0.63 cycles vs. 4.43 cycles between biological replicates), further supporting the idea that the population-level expression of individual TLO genes varied considerably.
Genes with high cell-to-cell variation in gene expression are often differentially expressed across a large number of environments [41]. To ask if this is the case for the TLO genes, we analyzed an RNA-Seq dataset for expression of all C. albicans genes under eleven different environmental conditions [50]. Across growth conditions, the 13 telomeric TLO genes generally had high CV values relative to the average for all C. albicans genes analyzed by RNA-seq (Fig. S1A), and, as a group, their mean CV value was significantly higher than for a set of 13 randomly chosen genes (determined by examining 50,000 simulated gene sets, p<0.025, Fig. S1B). Cells either mock-treated or exposed to a variety of stresses were equally variable (Fig. S1C). Interestingly, the CV value for TLOα34, the only non-telomeric member of the TLO gene family, had a lower CV than the average telomere-adjacent TLO gene (Fig. S1A, blue arrow).
We next asked if Tlo protein levels were also variable. To detect individual Tlo protein levels, we constructed strains with a single copy of GFP fused to a given TLO gene and detected the fusion protein with an antibody to GFP. Tlo-GFP levels were highly variable among biological replicates grown from single colonies under identical conditions. For example, when different colonies expressing Tloβ2-GFP were prepared for protein extraction from independent log-phase cultures on the same day, the levels of GFP were much more different than a similar comparison of two control proteins (Fig. 1D). Two other Tlo-GFP fusion proteins (representing all three Tlo protein clades [51] showed similar variability when examined under several growth conditions (Fig. 1E, F). Of note, differences in the protein levels of Tlos generally were less dramatic as those seen for transcripts. Nonetheless, individual Tlo protein levels varied considerably between different biological replicate populations.
Expression variability between isolates could be the result of expression differences between whole populations or due to cell-to-cell variation within a population. We hypothesized that this high level of variability from population to population could be due to TLO gene expression differences originating from variability between colonies grown on solid agar plates. Based on the assumption that colony growth on solid media subjects cells to intense founder effects and/ot different local environments [52], [53], we asked if Tlo expression differences become less evident after cells from single colonies were propagated in liquid medium, assumed to be a more uniform environment that is also less sensitive to founder effects because cells are continuously mixed. To address this question, we compared Tloα12-GFP expression profiles from 6 individual colonies, originating from a single parent colony, that were grown on solid media plates and the same six populations after two days of passaging in a constantly agitated liquid medium (Fig. 2A). The irregular shapes of expression profiles for cells from individual colonies that were prepared for flow cytometry (by propagation in liquid medium for two hours), suggested that these cultures contained mixtures of different subpopulations. Furthermore, these profile shapes were different for the six colonies, suggesting different founder effects. Because cells lifted from a colony are closely related both genetically and epigenetically (more likely to be in the same silencing state), we think variability in silencing states and, potentially, the local environments within a colony produce these profile differences. In contrast, passaging the same colony isolates in liquid medium for two days resulted in expression profiles that were more regularly shaped and more similar to one another (Figure 2).
Passaging in liquid for two days did not significantly alter Tloα12-GFP mean expression (t5 = 1.38, p = 0.29) or mean robust CV (t5 = 1.90, p = 0.12) among the five wild-type populations. However, the variance among populations was significantly reduced for both mean expression (F{5, 5} = 71.7, p = 0.0002) and robust CV (F{5, 5} = 32.7, p = 0.002) (Fig. 2C). This suggests that either the populations became more homogeneous because distinct subpopulations were better mixed in liquid culture, and/or because Tloα12-GFP expression was more uniform in a more homogenous environment.
In S. cerevisiae, Sir chromatin modifiers affect telomeric silencing, with the Sir2p NAD+-dependent histone deacetylase (HDAC) being the most evolutionarily conserved. To ask if Sir2 regulates the colony-to-colony variation observed, we performed flow cytometry on different colonies expressing Tloα12-GFP in a sir2Δ/Δ strain. Mean fluorescence of Tloα12-GFP in a sir2Δ/Δ background did not change (t5 = −2.13, p = 0.087), while Robust CV significantly decreased (t5 = 14.01, p<0.0001) after liquid passaging (Fig. 2B, C). As in the wild-type background, both fluorescence intensity and Robust CV show less population-to-population variability after liquid passaging (mean fluorescence: F{5, 5} = 10.93, p = 0.020; CV: F{5, 5} = 8.76, p = 0.035). Comparing the variance among populations of wild-type and sir2Δ/Δ cells, the wild-type populations were always more variable than the sir2Δ/Δ populations, regardless of the parameter or the timepoint (D0, mean fluorescence: F{5, 5} = 93.62, p = 0.0001; D2, mean fluorescence: F{5, 5} = 14.27, p = 0.011; D0, CV: F{5, 5} = 62.22, p = 0.0003; D2, CV: F{5, 5} = 16.65, p = 0.0078). Thus, the absence of Sir2 protein reduced the founders effect seen in WT populations isolated from different colonies, suggesting that the function of wild-type Sir2 is to mediate the variation in expression of Tloα12-GFP.
To further test the founder effect on Tlo expression, we examined expression of Tloα12-GFP protein in cells originating from opposite sides of the same colony. Interestingly, flow cytometry profiles (after 2 hours of liquid growth) differed for the different colony regions (Fig. 2D), suggesting that populations of cells within a colony have different degrees of expression and that each population can have different levels of cell-to-cell noise. It also implies that the reduction in noise following overnight growth in liquid is not a simple function of more uniform mixing in the liquid media. Thus, it appears that colony regions have different levels of expression and of cell-to-cell noise (Fig. 2A, C, D). In contrast, flow cytometry profiles of Tloα12-GFP expression from different parts of a single sir2Δ/Δ colony were similar (Fig. 2E). Therefore, expression variability between and within single colonies is Sir2p-dependent. Furthermore, although microenvironments may differ within a colony [52], expression levels do not vary considerably within sir2Δ/Δ colonies, suggesting that the variation seen in wild-type cells is either not due to microenvironmental differences or that Sir2 is required to sense those microenvironmental differences. We propose that the variation at Tlo genes is primarily a function of intrinsic noise rather than a response to the microenvironment.
To address the degree of heritability of Tloα12-GFP expression levels and expression noise, we analyzed the expression level of mother-daughter cell pairs by pedigree analysis. We isolated 10 mother-daughter pairs, dissected buds from mothers, and allowed them to grow separately on a plate for 18 hours (Fig. 3A). We compared populations of 50 cells from individual mothers to 50 cells from their own daughters to ask if these related populations were more similar to one another than expected by chance (Fig. 3B). The mean difference in absolute ln(expression) was 0.58 for the mother-daughter pairs and was 1.24 for randomized daughter pairs, with the 5% quantile at 0.96. Thus, the mother-daughter pairs were significantly similar to one another (p≤0.0001) than expected by chance (Figure 3C). Interestingly, two daughter populations (Colonies 2 and 10, Figure 3C) did not exhibit perfect overlap with their respective mother populations, indicating that expression similarity, although heritable, can diverge over a small number of generations.
The studies above analyzed primarily variation in mean and CV of populations of cells. Gene expression noise is studied at the level of cell-to-cell differences, so we next measured cell-to-cell variation using fluorescence microscopy of individual cells isolated from multiple populations (originating from single colonies). We analyzed the cell-cell variation (measured as CV) within each population (founded from a single colony), and also compared the CV between different populations. For microscopy studies we analyzed 50 cells from each population of five Tloα and Tloβ clade fusion proteins, which localize to the nucleus and are expressed at higher levels (and thus are more detectable by fluorescence microscopy than Tloγ-clade genes) [51].
Strikingly, the fluorescence signal for subtelomeric Tlo genes varied dramatically from cell-to-cell, ranging from very bright cells to cells with no obvious signal (Fig. 4A, B). The level of population-to-population variation was also higher for subtelomeric Tlo genes, consistent with the detection of expression plasticity at the population level (Fig. 1). Growth under stress conditions (5 mM H2O2 or cell wall stress) also resulted in high levels of Tloα12 cell-to-cell variation (Fig. S2; p<0.001; significance determined using a bootstrap procedure that compared the measured ratio of CVNup49-GFP/CVTloα12-GFP against the critical value obtained from 10,000 simulated datasets that randomized the background of measured cells). Consistent with the RNA-seq results, the non-telomeric Tloα34-GFP gene, exhibited minimal cell-to-cell and population-to-population variation (Fig. 4A, B).
To measure gene expression levels for much larger numbers of cells, we analyzed GFP expression levels using flow cytometry (100,000 cells per population). Nup49, which encodes a nuclear pore component expressed at similar average levels to the Tloα and Tloβ proteins, exhibited minimal variation between cells within a population (evident by examining the peak width) and between populations (Fig. 4C, S3). In contrast, both cell-cell and population-population variability was much greater for Tlo-GFP than for Nup49-GFP fluorescence levels (Fig. 4C).
Two general sources of cell-to-cell variation have been explored extensively in many different species [1], [3], [12], [13], [40]. Extrinsic noise is due to conditions that differ between cells, such as a general level of ribosome or a local exposure to different growth conditions (Fig. 2). In contrast, intrinsic noise operates independently on different alleles of the same gene or promoter. The classic method to distinguish between extrinsic and intrinsic noise is to tag two different alleles of the same gene/promoter with two different fluorescent proteins and to observe the relative levels of each on a cell-by-cell basis. Accordingly, we tagged both alleles of TLOα12 or TLOβ2, using GFP for one allele and mCherry for the other, and determined the degree to which each of the alleles was expressed in individual cells by fluorescence microscopy (Fig. 5A). Extrinsic noise manifests as variable yet correlated expression of the two alleles, while intrinsic noise results in independent, allele-specific expression levels.
The relationship between mCherry and GFP expression in Nup49 (control), Tloα12, and Tloβ2 were clearly different, based on fluorescence intensities (Figs. 5B, S4). In each individual population (12 populations for each tagged gene, see methods) a simple correlation test between the two fluorophores indicated that there were considerable differences for the three tagged genes (Fig. S4, Table S5). We considered the 12 populations for each gene as independent because different colonies and different locations within colonies were different enough from one another that they were not good predictors of the degree of either intrinsic or extrinsic noise (Table S5). The levels of both intrinsic and extrinsic noise (extrinsic: F2, 33 = 12.8, p<0.0001; intrinsic: F2, 33 = 26.5, p<0.0001, Fig. 5C) were different for the different genes measured. Post-hoc Tukey tests indicated the difference between the two types of noise; the two TLO genes both had significantly higher intrinsic noise than Nup49. On the other hand, extrinsic noise levels were not specific to TLO genes. Tloβ2 has significantly less extrinsic noise than Nup49 or Tloα12 (which were not different from each other). Furthermore, for both TLO genes, the contribution of intrinsic noise to total noise was significantly greater than the contribution of extrinsic noise (Tloβ2: t11 = −16.8, p<0.0001; Tloα12: t11 = −6.5, p<0.0001, Nup49 t11 = 0.056, p = 0.96, Fig. 5C).
To investigate whether increased expression plasticity is a general property of telomere-proximal genes, we examined the expression of sets of 16 genes starting with the most telomere-proximal and stepping sequentially into chromosome internal genes using the available C. albicans RNA-Seq dataset [50]. Both sets of the 16 most telomere-proximal genes (including 9 of 13 subtelomeric TLOs) and the set of 16 penultimate telomere-adjacent genes (including 4 of 13 subtelomeric TLOs) were significantly more transcriptionally variable than sets of 16 random genes (Fig. S5A, B; significance determined by a bootstrapping procedure as described above; p<0.025 in both cases). A similar trend was seen for the genes in the third-most telomere-proximal position (Fig. S5B). However, this pattern did not continue as a general trend along the chromosome (Figure S5C), indicating that any ‘spreading of TAGEN’ inwards from the telomere does not propagate more than ∼8 kb into the chromosome arms.
Many studies of S. cerevisiae found that differences in promoter structure correlate with differences in the amplitude of gene noise [54], [55]. To determine the extent to which telomere position and promoter structure affect the variability of TLO gene expression, we constructed two TLO-NUP49 swap strains (Fig. 6A): NUP49-GFP@TLO, in which the control gene NUP49-GFP, together with its native promoter, was moved to the sub-telomeric TLOα9 locus on the left end of Chromosome 4 (YJB12963); and TLOα9-GFP@NUP49, in which TLOα9-GFP, together with its native promoter, was moved to the internal NUP49 locus on the right arm of Chromosome 1 (YJB12966). Importantly, when either Nup49-GFP or Tloα9-GFP were expressed at the NUP49 locus, noise (as measured by fluorescence microscopy) was significantly lower than when either of these proteins was expressed from the TLOα9 locus (Fig. 6A–C, Fig. S6; p<0.05). Expression of Nup49-GFP and Tloα9-GFP was also significantly lower at the TLOα9 locus compared to the NUP49 locus (Fig. 6A–C; NUP49: t85.42 = 16.43, p<0.00001; TLOα9: t85.44 = 4.71, p<0.00001). Flow cytometric analysis of the four strains (two with tagged genes at their native loci and two with swapped loci) also indicated that genes at the subtelomeric TLOα9 locus exhibit a significant decrease in the mean fluorescence signal (position: F1 = 5.04, p = 0.038, gene: F1 = 0.93, p = 0.35) and an increase in the level of gene noise (Robust CV; position: F1 = 10.12, p = 0.005, gene: F1 = 2.10, p = 0.17) relative to the internal NUP49 locus (Fig. 6D). This suggests that the subtelomeric TLOα9 locus is sufficient to cause increased noise because it is telomere-adjacent and affected by Telomere-Adjacent Gene Expression Noise (TAGEN), which influences both population-to-population (expression plasticity) and cell-to-cell (noise) variability. Furthermore, TAGEN appears to be independent of the promoters tested.
The Sir2p HDAC was required for TLO expression variability between colonies. Therefore, we hypothesized Sir2 may also influence TLO noise among cells in a single population. We first asked whether addition of nicotinamide (NAM), an inhibitor of NAD+-dependent HDACs, or deletion of SIR2 had an effect on TAGEN at TLO genes using qRT-PCR. Addition of NAM or the lack of Sir2p significantly reduced expression plasticity (measured with qPCR, Fig. 7A, S7; background: F1 = 6.44, p = 0.020; NAM: F1 = 7.79, p = 0.011; interaction: F1 = 3.25, p = 0.086), while neither NAM nor the absence of Sir2p significantly influenced mean TLO gene expression (Fig. S7; Sir2 background: F1 = 0.03, p = 0.86; NAM: F1 = 1.42, p = 0.25). Furthermore, the effect of deleting SIR2 together with NAM exposure affected expression and plasticity to a similar degree as either NAM or deletion of SIR2 alone: reduced variability with little effect on expression levels (interaction; CV: F1 = 3.25, p = 0.086). Similar results for wild-type vs sir2Δ/Δ mutants were obtained by microscopy (Fig. 7B, S8; p<0.05) as well as by flow cytometry of Tloα10-GFP or Tloα12-GFP (Fig. 7C; Robust CV; gene: F1 = 1.21, p = 0.29; Sir2 background: F1 = 5.44, p = 0.03; interaction: F1 = 0.165, p = 0.69). Thus, Sir2p makes a significant contribution to expression plasticity of Tloα10-GFP and Tloα12-GFP.
To ask if Sir2p contributes to the position-dependent aspect of TLO TAGEN, we compared the level of expression noise for the Nup49-GFP@TLOα9 locus in a sir2Δ/Δ strain relative to the level of expression noise for the Nup49-@TLOα9 locus in a wild-type strain. Importantly, the expression noise for Nup49-GFP was decreased in a sir2Δ/Δ strain only for Nup49-GFP@TLOα9 locus and not for Nup49-GFP at its native locus (Fig. 7D, Robust CV; position: F1 = 11.38, p = 0.005, background: F1 = 5.10, p = 0.042, interaction: F1 = 6.91, p = 0.021). Thus, the position-dependent and promoter-independent TAGEN seen at TLO genes is dependent upon Sir2p and, most likely, dependent upon its activity as a NAD+-dependent HDAC.
Telomeric silencing is considered to be a process by which telomeres toggle between “OPEN” and “CLOSED” chromatin states. Such a biphasic switch would be expected to generate two subpopulations of cells that would be distinguishable by flow cytometry as having different expression peaks. Yet, expression profiles of specific Tlo-GFP fusion proteins did not exhibit two clear peaks. This could be due to regulation of TLO expression by multiple factors [4] or a relatively fast rate of switching between two expression states [56]. Thus, we explored the role of additional chromatin modifiers in the regulation of TLO expression levels and the degree of TLO TAGEN. Nine modifiers were analyzed by qRT-PCR. HST1 and SET1 influenced expression plasticity (HST1: t6 = −2.89, p = 0.028, SET1: t6 = −2.60, p = 0.041) but not expression levels (HST1: t6 = −0.99, p = 0.36, SET1: t6 = 1.20, p = 0.27), while HDA1, HOS2, HST2, PHO13, NAT4, RPD31, and SET3 had no effect on expression level or plasticity (Fig. S9 and data not shown). Consistent with the qRT-PCR results, deletion of HST1, a SIR2 paralog that affects some telomere-associated genes in S. cerevisiae [57], [58], resulted in decreased fluorescence signal for two GFP-tagged Tlo proteins, Tloα10 (t187.2 = 7.03, p<0.0001) and Tloα12 (t139.4 = 5.30, p<0.0001) (Fig. S10A, B), as measured by fluorescence microscopy. Consistent with a role for Hst1 protein at internal as well as telomeric loci, the expression noise for Nup49-GFP at its native locus was reduced in the hst1Δ/Δ strain (p<0.05). Cell to cell noise in the hst1Δ/Δ strains was reduced at Tloα12 (p<0.01) but not at Tloα10 (Fig. S10A, C), relative to noise levels in the wild-type HST1 parent strains. Thus, unlike Sir2p, which has a major position-dependent role in enhancing noise at telomere-adjacent loci, Hst1p affects expression noise at internal as well as telomere-proximal regions and it affects expression plasticity and noise of different TLO genes differently.
We next asked if TAGEN and TPE are functionally related by measuring TLO expression variability in cells selected for constant expression or constant silencing of a TLO-adjacent selectable marker, URA3. We measured levels of the adjacent TLO (in cis) as well as an unlinked TLO (in trans), when cells were selected for expression of URA3 (ON state selected on medium lacking uridine) or when cells were selected for repression of URA3 (OFF state selected on medium containing 5-FOA) vs cells being free to ‘toggle’ between the two states (ON and OFF states, no selection on YPAD medium). We first constructed two strains, each with URA3 inserted head-to-head at a TLO-adjacent position (adjacent to TLOα9 or TLOα12; Fig. 8A) in the subtelomeres. These strains enabled the selection of cells expressing URA3 (by growth in media lacking uracil (“-ura”)), or to select for silencing of URA3 (by growth in the presence of 5-floroorotic acid (“5-FOA”)). Growth of TLO-adjacent URA3 strains on media lacking uracil or with 5-FOA reduced or increased transcript abundance of URA3, respectively (data not shown). We then asked if selection in –ura or 5-FOA influenced TLO expression plasticity (Fig. 8B). Importantly, in both strains, selection either for or against URA3 expression significantly reduced variability of the URA3-adjacent TLO transcript levels, yet it did not affect the transcript variability at an unlinked TLO (Fig. 8C; presence of selection: F1, 20 = 40.4, p<0.0001, gene: F1, 20 = 0.28, p = 0.60, interaction: F1, 20 = 0.174, p = 0.69). This occurred without a significant effect on expression levels (Fig. S11; presence of selection: F1, 20 = 0.03, = 0.87, gene: F1, 20 = 2.48, p = 0.13, interaction: F1, 20 = 0.145, p = 0.71). Thus, TAGEN at a specific TLO locus requires that cells toggle between the ON and OFF states and is lost if expression of an adjacent gene is constitutively ON or OFF. Furthermore, the effect of telomeric silencing on TAGEN occurs in cis and does not affect silencing or TLO expression at other subtelomeres.
Here, we discovered and characterized Telomere-Associated Gene Expression Noise (TAGEN), which is detectable not only as intrinsic variation at the cell-to-cell level but also generates variation at the population level. TAGEN is position-dependent, affecting only the most telomere-proximal genes, and it is reduced when cells are locked in a constant chromatin state or when cells are grown for multiple passages in liquid medium. TAGEN is subject to regulation by Sir2p in a position-dependent manner and also to other position-independent chromatin modifiers and transcription factors, e.g., Hst1p, which affect different TLO genes differently. Importantly, TAGEN is largely promoter-independent and it is tightly associated, in cis, with telomere position effect dynamics. Thus, TAGEN and TPE appear to reflect different aspects of the same phenomenon—the chromatin structure and its impact on gene expression at telomeres is dependent upon proximity to a telomere. Furthermore, increased expression plasticity and noise at telomere-adjacent genes (TAGEN) requires the dynamic process by which telomere-adjacent genes toggle between the ON and OFF states of expression presumably due to the OPEN and CLOSED states of telomeric chromatin.
At most genomic loci, noise is a phenomenon detectable only when cells are analyzed as individuals [13]. In contrast, TAGEN is detectable in populations of cells isolated from different colonies and also as a cell-to-cell variability largely due to intrinsic noise. The inherited epigenetic expression state is dependent upon telomere-adjacent position, SIR2, and the initial level of expression appears to exert a founder effect. Importantly, toggling or switching between the ON and OFF epigenetic state of cells in each population likely drives colony-to-colony variation seen at the population level (Fig. 8). A similar effect was seen for one telomere-adjacent gene, EPA1, in C. glabrata [48].
TAGEN is detected as large variations in levels of transcripts, measured by either qRT-PCR or by RNA-Seq (Figs. 1, S1). TAGEN is also evident at the individual cell level, when levels of GFP fusion proteins are measured by fluorescence microscopy or by flow cytometry (Figs. 2–7). This suggests that some of the transcriptional plasticity that affects TLO gene expression is buffered by post-transcriptional mechanisms, although we cannot rule out that the long half-life of GFP fusion proteins may contribute an additional buffering mechanism [59]. Since Tlo proteins are produced at levels far higher than they are needed [31], and since all TLOs encode a related subunit present in a single copy per Mediator complex, we suggest that excess Tlo proteins are likely subject to proteasome degradation [60]–[62].
Amplified gene families that promote growth within a relatively new environment are often located at telomeres. For example, S. cerevisiae strains used to produce wine, sherry or beer carry amplified MEL, SUC, and MAL genes, respectively, which promote breakdown of the predominant sugars in the respective fermentation processes. It is thought that the cost of amplification and diversification of gene family members is lower near telomeres [63]. In addition, the work here suggests that noise at telomeric loci may be exacerbated in a non-uniform environment (Fig. 2C). The fact that this noise is Sir2p-dependent suggests that it is a function of both TAGEN and TPE. Increased gene noise is also associated with duplicated genes [64], a common feature of expanded gene families at telomere ends. Based on this idea, subtelomeric loci populated with gene families would be expected to be transcriptionally noisy because of the reduced fitness costs associated with noise when multiple functional homologs are present. Bioinformatic analysis of gene expression in S. cerevisiae found that telomere-adjacent loci were expressed with higher levels of transcriptional noise [39].Thus, telomeres are not only safe neighborhoods for gene amplification but they are noisy neighborhoods for gene expression. We suggest that, because increased noise in non-uniform conditions is Sir2p-dependent, that it is intrinsic feature of TAGEN and, most likely, of TPE as well.
Intrinsic noise is generally thought to be influenced by the chromatin state at a given locus and is often ascribed to specific promoter structures or to interactions with specific components of the transcription regulation machinery. Consistent with this, most chromatin modifiers affect either the transcription burst frequency (frequency with which a promoter switches into a transcriptionally active state) and/or the transcription burst size (the total number of transcripts or proteins produced during each transcriptionally active state) [40], [56], [65]. Interestingly, mutations affecting TAGEN often reduced the noise level without causing a substantive change in gene expression levels (Figs. 7, S7, S9). We suggest that regulating the rate of switching between silent and active chromatin at telomeres will reduce the noise, even if it does not affect the net expression levels [3], [56]. Thus, TAGEN levels are dependent upon the frequency with which telomeric silencing opens and closes the chromatin.
TAGEN is dependent upon NAD+-dependent HDACs. Sir2p and the Sir2-like Hst1p contribute to TPE in S. cerevisiae as well as in Schizosaccharomyces pombe, Plasmodium falciparum and Drosophila melanogaster [66]–[68]. This provides further support for the idea that both processes are likely related to one another. TAGEN shows fairly smooth distributions of different expression levels per cell through a population (Figs. 2, 4, 6–7), yet TPE is considered a biphasic switch between two states [34], [69]. This is likely because TPE is often measured as a growth phenotype that must cross a specific threshold to be detected [33], [70] and has been considered as a largely population effect. In contrast, we measured TAGEN at the molecular level and, thus, detected a continuous distribution of expression levels and high levels of intrinsic noise. Importantly, the two processes appear to be inextricably linked: when cells with a TLO-adjacent URA3 gene were selected for URA3 expression to be either in all “OFF” or all “ON”, expression levels for the adjacent TLO gene were less variable than when no selective pressure was applied (Fig. 8D). This supports the idea that TAGEN is a consequence of dynamic switching between TPE states, rather than a consequence of silencing or depression of telomere gene expression per se.
In C. albicans, TLOs all encode the Med2 subunit of Mediator. In S. cerevisiae, Mediator interacts with Sir2 to modulate TPE [71], [72]. If a similar relationship exists in C. albicans, then one would expect Tlo proteins to be components of the silencing machinery itself. Consistent with this, a strain lacking Med3p, which interacts with Tlo proteins in the C. albicans Mediator complex tail, exhibits lower levels of TAGEN (data not shown). Thus, noisy TLO expression may contribute to TAGEN, and may proscribe an interesting feedback circuit. Whether the amplification of TLO genes has been an important adaptation for the recently evolved virulence features of C. albicans, and whether TAGEN and Mediator feedback play a role in this process remains to be determined.
N/A
Yeast cells were grown in standard conditions in rich medium (YPAD) at 30°C [73] unless noted otherwise. Assays were performed by diluting an overnight culture 1∶100 in fresh YPAD and grown at 30°C, 39°C, with 10% fetal bovine serum, with 5 mM H2O2, with 100 µg/µl Congo Red, or with 2 mM nicotinamide for 4 hours, as indicated.
Strains are listed in Table S1. Transformations were performed using lithium acetate as previously described [73]. Strains carrying NUP49 and TLO tagged with GFP or mCherry at the C-terminus were constructed by PCR amplification from plasmid p1602 [74], p2120, or p2343 [75], which contain GFP and URA3, GFP and NAT1, or mCherry and NAT1, respectively, using primers with at least 70 bp of homology to the target gene (Table S2). Correct insertion of the fluorescent protein in frame with the relevant TLO gene was first detected as described previously [51]. Only strains in which insertion was detected as a single unambiguous PCR fragment from a single chromosome arm were analyzed further. Integration of the construct at the expected locus was confirmed by PCR, Sanger sequencing, and Southern Blot analysis as described [51].
Locus swapping strains (Fig. 4) were constructed using a PCR amplicon containing the full open reading frame (ORF) to be moved, including either all sequences up to the adjacent open reading frames or 1 kb upstream and 1 kb downstream, whichever was shorter, the fluorescent tag, and the selectable marker from previously constructed strains. Transformation and screening were performed as described above.
Transcript abundance measurements by qRT-PCR were performed as described [51] with primers listed in Table S3. Absolute quantification of SYBR fluorescence using the 2nd derivative maximum value was used to calculate ΔCT values using SEC14 as a control. All qRT-PCR results represent the average abundance of at least four independent cultures for each strain of interest.
RNA-Seq data for C. albicans grown under 11 different conditions in biological duplicates was obtained from Bruno et al [50]. We determined the coefficient of variation (CV = standard deviation divided by the mean) for each gene in each of the eleven environments that data were available for. We then averaged across all environments to determine the average CV for each gene. To determine whether a group of genes was significantly more transcriptionally variable than average, we conducted a bootstrap procedure to obtain a distribution of mean CV values for a group of genes of the appropriate size (i.e., 13 to examine TLO expression plasticity, 16 to examine position effects). We simulated 50 000 gene groups using the ‘sample’ function in the R Programming Language on the 6006 ORFs measured in the Bruno dataset; the 97.5% quantile of these 50,000 datasets was used to determine the critical value.
Protein lysates were collected as previously described [76]. Briefly, cells were inoculated into liquid YPAD cultures and grown overnight to stationary phase at 30°C with constant shaking. A 1∶100 dilution was then transferred to fresh YPAD and grown for four hours at 30°C with constant shaking prior to collecting lysates. Proteins were separated on a 12% polyacrylamide geland transferred to PVDF membrane (Immobilon-P, Millipore, Billerica, MA) as previously described [75]. Western blots were performed with mouse anti-GFP (Roche, Penzberg, Germany), rabbit anti-H4 (Santa Cruz Biotechnology, Santa Cruz, CA), and mouse anti-PSTAIR ab10345 (abcam, Cambridge, MA) followed by HRP-anti mouse or HRP-anti rabbit antibody (Santa Cruz Biotech, Santa Cruz, CA). Densitometry of band intensities was quantified using Fiji/ImageJ v1.46 (NIH, Washington D.C, District of Columbia).
TLOα12-GFP cells were struck onto SDC agar plates. Ten single cells were isolated using an Olympus BX40 dissecting microscope and followed during growth and division. Following the first division the mother and daughter cells were separated and allowed to grow up for 18 hours on the SDC agar plate. Tloα12-GFP expression was visualized by microscopy. We compared the mean difference in absolute ln(expression) values from colonies of daughter cells with 10,000 randomized affiliations.
Overnight cultures in YPAD were diluted 1∶100 in fresh SDC medium and grown at 30°C for 3–4 hours. DNA was stained with DAPI (4′,6-diamidino-2-phenylindole) (Sigma, St. Louis, MO) diluted 1∶1000 for 25 minutes, washed twice in fresh SDC, and imaged using differential interference contrast (DIC) and epifluorescence microscopy with a Nikon Eclipse E600 photomicroscope (Chroma Technology Corp., Brattleboro, VT). Digital images were collected using a CoolSnap HQ camera (Photometrics, Tucson, AZ) and MetaMorph software, version 6.2r5 (Universal Imaging Corp., Downingtown, PA). A total of 8 fields, were collected with 8 fluorescent images along the z axis, in 1-µm increments, for each cell to insure that any signal present was captured throughout the diameter of the cell. Exposure times were 500 ms for Nup49 and Tlo fluorescent fusion proteins. Projections of the z series were constructed with the stack arithmetic/sum function of MetaMorph for analysis and presentation.
Fluorescent-tagged protein abundance for each cell was measured by subtracting the average pixel intensity of three 4×4 regions of adjacent background from each of three 4×4 pixel regions within each nucleus. The signal intensity was defined as the average of the three background-subtracted nuclear regions. Nuclear signal intensity was determined for all cells in a minimum of 50 cells for each strain of interest. For all experiments, an equal number of cells were examined for expression and noise; for strains where data from more than the minimum number cells was collected, we used the ‘sample’ procedure in the R programming language [77] to randomly select cells to be analyzed.
Extrinsic and intrinsic noise was calculated as in Elowitz et. al [13]. Three strains (NUP49-GFP/NUP49-mCherry, TLOβ2-GFP/TLOβ2-mCherry, and TLOα12-GFP/TLOα12-mCherry) were streaked onto YPAD solid agar plates. Three colonies were chosen for each strain and cells from four regions of each colony were sampled (two from the edges of the colony and two from the center). These cells were suspended in liquid and the expression of the GFP and mCherry tagged genes was quantified by fluorescence microscopy for 50 cells using the method described above. The cells were also then cultured in liquid YPAD media for two days with passaging every 24 hours. Cells were taken in logarithmic growth (OD600∼0.5) after two days and 50 cells were measured again for GFP and mCherry fluorescence signal by microscopy.
Cells for flow cytometry were prepared using a modified protocol from Sudbery [78]. An overnight culture in YPAD was diluted 1∶100 in fresh SDC media and grown at 30°C for 3–4 hours. Cultures in mid-logarithmic growth (OD600∼0.5) were collected at 1500×g, resuspended in 4% methanol-free formaldehyde (Thermo Scientific, Rockford, IL), and incubated on a tube rotator for 30 minutes. Cells were then spun down and resuspended in ice cold methanol for 3 minutes, washed three times in 55 mM HCl, resuspended in 500 µl of 5 mg/ml pepsin in 55 mM HCl, and incubated for 30 minutes at 37°C with gentle shaking. Cells were collected by centrifugation, washed three times with 1 ml of 10 mM Tris (pH 7.5), and resuspended in 460 µl Buffer A [78]. Cells were incubated in 40 µl of 1 mg/ml Zymolyase-20T (ICN Biomedicals, New York, New York) in 0.1 M phosphate buffer (pH 7.5) and 1 µl β-mercaptoethanol for 30 minutes at 37°C with gentle shaking and washed 5 times with 1% bovine serum albumin (BSA) in phosphate-buffered saline (PBS). Cells were resuspended in 500 µl of primary antibody polyclonal anti-GFP, ab290 (abcam, Cambridge, UK) diluted 1∶1000 in 1% BSA in PBS, and incubated overnight on a rotisserie at 4°C, washed 5 times in PBS. Secondary antibody (500 µl Alexa Fluor 488 donkey anti-rabbit (Invitrogen, Carlsbad, CA) diluted 1∶2000 in 1% BSA in PBS) was added, samples were incubated 45 minutes in the dark, cells were washed 5 times with PBS, resuspended in 1 ml PBS, and sonicated at 20% duty cycle three times.
Flow cytometry was performed using a FACSCalibur (BD Biosciences, San Jose, CA). Measurements were collected for 100,000 events and analyzed using FlowJo (Ashland, OR). Events were initially examined on a plot of SSC by FSC and gated to include all events (cells) that had measurable FSC and SSC. Mean expression and the Robust CV (100*0.5*(Intensity [at 84.13 percentile] – Intensity [at 15.87 percentile])/Median) of the gated population were collected using cell fluorescence measurements from the FL1 (fluorescein/GFP) channel. These measurements were the basis for further analysis.
Subtelomeric URA3 was inserted in a head-to-head orientation immediately upstream of the TLO promoter (∼600 bp upstream of the TLO start codon) to produce a subtelomeric, TLO-adjacent URA3. Insertion sites were identified by PCR and sequencing as well as separation of chromosomes on contour-clamped homogenous electric field (CHEF) karyotype gens and Southern blotting.
Strains containing a TLO-adjacent URA3 were grown in liquid YPAD and plated for single colonies onto YPAD for no selection of URA3 expression, synthetic complete media (SDC) lacking uracil to select for URA3 expression, and 5-floroorotic acid (5-FOA) to select for URA3 silencing [70]. Five colonies from each condition for three different experiments were assayed for gene expression by qRT-PCR as described above.
|
10.1371/journal.pbio.0060010 | Strong Purifying Selection in Transmission of Mammalian Mitochondrial DNA | There is an intense debate concerning whether selection or demographics has been most important in shaping the sequence variation observed in modern human mitochondrial DNA (mtDNA). Purifying selection is thought to be important in shaping mtDNA sequence evolution, but the strength of this selection has been debated, mainly due to the threshold effect of pathogenic mtDNA mutations and an observed excess of new mtDNA mutations in human population data. We experimentally addressed this issue by studying the maternal transmission of random mtDNA mutations in mtDNA mutator mice expressing a proofreading-deficient mitochondrial DNA polymerase. We report a rapid and strong elimination of nonsynonymous changes in protein-coding genes; the hallmark of purifying selection. There are striking similarities between the mutational patterns in our experimental mouse system and human mtDNA polymorphisms. These data show strong purifying selection against mutations within mtDNA protein-coding genes. To our knowledge, our study presents the first direct experimental observations of the fate of random mtDNA mutations in the mammalian germ line and demonstrates the importance of purifying selection in shaping mitochondrial sequence diversity.
| Mammalian mitochondrial DNA (mtDNA) is maternally transmitted and does not undergo bi-parental recombination in the germ line. This asexual mode of transmission, together with a high rate of mutation, should eventually lead to the accumulation of numerous deleterious mtDNA mutations and a “mutational meltdown” (a phenomenon know as Muller's Ratchet). In this study, we utilized a genetic mouse model, the mtDNA mutator mouse, to introduce random mtDNA mutations, and followed transmission of these mutations. Maternal transmission of mtDNA is typically subjected to a bottleneck phenomenon whereby only a fraction of the mtDNA copies in the germ-cell precursor are amplified to generate the approximately 105 mtDNA copies present in the mature oocyte. As a consequence of this phenomenon, the established maternal mouse lines carried high levels of a few mtDNA mutations. We sequenced the entire mtDNA to characterize the maternally transmitted mutations in the established mouse lines. Surprisingly, mutations causing amino acid changes were strongly underrepresented in comparison with “silent” changes in the protein-coding genes. These results show that mtDNA is subject to strong purifying selection in the maternal germ line. Such selection of functional mtDNA genomes likely involves a mechanism for functional testing to prevent transmission of mutated genomes to the offspring.
| Mammalian mitochondrial DNA (mtDNA) has a high mutation rate and is inherited in a non-Mendelian manner only from the mother [1,2]. Though there are reports of mitochondrial recombination in mammals, it is thought to be quite rare, and it is currently not known whether this phenomenon would be at a sufficient frequency to leave a signature in the population [3–6]. This asexual mode of transmission should leave the mitochondrial genome vulnerable to mutational meltdown by Muller's Ratchet, a process leading to deleterious mutation accumulation in asexual, nonrecombining lineages. The bottleneck phenomenon, which was first proposed after observation of rapid fixation of mitochondrial DNA variants in Holstein cows [7,8], allows for rapid exposure of variant mtDNAs to selection at the level of the individual [9], and may thereby, at the level of the population, protect against mutational meltdown.
The over 100,000 mtDNA molecules in the mammalian oocyte do not undergo replication through the early stages of embryogenesis [2]. Therefore, these maternally derived mtDNA molecules are segregated through cell division events in the developing embryo to generate primordial germ cells with approximately 950–1,550 mtDNA copies [10]. Replication of mtDNA is reinitiated as the primordial germ cells migrate and differentiate to generate oocytes transmitting mtDNA to the next generation [2,11]. The mtDNA bottleneck appears to result from the replication of only a small subset of the mtDNA molecules as the primordial germ cells differentiate to generate oocytes [10].
It has long been thought that animal mtDNA is an essentially neutral marker of sequence evolution [1], but evidence of the selective constraints on mtDNA is accumulating. Studies of animal mtDNA sequence variation within natural populations or in interspecies comparisons consistently show the signatures of negative selection (e.g., [12–14]). For humans, a considerable amount of mtDNA sequence is available from individuals as a result of studies into human evolution and human mtDNA diseases [15,16]. Analyses of the variation in human mtDNA sequences have led to a debate whether random genetic drift (dependent on demographic history), positive selection, or purifying selection is important in the transmission and maintenance of this variation. [17,18]. Consensus is forming that selection is an important part of mtDNA sequence variation in human mtDNA, but the strength and nature of this selection are unresolved [18]. Population-level studies detect signatures of purifying selection in mtDNA sequence variation [19–26], but recent accumulated variation within human populations implies neutrality or weak selection on these variants [25]. Findings from studies of mtDNA mutation inheritance in families with mtDNA-associated disease are compatible with the occurrence of only very weak or no selection on these mtDNA mutations [17,27]. Positive selection facilitated by climatic variation has recently been proposed for human mtDNA [28–30] and would profoundly affect the reliability of mitochondrial molecular clocks and drastically alter our understanding of human divergences.
In an attempt to elucidate the mechanisms of mammalian mtDNA segregation in the germ line, several groups generated transmitochondrial mouse strains carrying two distinct mtDNA sequences (a condition known as heteroplasmy). These transmitochondrial mice are generated by embryo–cytoplast fusions and exhibit germ line segregation patterns explainable by random drift [31–34]. However, tissue-specific segregation patterns within the offspring imply strong nuclear–mitochondrial interactions and suggest that molecular mechanisms exist that could allow for strong selection of mitochondrial variants within offspring [33,35–37]. Unfortunately, the technical complexities of the transmitochondrial technologies have much limited their use by research groups, and so far only a few sequence variants have been investigated.
The mtDNA mutator mice are homozygous for a knock-in allele (PolgAmut/PolgAmut) expressing a proofreading-deficient catalytic subunit of mitochondrial DNA polymerase [38]. These mice have a substantial increase in the levels of mtDNA mutation in all investigated tissues. The somatic mutations generated are evenly distributed along an amplified fragment of the protein-coding mt-CYB gene of mtDNA, and all three codon positions are mutated at equal frequency, though transition mutations were more frequently observed than transversions [38]. In this study, we took advantage of this high mtDNA mutation rate to study the transmission of random mtDNA mutations in the mouse germ line. Female lineages were derived from mtDNA mutator mice by continuous backcrossing, allowing us to isolate, segregate, and characterize germ line mtDNA mutations.
We used eight mtDNA mutator founder females to establish independent maternal lines through 13 F1 females. The breeding scheme used (Figure 1) takes advantage of the bottleneck phenomenon and allowed us to segregate the mtDNA mutations on a wild-type PolgA nuclear background from generation N2 and onwards. Sequencing was conducted from N2 onwards to sample only animals of wild-type PolgA nuclear background, and because levels of individual mutations in mtDNA mutator mice and N1 animals were too low to be detected by the sequencing methods employed.
We sequenced the entire mtDNA of 190 animals from generations N2 to N6 and identified 1,069 unique mutations (Dataset S1). The typical animal carried approximately 30 mtDNA mutations (mean = 29.8 mutations, standard deviation = 9.2). In each line, a large proportion of the identified mtDNA mutations (38.48%) were transmitted to the descendents of that particular N1 female, similar to the propagation of mtDNA haplogroups in human pedigrees. Other mutations were only observed in the siblings of a single litter (17.98%) or in a single mouse (43.54%), but not in their offspring.
Consistent with purifying selection acting on the mtDNA of these mouse lines, synonymous mutations were observed more frequently than nonsynonymous mutations. The ratio of nonsynonymous substitutions per site to synonymous substitutions per site for the protein-coding regions gave a value of 0.6035, signifying purifying selection against amino acid changes in the protein-coding genes (values less than 1.0 signify purifying selection). When mutations that occur only in an individual or a single litter were removed from the dataset, the ratio dropped to 0.4617. The McDonald-Kreitman test of neutral evolution [39] and the accompanying Neutrality Index [40] were calculated for our mtDNA mutator lines, and gave values consistent with excess polymorphisms within our mtDNA mutator lines compared to either Mus musculus molossinus or the NZB mouse strain mtDNA sequences (see Table S1).
We found a strong decrease in the number of mutations at the first and second codon positions of the protein-coding genes when compared to the third codon positions (Figure 2A and Table S2A). This distribution of mutations is a hallmark of purifying selection, because changes in the first and second codon position usually result in an amino acid substitution, whereas many third codon position changes do not. This purifying selection is strong and rapid as the same codon distribution bias is evident in the N2 generation (Figure S1). The observed nucleotide mutational bias in protein-coding genes varied significantly from those observed for the other sites in the mtDNA molecule (chi-square contingency table, p = 0.0023) thus showing differential selection pressures on the protein-coding genes versus other sites (Table 1).
We observed the same selective signature against first and second codon positions when we compared 21 mouse-strain mtDNA sequences obtained from GenBank (Figure 2B) and human mtDNA sequence data obtained from the mtDB database [16] (Figure 2C and Table S2B). There was a similar level of reduction of first codon position mutations in comparison with third codon position mutations in mtDNA mutator lines (2.0-fold reduction; Figure 2A) and humans (2.6-fold reduction; Figure 2C). This striking similarity is surprising because the mtDNA mutator strains have undergone selection for at most six generations, whereas human sequence variation is the consequence of a much larger number of generations to act on these less deleterious substitutions. This illustrates the speed and strength of the selection on the mtDNA and its importance in sculpting modern mtDNA variation in natural populations. It also demonstrates this experimental model can be a powerful tool in investigation of mtDNA evolution. The smaller number of nonsynonymous changes in the mouse strains (Figure 2B) in comparison with mtDNA mutator lines (Figure 2A) can probably be explained by the limited sampling of only one individual from each of the 21 different mouse strains. In addition, it should be emphasized that the mtDNA mutator mice have been exposed to the effects of purifying selection for only a few generations.
We further investigated the observed selection on protein-coding genes in the mtDNA mutator lines by separating the mutations at 4-fold degenerate sites (third codon positions for amino acids L2, V, A, T, P, S1, R, and G) from all other protein-coding mutations. The 4-fold degenerate sites can mutate to any nucleotide without changing the encoded amino acid and should therefore be subject to less selective constraint than other protein-coding sites. Expected values were calculated based on an assumption of an equal distribution of the observed mutations of these two classes, across the genes and corrected for their coding size. The ratios between observed and expected mutation frequencies at 4-fold degenerate sites were approximately equal in all of the protein-coding genes except for mt-ND2 and mt-ATP8 (Figure 3A and Table S3A). In contrast, mutations at the non–4-fold degenerate sites deviated profoundly from the ratios predicted by equal distribution of mutations (Figure 3A). Contingency table analysis was carried out to detect changes in the proportion of 4-fold degenerate site mutations to other sites within the protein-coding genes. The 13 protein-coding subunits were grouped by the oxidative phosphorylation (OXPHOS) enzyme complex to which they belong. Only complexes III (containing mt-CYTB) and complex IV (containing mt-CO1–3) showed statistically significant changes in the ratio of 4-fold to non–4-fold sites (see Table 2). After correcting for multiple tests, only the complex IV data remained significant. Interestingly, the mt-ATP8 and mt-ATP6 subunits appear to allow for excess changes at all sites relative to the expected values, though the ratio of 4-fold to non–4-fold sites did not vary significantly (Figure 3A). Analyses of human mtDNA sequences have shown a similar occurrence of excess sequence variation in the mt-ATP8 and mt-ATP6 genes, particularly evident for the mt-ATP6 gene [23–25,28]. These previous reports lead us to investigate available human sequence data, and we found a strong selection against non–4-fold degenerate changes in mt-CO1 versus the weaker selection in mt-ATP6, mt-ATP8, and mt-CYTB (Figure 3B and Table S3B). Observed mutations for each protein gene for mtDNA mutator lines and human population showed the same variation from expected in 11 of 13 cases (Figure 3C). Thus, sequence variation in protein-coding genes of this experimental mouse model demonstrates similar patterns to those seen in human populations.
In contrast to the patterns found in the protein-coding genes, we found higher levels of mutations in tRNA and rRNA genes in the mtDNA mutator lines (Figure 4A) in comparison to the levels in mouse strains and humans (Figure 4B and 4C). There are several observations from human mtDNA disease that imply that tRNA genes may be subject to a less rapid form of purifying selection than that observed for the protein-coding genes of our mtDNA mutator mouse lines: (1) Population-level sampling has revealed an increase in recent mtDNA sequence variation within tRNA and rRNA genes in humans [30,41]; (2) 58.2% of the known pathogenic human mtDNA mutations are located in the tRNA genes, although these genes only occupy 9.1% of the genome [42], implying that these changes are, at low levels of heteroplasmy, more compatible with life than some protein-coding mutations; and (3) disease-causing tRNA gene mutations reach high heteroplasmic levels, or sometimes must be homoplasmic, before the onset of disease [43,44]. A less acute, but equally important, mechanism of purifying selection appears to be acting on tRNA genes. This may explain why the corresponding mutations are not removed as rapidly from the mtDNA mutator mouse lines.
A very low mutation rate was observed for the control region of mtDNA mutator strains (Figure 4A), despite the fact that the control region sequences are normally the most variable regions in mtDNAs. A reduction in the number of mutations within the control region was also reported in the somatic tissues of the mtDNA mutator mice [38], though a mechanism to explain this observation is still elusive.
We present experimental evidence for strong purifying selection against nonsynonymous mutations in protein-coding genes during maternal transmission of mutated mtDNA in the mouse. The drastic reduction of mutations in the amino acid changing first and second codon positions of protein-coding genes are a direct result of purifying selection against deleterious mtDNA mutations at some stage within the reproductive cycle of these mice. This bias occurs rapidly and is evident as early as the N2 generation. These findings have profound implications for our understanding of how mutated mtDNA is transmitted between generations. It is important to recognize that this strong purifying selection against nonsynonymous changes that we observe is likely to be a universal phenomenon in mammals, but the rapid nature of this selective force would render these mutations difficult to detect in population studies. These findings have profound implications for our understanding of how mutated mtDNA is transmitted between generations.
Within studies of human mtDNA evolution, the observation is that many substitutions are not ancient changes shared deep within human haplogroups, but rather are new variants clustered within the tips of phylogenetic networks and found only in a small number of individuals. This implies they are mildly deleterious variants not yet selected against [23,26,30]. Studies of disease-causing mtDNA mutations show they are often heteroplasmic, and can be present at high levels without consequence for the carrier. However, once the levels exceed a specific threshold, the respiratory chain function will be impaired, causing a clinical phenotype [11]. Based on these observations, when using the mtDNA mutator mouse to study germ line transmission of mtDNA mutations, one could expect to observe the inheritance of high numbers of mutations at all sites in the early generations, which would eventually be removed from the mouse lines once their phenotypic thresholds had been crossed. Whereas the inheritance of the tRNA, rRNA, and third codon position mutations appear to be following this expected pattern (see Figures 2 and 4), this is not the behaviour of mutations at the nonsynonymous first and second codon positions in our mouse lines (Figure 2).
The strongest signature of purifying selection can be observed within mt-CO1 and mt-CO2, consistent with the very high levels of sequence conservation in these genes. The strength and speed of this purifying selection could have other effects on the mutation patterns observed in our model. The consensus view is that bi-parental recombination of mammalian mtDNA is at most extremely rare [3–6] and therefore selection acting at any one site in the mtDNA will affect the entire mtDNA molecule. The observed strong and rapid selection of mtDNA mutations could therefore also reduce the number of neutral variants observed, due to their linkage to deleterious mutations. This means that 4-fold degenerate sites or even noncoding mutations might not be the reliable measure of the mitochondrial neutral mutation rate. Such an underestimation of the mtDNA mutation rate using phylogenetic or population methods relative to pedigree-based observation has been reported previously [45–47]. If this is the case, the models based on this assumption require recalibration.
This point is also important in interpreting the excess change observed for the mt-CYB, mt-ATP6, and mt-ATP8 genes in mtDNA mutator lines. Similar gene-specific increases of mutations have been reported in human mtDNA, especially in mt-ATP6 [23–25,28,29]. Though some argue that this signifies positive selection, the pattern may also be due to less-intense purifying selection on these specific genes. If mutations at mt-ATP6 experience less-selective constraint, mutations at these sites will be allowed to accumulate and persist in the mtDNA pool. Meanwhile, mutations at strongly selected sites, such as mt-CO1 and mt-CO2, are eliminated, leading to the relative increase in the observed frequency of mt-ATP6 and mt-ATP8 mutations in our model organisms.
In contrast to the rapid selection against nonsynonymous changes, rRNA and tRNA genes experienced less-intense purifying selection in our mtDNA mutator lines. Though tRNA genes also have high levels of sequence conservation, the frequency of observed mutations at these sites in our mouse lines was quite similar to the rate observed at third codon positions (Figures 2A and 4A). Some of the identified tRNA mutations, e.g., the deletion of one base in the anticodon loop of mt-TM (3873delC mutation) can be predicted to have a biochemical effect if present at high levels. Previous models have mainly been based on observational studies of transmission of mutated mtDNA in human pedigrees affected by mitochondrial disease. Such threshold-mediated protection from selection should lead to slower purifying selection of the mtDNA variant, which may be reflected in the essentially neutral segregation patterns observed for disease-causing mutations prior to clinical manifestation [11,17,27,48].
It is plausible that these tRNAs, as well as a number of the nonsynonymous changes in the protein-coding genes in our model system, may eventually behave like human mtDNA disease mutations in that these mutations are transmitted and cause no obvious phenotype at low levels, but may be selected against or cause a disease-like phenotype at higher levels. We will continue sampling our lines to investigate the long-term fate of the observed tRNA gene mutations, as well as the stably transmitted nonsynonymous protein-coding gene changes.
In the mtDNA mutator mice, the mutations within protein-coding genes are equally distributed across all three codon positions [38,49], whereas the pattern of mutation accumulation is different in the mtDNA mutator lines. It has previously been proposed that mitochondrial fitness may be selected for during oocyte development [50], and it is therefore quite possible that mtDNA in germ cells is under a different selective regime than the mtDNA in somatic cells. There is a massive proliferation of mtDNA during oogenesis, whereby a small number of mtDNA copies in the primordial germ cells are extensively amplified to generate the approximately 105 mtDNA copies in the mature oocyte [2,50]. This mechanism provides ample opportunities for functional testing of mtDNA during female germ-cell development, and future research is required to unravel molecular mechanisms responsible for this selection.
Our experimental strategy has allowed us to look at the fate of a broad spectrum of mtDNA variation, and we report evidence of strong purifying selection in the mouse female germ line. All of the generated mtDNA mutator mouse lines showed the same strong reduction in nonsynonymous substitutions, exemplified by the reduction in first and second codon position mutations. This pattern is also seen in human populations and implies that purifying selection has a similar, drastic impact on the mtDNA variation in humans despite different demographic histories. The RNA genes, in contrast, appear to accumulate at levels approximating the synonymous third codon positions. These mutations are expected to eventually raise to high-enough levels and lead to impaired mitochondrial function in a manner similar to the threshold effect seen in human mtDNA disease. The data generated from this experimental model will allow us to build more accurate molecular models of mtDNA evolution and aid the understanding of inheritance patterns of human mtDNA disease mutations.
Heterozygous knock-in male mice (PolgAmut/+) [38] were crossed to C57Bl/6NCrl females (Charles River Laboratories). Resulting heterozygous mice were intercrossed to obtain female homozygous knock-in mice (PolgAmut/PolgAmut). We performed crosses of mtDNA mutator females to wild type C57Bl/6 males to produce N1 females. Maternal mouse lines were then established by successive backcrossing of females to C57Bl/6 males (Figure 1). The genotype at the PolgA locus was determined as described previously [38], and from the N2 generation onwards, only mice homozygous for the wild-type PolgA-allele were used in the study. In two cases, we bred heterozygous knock-in females from the N2 generation because of small litter sizes. In these two lines, all animals were homozygous for the wild-type PolgA-allele from generation N3 and onwards.
This animal study was approved by the animal welfare ethics committee and performed in compliance with Swedish law.
Pups were weaned at 21–25 d, and tissue from an ear punch was used for DNA isolation, as previously described [51], except that the DNA was purified by phenol:chloroform:isoamyl alcohol (25:24:1) and chloroform extraction, followed by sodium acetate salt and ethanol precipitation. DNA was dissolved in an appropriate volume of deionised water.
The mtDNA genome of each animal was amplified in 29 overlapping PCR reactions (Table S4). All PCR primers contain 5′ M13F or M13R sequence to use as sequencing primers. PCR samples were cleaned using the Agencourt AMPure PCR purification and directly sequenced in both directions using BigDye version 3.1 sequencing kit (ABI). Cycle sequencing reactions were cleaned by precipitation (ABI protocol) or by the Agencourt CleanSEQ Dye-terminator removal kit.
The sequence reactions were analysed on a 3130xl capillary sequencer (ABI), and assembled and analysed for heteroplasmies and substitutions using SeqScape V2.5 software (ABI) and compared to our C57Bl/6 mtDNA reference sequence. The software identified heteroplasmy or substitutions at ≥25% signal intensity on both strands of DNA sequence. All heteroplasmies were confirmed by eye. All mutation sites are potentially heteroplasmic in this analysis, due to these detection thresholds in sequencing technology [52]. The wild-type C57Bl/6 mtDNA sequence used in this study varied from that presented in GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession number NC_005089.1 in two sites; position 9,461 was C (synonymous change at amino acid 1 of mt-ND3) and position 11,515 was A (S450N, nonsynonymous change in mt-ND4). Two poly-A tracts in the mouse genome could not be reliably scored for insertion and deletions due to sequencing complications. We therefore ignored insertions/deletions at positions 5,160–5,191 (origin of light strand replication) and 9,821–9,830 (polymorphic region of mt-tR).
The following mouse sequences were used for comparison of strain variation, GenBank accession numbers: AB042432, AB042523, AB042524, AB042809, AB049357, AJ489607, AJ512208, AY339599, AY466499, AY533105, AY533106, AY533107, AY533108, AY675564, AY999076, DQ106412, DQ106413, J01420, L07095, L07096, and V00711. Sequences were aligned to the wild-type C57Bl/6 mtDNA sequence, and all variations were scored.
Human mtDNA sequence data were obtained from the mtDB human database (http://www.genpat.uu.se/mtDB/; accessed February 2007) [16]. Variants were classified as all observed sequence changes from the most prominently observed nucleotide at the given position in the database.
Codon usage for our C57Bl/6 line was calculated using CodonW version 1.3 (John Peden, http://codonw.sourceforge.net//). The codon usage was used to calculate the number of synonymous and nonsynonymous substitution sites in the C57Bl/6 mtDNA genome.
The number of 4-fold degenerate sites versus other sites for the protein-coding genes was derived from codon usage calculations on each protein-coding gene. Expect values (Figure 3A and 3B) made the assumption of a random distribution of observed mutations across the mtDNA molecule. The genome total of observed mutations within each class was multiplied by the proportion of those sites encoded for each gene. Observed mutations are reported as the number of mutations for that gene detected, multiplied by the proportion of sites for that class within that gene.
Comparison of nucleotide biases was conducted using a contingency analysis on the data supplied in Table 1. The analysis was performed on the eight substitution classes, where all observed values were greater than six. The variation was observed to be significant (chi-square test, p = 0.0023, 7 df).
Comparisons of the 4-fold versus non–4-fold mutations for each mitochondrial OXPHOS enzyme complex were carried out using the Fisher exact test on a contingency table comparing each enzyme subunit to the sum of mutations for the remaining complexes. The Bonferroni correction for multiple testing of the data was applied when determining significance.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the mouse sequences used for comparison of strain variation to the wild-type C57Bl/6 mtDNA sequence (DQ106412) are as follows: AB042432, AB042523, AB042524, AB042809, AB049357, AJ489607, AJ512208, AY339599, AY466499, AY533105, AY533106, AY533107, AY533108, AY675564, AY999076, DQ106413, J01420, L07095, L07096, and V00711.
|
10.1371/journal.pntd.0000784 | The Echinococcus granulosus Antigen B Gene Family Comprises at Least 10 Unique Genes in Five Subclasses Which Are Differentially Expressed | Antigen B (EgAgB) is a major protein produced by the metacestode cyst of Echinococcus granulosus, the causative agent of cystic hydatid disease. This protein has been shown to play an important role in modulating host immune responses, although its precise biological function still remains unknown. It is generally accepted that EgAgB is comprised of a gene family of five subfamilies which are highly polymorphic, but the actual number of genes present is unknown.
Based on published sequences for the family, we designed specific primers for each subfamily and used PCR to amplify them from genomic DNA isolated from individual mature adult worms (MAW) taken from an experimentally infected dog in China and individual larval protoscoleces (PSC) excised from a single hydatid cyst taken from an Australian kangaroo. We then used real-time PCR to measure expression of each of the genes comprising the five EgAgB subfamilies in all life-cycle stages including the oncosphere (ONC).
Based on sequence alignment analysis, we found that the EgAgB gene family comprises at least ten unique genes. Each of the genes was identical in both larval and adult E. granulosus isolates collected from two geographical areas (different continents). DNA alignment comparisons with EgAgB sequences deposited in GenBank databases showed that each gene in the gene family is highly conserved within E. granulosus, which contradicts previous studies claiming significant variation and polymorphism in EgAgB. Quantitative PCR analysis revealed that the genes were differentially expressed in different life-cycle stages of E. granulosus with EgAgB3 expressed predominantly in all stages. These findings are fundamental for determining the expression and the biological function of antigen B.
| Antigen B (EgAgB) is a major protein produced by the metacestode cyst of Echinococcus granulosus and plays an important role in modulating host immune responses, although its precise biological function still remains unknown. Previous studies suggested the EgAgB gene family is variable between isolates and genotypic strains of E. granulosus. We designed specific primers to amplify and determine the number and variation of the genes using genomic DNA from individual worms. Based on sequence alignment analysis, we found that the gene family comprises ten unique genes. Each of the genes was identical in both larval and adult E. granulosus and in isolates collected from the two distinct geographical areas. We showed that the genes were differentially expressed in different stages of E. granulosus with one gene, EgAgB3/1, expressed predominantly in all stages. This is the first study to report such a large number of unique and conserved genes in the EgAgB gene family and their differential expression in different life cycle stages of E. granulosus. These findings are fundamental for determining the expression and regulation of this gene family in E. granulosus and the biological function of antigen B.
| Antigen B (EgAgB) is the most abundant protein generated by the pathogenic larval stage (hydatid cyst or metacestode) of Echinococcus granulosus, the cause of cystic echinococcosis (CE). Synthesized and secreted by both cyst germinal layer and protoscoleces [1], the protein is highly immunogenic and can be recognised by more than 80% of sera from patients with CE [2], [3]. Nevertheless, its precise biological function remains undetermined, although one report suggests that EgAgB might have lipid-binding properties [4]. It has been as well hypothesised that EgAgB plays a key role in the interaction between parasite and host based on studies showing it functions as a serine protease inhibitor that impairs neutrophil chemotaxis [5] and as an immune modulator that skews Th1/Th2 cytokine ratios to Th2 polarized responses [6], benefiting parasite survival in the mammalian host [7]. A number of previous studies have also indicated that the protein is encoded by a gene family [8], that is highly variable between isolates and strains of E. granulosus [5], [8]–[10].
We believe the high levels of variation reported by others was based on comparisons of paralogs, amplified using conserved primers and assumed to be orthologs. Until now, there have been no data showing how many genes are represented in the EgAgB family, although it is known that there are five subfamilies (EgAgB1-5) present [5], [8], [9], [11]–[13]. Genomic Southern blots revealed that the gene family should include at least seven genes [14]. However, as these genes are highly similar, especially at the subfamily level, it has proven difficult to generate clear data from the Southern blot analysis. Determining the number of the genes in the family is fundamental for further exploring the expression and regulation of EgAgB in E. granulosus. This will provide insight to more fully understanding its biological function in this and other taeniid species, which share similar gene sequences to those found in E. granulosus [15]–[18].
We cloned and sequenced ten unique genes from individual worms (adults and protoscoleces) of E. granulosus and show that each is conserved in parasites originating from different geographical areas and hosts. Further, we show the differential expression of all of the family of genes in five developmental stages of E. granulosus by real time PCR and cDNA sequencing.
Protoscoleces (PSC) of E. granulosus were aspirated from a fertile hydatid cyst collected from a kangaroo (Macropus giganteus) from eastern Australia. The cyst was kindly provided by Dr. Tamsin Barnes from a previous study [19]. Mature adult worms (MAW) were collected from a dog from Xinjiang, China [20]. The parasite materials were stored until use in liquid nitrogen as described [20].
PSC and MAW were thawed in RNAlater (Ambion, Austin, USA) and diluted with water. Individual PSC and MAW were respectively pipetted into plastic mortar microtubes (Sigma–Aldrich, St. Louis, USA) under microscopy to make sure that each tube contained a single parasite. After a brief centrifugation to spin-down the parasite, 50 µl of PrepMan Ultra Sample Preparation Reagent (Applied Biosystems, Foster, USA) was added to each of the tubes. The single parasite was ground with a micro-grinder using a plastic pestle. The homogenate was heated at 100°C for 10 min and centrifuged at 16,000 g for 5 min. The supernatant was precipitated with 1× vol of isopropanol. The invisible pellet was washed with 70% (v/v) ethanol, dissolved in 50 µl water and used as DNA template.
PCR reactions were performed with a Taq polymerase kit (Promega, Madison, WI) with 5 µl of the DNA template preparation and 20 pmol of each PCR primer in a final volume of 50 µl. To amplify the EgAgB gene fragments from genomic DNA, we designed two forward primers, EgAgBF1, specific for subfamily EgAgB1 and EgAgB3, and EgAgBF2, specific for subfamily EgAgB2, EgAgB4 and EgAgB5 based on previous studies [8], [9], [11],[12],[21]. The forward primers were based on the first exonic sequences of the EgAgB gene family. We designed eight down-stream primers, which were specific for each of the gene subfamilies (the primers for EgAgB1 and EgAgB3 were within the second exons). All primers used to isolate the EgAgB gene variants are listed in Table 1. Amplification was performed with 35 cycles of 94°C for 30 s, 54°C for 30 s and 72°C for 30 s, followed by a denaturing step at 94°C for 1 min, and a final extension step at 72°C for 7 min on a Mastercycle Gradient thermocycler (Eppendorf, Hamburg, Germany).
PCR products were purified using PCR Purification Kits (Qiagen, Hilden, Germany). Fifty ng of the PCR products were ligated with 50 ng of pGEM-T vector (Promega) in a final volume of 20 µl according to the manufacturer's instructions. One microlitre of the ligation reaction was used to transform 20 µl of competent E. coli strain JM109 cells (Promega). White colonies containing inserts were selected on LB agar plates containing ampicillin and X-gal. As each pair of primers is specific to each subfamily, and may amplify gene fragments with different sized PCR products, a quick plasmid extraction/PCR step was performed to determine the size of inserts before selecting clones for sequencing. In brief, after the white colonies had grown to about 0.5 mm in diameter, 30–50 of these colonies from each transformation were individually transferred to microtubes containing 50 µl of water. After vortexing and centrifugation at 12,000 g for 1 min, 10 µl of the supernatant from each tube was used as DNA template for PCR using the same original primers. For each transformation, 3–10 colonies with the same sized insert were selected for sequencing, performed using a Big-Dye Version 3.0 kit on an ABI 377 sequencer (Applied Biosystems) after purification with QIAprep Spin Miniprep Kits (Qiagen).
Inspection of the amino-acid sequences inferred from data collected during this study and obtained from the public databases showed that some members of the EgAgB subfamily could be aligned with ease. However, sequences from other subfamilies of EgAgB and sequences from other cestodes proved more difficult to align. Furthermore any alignment would be short: 54 amino acids being the length of the shortest sequence. However, to produce a graphical representation of the data, we constructed a simple phylogenetic tree to show the different clusters clearly, including relationships among members of each subfamily. We accept that such a tree does not provide robust inference for the deeper nodes. Bioedit (http://www.mbio.ncsu.edu/BioEdit/bioedit.html) was used to align sequences. Molecular Evolutionary Genetics Analysis version 4 (MEGA v4) [22] program (http://www.megasoftware.net/) was used to construct the tree from amino acid sequences translated from the second exonic sequences of EgAgB amplified and cloned from E. granulosus in this study and homologous protein sequences from other cestode parasites deposited in the GenBank, EMBL and DDBJ databases, after removal of the signal peptides at their N terminal. A distance matrix was constructed using a Poisson correction method before a mid-point rooted tree was constructed by the minimum-evolution method. One thousand bootstrap cycles were used.
We used quantitative PCR to determine the expression level of each of the EgAgB family of genes in five life cycle stages/structural compartments of the cyst of E. granulosus. These were: protoscolex (PSC), cyst germinal membrane (CM), immature adult worm (IAW), mature adult worm (MAW) and oncosphere (ONC). Sheep livers containing hydatid cysts were collected from a slaughterhouse in Urumqi, Xinjiang, China. The inner parasite cyst membrane was carefully released from the outer host capsule under sterile conditions. PSC and brood capsules containing PSC were aspirated and then treated with 1% (w/v) pepsin in saline, pH 3 [23], to remove capsule membranes and immature PSC. After 3 washes, the precipitated PSC were stored in liquid nitrogen until use. To prepare the CM, the inner cyst membrane was rinsed several times with PBS to remove any remaining PSC, and the membrane was divided into small pieces. These were pooled and stirred at 4°C for 30 min to release the germinal layer from the laminated layer. After leaving 1 min on ice, the laminated membranes were precipitated and the supernatant transferred to a fresh tube. After centrifugation at 3000 g at 4°C for 15 min, the pellet (CM) contained mostly germinal cells that were stored in liquid nitrogen until use. IAW and MAW (from dogs infected with sheep PSC) and activated ONC were prepared as described [20], [24].
Total RNA was extracted from the different stages/compartments of E. granulosus using TRIzol reagent (Invitrogen, Carlsbad, CA, USA), according to the supplier's instructions. The RNA was treated with DNase I (Promega) to remove possible genomic DNA contamination. All the RNA samples were of high quality (A260/A280 nm>1.8 and <2.0 in nuclease-free water) assessed using a Bioanalyzer RNA Pico LabChip (Bioanalyer). First-strand cDNA synthesis was carried out with oligo (dT) 12–18 using a Superscript Reverse Transcription kit (Qiagen) with 45 ng of total RNA, according to the manufacturer's instructions. For real time PCR, all cDNA samples were diluted to a concentration of10 ng/µl. Subsequently, 5 µl aliquots were combined with 10 µl of SYBR Green, 3 µl of water and 2 µl (5 pmol) of the forward and reverse primers listed in Table S1. Each experiment was performed in triplicate. Expression profiles of EgAgB1-5 in the different stages/compartments were obtained by real time PCR using a Rotor Gene (6000) real time PCR machine (Qiagen) and data were analysed by Rotor Gene 6 Software. To identify the expression profile of EgAgB3, we used a pair of primers, EgAgBF1 and EgAgB3R (Table 1), to amplify cDNAs obtained by reverse transcription from total RNA isolated from the five E. granulosus stages/compartments. The resulting PCR products were ligated into pGEM-T (Promega) and then transformed into E. coli strain JM109. We randomly selected 30 colonies from each of the transformations for sequencing.
With the eight combinations of primers shown in Table 1, we successfully amplified gene fragments with genomic DNA extracted from six individual MAW isolated from a dog (from China) and five individual PSC isolated from a single cyst from a kangaroo (from Australia). Fig. 1 shows representative examples of the amplified bands from one MAW (ZGA2) and one PSC (ZGP5). The sizes of the PCR products matched the predicted sizes (315 to 387 bp) (Table 1). In total, we generated 435 clones with validated sequences including 234 from MAW and 201 from PSC. Alignment of all the sequences showed ten clusters (data not shown) representing ten genes. Figs. 2–4 show alignment s of intronic, exonic and amino acid sequences of ten gene representatives isolated from MAW ZGA2 and PSC ZGP5, respectively. The terminology for each subfamily follows previous studies [5], [8], [9], . Each pair specific to EgAgB1, 2 and 5 generated only one sequence cluster, respectively, indicating only one gene in the three subfamilies, comprising subfamily 1 (EgAgB1/1; accession numbers HM237302 (PSC) and GU166202 (MAW)), subfamily 2 (EgAgB2/1; accession numbers GU166200 (PSC) and GU166201 (MAW)) and subfamily 5 (EgAgB5/1; accession numbers GU166215 (PSC) and GU166216) (MAW)). In contrast, primers specific to subfamily 3 amplified four genes in the subfamily (EgAgB3/1–4, accession numbers GU166204-GU166214) whilst primers specific for EgAgB 4 generated three genes in the subfamily (EgAgB4/1–3, accession numbers GU166196-GU166199). Almost all the sequences in each gene cluster were identical to the EgAgB sequences deposited in GenBank (Figs. 2–4), which were obtained from isolates of E. granulosus from different geographical areas. Table 2 shows a comparison of each of the EgAgB DNA (intron and second exon) sequences and amino acid (aa) sequences encoded by the second exon, which are likely to be the mature and secreted proteins comprising 65–71 residues in length (Table S2). The degree of identity between the EgAgB protein family varies from 26.3% to 97.1%; the DNA sequences vary from 19% to 91% (Table 2). The lowest aa similarity occurred between EgAgB4/2 and EgAgB5/1. Although EgAgB3/1 has the highest identity (97.1%) with EgAgB3/2 at the aa level, the difference in their intronic sequences showed that they are different genes with a DNA identity of 57.1%.
The major differences between the EgAgB genes appear in their introns (Fig. 2) and the second exons (Fig. 3) which encode different protein sequences (Fig. 4). The intronic sequences can be used for distinguishing all subfamilies and four genes in the EgAgB3 subfamily as they have different sizes and variable sequence (Fig. 2). Based on alignment analysis with sequences from GenBank, EgAgB1 has two clusters of intronic and exon sequences shown in Fig. S1. They are likely to be encoded by different alleles. However, in our study, only one unique sequence (EgAgB1/1) was amplified from individual worms and it has 89 bp of intronic sequence. The second exonic sequence comprises 198 bp (Fig. 3) encoding 65 aa (Table S2 and Fig. 4).
Cluster analysis of 99 cloned fragments of EgAgB2 with intronic and the second exonic sequences isolated from individual PSC and MAW (30 sequences aligned in Fig. S2) showed that the subfamily EgAgB2 comprises only one gene cluster, indicating there is only one gene in the EgAgB2 subfamily. The intron is 68 bp in length and the second exon is composed of 213 bp encoding 70 aa (Table S2).
We designed a pair of primers to amplify the EgAgB3 gene subfamily by PCR from genomic DNA. Based on the size of inserts in clones and subsequent sequence analysis, we isolated four clusters of fragments representing four genes in the subfamily. EgAgB3/1, EgAgB3/2, EgAgb3/3 and EgAgB3/4 had introns of 137 bp, 140 bp, 152 bp and 140 bp respectively (Fig. 2 and Table S2). Although EgAgB3/2 had the same sized intron (140 bp) as EgAgB3/4, there were 26 substitutions between the two sequences. The second exonic sequences of the AgB3 subfamily also exhibited four types of sequences matching the intronic differences (Figs. 2–4). The amplified second exons of EgAgB3/1 and EgAgB3/3 encode 54 aa, but they are distinguishable from each other by differences in their intronic sequences of 137 and 152 bp, respectively (Fig. 2 and Table S2). In addition, there are eight aa substitutions in EgAgB3/1 compared with EgAgB3/3 (Fig. 4). The amplified regions of EgAgB3/2 and EgAgB3/4 encode 55 aa and 53 aa, respectively. The major difference in protein sequence encoded by the EgAgB3 subfamily occurs in the region immediately linked to the signal peptide, which is a region rich in aspartic acid (D). EgAgB3/1 has 5Ds, EgAgB3/2 has 6Ds, EgAgB3/3 has 3Ds and EgAgB3/4 has 4Ds. Highly conserved sequences were found in the remainder of the second exonic sequences (Figs. 3, 4).
We designed four primers based on the 3′ terminal sequences of the EgAgB4 subfamily including one for the 3′ UTR sequence. Combined with forward primer, EgAgBF2, the four pairs of primers allowed us to amplify three clusters of sequences from individual MAW, indicating that there are three genes (EgAgB4/1–3) present in the subfamily. EgAgB4 is very similar to EgAgB2 both in intronic and exonic sequence (Figs. 2, 3). The two subfamilies have the same sized 68 bp intron but there are ten nucleotide substitution differences in their intronic sequences (Fig. 2). In addition, there are 14–17 bp differences in the second exon of EgAgB4 compared with EgAgB2 (Fig. 3), resulting in 17 aa changes at the protein level (Fig. 4). The second exon of EgAg4/3 is composed of 216 bp encoding 71 aa, while the second exons of EgAgB 4/1 and EgAgB4/2 encode 70 aa and 69 aa, respectively (Fig. 4). The intronic sequences of EgAgB4 are identical. A major difference among the subfamily of genes is in their 3′ terminal exonic sequences, encoding different aa sequences rich in glutamic acid (E) and D residues (Fig. 4).
EgAgB5 is a unique gene consisting of an intron of 67 bp and its second exon encodes a peptide of 66 aa (Fig. 4). Its DNA sequence is considerably different from those of the other EgAgB subfamily members (Table 2 and Figs. 2, 3). Consequently, the protein sequence of EgAgB5/1 has the lowest identity to the other proteins (Table 2).
We used MEGA methods for phylogenetic analysis of the inferred amino acid sequence of the EgAgB family of proteins to illustrate the evolutionary relationships within the family and, particularly, with those present in species from the confamilial genus Taenia. We confirmed the results with Bayesian analysis (Mr Bayes 3.1) [25] (data not shown) and the two methods showed a very similar evolutionary pattern. The minimum evolution tree (Fig. 5) has very low bootstrap values for deeper nodes, as anticipated because of dissimilarities between sequences from different subfamilies, and especially different species. The “Taeniidae antigens,” [26], commonly found in taeniid cestodes (and one example from Hymenolepis diminuta) form an outgroup in this mid-point rooted tree. All sequences from the genus Echinococcus, including sequences from E. granulosus (EgAgB), E. multilocularis, E. vogeli, E. oligarthrus, E. ortleppi and E. canadensis form a monophyletic clade (Fig. 5). This implies that these genes have radiated in the Echinococcus lineage after separation from the other taeniids.
For Echinococcus, the majority of the protein clusters include representative sequences from several species (Fig. 5), indicating the encoding genes were likely present in the most recent common ancestor of the genus suggesting the antigen B family has been important in its evolution.
It is important to note that we treated all RNA preparations for analysis with RNase-free DNase prior to reverse transcription. To determine whether the RNA samples contained DNA after treatment, we added a PCR control that comprised the cDNA synthesis reaction comprising all components but without the addition of reverse transcriptase (RT). Both normal RT and real-time PCR analysis showed there were no amplicons generated from these control samples (data not shown).
For normalizing the real time PCR data, we initially used actin II as a house-keeping gene to profile gene expression in the different stages of E. granulosus. However, as actin II was shown to be significantly up regulated in MAW and variable in the other stages, we used an eukaryotic translation initiation factor (Eg-eif) of E. granulosus as an alternative house-keeping gene, which was identified by microarray analysis and confirmed by real-time PCR and normal reverse transcription PCR analysis (data not shown). Figure 6 shows the results of the expression levels of 5 subfamilies of the EgAgB genes and actin II after normalization using Eg-eif in the 5 E. granulosus stages and a pooled mixture of the 5 stages as a PCR control with different combinations of primers. EgAgB1, EgAgB2 and EgAgB5 were expressed at very low levels in all stages. EgAgB3 was expressed in all stages of the parasite, with the highest in IAW and MAW. Except for EgAgB3, the EgAgB genes were almost undetectable in PSC and ONC. EgAgB4 was expressed in CM, IAW and MAW, but at a low level. It is worth noting that EgAgB3 was highly expressed in MAW (3–10 times higher than in the other stages), suggesting this gene subfamily may play a role in worm development in the gut of the definitive host. We used EgAgBF1 (Table 1) and EgAgB3R (Table S1) sequences as universal primers to amplify cDNA which showed (Table 3) that EgAgB3/1 was the most highly expressed gene in all stages, followed by EgAgB3/2. EgAgB3/3 and EgAgB3/4 were lowly expressed.
All the genes in the E. granulosus antigen B (EgAgB) gene family have a similar gene structure with one intron flanked by two exons [27]. Furthermore, the first exonic sequence of EgAgB encodes a signal peptide. We did analysis of all EgAgB sequences deposited in the GenBank databases and showed that the sequences in this region are highly conserved (data not shown) with two clusters. This allowed us to design two forward primers, one for subfamily 1 and 3, and another for subfamily 2, 4 and 5 (Table 1) in the first exonic region of the gene family. The variable sequences occur at the 3′ terminal ends. Consequently, we designed eight downstream primers specific to the 3′ terminal sequences to cover all possible genes in the five recognised gene subfamilies. Primer EgAgB24UTR (Table 1) was designed based on the identical sequences of the 3′ terminal UTRs of subfamilies EgAgB2 and EgAgB4, which allowed us to amplify the entire second exonic sequences in the subfamilies. With the designed primers, the PCR amplified fragments therefore contained both the intronic and the second exonic sequences of genes in the EgAgB family. Since eight pairs of primers were used to amplify genomic DNA from 11 MAW/PSC, instead of using random selection of clones for direct sequencing, we used a new strategy (described in detail in the Methods and Materials section) to select clones for sequencing. With this selection strategy, we chose 3–9 clones from each transformation for further sequencing. This strategy minimized the number of clones for sequencing and covered all possible EgAgB sequences. In total, we generated 435 clones with sequence information, which represents the largest reported number of EgAgB gene family sequences amplified from genomic DNA isolated from individual E. granulosus MAW and PSC. We isolated genomic DNA from individual PSC collected from a single hydatid cyst obtained from an infected kangaroo. The PSC clones allowed us to determine whether any apparent gene variation was caused by a different gene or by a mutation. As the PSC were collected from a single hydatid cyst, their genomic DNA should be identical [28], and, indeed, we showed the sequences for each gene were indistinguishable. Two conclusions resulted from this sequence analysis: 1). E. granulosus genomic DNA contains at least ten genes comprising the EgAgB family; and 2). each of the genes is highly conserved. We isolated all ten genes from each of six MAW. The MAW were collected from a dog experimentally infected with pooled PSC originating from a number of hydatid cysts obtained from three individual sheep. The worms could, therefore represent different genotypes, but the sequence analysis showed that each gene was identical, confirming, therefore, the conservation of each gene in the EgAgB gene family, which was further supported by alignment with sequences deposited in the GenBank databases (Fig. S3).
In addition, we showed that each of the ten EgAgB genes was identical in isolates collected from two distinct geographical areas, China and Australia. Macropods have only recently acquired E. granulosus as the parasite is believed to have been introduced into Australia by European immigrants about 200 years ago [29]. The conservation in sequence of the EgAgB genes isolated from a recently acquired new intermediate host, this case, a macropod, suggests that the EgAgB genes may play a fundamental role in parasite survival.
EgAgB has been considered to be a polymorphic gene [5], [8], due likely to host selection for adaption given that E. granulsous strains are generally specific for the intermediate hosts they infect [28]. Accordingly, different stains have been presumed to have different genomic isoforms or alleles for some of their EgAgB genes [10], [30]. An alignment with sequences from GenBank showed that EgAgB1 has two or three major clusters of intronic and second exonic sequences (Fig. S1). As the sequences have the same intronic and exonic sequence lengths and several nucleotide substitutions, they are likely to be encoded by a polymorphic gene that is strain-related [31]. It is not clear whether the variation of the sequence is due to heterozygosity, which has been shown in the Echinococcus malate dehydrogenase (MDH) gene [32], or to the presence of host-specific alleles. We isolated one cluster of EgAgB1 sequence from the MAW and larval PSC of E. granulosus. The parasite samples were collected from different hosts from two continents. One sequence (GU166203) was identical to one of the cluster sequences (AF143813 cluster, Fig. S1) that is related to a sheep strain sequence [31]. Another two clusters in the EgAgB1 subfamily are related to those from E. granulosus cattle (FJ696924-FJ696928) and buffalo (FJ696936, FJ696923) strains [30], [31]. Further study is required to determine whether EgAgB1 can be used as a universal probe for distinguishing the recognized genotypes of E. granulosus [33].
It is not surprising that EgAgB comprises a multigene family. Southern blotting analysis showed several bands present in hybridizations with genomic DNA from E. granulosus [8], [34] indicating the family has different genomic loci. With genomic DNA extracted from a single cyst, Chemale et al. [35] suggested there are three genes in the EgAgB gene family. Southern analysis, however, does not indicate precisely the number of genes in the family, which can only be determined by a sequencing approach.
We performed a phylogenetic analysis of inferred amino acid sequence of EgAgB family proteins to illustrate the evolutionary relationships within the family and particularly with those of the confamilial genus Taenia spp. (Fig. 5). The Taenia proteins have been termed “Taeniidae antigens,” as the encoding genes are found mostly in taeniid cestodes [26], with one sequence (AF249884) isolated from Hymenolepis diminuta, a member of the cyclophyllidean family Hymenolepididae. The proteins were classified into several major and distinct clusters. All sequences from the genus Echinococcus, including sequences from E. granulosus (EgAgB), E. multilocularis, E. vogeli, E. oligarthrus, E. ortleppi and E. canadensis form a monophyletic clade (Fig. 5), which is separated from those of the large tape worms, such as Taenia and Hymenolepis. This suggests that these genes have radiated in the Echinococcus lineage after its separation from the other taeniids. This radiation might be correlated with the unique biological features of the Echinococcus genus such as the extensive asexual reproductive capacity of the multi-compartmentalized metacestode stage, the use of different hosts and organs for cystic development, small MAW with few segments and low definitive host specificity; perhaps some or all of these traits are indicative of a role for the antigen B proteins.
Mumuti et al. [21] showed, using specific antibodies against each of 5 gene products in E. multilocularis, that the EmAgB genes were differentially expressed in the adult and larval cyst and PSC stages, with EmAgB3 being predominantly expressed; however, the ONC stage, which is responsible for human infection, was not included in the analysis. To determine the expression of the EgAgB family of genes in E. granulosus, we used quantitative PCR to measure their expression levels in the PSC, CM, IAW, MAW and ONC. As the genes in each of the subfamilies have very similar sequences, it was challenging to design PCR primers to readily distinguish them individually. However, the differences in sequences between the subfamilies allowed us to design specific primers to amplify cDNA fragments to distinguish the genes at the subfamily level. We initially used E. granulosus actin II (accession no. L07773) as a house-keeping gene as used in other studies with Echinococcus [31], [36]–[39] but this gene proved to be highly variable between different stages of the parasite at the transcription level, being expressed 35 and 20 times higher in MAW than in the ONC and CM, respectively (Fig. 5). Our results showed that the EgAgB gene family members were expressed differentially, with the EgAgB3 genes predominantly expressed in all life-cycle stages investigated, including the ONC. The expression profiles obtained were similar to these obtained by by Mamuti et al. [21], for E. multilocularis, who used specific antibodies against the EmAgB protein family. We were able to demonstrate that there are 4 genes in the EgAgB3 subfamily. However, it is difficult to use normal real time PCR to distinguish their expression in E. granulosus due simply to the high similarity in their transcription levels. We expressed all the second exonic sequences of EgAgB3 and subsequent analysis showed that they cross reacted strongly (data not shown), indicating neither normal real time PCR, nor Western blot analysis can be used for distinguishing each of the genes in the subfamily. Although not accurate, sequencing mRNAs from different stages of E. granulosus may be a way to predict the expression profiles of the EgAgB3 genes based on the transcription frequency of the genes. We demonstrated that EgAgB3/1 is the most predominant subfamily gene expressed in the intermediate host cyst and PSC stages, suggesting that EgAgB3/1 may be a suitable serodiagnostic target molecule.
It is almost 40 years since the EgAgB protein was identified in E. granulosus hydatid cyst fluid [40], but its precise biological function(s) still remains unknown. Here, we have shown that the E. granulosus antigen B family contains at least 10 genes. We believe these new findings are important for addressing the expression and regulation of the EgAgB genes, as they may provide new insights for determining the biological features and characteristics of the proteins encoded by this complex gene family, notably its potential role in the interaction between parasite and host as an immune modulator, benefiting parasite survival.
|
10.1371/journal.pntd.0003572 | The Effect of Meteorological Variables on the Transmission of Hand, Foot and Mouth Disease in Four Major Cities of Shanxi Province, China: A Time Series Data Analysis (2009-2013) | Increased incidence of hand, foot and mouth disease (HFMD) has been recognized as a critical challenge to communicable disease control and public health response. This study aimed to quantify the association between climate variation and notified cases of HFMD in selected cities of Shanxi Province, and to provide evidence for disease control and prevention. Meteorological variables and HFMD cases data in 4 major cities (Datong, Taiyuan, Changzhi and Yuncheng) of Shanxi province, China, were obtained from the China Meteorology Administration and China CDC respectively over the period 1 January 2009 to 31 December 2013. Correlations analyses and Seasonal Autoregressive Integrated Moving Average (SARIMA) models were used to identify and quantify the relationship between the meteorological variables and HFMD. HFMD incidence varied seasonally with the majority of cases in the 4 cities occurring from May to July. Temperatures could play important roles in the incidence of HFMD in these regions. The SARIMA models indicate that a 1° C rise in average, maximum and minimum temperatures may lead to a similar relative increase in the number of cases in the 4 cities. The lag times for the effects of temperatures were identified in Taiyuan, Changzhi and Yuncheng. The numbers of cases were positively associated with average and minimum temperatures at a lag of 1 week in Taiyuan, Changzhi and Yuncheng, and with maximum temperature at a lag of 2 weeks in Yuncheng. Positive association between the temperature and HFMD has been identified from the 4 cities in Shanxi Province, although the role of weather variables on the transmission of HFMD varied in the 4 cities. Relevant prevention measures and public health action are required to reduce future risks of climate change with consideration of local climatic conditions.
| Understanding of the impact of weather variables on HFMD transmission remains limited due to various local climatic conditions, socioeconomic status and demographic characteristics in different regions. This study provides quantitative evidence that the incidence of HFMD cases was significantly associated with temperature in Shanxi Province, North China. The delayed effects of weather variables on HFMD dictate different public health responses in 4 major cities in Shanxi Province. The results may provide a direction for local community and health authorities to perform public health actions, and the SARIMA models are helpful in the prediction of epidemics, determination of high-risk areas and susceptible populations, allocation of health resources, and the formulation of relevant prevention strategies. In order to reduce future risks of climatic variations on HFMD epidemics, similar studies in other geographical areas are needed, together with a longer study period to enable trend analysis which takes into consideration local weather conditions and demographic characteristics.
| Hand, foot and mouth disease (HFMD) is an emerging infectious disease mainly caused by highly contagious intestinal viruses human enterovirus 71 (EV71) and coxsackievirus A16 (Cox A16) [1–3]. It is a human syndrome characterized by a distinct clinical presentation of fever, accompanied by oral ulcers and maculopapular rash or vesicular sores on the hands and feet, and sometimes the buttocks. HFMD transmission is through close personal contact, exposure to feces, contaminated objects and surfaces of an infected person. In recent decades, HFMD has become a growing public health threat to children, particularly those under the age of 5 [4–6]. Epidemics of HFMD are frequent and widespread in Asian countries, especially in China, Singapore, Malaysia and Japan, which have documented many large outbreaks of HFMD with severe complications and deaths predominantly among children [2,7–12]. At present, there is no specific curative treatment, and vaccine development is still in progress [13]. Weather variables might play a certain role in the transmission of the disease, as time series analysis in Guangzhou and Shenzhen, in China, showed that weather variation could affect the disease occurrence with a short lag period [14,15]. Under the context of global environmental change, the frequency of HFMD epidemics may be projected to increase in the future due to continued viral mutation, climate change, and the lack of health resources and effective surveillance systems in some regions [14,16–18]. Therefore, risk detection, early warning of HFMD cases with the capacity to predict a possible epidemic, and efficient public health response will be important to minimize the risk of epidemics and adverse impacts of HFMD.
Intestinal viruses have a worldwide distribution. In tropical and semitropical areas, they are present throughout the year, whereas in temperate climatic zones, they are more common during summer and fall [19]. A previous literature review has indicated that HFMD typically occurs in the summer and early autumn [20]. Such seasonal distribution suggests climatic variations may play a certain role in the transmission of the disease, particularly in temperate areas. The fifth assessment report of the Intergovernmental Panel on Climate Change (IPCC AR5) pointed out that the globe has experienced surface warming and projections of annual average temperature changes for 2081–2100 under Representative Concentration Pathways (RCPs) 2.6 and 8.5, relative to 1986–2005 [21]. In China, increases in temperature have been observed in most regions from south to north during the last decade [22]. Climate change has also been identified as an important risk factor for transmission of infectious diseases, especially vector and food borne diseases [23].
Since 2008, the China Ministry of Health has listed HFMD as a notifiable Class-C communicable disease, which has been included in the national communicable disease surveillance system and reporting network. Over six million cases had been reported up to the end of 2012 in China [17]. Studies have examined the association between HFMD and climate variables in selected regions [14,24–26]. Meteorological parameters, such as temperature and relative humidity, may affect the transmission and the frequency of HFMD. However, the effects of climate variables are not consistent in published studies, which could be due to various local climatic conditions, socioeconomic status and demographic characteristics in different regions. In particular, an understanding of the impact of seasonality and meteorological variables on disease transmission remains limited. A comparison among different cities within a Province may minimize potential confounding effects of socioeconomic inequalities and demographic differences, which will be used in this study to examine the role of climate variation on the incidence of HFMD, using Seasonal Autoregressive Integrated Moving Average (SARIMA) models.
The purpose of this study was to identify, with SARIMA models, the impact of meteorological variables on HFMD in 4 major cities of Shanxi in northern China, using existing surveillance data, and to quantify the relationship between climate variation and the incidence of HFMD. This study will provide scientific evidence to assist public health policy-making to carry out efficient prevention and control of HFMD.
The study was approved by the Ethics Committee of Shanxi Medical University (No. 2013091), China, and conducted in accordance with its guidelines. Shanxi HFMD data were provided by the Shanxi Center for Disease Control and Prevention and were obtained from the National Surveillance System. No informed consent was required because no individual-level analysis was performed. The information contained in the patients’ records was anonymized and de-identified prior to analysis. Only aggregated data were analyzed and reported.
Shanxi Province, which is located in North China, has a temperate, continental, monsoonal climate with four distinct seasons. The average temperature in January is in the range -16°C to -2°C and in July between 19°C and 28°C, the average rainfall is between 350 to 700 mm, and the average daily sunshine is between 7 to 9 hours. This study selected 4 major cities from north to south (Datong, Taiyuan, Changzhi and Yuncheng), which have similar socioeconomic and demographic conditions (Fig. 1). Datong (latitude 40°2′30" N and longitude 113°35′50" E) is the northernmost prefecture-level city of Shanxi Province, with a population of 3.36 million by the end of 2012 (data from the Shanxi Bureau of Statistics). Taiyuan (latitude 37°43' 36" N and longitude 112°28′14" E) is the capital and largest city of Shanxi, which is located at the centre of the province with an East-West span of 144 km and a North-South span of 107 km, and a population of 4.26 million in 2012. Changzhi (latitude 36°11′0" N and longitude 113°6′0" E) is the southeast city of Shanxi Province, with a population of 3.37 million in 2012. Yuncheng (latitude 35°1′33" N and longitude 111°0′19" E) is a southwestern city in Shanxi, with a population of 5.19 million in 2012.
Daily meteorological data including precipitation, average temperature, maximum temperature, minimum temperature, average relative humidity, and hours of sunshine for the study period from 1 January 2009 to 31 December 2013, were obtained from the Shanxi Meteorological Administration. Daily meteorological data were aggregated on a weekly basis which comprised a total period of 261 weeks.
Being a notifiable disease [17], all clinical and hospital doctors are required to report cases of HFMD to the local Center for Disease Control and Prevention. The diagnosis criteria for HFMD cases were provided in a guidebook published by the Chinese Ministry of Health [27,28]. Patients with HFMD have the following symptoms: fever, papules and herpetic lesions on the hands or feet, rashes on the buttocks or knees, inflammatory flushing around the rash and fluid in the blisters, or sparse herpetic lesions on the oral mucosa. A recent data quality survey report has demonstrated that the data are of high quality in China, with reporting completeness of 99.84% and accuracy of the information reported to be 92.76% [29]. In addition, in order to reduce apparent underreporting and a large number of missing information of patients from the early stage of the surveillance system in 2008, only the data from 2009 to 2013 were used for analysis. The weekly data of HFMD cases for the period were obtained from the China Information System for Disease Control and Prevention in Shanxi. According to our data, 90.9%, 91.0%, 90.3% and 97.8% HFMD cases were children aged 0–5 years in Datong, Taiyuan, Changzhi and Yuncheng, respectively. Therefore, we focused analysis on the incidence of HFMD among children aged 0–5 years in this study.
The analysis includes descriptive, correlation and time series regression analyses. The meteorological variables data were calculated for intervals of 7 consecutive days, and transformed into a time series format. Descriptive analysis was performed by describing the distribution of climate variables and HFMD cases. Spearman rank correlation and partial correlation analysis were used to examine the association between each meteorological variable and the incidence of HFMD. In addition, given the potential lagged effect of the meteorological variables on disease transmission, cross-correlation analysis was also performed with relevant time lag values. Time series analysis was used to assess the effect of climatic variables on HFMD incidence.
The plot of the observed HFMD incidence showed most of the cases occurred from May to July in the 4 cities (Fig. 2A-2D). Furthermore, the plots of autocorrelation function (ACF) and partial auto correlation function (PACF) of HFMD cases (Fig. 2E-2H) showed the time series was non-stationary. With the temporal dependence of HFMD incidence, the need to use a SARIMA model was evident. The Seasonal Autoregressive Integrated Moving Average (SARIMA) model (Box and Jenkins method) has been recently applied in epidemiological studies [24,30–33], and was used to describe current (and future) incidence of HFMD in terms of their past values in this study. SARIMA models extend basic ARIMA models and allow for the incorporation of seasonal patterns. A SARIMA model, which includes seasonal and non-seasonal components, is typically represented by (p,d,q)(P,D,Q)s, where p represents the order of autoregression (AR), d is the order of differencing, and q is the order of the moving average (MA). P, D, and Q are their seasonal counterparts, and s is the seasonal lag [34]. The long-term trend and seasonal components of each time series can be removed using SARIMA models [24]. In addition, we considered the weather during holidays may also affect the occurrence of HFMD.
The development of a SARIMA model is a four-step process. Therefore, for the HFMD time series analysis, it was firstly necessary to stabilize the variance of the series by square root transformation, and seasonal and regular differencing was also applied. Secondly, in order to identify the order of MA and AR parameters, the structure of temporal dependence of stationary time series was assessed respectively, by the analysis of autocorrelation (ACF) and partial autocorrelation (PACF) functions. From the correlograms of the series, the p value may equal 0, 1 or 2 for autoregressive parameters and q value may equal 1, 2 or 3 for moving average parameters (Fig. 3A-3D). Thirdly, parameters of the model were estimated by using the maximum likelihood method. The goodness-of-fit of the models was determined for the most appropriate model (the lowest normalized Bayesian Information Criteria (BIC) and the highest stationary R square (R2)), using the Ljung-Box test that measures both ACF and PACF of the residuals, which must be equivalent to white noise. The significance of the parameters should be statistically different from zero. Finally, the predictions were performed by using the best fitting model. The predictive validity of the models was evaluated by calculating the root mean square error (RMSE), which measures the amount by which the fitted values differ from the observed values. The smaller the RMSE, the better the model is for forecasting. Therefore, the SARIMA model was developed and verified by dividing the data file into two date sets: the data from the 1st calendar week of 2009 to the 52nd calendar week of 2012 were used to construct a model; and those from the 1st calendar week to the 52nd calendar week of 2013 were used to validate it. For statistical analysis SPSS version 19.0 and Stata version 12.0 were used.
The sensitivity analysis was conducted based on daily unit to check whether the number of weekly HFMD cases could affect the result estimates. Meanwhile, we controlled for day of the week and public holidays using categorical indicator variables. In addition, we graphically examined the exposure-response curves derived using a smoothing function [28,35–37], and natural cubic splines [35] to control long-term trend and seasonality with 6 df per year for time [37], which was done using the distributed lag non-linear models (dlnm) package in the software R.
Meteorological variables and number of HFMD cases show differences from the northern city to the southern city (Table 1). The northern city (Datong) has a lower temperature and relative humidity than the central city (Taiyuan) and southern cities (Changzhi and Yuncheng); while the southern cities have less sunshine than the central city and northern city. During the study period, the number of HFMD cases in Taiyuan and Changzhi were more than that in Datong and Yuncheng.
In the 4 study cities, precipitation, temperature, relative humidity and hours of sunshine were positively correlated with incidence of HFMD (p < 0.05). Different meteorological variables may also be correlated with each other. For example, average temperature was positively correlated with maximum temperature in the 4 cities from north to south (rs = 0.994, 0.990, 0.990, 0.981; p < 0.001, respectively), and also correlated with minimum temperature (rs = 0.993, 0.988, 0.987, 0.989; p < 0.001, respectively). Accounting for these correlations, the association between meteorological variables and the number of HFMD cases were then analyzed using partial correlations. Results showed the associations of increased number of HFMD cases with increasing atmospheric temperature in the 4 cities (p < 0.05). In addition, the results showed statistically significant but weaker correlation for the association between relative humidity and the incidence of HFMD in Taiyuan, as well as the weaker correlation between precipitation and HFMD in Changzhi (Table 2).
In order to estimate the values of parameters in fitted models, these models were diagnosed by analyzing the data with several SARIMA models without the weather variables, and the models in which the residual was not likely to be white noise were excluded. Therefore, the univariate SARIMA (0,1,1)(2,0,1)52 model for Datong; SARIMA (2,1,3)(1,1,1)52 model for Taiyuan; SARIMA (0,1,1)(0,1,1)52 model for Changzhi; and SARIMA (0,1,1)(1,1,2)52 model for Yuncheng had both the lowest Bayesian information criterion (BIC) and the highest R2 values and were the best to fit the HFMD cases, respectively (Table 3). The Ljung-Box test confirmed that the residuals of the time series were not statistically dependent (p > 0.05) and the residuals on ACF and PACF plots showed the absence of persistent temporal correlation (Table 3) (Fig. 4). The selected SARIMA model fitted the observed data from 2009 to 2012. Then, the model was used to project the number of HFMD cases between January to December 2013, and was validated by the actual observations. The validation analysis suggested that the model had reasonable accuracy over the predictive period in the 4 cities (root-mean-square error (RMSE) = 1.763, 4.505, 4.907 and 2.817, respectively) (Table 3).
The cross-correlation analyses showed the lag effects of the meteorological variables on the number of HFMD cases were different in the 4 cities. In Datong, HFMD was significantly positively associated with average temperature at lag 0 (coefficients = 0.540, p < 0.05), mean maximum temperature at lag 0 (coefficients = 0.520, p < 0.05), mean minimum temperature at lag 0 (coefficients = 0.542, p < 0.05). In Taiyuan, the cases were significantly positively associated with average temperature at lag 1 week (coefficients = 0.615, p < 0.05), maximum temperature at lag 1 week (coefficients = 0.612, p < 0.05), minimum temperature at lag 1 week (coefficients = 0.606, p < 0.05), relative humidity at lag 3 weeks (coefficients = 0.224, p < 0.05). In Changzhi, the disease was significantly positively associated with average temperature, maximum and minimum temperature at lag 1 week (coefficients = 0.479, 0.472 and 0.465, p < 0.05), respectively. In Yuncheng, HFMD was significantly positively associated with average temperature at lag 1 week (coefficients = 0.351, p < 0.05), maximum temperature at lag 2 weeks (coefficients = 0.347, p < 0.05) and minimum temperature at lag 1 week (coefficients = 0.352, p < 0.05).
To reduce potential multicollinearity, weekly average temperature, weekly mean maximum temperature and weekly minimum temperature were put into separate regression models (Models 1, 2 and 3). In Model 1, 2 and 3, temperature (average, maximum, and minimum, respectively) was included, along with other meteorological variables. The results indicated that average temperature, and maximum and minimum temperatures with different lag times were significant in the SARIMA Model 1, Model 2 and Model 3, respectively (Table 4). Overall, SARIMA models with temperature were a better fit and validity than the models without the variable (Stationary R-squared (Stationary R2) increased, while the BIC decreased) (Table 3 and Table 4). Other meteorological variables were not significantly included in the models, indicating their contribution was not statistically significant in this study.
The models suggest that in Datong, a 1°C rise in weekly average temperature, weekly mean maximum temperature and weekly mean minimum temperature may be related to an increase in the weekly number of cases of HFMD of 0.8% (95%CI: 0.3%-1.2%), 0.6% (95%CI: 0.1%-1.0%) and 0.7% (95%CI: 0.2%-1.3%) respectively. In Taiyuan, a 1°C rise in weekly average temperature, weekly mean maximum temperature and weekly mean minimum temperature may be related to an increase in the weekly number of cases of HFMD of 1.4% (95%CI: 0.5%-2.7%), 1.0% (95%CI: 0.3%-1.8%) and 1.1% (95%CI: 0.2%-2.1%) respectively. In Changzhi, a 1°C rise in weekly average temperature, weekly mean maximum temperature and weekly mean minimum temperature may be related to an increase in the weekly number of cases of HFMD of 1.1% (95%CI: 0.1%-2.1%), 1.6% (95%CI: 0.4%-2.7%) and 1.5% (95%CI: 0.3%-2.7%) respectively. Finally, in Yuncheng a 1°C rise in weekly average temperature, weekly mean maximum temperature and weekly mean minimum temperature may be associated with an increase in the weekly number of cases of HFMD of 2.1% (95%CI: 0.4%-4.7%), 1.4% (95%CI: 0.3%-8.4%) and 1.9% (95%CI: 0.1%-5.8%) respectively (Table 4). Table 4 indicates that the incidence of the disease may rise with the increase of temperature, but only within a certain range of temperatures. The selected SARIMA model was used to project the number of HFMD cases in each city for the 52 weeks between January and December 2013. The validation for January to December 2013 data showed a good fit between observed and predicted data (Fig. 5).
A comparison between the models for the 4 cities, reveals that although a 1°C rise in temperature may cause a similar relative increase in the number of cases, the lag times for the effects of temperatures were shorter in Datong (at lag 0) than those in Taiyuan, Changzhi and Yuncheng (at lag 1 week). The lag times for the effects of maximum temperature were longer in Yuncheng (at lag 2 weeks) than those in Changzhi, Taiyuan and Datong (Table 4).
In the sensitivity analysis, the results using daily data indicated that a 1°C rise in daily average temperature may be related to an increase in the daily number of cases of HFMD of 1.2% (95%CI: 0.1%-2.3%) at lag 1 day, 1.6% (95%CI: 1.0%-2.2%) at lag 8 days, 1.5% (95%CI: 0.5%-2.5%) at lag 6 days, and 2.4% (95%CI: 0.1%-4.7%) at lag 8 days in Datong, Taiyuan, Changzhi and Yuncheng, respectively. Although the analysis of daily data is likely to yield a more precise estimate compared to weekly data, the results were similar. Fig. 6 shows non-linear dose-response relationships for temperature with HFMD occurrence in the 4 cities, and an increase in HFMD occurrence within a short interval.
HFMD has become an important public health concern in the affected countries and has attracted an increasing research interest [38–40]. In China, HFMD has been a notifiable infectious disease since 2008 and appropriate prediction of risk may aid in the effective control of the disease nationwide. Despite the evidence that shows that climate variation is closely linked with a huge burden of vector and food borne diseases at the global scale, the role of weather variables in HFMD transmission is still not well understood. This is because its epidemiological characteristics may vary over different geographic locations due to various climatic and socioeconomic situations in China. This study is the first to compare the association between climatic variables and the HFMD transmission among different cities within a Province. The results suggest that temperature may play an important role in the transmission of HFMD in the 4 cities although the lag times for the effects of temperatures differed. This may aid in the projection of the disease and in disease control and prevention, especially in an era of global environmental change.
HFMD has a seasonal distribution, and there are different epidemiological features in different regions due to various climatic, geographic and socioeconomic factors. In Singapore, there is a large peak from mid-March to the end of May and a smaller peak from early October to early December 2008 [41]. In Japan, one single peak was detected in July 2011, particularly in western Japan [12]. In Taiwan, a study showed HFMD peaked in June and disappeared after August 2008 [42]. In Hong Kong, a seasonal peak was detected in the warmer months (May-July), along with a smaller winter peak (October-December) from 2001 to 2009 [43].
In mainland China, the number of reported cases of HFMD and its associated morbidity and mortality varies remarkably among various Provinces. The seasonal peaks can vary from a single peak to double peaks in some Provinces (such as Henan, Shandong, Guangdong) [24,44,45]. There is also a trend for the disease to move from the south to the north between May and June [46]. In this study, HFMD incidence varied seasonally with the majority of cases in the 4 cities occurring from May to July, which is typically one month later than that observed in other Provinces [45]. Specifically, the seasonal patterns of HFMD epidemics were different in the 4 cities. In the central city of Taiyuan, two peaks per year were observed in the warmer months and winter months, and the highest peak was present from May to July, accompanied by a smaller peak from October to December. However, in the northern city and the southern cities, HFMD presented once per year and biannual patterns were evident. The cases peaked in 2009, 2009 and 2010 in Datong, Changzhi and Yuncheng, respectively, indicating HFMD cases occur in spatio-temporal clusters [17]. The SARIMA models also showed the seasonal variation in the 4 cities. These findings are similar to those of other studies investigating the epidemiological features of HFMD [47,48].
Global climate projections have suggested an increase in the distribution and prevalence of infectious diseases in association with climate change [49]. Climate variations may change the reproductive capacity of infectious pathogens and vectors, alter the survival of viruses in the physical environment, alter patterns of water and food use, change human behaviour, and increase disease prevalence [50,51]. It was observed from this study that HFMD incidence was significantly associated with meteorological variables, which is consistent with previous studies [7,14,24–26]. However, previous studies have not excluded the potential confounding factor of socioeconomic status, and findings have been inconsistent. In this study, we compared the time series data within Shanxi Province, using 4 major cities with similar socioeconomic status. Our results indicated that temperature could be the key climatic indicator in the transmission of HFMD. Moreover, the SARIMA models suggested that although the pattern differed in each city, temperature had a significant impact on the transmission of HFMD in the 4 cities. The SARIMA models, which allow the integration of external factors (e.g. climatic variables), are a useful tool for interpreting and applying surveillance data in disease control and prevention [24]. Our study shows that, within a certain range of temperature variation, a 1°C rise in average temperature may lead to 0.8%, 1.4%, 1.1% and 2.1% increase in the number of cases of HFMD in Datong, Taiyuan, Changzhi and Yuncheng, respectively. This result is similar to other research on the effects of temperature on enteric infectious diseases [25,31,52]. In addition, the results show that a 1°C rise in maximum and minimum temperatures may also be related to increases in the number of cases in these regions, indicating temperature could be used as a forecasting factor in public health practice.
Although the associations between HFMD and other climate variables including precipitation, average relative humidity and hours of sunshine are statistically significant in the correlation analyses, they were not significant in the SARIMA models. In the context of global warming, precipitation, relative humidity and hours of sunshine may contribute to the transmission of enteric infectious diseases by affecting the ecological environment of pathogens, exposure probability and host susceptibility, thus maybe resulting in the incidence of diseases [49]. Therefore, further studies are necessary to better understand the role of these climatic factors on HFMD.
The time lag effect refers to the delay between the time of an exposure and the subsequent development of a disease. The time lag effects of temperature were observed in the central and southern cities of Taiyuan, Changzhi and Yuncheng, but no lag effect was detected in the northern city of Datong. These findings suggest HFMD cases may be forecast 1 week ahead in Taiyuan, Changzhi and Yuncheng according to weekly average temperature and minimum temperature, and 2 weeks ahead in Yuncheng according to weekly maximum temperature. This result is similar to a previous study conducted in Japan investigating the relationship between temperature and HFMD, with lag period 0–3 weeks [25]. Other studies showed the role of temperature and relative humidity at 2 weeks lag time in Hong Kong [26], and the role of temperature and relative humidity at 1 week lag time in Guangzhou [14]. On the contrary, our study has not detected the role of relative humidity on HFMD. These various impacts of climatic variables on HFMD could be attributed to local climate conditions and other socioeconomic characteristics. Therefore, local climate factors, together with other variables should be carefully considered for public health professionals to prevent or reduce future risks of HFMD incidence. Due to the different lag time effect in these cities, the response from disease surveillance systems should be different.
The limitations of this study should be acknowledged. Firstly, weekly data rather than daily data may underestimate the relationship between climate variables and HFMD incidence. This may affect the accuracy of exposure assessment. Secondly, the Seasonal Autoregressive Integrated Moving Average (SARIMA) model has been recently applied in modeling and projecting for both non-linear and non-stationary time series, but other functions are required to assess the dose-response curve for temperature and HFMD occurrence. Thirdly, we were unable to differentiate the pathogens of HFMD cases reported to CDC surveillance system. Therefore, it was not possible to examine the specific impacts of climatic conditions on different pathogens. In addition, our analysis is exploratory, and we are unable to exclude the possibility of a spurious finding or unmeasured confounding factors that may be associated with both weather/ecological variables and HFMD occurrence, which has been a common challenge for ecological studies. Although we considered that holidays may affect the occurrence of HFMD, it was difficult to define a holiday variable using weekly data, and it was not possible to combine holidays (in summer and winter). Despite these shortcomings, these findings are useful as they provide information to better understand the effect of climate variation on HFMD, and this study may inform policy makers in the development of efficient prevention strategies.
Although there are some time series studies in China on HFMD, these studies scarcely considered the potential impact of socioeconomic status. The findings in this study indicate that the occurrences of HFMD were positively associated with temperature in 4 major cities which have similar socioeconomic status and demographic characteristics. In addition, different lag effects of temperature were observed in selected regions from north to south. The results will be useful to assist public health responses in the different regions and informing local community and health authorities to better predict disease outbreaks.
|
10.1371/journal.pcbi.1006008 | Same but not alike: Structure, flexibility and energetics of domains in multi-domain proteins are influenced by the presence of other domains | The majority of the proteins encoded in the genomes of eukaryotes contain more than one domain. Reasons for high prevalence of multi-domain proteins in various organisms have been attributed to higher stability and functional and folding advantages over single-domain proteins. Despite these advantages, many proteins are composed of only one domain while their homologous domains are part of multi-domain proteins. In the study presented here, differences in the properties of protein domains in single-domain and multi-domain systems and their influence on functions are discussed. We studied 20 pairs of identical protein domains, which were crystallized in two forms (a) tethered to other proteins domains and (b) tethered to fewer protein domains than (a) or not tethered to any protein domain. Results suggest that tethering of domains in multi-domain proteins influences the structural, dynamic and energetic properties of the constituent protein domains. 50% of the protein domain pairs show significant structural deviations while 90% of the protein domain pairs show differences in dynamics and 12% of the residues show differences in the energetics. To gain further insights on the influence of tethering on the function of the domains, 4 pairs of homologous protein domains, where one of them is a full-length single-domain protein and the other protein domain is a part of a multi-domain protein, were studied. Analyses showed that identical and structurally equivalent functional residues show differential dynamics in homologous protein domains; though comparable dynamics between in-silico generated chimera protein and multi-domain proteins were observed. From these observations, the differences observed in the functions of homologous proteins could be attributed to the presence of tethered domain. Overall, we conclude that tethered domains in multi-domain proteins not only provide stability or folding advantages but also influence pathways resulting in differences in function or regulatory properties.
| High prevalence of multi-domain proteins in proteomes has been attributed to higher stability and functional and folding advantages of the multi-domain proteins. Influence of tethering of domains on the overall properties of proteins has been well studied but its influence on the properties of the constituent domains is largely unaddressed. Here, we investigate the influence of tethering of domains in multi-domain proteins on the structural, dynamics and energetics properties of the constituent domains and its implications on the functions of proteins. To this end, comparative analyses were carried out for identical protein domains crystallized in tethered and untethered forms. Also, comparative analyses of single-domain proteins and their homologous multi-domain proteins were performed. The analyses suggest that tethering influences the structural, dynamic and energetic properties of constituent protein domains. Our observations hint at regulation of protein domains by tethered domains in multi-domain systems, which may manifest at the differential function observed between single-domain and homologous multi-domain proteins.
| A large proportion of proteins, coded in the genomes of diverse organisms, is constituted of more than one domain [1, 2]. Multi-domain proteins have evolved from single-domain proteins through many duplication and adaptive events [3]. Duplication and shuffling of domains have led to the emergence of various unique and novel functions using an existing repertoire of domains [3–5]. Presence of multiple domains in proteins has been reported to confer structural stability [6] and folding and functional advantages [7]. Proteins can be decomposed into domains based on various criteria namely sequence, structure, function, evolution and mobility [8, 9]. At the sequence level, domains are defined on the basis of conservation of residues over significant length; structural domains are defined on the basis of globularity and compactness; functional domains are modules in proteins which can function independently of other modules in the protein; evolutionary domains are protein modules propagating through evolution by recombination, transposition, shuffling etc. and protein modules with high correlated mobility are identified as domains according to the mobility definition [8]. It is important to note that a given protein may have different but equally valid domain annotations depending upon the basis of domain annotation [9].
Often domains in multi-domain proteins interact with one another. The role of domain-domain interfaces has been implicated in long-range allostery regulation [10–12], the emergence of a new function [13], the regulated mobility of the proteins [14] etc. In comparison to protein-protein interfaces, geometrical and chemical properties of domain-domain interfaces have been observed to be intermediate to interfaces in permanent and transient protein-protein complexes [15]. Domain interface size and linker length have been observed to influence the folding and stability of domains in multi-domain proteins [16]. The physiochemical nature of the domain-domain interface [15], the associated energetic of domain-domain interface [6] and its influence on folding in multi-domain proteins [16, 17] is well described. A recent review covers extensively the effect of domain tethering on the thermodynamics of the protein and its influence on the protein stability and folding [18]. But how protein domains behave in multi-domain proteins in comparison to single-domain proteins, has largely been unexplored and unaddressed, except some studies on the influence of tethering on the folding pathway [16, 17, 19, 20]. In the current study, we have explored how protein domains behave in multi-domain systems in comparison to single-domain systems.
For this, identical protein domains crystallized in two forms (a) tethered to other protein domains and (b) tethered to fewer protein domains than (a) or not tethered to any protein domain were studied. For example, full-length rat DNA polymerase β consists of three domains (DNA polymerase β N-terminal; DNA polymerase β and DNA polymerase β catalytic). Crystal structures are available for full-length protein (PDB id: 1BPD) and the two C-terminal domains (PDB id: 1RPL) (Fig 1). For the study, we have compared the properties of the second and third domains in the two crystal forms. This comparison allowed us to study the influence of the first domain on the second as well as the third domains. Further comparative dynamics analyses of homologous protein domains were carried out to understand the functional relevance of tethering of domains. Analyses reveal an intricate coupling between the domains in multi-domain systems leading to alteration in dynamics in 18 protein pairs. Structural and energetics differences were observed in half the numbers of cases studied. Differential dynamics were observed for identical and structurally equivalent functional residues of the homologous protein domain pairs. Our observations strongly suggest that tethering of domains in multi-domain proteins changes the properties of constituent domains, thus regulating the function of the entire protein.
Differences in the conformation of domains were observed in comparative structural analyses of identical protein domain pairs crystallized in two forms (a) tethered to other protein domains (henceforth referred as MD) and (b) tethered to fewer protein domains than (a) or not tethered to any protein domain (henceforth referred as ID). Distributions of RMSD and GDT values for the 20 protein domain pairs are shown in Fig 2A. To delineate the differences arising due to differences in crystal packing, RMSD and GDT distributions of the protein domain pairs were compared with the control dataset 1. The control dataset 1 consists of pairs of identical monomeric proteins. Distributions of structural deviation of the protein domain pairs and control dataset 1 were observed to be significantly different (two-sample KS test, p-value: 1.26e-06 (RMSD), p-value: 8.14e-06 (100-GDT), Fig 2B and 2C). This suggests that structural deviations observed in the protein domain pairs are likely to be due to tethering of domains and not due to crystallization artefact. The upper quartile limits of RMSD (RMSD > 1Å) and GDT (100-GDT > 5) distributions of the control dataset 1 were taken as a cut-off to identify the protein domain pairs with significantly different conformations. RMSD and GDT distributions of the protein domain pairs suggest subtle changes in global conformation of the common protein domains for 10 cases (100-GDT ≤ 5) while 10 cases show substantial changes in the conformation (100-GDT > 5) (Fig 2A).
Since GDT and RMSD give an estimation of structural deviation over the entire length of a protein domain, significant structural deviations at local short stretches can be missed out. All the protein domain pairs were analyzed to identify stretches of residues showing significant structural deviation (refer structural analysis section in materials and methods). Four categories of pairs were observed: (i) only domain-domain interface showed significant structural deviation, (ii) regions other than the domain-domain interface showed structural deviation but no structural deviations were observed at the domain-domain interface, (iii) structural deviations were observed both at domain-domain interface and regions other than the domain-domain interface and, (iv) no significant structural deviation was observed between the protein domain pairs. Representative examples of the 4 case types are shown in Fig 2E. 9 out of the 20 protein domain pairs showed structural changes at regions other the domain-domain interface (S1A Fig). Further analysis of the regions with significant structural deviation shows ~14% of such regions harbors functional residues while ~24% harbors domain-domain interface residues. Functional significance of ~62% of the residues cannot be commented upon (Fig 2D). It has to be noted that structural deviations were observed independently of the number of domain-domain interface residues. For example, despite no interaction between the domains in fibronectin, structural deviations are observed (S1B Fig). The observations suggest that tethering of domains can alter the conformation of the constituent domains, with many residues apart from domain-domain interface residues showing significant structural deviation.
Previous analyses by del Sol et al. have shown that network property, namely residue centrality of hemoglobin and NtrC differ between the inactive and the active state of the proteins [21]. Residue centrality measures the importance of the residue in maintaining the residue-residue communication network within the protein structure. Domains in conjunction with other domains can be treated as one of the states of the protein domain and the domains in the absence of tethered domains can be treated as another state of the protein domains. Hence, a network approach was undertaken to understand the differences in residue-residue contacts, if any, for the 20 protein domain pairs. To represent residue-residue communication numerically, a network parameter namely communicability centrality (henceforth referred as coc) is used. High communicability centrality measure of a residue implies its importance in residue-residue communication in the structure. The distribution of the coc score of ID is observed to be significantly different from MD (two-sample KS-test, p-value < 2.2e-16) (Fig 3A). Interface residues also show differences in the coc score between MD and ID (two-sample KS-test, p-value: 1.05e-08) (Fig 3B). Since interface residues form intensive contacts at the domain-domain interface in MD, we expected the coc scores to be lower for interface residues in ID than MD, but ~ 30% of the residues show higher coc in ID than MD (Fig 3B). This observation suggests that rewiring of intra-domain residue-residue contacts of interface residues results on tethering of domains. Distribution of coc scores of non-interface residues is also observed to be different between MD and ID (two-sample KS-test, p-value < 2.2e-16) (Fig 3C), implying that on tethering of domains in a multi-domain system, many residues which are not part of interface region also undergo changes in the residue-residue contacts. The functional residues did not show a significant difference in the coc distribution (S2A Fig, two-sample KS-test, p-value: 0.04). It has to noted that ~7% (291 residues out of 4284 residues) of residues show significant differences in centrality score (|centrality score (MD)–centrality score (ID)| > 1.5) (Fig 3A). These 291 residues belong to 12 domain pairs in the dataset. Only ~3% of these 291 residues form a part of domain-domain interface regions. Many residues showing a significant difference in coc score (|centrality score (MD)–centrality score (ID)| > 1.5) showed low structural deviation (S2B Fig) implying that rewiring of the residue-residue contact can happen without any significant structural deviation. An example of the coc distribution of a domain pair (fibronectin) is shown in Fig 3D. Fibronectin domain shows differences in centrality score both at the domain-domain interface residues (boxed as black in Fig 3D) as well as residues other than domain-domain interface residues (boxed as red in Fig 3D).
Normal mode analysis was used to study the extent of influence of tethering on the dynamics of the constituent domains. Normal modes, accounting for 80% variance of the protein motion, were calculated for each MD and ID of the 20 protein domain pairs. To compare the flexibility of MD and ID normalized summed square fluctuation values were compared. The flexibility profiles were observed to be statistically different for all the domain pairs, except two (Fig 4A and 4B, two-sample KS test, p-value < 2.2e-16). To ensure that the differences are not an artefact of crystal packing, flexibility profiles of ID and MD were compared with two control datasets namely control dataset 2 and control dataset 3 respectively. The control dataset 2 was generated by in silico removal of the tethered domains from MD. The domains in the control dataset 2 (referred to as AD) are essentially identical to ID in sequence as well as length. The flexibility profiles of ID and AD were observed to be similar (S3A Fig). The control dataset 3 was generated by in silico ligation of the ID with the tethered domain of MD. This was achieved by superimposing the ID onto MD, followed by in silico removal of the common domain from MD and then ligation of the remaining domains of MD with ID. The multi-domains in the control dataset 3 (referred to as swapped domain) are essentially identical to MD in sequence and length. The flexibility profiles of MD and swapped domains were also observed to be similar (S3B Fig). The similarity of the flexibility profiles of the protein domain pairs and the control datasets ensured that the differences observed in the flexibility profiles of MD and ID are a consequence of the tethering of the domains in multi-domain systems than a crystallization artefact.
The flexibility of the residues was observed to be different in MD and ID (Fig 4A). ~32% of the residues show higher flexibility in ID than in MD, while ~22% of the residues have higher flexibility in MD than in ID. The rest of the residues have comparable flexibilities. Higher variance in the distribution of flexibility of residues is observed for MD than ID (Fig 4B). The higher variance of the flexibility of residues in MD implies that many residues in MD show higher/lower flexibility than the mean flexibilty. To ascertain further, how the flexibility profiles of interface residues and functional residues differ in MD and ID, the flexibility distribution of the interface residues and functional residues were compared. The interface residues generally show higher flexibility in ID than MD (Fig 4C). A majority of interface residues (~70%) have higher flexibility in ID than in MD. But interestingly, ~30% of interface residues have comparable flexibility in MD and ID. Thus, some of the interface residues retain their rigidity in the isolated state as well. ~36% of the functional residues have higher flexibility in ID than MD while ~18% have higher flexibility in MD than ID (Fig 4D). Hence many functional residues are rigid in MD than ID. Many residues which are neither part of interface nor functional residues show differences in the flexibility profile (Fig 4D and S3C Fig). To ascertain whether the residues showing differences in fluctuation in MD and ID show structural deviation as well, we calculated the correlation between the two. A poor correlation (Spearman correlation coefficient: 0.25, S3D Fig) was observed between the differences in fluctuation and structural deviation, suggesting tethering of domains can alter the dynamic properties of protein domain without significant structural conformation change.
Residue-residue communication in protein domains is important for the function and structural integrity of proteins. Residues can relay information to other residues either by forming contacts or through synchronization of dynamics. To understand the influence of tethered domain on the synchronization of dynamics of residues in protein domain, the extent of correlation of fluctuation among residues (henceforth referred as cross-correlation) was studied. Higher number of residues with high cross-correlation value (|cross-correlation| ≥ 0.7) was observed for MD (~22%) as compared to ID (~10%) (Fig 5A). This observation implies that residues show tight coupling (|cross-correlation| ≥ 0.7) in the case of MD but no or weak coupling in the ID (|cross-correlation| < 0.7). Moreover, clusters of high correlation were observed in the case of MD; which often corresponded to sub-domains or domains or super-secondary structures in the spatial coordinate. The matrices of MD and ID were observed to have a low similarity (low Rv coefficient) for all the domain pairs except two (Fig 5B). A representative example (fibronectin) is shown in Fig 5C and cross-correlation matrices for 20 protein domain pairs are shown in S4 Fig. To ensure that differences are not observed due to crystal packing or other artefact, Rv coefficient between cross-correlation of ID and control dataset 2 and cross-correlation between MD and control dataset 2 were calculated (S5 Fig). The comparison ruled out any other factor apart from tethering for the behavior observed. An important point to note here is that this characteristic has been observed irrespective of the number of interactions between the domains. For example, the domains in rat DNA polymerase β do not interact with each other but still, low Rv coefficient is observed (1BPD in Fig 5B).
Molecular dynamic studies were carried out for 3 domain pairs from the dataset to study the synchronization of motions in the domain at all-atom level. These 3 pairs of domains were selected based on the number of interfacial residues between the domains. Tight coupling of motions was observed not only between the C-alpha of residues but also between the side-chains of residues in MD (S6 Fig). While weak or no coupling was observed for side-chains of residues in ID. Thus, molecular dynamics analysis for 3 pairs showed that higher cross-correlation between residues in MD is manifested not only at the backbone level, as observed also from NMA, but also at the side-chain level. All the observations imply that tethering of domains in multi-domain proteins alters the flexibility as well as the synchronization of the fluctuations of residues of the constituent domains.
From the network analysis of the structure of the 20 protein domain pairs, it was observed that certain residues show significant differences in the communicability centrality score. We further wanted to study whether this rearrangement in the intra-domain residue-residue contacts, as represented by communicability centrality score, changes the energetic stability of the residues and residue-residue contacts. Frustratometer algorithm [22] was used to study the effect of tethering on energetics distribution of residues. The algorithm calculates a parameter, single residue level frustration (SRLF), for each residue in the structure. Two parameters, configurational frustration index and mutational frustration index, are calculated for all the contact pairs in the structure. SRLF measures the energetic stability of the residue with respect to every other amino acid at that position. Configurational frustration index measures the stability of the contact pair with respect to every other configuration the contact pair can take during the folding process. Mutational frustration index measures the stability of the contact pair with respect to every other amino acid combination at that position. Mathematically, frustration index is the Z-score of the energy of the native with respect to the decoys. A residue or a contact is considered as minimally frustrated if the frustration index is greater than 0.78, highly frustrated if the frustration index is less than -1 and neutrally frustrated if frustration index is in between -1 and 0.78 [22].
The frustration indices were calculated for the 20 protein domain pairs. Though the distribution of SRLF of MD and ID were observed to be largely comparable (two-sample KS test, p-value: 0.98) (Fig 6A) but ~12% (region II, III, IV, VI, VII and VIII of Fig 6A) of the residues showed differences in the single residue level frustration (SRLF) with 5 residues (region III and region VII of Fig 6A) showing drastic substitution from high frustration to minimal frustration and vice-versa. These 12% residues are distributed over the entire domain dataset i.e. each domain pair have at least one residue showing different frustration indices. Residues apart from domain-domain interface residues and functional residues were also observed to differ in the frustration index (S7A Fig). Moreover, differences in the frustration index of MD and ID were observed to be independent of the structural deviation observed. Equivalent numbers of substitutions were observed at structural deviation greater than 1Å and lower than or equal to 1Å (Fig 6D). Similar trends as that of SRLF were observed for configurational frustration and mutational frustration (Fig 6B, 6C, 6E and 6F) but a higher number of contacts showed differences in configurational frustration type as compared to mutational frustration type. Many residues which are neither domain-domain interface residues nor associated with function showed differences in the frustration type of contact (S7B and S7C Fig). The differences suggest that when a protein domain tethers to another domain not only the stability of entire domain [6] or the folding rates differ as reported earlier [16] but the stability of the residues as well contact pairs changes for few cases. Since a larger number of contacts were observed to be configurationally frustrated (higher the configurational frustration index; more stable the conformation during the folding process) in comparison to mutationally frustrated, it hints that the domains may sample different conformations during the folding process in MD and ID, as have been reported earlier in literature for some multi-domain proteins [16–20].
To understand further the influence of tethering of domains on the function of proteins, a comparative analysis was performed for homologous domain pairs, where one member is a single-domain protein while the other member is a part of a multi-domain protein. Both the members are full-length gene products. Four pairs of proteins namely (a) phosphoribosylanthranilate isomerase from E. coli (PDB id: 1PII) and Jonesia denitrificans (PDB id: 4WUI), (b) cyclophilin from Bos taurus (PDB id: 1IHG) and Homo sapiens (PDB id: 3ICH), (c) sialidase from Micromonospora viridifaciens (PDB id: 1EUT) and Homo sapiens (PDB id: 1SO7) and, (d) hexokinase-1 from Homo sapiens (PDB id: 1HKC) and Saccharomyces cerevisiae (PDB id: 3B8A) were studied. The four domain pairs have sequence identity in the range of 27–56% with RMSDs in the range of 1.3–2.2Å (Fig 7). Since the homologous proteins differ in their amino acid sequences, only the dynamic properties of the protein were compared. The dynamics of the proteins were studied using normal mode analysis. For the comparative analysis, in-silico multi-domain chimeras of the single-domain proteins were generated. This was achieved by superposing the single-domain protein on the multi-domain protein, followed by in-silico removal of the homologous domain from the multi-domain protein and ligation of the domains. This in-silico protein will henceforth be referred as a chimera. For the hexokinase-1 protein, since the two-functional domains show gene duplication, the chimera was generated by superposing the single-domain on both the domains of multi-domain. Thus the two halves of the chimera of hexokinase-1 are identical. The flexibility and the cross-correlation coefficient of the functional residues were compared between single-domain proteins, multi-domain proteins and the chimeras for understanding the influence of tethering of domains on the function of proteins. Only topologically equivalent and identical functional residues of the homologous domain pairs were compared to minimize the influence of nature of residues.
Normalized square fluctuations of functional residues were compared between the single-domain and multi-domain proteins. The functional residues have lower flexibility (normalized square fluctuation < 0) in both single-domain and multi-domain proteins (Fig 8). Residues important for function or structural integrity are known to show lower flexibility [23]. Nonetheless, the flexibility of functional residues is lower in the multi-domain proteins as compared to the single-domain proteins (Fig 8). The flexibility of functional residues in the multi-domain protein and the chimera is observed to be similar (S8 Fig) except in the case of sialidase. This observation implies that increase in the rigidity of functional residues is a consequence of tethering of domains in multi-domain proteins. The differences in the flexibility of the functional residues can contribute towards differences reported in the functions of homologous protein domains, which are discussed later. To further understand the alteration in the dynamic properties of the domain, cross-correlation of the functional residues were studied. High correlation of motions was observed among functional residues for multi-domain protein in comparison to single-domain proteins (Fig 9, upper row). The single-domain proteins showed weaker cross-correlation among residues for all the cases (Fig 9, middle row). The cross-correlation between functional residues was comparable between the multi-domain and chimera for all the cases, except hexokinase-1 (Fig 9, lower row). The observations suggest that alteration in the synchronization of motion is a consequence of tethering.
For cyclophilin, the multi-domain protein is known to be less sensitive to cyclosporin as compared to single-domain cyclophilin [24]. Detailed analysis of cyclophilin single-domain protein showed the cyclosporin binding pocket shows low cross-correlation because of the closing movement of the pocket; but the multi-domain cyclophilin is superseded by domain-domain motion, where the functional residues move in the same direction resulting in high cross-correlation values (S9 Fig). This differential dynamics can provide a rationale for the lower sensitivity towards cyclosporin of the multi-domain protein in comparison to single-domain protein. The closing movement of the functional residues in the single-domain protein can hold the ligand better than the observed motion of the residues in the multi-domain protein. The single-domain hexokinase-1 protein has higher Km (300 μM) [25] as compared to multi-domain protein (32 μM) [26]. The glucose-binding pocket is at the interface of sub-domains for both multi-domain and single-domain protein. The sub-domain movement in single-domain protein is superseded by the domain movement in multi-domain protein. The low-frequency global motion in multi-domain protein allows better-synchronized motion of the binding pocket as compared to single-domain protein (Fig 9D). Weaker correlation between residues in single-domain hexokinase-1 as compared to multi-domain hexokinase-1 can explain the different Km, despite identical binding protein. From these analyses, we argue that tethering of domains influences the function of the constituent domains.
Chimera hexokinase-1 also exhibited an interesting feature. Though the structure and sequence of the two protein domains in the chimeric hexokinase-1 is identical, the domains exhibited different flexibility profile (S10A Fig). It has to be noted that while constructing the chimera of the yeast hexokinase-1, a stretch of 9 amino acids from the C-terminal of the first domain and a stretch of 9 amino acids from the N-terminal of the second domain were removed to relieve short contacts at the domain-domain interface region and linker region. To ensure that the differences are not observed due to this specific amino-acids deletions, the flexibility profile of the natural single-domain yeast hexokinase-1 (3b8a in S10B Fig) was compared with the in-silico generated model of the yeast single-domain hexokinase-1 with 9 amino acids deleted from the N-terminal (3b8a_N in S10B Fig) and the in-silico generated model of the yeast single-domain hexokinase-1 with 9 amino acids deleted from the C-terminal (3b8a_C in S10B Fig). The flexibility profiles were observed to be identical (S10B Fig), implying that the differences in the flexibility profile are only due to the tethering of domains and not due to deletion of the amino-acids. The observations suggest that the differences observed in the constituent domains of multi-domain protein depend on the order of the domain in the multi-domain proteins. The cross-correlation between the functional residues in the N-terminal and C-terminal domain also differs (S10C Fig). A number of positively correlated motions were observed in the C-terminal domain than in N-terminal domain. 6 pairs of functional residues viz. 173–210, 173–211, 174–210, 173–211, 176–210 and 176–211 exhibit anti-correlation motion in the N-terminal domain while the same residue pairs exhibit positively correlated motion in the C-terminal domain. We hypothesize that the differences in the nature of correlation of the fluctuation of the functional residues in the N-terminal and C-terminal domain may have given rise to the differential functional activity of the two domains in human hexokinase-1 at the first duplication event during evolution. The C-terminal of the human hexokinase-1 is catalytically active while N-terminal is catalytically inactive.
Conformational and structural alterations have been observed in proteins as they bind to other proteins [27, 28]. This line of thinking is extended in the current work to understand the structural, dynamic and energetic effects of tethering of protein domains in multi-domain proteins on the constituent domains. The extent of similarity between the physical and geometrical properties of protein-protein interaction and domain-domain interaction in multi-domain proteins [15] motivated us for the study. A dataset of 20 protein domain pairs of known 3-D structure has been used in the analysis. Each pair comprises of an entry with one or more domains of a multi-domain protein and the other entry has at least one additional domain tethered. Fifty percent of the protein domain pairs show differences in the global conformation on tethering. Rewiring of some intra-domain residue-residue contacts was observed in 12 protein domain pairs. Normal mode and molecular dynamics analyses of the domain pairs suggested that the flexibility of residues differs between domain in isolation and domain in multi-domain protein. Tight coupling of fluctuation was observed between residues in multi-domain proteins as compared to domain in isolation for all the domain pairs except one. These differences in the fluctuation and coupling of fluctuation are observed due to the shift from low-frequency local motion in isolated domain to low-frequency global motion in multi-domain systems. The stability of ~12% of residues and residue-residue contacts changed on tethering in all the domain pairs. Many of the differences in the intra-residue contacts, dynamics and energetics of the residues were observed without any significant structural deviation. These results strongly suggest that tethering of domains in multi-domain proteins influences the conformation, intra-domain residue-residue contact map, dynamics and the stability of residues and residue-residue contact of domains. Structural, dynamic and energetic differences were observed for many residues apart from domain-domain interacting residues in many domain pairs. These differences at regions spatially away from domain-domain interface could have allosteric origin; where the domain-domain interface region is the orthosteric site, the regions showing alteration are the allosteric site and the perturbation being tethering of domains. Allosteric alteration of proteins by altering the flexibility or correlated motion of the side-chains has been reported for some proteins [12, 29–32]. For example, the isolated WW domain and PPIase domain of human Pin1 protein has been shown to retain substrate binding and isomerase activity in vitro; but genetic studies showed that the WW domain is essential for in vivo Pin1 activation [12, 29, 30]. The WW domain regulates the activity of the PPIase domain by altering the flexibility and the extent of correlation of motion of side-chain of the three catalytic loops without much conformational changes [12, 29, 30].
To gain further insights on how tethering of domains influences the function of proteins, comparative dynamics analyses were carried out for 4 pairs of homologous domains, where a member in a pair is a multi-domain protein and the other member is a single-domain protein which is a homologue of one of the domains in the other protein in the pair. In each pair, only identical and structurally equivalent functional residues were analyzed. Functional residues were observed to be more rigid in all the multi-domain proteins than the single-domain proteins. This rigidity of functional residues is observed due to superseding of the low-frequency local motion of the single-domain protein by the low frequency global domain-domain motion in the multi-domain proteins. The low-frequency global domain motion alters the synchronization of residue-residue motion of functional residues in multi-domain proteins as compared to single-domain homologues. Differences in the catalytic activity reported for these homologous domain pairs can be a manifestation of these alteration in fluctuations. Combined with our observations on the identical domain pairs, it can be concluded that tethered domains in multi-domain proteins influence the function of domains by affecting the dynamics of the domains. Identical functional residues were observed to have different dynamics depending on the domain order, as exemplified by the chimera hexokinase-1 in our study. The N-terminal and the C-terminal domains of the chimera hexokinase are identical in sequence and conformation, but the flexibility and the synchronization between functional residues differ between the two domains. Similar observation was made by Kirubarkaran et. al. Artificial two-domain proteins were generated by fusing the natural protein domains PDZ3 and SH3 with five artificial domains. Observed differences in the fluctuation of the residues in PDZ3/SH3 domains were found to be dependent on the order of the domain construct for many cases [33]. These observations suggest that domains are not tethered during evolution at random but as a design to modulate the function of the constituent domains. Since dynamic alterations are observed in all the domain pairs; irrespective of the number of interface residues, size of the constituent domains, directionality of domain order or the fold (as defined in SCOP) of the domains (Table A and B in S1 Text), it can be concluded that dynamic allosteric regulation of domains is an intrinsic property of multi-domain proteins. This observation reinforces reports by others in literature that allostery is an intrinsic property of globular protein and allosteric regulation is prevalent in many multi-domain proteins [11, 34–36]. Alteration in the dynamics of the domain without any significant conformational difference by the tethered domain can be a great tool by evolution to modulate the function of same domain in different multi-domain proteins without altering the fold or structure of the domain, which otherwise can be an expensive process.
Alterations in the covalent structure of proteins such as post-translational modifications are known for causing changes in the conformation and/or nature of dynamics at the site of modification and around [37–40]. For example, phosphorylation of the activation loop of kinases such as cAMP-dependent kinase and CDK is well known to alter the conformation of the kinase extensively, enabling transition between inactive and active forms [41–43]. In our work, we considered pairs of identical domains, one in isolation and the other tethered to another domain. This pair can be viewed as though the domain in isolation is “modified” covalently in the other structure in the pair i.e. a domain and a domain linker region is covalently attached at one of N or C-terminus of the domain of interest. Clearly, this “covalent modification” in the terminus will have an influence on the structure/dynamics of the domain in the neighborhood of covalent attachment or possibly, even at a distant site. Interactions between the domain-domain linker and the flanking domains are common for all the examples studied in this work. Indeed, such interactions are present even in the examples where the direct domain-domain interactions are not present as the two domains are spatially well separated, for example cyclophilin and hexokinase-1. We believe that interactions between domain-domain linker and the domain of interest play a significant role in conferring alterations in structure, dynamics and correlated motions we observe in comparison with isolated domains. Since alterations in dynamics were observed independent of the number of amino acids in the linker (Table A and B in S1 Text), we believe that the effects depend on the presence of linker than the length of the linker. Role of linker residues in the allosteric communication between domains has been suggested by others as well in the literature [11, 33, 44–47]. All these observations suggest that the tethered domain and linker region can act as a scaffold for allosteric modulation of domains. The study presented here can be further exploited in designing new domain combination with desired activity.
Proteins in the datasets were structurally aligned using TM-align [51]. For the same-domain dataset, the structural variations were studied at the global and local level. Global Root Mean Square Deviation (RMSD) and Global Distance Test–Total Score (GDT-TS) [52] score were used to define global deviations. GDT-TS, henceforth mentioned as GDT, is used to define structural similarity between domains of identical sequences. Unlike RMSD it is largely insensitive to outliers arising especially due to differences in loop conformations. It is defined as the number of alpha carbons falling within a distance cut-off from the corresponding Cα of the other structure. MAXCLUSTER, an improved version of the maxsub algorithm [53], with a cut-off of 4Å was used for calculation of GDT score. High GDT scores are indicative of a low structural deviation between the proteins.
For studying structural variation at the local level, regions of residues that show significant structural deviation as compared to other regions of the structure were compared. For this, the distance between corresponding Cα atoms of the protein pairs after superimposing the structure onto each other was calculated. All the residues, whose distance between corresponding Cα(s) is more than twice the standard deviation from the mean of the distance distribution of all the residues, were identified as region showing significant structural deviation. For homologous protein domain pair, only RMSD has been calculated to quantify the structural differences.
To capture differences, if any, in residue-residue communication within proteins; undirected and unweighted networks of protein structures were constructed. The network was constructed for repaired structures (refer following section on dynamics). Each node in the network represents Cα and each edge represents the interaction between the nodes provided the distance between Cα atoms is less than or equal to 5Å. Network property namely communicability centrality was calculated using NetworkX [54] module of python. Communicability centrality quantifies the extent to which a node communicates with its neighbour. High communicability centrality measure of a residue implies its’ importance in inter-residue communication in protein structure. Numerically, it is the summation of all the closed walks of all lengths starting and ending at a node.
To study dynamics of domains, we have used two approaches namely normal mode analysis (NMA) and molecular dynamics. Crystal structures were energy minimized using GROMACS package [55] with conjugate gradient as the energy minimization method. Prior to energy minimization, the structures were repaired for missing residues and missing atoms. The missing residues were modelled using Rosetta 3.4 [56] and missing atoms were built using WHAT IF 10.1a algorithm [57]. Normal modes were calculated by generating coarse-grained anisotropic network model (ANM) for proteins, with 15 Å as the cut-off for connecting the nodes. Distance-dependent spring constants (the closer the nodes, stiffer is the edge) were used for the edges. Calculation of normal modes as well as the associated calculations and analyses were done using the ProDy package [58]. For the analyses, only the normal modes contributing to 80% variations were studied and fluctuation values contributed by first five N-terminal residues and last five C-terminal residues were removed. Furthermore, correlation of fluctuation between each residue pairs, termed as cross-correlation, was compared. The similarity between cross-correlation matrices has been measured using distance independent measure called Rv coefficient [59]. Rv coefficient measures the closeness of a set of points represented as a matrix. It is a multivariate generalization of Pearson correlation coefficient.
Molecular dynamics was performed to study the correlation of fluctuation at the all-atom level for 3 pairs. Molecular dynamics was performed using GROMACS package. The proteins were simulated using Charmm 27 force field [60] and SPC water model [61] in a dodecahedron box. The system was energy minimized using steep descent after addition of appropriate counter ions to balance the charges. The system was appropriately equilibrated for 100 ps using V-rescale and 100 ps using Parrinello-Rahman. The final production run was performed once for 400ns.
Energetics calculation was performed only for a dataset of identical protein domains. As the homologous protein domains differ in sequence identity, it is futile to compare their energetics. Frustratometer algorithm [22] was used to perform the energetics calculation. The algorithm systematically perturbs each residue and contact to generate the decoys and compute energy according to Associative Memory Hamiltonian with Water-mediated interaction energy function (AMW) [62]. Then the energy of the native protein is compared with the energy distribution of the decoys to calculate the frustration index, which is the Z-score of the energy of the native with respect to the decoys. A residue or a contact is considered as minimally frustrated if the frustration index is greater than 0.78, highly frustrated if the frustration index is less than -1 and neutrally frustrated if frustration index is in between -1 and 0.78 as defined in [22].
AMW is a coarse-grained energy function where the backbone is represented as Cα, O and the side chain is reduced to Cβ, the position of N and C is generated considering the ideal geometry of the peptide bond. AMW energy function consists of five non-local energy terms namely Lennard-Jones 6–12 potential, H-bond potential, compactness potential, burial potential and water-mediated interaction potential. A pair of amino acids is considered to form a contact if the inter Cα distance is less than or equal to 5Å. Each contact is perturbed either by mutating each interacting residue pair to every other amino acid pair but keeping all other interaction parameters same as the native structure. Then the effective energy of the native contact is compared with the decoys to access the energetic stability of the contact to mutation. So, it provides a qualitative measure of the energetic feasibility of mutation of such contacts. The frustration index calculated by this method is termed as mutational frustration index. Another way of perturbing the contacts is by displacing the location of each contact thus sampling the possible configurations which can be taken by the contacts during folding. The frustration index calculated in such a way is termed as configurational frustration index. Similar to contacts, each residue is perturbed to every other amino acid and other configurations to evaluate the stability of residue in the native structure to all these perturbation. The frustration index calculated by this method is termed as Single Residue Level Frustration (SRLF).
All the statistical analyses were performed using R package.
|
10.1371/journal.pcbi.1006761 | Phylogenies from dynamic networks | The relationship between the underlying contact network over which a pathogen spreads and the pathogen phylogenetic trees that are obtained presents an opportunity to use sequence data to learn about contact networks that are difficult to study empirically. However, this relationship is not explicitly known and is usually studied in simulations, often with the simplifying assumption that the contact network is static in time, though human contact networks are dynamic. We simulate pathogen phylogenetic trees on dynamic Erdős-Renyi random networks and on two dynamic networks with skewed degree distribution, of which one is additionally clustered. We use tree shape features to explore how adding dynamics changes the relationships between the overall network structure and phylogenies. Our tree features include the number of small substructures (cherries, pitchforks) in the trees, measures of tree imbalance (Sackin index, Colless index), features derived from network science (diameter, closeness), as well as features using the internal branch lengths from the tip to the root. Using principal component analysis we find that the network dynamics influence the shapes of phylogenies, as does the network type. We also compare dynamic and time-integrated static networks. We find, in particular, that static network models like the widely used Barabasi-Albert model can be poor approximations for dynamic networks. We explore the effects of mis-specifying the network on the performance of classifiers trained identify the transmission rate (using supervised learning methods). We find that both mis-specification of the underlying network and its parameters (mean degree, turnover rate) have a strong adverse effect on the ability to estimate the transmission parameter. We illustrate these results by classifying HIV trees with a classifier that we trained on simulated trees from different networks, infection rates and turnover rates. Our results point to the importance of correctly estimating and modelling contact networks with dynamics when using phylodynamic tools to estimate epidemiological parameters.
| Understanding whether and how transmission patterns are revealed by branching patterns in phylogenetic trees for pathogens remains a challenging research question. Besides the diversification of the pathogen, branching patterns depend strongly on the host contact structure as it shapes opportunities for the pathogen to reproduce. However, the host contact network is often difficult to study, in particular as it evolves in time. In this paper we perform a simulation study on three different dynamic networks, on which we simulate transmission trees. We use a simple Erdős-Renyi random network and two more realistic networks with skewed degree distribution, where one is also clustered. We convert transmission trees into phylogenetic trees and analyze them with different tree statistics like imbalance measures, counts of small substructures, and measures containing the branch lengths. We study the tree features with principal component analysis and with supervised learning methods, and find that network dynamics and network type can strongly influence the shape of phylogenetic trees. This implies that using phylogenetic trees from a mis-specified network type and dynamic can lead to poor phylodynamic estimation of transmission parameters. We illustrate this with HIV phylogenetic trees constructed from viral sequences of patients in the Dutch ATHENA cohort, and from sequences of the Los Alamos Sequence database.
| Understanding whether and how the transmission patterns of a pathogen are revealed by branching patterns in pathogen phylogenetic trees remains a challenging research question. Alongside the stochastic diversification of the pathogen on the short time scales of an infectious disease outbreak, branching patterns in the pathogen’s phylogenetic tree also depend strongly on the underlying transmission pattern [1] and the host contact structure, as these shape the pathogen’s reproductive opportunities.
The role of networks in epidemic spreading has been studied extensively in past decades [2–12]. The topology of the host contact network plays a crucial role in setting the epidemic threshold, the epidemic size and the most effective interventions. Network properties also play a role in determining which individuals are at high risk of infection. Naturally, modellers seek to inform simulated networks with individual-level data from real populations. Respondent-driven sampling [13, 14], snowball sampling or questionnaires [15] are several approaches to gathering these data, but all are challenging: people do not always remember how many people they have been in contact with, and in some contexts (such as injection drug use or sexual behaviour), contact is stigmatized or even illegal. As a result, individuals may not wish to report contacts to public health practitioners.
Recently there has been interest in using genetic data from pathogens, together with phylogenetic and phylodynamic tools, to estimate the parameters of human contact networks [16–19]. This is appealing, in that data now accessible with high-throughput sequencing technologies (pathogen sequences, at a level of resolution that makes detecting even small amounts of genetic variation feasible) can reveal information about a fundamental population-level structure (the network). Sequences can show patterns of descent, and pathogens transmitted directly from human to human need human contact networks to have descendants. Since networks are difficult to observe directly and phylogenetic trees in principle contain some information about them, researchers have used a variety of tools to relate pathogen phylogenetic trees to the underlying contact network’s degree distribution, connectivity and clustering [17, 20]. This method has been of particular interest for HIV phylogenies [21–24].
Studies have reported varying strengths of the effect of the contact network on the phylogeny. For example, [25] found a very weak influence of the network’s clustering coefficient when the degree distribution is held constant, [26] studied the shapes of phylogenies from simulated genetic data and found a moderate influence of the underlying network degree distribution, though “clustering” in phylogenetic trees did not parallel the heterogeneity in the degree distribution, and network dynamics shape phylogenies as well. [21] found a relatively stong effect of the variance in degree distribution and of the average pathlength of the network on the shapes of phylogenies. Also, within-host viral diversity affects the link between network structures and phylogenies [23], as do the basic reproduction number and other details of the process [27, 28]. It is therefore reasonable to assume that details of timing of infection, in-host selection, selection at the population level and other factors may also affect the relationship between contact networks and phylogies.
Human contact networks are self-organizing systems with certain general characteristics; one approach to modelling human host networks is to perform simulations that are able to reproduce those characteristics. Key characteristics include a short average pathlength (small-world property) [29], clustering [30] and a scale-free (or at least highly skewed) degree distribution [31, 32]. In particular, networks with a skewed degree distribution have received much attention for epidemic spreading, as they yield significantly different transmission patterns from a homogeneously mixed population. Depending on the transmission pathway, there is evidence that networks can have an exponential degree distribution [13, 33] or a scale-free degree distribution, found in various social networks [34–36], and in human contact networks [37–39]. The Barabasi-Albert model [40] in particular is a much-studied process by which scale-free degree distributions may emerge. It is based on the idea of preferential attachment: nodes attach preferentially to existing nodes that already have many links.
Preferential attachment is a plausible rationale for many applications (fame, publicity). It describes a constantly growing network, or a static network if the growth is halted. In contrast, human host contact networks are often dynamic, but may not be growing in size over time. Instead, they have population turnover [5, 41], with individuals entering and leaving a network as time goes on. Especially for chronic infections like TB, HCV or HIV [42], people may enter and exit the network over shorter timescales than the length of the infectious period. The number of contacts that individuals accumulate over time is significantly larger than the number of contacts at one point in time.
Furthermore, many of the observations underlying reports of scale-free degree distributions in human contact networks are derived from reports of the cumulative numbers of contacts that individuals have over a long period (for example over one year [32, 43], or accumulated to date). Accordingly, it may not be appropriate to compare simulated transmission dynamics in models where individuals’ degrees are modelled from observed accumulated numbers of contacts to transmission where degrees are taken as the instantaneous (or even shorter-term) numbers of contacts. The static network (with degrees modelled on data for the number of contacts accumulated over long time periods) can be a very poor approximation of the true dynamic network; outbreaks can spread faster in such a static network due to the potentially very high numbers of simultaneous contacts.
In using phylodynamic tools to estimate network parameters from pathogen phylogenies, it is typically assumed that the contact network is static in time; one seeks network parameters that produce pathogen phylogenetic trees that are similar to observed trees, conditional on the static network assumption (and perhaps also on assumptions about the degree distribution, clustering patterns and other network attributes). Whatever the details, inferred quantities such as degree distribution, the average number of partners and the infection rate are influenced by assumptions about the network, including the static assumption.
The duration of infectiousness and the time scale of the network dynamics must affect the relationship between pathogen phylogenies and network parameters. Clearly, no individual has thousands of contacts over a week; reports of degrees that are orders of magnitude higher than the average are from data aggregated over long time periods; where an infectious duration is of the order of weeks or a few months, the scale-free property is unlikely to hold. These issues are presented briefly in [26] and [44, 45].
In this paper, we investigate the effect of human host network dynamics on pathogen phylogenies. Our study focuses on simulations, and on the relationship between network assumptions and estimates of transmission parameters. We compare simulated phylogenies from outbreaks on static and dynamic networks, and we explore the effect of the turnover rate at which individuals enter and leave the system. We also study the effect of the network characteristics on the phylogenies. For this, we use networks with binomial degree distribution and skewed degree distribution, as well as clustered and unclustered networks. We explore the effect of the infection rate and the mean number of contacts. We study how the features of the underlying networks affect phylogenetic trees with various tree statistics. Finally, we turn to phylodynamic inference of HIV transmission parameters and illustrate our main results using HIV sequence data from the Dutch ATHENA cohort and Los Alamos. In particular, we characterise the impact of alternative assumptions on human contact network dynamics on estimation of key transmission parameters including R0.
We simulate the human contact network with the algorithms described in section. First, we allow the networks to converge to a stationary state in terms of degree distribution; in this stationary state, networks are still dynamic in the sense that people enter and exit. Then, an outbreak is simulated on the networks while they continue to evolve. One person is infected and, with a constant infection rate per contact, the infection can spread. The resulting infection trees are converted into a phylogenetic trees (see section). Unlike the Barabasi-Albert (BA) model, our approach allows a skewed degree distribution to emerge while keeping the size and total degree fixed. Throughout, individuals enter and leave the network and links are formed and dissolved. In contrast, in the BA model, nodes and links are continuously added and remain in the network. We set a constant number of tips in our trees. We use tree shape and length statistics, detailed in section, to compare phylogenetic trees.
We use an algorithm for a “skewed-clustered” network which generates a network with a skewed degree distribution and positive transitivity [38]. To understand what these features add, we also use skewed (but not particularly clustered) networks, and an Erdős-Renyi random network. These all have a stationary average number of contacts and stationary degree distribution, while people are entering and exiting the network. This entry and exit happens with a turnover rate δ, which is the ratio between the number of people entering per time step to the number of people in the network. Networks are simulated in discrete time. In each time step the following steps happen:
In our simulations, we begin with one infected individual who then infects neighbours at a constant infection rate per contact, after which the neighbours can infect their respective neighbours in the next time step, and so forth. Infected individuals stay infected throughout the simulation, modelling a long-term infection. This simulates an outbreak on these dynamic networks. There is at least one time step between an individual becoming infected and infecting a neighbour, and we model a positive time between any two infection events by adding a small positive time to the infection events of one iteration, such that they occur with equal time lapses.
We extract what would be the “true timed phylogeny” of the pathogen given the transmission tree in our network, under the assumption that hosts carry a single pathogen lineage. To do this we form a binary branching tree in which each host corresponds to a tip in the phylogeny and branch lengths correspond to time. Since we know the true transmission tree and its timing, this can be done by tracking the infectors, infectees and the time between infection events. This is available in the getLabGenealogy function in the R package PhyloTop [47]. The simulation of the outbreak is stopped after a time such that the phylogenetic trees all have the same number of tips.
We compute features of the phylogenies with software sources listed in Table 1.
Number of substructures
Cherries: Substructure consisting of two tip descendants
Pitchforks: Substructure consisting of three tips
Imbalance measures
Sackin Index: Average number of internal nodes Ni between each tip i and the root of the phylogenetic tree S n = 1 n ∑ i = 1 n N i, [48, 49]
Colless Index: It compares the number of tips that descend on the left and right (L and R) from each internal node, and averages over these differences |L − R| [49, 50].
Other tree measures
Maximum Height: Maximum height of tips in the tree.
Average Size of Ladder: Ladder structures [1] consisting of a connected set of internal nodes with a single tip descendant
IL numbers: Number of internal nodes with a single tip child.
Centrality measures and general network measures
Maximum Betweenness: Maximum number of shortest paths that pass through a particular node.
Wiener Index: Sum of the lengths of the shortest paths between all pairs of nodes.
Maximum Closeness: Sum of lengths of the shortest paths between one node and all other nodes (maximum thereof).
Average Pathlength: Average distance between two nodes.
Diameter: Longest possible path between two nodes in the tree.
Tree measures that use the edge length
Branching next index (BNI): We compare the extent to which a node that branches at time t is chronologically next to branch; in other words, does branching now make it more or less likely that a node will branch next? If a node’s child is chronologically next to branch following the node itself, we say the node has the ‘branching next’ property (si = 1). We add and rescale the sum of si over all internal nodes i in the tree (except the root and the last node to branch). si is a Bernoulli random variable whose expected value is pi = 2/ki, where ki is the number of lineages in the tree that exist at time ti + ϵ, in the limit as ϵ → 0, where ti is the time of node i and ϵ > 0. We define the BNI as ∑ i s i - p i ∑ i p i ( 1 - p i )
Generalised branching next (MNI): Extending the BNI concept, we ask whether one of the next m branching events (chronologically) in the tree descends from the current node, in which case we set di = 1 for node i. We sum and rescale di, as with si, over the tree to create this summary statistic. We let kij, j = 1, …, m be the numbers of lineages immediately after the j′th branching event following node i (in the entire tree). We define qi = ∏j(1 − 1/kij) and normalise by setting MNI to ∑ i d i - q i ∑ i q i ( 1 - q i ). Since now they are not independent we use every m′th node i rather than every node.
Length statistics We use the mean of the path length from the internal nodes of the tree to its root, as well as the median, variance, skewness and kurtosis of this set of path lengths.
We use two approaches to understand how the underlying contact network affects the tree features. The first is to visualise the results using principal components analysis (PCA) on the matrix of features described above. The matrix values are scaled such that the mean is zero, and normalized such that variance is 1, as is standard in PCA. This visually illustrates the extent to which these features discriminate between phylogenetic trees derived from different contact networks. However, visual separation on a 2-dimensional PCA plot is a limited measure of how informative the features are of the contact network. Thus, we also explore this quantitatively using both K-nearest neighbours and random forest classification. We attempt to classify the network (random, skewed or skewed-clustered) based on the features. We assess accuracy in these binary and categorical classifications when the underlying network model correct, and when it is mis-specified. We also attempt to classify the transmission rate. For this goal we use trees from simulated outbreaks where we distributed the transmission rate β uniformly. We grouped these trees into bins depending on the underlying β and train classifiers on the tree features with the aim of predicting the bin of β for a test set. We study a scenario where turnover rate δ and mean degree d ^ are distributed uniformly, and a scenario where they are kept constant.
Partial nucleotide HIV-1 polymerase sequences were obtained as described previously from patients in the ATHENA national observational HIV cohort in the Netherlands (by June 2015) [52]. We used the first sequence per patient, with a minimum of 750 nucleotides length. No patient information was included in the analysis. Sequences were aligned with Clustal Omega 1.1.0 [53] and manually checked and adjusted. HIV-1 subtyping was performed with COMET v1.3 [54] and 6912 subtype B sequences were considered for further analysis. In addition we retrieved 19,459 HIV-1 subtype B sequences from the Los Alamos database (by September 2017) [55], with a minimum length of 1000 nucleotides overlap to the ATHENA alignment. Excluding sites with less than 75% coverage, and with IAS resistant mutations 2015 removed This resulted in a sequence alignment of 1,128 nucleotides length [56]). Viral phylogenies were reconstructed with FastTree version 2 [57].
From this tree we identify 90 non-intersecting clades in the specified size range 100-151, using a depth first search approach. The mean number of tips in the clades was 127. 86 out of 90 clades contained samples from the ATHENA cohort, with a fraction between 0.01 and 0.97. Overall, the clades we extracted contained 8326 sequences from the Los Alamos data and 3186 from the Dutch HIV-1 ATHENA cohort. We compared the HIV clades with simulated trees from different networks and to trees simulated on the same network, but with varying infection rates. We trained random forest and K-nearest neighbour classifiers on tree features from the simulated networks, and used the features from the HIV clades as a test set. The simulated trees (the training set) had 100 tips. We then used the classifiers to predict the network type or infection rate for the HIV clades.
We used principal component analysis to study different types of networks, different mean degree and infection rate for a given network, as well as different turnover rates and a time-integrated static network (see all scenarios in Table 2). We also trained classifiers on the networks in order to predict infection rate, turnover rate and network type (see all scenarios in Table 3).
The network structure and dynamics both affect features of phylogenetic trees of pathogens spreading on the networks. However, the effects are modulated by the transmission rate and the turnover rate. These relationships are sufficiently strong as to disrupt the signal of the network type in the pathogen phylogeny. A summary of results for the different network structures is given in in the discussion and the trees are given in supporting information.
Fig 2 shows a principal component analysis based on phylogenetic trees simulated on dynamic networks with three different topologies. Phylogenies from the Erdős-Renyi network differ strongly from the two others. This holds even for relatively small trees (100 tips), whereas for clustered and unclustered networks, the discrimination improves with the size of tree (up to 250 tips). The same results hold for a wide range of infection rates (β = 0.025 to β = 0.2) and higher turnover rates (δ = 0.1). Overall, the discrimination between networks improves with tree size. The distinction between trees from different underlying networks improves if additional features are used that take into account the lengths of edges. Skewed and skewed-clustered network have a lower number of small substructures (cherries and pitchforks), and a higher value for all imbalance measures. Most network measures (except betweenness) are also positively correlated with imbalance measures.
The network structures become more distinct with a higher rate of infection per contact and with a higher rate of turnover (eg β = 0.2, δ = 0.1), and in particular the numbers of cherries and the path lengths become more distinct as these parameters increase. Differences in the path lengths and the imbalance between the networks are also more pronounced with higher β and δ. In contrast, however, there are a few features for which differences are more pronounced at low infection rates (including the ‘ILnumbers’ and the Wiener index for clustered vs unclustered networks). In other words, given fixed values of the transmission and turnover rates, it is possible to separate, and estimate, the underlying network structure based on phylogenetic tree features, for example by discriminant analysis, classification methods, or by Approximate Bayesian Computation.
However, the details—which phylogenetic features point to which kinds of networks—are specific to the transmission and turnover rates, and mis-estimation seems likely if these are mis-specified. Furthermore, for some choices of parameters, the networks are no longer well-separated in the PCA analysis; for example, if β = 0.05 and δ = 0.1 (so β < δ), the clustered network overlaps with the random network, whereas if β > δ, they do not overlap, but the two skewed networks (clustered and unclustered) begin to overlap.
When infection rate per contact β increases, so does the variance of tree features, and the following tree features increase on average: Colless index, Sackin index, IL numbers (nodes with single tip child), average ladder size, maximum height, average pathlength, Wiener index and diameter. The number of cherries, pitchforks and maximal closeness decrease with increasing infection rate, as shown in Fig 3 for the skewed-clustered network.
The same features increase as the mean degree increases (red and green vs. turquoise and purple in Fig 3), which is expected, as both increasing β (infection rate per contact) and increasing the number of contacts increase the basic reproduction number R 0 = β d ¯ τ (τ being the duration of infection and d ¯ the median degree) of an outbreak. The phylogenies from the four outbreak hypotheses in Fig 3 may therefore correspond to different pathogens or to a pathogen in rather different epidemiological settings, as in these scenarios R0 values may differ substantially. However, the tree features that discriminate these scenarios are also affected by the nature of the contact network (Fig 1) and by the turnover rate (Fig 4). This comparison highlights that the network type and turnover are likely to affect estimation of the mean degree and the infection rate from phylogenetic trees.
Simulated trees to figure 4 are found in S3 File.
Fig 4 shows a PCA of phylogeny features derived from skewed-clustered networks with same mean degree but different turnover rates (i.e. rates at which people enter and exit the system), and from a time-integrated static network of same mean degree d ^. Higher population turnover of the network increases the following features of the simulated phylogenetic trees: Sackin index, Colless index, average ladder sizes, IL number, maximum height, average pathlengths, diameter, Wiener index, and betweenness, and decreases the number of cherries and pitchforks as well as maximum closeness.
Higher turnover gives similar results to a higher mean degree or a higher infection rate (see Fig 3). The static time-integrated network has no turnover, but contacts have a longer duration, presenting the opportunity to transmit comparably to a dynamic network with much higher turnover than the one used for the time integration. In dynamic networks, links get rewired often and therefore many opportunities for transmission exist. The static network has higher mean degree as the temporally existing links are accumulated (see Fig 4). Instead of resembling those from very low turnover, the phylogenies from static networks have therefore features similar to those from networks with very high turnover.
This effect holds for different infection rates β, but the higher the infection rate, the more the phylogenies from a time-integrated network differ from those from networks with low turnover.
Results for varying infection rate, mean degree, turnover and time-integration are qualitatively the same for the skewed-clustered and skewed-unclustered network, but since the unclustered network has shorter average pathlength than the clustered network of same mean degree, the effects are more pronounced.
Imbalance measures are always anticorrelated with the counts of small substructures (pitchforks and cherries). The fact that network skewness increases tree imbalance (and decreases substructures) could be due to the fact that high heterogeneity in the network degree is passed on to high heterogeneity in the number of secondary infection, resulting in an imbalanced tree (measured e.g. by Sackin and Colless index). On the other hand, increased network clustering may have the opposite effect, as it results in fewer nodes being connected to hubs in the network, which may cause the infection tree and resulting phylogenetic tree to be more balanced and to exhibit more pitchforks and cherries. However, an imbalanced phylogenetic tree could in principle also result from long chains of person-to-person transmission, in which each individual infects exactly one other: imbalanced trees do not necessarily require heterogeneous contact numbers or heterogeneous numbers of secondary infections.
For simulations with distributed values for β, δ and mean degree of the network, we calculated all of our features of phylogenetic trees and used these to train classifiers, which we then tested. We used K nearest neighbours (KNN) [58] which classifies an object based the the class of the majority of its nearest neighbours, and random forests [59] which use decision trees to classify the test data.
We simulated 1549 phylogenetic trees on the three types of networks, with random uniformly distributed values of the turnover and transmission rate parameters (both in [0.05, 0.15]) and mean degrees (in [4, 9]). We trained classifiers on 1040 instances to classify from which type of network a phylogeny was derived. We compute the mean and standard deviation of the accuracy using 10-fold cross-validation. The classification is successful in the sense that it is possible to classify the dynamic network type based on the phylogenetic features, given a range of transmission parameters and turnover rates in the training data. Table 4 lists the results when we choose the key parameters β (transmission rate), mean degree and turnover δ uniformly at random over the specified ranges. Both classifiers predict the network type with high accuracy, using the phylogenetic features. This means that even with the additional complications of dynamic networks and unknown underlying parameters, phylogenetic trees encode information about the nature of the network.
We also asked how varying the underlying (and in general unknown) dynamic contact network would affect estimation of the transmission parameter β (also in Tables 4 and 5). Estimation of β is much worse than estimation of the network, and strongly depends on the assumed network. The performance is best for random forests with either all three networks present in the data (accuracy 0.47) or with a single, correctly-specified, skewed or random network used to train the model (accuracy 0.55, 0.44 respectively). Mis-specification of the network worsens predictions.
Discrimination between skewed and skewed-clustered networks remains difficult, as these networks are quite similar. The difference between skewed and random networks is more pronounced (as also seen in the PCA analysis in Fig 2). In that sense our results are similar to the results in [60–62], who successfully predicted contact rates with Approximate Bayesian Computation (ABC) on static networks, where the phylogenetic trees separate well in a PCA plot of extracted tree measures.
Given the poor ability to predict β when the mean degree and turnover are randomly sampled, we explored whether keeping these parameters fixed would improve the estimation: if we knew these parameters and had pathogen phylogenies, would we then be able to estimate the transmission rate in the context of dynamic networks? Here, the accuracy is only good in the case of the random network (0.7, 0.82 for KNN, random forests respectively). Random forests give consistently slightly higher accuracy, with an accuracy over 0.5 where (1) all three networks (skewed, skewed-clustered an random) were present in the training data, or (2) the model was trained on the skewed or random networks. If the network is mis-specified or skewed, neither approach is able to predict β. We suggest that this may have adverse consequences for analyses using static or other assumed network models in phylodynamics; these may draw erroneous conclusions about the rate of transmission or other parameters due to mis-specification of the underlying network.
We trained classifiers on phylogenetic trees simulated with different network hypotheses, in order to predict the network type for HIV clades from sequences of patients in the Dutch ATHENA cohort and from sequences of the Los Alamos Sequence database [55]. The Dutch sequences predominantly capture the Dutch national HIV epidemic (cite Bezemer PLoS Med), whereas the sequences in the Los Alamos database are from cases worldwide and capture many diverse HIV epidemics. Our network predictions are consistent with this: the higher the fraction of tips from the Netherlands, the more HIV trees are predicted to arise from skewed or skewed-clustered networks, rather than random (see Table 6); this signal is consistent in the K-nearest neighbour and random forest classification.
We also trained the classifiers on simulated trees from a skewed-clustered network with two different infection rates (β = 0.05 and β = 0.2), in order to predict the infection rate for the HIV trees (see Table (7). We did the latter both with trees from static networks and dynamic networks with turnover rate δ = 0.1. For the static network, roughly two thirds of the HIV trees are predicted to have infection rate β = 0.05 and one third β = 0.2. In contrast, all of the HIV trees are predicted to have the higher infection rate of β = 0.2 on the dynamic network.
It is not surprising that more HIV trees were predicted to have the higher infection rate β = 0.2 when the classifiers were trained on the dynamic network. On dynamic networks, not all links are present at any moment, which slows down the outbreak. A higher infection rate could compensate to attain the same R0. This result was very robust even when fewer tree features were used to train the classifier. However, if only imbalance measures were used, a low fraction of HIV trees were predicted to have β = 0.05 by dynamic-network-based classifiers. This suggests that using a variety of tree features is important for specification of network parameters from phylogenies.
We have also listed separate predictions for clades in which more than 50% or 70% of the tips are from the ATHENA dataset; these are geographically linked, may include more recent transmission and are likely to have a higher sampling density than background clades from the Los Alamos database. Compared to the whole set of 90 HIV clades, these clades are more likely to be classified to have come from a skewed (clustered) network and to have a high transmission rate (β = 0.2). However, the certainty on this prediction depends on the underlying network assumption, with classifiers trained on dynamic models showing a completely consistent set of predictions while those trained on static models leave considerable variation (Table 7). In contrast, clades with fewer Dutch sequences were classified predominantly to have a lower transmission rate if classifiers were trained using static networks, but a higher transmission rate using dynamic networks. The fact that the results differ considerably depending on the underlying network assumption indicates that a mis-specified network, via an incorrect turnover rate or indeed the assumption of a static network, can have a strong effect on predicted transmission rates.
We used models of different human host contact networks to simulate outbreaks of pathogens, and convert the infection trees into phylogenetic trees. We showed that it is possible to discriminate with tree statistics between different contact network hypotheses, different turnover rates, different mean degrees and different infection rates. Table 8 sumarizes the network effect on tree statistics. The underlying contact network hypothesis (random, skewed or skewed-clustered) is clearly identifiable in statistics of the simulated phylogenetic trees, if β and δ are the same. This indicates that simple networks such as the Erdős-Renyi model are likely to be unsuitable models for human host contact networks where there is evidence for a skewed degree distribution and clustering.
Nevertheless, in our simulations, phylogenies from skewed-clustered networks are slightly more similar to those from random networks than those from unclustered networks of the same degree distribution. Phylogenetic trees from outbreaks on the same static network, but with different infection rates or different mean degrees, can be distinguished clearly in PCA plots. This result holds also on dynamic networks, and suggests, in keeping with previous work, that phylogenetic tree features can be used to estimate epidemiological parameters. However, the relationships between the epidemiological parameters, networks and phylogenetic trees are complex. We tested the strength of some of these relationships using supervised learning methods, and found that both network mis-specification and variability in other parameters (modelling uncertainty about the values of these parameters) have a strong impact on the ability to estimate the transmission parameter. Our results indicate that consistent network mis-specification and parameter uncertainty may have an adverse impact on phylodynamic studies estimating parameters from data.
Population turnover in dynamic networks has a measurable effect on pathogen phylogenies; phylogenetic tree features can discriminate between different turnover rates at which the underlying network is evolving. Overall, the higher the turnover, the higher the imbalance measures and the lower counts of small substructures. No single feature captures the differences between contact network hypotheses entirely, and a combination of many different features yields the best visual separation between the groups in a PCA plot. Features that take into account the branch length of the phylogenetic trees improve the separation slightly. Very different patterns are obtained from a static time-integrated network as compared to dynamic networks, on which transmission happens slower. This suggests that in the phylodynamic setting, static networks are a poor approximation for dynamic networks, highlighting the need for dynamic network models. This also highlights the need for investigating turnover and dynamics in empirical networks to obtain the data necessary to develop dynamic models. We illustrated this result by predicting the infection rate β of HIV trees, and showed that the predictions strongly underestimate β if a static network is used instead of a dynamic one. Comparison to HIV data also showed that clades with tips predominantly from the Dutch sequence dataset with high sampling fraction of infected individuals are more likely to be predicted to have come from a skewed or skewed-clustered network than those with tips mainly from the even sparser sampled Los Alamos database.
Although the dynamic skewed-clustered network is likely to be a more realistic approximation to real networks than static or unclustered networks, it still might not be as clustered as a given real contact network. The details of the relevant network for a study of real data will depend on the pathogen and also on the nature of the community in which that pathogen is being studied. The dynamic models we have used here are still relatively simple and tractable, and real networks are likely to be even more heterogeneous.
|
10.1371/journal.pbio.0050210 | Precise Positioning of Myosin VI on Endocytic Vesicles In Vivo | Myosin VI has been studied in both a monomeric and a dimeric form in vitro. Because the functional characteristics of the motor are dramatically different for these two forms, it is important to understand whether myosin VI heavy chains are brought together on endocytic vesicles. We have used fluorescence anisotropy measurements to detect fluorescence resonance energy transfer between identical fluorophores (homoFRET) resulting from myosin VI heavy chains being brought into close proximity. We observed that, when associated with clathrin-mediated endocytic vesicles, myosin VI heavy chains are precisely positioned to bring their tail domains in close proximity. Our data show that on endocytic vesicles, myosin VI heavy chains are brought together in an orientation that previous in vitro studies have shown causes dimerization of the motor. Our results are therefore consistent with vesicle-associated myosin VI existing as a processive dimer, capable of its known trafficking function.
| Myosin VI is a molecular motor implicated in diverse cell processes, including trafficking endocytic vesicles into the cell, transporting proteins to the leading edge of a migrating cell, and anchoring stereocilia to the hair cells of inner ear sensory epithelia. The motor has been studied in both a monomeric and dimeric form in vitro and is reported to exist as a monomer in the cytoplasm of cells. Because the functional characteristics of the motor are dramatically different for these two forms, an understanding of the activity of myosin VI requires an understanding of its functional form in vivo. To probe the role of myosin VI in vesicle trafficking, we labeled myosin VI truncations with a fluorescent protein and studied the positioning of these constructs on endocytic vesicles. We observed nonradiative transfer of energy between the fluorescent proteins, a process that can only occur if they are brought extremely close together. Our results indicate that, when myosin VI heavy chains bind to endocytic vesicles, they are precisely positioned very close together. Work from other laboratories indicates that myosin VI heavy chains brought together in this manner are capable of dimerization. Our results are therefore consistent with vesicle-associated myosin VI existing as a processive dimer, capable of myosin VI's known trafficking function.
| Class VI myosins are found in a variety of organisms from Caenorhabditis elegans to human, and in a variety of cell types (reviewed in [1,2]). Unlike other characterized myosins, they move toward the pointed end of an actin filament [3], and so are capable of functions unique from other myosins. For example, during clathrin-mediated endocytosis, myosin VI is implicated in trafficking vesicles that have recently shed their clathrin coat, denoted uncoated vesicles (UCV). The motor transports UCV from the periphery of a cell to its interior, presumably along actin filaments in the cell periphery that are oriented with pointed ends directed toward the cell interior [4–6].
The motor's heavy chain contains an N-terminal catalytic head followed by a unique myosin VI insert and an IQ motif, each of which can bind a single calmodulin [7,8]. The calmodulin binding domains are followed by a tail domain (TD) that is predicted to be highly α-helical. The C-terminal domain is the motor's cargo-binding domain (CBD), a region implicated in association of the motor with its protein cargo [9–11] (Figure 1A).
Myosin VI heavy chains have been hypothesized to dimerize [7,12,13]. This model is supported by single-motor optical trap assays that utilized a motor construct containing a GCN4 leucine-zipper domain in the C-terminal region of the TD, ensuring dimerization of the motor even under the dilute (pM) conditions of single-molecule assays [14]. This dimer walks processively along actin, meaning it takes numerous steps along a filament before dissociating [15]. Its stepping is highly coordinated, with mechanical strain regulating the biochemical behavior of the molecule, resulting in head-to-head communication and proper in vivo function [16,17].
Surprisingly, however, Lister et al. [18] demonstrated that the myosin VI heavy chain, when purified from a baculovirus expression system or observed in extracts from rat kidney fibroblastic tissue culture, exists as a monomer. Though an ensemble of monomeric motors may be capable of myosin VI's predicted trafficking function, such an ensemble is not ideal for trafficking because the motor has a high duty ratio [16], and so monomers attached to actin would work against newly attached and stroking motors. On the other hand, as a coordinated processive dimer, the motor would be well suited to traffic cargo efficiently with relatively few motors, as demonstrated by in vitro studies of a myosin VI dimer [15,17]. We thus speculated that, in regions of the cell where myosin VI performs trafficking function, dimerization of the motor occurs in a regulated manner.
Park et al. [19] demonstrated that monomeric myosin VI motors lacking the CBD can dimerize in vitro if they are brought into close proximity, suggesting that myosin VI may be capable of in vivo dimerization in regions of high local motor concentration. However, the CBD appears to somewhat inhibit this dimerization, indicating that dimerization may require a proper positioning of the monomers.
Given the above considerations, we hypothesize that two myosin VI CBDs are precisely positioned close together when loaded onto a vesicle and that this positioning orients the motor appropriately for dimerization (Figure 1B). This would then allow the motor to perform its predicted trafficking function as a processive dimer [5].
Similar regulation of a motor protein between a monomer and dimer has been proposed for the C. elegans kinesin Unc104 [20,21]. However, mechanisms of dimerization for Unc104 and myosin VI are both inferred from in vitro data, and there is a lack of evidence indicating that this dimerization occurs in vivo. Here, we provide evidence for formation of a myosin VI dimer in vivo.
We conducted our studies in ARPE-19 cells, a human retinal pigment epithelial cell line that was one of the lines used by Dance et al. [6] in their in vivo studies of myosin VI. Dance et al. demonstrated in various cell lines that, during clathrin-mediated transferrin uptake, endogenous myosin VI colocalizes with transferrin-containing UCV [6]. Figure 1C shows a green fluorescent protein (GFP) image of an ARPE-19 cell expressing full-length myosin VI with an N-terminal GFP (GFP-FL). After transfection with the GFP-FL construct, cells exhibited two distributions of GFP fluorescence: a homogeneous GFP haze throughout the cytosol and small, bright GFP puncta that exist throughout the cell, though are often more dense in the cell periphery (Figure 1C). Dance et al. [6] observed similar colocalization for endogenous myosin VI.
To verify that the myosin VI puncta correspond to UCV, we observed endocytosis of transferrin conjugated with Alexa 647 dye (Alexa647-Tfn). Transferrin is known to be internalized via the clathrin-mediated endocytic pathway. GFP puncta showed a high degree of colocalization with internalized fluorescent transferrin immediately after internalization (Figure 1C).
We next sought evidence for precise positioning of myosin VI motors on UCV. As discussed in the introduction, we speculated that myosin VI heavy chains, although likely to be monomeric in the cytoplasm [18], are brought into close proximity on its cargo, allowing the motor to function as a dimer.
Two primary myosin VI–truncated CBD constructs were used for these studies (Figure 1A). The first is a myosin VI containing the CBD as well as 17 residues from the TD N-terminal to the CBD (we refer to this simply as the CBD construct). The second construct is the CBD construct with a leucine zipper (GCN4) attached at its N-terminus (GCN4-CBD), which forces it to dimerize [14]. These constructs were made fluorescent by inserting a monomeric GFP isoform [22] at the N-termini of CBD (GFP-CBD) and of GCN4-CBD (GFP-GCN4-CBD).
After transfection of ARPE-19 cells with these constructs, cells exhibited the same two distributions of GFP fluorescence as cells transfected with GFP-FL: a homogeneous GFP haze throughout the cytosol and small, bright GFP puncta throughout the cell (Figure 1D and 1E). We also observed endocytosis of transferrin conjugated with Alexa 647 dye in these cells. GFP puncta showed a high degree of colocalization with internalized fluorescent transferrin immediately after internalization (Figure 1D and 1E). Thus, our constructs have maintained their ability to associate with UCV similarly to endogenous myosin VI [6].
In movies of cells expressing our GFP-tagged myosin VI–CBD constructs, UCV exhibited motion throughout the cell, with UCV toward the periphery of the cell typically exhibiting slower velocities relative to those further into the cell (see Videos S1–S3). For cells expressing GFP-FL, the slower motion likely corresponds to myosin VI–dependent movement of UCV through the thick actin-mesh at the cell periphery. This peripheral mesh is particularly thick in ARPE-19 cells, and UCV travel a net distance of approximately 2 μm through the actin in a process that takes on the order of 5 min [5]. For cells expressing myosin VI constructs lacking the catalytic head, the GFP construct competes with the endogenous motor for binding to the UCV, and acts as a dominant negative. In these cells, the slower motion corresponds to Brownian-like motion with a slow drift toward the interior of the cell [5]. The faster motion observed deeper in the cell may result from UCV moving on microtubules, consistent with predictions that a microtubule network is involved in intracellular trafficking of UCV from the early to late endosomes [23].
Fluorescence resonance energy transfer (FRET) is the nonradiative transfer of energy between fluorophores occurring when the emission spectrum of an excited fluorophore overlaps with the absorption spectra of a fluorophore in very close proximity (within 10–100 Å) [24]. FRET between identical fluorophores (homoFRET) serves as an ideal way for detecting homo-oligomeric protein configurations [25,26].
According to our proposed mechanism for myosin VI function when bound to its cargo (Figure 1B), the CBDs of UCV-associated myosin VI heavy chains are positioned to bring together the heavy chains. By analogy, this mechanism also predicts that, for two UCV-associated GFP-CBD constructs, the CBDs are positioned to bring into close proximity their associated GFPs. Thus, homoFRET of GFP-CBD on the UCV serves as a readout of our proposed mechanism.
GFP-GCN4-CBD serves as a positive control for detection of homoFRET. The leucine zipper forces the construct to form a constitutive dimer, resulting in close association of the GFPs adjacent to the GCN4 coiled coil and subsequent homoFRET. As a negative control, we used a construct that is identical to GFP-CBD, except the GFP is located at its C-terminus (CBD-GFP). For this construct, we expect that GFPs are not likely to be positioned to undergo homoFRET, even if the CBDs are brought close together.
To quantify the levels of GFP-FL, GFP-CBD, GFP-GCN4-CBD, and CBD-GFP in the cytosol and on UCV, we used multiphoton microscopy to gather confocal images of GFP fluorescence from ARPE-19 cells expressing these constructs. A sample image collected for GFP-CBD is shown in Figure 2 (top). Fluorescence emission was collected with multichannel plate photomultiplier tubes capable of photon counting. For an imaged cell, the mean photon count was calculated in numerous regions corresponding to the UCV and the cytosol (for example, see Figure 2, top), and these values were averaged to arrive at the cell's mean fluorescence intensity at each localization. The mean fluorescence intensities for multiple cells were then averaged to arrive at the overall mean fluorescence intensities at UCV and in the cytosol for each construct (see Materials and Methods). The total expression of all constructs was similar, as were their levels on UCV and in the cytosol (Figure 2, middle).
We excited the fluorophores of our GFP-tagged CBD constructs with pulsed, polarized excitation, and observed subsequent changes in fluorescence emission polarization, as quantified by the fluorescence anisotropy, over time. The emission polarization is initially aligned with the excitation polarization, resulting in a high initial anisotropy, and becomes randomized over the lifetime of the fluorophore through two processes: (1) rotational diffusion of the GFPs and (2) energy transfer to GFPs in close proximity (on the order of the Förster's radius) [27]. Rotational diffusion and homoFRET each result in exponential decays in anisotropy which, for large proteins, occur on very different time scales [26,28]: homoFRET results in a rapid anisotropy decay, and rotational diffusion results in a slower decay (Figure 3, top). From the former, we can detect processes that bring GFPs into close proximity, and from the latter, we can infer the size of the rotating object.
We measured fluorescence anisotropy following polarized multiphoton excitation with a pulsed laser (∼12-ns repetition rate) using time-correlated single-photon counting (TCSPC) [26,29] (see Materials and Methods). Pico-second time-resolved anisotropy decays were measured for our three GFP-tagged CBD constructs in the cell periphery, both in the cytosol and at the UCV. An example of the regions selected for these measurements is shown in Figure 2 (top). We selected UCV in the cell periphery to be sure the construct is associated with UCV in the peripheral actin network. To further ensure that we selected for these vesicles, we collected an image encompassing the area of the measured UCV both before and after the measurement, an interval lasting approximately 1 min. UCV that remained in the observation volume before and after the measurement corresponded to slowly moving UCV that were associated with the peripheral actin network and so were selected for analysis.
We fit two decay models to each empirical anisotropy decay: (1) a single exponential decay and (2) the sum of two exponential decays (see Materials and Methods and Figure 3). These fits revealed two classes of decay profiles. For the first class, the profiles were well fit by a single exponent; the addition of another exponent had little effect on the fit. These profiles describe decay in anisotropy through only a single process, presumably rotational diffusion. For a second class, the decay was not fit well by a single exponent, but the addition of a second exponent resulted in a good fit (for example, see Figure 3, bottom). These profiles describe decay in anisotropy through two exponential processes, both homoFRET and rotational diffusion. In this manuscript, we describe in detail the best fits for all empirical decays, the first class of decays to a single exponent and the second class to the sum of two exponents (Figure 4 and Table 1).
As a control to test the instrumentation, we transfected cells with monomeric GFP. The GFP homogeneously filled the cytoplasm, and anisotropy decay profiles collected from cytosolic GFP were well fit by a single exponent (Figure S1). The time scale of this decay (∼25 ns) is consistent with previous measurements of GFP tumbling in the cytosol [30].
For CBD-GFP, anisotropy decays measured both in the cytosol and at UCV were well fit by a single exponent (Figure 4, middle), consistent with our expectation that the construct does not undergo homoFRET and that anisotropy decreases only through fluorophore rotation. The time scale of this decay in the cytosol is consistent with tumbling, and the decay is considerably slower at the UCV, consistent with a slowed rotation due to association of the CBD-GFP with a UCV (Table 1).
Anisotropy decays collected for GFP-GCN4-CBD both in the cytosol and at UCV could not be fit by a single exponential but were well fit by the sum of two exponents (Figure 4, right), consistent with our expectation that anisotropy decreases both through rotational diffusion and through homoFRET of the dimeric construct. The time scales for the fast and slow decays are consistent with homoFRET and tumbling, respectively. As with CBD-GFP, the decay corresponding to rotation is slower at the UCV relative to the cytosol due to association with the UCV (Table 1).
Anisotropy decays collected from GFP-CBD in the cytosol were well fit by a single exponent. Anisotropy decays collected at UCV, however, were only well fit by the sum of two exponents (Figure 3, bottom, and Figure 4, left). The rapid anisotropy decay at UCV is consistent with a homoFRET process. Both the slower decay at the UCV and the decay in the cytosol are consistent with rotational diffusion. As with the other GFP-constructs, the decay describing rotation is slower at the UCV compared to the cytosol (Table 1).
From these data, we infer that GFP-CBDs are positioned on UCV to bring their N-termini together (Figure 1B). The lack of homoFRET in the cytosol confirms that this precise positioning requires the construct to be loaded onto the vesicle. This result is consistent with our prediction that a precise positioning of CBDs on a vesicle orients heavy chains in close proximity.
For anisotropy decays measured in the cytosol, the time scale describing rotational diffusion (Table 1) provides information about the size of the GFP construct. Because these decay times are longer than the fluorescence lifetime of GFP (∼3 ns), unambiguous molecular weights cannot be determined. However, we can infer relative sizes of our constructs from these decays.
Rotation of GFP-CBD is slower than for GFP alone (Figure S1), consistent with slowed rotational diffusion of the fluorophore when attached to the myosin VI construct. The 2-fold difference in the decay time suggests that the molecular weight of GFP-CBD is twice that of GFP (molecular weight, 27 kDa), consistent with a monomeric form of GFP-CBD (molecular weight, 58 kDa).
The rotational decay time of GFP-CBD is also similar to that of CBD-GFP, and both constructs rotate faster than GFP-GCN4-CBD (Figure S1). Thus, both GFP-CBD and CBD-GFP appear to be smaller than a similarly sized GFP construct known to dimerize, supporting our prediction that both are monomeric in the cytosol.
To further demonstrate that homoFRET occurs when GFP-CBD is associated with UCV, we determined both the steady-state fluorescence emission and steady-state fluorescence anisotropy throughout cells expressing GFP-CBD (Figure 5A). Steady-state anisotropy represents the integral over time of an anisotropy decay profile, and so it is reduced by both rotational diffusion and homoFRET (Note the areas under the curves in Figure 3, top), though these processes cannot be distinguished by steady-state analysis [31]. In all cells analyzed, steady-state anisotropy for GFP-CBD was clearly lower at UCV relative to the surrounding cytosol (Figure 5A). Considering only the effects of fluorophore tumbling, we would have expected steady-state anisotropy to be lower in the cytosol, where rotational diffusion is more rapid. The observed pattern of steady-state anisotropy is thus consistent with a further reduction in steady-state anisotropy at the UCV due to homoFRET.
To quantify this, we manually selected from the periphery of each cell numerous UCV as well as 30 regions in the cytosol, similar in size to the UCV, and calculated the steady-state anisotropy at these regions (see Materials and Methods). We observed that the mean steady-state anisotropy of these regions at either localization is not consistent from cell to cell, due to a variety of factors such as cell thickness. However, for all ten cells analyzed, we observed that the mean anisotropy at the UCV was consistently lower than the mean anisotropy in the cytosol. An example of analysis of a single cell is shown in Figure 5B.
As a control, we also determined the steady-state fluorescence anisotropy and emission throughout cells expressing CBD-GFP. Because CBD-GFP does not exhibit homoFRET at UCV, its steady-state anisotropy should be dictated solely by its rotational diffusion. Thus, we expect a higher anisotropy at UCV relative to the surrounding cytosol, in contrast to the pattern observed for cells expressing GFP-CBD. We observed this expected pattern of steady-state anisotropy for all cells analyzed (Figure 5C), and again we quantified this by calculating steady-state anisotropy at numerous UCV regions and at 30 regions corresponding to the cytosol. For all nine cells analyzed, we observed that the mean steady-state anisotropy at the UCV was higher than the steady-state anisotropy in the cytosol. An example of analysis of a single cell is shown in Figure 5D.
Our time-resolved and steady-state anisotropy experiments demonstrated that GFP-CBD undergoes homoFRET when localized to UCV. Though we hypothesize that this is the result of precise positioning of the construct on a UCV, we must also consider the possibility that GFPs are brought into close proximity simply due to the crowding of a high density of GFP-CBD on the UCV surface.
From our analysis of steady-state GFP-fluorescence emission images of cells expressing GFP-CBD and CBD-GFP, we observed that both constructs are expressed to similar levels in our cell line (Figure 2, middle). Furthermore, both constructs exhibit similar ratios of GFP intensities on UCV and in the cytosol, indicating that they are loaded onto vesicles at similar densities and so are similarly crowded (Figure 2, bottom). Thus, if homoFRET of GFP-CBD were the result of crowding, we would also expect CBD-GFP to exhibit homoFRET when loaded onto UCV. Since this is not the case (Figure 4, middle), crowding cannot be the cause of homoFRET. Instead, the CBD must be positioned on a UCV so that its N-termini are brought together, resulting in homoFRET from an N-terminal (and not a C-terminal) GFP.
We confirmed this conclusion by examining UCV associated with varying densities of GFP-CBD. If homoFRET were the result of crowding, then reducing the construct concentration on the UCV would reduce close packing of fluorophores and subsequently reduce the occurrence of homoFRET. On the other hand, if homoFRET results from precise positioning of GFP-CBD on UCV, then, even at low densities, the construct will undergo homoFRET.
To differentiate between these mechanisms, we calculated the steady-state fluorescence emission intensity at the UCV regions selected from the previous steady-state anisotropy analysis (see Figure 5 and Materials and Methods). Using these measurements, we probed for effects of GFP-CBD density on homoFRET by looking for effects of varying steady-state fluorescence emission intensity on steady-state fluorescence anisotropy (Figure S2). To determine the degree to which these measured values of anisotropy and intensity are related, we calculated the Pearson product-moment correlation coefficient (r).
For ten cells expressing GFP-CBD, the UCV regions of nine cells showed no significant correlation between steady-state fluorescence emission intensity and anisotropy (p > 0.05 for nine cells, p = 0.02 for one cell). The lack of correlation indicates that the extent of homoFRET does not depend on the density of UCV-associated GFP-CBD. This supports our conclusion that homoFRET results from precise positioning of GFP-CBD on the vesicle, and not crowding of the fluorophores.
As expected, when we performed a similar analysis for GFP-CBD regions in the cytosol, where the construct does not undergo homoFRET, we observed no correlation between fluorescence emission intensity and anisotropy. The same is true for the UCV and cytosolic regions of cells expressing CBD-GFP, which does not undergo homoFRET at either localization, and for GFP-GCN4-CBD, which undergoes homoFRET at both localizations due to precise positioning of its fluorophores (unpublished data).
In summary, understanding the in vivo functional form of a molecular motor is essential to understanding its function. Our data suggest that, although myosin VI exists as a monomer in the cytosol, heavy chains are brought into close proximity on UCV, allowing the motor to function as a dimer. Consistent with our model, Spudich et al. [13] reported that a myosin VI tail construct, when bound to artificial lipid vesicles in vitro, can be linked as dimers upon addition of a zero-length cross-linker. Through this mechanism, myosin VI is able to processively traffic its vesicular cargo through the actin meshwork in the cell periphery [32].
GFP constructs were derived from the GFP-HM6Tail+LI construct from Dance et al. [6], which consists of a myosin VI–CBD construct in the pEGFP-C3 expression vector (Clontech, http://www.clontech.com). GFP-CBD was made from GFP-HM6Tail+LI by changing residue 206 of the GFP from alanine to lysine (A206K), which reduces the proclivity of GFP to dimerize [22]. This was achieved through site-directed mutagenesis using the primer 5′-CCTGAGCACCCAGTCCAAGCTGAGCAAAGACCCCA-3′ and the QuikChange Site-Directed Mutagenesis Kit (Stratagene, http://www.stratagene.com).
To make GFP-GCN4-CBD, the leucine zipper from the myosin VI/GFP plasmid described in [33] was amplified using the primers 5′-CCCGAATTCTGGAAGACATGAAACAGCTCGAGGACAAAGTAGAGGAGCTGCTGTCCAAG-3′ and 5′-GCCCGCGGCTCCCCGACCAGCTTCTTAAGTCTCGCAACCTCATTTTCTAGATGG-3′.
The resulting PCR product was cut with EcoRI and SacII, and inserted into the MCS of the GFP-CBD plasmid.
To make CBD-GFP, CBD was amplified from the GFP-CBD plasmid using the primers 5′-CGCCGCGGATGAGGATTGCCCAGAGTGAAGCCGAGCTCATCAGTGATGAGGCCC-3′ and 5′-TTGGATCCGCCTTTAACAGACTCTGCAGCATGGCTGTTGCATAGGTGGGCCGAGCCTG-3′.
The resulting PCR product was cut with SacII and BamH1, and inserted into the multiple cloning site (MCS) of the pEGFP-N1 expression vector (Clontech) containing the A206K GFP mutation. The A206K mutation was made in pEGFP-N1 using the site-directed mutagenesis described above.
ARPE-19 cells were purchased from American Type Culture Collection (ATCC, http://www.atcc.org). Cells were grown at incubating conditions (37 °C and 5% CO2) in medium + serum (DMEM/F-12 [GIBCO-Invitrogen, http://www.invitrogen.com], 1% fungizone [GIBCO], 1% L-glutamate [GIBCO], 10% FBS [GIBCO], 1.5 M HEPES, 100 U/ml penicillin, and 100 mg/ml streptomycin). Before transfection, cells were grown in imaging dishes that were polylysine-coated, with a translucent bottom appropriate for fluorescence imaging. To transfect cells, a transfection mixture, consisting of 75 μl of serum-free media (SFM; medium+serum lacking FBS), 6 μl of TransIT Transfection Reagent (Mirus Bio Corporation, http://www.mirusbio.com), and 1 μg of plasmid DNA, was added to the cell culture, which is in 0.75 ml of medium + serum. Cells were imaged 10–20 h after transfection.
Transferrin was labeled with Alexa 647 (Alex647-Tfn) using the Alexa Fluor 647 Protein Labeling Kit (Molecular Probes, http://probes.invitrogen.com). To observe uptake of Alexa647-Tfn, cells grown in imaging dishes were starved in SFM for 2 h at incubating conditions. The media was removed, and 150 μl of 10 μg/ml Alexa647-Tfn in SFM was applied to the cells. The cells were left at incubating conditions for 30 min.
Cells were fixed by washing in M1 buffer (150 mM NaCl, 5 mM KCl, 1 mM MgCl2, 1 mM CaCl2, 20 mM HEPES [pH 7.4]) and then adding 150 μl of 4% paraformaldehyde in M1 buffer. Cells were incubated for 25 min at room temperature, washed with M1 buffer, and imaged.
For colocalization experiments, the transferrin-internalization protocol was begun 10.5 h after transfection of cells with the GFP construct.
Briefly, experiments were done on a Nikon TMD fluorescence microscope with a cooled back-illuminated, 16-bit charge-coupled device (CCD) camera (Nikon, http://www.nikonusa.com). Different filter sets were used to image Alex 647 and GFP. Images were collected using Metamorph software (Molecular Devices, http://www.moleculardevices.com). Fluorescence imaging was carried out exactly as described [34].
Live-cell measurements of fluorescence anisotropy were made using TCSPC and pulsed multiphoton excitation. Details of the method and analysis will be described elsewhere (D. Goswami, K. Gowrishankar, M. Rao, and S. Mayor, unpublished data). Briefly, steady state and time-resolved anisotropy measurements of fluorophores excited by multiphoton excitation were made on a Zeiss LSM 510 Meta microscope (Carl Zeiss, http://www.zeiss.com) with 63× 1.4 numerical aperture (NA) objective coupled to the femtosecond-pulsed Tsunami Titanium:Sapphire tunable pulsed laser (Newport, http://www.newport.com). Parallel and perpendicular emissions were collected simultaneously into two Hamamatsu R3809U multi-channel plate photomultiplier tubes (PMTs; Hamamatsu Photonics, http://www.hamamatsu.com) using a polarizing beam splitter (Melles Griot, http://www.mellesgriot.com) at the non-descanned emission side. TCSPC was accomplished using a Becker & Hickl 830 card (Becker and Hickl, http://www.becker-hickl.de), operating in a stop–start configuration [35]. For multiphoton excitation of GFP or fluorescein in cells, we used 920-nm excitation wavelength. At this wavelength, the two-photon absorption cross section for GFP is higher, enabling lower laser excitation power, and autofluorescence signals are minimized. The repetition rate of the pulsed laser is 80.09 MHz (12 ns).
Steady-state imaging was accomplished using a pixel residence time of 102 μs/pixel, setting the detection time resolution in the Becker and Hickl card to one. Thus, a full image (512 × 512 pixel) was collected over 62 s. For time-resolved anisotropy measurements, the time resolution was 12.2 ps. The beam was “parked” at a single point using routines available in the Zeiss software. The parked beam was placed at the center of the field to maintain uniformity of G-Factor, and photons were collected for 30–50 s. Photons were collected at a maximum rate of 0.1 MHz to ensure that TCSPC conditions were strictly met. Because of the low laser power, less than 10% bleaching was observed during a measurement. The instrument response function (IRF) was measured using 10–16-nm gold particles dried on a coverslip as a second harmonic generator; full width at half maximum (FWHM) of IRF is approximately 60 ps.
In our experimental setup, the steady-state anisotropy measured while the laser beam is parked at a single point (Table S2) was always higher than the steady-state anisotropy measured using the scanning mode (Figure 5B and 5D). This is attributed to a small but detectable depolarization of the excitation laser beam in the scanning mode when using high NA objectives, which is also seen for measurements of a monomeric GFP solution. This effect is negligible for objectives with NA less than 0.8 (unpublished data). A high NA objective was required to discern GFP associated with UCV as puncta distinct from the GFP cytosolic haze.
Fluorescence lifetime and anisotropy decay analyses were done essentially as described [26,36,37], with minor modifications in the analysis procedure. Briefly, the experimentally measured fluorescence decay is a convolution of the IRF with the intensity decay function. The intensity decay data were fit to the appropriate equations by an iterative reconvolution procedure using a Levenberg-Marquardt minimization algorithm.
When fitting the models to the decay profiles, ro was constrained to a small window to improve the ability of the fitting algorithm to find the optimal fit. A constrained range of values for ro (0.43 ± 0.3) was used for all fits described in the manuscript. This range of values was obtained from unconstrained fitting of cytoplasmically expressed GFP fluorescence emission anisotropy decays (n = 8). These fits provided reliable values for the initial anisotropy because the fluorophore does not undergo homoFRET and because the time scale of rotation is much slower than the time scale of the measurement.
When fitting a model describing two exponential decays to the decay profiles, the two decay times were somewhat constrained to a wide range of values. These constraints involved large windows centered on the expected decay times for the physical processes involved (homoFRET and rotational diffusion). Again, these constraints improved the ability of the fitting algorithm to find the optimal fit.
It is important to note that the empirical anisotropy decay profile is the convolution of the real-time behavior of the fluorophore with the IRF. This distortion results in an apparent fast decay at the start of all measurements that is an artifact and does not represent anything physical. This artifact is apparent because our sampling rate of 12.2 ps is smaller than the width of the instrument's IRF (∼60 ps). This effect is also apparent in the empirical decay for a monomeric GFP in the cytoplasm (Figure S1).
The G-Factor was estimated using a fluorescein solution and setting the anisotropy at late times to 0.005. Fluorescence and anisotropy decays were considered well fit if three criteria were met: reduced χ2 was less than 1.4, residuals were evenly distributed across the full extent of the data, and visual inspection ensured that the fit accurately described the decay profile.
For analysis of steady-state images, N nearest-neighbor averaging of a 512 × 512 array of pixel values refers to the following calculation: the pixel value at location (row = i, column = j) was set to the mean value of pixels spanning rows i − N to i + N and columns j – N to j + N. This calculation was done using software developed in Matlab (The MathWorks, http://www.mathworks.com).
Steady-state anisotropy was calculated from steady-state parallel- and perpendicular-polarization images. To account for differences in the optical paths traversed by the perpendicular and parallel emissions, a G-Factor correction was applied to the data as follows. We collected steady-state emission images from a fluorescein sample. Because fluorescein tumbles rapidly relative to the time scale of our measurements, fluorescein provides a pixel-by-pixel readout of the detector output from a source emitting identically in both polarizations. We created a G-Factor image from the parallel and perpendicular fluorescein emission images, and perpendicular images from subsequent experiments were multiplied by this G-Factor image to apply the appropriate correction.
To create this G-Factor image, dividing the parallel and perpendicular fluorescein images pixel by pixel is insufficient, and results in a correction that is artificially too large. This is because pixel values in the parallel and perpendicular images exhibit Poisson photon noise. To reduce artifacts arising from dividing signals containing noise, the images must first be averaged so as to increase the signal-to-noise ratio. Using simulated data, we determined the extent of averaging required to sufficiently reduce the G-Factor artifact while retaining information about G-Factor variation across the image (unpublished data). To create our G-Factor image, we first applied three-nearest-neighbor averaging to the parallel and perpendicular fluorescein images and then divided the averaged images pixel by pixel. A new G-Factor image was created for each day of experiments.
After applying the G-Factor correction to the data, anisotropy was calculated at UCV and in cytosolic regions using software developed in ImageJ [38]. Regions were manually selected in the parallel image and were transferred to the perpendicular image. The mean perpendicular- and parallel-polarization emission intensities were calculated for each region, and from these, the steady-state fluorescence anisotropy and fluorescence emission intensity were calculated using the relations:
and
where I‖ and I⊥ are the calculated intensities of fluorescence emission with polarization parallel and perpendicular to the excitation polarization, respectively, I is the total fluorescence emission intensity, and r is the fluorescence anisotropy.
|
10.1371/journal.pcbi.1000708 | Slower Visuomotor Corrections with Unchanged Latency are Consistent with Optimal Adaptation to Increased Endogenous Noise in the Elderly | We analyzed age-related changes in motor response in a visuomotor compensatory tracking task. Subjects used a manipulandum to attempt to keep a displayed cursor at the center of a screen despite random perturbations to its location. Cross-correlation analysis of the perturbation and the subject response showed no age-related increase in latency until the onset of response to the perturbation, but substantial slowing of the response itself. Results are consistent with age-related deterioration in the ratio of signal to noise in visuomotor response. The task is such that it is tractable to use Bayesian and quadratic optimality assumptions to construct a model for behavior. This model assumes that behavior resembles an optimal controller subject to noise, and parametrizes response in terms of latency, willingness to expend effort, noise intensity, and noise bandwidth. The model is consistent with the data for all young (n = 12, age 20–30) and most elderly (n = 12, age 65–92) subjects. The model reproduces the latency result from the cross-correlation method. When presented with increased noise, the computational model reproduces the experimentally observed age-related slowing and the observed lack of increased latency. The model provides a precise way to quantitatively formulate the long-standing hypothesis that age-related slowing is an adaptation to increased noise.
| In a hand-eye coordination task that requires continuous movement to correct for a disturbance, it turns out that signs of response to the disturbance appear no later in the elderly than in the young. The elderly motion is noisy and less efficient, however, and once movements in response to a disturbance begin, they are at a lower speed. One can model subject response by assuming that it results from combining noise and a response that is mathematically optimal given this noise, delay, and a least-squares sort of control objective. This modeling approach is appropriate for young and most elderly subjects. The model holds that increased noise should lead to no change in delay until response gets underway, but should make the response itself proceed at a slower speed. This is consistent with the data and with a causal link from the observed noise and disorder in elderly motor function to the observed age-related slowing.
| The existence of a general phenomenon of age-related impairment of the sensorimotor system is widely accepted [1]–[3], but its causes are incompletely understood and its expression varies depending on task. Conventional wisdom is that impairment is the result of slowing and increased sensorimotor noise. In this paper we show that in a compensatory tracking task, age-related slowing is confined to the later phases of the response, and that this slowing is consistent with the adaptation of a relevant optimal control strategy to increased noise. This suggests that slowing in this task is consistent with adaptation to the cause of impairment, i.e. increased noise, rather than a cause of impairment.
The hypothesis that age-related slowing is an adaptive response to increased cortical disorder rather than a primary cause of impairment was put forward in a survey of work on aging and reaction times [2]. It is suggested in [2] that the elderly average sensory data using longer timescales in order to reduce the effect of cortical noise. Recent data confirms that age related slowing is correlated with increased disorder in cortical Event Related Potentials [4],[5], and in [5] the same hypothesis about averaging noise away is formed. In [6] data is presented suggesting that slowing is too great to be accounted for by then-current empirical speed-accuracy tradeoff models. In [7] the modeling concepts of submovements and signal-dependent motor variability [8] were used to analyze age effects in a target acquisition task and the authors suggested that elevated motor noise is a primary cause of slower performance, based on empirically parametrized speed-accuracy tradeoffs. In this work we use a continuous compensatory tracking task and a model based on Bayesian and quadratic optimality consistent with past work on the sensorimotor application of this theory [9]–[26]. This modeling approach provides a quantitative way to relate noise levels to the timescales involved in averaging data. Our task elicits responses in the form of continuous time series that are dynamically rich compared to, for example, static force production or reaction time data. It also resembles dynamic activities of daily living such as driving a car on a windy day. The nature of the controlled dynamics and perturbations enable use of well-developed optimal control theory.
Age-related sensorimotor impairment and slowing have been heavily investigated [1]–[3],[27]. One hypothesis is that of generalized slowing, as reviewed in [1],[28]. Published results show wide variability. For example, some experiments involving simple reaction time [4],[5] fail to show a significant age-related increase, but slowing is strongly expressed in choice reaction experiments [4],[5] and is the rule rather than the exception experimentally [2]. Subtle changes in force stimuli can evoke or eliminate age effects on response latency during fingertip force generation [29]. Age-related degradation and delay in the peripheral components of the sensorimotor system has been documented in cutaneous mechanoreceptors [30], transmission of signals from visual to motor areas [31], and basic visuomotor processes such as saccades [32]–[34]. Nevertheless, it has been known for decades that age-related sensorimotor impairment in common experimental paradigms is dominated by central and not peripheral effects [2],[27].
We distinguish between two different kinds of “slowing” that are well characterized in the engineering field: (i) response latency, that is, delay until the onset of motor response, vs. (ii) slowed timescales of post-onset response. These are shown in Fig. 1. This distinction is closely related to the classical behavioral classification of response in reaction time studies into “reaction time” vs. “movement time” after the onset of movement [35], but we avoid this terminology to avert confusion with other published uses of those terms. Our analysis is specifically designed to disambiguate between these two kinds of slowing using complementary analytical approaches. The first is a cross-correlation approach that is strictly data-driven, phenomenological and free of assumptions; and the second is a model-based computational approach based on justifiable assumptions of optimal control theory. The advantage of the model-based approach is that it improves precision and provides a framework for mechanistically relevant analysis of response.
It is important to note that our definition of endogenous noise is a more general concept than physiological motor noise. For the purposes of this paper we define endogenous noise to be any deviation from optimal behavior (possibly of sensory, motor, conduction, or neural processing origin). We address the roles of specific sources such as muscular “motor noise” or cortical “functional dysregulation” [4] in the Discussion.
The experimental procedure was identical to that described in [25]. The paradigm is compensatory tracking with a controlled dynamical system implemented in software. A band limited Gaussian perturbation [36] continuously moves a displayed cursor on a horizontal line and the subject is asked to provide a corrective control input to make the cursor track the stationary midpoint of the line. Ideally the subject would provide an input that perfectly canceled the perturbation, but they cannot due to lack of foreknowledge of the perturbation, delays, and noise in the sensorimotor system. Fig. 2 shows a brief representative section of the time histories of the perturbation , the control input , and the displayed cursor error .
The cursor's motion results from simple but marginally unstable dynamics:(1)
Equation 1 states that the horizontal velocity of the cursor is the sum of the subject's control input (defined as the left-right deviation of the manipulandum from its initial position) and the software-supplied perturbation . Thus the horizontal position of the cursor (i.e., error ) is the time integral of this velocity. These cursor dynamics are marginally unstable, similar to [37]. They resemble driving a car on a windy day such that, if the subject does not correct, then the cursor will wander off the screen in approximately 10 seconds. These dynamics are a compromise between the static and the excessively difficult , for which total loss of control was seen even in some young subjects.
The data for eleven of the young subjects appeared in [25]. Repeated testing was performed eight months later on eight of the young subjects as described in [25]. Subjects 1 and 10 participated and no longer displayed the outlying behavior described in this paper, and other subjects remained well described by the optimal control model. All the young subject data used in this paper for comparison to the elderly is from the first session.
We used a computational model to improve the precision of the response latency estimates and to investigate quantitatively the slowing of the post-onset response. The structure of this continuous linear feedback control task is suited to modeling using concepts from Linear Quadratic Gaussian (LQG) optimal control [38]. We need only make standard assumptions of quadratic optimality and additive Gaussian noise, and minimalist assumptions about a feedback loop structure as shown in Fig. 3, to arrive at our modeling framework. The quadratic optimality assumption is that the expected value of the following cost, integrated over an arbitrarily long time frame, is to be minimized:(2)where is a weight, is the time derivative of the ideal control input (that is, hand motion before adding endogenous noise ), and is the tracking error. The resulting LQG control strategy is to choose a control input by multiplying a problem-dependent static gain matrix and the Bayesian optimal estimate of the state of the system. These are referred to as the Linear Quadratic Regulator (LQR) gain matrix and the Kalman state estimate [38]. For a survey of applications of optimality as an organizing principle of animal behavior see [16],[17]. This approach is particularly relevant to the study of sensorimotor behavior because the delays and noises under which the task is performed are taken into account when computing the optimal control strategy. The use of this model to analyze visuomotor response is not a claim that it is the best possible model; its advantages are simplicity and parsimony [25]. We have not found a better model for response in our task in the literature, and we show later that a standard way of forming linear model fits (without any restrictions to optimality) does not do substantially better.
The standard LQG model introduced above is based on an additive noise model. It is well known that sensorimotor noise is signal-dependent, that is, its variance is a function of the amplitude of the corrupted signal [8],[11]. Considering the effects of signal-dependent noise will require us to make some mild assumptions, but ultimately leave us with the same model as the additive noise model, with a different interpretation of the parameter. The remainder of this paragraph describes our approach. The most straightforward model of signal-dependent noise is to assume that the variance is proportional to the variance of the corrupted signal, that is, multiplicative noise. Two approaches may then be taken to solve the control design problem: either adopt time-varying control policies [18], or restrict ones' attention to much simpler time-invariant control strategies and examine the role of multiplicative noise in that tractable case. We do the latter. The well-known principle that estimation and control can be treated separately for the additive noise LQG problem does not hold for signal dependent noise [18], but it proves tractable and consistent with the data to continue treating the problem as having separate control and estimation components. First, in terms of optimal control given a state measurement, it has been shown that under the relatively mild time-invariance and multiplicative noise assumptions, control cost and multiplicative noise levels have a summed joint effect during optimal LQR control design [39]: a positive coefficient describing the intensity of the multiplicative noise is added to a freely chosen positive control cost during control design. This leads to a situation where we cannot disambiguate these two effects during model fitting unless we make a poor assumption that the noise is purely multiplicative. Second, in terms of estimation, the restriction to time-invariant control implies a time-invariant Kalman filter, which is equivalent to designing the filter with some appropriate level of additive noise. Ultimately, the approach described in this paper is to note that the tractable approximate approach to signal-dependent noise under mild assumptions is to use the same model as in the additive noise case, and then to note that the remaining difference is only in the interpretation of the parameter : in an additive noise model, it is purely a freely chosen control cost, while in the approximate time-invariant solution to the multiplicative noise model, the parameter is the sum of a coefficient of noise intensity, and a freely chosen control cost.
The optimal control model structure is shown in Fig. 3 and its components are described in the caption. In our approach, the four parameters used to fit the optimal control model to the data from any given trial are:
Once we fit parameters in the assumed modeling framework, the model is fully defined and we can predict properties of the control output . Iterative Nelder-Mead [40] minimization of the squared discrepancy between predicted and observed subject response was used to fit models. The resulting controllers have nine states because the plant has nine states (the delay is modeled with a fourth order approximation, one state in the computer, three states in noise filters, and one state in ), but less than five states are significant in the fitted controllers, based on analysis of Hankel Singular Values (HSVs) [38].
The first two results make no use of the optimal control model. The third result is a set of tests of the relevance of the optimal control model. The remaining results involve properties of the fitted optimal control models.
To determine the response latency in this task, we examined the cross correlation of the random signal driving the perturbation, and the subjects' response. The cross correlation [36] is a simple phenomenological way to analyze the temporal relationships between the discretized signals (the perturbation) and (the control input or “correction” by the subject) without making any modeling assumptions. The perturbation was a discrete-time approximation to band-limited Gaussian noise [36], created by passing a sequence of normally distributed random numbers (discrete time white noise) through a first order filter with a cutoff frequency of Hz. This filter is required to ensure that the perturbation is not “too jumpy” for the subject to react. This filter induces autocorrelation in [36], thus in our analysis we use the pre-filtered or “whitened” [36] signal, to remove spurious effects. We use indexing subscripts to denote samples at time through and compute the cross-correlation(3)Using the cross correlation to infer response latency is straightforward because it wanders randomly near zero for low values of lag index until the effects of the perturbation are seen in . These effects occur because the subject attempts to negate the effects of . Therefore the cross-correlation begins to show a negative (i.e., corrective) trend at a value of corresponding to the response latency, and we use this to measure latency.
Application of our model-free cross-correlation method yields mean inferred response latencies of 267 and 263 ms for young and elderly subjects. The precision of the inferred quantities is evidenced by the median intrasubject standard deviations of 36 and 44 ms. A two-tailed t-test [41] was used to compare the 12 young to the 12 elderly using the mean from each subject to avoid repeated measures bias. The measurements' distribution was consistent with normality with or without using logarithms. The t-test indicates that the 4 ms difference between the two means is not statistically significant . The averaged cross-correlation data of young and elderly subjects is presented in Fig. 4.
We observed increased disorder in the response of elderly subjects. Good performance in these tests corresponds to low root mean squared (RMS) cursor error . A subject's behavior is more efficient if they obtain the same level of performance using less corrective motion, that is, low RMS control input velocity . Significantly, young subjects choose different strategies in this task, leading to a Pareto optimal curve in Fig. 5. We calculated the line fit to the data of the young subjects using the optimal control model by varying the control cost and noise intensity parameters in a coupled way according to the multiplicative/signal-dependent nature of noise in sensorimotor tasks [9],[11]. Despite uniform physiological health among the young, and the instructions to “keep the error as small as possible,” some provided less corrective effort and tolerated larger amounts of error. In repeated trials spaced months apart, some subjects shifted location along this Pareto front, and the outlying young subjects 1 and 10 moved onto this front. In Fig. 5 the elderly fall almost completely to one side of the curve fit to the young. When compared to the young, they move their hand more but accomplish less. Without assuming any model, it is reasonable to conclude that this is not intentional, and that it demonstrates increased noise or disorder in their response.
The elderly also have a lower average RMS control input rate - that is, they move more slowly. The effect is visible in Fig. 5. The value of a two-tailed t-test on the logarithm of the subject means across twelve young vs. twelve old subjects is . Removing subjects that were always poorly fit by the optimal control model (1,10,19,24) yields . The existence of age related slowing is well established elsewhere. In this experiment, it is reasonable to suppose that the high value results from the effort/performance tradeoff seen in the young subjects' data in Fig. 5.
Despite the use of a fitting process, there exist opportunities for the optimal control model to be tested. To be precise, the model parametrizes an infinitesimal fraction of all possible functional control strategies, and thus despite the fitting process it has opportunities to be inconsistent with either the data or with the results of accepted analysis methods. These are enumerated below.
First, a fitted model implies a certain level of RMS cursor error . This is not a prediction outside of the data set, as all the measured variables are coupled by the feedback loop. However, it is evident from theory and experiment that it is possible for the observed RMS to be significantly different from that which would be expected based on the fitted model, and this provides an opportunity to prove the model false. This is can be understood with the following equations, based on zero assumed initial conditions:(4)(5)This indicates that the accuracy of the model's predicted RMS error level rests on the accuracy of its model of the interaction of and - that is, the subject response to the perturbation. By restricting our models to be optimal controllers, we are restricting the choices available to our fitting method to represent this interaction. The ratio of predicted and observed RMS is near one as shown in Fig. 6. There is some loss of accuracy due to the use of a statistical model of endogenous noise over a finite time frame. Again, the outlying behavior by young subjects 1 and 10 was not reproduced on repeated testing.
Second, the model would be inconsistent with the data if the fitted models did not obtain significantly different parameters for different subjects (consistent with their supposed meaning), in cases where direct inspection of the data shows that such differences exist. This is not a difficult test, as the modeling method is of course designed to capture this difference, but it is still useful to inspect the results. Summary statistics of the model parameters in Table 1 indicate that the method is able to detect intersubject differences and yield sensible results. The base 10 logarithm of all parameters was taken for normality. The table gives mean values for the young and old, the aggregate SD, the median intra-subject SD, and the median intra-trial SD. After the four parameters, a parameter proportional to motor noise amplitude normalized to control input amplitude is given. This demonstrates that the age-related reduction in signal to noise ratio implied by Fig. 5 is observed in our parameters. The two-tailed t-test p-value for logarithms of the twelve young subject means vs. twelve elderly subject means for this quantity is . It decreases to if the subjects that were poorly fit by the optimal control model (1,10,19,24) are excluded.
Third, the agreement between observed and fitted Fourier transformed auto- and cross- correlations [36] of input and perturbation for the optimal control model are good. It is difficult to make this statement precise because spectral analysis is an engineering technique to guide model creation rather than a statistical tool for testing hypotheses [36]. There exists a metric resembling the measure, coherence, but it is valid only for a general form of empirical linear model fitting [36]. In Fig. 7 we show spectral data from the trials with median fit cost from young and elderly subjects. The spectral fits of (1,10,19,24) were generally poor and those of 22 were poor for some trials. Remarks: There exist multi-taper methods for constructing experimental spectral estimates with confidence intervals [42] but these rely heavily on assumptions about stationarity of statistical processes. Due to the signal-dependence of noise, these assumptions are invalid. Another commonly used measure of linear dynamical model fit quality is coherence. Conceptually, this is a frequency-dependent -type goodness-of-linear-fit measure [42] but coherence is well-defined (such that it falls in [0,1]) only for empirical data smoothed using certain tapers [36], and not for arbitrary linear models including ours. The experimental spectral data shown in this paper are such smoothed empirical data, and the coherence is invariably nearly one at low frequency and nearly zero over 10 radians per second. In terms of our model, this loss of coherence is due to dominance of endogenous noise. Were the behavior at higher frequencies actually productively structured, and the label of endogenous noise therefore incorrect, then we would expect to see over-predicted RMS error levels. This is not the case generally. Thus low coherence at high frequencies is not a failing of our model, but a manifestation of the limited ability of the subjects to execute productive motion at those frequencies.
Fourth, the optimal control model would be in doubt if its inferred latency values were inconsistent with the result from the cross-correlation method; this is not the case, as described in detail in the next subsection. Along similar lines, the increase in disorder visible in Fig. 5 should be captured in our model's parameters as increased multiplicative noise and therefore reflected in the fitted parameter. Specifically, direct inspection of the data in the left panel of Fig. 5 shows significant differences in RMS control input that should be correlated with age-related changes in the multiplicative noise/control cost parameter . Fig. 8 shows that the expected relationship holds for the young, where it is reasonable to assume that the effect of multiplicative noise on this parameter is relatively uniform due to uniform health, and variations in result only from altered willingness to expend effort. What is more interesting is that Fig. 7 also shows that tends to be larger for the elderly given some level of observed RMS control input . This is consistent with the empirical observation of increased multiplicative noise.
Fifth, the optimal control model would be in doubt if more general linear system identification techniques were able to better fit the data. Even with a restriction to the set of all stabilizing linear controllers with similar state dimension, the optimal controllers represent an infinitesimal fraction of that set, due to restrictions on the LQG cost function implied by our parametrization [38]. We therefore compared the fit costs attained by our model fitting method to those obtained using the more general n4sid [43] model fitting method available in Matlab [40]. The mean ratio of the fit costs was 1.009, and the standard deviation was 0.0062.
Sixth, the optimal control model predicts that there is a volitional degree of freedom in response, parametrized by the control cost component of the parameter . Therefore a population with uniform multiplicative noise properties should fall along a corresponding Pareto front. This is indeed observed among the young as shown in Fig. 5.
Last, we have a catch-all that if some phenomenon exists in the data that is not qualitatively consistent with the model, the model is flawed in that sense. We observe intermittent periods of stillness that are not predicted by our model, as is visible in Fig. 2. These tended to occur in lower-performing subjects, and became very pronounced for the worst-performing elderly subjects. They appear to be either periods of inattention or control deadbands or deadzones similar to those reported in [44]. This discrepancy reflects the approximate nature of our assumption of quadratic optimality - evidently, if the appropriate corrective motion is sufficiently small, it is sometimes ignored. This flaw does not seem important to the phenomena we address in this paper.
Besides consistency with the data, a good model should be parsimonious, and we demonstrate this for young subjects in [25]. Delayed proportional control of limited bandwidth would be a conceptually simpler model of response. For high values of the multiplicative noise/control cost parameter , our control model becomes very similar to this strategy, and indeed the data for lower performing subjects is well described in this way. Proposed methods to fit both optimal and less parsimonious non-optimal models to the behavior of healthy young subjects are surveyed in [45]. Rather than adopting different modeling approaches for different cohorts we simply use the optimal approach throughout.
Based on the above results, we conclude that the optimal control model is appropriate as a model of the behavior of all young subjects, and for the elderly except subjects 19, 24, and marginally 22. The model was consistent with the behavior of all but two young subjects during a first session, and with the behavior of all young subjects upon repeated testing [25].
The inferred mean latencies using the model-based method for young and elderly groups are 260 ms and 247 ms, therefore we cannot attribute elderly impairment to increased response latency. The median intra-subject standard deviations are 14 ms for the young and 20 ms for the elderly, which is significantly less that the variability of the cross-correlation method applied to the same data. The difference in means is not significant: Averaging inferred latencies within each subject, grouping them into 12 young and 12 elderly, and applying the two-tailed t-test yields a p-value of 0.175. A histogram of the model-based inferred response latencies is presented in Fig. 9, aiding interpretation and confirming normality.
This is a computational result obtained by taking the models fitted to the young and increasing the noise bandwidth parameter, the noise intensity parameter, or the multiplicative noise/control cost parameter . The magnitude of the predicted closed loop subject response to the perturbation decreases in each case, and particularly at higher frequencies. This implies slowed response. Within the structure of the optimal controller, this is associated with smaller elements in the LQR gain matrix (in the case of ) and longer timescales in the Kalman Filter (in the case of the noise parameters). This is consistent with both classic results concerning signal-dependent or multiplicative noise [7],[8],[11],[39] and the intuitive suggestion about averaging timescales in [2].
Within our data, the behavioral changes seen in nine of the twelve elderly can be replicated by retuning of the optimal models found in the young to increase noise. This follows directly from the fact that nine of twelve elderly subjects' behavior is well fit by our model without significant changes from latency values from the young group.
We examined the normalized Hankel Singular Values (HSVs) of the optimal control models that we fit to the behavior of our subjects. We found that models fit to the elderly subjects have smaller third and fourth normalized HSVs as shown in histograms in Fig. 10. Taking the logarithms of the mean normalized third and fourth HSVs for each subject and again applying t-tests on young and elderly groups yields p-values of 0.018 and 0.026, indicating that simpler models are fitted by our method to the behavior of the elderly group. Higher HSVs were negligible.
The principal experimental result of this work is that age-related impairment in a dynamic compensatory tracking task takes the form of reduced efficiency of corrective motion without an increase in response latency. This is a robust result in that it can be obtained with a phenomenological cross-correlation method free of assumptions, and reproduced with greater precision by a more sophisticated model-based approach. The optimal control model presented in this paper has testable implications for behavior in this task as surveyed in the results. These indicate that it is relevant as a model of behavior for all young subjects and most elderly subjects. Therefore the model is relevant to analysis of at least the initial phases of age-related changes in sensorimotor response. We then have the computational result that if one adjusts the parameters used in the optimal control models fitted to the young to account for increased noise levels, the control strategy changes in a way consistent with slowing. Thus analysis of the computational model is consistent with the hypothesis that slowing of post-onset response is an optimal adaptation to increased endogenous noise. This hypothesis has been suggested before without formulating an explicit model [2]. This hypothesis has also been suggested based on an analysis involving hypothesized submovements and empirical speed-accuracy tradeoffs [7], rather than Bayesian and quadratic optimality.
Both our empirical and model-based results are inconsistent with the generalized slowing hypothesis of aging as reviewed in [1],[28]. Fig. 5 shows that the elderly are unable to obtain as favorable a speed/accuracy tradeoff as the young. This is inconsistent with the hypothesis that slowing might be due to a choice involving this tradeoff put forward in [46].
Elderly impairment is shown in Fig. 5 to manifest as inefficient, disorderly movement; therefore we try to clarify the definition and role of endogenous noise in this paper. As described in the introduction, we define “endogenous noise” as deviations from optimal behavior, and therefore include suboptimal performance due to sensory, motor, conduction, or neural processing imperfections. A similar definition is unavoidable when defining noise in dynamic response, and it is best viewed as simply self-consistent. For example, in static force production tasks or target reaching tasks, there is implicitly a reasonable and simple model of what the subject “wants” to do, and noise is defined similarly as a deviation of behavior from that model. Any distinction between (a) involuntary cortical “dysfunction,” and (b) voluntary suboptimal or non-functional behavior, is bound to be scientifically unsatisfactory absent some quantitative measure of subject intent. Therefore we make no such distinction. This does not mean that our claim that the elderly have worsened endogenous signal-to-noise rests on our model. Instead we refer the reader to the deviation of the elderly from the Pareto front in Fig. 5, an indication that the effort/performance tradeoff has worsened with age. This can only be due to a worsened mixture of productive behavior (i.e. signal) to unproductive behavior (noise).
We emphasize that optimal modeling does not assume that the subjects behave perfectly. It is better understood as a modeling approach that treats behavior as the sum of the effects of an optimal controller and a random noise source, acting in a feedback loop. The series of successful predictions of the optimal model were detailed in the Results section. Importantly, our report of outlying subjects shows that significantly suboptimal strategies are viable but atypical, and thus that our assumed form of optimality is not a trivial implication of success in the task.
Adaptation of motor control strategies with age has been reported elsewhere [32]. Evidence for altered control strategies in the elderly is presented in [47],which showed that different neural adjustments are used by the elderly during learning in an isometric contraction reference tracking task. In addition, [48] reports the use of co-contraction to suppress noise using EMG measurements. This co-contraction was associated with slowing, and thus suggests one way to implement a slowed and less complex controller.
We investigated the possibility that reduced complexity of elderly control strategies in this task caused reductions in response latency that offset expected aging effects, as suggested by well-documented age-complexity-latency interactions [1],[4],[5],[27],[28],[49],[50]. Complexity has proven to be a “slippery” concept in the psychology of aging [49]. Furthermore, quantifying complexity in a dynamic task is intrinsically more difficult that in a discrete choice task. In our experimental paradigm, unlike in a choice reaction task, the complexity of the subject's response is not constrained. We consider our task as at least as complex as a one-bit choice reaction task in the sense that there are two possible directions that the subject may move their hand, and furthermore the amplitude of the motion must be determined. There have been reports of methods to measure the complexity in a time series (the data format of our dynamic perturbation-rejection task) such as entropy based methods [50],[51]. In the time series entropy paradigm, complexity is a property of the signal rather than the neural control system, and it is determined by the output signal of the system of interest. These entropy methods based on measuring only the output of a system imply a paradoxical assertion that a system that outputs noise (such as thermal noise from a resistor) is more complex than a system that outputs structured signals (such as a brain). The paradox can be resolved by viewing complexity as a property of the system rather than of the signal. In the experiment, we can only infer a system model from the signals input to and output by the subject. This justifies treating complexity in terms of systems' input-output relationships as identified in dynamical modeling. The linearity of optimal response in our experimental paradigm made available the well-developed tools of linear system theory, in particular, the Hankel Singular Value (HSV). Our results show that the elderly employ strategies of reduced controller complexity according to this metric. We report a computational result but draw no immediate physiological conclusions about this for two reasons. First, we can only speculate on the relevance of this result to response latency because the dependent variables of response latency and inferred HSVs have too much variance (even within the homogenous young group) to allow us to demonstrate a HSV-latency effect within our data. Second, higher order terms in computational models can display epiphenomena and generally call for independent confirmation.
Concepts from our modeling approach resemble those in contemporary work involving Parkinson's disease. It was found in [52] that in a pointing task with a penalty associated with deviations from a desired time-to-completion, affected patients were slower than healthy subjects to adjust to requests for increased speed - without any change in endpoint accuracy. This result contrasts with our empirical observation of increased disorder in elderly subjects, though the comparison is necessarily indirect due to the differences in tasks and measures.
In conclusion, the major contribution of the optimal control model is that it quantitatively specifies the extent to which subjects should use slower control strategies to reduce the effects of noise. This is subtly different from less mathematically precise ways of arriving at this reasonable strategy. Specifically our model holds that the slowing should take the form of longer timescales during the response dynamics, and makes no prediction of longer latency until the onset of these response dynamics. This is consistent with the surprising experimental result that there is no increase in latency until onset of response in this relatively complex task.
All subjects completed consent forms approved by Cornell University's Committee on Human Subjects and brief health questionnaires.
The young subjects numbered 1 through 12 were healthy volunteers between 20 and 32 years of age. The elderly subjects numbered 13 through 24 were between 65 and 92 years of age. The manipulandum was a calibrated optical mouse modified to only record lateral hand motion, and mounted on ball bearings to reduce friction. Subjects were free to use the hand with which they felt most comfortable. Two self-reported left-handed subjects chose to use their right hands because they are accustomed to using computer mice with the right hand. This did not seem to affect the data, as the outliers were all using their right-dominant hand. There were ten 60-second trials. The first 9 seconds of data from each trial were discarded. Trials started every two minutes to prevent fatigue. Each trial consisted of a different random perturbation time history, and all subjects had the same perturbation time histories in the same order. The results of the first two trials are neglected to allow for learning effects. All young and all but one elderly subjects' behavior converged in this time, as measured by RMS cursor error and input velocity. The software sampled the control input , added the perturbation , and updated error at a rate of 100 samples/second. This rate ensures that closed loop behavior is unaffected by software-induced delay, and that measurement resolution covers the timescales of motor behavior in the hand and arm.
To assign a specific time value for response latency based on cross correlations, we looked at the peak of the second derivative of the cross correlation with respect to lag index (see Eqn. 3. Because derivatives introduce spurious numerical noise, we smoothed by forward and backwards discrete filtering to obtain before calculating the second derivative. We used a second order Butterworth filter [40] with a 20 Hz cutoff frequency.
Consistent with today's literature, we assume quadratic optimality and Gaussian noises. We present a linear plant to be controlled. The dynamics of our task and the basic dynamics assumed to exist in the human are jointly described by the state space matrix equation:(6)(7)(8)(9)The state includes states from each sub-system: a state from the controlled dynamics in the software, one state for the pre-filter on the external perturbation, two states (vector ) for the pre-filter on the motor noise, four states (vector ) for a fourth order Pade approximation to some delay that is assumed to exist in the human, and one state needed to implement differentiation of the control input over the frequency range of interest for use in the cost function. The inputs to (7) are an ideal noise-free control input , a pre-filter white perturbation with known statistical properties, and a pre-filter white endogenous noise with known intensity. Intensity is defined as variance per sampled bandwidth. The measurement output available to the subject is , where is such that the states involved in the approximation to the delay affect the measurement so that the cursor error (identical to in the general notation) can only be observed after the delay. The measurement noise is . Each of the subsystems is represented above in standard state-space form by , and matrices [38] with appropriate subscripting. The matrices are zero except for and . For example the perturbation pre-filter subsystem in isolation is given by(10)(11)The above equations are in matrix form; sometimes the matrices are scalars, and sometimes is actually a matrix of zeros, as appropriate. The cursor dynamics are given by so that . The control-effort differentiation dynamics are given by . These dynamics implement differentiation of the control input across the frequency range of interest while avoiding technical problems during LQG synthesis.
The other sub-systems (endogenous noise prefiltering and latency ) are now presented. The pre-filters and allow us to solve the relevant equations without unreasonably assuming white noise disturbances in the physical system. The endogenous noise prefilter is a critically damped second order system(12)(13)Our approach lumps all latencies in the system into one system , but in our simple single-input-single-output task this makes no difference. The state space matrices and for vary depending on the parameter, and are non-unique and unenlightening. This is tolerable because they are easily created with commercial software implementation of the Pade approximation [40] for any given latency value and approximation order.
The measurement noise was estimated to be at a negligible level for all subjects by having subjects read the instructions in a very small font. If the measurement noise is taken to be white, this acuity corresponded to a level of noise low enough that optimal response is insensitive to it. Non-white measurement noise models were not pursued because their effects are ultimately indistinguishable in their effects on subject response from changes in the main endogenous noise .
The matrix is large and complex, but commercial software can assemble it automatically given a block diagram and straightforward equations for its components [40].
We assume that the subject behaves so as to minimize the following quadratic cost:(14)where is the expectation operator [38]. This assumption is chosen because it is tractable, consistent with contemporary methods, and a simple way to capture the concept of “reasonable” control strategies. In conjunction with the linear model, this assumption implies Linear Quadratic Gaussian (LQG) optimal control [38]. LQG control is implemented by the series combination of a Kalman filter (a Bayesian optimal state estimator) and a controller that minimizes a cost defined as a quadratic function of input and state variables. The Kalman filter is of the form(15)This equation holds that the state estimate evolves based on a known state space dynamical matrix , the input matrix describing the effects of control inputs on the state, and a Kalman gain times the difference between the observed measurement and the expected measurement. The assumption that the matrices and the noise statistics are known implies an assumption that response is practiced and an internal model is formed. The Kalman gain is obtained by solving a Riccati equation involving and noise variances [38], a process automated in commercial software [40]. The ideal control input is then(16)where is a static gain matrix, similarly obtained by solving a Riccati equation [38] involving and the control cost.
Our model of the response, , is then defined by(17)(18)(19)
Forms of LQG control are widely used in both control engineering and the modeling of sensorimotor response, as described in [9],[17],[22],[25]. In any such approach, the optimal controller is uniquely defined when the controlled system and the statistical properties of all disturbances are specified.
Our model fitting procedure is done in the frequency domain. It is technically equivalent to using a weighted function of the discrepancy between the experimental and predicted values of the cross- correlation of and as well as those of the autocorrelation of in the correlation- or lag-domain. We chose to use the frequency domain for easy comparison of the smoothed experimental cross correlation to the predicted closed loop transfer function from to , because the effects of added noise are visible in a straightforward way in the power spectrum of , and because it lends itself to frequency weighting in the cost function as explained later. A purely time domain fitting process was not used because the expected effect of endogenous noise on the time series is zero, making this key aspect of response invisible to the fitting process. The predicted closed loop transfer function and predicted power spectrum are specified by our model and the four parameters. In order to compare it to experimental data they can be evaluated at the sampling rate multiplied by the discrete frequencies corresponding to the Discrete Fourier Transformed experimental data [36]. This is done in the weighted cost function(20)In this cost function, is the smoothed Discrete Fourier Transformed cross-correlation of and , is the predicted power spectrum of , and is the un-smoothed experimental power spectrum of [36]. The parameter set is represented by . It is appropriate to not smooth the experimental power spectrum because response at low and middle frequencies is dominated by linear response to a perturbation that is known during analysis. The frequency weighting in the cost function emphasizes data at crucial frequencies near the closed loop bandwidth, rather than low frequency behavior that is not sensitive to key features such as latency, thereby reducing the variance of inferred parameters and the quality of the spectral fit in that frequency range. Results were insensitive to varying the weight of across the range of to ; some weight is required to avoid over-fitting at the expense of .
The sensitivity of the predicted closed loop transfer function and the predicted power spectrum to the parameters can be summarized as follows. One can distinguish between direct effects that occur regardless of whether the control strategy adapts, and indirect effects due to changes in the control strategy that are assumed to result from the subjects' understanding of their own latency and endogenous noise. Our method assumes that both direct and indirect effects manifest in the data. Increasing the control cost/multiplicative noise parameter indirectly but strongly decreases and at all frequencies, especially high frequencies. This occurs solely through the its effect on . Increasing the latency parameter has a direct effect on the phase of at higher frequencies, and indirect effects through the Kalman filter design that tend to cancel out some of the changes in phase of across a lower frequency range. These indirect effects are due to the effects of changing (which expresses changes in and ) when solving the Riccati equations for the Kalman filter. Increasing the endogenous noise intensity parameter has a direct effect of increasing , especially in a range near 1–3 radians per second depending on , and an indirect effect of decreasing at higher frequencies, again through effects the Kalman filter design. Increasing the endogenous noise bandwidth parameter directly causes an increase in the predicted at higher frequencies and has relatively small indirect effects on , again through similar effects on the Kalman filter design.
Further details can be found in [25].
Our measure of control strategy complexity is size of the Hankel Singular Values (HSVs) [38] of the fitted optimal subject response. The number of significant HSVs is the effective state dimension or order of a linear system model. For example, one could take reasonably noisy data from a second order system and fit a higher order model to it, but the first two HSVs would be much larger than the rest. In this sense the method resembles Principal Components Analysis for linear dynamical systems. Where Principal Components Analysis involves the Singular Value Decomposition (SVD) of a covariance matrix, HSV analysis looks at the SVD of the product of the controllability and observability Gramians [38]. The controllability Gramian indicates the sensitivity of the system's states to the inputs, and the observability Gramian indicates the extent to which changing the system's states leads to measurable outputs. System states are variables that capture all information from the past relevant to future system behavior. State dimension is essentially identical to a concept already introduced to the aging literature as a measure of complexity, i.e., the number of dynamical degrees of freedom that can be regulated independently [49]. A system state variable is precisely such a dynamical degree of freedom.
The controllability Gramian of the LQG controller in Eqns. 17, 18 is(21)The observability Gramian is(22)
Without altering input-output relationships from to , one can express the LQG controller using a linearly transformed state such that the state equations of are expressed in so-called “balanced” form, characterized by the Gramians being equal, diagonal, and monotonically decreasing along the diagonal. This process is described in [38] and automated in commercial software [40]. Their SVD is then trivial, and the singular value corresponding to each state in indicates its relevance to input-output behavior.
While, to our knowledge, we are the first to use HSVs in the neurophysiological field, they have a long history of use in system and control theory, where they are the standard method to make decisions on what order of model is required to approximate the behavior of dynamical systems [53]. It is often the case that a few states are influential and the remainder may be neglected.
Based on the preceding development, the complexity of implementing the inferred controllers may be quantified by examining the extent to which high order models are required to describe the behavior of the system - that is, whether the higher order HSVs are large. We normalized the magnitude of all HSVs by the subject's mean first HSV in order to eliminate the gross effect of altered gain, which we do not consider a form of complexity (omitting this refinement only amplifies the age differences). The dynamical significance of larger third and fourth HSVs in models fit to young subjects is that simpler low order models are less able to describe their response.
The HSVs are sensitive to all of our model parameters, but respond most strongly to multiplicative noise/control cost : reductions in this parameter are associated with high order models. The HSVs are unique for a given set of the parameter values, because the parameter values fully define the modeled system, and the HSVs are a property of the system.
|
10.1371/journal.pgen.1002148 | The Epistatic Relationship between BRCA2 and the Other RAD51 Mediators in Homologous Recombination | RAD51 recombinase polymerizes at the site of double-strand breaks (DSBs) where it performs DSB repair. The loss of RAD51 causes extensive chromosomal breaks, leading to apoptosis. The polymerization of RAD51 is regulated by a number of RAD51 mediators, such as BRCA1, BRCA2, RAD52, SFR1, SWS1, and the five RAD51 paralogs, including XRCC3. We here show that brca2-null mutant cells were able to proliferate, indicating that RAD51 can perform DSB repair in the absence of BRCA2. We disrupted the BRCA1, RAD52, SFR1, SWS1, and XRCC3 genes in the brca2-null cells. All the resulting double-mutant cells displayed a phenotype that was very similar to that of the brca2-null cells. We suggest that BRCA2 might thus serve as a platform to recruit various RAD51 mediators at the appropriate position at the DNA–damage site.
| Mutations in BRCA1 and BRCA2 predispose hereditary breast and ovarian cancer. Such mutations sensitize to chemotherapeutic agents, including camptothecin, cisplatin, and poly(ADP-ribose) polymerase (PARP) inhibitor, since RAD51 mediators including both BRCA proteins promote repair of DNA lesions induced by these drugs. Little is known of the functional relationships among RAD51, BRCA2, and other RAD51 mediators, because no brca2-null cells were available. Furthermore, the phenotype of sws1 mutants has not been documented. We here disrupted every known RAD51 mediator and analyzed the phenotype of the resulting mutants in both BRCA2-deficient and -proficient backgrounds. The understanding of the function of individual RAD51 mediators and their functional interactions will contribute to the accurate prediction of anti-cancer therapy efficacy.
| Homologous recombination (HR) maintains genome integrity by accurately repairing double-strand breaks (DSBs) that arise during the mitotic cell cycle or are induced by radiotherapy [1], [2]. HR also plays an important role in releasing the replication forks that stall at damaged template DNA strands [3], [4]. Thus, effective HR makes tumor cells tolerant to the chemotherapeutic agents that damage DNA and stall replicative DNA polymerases. Such chemotherapeutic agents include cis-diaminedichloroplatinum(II) (cisplatin), camptothecin, and poly(ADP-ribose) polymerase (PARP) inhibitors, including olaparib (AstraZeneca). Cisplatin is a crosslinking agent that generates intra- and inter-strand crosslinks and thereby stalls replicative DNA polymerases. Camptothecin inhibits the ligation of single-strand breaks (SSBs) that are formed during the normal functioning of topoisomerase 1. Resulting unrepaired SSBs are converted to DSBs upon replication. Similarly, PARP inhibitors interfere with SSB repair [5]. Since HR plays a major role in repairing DNA lesions generated by camptothecin, cisplatin, and PARP inhibitors [6], measuring HR efficiency in individual malignant tumors may help predict the efficacy of these chemotherapeutic treatment for each tumor [7]–[9].
HR-dependent DSB repair is accomplished by the following step-wise reactions [10]. DSBs are processed by the Mre11/Rad50/Nbs1 complex and the CtIP, Exo1 and DNA2 nucleases to develop 3′ single-strand DNA (ssDNA) tails [11]–[17]. RAD51, an essential recombinase, polymerizes on these ssDNAs, leading to the formation of nucleoprotein filaments. These filaments undergo homology search and subsequent invasion into homologous duplex DNA to form a D-loop structure, where they serve as a primer for DNA synthesis [18], [19]. After the extended end is displaced from the D-loop, it anneals to its partner-end to complete DSB repair. We know that RAD51 plays a key role in HR in vertebrate cells, as inactivation of RAD51 results in the accumulation of chromosomal breaks in mitotic cells and inhibits the completion of even a single cell cycle [1], [2]. The polymerization of RAD51 at damage sites is strictly regulated by a number of accessory factors (hereafter called RAD51 mediators), including the five RAD51 paralogs, SWS1, RAD52, SFR1, BRCA1, BRCA2, and PALB2 [3], [20]–[27]. The functional relationships of these RAD51 mediators are poorly understood, because cells deficient in multiple RAD51 mediators have not been established.
BRCA2 was originally identified as a tumor suppressor, as germline mutation of the BRCA2 gene results in a high risk of developing breast, ovarian, pancreatic, prostatic, and male breast cancer [3], [20], [28], [29]. BRCA2 is recruited to processed DSBs, and facilitates the assembly of RAD51 at the single-strand tail. The middle of the BRCA2 protein has eight BRC repeats, comprising 26 amino acids. Biochemical studies have revealed that individual BRC repeats prompt the loading of RAD51 on ssDNA [30], [31]. Since no brca2-null cells have been established, the function of BRCA2 has been postulated from the phenotypic analysis of mice carrying an allele, extending from the N-terminus to the BRC3 motif (hereafter brca2tr) that encodes a truncated form of BRCA2. Cells derived from brca2tr mice and brca2tr DT40 cells are able to proliferate and exhibit increased sensitivity to ionizing radiation, camptothecin, and cisplatin [32], [33]. It remains unclear whether brca2 null cells display the same phenotype as do rad51 cells, or a milder phenotype.
The roles of the RAD51 mediators have been characterized by phenotypic analysis of their mutants. Mammalian brca1-deficient cells show normal focus formation of the RPA ssDNA binding protein but diminished RAD51 focus formation at DSBs, indicating that BRCA1 facilitates the polymerization of RAD51 after the resection of DSBs [15], [34]. DT40 cells deficient in any one of the five RAD51 paralogs show a very similar phenotype, including compromised RAD51 focus formation and the same degree of DNA damage sensitivity [35], [36], suggesting that these five proteins form a functional unit in the promotion of RAD51 polymerization in which each RAD51 paralog is essential for its function. No biochemical studies have yet defined the molecular mechanisms underlying the promotion of RAD51 assembly by BRCA1, the RAD51 paralogs, or SWS1. SWS1 is another RAD51 mediator [26], and sws1-deficient vertebrate cells have not yet been reported. SFR1 was originally identified in fission yeast [25], and its role in in vitro HR reactions [37], [38] as well as the phenotypic analysis of sfr1-deficient mice have been recently reported [39]. The biochemical character of full-length human BRCA2 has been recently documented, whereas no biochemical studies have defined its functional interaction with other RAD51 mediators or the molecular mechanisms underlying the promotion of RAD51 assembly by BRCA1, the RAD51 paralogs, or SWS1 [40]–[42].
In this paper, we addressed the function of Rad51 mediators and their relationship in DNA-damage responses. We generated the single Rad51 mediator mutant cells, including brca2 null, sfr1 and sws1 deficient cells from DT40 chicken cell line [43]. We also disrupted the BRCA1, one of the RAD51 paralogs (XRCC3), RAD52, RAD54, SWS1, and SFR1 in brca2 null deficient DT40 cells. The phenotypes of these cells were analyzed to reveal hierarchical relationship of RAD51, BRCA2, and the other RAD51 mediators, where RAD51 is able to operate HR without BRCA2 while BRCA2 is required for the functioning of the other RAD51 mediators. Hence, BRCA2 might serve as a platform to recruit various RAD51 mediators at the appropriate position of DNA-damage sites. Our study sheds light on the functional relationship of RAD51 and every known RAD51 mediators for the first time, and thereby significantly contributes to the development of effective anti-cancer therapies.
To analyze SFR1 and SWS1, we disrupted the SFR1 and SWS1 genes in DT40 cells (Figure 1A–1D). Table 1 summarizes the selection marker genes we used to disrupt genes in this study. The resulting sfr1 and sws1 mutant clones proliferated with nearly normal kinetics (Figure 2) and exhibited an increase in cellular sensitivity to cisplatin. The sfr1 mutant was sensitive also to camptothecin and olaparib (Figure 3). Both mutants showed a slight but significant decrease in ionizing-radiation-induced RAD51 focus formation (Figure 4). We conclude that SFR1 and SWS1 indeed work as RAD51 mediators, though their contribution to HR-dependent repair is less significant than that of BRCA1, BRCA2, and the RAD51 paralogs.
To create brca2-null cells, we generated compound heterozygous mutant cells (hereafter called BRCA2−/con1 cells) (Figure 1E). The whole coding sequence was deleted in the minus (−) allele of the -/conditional-null allele-1 (-/con1) genotype of the BRCA2−/con1 cells (Coding sequence deletion allele in Figure 1F). We conditionally deleted the con1 allele of the BRCA2−/con1 cells by adding tamoxifen, which activated the chimeric Cre recombinase [28] and thereby eliminated the promoter and coding sequences, including exons 1 and 2 of the con1 BRCA2 allelic gene (BRCA2 conditional-null allele-1 in Figure 1F). At day two of continuous tamoxifen treatment, the vast majority of cells lost the intact BRCA2 gene in the conditional allele, and a substantial fraction of cells began to die. To our surprise, we were able to reproducibly establish clonally expanding cells wherein the conditional BRCA2 allele was deleted (BRCA2−/− cells, hereafter called brca2-null cells) from 10 to 20% of the tamoxifen-treated populations. We verified deletion of the BRCA2 allele in the established clones by Southern-blot (Figure 1G) and western-blot (Figure 1H) analysis. The ability of the brca2-deleted cells to proliferate is in marked contrast to the immediate cell death observed in rad51-deleted cells [1]. The plating efficiency of the brca2-null clones was around 20%, which is significantly lower than that of the wild-type (100%) and brca2tr (60%) cells [33].
One obvious concern with this experiment was that expression of the N-terminal-truncated BRCA2 protein in the Cre-mediated deletion lines could allow for residual function. We therefore created a second version of the conditionally inactivated BRCA2 allele, wherein sequences spanning from the promoter to intron 12 could be eliminated by induction of Cre (BRCA2 conditional-null allele-2 in Figure S1). We exposed the resulting compound heterozygous mutant cells to tamoxifen and confirmed reproducible establishment of BRCA2-deleted clones (Text S1). This second brca2 conditional-null allele supported proliferation with generation times very similar to those of the first version of the brca2-null cells (data not shown). The more extensively deleted brca2 clones showed a phenotype indistinguishable from that of the smaller deletion clones, indicating that both deletions confer the null phenotype.
We assessed the proliferative properties of brca2-null clones by monitoring their growth curve (Figure 2A and 2B) and cell cycle (Figure 2C). Wild-type cells doubled every 8 hours and increased in cell number by 64 times over 48 hours. The brca2-null cells increased by 17 times over 48 hours (Figure 2A), calculated as a 1.6-fold increase over 8 hours (1.66 = 17) (Figure 2B). The number of brca1, brca2tr, and xrcc3 clones [36] increased 1.6, 1.9, and 1.8 fold, respectively, over eight hours. The reduced growth kinetics of the brca2-null cells is partly due to apoptosis in a substantial fraction of the cell population, as evidenced by the accumulation of cells in a sub-G1 fraction (Figure 2C).
To understand the cause of this cell death, we measured the number of chromosome breaks in mitotic cells. Forty-six chromosomal aberrations were detectable in 100 mitotic brca2-null cells, larger than the number of aberrations observed in any other RAD51-mediator mutants (Table 2). We therefore conclude that spontaneously arising DSBs resulted in cell death in a fraction of brca2-null cells, thus accounting for the reduced growth kinetics (Figure 2A and 2B). The viability of the brca2-null cells reveals that BRCA2 plays a less important role than does RAD51 in the maintenance of genome integrity [1]. The high number of spontaneously arising chromosomal breaks in the brca2-null cells (Table 2) shows that BRCA2 plays a more fundamental role in genome maintenance than do any of the other RAD51 mediators.
We analyzed cellular tolerance to camptothecin, cisplatin, and olaparib by measuring cellular survival at 72 hours (7–9 cell cycles) after continuous exposure to these agents in a liquid medium. We did not use the conventional colony-formation assay for this analysis, because the plating efficiency of the brca2-null cells was only 20%, 5-fold lower than that of wild-type cells. Figure 3A presents an example of cellular sensitivity to camptothecin, a DNA-damaging agent. Subsequent figures illustrate the sensitivity of each mutant, assessed by LC50 values, i.e., the dose that reduces cell survival to 50% relative to the LC50 value of wild-type cells, which is defined as 100% (Figure 3B–3E).
In the cellular-survival analysis, the brca2-null cells showed an increased sensitivity to camptothecin (Figure 3B), cisplatin (Figure 3C), and olaparib (Figure 3D and 3E). Moreover, sensitivity to cisplatin and olaparib was higher with the brca2-null cells than for any of the other HR mutant cells, including the brca1, rad52, rad54, and xrcc3 clones (Figure 3). We therefore conclude that BRCA2 plays a more important role in HR-dependent repair than do the other RAD51 mediators, including BRCA1, RAD52, RAD54, the RAD51 paralogs, SFR1, and SWS1.
The less prominent phenotype of the brca2tr cells compared to the brca2-null cells indicates that the BRCA2 BRC3-truncated protein retains significant residual HR function. Although brca1 cells were less sensitive to cisplatin and olaparib than were brca2-null cells, the brca1-null cells exhibited a slightly higher sensitivity to camptothecin than did the brca2-null cells (Figure 3A). The greater contribution of BRCA1 to cellular tolerance to camptothecin might be attributable to the role played by BRCA1 in DNA-damage responses other than HR, such as collaborative action with CtIP to eliminate covalently bound oligo-peptides from DSBs [15].
We next measured the frequency of HR-dependent repair of I-Sce1-mediated DSBs in a recombination substrate, SCneo, inserted into the OVALBUMIN locus [44], [45] (Table 3). The frequency of HR-dependent DSB repair was calculated as the number of neomycin-resistant (neo+) colonies relative to the number of plated cells. The frequency of HR in the brca2-null, brca2tr, and brca1 cells was decreased by 1.5×104-, 1.5×102-, and 4.5×103-fold, respectively, compared with wild-type cells. We conclude that the brca2-null cells retain residual HR activity, which may account for their viability even in the complete absence of BRCA2.
Since BRCA2 promotes the loading of RAD51 at damage sites, we measured RAD51 focus formation at 3 hours after ionizing radiation. The number of RAD51 foci was reduced but not eliminated in the brca1 and brca2tr clones, compared with wild-type cells (Figure 4). These findings are consistent with previous observations [33], [46]. By contrast, we hardly detected any RAD51 focus formation in the brca2-null cells. In conclusion, the BRCA2 protein plays a key role in the efficient recruitment of RAD51 to DNA-damage sites, but is not essential for every HR reaction.
The idea that RAD51 carries out HR even without BRCA2 led us to investigate whether or not other RAD51 mediators substitute for BRCA2 in the promotion of RAD51-dependent HR. To this end, we deleted the BRCA1, RAD52, SFR1, SWS1, and XRCC3 genes in the conditional brca2-null background, then inactivated the BRCA2 gene by treating the cells with tamoxifen (Figure 1E). We also disrupted the RAD54 gene in the conditional brca2-null background (Table 1). The RAD54 protein promotes HR after the assembly of RAD51 at DNA-damage sites [47]. To our surprise, we were able to reproducibly establish all resulting double mutants, although a substantial fraction (∼30%) of the brca2-null cells died each cell cycle.
The growth kinetics for the brca1/brca2-null, rad52/brca2-null, sfr1/brca2, sws1/brca2-null, and xrcc3/brca2-null double-mutant clones was similar to those of the brca2-null single mutant (Figure 2). Taking the very severe phenotype of brca1 cells into account, the viability of the brca1/brca2-null cells was surprising. The cloning efficiency of the brca1/brca2-null cells was slightly higher than that of the brca2-null single-mutant cells (30% compared to 20%). Accordingly, the number of spontaneous chromosomal aberrations in the brca1/brca2-null cells was consistently slightly lower than that in the brca2-null cells (Table 2). In summary, although the loss of either BRCA1 or BRCA2 greatly increased the number of spontaneous chromosomal breaks, inactivation of BRCA1 in the brca2-null cells resulted in a slight reduction in the severity of the brca2 phenotype.
An early study shows rad52/xrcc3 double-mutant cells are synthetic lethal and exhibit numerous chromosomal breaks [24], whereas we here found that the brca2/rad52 and brca2/xrcc3 double-mutant cells were viable. Thus, the synthetic lethality might be attributable to BRCA2 mediated formation of toxic HR intermediates, because the brca2/xrcc3 cells exhibit spontaneously arising isochromatid breaks, where two sisters are broken at the same site due to defective completion of HR [48]. To test this hypothesis, we conditionally inactivated the BRCA2 gene in the rad52/xrcc3 cells (Text S1). We found that the inactivation of the BRCA2 gene indeed rescued the rad52/xrcc3 cells (Figure S2). This observation indicates that the synthetic lethality of the rad52/xrcc3 cells does not argue against the idea that the functioning of RAD52 and XRCC3 depends on BRCA2. Likewise, the formation of toxic HR intermediates might explain the apparent discrepancy between the viability of rad52/brca2-null DT40 cells and the mortality caused by shRNA mediated depletion of RAD52 in brca2 deficient mammalian cells [49], as the latter cells express a residual amount of RAD52 and truncated BRCA2 proteins perhaps leading to the formation of toxic HR intermediates.
We next measured the sensitivity of the brca1/brca2-null, rad52/brca2-null, rad54/brca2-null, sfr1/brca2-null, sws1/brca2-null, and xrcc3/brca2-null double-mutant clones to camptothecin, cisplatin, and olaparib (Figure 3 and Figure 5). Remarkably, inactivation of any gene did not increase cellular sensitivity to the three damaging agents by more than two-fold. This observation indicates that the contribution made by BRCA1, the RAD51 paralogs, RAD52, RAD54, SFR1, and SWS1 to HR depends mostly on BRCA2. Interestingly, the loss of BRCA1, SFR1, and SWS1 somewhat increased the cellular tolerance of the brca2-null cells to cisplatin. Similarly, the loss of SWS1 increased the cellular tolerance of the brca2-null cells to camptothecin and olaparib. This increased tolerance was not accompanied by the upregulation of RAD51 focus formation (data not shown). We therefore suggest that, in the absence of BRCA2, SWS1 has a moderately antagonistic effect on HR-dependent repair. By contrast, the loss of RAD52 and XRCC3 significantly increased the cellular sensitivity of the brca2-null cells to olaparib. In summary, BRCA2 is required for all the analyzed RAD51 mediators to function, and the functional relationships between BRCA2 and the other RAD51 mediators in HR-mediated repair differ slightly depending on the type of DNA damage.
In this study, we established brca2-null cells as well as cells deficient in each of the RAD51 mediators. We show that BRCA2 plays a more important role in the promotion of both RAD51 polymerization at DNA-damage sites and HR-dependent repair than does any other RAD51 mediator, including BRCA1, the RAD51 paralogs, RAD52, SFR1, and SWS1. The ability of brac2-null cells to proliferate is in marked contrast with the immediate cell death that occurs upon depletion of RAD51 [1]. Therefore, RAD51 is able to perform HR even in the absence of BRCA2. To explore the question of which RAD51 mediators might substitute for BRCA2 in the promotion of RAD51-dependent HR repair, we inactivated the RAD51 mediators in brca2-null cells. Loss of any one of the other RAD51 mediators did not further reduce the viability of brca2-null cells. In a related study, we also found that the brca2-null mutant and the palb2/brca2-null double mutant showed the same phenotype with respect to both spontaneous chromosomal aberrations and increased sensitivity to DNA-damaging agents (manuscript in preparation). Thus we conclude that BRCA1, PALB2, the RAD51 paralogs, RAD52, SFR1, and SWS1 all require BRCA2 to contribute to HR.
Data on Ustilago maydis [50] and Arabidopsis thaliana [51] suggest that BRCA2 might be essential for RAD51 to function in any HR reaction. However, we here report that RAD51 can form HR products even in brca2-null cells, indicating that RAD51 plays a more important role than BRCA2 in HR. This hierarchy between RAD51 and BRCA2 is supported by previous reports of experiments with mice, as rad51 null embryos died earlier (∼E6.5) than did BRCA2 null (∼E8.5) embryos [52], [53]. The viability of brca2-null DT40 cells is consistent with the clonal expansion of BRCA2-deficient cells derived from mammary epithelial lineage-specific or T cell lineage-specific BRCA2-null-deficient mice [54], [55]. Adding to these findings, we here show solid evidence that vertebrate RAD51 is capable of functioning in the absence of BRCA2.
The phenotypic analysis of brca1, brca2-null, and brca1/brca2-null clones, combining with the previous study of rad51-null cells, reveals the functional relationship described as follows. The capability of HR was dramatically diminished when either BRCA1 or BRCA2 was absent, indicating that the collaboration of BRCA2 and BRCA1 is required for efficient HR events. brca2-null cells exhibited more prominent defects in HR than did brca1-null cells, indicating that BRCA2 can function in HR independently of BRCA1. Moreover, BRCA2′s contribution to the repair of cisplatin-induced interstrand crosslinks is more significant than BRCA1, which is likely attributable to the fact that BRCA2, but not BRCA1, functions in the Fanconi anemia repair pathway [56]. BRCA1 has additional functions other than in HR, such as mediating the damage checkpoint and processing DSBs [15], [57]. The fact that rad51-null cells have a considerably stronger phenotype than brca2-null cells indicates that RAD51 could still perform HR-dependent repair in brca2-null cells.
The phenotypic similarities between the brca2-null and the brca1/brca2-null clones indicate that BRCA1 contributes to HR by collaborating with BRCA2. Presumably, the two BRCA proteins form a functional unit and collaborate intimately to load RAD51 at damage sites. This idea is supported by the fact that BRCA1 physically associates with BRCA2 through the PALB2 protein [58]. However, this idea is challenged by recent studies that suggest that BRCA1 plays a role in the resection of DSBs [14], [59]. One possible scenario is that the complex formation of BRCA1 and BRCA2 may allow for close collaboration between the BRCA1-dependent resection of DSBs and the subsequent loading of RAD51 on the resulting 3′ overhang. Such an interaction interface might be shared by the E. coli RecBCD complex, which serves as the DSB resection complex and also interacts directly with RecA following chi site recognition [60]. In summary, the phenotypic analysis of brca1, brca2-null, and beca1/brca2-null DT40 clones demonstrates that BRCA1 controls RAD51 in HR, mainly through collaboration with BRCA2.
Our study reveals that rad52/brca2-null, sfr1/brca2-null, sws1/brca2-null, and xrcc3/brca2-null clones exhibit a phenotype very similar to that of brca2-null cells (Figure 5). In a separate study, we conformed phenotypic similarity between brca2-null and palb2/brca2-null clones (data not shown). We therefore suggest that, like BRCA1, PALB2, the RAD51 paralogs, RAD52, SFR1, and SWS1 are also able to participate in HR, mostly depending on BRCA2. One possible scenario is that BRCA2 is recruited to DNA-damage sites through PALB2 or by directly interacting with the junction between the duplex DNA and the single-strand sequences [61]. BRCA2 might thus serve as a platform to recruit various RAD51 mediators to the appropriate positions of DNA-damage sites (Figure 6).
DT40 is a unique cell line that offers a panel of DNA-repair-deficient isogenic mutants derived from a stable parental line. DT40 cells have several characteristics that affect cellular responses to anti-cancer agents. First, DT40 appears, for unknown reasons, to possess a significantly higher HR efficiency than any mammalian cell line [43]. The efficient HR in DT40 cells is prominent particularly in HR between diverged homologous sequences such as Immunoglobulin V gene diversification [62] and gene targeting, where the selection marker genes of gene-disruption constructs may interfere with HR as heterologous sequences. Second, like many cancer cells, DT40 lacks the functional p53, and as a result has no G1/S damage checkpoint [63]. In addition, 70% of the DT40 cell cycle takes place in the S phase. Thus, DNA damage at any phase of the cell cycle may have a direct impact on DNA replication. These characteristics, specific to DT40, suggest that a defect in DNA repair associated with DNA replication, including HR-mediated DNA repair, may display a more prominent phenotype in DT40 cells than in other cell lines that have a longer G1 phase and/or a normal G1/S checkpoint. Bearing this in mind, DT40 is revealed as a unique and valuable tool and has been used extensively to explore the role of individual HR factors responsible for cancer therapy.
Cells were aquired and cultured as described previously [1], [43]. All mutants were isolated from single colonies. DNA transfection and selection were performed as described previously [43], [64]. Details of the cell lines used in this study are shown in Table 1.
To disrupt the SFR1 gene, we generated SFR1-puro and SFR1-bsr disruption constructs by combining two genomic PCR products with the puro- or bsr-selection-marker cassette. Genomic DNA sequences were amplified using the 5′-CCCGGTACTGAGGGGTGCGATTGCTTGCAGG-3′ and 5′-CCCTTAGAGTTGCACTCATTGGCTAAAG-3′ primers for the upstream arm, and the 5′-GGCTCAAACTGGTCAAGATGTACCGATCTAAGG-3′ and 5′-CCACCAGCATCCACTAAAGGGCAAGGAACG-3′ primers for the downstream arm. Amplified PCR products were cloned into pCR2.1-TOPO vector (Invitrogen). The 1.7 kb fragment of the upstream arm was cloned into the KpnI site of pCR2.1 containing the 3.0 kb downstream arm. Marker-gene cassettes were inserted at the BamHI site of the resulting plasmid.
To generate SFR1−/− cells, SFR1-puro and SFR1-bsr disruption constructs lineralized with NotI were transfected sequentially by electroporation (Bio-Rad). The genomic DNA of the transfectants was digested with both BamHI and EcoRI, and gene-targeting events were confirmed by Southern blot analysis. The probe was prepared from a PCR-amplification of DT40 genomic DNA using primers 5′-GAACAGCACCACGCAATTCA-3′ and 5′-CCTTAGAGTTGCACTCATTGG-3′.
Chicken SFR1 cDNA was isolated by PCR amplification of the primary cDNAs using the 5′-GTTGAGATGGAGGAAGCAGCGTGTGGTAAA-3′ and 5′-CACCACTCAATTCCACTTCAAAGAG-3′ primers. The gene bank accession number of the chicken SFR1 gene is XM-001234167.
To disrupt the SWS1 gene, we generated SWS1 gene-disruption constructs containing the 2.6 kb upstream and the 3.0 kb downstream genomic fragments. The 2.6 kb fragment was PCR-amplified using the 5′-ggggacaactttgtatagaaaagttgTTCTTACGTCACTCCAGAAGAACA-3′ and 5′-ggggactgcttttttgtacaaacttgCCAAGTCTGTGAATCGCAGAAGCA-3′ primers. The 3.0 kb fragment was PCR-amplified using the 5′-ggggacagctttcttgtacaaagtggAATTCCAAGCAGTTCCACATCTCT-3′ and 5′-ggggacaactttgtataataaagttgGTATGGCTCCTGTCAGGTTAGAGT-3′ primers. Note that the underlined sequences denote the recognition sequences in the Gateway system (Invitrogen). Using the MultiSite Gateway system with pENTR-lox-his, pENTR-lox-puro and pDEST-DTA-MLS [65], a floxed his or puro gene was inserted between the upstream and downstream arms on a plasmid carrying a diphtheria toxin A (DT-A) gene, thus yielding the two targeting vectors, SWS1-his/loxP and SWS1-puro/loxP.
To generate SWS1−/− cells, SWS1-his/loxP and SWS1-puro/loxP gene-disruption constructs linearized with AscI were transfected sequentially into DT40 cells (Bio-Rad). The genomic DNA of the transfectants was digested with both EcoRV and NotI, and gene-targeting events were confirmed by Southern-blot analysis. The probe was prepared by PCR-amplification of chicken genomic DNA using the 5′-GCTCGCAGGAACACAACTCCTT-3′ and 5′-GTACAGGAGTGTTTCTCTGCGG-3′ primers.
The gene bank accession numbers for the human and chicken Sws1 genes are XP-058899 and XP-415841, respectively. RT-PCR of DT40 transcripts was done using the 5′-CGCGTCGACATGGATAGCACCTTACCAGCT-3′ and 5′-CGCGGATCCTCATCCTTCATCCTCTTCCTC-3′ primers.
The brca2-null mutant cells were generated as follows (Figure 1). We inserted conditional brca2 heterozygous cells (BRCA2+/con1) harboring two loxP signals into the other allele upstream of the promoter and downstream of exon 2. Construction of the BRCA2 conditional-null targeting vector was carried out as described previously [33]. To delete the intact allele of the BRCA2+/con1 cells, we constructed a targeting vector to delete all exons of the BRCA2 gene. The ∼6. kb and ∼3.5 kb fragments at the BRCA2 locus [66] were amplified from DT40 genomic DNA by using the 5′-CCGCTCGAGTTTTGTTAGTTGTGAGATGTG-3′ and 5′-TTATCGGGGCTTTGTCAGCTTTAGCTTCTC-3′ primers and the 5′-CGGAGTTGAATAATGGTACATTTCTGGCAC-3′ and 5′-GTTGAATTTGAAACTGGCTGAACAGAAGAG-3′ primers, respectively. Both fragments were cloned into TOPO-pCRXL cloning vector (Invitrogen, Carlsbad, California) to make the topo/6.0 kb and topo/3.5 kb vectors. The ∼5.2 kb NotI fragment from the topo/6.0 kb vector was inserted into the NotI site in the multicloning site of the topo/3.5 kb vector, resulting in the pUpper/Lower vector. Finally, a loxP-flanked puro-resistance cassette was inserted into the BglII site in the pUpper/Lower vector. The resulting targeting construct was transfected into the BRCA2+/con1 cells followed by selection with puromycin. The genomic DNA of the transfectants was digested with XbaI, and gene-targeting events were confirmed by Southern-blot analysis with a probe that was amplified from genomic DNA using the 5′-ATCCATGTCACTGTTGACATCCTGACTGCC-3′ and 5′-AGATACAAACCCAATGGGAAGCCAGGTGTG-3′ primers. The bands detected by the probe were 8.6 kb from the wild-type allele and 5.2 kb from the targeted allele.
Upon exposure of the BRCA2−/con1 cells to tamoxifen, an estrogen antagonist, nucleotide sequences, including promoter and coding sequences encoding the initiation codon to the 67th amino acid, were excised by a chimeric Cre recombinase fused to the estrogen-receptor ligand-binding domain [24], leading to the complete disruption of the BRCA2 gene.
To disrupt RAD51 mediator genes in BRCA2−/− DT40 cells, we disrupted each gene in the BRCA2−/con1 cells (Table 1). We exposed the resulting RAD51 mediator−/−/BRCA2−/con1 cells to tamoxifen and isolated the RAD51 mediator−/−/BRCA2−/− cells.
Western blotting was performed as previously described [33]. The rabbit polyclonal primary antibody, which recognizes the N-terminal 203 amino acids of chicken BRCA2, was diluted 1∶100 with blocking buffer. The anti-rabbit IgG HRP conjugated antibody was diluted 1∶5000 with blocking buffer.
To measure growth kinetics, cells were counted daily using flow-cytometric analysis, as described previously [7]. To measure cell-cycle distribution, cells (5×105/ml) were labeled for 10 minutes with 20 µmol/L 5-bromo-2′-deoxyuridine (BrdU) and subsequently harvested. Harvested cells were fixed and analyzed as previously described [7].
To measure cellular survival, cells (1.5×103–1.5×104) were incubated in 1 ml culture medium per well containing various concentrations of the DNA-damaging agents. At 72 hours, the ATP in the cellular lysates was measured to assess the number of live cells. The camptothecin (TopoGen Inc., Colombus, OH) and olaparib (AstraZeneca) were diluted with DMSO, and the cisplatin (Nihonkayaku, Tokyo, Japan) was diluted with PBS. To measure the sensitivity of the DT40 cell lines to these agents, cells were continuously exposed to various concentrations of the drug and the number of cells was measured at 72 hours. At least three independent experiments were carried out. Sensitivity was calculated by dividing the number of cells treated with the drug by the number of untreated cells [8].
To assess cell numbers after treatment with the genotoxic reagents, we measured the amount of ATP in the whole cell lysate [67].
Cells were harvested at 3 hours after gamma irradiation. Cells were spun onto slides using a Shandon Cytospin 3 centrifuge (Shandon, Pittsburgh, Pa.). Staining and visualization of RAD51 foci were carried out as previously described [34] using rabbit polyclonal antibody, which recognizes human RAD51, at a dilution of 1∶500 (Calbiochem, San Diego, CA San Diego, CA), and Alexa Fluor 488 goat anti-human IgG antibody at a dilution of 1∶1000 (Molecular Probes Inc., Eugene, OR [34]).
Measurement of chromosomal aberrations was performed as described previously [68].
Measurement of recombination frequencies for I-SceI-induced DSB repair was performed as described previously [34], [44]. Modified SCneo was inserted into the previously described OVALBUMIN gene construct and targeted into the OVALBUMIN locus in wild-type, xrcc3, brca1, brca2, and brca2tr DT40 clones. For transient transfections, 1×107 cells were suspended in 0.5 ml of phosphate-buffered saline, mixed with 30 µg of I-SceI expression vector (pCBASce) or pBluescript KS without linearization, and electroporated at 250 V, 960 microfarads. At 24 hours after electroporation, the cells were plated in 96-well plates with or without 2.0 mg/ml neomycin analog (G418). The cells were grown for 7 to 10 days, after which formed colonies were counted. HR frequency was calculated by dividing the number of neomycin-resistant colonies by the number of plated cells.
Survival data were log-transformed giving approximate normality. Analysis of covariance (ANCOVA) was used to test for differences in the linear dose-response curves between wild-type and a series of mutant cells or brca2-null cells and a series of double-knockout mutant cells. Viability of the DT40 cells was estimated using regressing curves. Regression-curve equations were used to calculate LC50 (50% lethal concentration) values. Relative LC50 values were normalized according to the LC50 value of the parental wild-type cells.
|
10.1371/journal.pcbi.1000059 | The Emergence and Fate of Horizontally Acquired Genes in Escherichia coli | Bacterial species, and even strains within species, can vary greatly in their gene contents and metabolic capabilities. We examine the evolution of this diversity by assessing the distribution and ancestry of each gene in 13 sequenced isolates of Escherichia coli and Shigella. We focus on the emergence and demise of two specific classes of genes, ORFans (genes with no homologs in present databases) and HOPs (genes with distant homologs), since these genes, in contrast to most conserved ancestral sequences, are known to be a major source of the novel features in each strain. We find that the rates of gain and loss of these genes vary greatly among strains as well as through time, and that ORFans and HOPs show very different behavior with respect to their emergence and demise. Although HOPs, which mostly represent gene acquisitions from other bacteria, originate more frequently, ORFans are much more likely to persist. This difference suggests that many adaptive traits are conferred by completely novel genes that do not originate in other bacterial genomes. With respect to the demise of these acquired genes, we find that strains of Shigella lose genes, both by disruption events and by complete removal, at accelerated rates.
| Changes in genetic repertoires can alter the adaptive strategy of an organism, especially in bacteria, in which genes are continually gained and lost. Mapping the gains and losses of genes in the densely sequenced clade of Escherichia coli and Shigella shows that these genomes harbour two types of acquired genes: HOPs, which are those acquired genes with homologs in distantly related bacteria; and ORFans, which are genes without any known homologs. Surprisingly, the two classes of acquired genes display very different patterns of gain and loss. HOPs are acquired more frequently, though they rarely persist in the recipient genomes. In contrast, ORFans are much more likely to be maintained over evolutionary timescales, suggesting that despite their unknown origins, they will more often confer novel and beneficial traits to the recipient genome.
| The wide variation in bacterial genome sizes was originally detected in the 1960s by DNA reassociation analyses [1]. And because bacteria have gene dense chromosomes, the differences in genome sizes implied that there were likely to be vast differences in the gene contents of bacterial species. With the current availability of hundreds of complete genome sequences, it is now possible to establish exactly which genes are present in, as well as those that are absent from, a genome. Among sequenced bacterial genomes, gene sets vary over 40-fold, ranging from 182 genes in the gammaproteobacterial symbiont Carsonella ruddii [2] to almost 8000 genes in the soil-dwelling acidobacterium Solibacter usitatus (jgi.doe.gov).
The wide variation in genome sizes and gene contents can also be observed between strains within individual bacterial genera or species. For example, isolates of Frankia that are more than 97% identical in their rRNA sequences–the conventional cutoff value for a bacterial species–can differ by as many as 3500 genes, which represents nearly half of their 7.5 Mb genome [3]. Even among bacterial strains of similar genome sizes, there can be substantial differences in gene repertoires [4]. Unlike mammals, in which only about 1% of the genes in a genome are unique to a taxonomic order (e.g., mouse vs. human [5]), the gene contents of bacterial genomes can change rapidly over relatively short evolutionary distances.
The generation of novel gene repertoires is a consequence of the ongoing processes of gene acquisition and gene loss [6]–[10]. Although several mechanisms can generate new genes [6],[11],[12], the novel gene sets observed in closely related bacterial strains result largely from gene transfer from distant sources, as duplications and gene rearrangement only rarely produce entirely unique genes in the short timescales in which bacterial gene sets evolve. Although homolog searches indicate that many genes arise from lateral transfer from other bacteria, most bacteria also contain genome-specific sets of genes that lack any homologs in the known databases (termed “ORFans”) [4],[13],[14]. Counteracting the augmentation of bacterial genomes by gene acquisition, gene loss occurs both through large-scale deletions [15] as well as by smaller changes that erode and inactivate individual genes [7],[9],[16]. As observed for acquired sequences, prokaryotes also contain genome-specific sets of inactivated genes (i.e., pseudogenes), which can comprise up to 41% of their annotated genes [17].
Taken together, these lineage-specific gene repertoires indicate the need to monitor bacterial genome dynamics–i.e., the manner in which genes are gained and lost–over short evolutionary timescales. To this end, comparisons of closely related strains of Bacillus [18], Staphylococcus aureus [19] and E. coli [7],[20] have shown that gene acquisitions are prevalent at the tips of the phylogeny and that recently acquired genes seem to evolve more quickly. However, few studies have examined the fate of these genes within a bacterial lineage or have asked how many or which classes of genes, once acquired, are maintained, disrupted or removed from a genome. We address these questions by assessing the differences in gene repertoires among 13 sequenced strains of E. coli/Shigella clade. These strains are closely related, yet display substantial differences in genome size and gene content [21],[22], allowing us to pinpoint the introduction and persistence of genes in the lineages leading to these genomes.
We reconstructed the phylogeny of 13 sequenced strains of E. coli and Shigella species based on the concatenated sequences of 169 conserved, single-copy genes. The relationships and branching orders are well-resolved, well-supported, and congruent with previous studies [23]. The overall branching order of the resulting phylogeny is very similar to those based on other characters or for more limited sets of sequenced strains [24]–[26]: the uropathogenic E. coli (UPEC) form a monophyletic cluster at the base of the tree, and Shigella strains are polyphyletic, with a major lineage derived from the clade containing E. coli K-12 [27]. Based on this tree, we delineated 12 monophyletic clades of varied phylogenetic depths (of which we designate the corresponding ancestral branches as S1 to S4, SC, SCE, C1, E1, E2, U1, U2 and the ancestral branch, Figure 1), which were used to trace the evolutionary history of all genes in these 13 genomes.
By identifying the homologs of genes from the 13 E. coli and Shigella strains in each of 367 microbial genomes, and by mapping the gene distributions in a phylogenetic context, we could infer the ancestry (vertical or horizontally acquired) and dynamics (incidence of acquisition or loss) of genes among strains. Acquired genes were classified into two categories: ORFans, which are genes that have no homologs outside of the analyzed E. coli and Shigella strains, and HOPs, which are genes that have homologs outside of the analyzed E. coli and Shigella genomes but are not ancestral to all taxa containing the gene or whose phylogenetic distributions can not be most parsimoniously reconstructed solely through gene loss events.
From the 13 sequenced strains within the E. coli/Shigella clade, we identified a total of 1443 ORFan gene families and 652 HOP gene families (a family is a group of homologs). Gene family sizes ranged from one gene, for ORFans or HOPs present in a single genome (representing 11% and 32% of the total number of families, respectively) to 13 genes, for ORFans or HOPs with homologs present in all 13 genomes (representing 13% and <1% of the total number of families, respectively). We inferred the branch on which ORFans and HOPs originated by reconstructing the most parsimonious series of events that would give rise to their present-day distributions. By this approach, all HOPs could be assigned to a particular clade, but only 1177 of the 1443 ORFan families were assigned unequivocally, and together these constitute the set considered in subsequent analyses.
Only 8 ORFans (<1%) could be classified to a particular COG category, whereas 151 (23%) of the HOP families could be assigned to a COG other than ‘poorly characterized’: these included Metabolism (10%), Cellular Processes and Signaling (8%, mostly in the category Cell Wall/Membrane/Envelope Biogenesis) and Information Storage and Processing (5%) (Supplementary Table S1).
The numbers of acquired ORFans and HOPs vary substantially across strains and lineages (Figure 1), with the largest difference occurring in the gene set acquired by the ancestor to all tested strains in which ORFans are approximately four times more common than HOPs. This is in contrast to genes confined to a single E. coli or Shigella genome, where we identify ∼40% more HOPs than ORFans. This difference is not affected by the fact that 20% of ORFans could not be placed onto a specific branch, because singleton ORFans are among the easiest genes to assign. Taken together, these distributions suggest that HOP genes originate more frequently, but ORFans are more likely to persist.
Overall, ORFans constitute between 9% and 14% of the protein coding genes per genome, and HOPs account for at most 5% of the protein coding genes per genome. Cumulatively, ORFans outnumber HOPs; however, HOPs represent a larger proportion of the acquired DNA in all strains as they are, on average, longer than ORFans (853 bp vs. 308 bp respectively) (Table 1). There is an association between genome size and the amount of ORFan and HOP-derived DNA (r2 = 0.75 and r2 = 0.72, respectively) per genome; however, it is not simply a matter that the strains with the largest genomes have acquired the most DNA. For example, Shigella dysenteriae and E. coli EDL933 have gained identical amounts of DNA from ORFans and HOPs despite an 800 kb difference in their genome sizes.
ORFans are more A+T-rich than HOPs (44% vs. 47% G+C, respectively), and such differences in base composition are evident along most lineages (Supplementary Figure S1). When examining a single lineage at increasing phylogenetic depths, there is no clear trend towards increased G+C contents, G+C content of the third codon position or increased gene lengths of ORFans or HOPs with duration in the E. coli genome, although this has been observed previously for acquired genes assessed over substantially longer evolutionary timescales [20]. This indicates that the elapsed time since the divergence of the 13 tested strains from their common ancestor has been insufficient to adjust acquired genes to the nucleotide composition of their host genome.
Recently acquired ORFans and HOPs occur more often in multigene clusters than do those assigned to older branches. For example, in E. coli CFT073, which contains the largest numbers of both ORFans and HOPs, about half of the ORFans confined to this strain are adjacent to another ORFan of the same age. Going back to the next branch that subsumes this strain (U2), only a third of the ORFans reside next to another ORFan; and among those ORFans originating in the ancestor to the E. coli/Shigella clade, only 14% are situated next to another ORFan. The average cluster sizes of ORFans along these three branches are 1.59, 1.34, and 1.09 genes, indicating that ORFan genes are gained in clusters that subsequently shrink through fragmentation and gene loss. For the same lineages, a similar trend is observed for HOPs, although it is not as pronounced (with 1.34, 1.34 and 1.21 genes per cluster for singleton, U2-specific and ancestral HOPs, respectively). This decrease in gene cluster size is not due to the preferential insertion of new genes near older acquired genes, as we analyzed the cluster sizes of sets of ORFans and HOPs per introgression event (i.e., those originating on the same internal branch). A few of the clustered ORFans were located near genes of known phage functions, but a recent exhaustive study into viral ORFans has suggested that phages may play a lesser role in transferring ORFans to prokaryotes than previously thought [28].
Since the split from their common ancestor, the 13 E. coli and Shigella species have accumulated between 180 and 350 kb of foreign DNA per strain (Table 1). Aside from these additions, each of these strains has also lost between 30 and 190 kb of DNA that has been acquired and maintained in other strains. The two EHEC strains (E. coli EDL and Sakai) show the highest net gain of DNA, whereas the Shigella strains, E. coli K-12 and W3110 show the lowest.
To compare the rates at which lineages vary in rates of DNA gain and loss, we calculated the amounts of DNA acquired and lost in relation to the branch lengths in the tree relating the 13 tested genomes. The rates on individual branches indicate that closely related strains can differ by over two orders of magnitude in the rates at which newly acquired DNA is gained and retained (E. coli Sakai vs. S. dysenteriae) but less than 20-fold in the rates at which such DNA is lost (E. coli EDL933 vs. S. boydii) (Supplementary Table S2). It should be noted that branch lengths can also vary for other reasons (such as variation in substitution rates and differing rates of recombination), but these are most likely compensated due to the extensive gene set employed here.
Gene acquisition rates for both ORFans and HOPs are higher on the internal branches leading to the EHEC and UPEC strains, and in contrast, rates of loss for acquired DNA are highest on all branches descending from the SC ancestor leading to the Shigella species. Taken together, strains that gain the lowest amounts of DNA, lose the highest amounts of acquired DNA with the result that their genomes have lower numbers of unique genes.
There has been a continual gain and loss of ORFans and HOPs during the evolution and diversification of E. coli (Supplementary Figures S2 and S3), and based on the distribution of ORFans and HOPs in the 13 tested genomes, HOPs have a higher rate of origination, but ORFans are more likely to be retained.
Since many, possibly most, genes are transient and not present in any contemporary genome, it is not possible to monitor the full complement of genes that are gained and lost in these lineages by comparing their present-day gene repertoires. However, the patterns of retention of genes assigned to evolutionary lineages of different ages offer a glimpse into the fate of acquired sequences. Among genes that originated in the ancestor to all 13 strains examined, 78% (61% of the ORFans and 95% of the HOPs) were lost in one or more of the descendant lineages, whereas 96% (95% of the ORFans and 97% of the HOPs) of the genes acquired on the next older branch, SCE, were lost. Overall, genes acquired on the ancestral branch have higher retention rates than those genes acquired on more recent branches.
Combining the numbers of ORFans and HOPs, Shigella spp. (including S. dysenteriae) show significantly lower retention rates compared to E. coli strains (65% vs. 88%; p<0.01), which is not surprising since Shigella species have the highest rates of loss of recently acquired genes. The lower retention rates in Shigella spp. result from both significantly more gene inactivations (11% vs. 5% in E. coli, p<0.01) and gene losses (24% vs. 8% lost in E. coli, p<0.01), and though disruptions occur to a similar extent in both sets of acquired genes, HOPs are more frequently lost than retained as inactivated genes (Table 2, Supplementary Figure S3). Also, ORFans are inactivated predominantly by truncations, whereas HOPs are more often disrupted by insertion sequences (Supplementary Table S3).
Although pseudogenes have been shown to be largely genome-specific [7],[9],[16], it was expected that some would be retained in multiple lineages over the short evolutionary time-span examined in this study. However, more than half of the inactivated ORFans and HOPs exist only in a single genome, whereas their functional homologs are usually present in several genomes (data not shown). Similarly, over half of the losses of acquired genes are also genome specific (i.e., losses of the only member of a gene family), confirming the high turnover rate observed for inactivated DNA.
Gene gain and loss are ongoing processes in microbial genomes, resulting in a diversity in genome sizes, even among closely related strains within a bacterial species [3],[29]. By comparing the genome contents of sequenced representatives of the E. coli/Shigella clade, and by mapping the phylogenetic distribution of every gene present in these genomes, we find that the rates of change in novel genes can differ over 200-fold between strains and lineages. Moreover, genes of different phylogenetic origins arise and persist at very different rates. For example, ORFan genes, i.e., those with no homologs outside of the group of bacteria examined, emerge less frequently than do genes originating by acquisition from other bacteria (termed “HOPs”), but are, on average, about eight times more likely to be maintained. Of the genes acquired on the ancestral branch, nearly 39% of the ORFans, but only 5% of HOPs, are present in all 13 genomes indicating that they now provide functions integral to all strains.
The difference in the persistence of ORFans and HOPs is surprising because those genes acquired from other bacteria (i.e., HOPs) typically encode functional proteins in the donor and could be immediately useful to the recipient, whereas the ORFans, whose origins are less certain, have probably never served a function in a cellular genome prior to their acquisition. The disparity in the types of properties conferred by these two classes of genes is supported by their assignment to known functional categories: whereas nearly a quarter of the HOPs could be designated a COG category, less than 1% of ORFans could. Although ORFans are often poorly annotated and resist functional characterization by comparative approaches (partially due to their characteristically short length and atypical composition), several lines of evidence indicate that they encode functional proteins [20],[30], including structural in vitro analyses on E. coli ORFans (unpublished data). Therefore, the retention of ORFans may reside in the fact that they confer truly novel (but as yet unknown) functions, as opposed to traits that are apt to be redundant to the recipient organism. Alternatively, as ORFans are generally thought to be derived from selfish mobile elements (but see [28]), some might be perpetuated by encoding selfish functions themselves.
The distributions of ORFans and HOPs show that sequences that do not provide a useful function are eliminated and that bacterial genomes are not repositories of non-functional genes. This parallels the situation observed for pseudogenes, which, due to their rapid removal, are largely strain- or genome-specific [7]. Because the most-recently acquired genes are the least likely to supply an immediately useful function, we might expect that the newest genes in a genome are the most rapidly removed [18]. Indeed, comparing the two oldest branches indicates that while 33% of the genes gained on the ancestral branch are lost in each extant genome, 42% of the genes gained on a younger branch (SCE) are lost. From the present dataset, it is difficult to assess how this trend continues because relatively few genes are introduced on each branch (only 9 and 13 genes on SC and S4 respectively), and in younger clades, there are successively fewer genomes from which the gene can be eliminated. However, the low numbers of genes mapped to these internal branches probably reflects the fact that relatively few acquired genes are being maintained.
The density of sequenced genomes has allowed the use of phylogenetic methods to assess the dynamics of gene contents within several bacterial species, and has shown that rates of DNA gain and loss are often strain or lineage specific. Based on the same genomes analyzed in the present study, Hershberg and co-workers [26] found that the rates of gene loss in Shigella species were consistently higher than in related strains of E. coli, presumably due to reduced selection brought about by their small effective population sizes. Our data agree with these findings, and additionally, show that Shigella species also have lower rates of gene acquisition and lower rates of retaining acquired genes. Taken together, the inactivation and subsequent deletion of resident genes coupled with decreased levels of gene acquisition and subsequent persistence accounts for the reduced size of Shigella genomes.
Applying a similar approach, Vernikos and co-workers [31] analyzed the genes acquired by the strains of Salmonella enterica for which genome sequences are available. In Salmonella, most of the acquired genes have low GC-contents and are still “ameliorating”, i.e., adjusting their base composition towards that of the host genome [31],[32], similar to results observed for acquired sequences in the Gammaproteobacteria as a whole [20]. That amelioration has been observed in studies on Salmonella and the Gammaproteobacteria, but not in E. coli, is due to the fact that the sequenced strains of E. coli span a much shorter timescale and have not yet accumulated sufficient numbers of mutations to noticeably alter the average base composition of genes.
In addition to assessing genome dynamics by following the presence and absence of acquired genes, we also traced the formation of pseudogenes to more closely monitor the mechanisms by which genes are inactivated and eliminated from these genomes. Pseudogenes in our analyses have restricted distributions, and nearly half of the inactivated ORFans and HOPs occur in only a single genome. In that the formation of pseudogenes is an ongoing process, their very restricted distributions denote that inactivated genes are eliminated rapidly from the genome and imply that newly acquired genes that are not immediately functional are also subject to rapid removal. Although such assessments of gene contents are based only on those genes now present in contemporary genomes, the recognition of pseudogenes can provide additional insights into the evolution and dynamics of genomes. The inclusion of pseudogenes in the present analysis provides some indication that high numbers of genes are gained and lost without leaving traces of their introgression [7],[33].
In conclusion, comparative genomics of multiple closely related strains provides high-resolution assessments and quantifications of gene fluxes in an evolutionary context [31],[34], and allows specific estimations of the processes of gene inactivation and deletion. Within the sequenced strains of E. coli and Shigella spp., we detected large differences among closely related lineages in the rates of gene acquisition and loss, but also differences in gene retention rates due to the source of acquired genes. The higher retention rate observed in the functionally obscure ORFan genes suggests that there are unknown adaptive benefits to these small acquired genes.
To trace the history of each gene in the sequenced E. coli and Shigella genomes, it is first necessary to resolve the phylogenetic relationships among these 13 strains. We based this phylogeny on the core set of single-copy genes identified by Lerat et al. [35] as showing virtually no evidence of lateral gene transfer within the Gammaproteobacteria. The seven sequenced E. coli genomes (E. coli K-12 [36], E. coli W3110, E. coli Sakai [37], E. coli EDL933 [38], E. coli CFT073 [21], E. coli UTI89 [39] and E. coli 536) and six sequenced Shigella genomes (S. flexneri 301 [40], S. flexneri 2457, S. flexneri 8401 [41], S. dysenteriae [22], S. boydii [22] and S. sonnei [22]) were searched via BLASTP [42] for orthologs of these core genes, applying an E-value<1−10 and a match length >75%. Of the 203 genes identified by Lerat et al. [35], 169 single copy genes met these criteria of orthology and were used for phylogenetic reconstruction. Concatenated sequences of these 169 genes from all 13 E. coli and Shigella genomes were aligned using MAFFT [43] and the alignment was edited to remove gaps using Gblocks [44]. A maximum likelihood tree (using DNAML module of PHYLIP; http://evolution.genetics.washington.edu/phylip.html) was generated using the concatenated orthologous sequences of Salmonella enterica as the outgroup.
The genome sequences of the 367 prokaryotes (339 bacteria and 28 archaea) available at the time of this study were retrieved from GenBank (ftp.ncbi.nih.gov/genbank/genomes/Bacteria/; August 2006), and an in-house database was created by extracting protein sequences from all but the 13 E. coli and Shigella genomes.
Newly acquired genes can be of two types: ORFans, genes with no detectable homolog in the databases, and HOPs (heterogeneous occurrence in prokaryotes [20]), genes with homologs in distantly related species. ORFans in each of the 13 E. coli and Shigella genomes were identified as described previously in Daubin and Ochman [20]. In brief, all protein sequences from these genomes were compared with the database using BLASTP, applying an E-value cutoff of 0.01 to uncover distant homologs. Those genes without a match at this relaxed cutoff were considered to be potential ORFans. To eliminate possible artifacts due to annotation errors, we queried gammaproteobacterial genomes with all putative ORFans using TBLASTN and excluded those with matches having E-value cutoff<10−5 and alignment lengths >50%. The distribution of all remaining ORFans among the 13 strains of E. coli and Shigella were obtained by comparing the ORFans from each E. coli and Shigella genome with the remaining 12 genomes using TBLASTN with an E-value cutoff of 10−5. Based on their distribution among strains, ORFans were assigned to clades of the E. coli phylogeny. The orthology of ORFans present in more than one strain was confirmed by genome context.
In contrast to ORFans, HOPs have homologs in other prokaryotic genomes. To qualify as a HOP, a protein must be restricted to an E. coli clade, absent from closely related genomes, and have a homolog in a more distantly related prokaryotic genome. We performed BLAST analyses to identify genes that displayed such sporadic distributions. For example, the 9 HOPs restricted to clade S1 (Figure 1) were present in S. flexneri 301 and S. flexneri 2457, lacked homologs in the other E. coli and Shigella genomes, but had homologs in some distantly related genomes. We mapped the branch on which a gene was acquired by reconstructing the parsimonious scenario that explains the present-day gene distribution [18], such that the path that invokes the lowest number of events was viewed as the most evolutionarily plausible. In these reconstructions, gene gains and losses were viewed as individual and equally likely events. The genes acquired on each branch are listed in Supplementary Table S4. Classification of the identified ORFans and HOPs to Clusters of Orthologous Groups (COGs) [45] was performed using in-house scripts.
Pseudogenes were identified by using Ψ-Φ as described previously [7],[9],[16]. In this procedure, the annotated proteins from each genome were queried against the complete nucleotide sequence of every other strain with E-value cutoffs of 10−15 and sequence identities >75%. The Ψ-Φ program suite uses the TBLASTN output to return lists of predicted disrupted genes, which are manually curated. To identify gene-inactivating mutations, the predicted pseudogenes were aligned against their orthologs using CLUSTALW [46]. Gene-inactivating mutations were grouped into five classes: frameshifts (insertions or deletions of 1 or 2 nucleotides in length), deletions (>2 nucleotides in length), insertions (>2 nucleotides in length), truncations (large deletions at either or both ends of a coding sequence), nonsense mutations, or a combination of different classes. |
10.1371/journal.pbio.3000163 | Brain expansion promoted by polycomb-mediated anterior enhancement of a neural stem cell proliferation program | During central nervous system (CNS) development, genetic programs establish neural stem cells and drive both stem and daughter cell proliferation. However, the prominent anterior expansion of the CNS implies anterior–posterior (A–P) modulation of these programs. In Drosophila, a set of neural stem cell factors acts along the entire A–P axis to establish neural stem cells. Brain expansion results from enhanced stem and daughter cell proliferation, promoted by a Polycomb Group (PcG)->Homeobox (Hox) homeotic network. But how does PcG->Hox modulate neural-stem-cell–factor activity along the A–P axis? We find that the PcG->Hox network creates an A–P expression gradient of neural stem cell factors, thereby driving a gradient of proliferation. PcG mutants can be rescued by misexpression of the neural stem cell factors or by mutation of one single Hox gene. Hence, brain expansion results from anterior enhancement of core neural-stem-cell–factor expression, mediated by PcG repression of brain Hox expression.
| The central nervous system displays a pronounced anterior expansion that forms the brain. In the fruit fly Drosophila melanogaster, this expansion is driven by enhanced anterior cell proliferation. Recent studies reveal that cell proliferation in the brain is promoted by the Polycomb Group Complex, a key epigenetic complex. During development of the central nervous system, the Polycomb Group Complex acts to exclude Hox homeotic gene expression from the brain, thereby rendering the brain a Hox-free zone. Hox genes act in an antiproliferative manner, which explains the hyperproliferation observed in the brain, as well as the gradient of proliferation along the anterior–posterior axis of the central nervous system. Here, we find that Hox genes act by repressing a common neural stem cell proliferation program in more posterior regions, resulting in an anterior–posterior gradient of “stemness.” Hence, elevated anterior proliferation is promoted by the Polycomb Group Complex acting to keep the brain free of negative Hox input, thereby ensuring elevated expression of neural stem cell factors in the brain. Strikingly, mutants of the Polycomb Group Complex can be rescued by mutation of one single Hox gene, demonstrating that the primary role of the Polycomb Group Complex is indeed Hox repression. This study advances our understanding of how neural stem cell programs operate at different axial levels of the central nervous system and may have implications also for stem cell and organoid biology.
| During central nervous system (CNS) development, neural progenitor cells undergo repetitive rounds of asymmetric cell divisions, renewing themselves and generating daughter cells. After a certain number of cell divisions, neural progenitors subsequently exit the cell cycle. Daughter cells, in turn, may directly differentiate into neurons or glia or divide one or many times to expand any given lineage. Thus, lineage size depends upon two fundamental proliferation decisions: how many times should each progenitor divide, and how many times should its daughter cells divide? Studies have revealed profound differences in both progenitor and daughter cell proliferation behavior when comparing along the anterior–posterior (A–P) axis and over developmental time [1–7]. Such differences can result in striking alterations in lineage size when comparing between different progenitors, from a few to several hundred cells generated from one progenitor [8–13]. Modulation of lineage size can influence CNS regional size and contribute to the prominent anterior expansion of the CNS [3, 7, 14]. However, how neural progenitor and daughter cell proliferation is modified along the A–P axis and over developmental time to thereby result in changes in lineage and ultimately regional CNS size is still poorly understood.
The developing Drosophila CNS is a powerful model system for addressing these issues [15]. The Drosophila CNS, subdivided into the brain and the nerve cord, is formed by approximately 1,200 bilateral neuroblasts (NBs) and a smaller number of midline NBs, all of which form in the neuroectoderm during early-to-mid embryogenesis (Fig 1A) [10, 12, 13, 16–18]. After delaminating from the neuroectoderm, NBs undergo a series of asymmetric cell divisions, renewing themselves while budding off daughter cells with reduced proliferative potential. The majority of NBs initially generate daughters that divide once to make two neurons/glia, denoted Type I proliferation mode [9]. Subsequently, after a stereotyped number of NB divisions, many NBs in the nerve cord switch to budding off daughter cells that differentiate directly, denoted Type 0 mode, and hence they undergo a programmed Type I->0 daughter cell proliferation switch [19]. Finally, the majority of NBs in the brain and nerve cord appear to exit the cell cycle after a programmed number of divisions, unique to each NB subtype (Fig 1B). The Type I->0 daughter cell proliferation switch and NB exit occurs in a graded fashion along the CNS, and the majority of brain NBs appear to stay in the Type I mode throughout embryogenesis. This results in striking differences in the average lineage size along the CNS A–P axis (Fig 1B) [3, 7].
In the nerve cord, NB cell-cycle exit and the Type I->0 daughter cell proliferation switch are under the control of an elaborate program of proliferation “drivers” and “stoppers.” These include all members of the so-called temporal gene cascade: Hb->Kr->Nubbin/POU domain 2 (Nub/Pdm2; collectively referred to as Pdm)->Castor (Cas)->Grainy head (Grh), playing out in the majority of embryonic NBs (reviewed in [20]). The temporal cascade controls NB competence and governs the generation of specific glia and neurons at distinct stages of lineage progression [20]. However, the temporal genes also control proliferation, with the early temporal factors Hb, Kr, and Pdm acting in a pro-proliferative manner and the late temporal factors Cas and Grh acting in an antiproliferative manner [19, 21]. In addition to the temporal factors, the pan-NB factors of the sex determining region Y-box B (SoxB) family (SoxNeuro [SoxN] and Dichaete [D]), the Snail family (Snail, Worniou [Wor], and Escargot), and the basic helix-loop-helix (bHLH) factor Asense (Ase) also control NB generation and development [21–29] and play key roles in driving proliferation [21, 23, 30]. While most, if not all, NBs express one or several of each pan-NB factor family member [23–28, 30–33], detailed expression analysis revealed that their expression is down-regulated during embryonic neurogenesis [21]. Herein, we refer collectively to the three early temporal factors and the pan-NB factors as Early Factors (EFs). The EFs are necessary and partly sufficient for several aspects of the NB program, including asymmetric cell division [23, 30], the Type I daughter cell proliferation mode, and continuing NB proliferation [21]. By a complex regulatory temporal interplay, EFs are gradually and precisely replaced by the late factors Cas and Grh, which triggers the Type I->0 daughter cell proliferation switch and, ultimately, NB exit [19, 21]. In addition to this temporal program, there is critical control of NB and daughter cell proliferation mediated by A–P cues [4, 5]. These cues are provided by the Hox homeotic genes, which are activated late in nerve cord NBs and act with Cas and Grh to trigger the Type I->0 switch and NB exit [3, 19, 34]. Anteriorly, the Polycomb Group (PcG) complex, in particular Polycomb Repressor Complex 2 (PRC2), acts to keep the brain free of Hox homeotic gene expression, thereby preventing the Type I->0 daughter cell proliferation switch and promoting an extended phase of NB proliferation [7]. However, these studies raise the question of how the PcG->Hox program modulates EF activity along the A–P axis to promote the gradient in NB and daughter cell proliferation.
We find that EF expression is elevated and extended in brain NBs when compared to the nerve cord. EFs are necessary for brain proliferation, and EF co-misexpression overrides Type I->0 switches and NB exit, both in the nerve cord and brain, resulting in a striking increase in overall CNS size. Elevated EF levels in the brain and graded EF expression along the nerve cord are both gated by the PcG->Hox network. Strikingly, the effects of PRC2 mutation—including the nondetectable levels of Histone 3 K27 trimethylation (H3K27me3), the “invasion” of Hox gene expression into the brain, the reduction of EF expression, and the accompanied reduction of proliferation—can be rescued by mutation in the Hox gene Abd-B or by transgenic expression of EFs. These results demonstrate that the PcG->Hox network modulates a temporal neural stem cell program along the A–P axis, thereby allowing for the wedge-like development of the CNS, with its prominent anterior expansion.
We previously focused on thoracic segments T2–T3, finding that the three pan-neural factors Wor, SoxN, and Ase, as well as the three early temporal factors Hb, Kr, and Pdm, are expressed in early NBs but down-regulated during lineage progression [21]. This down-regulation is necessary for the Type I->0 daughter cell proliferation switch and the final NB cell-cycle exit. To address whether the expression levels of these six factors, collectively referred to herein as EFs, correlate with the gradient of the Type I->0 switch and NB exit observed in the developing Drosophila CNS, we analyzed EF expression levels in NBs in the brain (B1–B2) and thorax (T2–T3), as well as two abdominal regions (A5 and A8–A10) (Fig 1A and 1C). We analyzed three stages: Stage (St)11, St14, and St16, chosen because they represent the gradual accentuation of A–P proliferation differences [3, 7, 19, 21]. Specifically, at St11, there are only minor differences in NB and daughter cell proliferation along the A–P axis, with most NBs and daughter cells proliferating (Type I). At St14, many NBs in the A8–A10 region have stopped dividing, daughter cells are mostly nondividing (Type 0), and many T2–T3 NBs have undergone the Type I->0 switch. At St16, proliferation of NBs and daughter cells in both abdomen and thorax have largely ceased, while many NBs and daughter cells in the B1–B2 region continue dividing.
At St11, as anticipated from the minor differences in proliferation along the A–P axis at this stage, we observed that only four out of the six EFs showed a higher expression level in B1–B2 when compared to A8–A10 (S1A Fig). Wor, SoxN, and Kr furthermore displayed higher levels in B1–B2 also when compared to T2–T3 and/or A5. In contrast, Ase, Hb, and Pdm2 showed a more complex picture, with Ase and Pdm2 being highest in T2–T3, and Hb highest in A8–A10 (S1A Fig). At St14, in line with the accentuated A–P proliferation differences, levels of all six EFs were significantly higher in B1–B2 than A5 and A8–A10, with five of six also significantly higher in B1–B2 than T2–T3 (S1B Fig). At St16, similarly, all six EFs were significantly higher in B1–B2 when compared to one or several nerve cord regions (S1C Fig). A peculiar exception observed pertains to Pdm2 levels, which showed a gradient from B1–B2 to T2–T3 and on to A5 but was then elevated in A8–A10 (S1C Fig). To address EF expression over time, we analyzed St11, St14, and St16 on the same slide and focused on the levels in B1–B2. This revealed that all six EFs were significantly down-regulated between St11 versus St14 and/or St16 and St14 versus St16 (S1D Fig). Kr differed between St11 and St16 and St14 and St16 but not between St11 and St14 (S1D Fig). Setting the expression levels at St11 in B1–B2 as 100% and normalizing against this allowed for the generation of 3D graphs that illustrate the EF expression landscape over time and along the A–P axis (Fig 1D). The most salient feature was that during early stages, EFs were robustly expressed in NBs at all axial levels. As development progressed and A–P differences in NB and daughter proliferation play out, EFs were generally down-regulated in a graded manner in the nerve cord but maintained for longer in the brain.
Previous studies focused on the role of EFs in the thoracic segments T2–T3, revealing that they promote NB and daughter cell proliferation [21]. To determine if they play similar roles throughout the CNS, we analyzed brain segments B1–B2, as well as the nerve cord abdominal segments A1, A5, and A8–A10.
To analyze NB and daughter cell proliferation in the developing CNS, we used Prospero (Pros), Deadpan (Dpn), and phosphorylated Ser10 on Histone-H3 (PH3) as markers. This approach relies on the fact that dividing NBs are PH3+, Pros-cortical asymmetric, and Dpn+, while dividing daughter cells are PH3+, Pros cytoplasmic, and Dpn negative (S2A Fig) [19]. We analyzed proliferation at St14 in the EF mutants wor, SoxN, ase, hb, Kr, and nub,pdm2 (the function of the adjacent nub and pdm2 genes was addressed simultaneously by using a genomic deletion for both genes, referred to as nub,pdm2). We observed significantly reduced NB proliferation in B1–B2 in all mutants except SoxN (S2B–S2D Fig). In abdominal segments, all mutants showed reduced proliferation, with the exception of SoxN in A1 and nub,pdm2 in all segments (S2D Fig).
With respect to daughter cell proliferation, there was significantly reduced proliferation in B1–B2 in all mutants except nub,pdm2 (S2E Fig). The effects were generally weaker in the abdominal segments, with only wor and ase showing significant effects in A1, none of the genes showing effects in A5, and with wor, ase, and Kr showing effects in A8–A10 (S2E Fig). The role of hb could not be addressed in A8–A10 because of a loss of these segments in the compound mutant used herein.
We conclude that all of the six EFs are important for NB and/or daughter cell proliferation in the brain. With the exception of nub,pdm2, they are also important for NB proliferation in the abdominal segments, while their involvement in abdominal daughter cell proliferation is less pronounced.
We previously demonstrated that the brain is “hyperproliferative” when compared to the nerve cord, displaying an apparent lack of the Type I->0 daughter cell proliferation switch and an extended NB proliferation phase [7]. Herein, we observed elevated and extended EF expression in the brain and the importance of EFs for NB and daughter cell proliferation in the brain. In combination, these findings suggested that elevated and extended EF expression could be a contributing factor to the enhanced proliferation normally observed in the brain. To address this notion, we co-misexpressed all six factors (upstream activating sequence [UAS]-6xEF) during CNS development, using the inscuteable-Galactose4 (insc-Gal4) driver (S3A–S3Z Fig).
In control animals, at late embryogenesis stage (air-filled trachea [AFT]), PH3 staining was typically observed in only a few lineages in the brain and occasional cells in the nerve cord (Fig 2A and 2G). In contrast, in insc-Gal4/UAS-6xEF animals, we observed a striking increase in proliferation, with extensive numbers of PH3 cells in both the brain and nerve cord (Fig 2C and 2G). Previous studies demonstrated that the EFs regulate a number of key cell-cycle factors [21, 23, 30]. Therefore, we wished to compare the UAS-6xEF effects to that of direct misexpression of key cell-cycle genes. To this end, we co-misexpressed the four cell-cycle genes Cyclin E (CycE), Cyclin dependent kinase 2 (Cdk2), E2F transcription factor 1 (E2f1), and Dimerization partner (Dp) (denoted UAS-4xCC herein) that previous studies have shown to be critical for Drosophila embryonic CNS proliferation [19, 35]. This co-misexpression (insc-Gal4/UAS-4xCC) also resulted in significantly increased proliferation, but strikingly, to a lesser degree than UAS-6xEF (Fig 2B and 2G). We furthermore noticed that the ectopic proliferation evident in the stage AFT nerve cord involved both dividing NBs and daughter cells (Fig 2A′–2C′).
To address the effects of misexpression on CNS size, we used 4′,6-diamidino-2-phenylindole (DAPI) nuclear staining to quantify total nuclear (cellular) volume. This revealed striking increase in CNS volume for UAS-6xEF and a trend upwards for UAS-4xCC, albeit not a significant one (Fig 2D–2F and 2H). Quantifying the nerve cord and brain separately, as anticipated for UAS-6xEF co-misexpression, we observed cellular volume expansion in the nerve cord (Fig 2I). Somewhat surprisingly, the expansion was even more pronounced for the brain (Fig 2J). UAS-4xCC showed an upwards trend in volume both in the nerve cord and the brain, but this was not significant (Fig 2I and 2J). Hence, both with respect to proliferation and CNS volume, 6xEF co-misexpression is more potent than 4xCC co-misexpression.
Focusing further on the UAS-6xEF co-misexpression, the presence of dividing NBs and daughter cells indicated that both the NB exit and the Type I->0 daughter cell proliferation switch was over-ridden by co-misexpression (Fig 2C′). To further address this notion, we turned to single-NB–lineage analysis, using the NB5-6 specific lbe(K)-Gal4 driver [36]. NB5-6 NBs are generated by late St8 [12, 13] and commence proliferating in the Type I mode [3, 19, 36]; they then switch to Type 0 proliferation for several rounds until the NB exits the cell cycle and undergoes apoptosis (Fig 2N). In abdominal segments, Hox genes of the Bithorax Complex (BX-C) trigger an earlier Type I->0 switch and NB exit, resulting in smaller abdominal NB5-6 lineages (Fig 2K) [3, 12, 13, 34]. By St15, both thoracic and abdominal NB5-6 NBs have exited the cell cycle and undergone apoptosis. Hence, from St16 and onwards into AFT and L1 larval stages, there are no cell divisions observed in this lineage (Fig 2K, S4A Fig). In contrast, in lbe(K)-Gal4/UAS-6xEF, we observed ectopic NB and daughter cell divisions both at AFT and L1 (Fig 2L, S4B Fig). This resulted in a greatly expanded lineage size (Fig 2M, S4C Fig). These experiments demonstrate that EF co-misexpression can override both the Type I->0 switch and NB exit and drive aberrant continuing lineage progression (Fig 2N).
Previous studies addressing the interplay between EF and Hox genes focused on the thorax and the Antennapedia (Antp) Hox gene [21]. This revealed that EFs and Antp regulate each other. Here, we addressed the interplay between EFs and the BX-C Hox genes Ultrabithorax (Ubx), abdominal-A (abd-A), and Abdominal-B (Abd-B) by analyzing EF and BX-C mutants for EF and BX-C expression. Expression in NBs was quantified in the pertinent segment for each mutant (Fig 3A–3L). Of the 36 analyses performed, no fewer than 22 showed significant effects on protein expression (Fig 3L). The most clear-cut interplay was between Abd-B and the EFs, in which Abd-B acted as a repressor of all six EFs. This interaction fits well with A8–A10 displaying the earliest Type I->0 daughter cell proliferation switch and earliest NB exit [7], with the strong antiproliferative role of Abd-B [3], and with A8–A10 showing the lowest levels of EFs along the A–P axis of the CNS (Fig 1D). In contrast, three of the EFs were activators of Abd-B. We also observed that wor and nub,pdm2 repressed Ubx, and that wor, ase, and nub,pdm2 repressed Abd-A. In contrast, hb and SoxN activated Ubx, and surprisingly, abd-A activated Wor, Ase, and Pdm2.
In summary, combined with the previous study of EF-Antp interactions [21], we find that out of the 48 possible interactions between EFs and Antp/BX-C, we find 29 interactions (Fig 3M). The majority of these (16) are repressive and fit with the general theme of EFs repressing Hox genes and vice versa. Surprisingly, 13 interactions involved EF activation of Hox factors and vice versa, perhaps pointing to negative feedback.
Previous studies, using Chromatin immunopurification-sequencing (ChIP-seq) and DNA adenine methyltransferase identification-sequencing (Dam-ID-seq), reveal that several of the EF and Hox factors bind to each other’s genes [22, 37–39] (www.modencode.org) (S1 Table). These findings suggest that the cross-regulation observed herein may result from direct transcriptional regulation.
The Drosophila embryonic brain (B1–B2) does not express any of the Hox homeotic genes [7, 40]. A key regulatory system ensuring the repression of Hox genes in the brain is the PcG complex, and in particular the PRC2 [7, 41–44]. PRC2 is the key epigenetic complex responsible for adding the repressive H3K27me3 mark upon histone H3 [45, 46], and PRC2 is critical for the apparent lack of any Type I->0 daughter cell proliferation switches and the prolonged phase of NB proliferation observed in the brain [7]. To address the possible effect of PRC2 mutation on the elevated expression of EFs observed in the brain, we analyzed maternal and zygotic extra sex combs mutants (esc; mammalian Embryonic Ectoderm Development [EED]). Esc/EED is a key PRC2 component [45, 46], and esc maternal/zygotic mutants display nondetectable levels of H3K27me3 in the developing embryo accompanied by ectopic expression of Antp and BX-C in the entire brain (S5A–S5J Fig) [7]. As previously observed [7], in wild-type embryos, the brain showed stronger staining for the H3K27me3 mark (S5A Fig). In esc maternal/zygotic mutants, verified by nondetectable levels of H3K27me3 staining and ectopic Abd-B expression in the B1-B2 region (S5F and S5J Fig), we observed that all six EFs were significantly down-regulated in NBs (Fig 4A–4L).
Having found that EFs are down-regulated in esc mutants, we wanted to test whether EFs transgenic expression could cross-rescue esc mutants. To this end, we co-misexpressed all six EFs, driven from the insc-Gal4 driver, in an esc maternal/zygotic mutant background. Strikingly, we found that transgenic 6xEF coexpression could completely override the reduced NB and daughter cell proliferation observed in esc mutants (Fig 5A–5C, 5G and 5H). Hence, the strong antiproliferative effect of PRC2 upon brain development could be over-ridden by 6xEF transgenic coexpression.
The Type I->0 daughter cell proliferation switch and NB exit both occur first in the A8–A10 region, and Abd-B plays a prominent role in triggering these events [3]. The repressive role of Abd-B on all six EFs (Fig 3M) and the extension of Abd-B expression into B1-B2 in esc mutants (S5E and S5J Fig) [7], prompted us to address the extent to which reduced proliferation in esc mutants result from ectopic brain expression of Abd-B.
First, we addressed the potency of Abd-B in triggering the Type I->0 switch and the NB exit in the wild-type background by misexpressing it in the developing brain using insc-Gal4. We observed striking reduction of both NB and daughter cell proliferation in the B1–B3 segments (Fig 5E, 5I and 5J). This demonstrates that Abd-B is sufficient, at least in part, to impose a posterior nerve cord proliferation behavior on the brain. Next, we addressed the role of Abd-B in an esc mutant background by generating maternal and zygotic homozygous esc mutants simultaneously zygotically homozygous for Abd-B (esc; Abd-B). Strikingly, we found that the reduction of NB and daughter cell proliferation observed in esc mutants was rescued by Abd-B homozygosity (Fig 5F, 5I and 5J). In fact, surprisingly, Abd-B homozygosity restored NB and daughter cell proliferation back to wild-type levels (Fig 5I and 5J).
The nerve cord, specifically T1–A10, shows a gradient of NB proliferation and Type I->0 daughter cell proliferation switches [3, 4]. This is controlled by the graded expression of Hox genes and the increasingly antiproliferative role of posteriorly expressed Hox genes [3, 7]. In contrast, the brain is “hyperproliferative,” with prolonged NB proliferation and an apparent absence of the Type I->0 switch [7]. Here, we find significant differences when comparing EF levels along the A–P axis, with elevated and extended expression in the brain and a gradient along the nerve cord. Together with the mutant and misexpression data (herein; [21]), this supports a model in which graded EF levels are key for generating graded proliferation along the CNS A–P axis and, at elevated levels, drives brain-type proliferation (Fig 6).
The Type I->0 daughter cell proliferation switch and the precise NB cell-cycle exit depend upon balanced levels of four key cell-cycle genes: Cyclin E (CycE), string (stg; mammalian Cdc25a/b/c), E2f1, and dacapo (dap; Cdkn1a/b/c) [19, 35]. Previous studies of EFs, Hox genes, and PRC2 demonstrated regulatory links to these cell-cycle genes [3, 7, 19, 21, 23, 30]. This regulation results in both temporal and spatial changes in cell-cycle gene expression. Hence, cell-cycle “drivers” (CycE, E2f1, Stg) are elevated in early NBs when compared to late and elevated in brain when compared to nerve cord, while “stoppers” (Dap) show the inverse profile [7, 19]. Surprisingly, the co-misexpression of CycE, Cdk2, Dp, and E2f1 (UAS-4xCC) gave considerably weaker effects than that of 6xEF co-misexpression. One explanation for this weaker effect may be that EFs are involved in also regulating stg (“driver”) and dap (“stopper”) [21, 23, 30], which were not simultaneously misexpressed/mutated in these experiments.
In addition to Type I and Type 0, a third type of embryonic NB behavior was recently identified: Type II NBs [1, 6]. These NBs generate daughter cells that divide multiple times. Of the 1,200 NBs in the embryo, and 228 in the brain (B1–B3), only 16—eight in each B1 segment—have been identified. Strikingly, recent analysis reveal that Type II NBs do not express several of the EFs studied herein, e.g., Ase, Hb, and Kr, while they do express Pdm, Cas, Grh, and Pointed-P1 [1, 6]. This suggests that while Type I NB identity (including switching to Type 0 in the nerve cord) is specified by a common EF program, the Type II NBs are specified by a partly nonoverlapping genetic program.
The graded proliferation observed along the A–P axis of the CNS is controlled by graded Hox input in the nerve cord and an absence of Hox expression in the brain (B1–B2) [3, 4, 7, 19, 34]. The graded proliferation is mediated by graded EF expression, also under the influence of graded Hox gene expression. All three components—i.e., graded proliferation, Hox expression, and EF expression—are controlled by PRC2 (this work; [3, 7, 21]). To begin addressing this interplay, we conducted two types of cross-rescue experiments. First, we found that we could rescue esc by transgenic re-elevation of 6xEFs, demonstrating that in spite of nondetectable levels of the H3K27me3 mark and aberrant Hox expression in the brain, EFs can still restore brain-type proliferation. Second, we found that we could rescue the reduced brain proliferation in esc mutants by merely removing one Hox gene: Abd-B. While this second finding may at first glance be surprising, given the many possible roles of PRC2 and the ectopic expression of Hox genes in esc mutant brains, the rescue fits with several observations. First, of the four nerve cord Hox genes, Abd-B has the most prominent effect on proliferation in the nerve cord, explaining why A8–A10 display the earliest Type I->0 switch and NB exit, and hence the smallest lineages [3]. Second, Abd-B has strongly repressing effects on all six EFs, and, logically, EF expression is at its lowest in A8–A10. Third, in esc mutants, Abd-B is robustly expressed throughout the brain. Fourth, Abd-B misexpression in the brain results in reduced proliferation. Fifth, in the mouse, the posteriormost Hox gene in the B cluster, Hoxb13 (an Abd-B orthologue), has strong effects on spinal cord proliferation [47]. Moreover, misexpression of Hoxb13 in the chick telencephalon also suppressed proliferation and showed stronger effects than the more anteriorly expressed Hoxb9 gene [7]. In combination, the potency of Abd-B and perhaps the Abd-B paralog group in general (the most posteriorly expressed Hox group) combine to explain why esc could be rescued by removal of only one Hox gene. In combination, the regulatory interplay between Hox genes and EFs, the EF cross-rescue of esc, the Abd-B rescue of esc, and the PRC2 regulation of EF combine to support the idea that the main role of PRC2, with regards to promoting brain proliferation, is to repress Hox gene expression in the brain, thereby allowing for elevated EF expression.
In humans, three of the four main components of PRC2, including EED (Drosophila esc), EZH2, and SUZ12, are predicted to be haploinsufficient [48]. In line with this notion, several related human developmental syndromes, including the Weaver, Weaver-like, Cohen-Gibson, and Overgrowth and Intellectual Disability syndromes, have been linked to heterozygous, usually de novo, mutations in EED, EZH2, or SUZ12 [49–60]. In some cases, these mutations were shown to result in reduced H3K27me3 activity [51, 55]. These syndromes manifest with a range of peripheral overgrowth phenotypes but also neurological defects, involving delayed speech and psychomotor development as well as intellectual disability. MRI analysis has revealed a range of brain morphology anomalies, including polymicrogyria and white matter volume loss [51, 52]. It is tempting to speculate that our surprising finding that mutation in one single Hox gene could rescue esc mutants may point to new avenues of investigation for these syndromes.
A defining feature of bilaterians is the condensation of neural cells into a centralized and contiguous tissue: the CNS. In most, if not all, bilaterians a common property of the CNS is that the anterior part, the brain, is larger than the nerve cord. The brain and the nerve cord express distinct sets of highly conserved patterning genes, with the brain expressing a number of “brain-specific” genes and the nerve cord expressing Hox homeotic genes [44, 61–63]. The distinct patterning gene expression, as well as other aspects of nervous system development and evolution, has recently prompted the idea of a separate evolutionary origin of these CNS regions [61, 63–65]. Intriguingly, in arthropods such as Drosophila, the initial development of the brain and nerve cord is separate, only to merge during subsequent embryogenesis [66]. The finding of distinct proliferation patterns in the Drosophila brain versus nerve cord, with the brain using Type I and Type II daughter proliferation and extended NB proliferation [1, 6, 7] and the nerve cord displaying the Type I->0 daughter proliferation switch and earlier NB exit [3, 19], lends further support for the notion of separate evolutionary origins for the brain and nerve cord.
In parallel to A–P patterning of the developing CNS, neural stem cells are generated along the entire neuraxis. This study, and others, point to the fact that in Drosophila, there is a common program for establishing and driving NBs along the entire neuraxis. Studies in mammals have revealed that a number of the genes involved in the Drosophila neural stem cell program are evolutionarily conserved. Specifically, SoxN and D (SoxB family members) are orthologues of mammalian Sox1, -2, and -3, which are critical for CNS development [67]. The wor/sna/esg orthologue Snail1 is important for proliferation of mouse neural stem cells [68] and important for neural stem cell reprogramming [69]. The ase orthologue Achaete-scute complex homolog 2 (Ascl2) is critical for adult neurogenesis and proliferation [70, 71]. Herein, we demonstrate that the EFs, critical activators of the Drosophila neural stem cell program, are expressed in an A–P gradient along the developing CNS and that this gradient plays a key role in driving the wedge-like structure of the CNS, with its prominent anterior expansion. What is the evidence for a similar A–P gradient of the mammalian orthologues of the Drosophila neural stem cell program? To our knowledge, the A–P aspect of the mammalian neural stem cell program has not been extensively addressed. However, recent transcriptome analysis of the developing mouse forebrain versus spinal cord demonstrated elevated anterior levels of Sox1, -2, and -3 in the forebrain [7]. It is tempting to speculate that the conserved role of the PcG->Hox program in controlling proliferation [7] also extends into the conserved regulation of graded levels of an evolutionarily conserved neural stem cell (EF) program, acting to drive the expansion of the anterior CNS, common to most, if not all bilaterians.
From a cross between esc5 and esc21 mutants, esc5/esc21 females were selected. These females were crossed to escDf males (Df(2L)Exel6030) and set for embryo collection. esc homozygous mutant embryos, i.e., esc5/escDf or esc21/escDf maternal-zygotic mutant embryos, were identified by lack of the YFP-marked balancer chromosome.
Immunohistochemistry was conducted as previously described [21]. Primary antibodies were guinea pig anti-Dpn (1:1,000) and rat anti-Dpn (1:500) [74]; rabbit anti-phospho H3-Ser10 (PH3) (1:1,000) (cat. no. ab5176, Abcam, Cambridge, UK); rat anti-PH3-ser28 (1:1,000; cat. no. ab10543, Abcam, Cambridge, UK); chicken anti-GFP (1:2,000; cat. no. ab13970, Abcam); rat mAb anti-GsbN (1:10) [77] (provided by Robert Holmgren, Northwestern University, Evanston, IL, USA); rabbit anti-Abd-A (1:100) (provided by Maria Capovila, CNRS, Sophia Antipolis, France); mouse mAb anti-Ubx (1:10), mAb anti-Abd-A (1:10), mAb anti-Abd-B (1:10), mAb anti-Pros MR1A (1:10) (Developmental Studies Hybridoma Bank, Iowa City, IA, US); rabbit anti-Ase (1:1,000; provided by Yuh-Nung Jan, UCSF, San Francisco, USA); rabbit anti-SoxN (provided by Steve Russell, Cambridge University, Cambridge, UK); rat anti-Wor (1:1) (provided by Chris Q. Doe, University of Oregon, Eugene, OR, USA); rabbit anti-Hb and anti-Kr (1:500) (provided by Ralf Pflanz, MPI, Göttingen, Germany); mouse anti-Nub (1:100) (also detects Pdm2; provided by Steve Cohen and Hector Herranz, Temasek Life Sci, Singapore, Singapore); rat anti-Pdm2 (1:500; cat. no. ab201325, Abcam); rabbit anti-H3K27me3 (1:500; cat. no. 9733, Cell Signaling Technology, Danvers, MA, USA).
For fluorescent images, we used Zeiss LSM700 or Zeiss LSM800 confocal microscopes (Zeiss, Oberkochen, Germany) and merged confocal stacks using LSM software or Fiji software [78]. Images and graphs were compiled using Adobe Illustrator.
Embryos were fixed for 20 minutes in 4% PFA. After fixation, immunostaining was performed as previously described [21]. Embryos were stained with DAPI (cat. no. D9564, Sigma-Aldrich, Sweden AB, Stockholm, Sweden) at 1 μg/ml together with the secondary antibody incubation. An ImageJ language-based semiautomated macro was written and used to quantify the CNS volume using the Fiji software [78].
Embryos were dissected at precise stages, and mitotic NBs and daughter cells—as identified by PH3, Dpn, and Pros—were counted within each segment. An ImageJ language-based semiautomated macro was written and used for quantification with Fiji software [78].
Fluorescent images were analyzed using Zeiss LSM800 confocal microscopes, and Fiji software was used to visualize confocal stacks. To ensure identical staining conditions, control and mutant/misexpression embryos were dissected on the same slide and scanned using the same confocal settings. Dpn was used to identify the NBs. The integrated density (mean pixel intensity × area occupied by the signal) of individual cells was measured using Fiji software, focusing on a single 1-μm–thick confocal layer encompassing the center of the cell. The mean in control was set to 1, and experimental data were normalized to control.
Two-tailed Student t test was performed using Microsoft Excel 2016 or IBM SPSS V25.0 software (for specific statistical test used, see text and figures). Significance of p ≤ 0.05 is indicated with one star (*), p ≤ 0.01 with two stars (**), p ≤ 0.001 with three stars (***), and nonsignificant with “ns.” Microsoft Excel 2010 was used for data compilation and graphical representation. Figures and graphs were compiled using Adobe Photoshop and Adobe Illustrator.
|
10.1371/journal.pcbi.1003815 | Unbiased Functional Clustering of Gene Variants with a Phenotypic-Linkage Network | Groupwise functional analysis of gene variants is becoming standard in next-generation sequencing studies. As the function of many genes is unknown and their classification to pathways is scant, functional associations between genes are often inferred from large-scale omics data. Such data types—including protein–protein interactions and gene co-expression networks—are used to examine the interrelations of the implicated genes. Statistical significance is assessed by comparing the interconnectedness of the mutated genes with that of random gene sets. However, interconnectedness can be affected by confounding bias, potentially resulting in false positive findings. We show that genes implicated through de novo sequence variants are biased in their coding-sequence length and longer genes tend to cluster together, which leads to exaggerated p-values in functional studies; we present here an integrative method that addresses these bias. To discern molecular pathways relevant to complex disease, we have inferred functional associations between human genes from diverse data types and assessed them with a novel phenotype-based method. Examining the functional association between de novo gene variants, we control for the heretofore unexplored confounding bias in coding-sequence length. We test different data types and networks and find that the disease-associated genes cluster more significantly in an integrated phenotypic-linkage network than in other gene networks. We present a tool of superior power to identify functional associations among genes mutated in the same disease even after accounting for significant sequencing study bias and demonstrate the suitability of this method to functionally cluster variant genes underlying polygenic disorders.
| Plenty of gene variants have been associated with a disease, yet most of the heritability, along with the molecular basis, of common diseases remains unexplained. However, it is widely thought that the products of genes whose mutations are implicated in the same disease function together in the same biological pathways and it is the disruption of these pathways that underlies the disease. Such pathways are not well defined and their identification could help elucidate disease mechanisms. Consequently, groupwise functional analyses of gene variants to identify common disease-relevant pathways are becoming standard in next-generation sequencing studies, but we find that these analyses are confounded by coding-sequence length bias. We control for these bias and describe a phenotype-based approach which outperforms other methods in discerning functional associations among the disease-associated genes. We also demonstrate the suitability of this method to functionally dissect the gene variants underlying a complex disorder, the identified functional clusters offering insight into disease mechanisms.
| It is widely postulated that the products of genes whose variants are implicated in the same disease participate in the same biological function or process whose disruption leads to the disease [1], [2]. This concept is supported by examples of complex disease in which the proteins encoded by the implicated genes interact, form a molecular complex or function at different steps of the same biochemical pathway [3], [4]. As there is limited power to associate rare variants with disease by case–control studies, the use of functional-enrichment approaches that identify a shared function in a set of mutated genes is becoming standard in the interpretation of variants [5]–[8].
Since the function of many genes is not known and their classification to pathways is scant, functional associations between genes are often inferred from large-scale omics data [4], [6]–[9]. However, the suitability of such data types, including protein–protein interactions and gene co-expression networks, for functional-enrichment analysis remains unclear. Moreover, the inferred functional associations can be affected by confounding factors, potentially resulting in false positive findings. Thus, it is important to identify any bias affecting the implicated genes and control for them. Multiple exome-sequencing studies currently test variants for functional enrichment and yet there is no consensus concerning what to control for [6]–[9].
In this study, we have inferred functional associations between human genes from diverse data types and assessed the phenotypic agreement of the inferred gene–gene associations. We have examined different data types and networks and found that genes mutated in the same disease cluster more significantly in an integrated phenotypic-linkage network than in other gene networks. Examining the functional association between de novo gene variants, we have identified a confounding bias in coding-sequence length that we control for. We present a tool that identifies functional associations among genes mutated in the same disease even after accounting for significant sequencing study bias and demonstrate the power of this tool to functionally subcluster the gene variants underlying a polygenic disorder.
To test for functional associations among gene variants, we derived functional links between genes from diverse data types. For example, we calculated correlation coefficients from expression profiles, whereas gene annotation data were processed in the form of semantic similarity, which is a measure of relatedness between two genes assessed by the similarity of their annotations [10] (Figure 1A). The data were likely to include noise leading to false links and their reliability was unknown. To estimate and take into account the accuracy of the links, we evaluated the individual data types with a novel, phenotype-based method, by examining the semantic similarity between the mouse phenotypes of the genes they related to each other (Figure 1B). That is, each data type in turn indicated gene–gene linkages (gene pairs) and the accuracy of these links was assessed by considering the similarity of the phenotypes arising from the disruption of the unique mouse orthologs of these genes. We expected the data types to link together genes whose knockouts give rise to the same phenotypes, even if these mouse phenotypes were not necessarily expected to resemble human symptoms. The similarity of mouse phenotype annotations correlated with the similarity of human disease phenotypes (ρ = 0.223, P<2×10−16; Figure S1) and mouse phenotypes have been assigned to 6169 unique orthologs of human genes, 3.4-fold more than the 1801 genes annotated by the Human Phenotype Ontology (HPO; downloaded in 2012) [11]. Consequently, we used the phenotype annotations from the Mouse Genome Database [12] as the benchmark against which to evaluate other data types and set aside the HPO annotations for use as a test set for validation.
Integration of different data types into a combined network is expected to improve the accuracy of links and thus, in addition to considering individual data types, we also built an integrated gene network [13], [14]. For this, we selected data types that consistently linked together genes associated with similar mouse knockout phenotypes and that produced a positive correlation with the semantic similarity of mouse phenotypes (Table S1). For each data source suggesting functional links, we fitted regression curves in order to re-score the links so that any data-specific scores characterising the gene pairs were replaced with the semantic similarity that they corresponded to according to a regression function (see Figure 1B). When multiple data sources suggested functional linkage between the same two genes, we summed their link weights according to the approach of Marcotte and colleagues [15], thereby down-weighting less reliable data (see Methods). The resulting integrated gene network outperforms networks derived from the individual data types both in terms of coverage and accuracy (Figure 1C).
We corroborated the integrated phenotypic-linkage network by showing that genes whose perturbations are implicated in the same disease tend to be closely interlinked (Figures S2, S3, S4, S5, S6, S7). It is possible that their tendency to be closely interconnected is due to shared functional annotations assigned to them because they were implicated in the same disease in the literature. Also, we cannot assume that the associations of genes to phenotypes – forming the test sets – were made independently of any data type. Consequently, we turned to recently reported de novo mutations associated with developmental disorders that were identified independently of the data types included in the network.
Genes with de novo substitutions in patients with the same disorder [6]–[9], [16]–[18] showed a tendency to be more interconnected in the gene networks than random gene sets of the same size. However, as the interconnectedness of genes can be affected by confounding factors, it is important to identify any bias affecting the studied genes and control for them during the randomizations. We have found that the genes implicated through de novo sequence variants are biased in their coding-sequence (CDS) length, as longer genes are more likely to be mutated by chance (Figure 2). We also observe that genes with longer CDS tend to be interconnected (Figure 2C) and thus controlling for CDS length during the randomizations can significantly affect their relative degree of clustering (Figure 3). To control for coding-sequence (CDS) length during the randomizations, we have selected random genes the CDS length of which matched the CDS length of the studied candidate genes. Node degree has been previously identified as a confounding factor in functional analyses, particularly where an increase in degree results from study bias [19]. However, controlling for node degrees in a gene network does not correct the CDS length bias (Figure 3B). CDS length correlates very weakly with node degree (Spearman's ρ = 0.050). The length bias are highly significant in all the studied gene sets (Figure 2), while the node degrees are significantly different only in some of the candidate gene sets and there is no correlation between the node degree and mutational burden of genes (Figure S8). Having examined different data types and networks [15], [20], [21], we find that the disease-associated genes cluster more significantly in the integrated phenotypic-linkage network than in other gene networks (Figure 3).
We have inferred functional-association networks of human genes from diverse data types and assessed the phenotypic agreement of the inferred links. Having examined different data types and networks, we have found that genes mutated in the same disease cluster more significantly in an integrated phenotypic-linkage network than in other gene networks (Figure 3C). We note that another gene network, called NETBAG, has been developed by Gilman and colleagues [22]. We could not access NETBAG for the performance comparison. Nevertheless, Gilman and colleagues state the use of shared disease associations among 478 human genes as the gold standard in their network construction [22] and the used disease associations originate from a study published in 2001 [23]. By comparison, our method takes advantage of over 100,000 mammalian genotype–phenotype relations and fully exploits bio-ontologies by means of semantic similarity, with both advances expected to enhance greatly the phenotypic-linkage network that we explicitly present here.
Examining the functional association between de novo gene variants, we have identified a confounding bias in coding-sequence length that we control for to avoid false positive findings. Numerous implicated variants are in fact expected to be neutral mutations but they are more likely to appear in genes with longer CDS, leading to a tendency of the implicated genes to be interconnected in gene networks (see Figure 2). These bias have confounded functional analyses and likely led to an overestimation of functional clustering in former studies. We have found that the CDS-length bias were highly significant in all the studied gene sets, including the unaffected siblings, while the node degrees were not. The higher node degrees in some of the candidate gene sets may indicate a functional signal, as the same genes are significantly more conserved (Figure S8). We conclude that controlling for CDS length in functional analyses of gene variants is appropriate.
One way of controlling for CDS length is to compare the interconnectedness of the implicated genes with that of genes mutated in unaffected controls [9]. However, we observe that the control genes tend to be less interconnected than random genes (Figure S9), which suggests that our way of controlling for CDS length (see Methods) is more conservative.
The nature of the phenotypic-linkage network suggests that the clustered genes function together in the same disease-relevant cellular pathways (Figure S10). The functional convergence that we identify among the three sets of genes from independent exome studies of autism spectrum disorder demonstrates that the method is able to detect biological coherence among variant genes (Figures 3 and S10). Throughout, we have considered the larger class of non-synonymous variants which is likely to possess a more diluted signal than nonsense variants. As with all clustering methods, our method is sensitive to the number of variants identified and the likelihood of their causal relation. Half of our study sets included only 5–10 genes with nonsense variants, between which we either did not find any functional links or the sum of link weights was not significantly higher than expected after controlling for CDS length. For studies of rare or de novo variants derived from a single or small number of genomes, gene prioritizing methods based upon phenotypic similarity may be more appropriate [24]. Continuing efforts to systematically phenotype model organisms and to enrich the phenotype ontologies could further improve the resulting phenotypic-linkage networks that are constructed [25]. The integrated network toolkit is made available at http://groups.mrcfgu.ox.ac.uk/webber-group/resources.
To gain the most information about genes whose variants may be relevant to disease and to explore the functional relations between them, we collected large amounts of functional genomics data on human genes and their mouse orthologs. We wanted the data sets to inform us about the functional similarity of genes, therefore we processed them such that they indicated gene–gene links. For every data type except physical interactions, we derived a score characterising gene pairs, such as the correlation coefficient from expression profiles or semantic similarity from gene annotations.
All gene annotation data (such as GO, KEGG, Reactome, InterPro and mouse and human phenotype annotations) were processed in the form of semantic similarity, which is a measure of relatedness between two genes assessed by the similarity of their annotations [10]. The terms used to annotate genes have an information content (IC) defined as:where p(a) is the proportion of genes annotated with term a or its descendent terms among all genes with an annotation.
We used Resnik's [26] measure together with the GraSM approach [27] to calculate the similarity of terms organized in a hierarchical ontology, defining the semantic similarity between any two terms t1 and t2 as the average IC of their disjunct common ancestor terms (see Figure 1A):
To measure the functional relatedness of two genes, we compared their annotations with the maximum (max) and best match average (bma) methods [28]. Let T1 denote the set of terms annotated to gene g1and T2 denote the set of terms annotated to gene g2, the semantic similarity of their annotations according to the max approach is then given by:while the semantic similarity of their annotations according to the bma approach is defined as:
To estimate the reliability of the individual data sets, we evaluated them by examining the semantic similarity between the phenotypes associated with the unique mouse orthologs of the genes they linked to each other. For each data set, we derived gene–gene linkages (gene pairs) with data-specific scores characterizing the strength of a linkage and ordered the gene pairs by their score from largest to smallest. Next, we calculated and plotted the median semantic similarity of mouse phenotype annotations for bins of 1,000 gene pairs (see Figure 1B).
We tested if the data types linked together genes whose knockouts influence the same phenotypes. When the strongest linkages derived from a data set did not correspond to higher semantic similarities of phenotypes than expected by chance, we did not include the links from the given set in the integrated gene network (Table S2).
We selected data sets that produced a positive correlation with the semantic similarity of mouse phenotypes (Table S1) and fitted regression curves in order to re-score the links so that any data-specific scores characterising the gene pairs were replaced with the semantic similarity of phenotypes that they corresponded to according to a linear regression function (Figure 1B). Thus all gene pairs that had an original data-specific score were re-scored, including those that did not have phenotypic annotations.
By re-scoring the data types with a universal benchmark we weighted them in proportion of their relative accuracy. When multiple data sources suggested functional linkage between the same two genes, we summed their link weights (Figure S11) increasingly down-weighting less reliable data according to a formula used by Marcotte and colleagues [15]:where L represents a re-scored link weight from a single data set, L0 being the largest link weight among all the links between the given two genes, i is the index of the remaining links ordered by their weights for the gene pair and D is a free parameter. We optimized the value of this parameter and used D = 5 in integrating data types to create the final phenotypic-linkage network.
In testing for functional enrichment in a set of genes, the degree of functional association between the genes can be compared to that calculated for randomized gene sets. As the degree of functional association can be affected by confounding bias, it is important to identify such bias affecting the studied gene set and control for them. To control for coding-sequence (CDS) length during the randomizations, we selected random genes the CDS length of which matched the CDS length of the studied (mutated) genes. For each of the studied genes in turn we assigned a list of 100 genes with the same or most similar CDS length, using the longest CDS of each gene. Random gene sets were then assembled by selecting one random gene from each of these lists.
|
10.1371/journal.pntd.0002131 | Analysis of Schistosomiasis haematobium Infection Prevalence and Intensity in Chikhwawa, Malawi: An Application of a Two Part Model | Urinary Schistosomiasis infection, a common cause of morbidity especially among children in less developed countries, is measured by the number of eggs per urine. Typically a large proportion of individuals are non-egg excretors, leading to a large number of zeros. Control strategies require better understanding of its epidemiology, hence appropriate methods to model infection prevalence and intensity are crucial, particularly if such methods add value to targeted implementation of interventions.
We consider data that were collected in a cluster randomized study in 2004 in Chikhwawa district, Malawi, where eighteen (18) villages were selected and randomised to intervention and control arms. We developed a two-part model, with one part for analysis of infection prevalence and the other to model infection intensity. In both parts of the model we adjusted for age, sex, education level, treatment arm, occupation, and poly-parasitism. We also assessed for spatial correlation in the model residual using variogram analysis and mapped the spatial variation in risk. The model was fitted using maximum likelihood estimation.
The study had a total of 1642 participants with mean age of 32.4 (Standard deviation: 22.8), of which 55.4 % were female. Schistosomiasis prevalence was 14.2 %, with a large proportion of individuals (85.8 %) being non-egg excretors, hence zero-inflated data. Our findings showed that S. haematobium was highly localized even after adjusting for risk factors. Prevalence of infection was low in males as compared to females across all the age ranges. S. haematobium infection increased with presence of co-infection with other parasite infection. Infection intensity was highly associated with age; with highest intensity in school-aged children (6 to 15 years). Fishing and working in gardens along the Shire River were potential risk factors for S. haematobium infection intensity. Intervention reduced both infection intensity and prevalence in the intervention arm as compared to control arm. Farmers had high infection intensity as compared to non farmers, despite the fact that being a farmer did not show any significant association with probability of infection.
These results evidently indicate that infection prevalence and intensity are associated with risk factors differently, suggesting a non-singular epidemiological setting. The dominance of agricultural, socio-economic and demographic factors in determining S. haematobium infection and intensity suggest that disease transmission and control strategies should continue centring on improving socio-economic status, environmental modifications to control S. haematobium intermediate host snails and mass drug administration, which may be more promising approaches to disease control in high intensity and prevalence settings.
| Schistosomiasis is one of the great causes of morbidity among school aged children in the tropical region and Sub Saharan Africa in particular. It's mainly transmitted through contact with water infested with intermediate host snail Cercariae. Currently, over 200 million people are estimated to be infected in SSA alone. Here, we used robust and contemporary statistical methods in a two part application to analyse risk factors for S. haematobium infection intensity and prevalence. We found that S. haematobium was more common in younger children as compared to older children, thus making the infection and prevalence age dependent. We also found that mass chemotherapy reduced both infection prevalence and intensity. We found that dominance of agricultural, socio-economic and demographic factors in determining S. haematobium infection risk in the villages carries important implications for disease surveillance and control strategies. Therefore disease transmission and control strategies centered on improving strategies involving socio-economic status, environmental modifications to control S. haematobium intermediate host snails and mass drug administration may be more promising approaches to disease control in high intensity and prevalence settings.
| According to [1], Schistosomiasis infections affect an estimated 779 million people, with consequences in health nutritional and educational development of infected individuals [2]. The disease causes an annual loss of 4.5 million disability-adjusted-lifeyears (DALYs) [3]. In SSA alone, 207 million individuals are estimated to be infected with Schistosomiasis: S.haematobium and S.mansoni [1]. S.haematobium is reported to be endemic in 53 countries in the Middle east and most of the African continent including islands of Madagascar and Mauritius [4], whereas S.mansoni is mostly endemic in sub-Saharan Africa [4]. Schistosomiasis can be effectively treated with single dose oral therapies of praziquantel that are safe, inexpensive and required at periodic intervals [5]. Treatment is typically implemented through mass chemotherapy whereby the entire at-risk population is treated, as part of either school or community- based campaigns, referred to as mass drug administration (MDA).
The transmission intensity of Schistosomiasis is a function of parasitic worm load within a group of individuals, which can indirectly be quantified by the number of eggs that are excreted. Host heterogeneities in exposure and susceptibility to infection may lead to an aggregated distribution of worm burden across individuals [6]. For this reason, a few individuals would harbour large numbers of worms, whilst the majority of individuals are uninfected or only carry a low worm burden [6]. In addition, widely used diagnostic approaches for Schistosomiasis like the Kato-Katz technique for S.mansoni diagnosis fail to detect some infected individuals, particularly when only a single stool sample is examined and infection intensities are light [7]. Due to these two issues, often a large proportion of individuals are considered as “zero egg excretor” [6]. The standard Poisson distribution, which assumes equal mean and variance, commonly employed to model such count data, is inappropriate to fit observed egg counts since the variance of the counts is much larger than their mean, a case known as over-dispersion [8]. The use of negative binomial (NB) distribution has been proposed to model the extra-Poisson variation [9], and applications of NB in analysing helminth egg counts are many [8], [10].
Although NB models may be ideal for over-dispersion, they may not be suitable when data is zero-inflated. Other distributions like the hurdle models or zero inflated (ZI) or zero augmented models that may be more appropriate for modeling data with such excess zeros are reported [8]. These models can have more than one mode, including a mode at zero. ZI models attempt to account for excess zeros, i.e., zero inflation arises when one mechanism generates only zeros and the other process generates both zero and nonzero counts hence they can be expressed as a two-component mixture model where one component has a degenerate distribution at zero and the other is a count model [11]. ZI models estimate two equations, one for the count model and one for the excess zeros. ZI models assume that a proportion of individuals have no chance to be infected, as they are not exposed. In other words, there is a process which determines whether an individual is likely to be infected at all and a second process determining the number of excreted eggs among those who are at risk of infection. Zero inflated Poisson(ZIP) models assume that the number of excreted eggs follows a Poisson distribution. Zero-inflated negative binomial (ZINB) models assume that the number of worms among those who are at risk of infection has a negative binomial distribution [6].
ZI count data are common in a number of applications. Examples of data with too many zeros from various disciplines include agriculture, econometrics, patent applications, species abundance, medicine, and use of recreational facilities [8]. The zero-inflated Poisson(ZIP) regression models with an application to defects in manufacturing is described in [12], while zero-inflated binomial (ZIB) regression model with random effects into ZIP and ZIB models are defined in [13].
The idea for a hurdle model, a modified count model in which the two processes generating the zeros and the positives are not constrained to be the same, was developed in [14]. The two processes are modeled using a mixture of two models (i.e, two part or a hurdle model). The first part is a binary outcome model, and the second part is a truncated count model. Such a partition permits the interpretation that positive observations arise from crossing the zero hurdle or the zero threshold. The first part models the probability that the threshold is crossed, in our case thatan infection occurred. In principle, the threshold need not be at zero; it could be any value, and it need not be treated as known. The zero value has special appeal because in many situations it partitions the population into subpopulations in a meaningful way, one on infection status and the other for those infected it captures intensity. In contrast to the zero-inflated model, the zero and non-zero counts are separated in the hurdle model [15] which makes them very useful in inferential studies. Hurdle models are sometimes referred to as zero-altered models [16]. Zero-altered Poisson and negative binomial models are thus referred to, respectively, as ZAP and ZANB. They have also been termed overlapping models [17].
The application of hurdle or two part models in epidemiology has not been common so far. Use of ZI models have been reported. One such an application was in Cote d'Ivore, in which a ZINB model within a model-based geostatistics (MBG) framework for S. mansoni infection was applied [6]. This study showed that geostatistical ZI models produce more accurate maps of helminth infection intensity than the spatial negative binomial counterparts. However, to our knowledge, no hurdle or two part model has been applied in Schistosomiasis or geohelminth epidemiology.
This paper demonstrates the applications of hurdle models to helminth epidemiology (S. haematobium) and encourage its wider application in helminth disease control programmes. Its advantage is that it allows joint modeling of infection status and intensity. Although, a multinomial model maybe used [18]–[20], its limitation is that it involves stratifying egg counts, leading to a loss of information, whereas the negative binomial hurdle model approach makes full use of intensity data on a continuous scale, therefore, ideal to model latent infection intensity. In addition, hurdle models are robust when over-dispersion is present. In [8], it was concluded that the ZIP models were inadequate for the data as there was still evidence of over-dispersion. Moreover, the negative binomial hurdle model, which allows for over-dispersion and accommodates the presence of excess zeros through a two-part model has a natural epidemiological interpretation within the case study considered here.
The data which motivated this work were collected in 2004 in Chikhwawa district, in the Lower Shire Valley-southern Malawi. This is a rural area whose population is mainly engaged in subsistence farming. This area lies between 100 and 300 m above sea level. The rainy season extends from December to March. Temperatures can rise up to in months preceding rainy season. Malaria is known to be holoendemic [21].
Data were collected in eighteen villages, purposively selected from the control and intervention arms of a cluster randomized study design. There was only one round of treatment following community based and house to house approaches for mass drug administration (MDA). Over 90 percent of the eligible population were treated. All infected participants in non-intervention arm received appropriate treatment. After the follow-up assessment, both arms had mass treatment. In the study, polyparasitism was considered basing on the number of species an individual was hosting. The focus was on Hookworm, S. mansoni, S. haematobium and Ascaris. Polyparasitism is the epidemiology of multiple species parasite infections. Ten percent of the households were randomly selected from the villages for baseline survey using random number tables [22].
Subjects for geo-helminth survey were selected using a two stage-design. Briefly, at first stage, villages were selected, then at second stage, sample of households was listed and chosen. In the selected households all members aged one year and above were invited to participate. Consenting individuals had their demographic details completed and were given full body clinical examinations (except genitals for females) for chronic manifestations of human helminths. In addition they had anthropometric measurements taken and were asked to provide a single fresh stool and urine sample. All individuals (aged>1 year) were requested to provide a finger prick blood sample [22]. Further details are provided in [22].
Fresh stool samples were transported in a cooler box to the laboratory and processed within four hours of collection. A single Kato-Katz thick smear was prepared from each sample and immediately examined under a light microscope for parasite eggs (within 15–20 minutes). Standardized and quality controlled procedures were followed. Briefly 41.7 mg of sieved stool was placed on a microscope slide through a punched plastic template. Ova for each parasite observed were counted and expressed as eggs per gram (epg) of stool. Five percent of the slides were randomly selected for re-examination for quality control purposes [22].
Urine samples were processed on the day of collection. A measured volume (maximum 10 ml) was centrifuged at 300 rpm for five minutes. The sediment was then examined under a light microscope. The eggs seen were counted and the intensity of infection per 10 ml of urine accordingly determined. All those infected were treated with praziquantel at 40 mg/kg [22].
The study that collected data from Chikhwawa, Malawi received ethical clearance from Malawi's College of Medicine Research Ethics Committee (COMREC) [22]. Individual informed consent was orally obtained from each participant or (if they were aged<16) from one of their parents or a legal guardian. COMREC approved oral informed consent because the study was determined to be of minimal risk. The consent process was a four stage process. First stage, oral informed consent was obtained at the traditional authority (TA) level. Second stage, at village head level and third stage at the household level from the head of the household and fourth at individual level from each individual in the household (if applicable) else from parent/guardian if an individual was aged<16. Registers were kept for documentation whereby, for each individual in the selected household, a column was kept to indicate whether an individual had orally consented to participate in the study or not.
Various statistical models have been developed to model helminths disease burden as reviewed in the introduction. For purposes of this paper, we assumed a negative binomial logit hurdle (NBLH) model for joint analysis of infection prevalence and intensity of Schistosomiasis hematobium in Malawi. Following on [11], a NBLH model can be written as:(1)(2)where are observed counts taking values for each individual . The probability of infection is , such that indicates there are no zero counts and the model reduces to a truncated Negative Binomial distribution (TNegBinom); while means there are no infections. The observed counts are modelled by assuming two processes:(3)The first is assumed to model the infection prevalence (first hurdle) and the other the intensity of infection (second hurdle). The first hurdle assumes a binary outcome defining whether an individual is infected or not. This is modeled as a logit regression for a given set of risk factors . After determining infection status we are interested in analyzing the number of eggs - as a measure of intensity of infection, which is defined by the second hurdle. We model the second hurdle as a negative binomial regression model for a given set of risk factors . The NB model is suited for count data with over-dispersion. In many cases, the same risk factors are used in the logit and count regression models, i.e. . The two regression models, incorporating the risk factors, are given by:(4)(5)The model parameters and are estimated using maximum likelihood estimation in which the likelihoods (or log-likelihoods) are maximized separately.
The covariates included in the model are given in Table 1. Age and polyparasitsm were fitted as continuous variables, while sex, education levels, village type, fishing, gardening and occpuation were entered in the model as categorical variables, with the first category of each variable selected as the reference group. For both parts of the model we used the same set of covariates. We also fitted a number of count models, with the Poisson as the null model, for comparison and evaluated the number of zeros each model correctly predicts. We also compared model fit using AIC and zero capturing. A difference of 10 indicates the model with the smallest AIC is superior to others. Furthermore, deviance residuals were assessed for spatial correlation using variogram and were subsequently mapped using kriging to depict spatial variation in risk. Statistical model fitting was carried out using Political Science Computational Laboratory (PSCL) package [23] in R statistical software (The R Foundation for Statistical Computing, Version 2.14.0). Variogram analysis and kriging were implemented in geoR [24].
Table 1 gives summary statistics for study participants. The study had 1642 participants of which 55.4 % were female. The mean age (years) of 32.4 (standard deviation: 22.8). Of these, 324 had hookworm representing 19.7 % of sample population, 71 of these had S. mansoni representing 4.3 % and 233 had S. haematobium representing a prevalence of 14.2 %.
Figure 1 shows that a large proportion of individuals i.e. 85.8 % were “zero egg excretors” hence the data were inflated with zeros. The likelihood ratio test for overdispersion between Poisson and negative binomial at = 0.05 showed a critical value test statistic = 2.7 with a test statistic = 10606.5, p-value<0.001. Indeed, there was overwhelming evidence of overdispersion. This was confirmed by the presence of excess zeros (Figure 1).
Using the AIC and zero capturing, the predicted counts using the NBLH indicate a closer fit with the observed values. In Table 2, AIC results show that the NBLH offers a better fit compared to using Poisson Logit Hurdle (PLH) or a negative binomial (AIC = 3,482 for NBLH; AIC = 6,854 for PLH and AIC = 3,576 for NB respectively). The AIC further showed a difference of 10,700 for the NBLH compared to the Poisson and a difference of 19 comparing NBLH with ZINB, thus NBLH is superior among all competing models. With regards to zero capturing, the Poisson model was again not appropriate as it could only capture 515 of the zeros whereas the NB-Zero adjusted based models were much better in capturing the zero counts. The NBLH model captured 971 zeros which were equal to the observed (Table 3). Since NB logit hurdle model offered the best fit to zero inflated helminth data in terms of the AIC (minimum value for all the models fitted) as well as true zero count capturing, it therefore became a natural choice for fitting a final model to model helminth infection intensity and determination of factors that foster infections.
Table 4 provides estimates for the fixed effects. The probability of infection was found to be associated with age (Odds Ratio [OR] = 0.97, 95 % Confidence Interval [CI]: 0.96–0.99), the risk of infection was decreasing with age. This assumed a linear relationship with age; 6 years being the baseline age. The risk of infection was low in males than in females (OR = 0.61, 95 % CI: 0.41–0.89). The association between risk of infection with education at both primary level (OR = 1.18, 95 % CI: 0.81–1.71) and secondary level (OR = 1.37, 95 % CI: 0.41–4.60) relative to those with no education was not significant (p-value = 0.62). Infection probability was found to be associated with village type; whether one was in the intervention area or control area (OR = 0.38, 95 % CI: 0.26–0.54, p-value<0.001). Those in the intervention area were at a reduced chance of infection relative to those in control area. We observed a negative association between infection probability and fishing (OR = 0.73, 95 % CI: 0.44–1.20) though not significant; contrary to the expectation. Working in the garden was observed not to be significant albeit it was positive (OR = 1.34, 95 % CI: 0.90–1.99). Again, occupation (farmer/other) showed a negative association with infection probability though with marginal significance (OR = 0.61, 95 % CI: 0.35–1.06) with a p-value = 0.17. We also noted that chances of infection were increasing with number of parasite species an individual was hosting (Table 4) (OR = 7.30, 95 % CI:5.56–9.59).
From Table 4, it was observed that infection intensity reduced with an increase in age (Relative Risk [RR] = 0.96, 95 % CI: 0.95–0.98). Similar to infection prevalence, a linear relationship was assumed between infection intensity and age. There was no difference of infection intensity between males and females (RR = 1.03, 95 % CI: 0.72–1.47). Primary school children showed a high infection intensity relative to those that are in pre-school level (RR = 1.54, 95 % CI: 1.08–2.19) whereas those in secondary level showed a reduced infection intensity (RR = 0.34, 95 % CI: 0.11–1.06) though not significant. There was a reduced risk for those in intervention area relative to those in the control area, though, not significant (RR = 0.81, 95 % CI: 0.58–1.13). A positive association was also observed between those who did fishing in Shire river relative to those who did not fish (Table 4). We observed an increased infection intensity in those working in the gardens relative to those who did not (RR = 1.21, 95 % CI: 0.82–1.81), albeit not significant and also increased infection intensity for farmers compared to non-farmers (RR = 1.83, 95 % CI: 1.16–2.91).
Estimating the continuous surface using variogram analysis and kriging, spatial patterns in the residuals were observed and subsequently mapped. There was some degree of spatial dependence in residuals distribution across the study area, as evidenced by the spherical model (Figure 2). The magnitude of spatial correlation decreased with separation distance until at distance of 10 km. The predicted spatial surface, in Figure 3, showed a relatively increased risk of infection in the northern part of the study area compared to other areas. Low risk areas were in the southern parts, more especially in the south-eastern part of the study region (Figure 3).
The current study found a prevalence of 14.2 % for S. haematobium in Chikhwawa district. This prevalence was well below national estimates, which a previous study in Malawi indicated to be between 40 and 50 % [25]. The finding serves to highlight the fact that Schistosomiasis infections are highly localised and that nationwide surveys tend to overlook the focus of heterogeneity of infection. Indeed, in a study conducted in the northern lakeshore area [26], school children from four schools screened for Schistosomiasis reported a wide range of prevalence: 5 %–57 % of S. haematobium infection. A national survey, representative of all school children in the country, and undertaken just before the rainy season, showed far lower levels of 7 % for S. haematobium [25].
We used robust, contemporary statistical methods in a two part application to analyse risk factors for S. haematobium infection intensity and prevalence. This resulted in estimates of parasitic infection prevalence and intensity that could be used in control programme planning by channeling resources to areas with a known high disease burden. In this study we have looked at the intensity and prevalence of S. haematobium in relation to factors such as age, sex, education level, village type, fishing in Shire river, working in gardens, occupation and polyparasitism. Polyparasitism is the epidemiology of multiple species parasite infections. In the study, polyparasitism was based on the number of species an individual was hosting. The focus was on Hookworm, S. mansoni, S. haematobium and ascaris. The study confirms the critical importance of ascertaining the infection intensity.
We found that S. haematobium infection intensity reduced with age, this confirms what previous studies found. In common intestinal helminths such as Ascaris lumbricoides (large roundworms) and Trichuris trichiura (whipworm) and also Schistosomiasis, children are more heavily affected and infected than adults [27]. Several other studies have reported that school-aged children show high infection intensity and prevalence [25], [28], [29]. Fishing in Shire river and working in gardens along the river were potential risk factors for exposure to schistosomes and subsequent infection because transmission requires contact with the aquatic habitat of intermediate host snails [30]. This is in line with results from a study that was conducted in western Africa [20], that contact with water bodies that are a habitat for intermediate host snails is one of the main risk factors. Results showed low probability of infection for males compared to females. This could be explained by a number of factors including that Malawi being an agriculture based economy, and that mainly agricultural activities are carried out by females, hence they are more exposed to risk factors such as working in gardens and farming. Schistosomiasis is water dependent disease and the incidence is usually more amongst people who constantly get into contact with the schistosome infected waters through activities such as farming, fishing, swimming and washing [30].
Results from the study showed that individuals who had received chemotherapy cure for helminth showed reduced risk of infection as well as infection intensity as compared to those in the control area. Studies have shown that MDA significantly reduces Schistosomiasis infection [31], [32]. Evidence has shown that, following chemotherapeutic cure of S. mansoni or S. haematobium infection, older individuals display a resistance to re-infection in comparison to younger children [33]. Therefore there is need to channel integrated control and interventions for helminths to areas with diseases burden in order to reduce and/or eradicate the infections - more especially towards school age children. Several studies have shown that having one infection, is a risk factor for having other infections [34]. It is conceivable that the first parasite that establishes an infection may modulate the immune response in such a way that it makes it easier for the next [22].
Worthy noting were differences that existed in associations between infection probability and infection intensity. For gender, males had a reduced risk of infection as compared to females (negative association) but high infection intensity (positive association). This could possibly be explained by the fact that women were mostly involved in agricultural activities there by being more exposed. Also for those infected, many studies find that men visit public health care facilities much less frequently than do women [35] hence the high intensity. Poly-parasitism was positively associated with infection probability but had a negative association with infection intensity. This could be explained by the fact that having other parasites increases the chance of the body being susceptible to new parasite infections [34]. Again, secondary level of education had a positive association with infection probability but showed a negative association with infection intensity. This finding could be explained by the fact that an increase in education level corresponds to increase in age which comes with increased risk-behaviour of older school children who frequently contact schistosome-infested water for both domestic and livestock purposes relative to younger children [36] hence increased infection prevalence. At the same time, an increase in education may correspond to increased awareness and access to treatment [37] by this group hence reduced infection intensity. Those with the highest level of education, through high school, have showed the lowest mean infection intensity [37]. Being a farmer had a negative association with probability of infection and a positive association with infection intensity. The finding was in line with what was reported in [37]; farmers showed the highest levels of Schistosomiasis infection among occupational groups. Both education and occupation are proxies for the nature and intensity of water contact [37]. Individuals become infected by prolonged contact (like irrigating farm, bathing, washing or swimming) with fresh water containing free-swimming Cercariae [30].
We believe that the apparent dominance of agricultural, socio-economic and demographic factors in determining S. haematobium infection risk in the villages carries important implications for disease surveillance and control strategies. Prevalence of S. haematobium was highly associated with age of an individual as well as working in the garden and also number of parasites an individual hosted. Furthermore, S. haematobium infection intensity was associated with gender, education level, garden, occupation and village type (intervention). Cercariae control control through environmental modifications and strategies involving socio-economic status improvement and MDA may be more promising approaches to disease control in this setting.
Finally, zero adjusted methods represents a key advance in the epidemiological analysis of helminth disease data inflated with zeros. There are an increasing number of examples in the published literature where two part methods are being used for zero inflated data for helminths disease's control planning and implementation programmes [38], [39]. Ease of implementation and straightforward interpretation of the components and its direct link with the observed data, makes the negative binomial logit hurdle model definitely a valuable alternative for researchers analysing zero-inflated count data for helminths.
|
10.1371/journal.pbio.1001292 | Global Gradients in Vertebrate Diversity Predicted by Historical Area-Productivity Dynamics and Contemporary Environment | Broad-scale geographic gradients in species richness have now been extensively documented, but their historical underpinning is still not well understood. While the importance of productivity, temperature, and a scale dependence of the determinants of diversity is broadly acknowledged, we argue here that limitation to a single analysis scale and data pseudo-replication have impeded an integrated evolutionary and ecological understanding of diversity gradients. We develop and apply a hierarchical analysis framework for global diversity gradients that incorporates an explicit accounting of past environmental variation and provides an appropriate measurement of richness. Due to environmental niche conservatism, organisms generally reside in climatically defined bioregions, or “evolutionary arenas,” characterized by in situ speciation and extinction. These bioregions differ in age and their total productivity and have varied over time in area and energy available for diversification. We show that, consistently across the four major terrestrial vertebrate groups, current-day species richness of the world's main 32 bioregions is best explained by a model that integrates area and productivity over geological time together with temperature. Adding finer scale variation in energy availability as an ecological predictor of within-bioregional patterns of richness explains much of the remaining global variation in richness at the 110 km grain. These results highlight the separate evolutionary and ecological effects of energy availability and provide a first conceptual and empirical integration of the key drivers of broad-scale richness gradients. Avoiding the pseudo-replication that hampers the evolutionary interpretation of non-hierarchical macroecological analyses, our findings integrate evolutionary and ecological mechanisms at their most relevant scales and offer a new synthesis regarding global diversity gradients.
| Understanding what determines the distribution of biodiversity across the planet remains one of the critical challenges in biology and has gained particular urgency in the face of environmental change and accelerating species extinctions. Our study develops a novel analytical framework to jointly evaluate historical and contemporary environmental predictors of the latitudinal gradient in the diversity of terrestrial vertebrates. The number of vertebrate species is greater in warm, productive biomes, such as tropical forests, that have both a large size and a long evolutionary history. Using just a few key predictor variables—time, area, productivity, and temperature—we are now able to explain more than 80% of the variability in biodiversity among bioregions. By integrating each of these factors at both the regional and local scale in a hierarchical model, we are able to provide a consensus explanation for broad-scale diversity gradients that encompasses both ecological and evolutionary mechanisms.
| The uneven distribution of species diversity is a key feature of life on Earth and has myriad implications. While the scale-dependence of the determinants of the global variation in diversity is well acknowledged [1]–[6], to date a quantitative accounting of the roles of history and environment in generating and maintaining gradients in species richness is still lacking. Over the past three decades, increased data availability has facilitated analyses of contiguous geographic patterns in species richness at relatively fine spatial grains (100–200 km) at both continental [7]–[9] and global scales [10],[11]. At these spatial resolutions, environmental variables such as productivity or temperature have been shown to offer extremely strong statistical predictions of species richness [8],[11]–[18]. However, it has been difficult to connect these results directly with underlying evolutionary and ecological processes. One problem is that the ultimate drivers underpinning diversity, namely speciation and extinction [19], operate at scales much larger than the spatial resolution (e.g., 100 km grids) of most analyses. A number of studies have confirmed the strong effect of regional richness on local richness [1]–[3],[6],[20] and have speculated on the role of energy driving diversification at regional scales [21]–[24] as well as sorting at local scales [25]–[27]. But attempts to integrate them at the appropriate scale have been limited, and we know of no study that has quantified the effect of productivity on richness gradients jointly at regional and local scales and both in terms of evolutionary and ecological processes.
Another impediment to interpretations of gridded richness analyses has been that species' geographic ranges are generally much larger than, for example, 100 km×100 km grid cells, resulting in geographically non-random patterns of pseudo-replication, inflated spatial autocorrelation, and an overrepresentation of wide-ranging species and their respective climatic associations [8],[28]. These issues have to date precluded straightforward evolutionary and ecological interpretations of macroecological environment correlations of gridded richness patterns [5],[29]. While partly motivated by limits in the knowledge of fine-scale species distributions [30], macroecological analyses have also been conducted using, for example, ca. 800 ecoregions as spatial units [14],[31], but these regions still incur significant and geographically variable redundancy in species. We are not aware of a study on richness gradients that has successfully overcome this problem and thus truly have given each species equal weight.
Finally, while there is little doubt about the importance of time for diversification [32]–[34], attempts to date to invoke paleoclimate for understanding richness have been hampered by the lack of data, especially at deeper time scales. Several studies have linked relatively recent climatic oscillations, for example, those causing quarternary high-latitude glaciation, to geographic richness patterns [22],[35]–[37]. The geography of deeper time climate conditions and exactly how it relates to the tempo of past clade diversification is inherently difficult to estimate. But given deep conservatism in the environmental (e.g., biome) associations within clades [38],[39] compared to relatively dynamic geographic ranges [40], clades are expected to much more strongly track climatically defined regions, or biomes, rather than specific geographical locations over evolutionary history. The ages of biomes may thus offer a promising avenue for understanding the role of paleoclimate contributing to contemporary patterns of species richness and have recently been successfully correlated with both turtle and tree richness at the regional scale [41],[42]. To date, analyses connecting the age and area of regions to finer grain richness patterns have not been attempted.
Here, we aim to address these problems with a hierarchical framework that integrates the drivers of regional diversification of species with those of their sorting into finer grain assemblages at their respective scales of influence. We use this model to test the relative importance of past spatio-temporal variation of climatic conditions (specifically time-integrated area and productivity) versus contemporary environment for explaining both the regional and finer scale variation in the species richness of terrestrial vertebrates worldwide. Due to environmental niche conservatism, organisms are generally restricted to climatically defined bioregions, or “evolutionary arenas,” characterized by in situ speciation and extinction. We expect differences in species richness between such regions to arise from different levels of net diversification (speciation – extinction over time). The number of speciation and extinction events should vary among regions due to differences in the sizes of populations over time and the opportunities for reproductive isolation for all resident taxa [32]. We expect these drivers to be associated with today's area [29],[32],[43] and energy availability (i.e., productivity) [8],[11]–[14] of bioregions but, critically, also with the past levels of these factors—that is, how bioregions have varied in areal extent and productivity over time [42]. Furthermore, regional rates of diversification have been hypothesized to vary with temperature and its effects on activity and biological rates such as rates of molecular evolution or species interactions [15]–[17],[44]. We expect all of these drivers in concert to shape broad-scale gradients of diversity and predict that in an integrative assessment of regional differences in diversity, (i) models accounting for the temporal availability of area or productivity will outperform those without (i.e., regions that are older and/or have in the past been larger in extent will support higher vertebrate species richness than younger and/or smaller regions), (ii) area times energy availability (net primary productivity) will be a stronger predictor of richness than area alone [13],[45], and (iii) average bioregion temperature will positively affect richness above and beyond the effects of productivity and have a stronger effect in ectotherms compared to endotherms [15].
We test these predictions for the 32 main subdivisions, or “bioregions,” of the world based on vegetation type and major landmass (Figure 1, Tables S1 and S2) [46],[47]. We excluded montane regions (which exclusively harbor ca. 5% of vertebrate species and represent ca. 15% of global land area) due to their extremely steep environmental gradients and associated species turnover, which impedes reliable bioregional delineation and estimates of their extent over time. Over historical time-scales these climatically and geographically distinct bioregions have been characterized by similar environmental and climatic conditions, but have changed in size and shape over time within their respective realms [42]. All bioregions are within the range of scales over which allopatric speciation of terrestrial vertebrate speciation typically occurs (100–1,000 km scale [48]) and may thus be considered bio-climatically and geographically distinct “evolutionary arenas.” After deriving time-integrated models of bioregion species richness, we then in a second step assess their ability to predict the variation in richness at the scale of 110 km grid cell assemblages. We make these finer scale predictions first under a model of simple random sorting of species from those predicted for the bioregion and, second, under a model of sorting mediated by the relative productivity of a grid cell. The goals of this second step include (i) an evaluation of the ability of this hierarchical model to make strong fine-scale richness predictions (while including paleoclimate and avoiding regional-level conflation of sample size) and (ii) a demonstration of the separate roles energy availability has at different temporal and spatial scales.
Paleoclimatic data reveal dramatic variation in the age and spatial dynamics of different bioregions from the end of the Paleocene (55 MY bp) to the present day (Figure 1, Table S5). For example, grasslands are not thought to have covered large areas on earth until 8 million years ago, resulting in a much smaller area over time than observed for, for example, temperate or tropical moist forests that have a longer history (Figure 1). Linking estimates of the extents of bioregions over time allows the calculation of “time-integrated area” (TimeArea) [42], a synthetic index of area available to the bioregion's biota over time, varying from just 48×104 km2 integrated over 55 million years in the case of the Mediterranean bioregion at the southern tip of Africa to over 100,000×104 km2 in Eurasian temperate and African moist tropical forests. Unlike bioregion extent and position, climatic conditions of bioregions are assumed to be relatively static over time [49], which allows the determination of average bioregion net primary productivity (Productivity) and Temperature. Summed Productivity over bioregion Area yields total bioregion productivity (AreaProductivity)—that is, total annual carbon flux measured in kg/year over a whole bioregion, a measure that exhibits joint dynamics with bioregion Area. But integrated over time in the form of TimeAreaProductivity, it exhibits very different geographic patterns than TimeArea (Figure 1) with, for example, African and IndoMalay tropical moist forests experiencing a flux of over 8,000×1017 kg of carbon over the past 55 million years and the Mediterranean regions of the New World and Africa just under 3×1017 kg.
We summarized terrestrial vertebrate richness per bioregion as Total (every species found in a bioregion), Resident (species for which a given bioregion contains the largest portion of the range), and Endemic (species that are restricted to a single bioregion; Table S4). We find minimal overlap in Total species among bioregions (median Jaccard similarity among bioregions: 4% for birds, 0% for other taxa; Figure S1, Table S3), which confirms their relative evolutionary isolation in addition to climatic and spatial independence and a consistently strong pattern of biome conservatism [38],[50],[51]. It also confirms that across all four vertebrate groups these selected bioregions represent useful spatial units that avoid the pseudo-replication of species: for Resident species richness every species enters a given analysis exactly once, and the number of distribution records is equal to the global richness of species (13,860 endothermic mammals and birds, 11,836 ectothermic amphibians and reptiles; montane endemics excluded). For Endemic species (total of 13,111 species and records) bioregions are even more likely to represent the true regions of origin compared to Resident species. We therefore expect a stronger correlation of area and productivity integrated over time (TimeAreaProductivity) with the diversity of Endemic species.
All three predictions of our integrative model regarding the effect of time-area-productivity on richness are confirmed (Table 1). The models that account for time-integrated productivity and also include temperature as an additional predictor yield the strongest fits. For endotherms, the time-integrated measures of area outperform models that ignore time only for the Endemic richness dataset, which offered the more direct test of our hypotheses. Predictions of the two-predictor TimeAreaProductivity+Temperature model are consistently strong across all four vertebrate taxa, which represent independent replicates (Figure 2), explaining over 77% of the variation in richness (Figure 2, N = 128, see Tables S7 and S8 for more details). Models that fit TimeArea and Productivity as statistically separate terms do not on the whole yield stronger predictions (Table S9). This lends support to Wright's [45] parallel findings for large islands, which represent similarly closed systems, and contrasts with previous results reported for 110×110 km grid cells [13]. The shape of the Productivity-richness relationship is linear (in Endotherm Residents) or positive accelerating in linear space (in Ectotherm Residents and both Endemics groups). In contrast, the slopes of the AreaProductivity- and TimeAreaProductivity-richness relationships, whether fitted with or without Temperature, are all positive saturating—that is, species richness tends to increase more steeply in the low than in the high productivity ranges (coefficients in ln-ln space vary between 0.4 and 1, Table S7). We did not find evidence of a hump-shaped pattern for any measure of productivity and richness at the bioregional scale [52],[53].
As predicted, ectotherm richness increases much more steeply and strongly with temperature than endotherm richness, both when fitted singly and when controlled for TimeAreaProductivity (Figures S2 and S3, Table S7). This supports at the global scale the significant and complementary effect temperature may contribute to levels of regional ectotherm diversity (see also [4],[14],[44]). For ectotherms, higher temperatures in tropical regions may be promoting higher rates of genetic incompatibilities among populations or faster rates of biotic interactions, further accelerating speciation rates [44],[54],[55]. Alternatively, the thermal dependence of activity represents a strong constraint on ectotherm distribution [56], likely imposing limits on clade origination and diversification in high-latitude regions. Third, in warm regions, ectotherms are released from physiological and behavioral adaptations to cold stress promoting a greater diversity of life histories and metabolic “niches” [57],[58]. These factors are not mutually exclusive, and more work is needed for understanding the potential role of temperature and thermal physiology in driving diversification. Preliminary results from phylogenetic analyses suggest increased diversification rates at lower latitudes in both amphibians [59] and mammals [60], but with a much weaker and more equivocal trend in the latter.
Overall, our bioregion results support the hypothesized interactions of environmental conditions and area over time in influencing the speciation and extinction and ultimately species richness of biota in bioregions. We suggest that the bioregional variation in time-integrated productivity successfully captures key factors affecting both cumulative population sizes over time as well as the different opportunities for reproductive isolation. Large, productive areas like the Neotropical moist/wet forest biome have been characterized by high productivity and a continuously large extent, and thus have supported large populations of each of the four vertebrate clades, since before the Eocene (700×1012 km2 years and 663×1018 kg Carbon produced since 55 MY bp; Figure 1). Reproductive isolation has been facilitated by the large amount of time that vertebrate populations have had to encounter geographical barriers (such as rivers in non-volant mammals [61]) as well as heightened habitat heterogeneity related to the high productivity (i.e., multiple vertical forest strata) [12]. This contrasts with, for example, unproductive North American deserts, which have only come to cover a substantial area within the last few million years (12×1012 km years and 3×1018 kg Carbon; Figure 1) [62]. We suggest that the large TimeAreaProductivity seen in, for example, the Neotropical forest compared to the North American desert bioregion in Figure 1 reflects all factors affecting cumulative population sizes over time (which have affected both speciation and extinction probabilities) as well as opportunities for reproductive isolation. Together, these factors have led to the wide discrepancy in vertebrate diversity between these two bioregions.
Previous studies have employed phylogenies or sister-group comparisons to test whether the latitudinal diversity gradient derives from more evolutionary time [63], niche conservatism [38], or differences in speciation or extinction rates at different latitudes [22],[59],[60],[64]. Factors such as orbital forcing causing glaciation at high latitudes have been posited to elevate extinction rates and are expected to accentuate the observed disparities in species richness among bioregions, especially for endemics [35],[65]. The results reported here complement these studies and suggest that at the bioregion scale, and over an extremely large window of time (55 MY), diversification rates consistently vary with respect to the area, age, and productivity of a given bioregion (Figure 2). We thus view the time-integrated productivity of bioregions to be a general explanation for why so many clades originate at lower latitudes and correspondingly fewer have diversified into bioregions at higher latitudes. It is important to note that time alone is not sufficient to explain these patterns: temperate bioregions are just as ancient as tropical bioregions but strongly differ in their cumulative time-integrated area and productivity. In sum, the strong associations we find indicate a pathway toward first-order approximations of rates of net species production per bioregion, based on variation in area over time, productivity, and temperature. Future studies could integrate our approach with more detailed comparisons of clade-level diversification rates among bioregions or combine it with existing phylogenetic methods for quantifying correlates of diversification.
Having addressed key evolutionary drivers affecting the broad-scale variation in vertebrate diversity, we next assess how each bioregion's species sort into grid cell assemblages and how both processes combine to explain the finer scale geographic variation in richness (Figure 3A). We perform this assessment for the 18,467 bird, mammal, and amphibian species in the bioregion analysis and their 2,966,137 occurrences across the 9,253 110×110 km terrestrial grid cells encompassed by the bioregions (Figure 3B). Strong effects of regional- on fine-scale richness have previously been demonstrated [1],[2], and here we provide a first test of their pervasiveness at a global scale by evaluating the performance of bioregion models for explaining grid cell richness. We find that the two-predictor TimeAreaProductivity+Temperature model developed above (Table 1) alone explains 46%–60% and 32%–50% of the variation in Resident richness and Total richness, respectively (Figure 3B left column, Tables S11 and S12). This highlights how regional effects together with even simple null models of proportional sorting are able to explain much of the finer scale richness patterns. Fine-scale–regional richness relationships are known to be affected by spatial scale as well as by species' dispersal abilities [66]. In larger regions a grid cell of the same size represents a smaller portion of the regional area and, assuming similar levels of grid cell immigration/extinction, grid cell richness is expected to be smaller. This should apply whenever average species range sizes increase less than proportionally with bioregion size and should be particularly noticeable for taxa with relatively low dispersal rates or small within-bioregion range sizes (such as amphibians compared to birds or mammals), because with increasing bioregion size species will be progressively less likely to occupy a given grid cell. We find these expectations confirmed. Bioregion Area exhibits an additional negative effect and improves fine-scale predictions, especially for Total richness. It does so most strongly in Amphibians (Figure S4, Table S12), whose greater dispersal limitation (and on average by a factor of four smaller geographic ranges) compared to mammals or birds has been previously suggested as contributing to their strong patterns of species turnover [67].
Species vary strongly in the number of assemblages they occupy and the species richness of grid cell assemblages is a function of the drivers that affect species' sorting and resulting overlap in geographic ranges. One variable strongly associated with the sorting into assemblages, particularly by wide-ranging species, is local energy availability [8],[25]. We find that relative productivity in a grid cell (CellPropProductivity, i.e., the proportion of the maximum grid cell productivity observed in a bioregion) predicts a substantial additional amount of observed variation in grid cell richness (Figure 3B middle column, and S13) and confirms the expected greater tendency of species within a bioregion to occupy high-productivity grid cells. Allowing the shape of the richness–productivity relationship to vary among regions improves predictions (Tables S12 and S13), but only slightly so, suggesting a within-regional role of productivity that is globally fairly consistent. Nevertheless, the total amount of variation explained by the TimeAreaProductivity+Temperature model (58%–77%) is remarkable and similar to that found in previously published broad-scale gridded richness regression analyses [8],[11]. Notably, however, the hierarchical approach avoids the dual problems of species pseudo-replication and conflation of among- and within-regional processes—issues that have seriously impeded interpretations of all previous gridded biogeographic or macroecological analyses at broad scales.
Our results largely corroborate past studies that have hypothesized that net primary productivity should be a dominant predictor of fine-grain assemblage richness [8],[11],[16]. However, our hierarchical model is able to separate how productivity influences species richness at different temporal and spatial scales. At the bioregional scale, productivity should increase the cumulative population size and opportunities for reproductive isolation over time, promoting higher species richness in high-productivity bioregions [12]. At the fine scale productivity affects the occupancy of assemblages in relation to the regional pool [27],[68]. In addition to the sampling effects inherent with larger assemblage-level population sizes, increased productivity may promote greater richness due to an increased number of niches facilitating species coexistence [12],[25].
We consider the contributions of this study to be conceptual in addition to empirical and hope that its framework will inspire further consideration of diversity gradients that aims to integrate ecological and evolutionary mechanisms across scales. Our global hierarchical approach represents an analytical paradigm shift away from the traditional analysis of fine-scale assemblages as independent spatial units. But there are obvious limits to our analysis. While the strong association of vertebrates with dominant vegetation types and the observed biotic independence of bioregions support their delineation as major evolutionary arenas, challenges remain surrounding the demarcation of the exact boundaries of such regions, the accuracy of past climate reconstructions, and their comparability across clades. Future availability of higher resolution phylogenies of the four vertebrate clades will allow more rigorous comparative approaches within and across lineages, but even comprehensive, strongly supported phylogenetic reconstructions are unlikely to provide vital information regarding the estimation of ancestral distributions (or ranges) and extinction rates [69]. Thus, our model can be viewed as a template on top of which other processes surely influence the origin and maintenance of diversity. For example, glaciation cycles influence speciation and extinction rates [36] and play an important role in driving recent speciation over broad scales [70]. Historical climate dynamics along elevation gradients in particular are known to create opportunities for rapid climate-associated parapatric or allopatric speciation and contribute strongly to the high richness of many tropical mountain areas [71]–[73]. Furthermore, a multitude of trophic interactions are likely to interact with these large-scale processes to cause positive, coevolutionary feedback loops, thus further increasing fine-scale and regional diversity [15].
Our findings show that energy availability has a large effect on both the regional pool and local sorting of richness. This highlights its importance for both evolutionary and ecological processes and the critical need to integrate these effects. This is especially crucial today, given the attention paid to recent models predicting the effects of climate change on the richness of whole gridded assemblages. The redundancy of information and conflation of ecological and evolutionary processes in smaller scale models impede interpretation in a way that is overcome in our analysis. Here we have shown how history can be integrated into a model predicting diversity with area, productivity, and temperature at the global scale. The separate consideration of drivers of diversification and finer scale occupancy and their joint effects on observed gradients of species richness should help pave the way for a more integrated macro-evolutionary and -ecological understanding of the origin and maintenance of global richness gradients.
We selected 32 well-established, geographically and climatically distinct bioregions (Figure 1). These bioregions correspond to the biomes (tundra, desert, grassland, boreal forest, temperate forest, tropical moist/wet forest, tropical dry forest/savanna, and Mediterranean forest/shrublands) within the world's main biogeographic realms (Neartic, Paleartic, Neotropical, Australian, IndoMalayan, and Afrotropics) as described by Olson et al. [46] and also used in the Wildfinder vertebrate distribution database (see below) [74]. Although we do not have detailed, fine-scale records throughout every interval of time for the past 55 million years, enough information exists regarding the age of all biomes and directionality of their expansion and contraction to make reasonable estimates of the measures of their area integrated over time (Table S5). We excluded the “Mangroves” biome (Biome ID 14 in [46],[74]) and also the “Montane Grasslands & Shrublands” Biome (Biome ID 10 in [46],[74]). The latter was not included due to the difficulty in estimating areal and climate changes over their steep gradients over such a long time period. For example, in the Andes, different biomes occur at different elevations on the western and eastern slopes at different latitudes, and the available data are not sufficient to accurately estimate the elevations of the southern, central, and northern Andes at various time intervals since the Miocene, as each chain has uplifted at different rates and at different times [75]. This is critical information to be able to reconstruct the areal extent over time of each bioregion in the Andes and a general problem common to all of the world's mountain ranges, which is why they were excluded from our analysis.
The last 55 million years is an appropriate interval of time to measure the time-integrated area of the world's biomes within realms for two reasons. First, the beginning of the window of time is 10 million years after the massive extinction, which occurred 65 million years ago, causing major upheaval in the vertebrates. By 55 million years ago, the biosphere had recovered but its biota was very different from the plants and animals that had dominated the Cretaceous. Second, most of the “higher taxa”—that is, ancestors of modern lineages of vertebrates that now dominate the extant diversity of mammals, birds, amphibians, and reptiles (for example, fossils recognizable as extant genera)—are already represented in the fossil record by 55 million years ago [76]. Plant communities by the Eocene are, for the first time, composed of Angiosperms and Gymnosperms that are recognizable as the “genera” and “families” that are dominant in today's biomes [62],[77]. Thus, the biota in the Eocene has a “modern aspect” [76],[77].
The Earth's biomes have experienced large changes over the last 55 million years due to the consistent pattern of cooling and drying that has steadily taken place over this period of time [62],[78],[79]. Average global temperatures have plummeted from 27°C 55 million years ago to today's average of 15°C and precipitation has similarly dropped [62],[80]. For the moist/wet forest biomes (boreal forest, temperate forest, and tropical moist/wet forest) we used maps generated by Fine and Ree (2006) that were based on five sources: [49],[62],[81]–[83]. For the other biomes, our approach to estimate the time-integrated area of each biome was first to try to determine the paleobotanical consensus opinion for the age of each biome (Table S5). Then, we took the extant area of that biome and backcasted in time over the years that it has been present, reasoning that as tropical forests have receded during the past 55 MY years, dry and cold biomes such as tundra, desert, Mediterranean, grassland, and dry forest/savanna must have increased in size from the date of their origin to today's area.
We made two interpretations—a “wet” and a “dry” interpretation (Table S5). These two interpretations span the diverse opinions regarding the extent and age of the world's biomes over the last 55 million years and thus gauge the robustness of our results according to a range of expert opinions. For example, desert plants are absent in fossil records until about 2 Ma [77], even though it is hypothesized with molecular dating that plant lineages today found only in desert floras are at least 50 Ma old [62]. Thus, the consensus opinion is that deserts were probably present in the Eocene, but much restricted in size compared to today. For example, evaporite sediments point to extreme aridity in western Africa, Arabia, and central Asia in the late Miocene [82]. We thus made two estimates for the time-integrated area of desert biomes. The “wet” interpretation gives deserts an origin of 34 MYA but covering 10% of their current area from 34 MYA until 2 MYA, which is consistent with the lack of fossil evidence for any desert plant communities. The “dry” interpretation also gives the origin of deserts 34 MYA but has deserts covering the same areal extent as today since their origin, which is almost certainly an overestimate but is possible given the ancient age of some desert plant lineages and the difficulty of fossilization of desert environments (Table S5).
Our wet and dry interpretations both yield qualitatively similar results, and for simplicity, we focus on the “wet” interpretation throughout the article. The current-day extent of a bioregion as given in [46] yielded our predictor variable Area (units km2). Time-integrated area (TimeArea, in units year km2) was given as the integrated areal extent of a bioregion over 55 million years, or simply the sum of the area estimated for each of the 55 one-million-year periods. We acknowledge that this offers only a first order approximation. While exact values will be subject to change as paleoecological knowledge advances, we expect these changes to refine the details rather than radically alter overall patterns, which would have relatively little effect on our analyses, and thus we do not expect systematic biases in our results.
While topographic heterogeneity is expected to also influence the potential for reproductive isolation [32], in this dataset (which excludes montane regions) it is largely captured by bioregion Area and does not yield improved predictions (see Table S12).
We aggregated existing eco-regional terrestrial vertebrate species lists for the selected 32 bioregions from the Wildfinder distribution database [74]. We excluded all eco-regions in biomes not selected for analysis (see above), including all montane eco-regions (which have a total of 1,015 terrestrial vertebrate species restricted to them). This resulted in 54,122 bioregion occurrence records for 25,696 species (9,229 birds, 4,607 amphibians, 4,631 mammals, and 7,229 reptiles). We calculated terrestrial vertebrate richness (“vertebrate richness”) per bioregion in three different ways: Total, which includes every vertebrate species found within each bioregion; Resident, which only counts species in the bioregion with the largest proportion of its geographic range; and Endemic, which counts only species that are restricted to a single bioregion (see Table S2 for complete raw data). Assigning each species only to its dominant bioregion to eliminate pseudo-replication yields a Resident richness pattern very similar to that of Total richness (rS = 0.85, Table S4). For the analyses, vertebrates were divided into ectotherms (amphibians and reptiles) and endotherms (birds and mammals) and further separated into birds, mammals, reptiles, and amphibians. All richness values were natural log-transformed.
Species occurrence data across grid cells were compiled from global expert opinion range maps extracted across a 110×110 km equal area grid in a Behrman projection. For mammals [84], and amphibians, sources were the IUCN assessment (http://www.iucnredlist.org). For birds, breeding distributions were compiled from the best available sources for a given broad geographical region or taxonomic group [85]. For reptiles, global-scale expert range maps have not yet been compiled, and they were therefore not included in the grid cell assemblage analyses. We excluded all cells that were not >50% inside the selected bioregion boundaries as described above (and shown in Figure 1). Only cells with >50% dry land and with at least one species from each of the three vertebrate groups were included in the analysis, resulting in 9,253 cells. For each grid cell we summarized richness of Resident species (i.e., species were counted if they occurred in several grid cells only within the same bioregion) and of Total species (i.e., species were counted whether they occurred in multiple grid cells within the same or in a different bioregion). Values were log10-transformed before analysis. For Total species, the full database consisted of a total of 2,966,137 grid cell records (birds 2,010,091; mammals 695,133; and amphibians 260,913).
Bioregion-typical temperature estimates (Temperature) were based on average annual temperatures calculated from the University of East Anglia's Climatic Research Unit gridded climatology 1961–1990 dataset at native 10-min resolution [86]. For estimates of bioregion-typical annual net primary productivity, we used an average from 17 global models at a spatial resolution of 0.5 degrees latitude-longitude [87]. Average bioregion productivity (Productivity, units grams Carbon m−2 year−1) was calculated from all 0.5×0.5 degree grid cells that predominantly fall inside a bioregion, and summed productivity (AreaProductivity, units grams Carbon year−1) was then given by the product of this value and bioregion Area. With bioregions defined by their typical environmental conditions, we assumed average productivity characteristic of a bioregion to have been constant through time [49],[62]. Time-integrated productivity (TimeAreaProductivity, unit grams Carbon) was thus given as the product of Productivity and TimeArea. Values for all bioregion predictor variables are given in Table S1. All response and predictor variables were natural log-transformed for analysis, except for temperature, which was 1/kT transformed (where k is the Boltzmann constant, see [44]). We used the same global net primary productivity dataset [87] to estimate productivity at the level of 110×110 km grid cells. First, we calculated average grid cell productivity (NPP) across all encompassing 0.5×0.5 degree grid cells. Second, we normalized each grid cell by dividing by the maximum productivity grid cell value observed in a bioregion, resulting in a measure of proportional productivity (PropNPP) varying from 0 to 1.
We performed a total of nine GLM models on the bioregion data and used the Akaike criterion to identify those offering the best fit [88]. Six models were given in the form of single predictors (Temperature, Area, Productivity, AreaProductivity, TimeArea, and TimeAreaProductivity). An additional three models were formed by the combination of the latter three variables with Temperature. We performed a separate set of analyses to assess the potential additional effect of elevation range within a bioregion, but adding this variable to any of the three two-predictor models did not improve model fit, and thus we excluded the variable from further consideration. Because of the strong independence of sampling units both in terms of response (no overlap in species) and predictor variables (by definition each bioregion is environmentally highly distinct from neighboring bioregions), the usual concerns about spatial autocorrelation affecting model results [89],[90] do not apply to this analysis, and additional spatial regression analysis was not performed.
Having established models of bioregion richness, we assessed the success of predictions of resident bioregional richness to explain the species richness (Total and Resident, see above) of all 110×110 km grid cells within bioregions (for a conceptual overview of the analytical steps, see Figure 3). Note that unlike the bioregional tests described above, analyses at this scale do double-count species. In our study we make the simplifying assumption that diversification processes are sufficiently accounted for at the bioregional scale. The models at the within-bioregion scale then address the sorting of these species each into multiple grid cells, with multiple occurrences an integral part of the signal. We acknowledge that, depending on taxon and region, diversification processes may still exert influence on the within-bioregion patterns of distribution and richness, and we hope that our work will spur further research into additional approaches that can be integrated across all scales.
We first evaluated bioregion predicted resident richness alone (in essence testing for a random sorting of bioregion species into finer scale assemblages), then included bioregion Area as an additional predictor, and finally we added estimates of grid cell NPP as a finer scale predictor. We first performed simple GLM models with all 9,253 grid cells as sampling units, together with bioregion Resident richness as predicted by the TimeAreaProductivity+Temperature and AreaProductivity+Temperature models as a first predictor (BioregPred) and bioregion Area as a second predictor (Figure 3, Table S9). In the same GLM we then added grid cell proportional net primary productivity (CellPropNPP, i.e., relative productivity within a bioregion, see above) as an additional predictor. In preliminary post hoc analyses with a number of environmental variables CellPropNPP remained by far the strongest, in line with recent work on within-regional richness filters that also find productivity-related variables to be dominant [26],[27]. Given the nested nature of these analyses we focus on pseudo-r2 values (fit of observed versus predicted) and visual examination of results in the form of partial residual plots (Figure 3). For this first demonstration, focused on a single variable, we did not include further analyses additionally fitting the signal of spatial autocorrelation.
We performed a second set of analyses in an explicit mixed effects model setting (Table S10), with bioregion as a random effect (R library lme4, Version 0.999375-32, function lmer). As in the GLM model, grid cell richness is first fitted by the predictions for regional resident species richness (BioregPred, see Table 1), and then by area of the region (Area), and grid-cell-level NPP (NPP). Region was fitted as a random effect, and the slope and strength of BioregPred and BioregPred+Area as fixed effects were assessed (model formula in R: lmer (y∼BioregPred+Area+(1|Bioregion)). The additional effect of grid cell NPP was then evaluated by fitting it as an additional fixed effect with a globally constant slope (NPPconst) and by allowing the NPP–richness relationship to vary within regions as random slope (NPPvar) (model formula in R: lmer (y∼BioregPred+Area+(1|Bioregion)+(NPP|Bioregion)).
The data are deposited in the Dryad Repository (http://dx.doi.org/10.5061/dryad.45672js4).
|
10.1371/journal.pgen.1000456 | Rapid and Accurate Multiple Testing Correction and Power Estimation for Millions of Correlated Markers | With the development of high-throughput sequencing and genotyping technologies, the number of markers collected in genetic association studies is growing rapidly, increasing the importance of methods for correcting for multiple hypothesis testing. The permutation test is widely considered the gold standard for accurate multiple testing correction, but it is often computationally impractical for these large datasets. Recently, several studies proposed efficient alternative approaches to the permutation test based on the multivariate normal distribution (MVN). However, they cannot accurately correct for multiple testing in genome-wide association studies for two reasons. First, these methods require partitioning of the genome into many disjoint blocks and ignore all correlations between markers from different blocks. Second, the true null distribution of the test statistic often fails to follow the asymptotic distribution at the tails of the distribution. We propose an accurate and efficient method for multiple testing correction in genome-wide association studies—SLIDE. Our method accounts for all correlation within a sliding window and corrects for the departure of the true null distribution of the statistic from the asymptotic distribution. In simulations using the Wellcome Trust Case Control Consortium data, the error rate of SLIDE's corrected p-values is more than 20 times smaller than the error rate of the previous MVN-based methods' corrected p-values, while SLIDE is orders of magnitude faster than the permutation test and other competing methods. We also extend the MVN framework to the problem of estimating the statistical power of an association study with correlated markers and propose an efficient and accurate power estimation method SLIP. SLIP and SLIDE are available at http://slide.cs.ucla.edu.
| In genome-wide association studies, it is important to account for the fact that a large number of genetic variants are tested in order to adequately control for false positives. The simplest way to correct for multiple hypothesis testing is the Bonferroni correction, which multiplies the p-values by the number of markers assuming the markers are independent. Since the markers are correlated due to linkage disequilibrium, this approach leads to a conservative estimate of false positives, thus adversely affecting statistical power. The permutation test is considered the gold standard for accurate multiple testing correction, but is often computationally impractical for large association studies. We propose a method that efficiently and accurately corrects for multiple hypotheses in genome-wide association studies by fully accounting for the local correlation structure between markers. Our method also corrects for the departure of the true distribution of test statistics from the asymptotic distribution, which dramatically improves the accuracy, particularly when many rare variants are included in the tests. Our method shows a near identical accuracy to permutation and shows greater computational efficiency than previously suggested methods. We also provide a method to accurately and efficiently estimate the statistical power of genome-wide association studies.
| Association studies have emerged as a powerful tool for discovering the genetic basis of human diseases [1]–[3]. With the development of sequencing and high-throughput genotyping technologies, the number of single nucleotide polymorphism (SNP) markers genotyped by current association studies is dramatically increasing. The large number of correlated markers brings to the forefront the multiple hypothesis testing correction problem and has motivated much recent activity to address it [4]–[6].
There are two common versions of the multiple testing correction problem: per-marker threshold estimation and p-value correction. In a typical study which collects markers, at each marker, we perform a statistical test and obtain a p-value which we refer to as a pointwise p-value. We would like to know how significant a pointwise p-value needs to be in order to obtain a significant result given that we are observing markers. The per-marker threshold can be defined as the threshold for pointwise p-values which controls the probability of one or more false positives [6]. Similarly, we would like to quantitatively measure the significance of a pointwise p-value taking into account that we are observing markers. For each pointwise p-value, the corrected p-value can be defined as the probability that, under the null hypothesis, a p-value equal to or smaller than the pointwise p-value will be observed at any marker [7]. For example, the Bonferroni correction corrects a pointwise p-value to , or estimates the per-marker threshold as given a significance threshold .
While the Bonferroni (or Šidák) correction provides the simplest way to correct for multiple testing by assuming independence between markers, permutation testing is widely considered the gold standard for accurately correcting for multiple testing [7]. However, permutation is often computationally intensive for large data sets [4]. For example, running 1 million permutations for a dataset of 500,000 SNPs over 5,000 samples takes up to 4 CPU years using widely used software such as PLINK [8] (See Results). On the other hand, the Bonferroni (or Šidák) correction ignores correlation between markers and leads to an overly conservative correction, which is exacerbated as the marker density increases.
In this paper, we correct for multiple testing using the framework of the multivariate normal distribution (MVN). For many widely used statistical tests, the statistics over multiple markers asymptotically follow a MVN [9],[10]. Using this observation, several recent studies [4],[9],[10] proposed efficient alternative approaches to the permutation test, and showed that they are as accurate as the permutation test for small regions at the size of candidate gene studies (with <1% average error in corrected p-values) [4]. However, when applied to genome-wide datasets, they are not as accurate. In our analysis of the Wellcome Trust Case Control Consortium (WTCCC) data [11], these methods eliminate only two-thirds of the error in the corrected p-values relative to the Bonferroni correction. There are two main reasons why these methods do not eliminate all of the error. First, the previous MVN-based methods can be extended to genome-wide analyses only by partitioning the genome into small linkage disequilibrium (LD) blocks and assuming markers in different blocks are independent, because they can handle only up to hundreds of markers in practice [4],[9]. This block-wise strategy leads to conservative estimates because inter-block correlations are ignored (Figure 1B). Second, these methods do not account for the previously unrecognized phenomenon that the true null distribution of a test statistic often fails to follow the asymptotic distribution at the extreme tails of the distribution, even with thousands of samples.
We propose a method for multiple testing correction called SLIDE (a Sliding-window approach for Locally Inter-correlated markers with asymptotic Distribution Errors corrected), which differs from previous methods in two aspects. First, SLIDE uses a sliding-window approach instead of the block-wise strategy. SLIDE approximates the correlation matrix as a band matrix (a matrix with non-zero elements along the diagonal band), which can effectively characterize the overall correlation structure between markers given a sufficiently large bandwidth. Then SLIDE uses a sliding-window Monte-Carlo approach which samples a statistic at each marker by conditioning on the statistics at previous markers within the window, accounting for entire correlation in the band matrix (Figure 1C).
Second, SLIDE takes into account the phenomenon that the true null distribution of a test statistic often fails to follow the asymptotic distribution at the tails of the distribution. It is well known that if the sample size is small, the true distribution and the asymptotic distribution show a discrepancy [12],[13]. However, to the best of our knowledge, the effect of this discrepancy in the context of association studies has not been recognized, since thousands of samples are typically not considered a small sample. We observe that this discrepancy often appears in genome-wide association studies, even with thousands of samples, because of the extremely small genome-wide per-marker threshold (or pointwise p-value). The error caused by this discrepancy is more serious for datasets with a large number of rare variants, highlighting the importance of this problem for association studies based on next-generation sequencing technologies (See Materials and Methods). SLIDE corrects for this error by scaling the asymptotic distribution to fit to the true distribution.
With these two advances, SLIDE is as accurate as the permutation test. In our simulation using the WTCCC dataset [11], the error rate of SLIDE's corrected p-values is more than 20 times smaller than the error rate of previous MVN-based methods' corrected p-values, and 80 times smaller than the error rate of the Bonferroni-corrected p-values. Our simulation using the 2.7 million HapMap SNPs [14] shows that SLIDE is accurate for higher-density marker datasets as well. In contrast, the error rates of previous MVN-based methods increase with the marker density, since the dataset will include more rare variants. Computationally, our simulation shows that SLIDE is orders of magnitude faster than the permutation test and faster than other competing methods.
The MVN framework for multiple testing correction is very general, allowing it to be applied to many different contexts such as quantitative trait mapping or multiple disease models [4]. We show that the MVN framework can also correct for multiple testing for the weighted haplotype test [15],[16] and the test for imputed genotypes based on the posterior probabilities [17].
In addition to multiple testing correction, we extend the MVN framework to solve the problem of estimating the statistical power of an association study with correlated markers. There are two traditional approaches to this problem: a simulation approach constructing case/control panels from the reference dataset [4],[10],[17],[18], which is widely considered the standard but is computationally intensive; and the best-tag Bonferroni method [19]–[21], which is an efficient approximation but is often inaccurate.
The power estimation problem can be solved within the MVN framework because the test statistic under the alternative hypothesis follows a MVN centered at the non-centrality parameters (NCP). The vector of the NCPs turns out to be approximately proportional to the vector of correlation coefficients () between the causal SNP and the markers. This is a multi-marker generalization of the Pritchard and Preworzki [22] single-marker derivation of the NCP proportional to . Our method SLIP (Sliding-window approach for Locally Inter-correlated markers for Power estimation) efficiently estimates a study's power using the MVN framework.
Seaman and Müller-Myhsok [9] and Lin [10] pioneered the use of the MVN for multiple testing correction. Seaman and Müller-Myhsok described the direct simulation approach (DSA) method. Conneely and Boehnke [4] increased its efficiency by adapting an available software package called mvtnorm [23],[24]. Both studies primarily focused on datasets used in candidate gene studies and suggested the block-wise strategy as a possible approach for genome-wide studies.
Another approach for multiple testing correction is to estimate the effective number of tests from eigenvalues of the correlation matrix [25]–[27]. Recently, Moskvina and Schmidt [6] and Pe'er et al. [28] showed that the effective number of tests varies by the p-value levels, demonstrating that a method estimating a constant effective number can be inaccurate. Moskvina and Schmidt [6] proposed a pairwise correlation-based method called Keffective, which estimates the effective number taking into account the significance level. Keffective is a sliding-window approach similar to SLIDE, but it differs because within each window it uses the pairwise correlation to the most correlated marker, while SLIDE uses the conditional distribution given all markers. Fitting the minimum p-value distribution by a beta distribution [29] has been shown often to be inaccurate [6]. Kimmel and Shamir [30] developed an importance sampling procedure called rapid association test (RAT). RAT is efficient for correcting very significant p-values, but requires phased haplotype data.
Connecting the multiple testing correction and power estimation problems leads to the insight that the per-marker threshold estimated from the reference dataset for estimating power can be used as a precomputed approximation to the true per-marker threshold for the collected samples. In simulations using the WTCCC control data, we show that the per-marker threshold estimated from the HapMap CEU population data approximately controls the false positive rate.
Our methods SLIP and SLIDE require only summary statistics such as the correlation between markers within the window size, allele frequencies, and the number of individuals. Therefore unlike the permutation test, our method can still be applied even if the actual genotype data is not accessible. Our methods are available at http://slide.cs.ucla.edu.
Multiple testing correction is generally performed using the collected data and not the reference data. Recall that the difference between the per-marker threshold for multiple testing correction () and the per-marker threshold for power estimation () is that the former is estimated from the collected data, the latter from the reference data. We suggest that multiple testing can be approximately corrected using the reference data, by using as a substitute of . The advantage is that we can obtain an idea of the per-marker threshold even before the samples are collected. In Results, we show the accuracy of this approximation using the HapMap data and the WTCCC data.
We downloaded the HapMap genotype data (release 23a, NCBI build 36) from the HapMap project web site [14],[38] and phased the data into haplotypes using HAP [39], which can handle the trio information. We downloaded the case/control genotype data from the Wellcome Trust Case Control Consortium web site [11] and phased it into haplotypes using Beagle [40].
The URL for methods presented herein is as follows: http://slide.cs.ucla.edu
We compare four different methods for estimating genome-wide power: standard simulation, null/alternative panel construction, best-tag Bonferroni, and SLIP. We assume a multiplicative disease model with a relative risk of 1.2 and a disease prevalence of .01, and a significance threshold of .05. We use the CEU population data in the HapMap as the reference dataset. We use the genome-wide markers in the Affymetrix 500 K chip and assume a uniform distribution of causal SNPs over all common SNPs (MAF≥.05) in the HapMap.
We first perform the standard simulation, which we will consider as the gold standard. We construct a number of genome-wide ‘alternative’ panels from the HapMap data by randomly assigning a causal SNP for each panel. We permute each panel 1,000 times to estimate the panel-specific per-marker threshold. The power is estimated as the proportion of panels showing significance given its per-marker threshold. Conneely and Boehnke [4] used this procedure for power estimation.
Another panel construction-based approach is the null/alternative panel construction method. Instead of permuting each of alternative panels, this method constructs another set of ‘null’ panels under the null hypothesis. The null panel gives us a ‘global’ per-marker threshold that can be applied to all alternative panels. Since this method is as accurate as the standard simulation but is more efficient, it is widely used [17],[18],[21].
We apply SLIP and re-use the samples for the null MVN for estimating the alternative MVNs. Lastly, we apply the analytical best-tag Bonferroni method [19]–[21] which uses the Bonferroni correction for the per-marker threshold and estimates power for each causal SNP by using the most correlated marker (best tag SNP). This method can also be accelerated by sampling the causal SNPs and setting a window size.
For the standard simulation, we use 10 K alternative panels. For the null/alternative panel construction method, we use 10 K alternative panels and 10 k null panels. For SLIP, we use 10 K sampling points. For the best-tag Bonferroni method, we use 10 K samples for causal SNPs. For SLIP, we use a window size of 100 markers. For all other methods, we use a window size of 1 Mb.
Figure 9 shows that both SLIP and the null/alternative panel construction method are as accurate as the standard simulation. The best-tag Bonferroni method is inaccurate, underestimating power by up to 5%.
Table 2 shows the running time of each method for estimating genome-wide power. As shown, SLIP is very efficient. Since SLIP uses the correlation structure, the running time is approximately independent of the study sample size, whereas the running time of the standard simulation or the null/alternative panel construction method is linearly dependent on the sample size.
SLIDE and SLIP provide efficient and accurate multiple testing correction and power estimation in the MVN framework. SLIDE shows a near identical accuracy to the permutation test by using a sliding-window approach to account for local correlations, and by correcting for the error caused by using the asymptotic approximation. SLIDE can be applied to datasets of millions of markers with many rare SNPs, while other MVN-based methods become inaccurate as more rare SNPs are included. To the best of our knowledge, SLIP is the first MVN-based power estimation method.
Throughout this paper, we considered the classical multiple testing correction controlling family-wise error rate (FWER) [7], the probability of observing one or more false positives. SLIDE can be extended to control false discovery rate [41],[42] as well, using a similar approach to Lin [10]. In Text S1, we show that the MVN framework can be extended to the weighted haplotype test [15],[16] and the test for imputed genotypes [17]. SLIDE can be use for any multiple testing correction problem with a local correlation structure, as long as the covariance between statistics can be derived.
We considered the permutation test as the gold standard for multiple testing correction. The permutation test can be performed in two different ways: at each permutation, we can either assess the maximum statistics among the markers (max-T permutation), or assess the minimum pointwise p-value among the markers by performing another permutation for each marker (min-P permutation) [7],[42]. We used the former approach because the latter approach is computationally very intensive.
In Text S5 and Figure S3, we describe some additional insights obtained through the study. When marker frequencies do not follow the Hardy-Weinberg proportions (HWP), the use of an allelic test (e.g. allelic test) for unphased genotype data is not recommended due to the possible bias [43]. However, widely used software [8] often allows the use of an allelic test for genotype data under the reasoning that, as long as the permutation or an exact test is performed, the pointwise p-value will be the same as if we use a genotypic test (e.g. Armitage's trend test). Theoretically, this is due to the fact that the allelic and genotypic test statistics differ only by their variance [44]. However, for assessing corrected p-values, the permutation test does not provide this kind of “protection”. Even after a quality control process that excludes SNPs which significantly deviate from the Hardy-Weinberg equilibrium (HWE), still many SNPs may not follow HWP. Therefore, using an allelic test for genotype data for multiple testing correction can result in inaccurate estimates.
Recently, a different view of multiple testing correction has been introduced [5],[28], which suggest that we should correct for the uncollected or unknown markers as well as the collected markers, in order to take into account additional testing burdens such as the possible testings in a follow-up study. Pe'er et al. [28] estimates the per-marker threshold by extrapolating from the resequenced ENCODE regions, and Dudbridge et al. [5] estimates the per-marker threshold by subsampling the SNPs at an increasing SNP density. Although we employed the classical point of view that corrects for multiple testing only over observed SNPs, our method can also be applied to this alternative view. Our method can be used to estimate the effective number of tests for a representative resequenced region or for the set of subsampled SNPs. Since the SNP density of genotyping technology is dramatically increasing, we assume that the number of unknown and uncollected SNPs will decrease, causing the two different views to converge.
In our experiments, we used a constant block size for the block-wise strategy. In practice, it will be more reasonable to split the region according to the LD blocks. However, this is not always possible because LD blocks are often ambiguous and some blocks can be larger than the maximum block size of the method. For example, if we collect 10 million SNPs, a block size of 1,000 is required to cover 300 kb LD. However, the maximum block size of mvtnorm that allows an accurate estimate is currently 300 [4], and DSA with window size 1,000 often requires a prohibitively large memory in our simulations (data not shown). By contrast, SLIDE with window size 1,000 for the WTCCC chromosome 22 data requires ∼150 Mb memory and thus is feasible. Nevertheless, it should be noted that the block-wise strategy can always be implemented to have the same block size as SLIDE.
Recently, a method called PRESTO [45] was introduced, which increases the efficiency of the permutation test by applying several optimization techniques. Based on the claimed running time, SLIDE is ∼10 times faster than PRESTO, but PRESTO has an advantage that it does not depend on the asymptotic approximation but provides exactly the same result as the permutation test.
We considered the pairwise correlation between SNPs. There can also be so-called higher-order correlations, such as the correlation between a haplotype and a SNP. For example, even though three SNPs are pairwisely independent, the combination of the first two SNPs can be a perfect proxy to the third SNP. However, the multivariate central limit theorem proves that the joint distribution of the test statistics is fully characterized by the matrix of the pairwise correlations. Thus, the effect of the other correlation terms on the joint distribution is asymptotically negligible. Nevertheless, our method is not limited to the SNP test. If our method is applied to the weighted haplotype test [15],[16] as shown in Text S1, the pairwise correlation in the correlation matrix can be interpreted as the higher-order correlations between a haplotype and a SNP or between haplotypes.
In summary, SLIP and SLIDE are two useful methods for genome-wide association studies which provide accurate power estimation at the design step and accurate multiple testing correction at the analysis step. The software is available as a resource for the research community.
|
10.1371/journal.pgen.1004738 | POT1a and Components of CST Engage Telomerase and Regulate Its Activity in Arabidopsis | Protection of Telomeres 1 (POT1) is a conserved nucleic acid binding protein implicated in both telomere replication and chromosome end protection. We previously showed that Arabidopsis thaliana POT1a associates with the TER1 telomerase RNP, and is required for telomere length maintenance in vivo. Here we further dissect the function of POT1a and explore its interplay with the CST (CTC1/STN1/TEN1) telomere complex. Analysis of pot1a null mutants revealed that POT1a is not required for telomerase recruitment to telomeres, but is required for telomerase to maintain telomere tracts. We show that POT1a stimulates the synthesis of long telomere repeat arrays by telomerase, likely by enhancing repeat addition processivity. We demonstrate that POT1a binds STN1 and CTC1 in vitro, and further STN1 and CTC1, like POT1a, associate with enzymatically active telomerase in vivo. Unexpectedly, the in vitro interaction of STN1 with TEN1 and POT1a was mutually exclusive, indicating that POT1a and TEN1 may compete for the same binding site on STN1 in vivo. Finally, unlike CTC1 and STN1, TEN1 was not associated with active telomerase in vivo, consistent with our previous data showing that TEN1 negatively regulates telomerase enzyme activity. Altogether, our data support a two-state model in which POT1a promotes an extendable telomere state via contacts with the telomerase RNP as well as STN1 and CTC1, while TEN1 opposes these functions.
| Telomeres are required to stabilize the ends of linear chromosomes, and thus ensure genome integrity. Telomeric DNA is maintained though the action of both conventional and non-conventional DNA replication mechanisms. To ensure that chromosome ends are fully protected and fully replicated, telomeres dynamically oscillate between a closed (non-extendable) and an open (extendable) conformation throughout the cell cycle. The telomerase reverse transcriptase engages telomeres when they are in an extendable conformation. How this conversion occurs, how telomerase is recruited to the chromosome terminus and how telomerase action is terminated are unanswered questions. Here we provide evidence that POT1a, a telomerase accessory protein from the flowering plant Arabidopsis, helps to convert the telomere into a telomerase-extendable state through dynamic interactions with a critical telomere binding protein complex, and through stimulation of telomerase enzyme activity. The results of this study provide new insight into the regulation of telomeric DNA replication.
| Eukaryotes face end-protection and end-replication problems due to the linear nature of their chromosomes and the limitations of conventional DNA replication. Telomerase averts these crises using its RNA subunit (TER) as a template to reiteratively synthesize G-rich repeat sequences on the 3′ single-strand extension (G-overhang) of the chromosome terminus. Both the single (ss) and double-strand (ds) portions of the telomere are host to protein complexes that modulate telomerase action and distinguish natural chromosome ends from double-strand breaks [1]–[4].
Telomeres vacillate between a telomerase extendable and a telomerase un-extendable state during the cell cycle [5], [6]. In G1, the G-overhang is sequestered, preventing the DNA terminus from eliciting a damage response, but also preventing telomerase access. In late S/G2 phase, telomerase is recruited to chromosome ends for DNA synthesis. Once telomerase extends the G-rich strand, the C-strand is replicated by DNA Polymerase α/primase [7], [8], followed by terminal DNA processing to create the 3′ G-overhang [9]. The terminus is then sequestered once again. These reactions are highly coordinated, and driven by the exchange of large replication/processing complexes on the G-overhang.
One telomere complex under intensive scrutiny is CST (Cdc13/CTC1, Stn1, Ten1), an RPA-like heterotrimer [10], [11] first identified in budding yeast. Cdc13 anchors CST to ss telomeric DNA via its central oligosaccharide-oligonucleotide binding domain (OB-fold) [12]. Genetic analysis of separation-of-function alleles reveals that Cdc13 maintains genome integrity and regulates telomere maintenance [13], [14]. Stn1 and Ten1 are also essential for telomere integrity, and their association with Cdc13 renders telomeres into an un-extendable state [15]–[17]. However, the CST heterotrimer is not static, and recent data show that Stn1 and Ten1 make contributions distinct from Cdc13 [18]. In addition, phosphorylation of Cdc13 in late S phase shifts the binding preference from Stn1 and Ten1 to the telomerase accessory factor Est1 [19], [20], converting the telomere into an extendable conformation. Est1 is a multifunctional protein that directly binds the TER subunit (Tlc1) as well as Cdc13. This interaction recruits telomerase to the chromosome end [21]–[24]. Consistent with its critical role in telomere maintenance, Est1 deletion causes progressive telomere shortening [25]. Est1 also stimulates the activity of telomerase on telomeric DNA [23], [26] likely through contacts with Cdc13 [27].
Mammalian telomeres are protected by an alternative complex termed shelterin. The six shelterin subunits include TRF1, TRF2, and RAP1, which are tethered to ds telomeric DNA and are bridged by TIN2 and TPP1 to the ss DNA binding protein POT1 [1], [28]. All shelterin components are critical for genome stability, and like budding yeast CST, may shift between sub-complexes during the cell cycle [29]. POT1 inhibits telomerase elongation in vitro by preventing substrate access [30], [31]. In contrast, the POT1-TPP1 heterodimer stimulates telomerase repeat addition processivity (RAP) by promoting substrate association and template translocation during telomerase extension [32]–[34]. In addition, TPP1 appears to directly contact the telomerase catalytic subunit TERT and thereby recruits telomerase to telomeres [35]–[37].
CST also exists in vertebrates and plants, although Cdc13 has been replaced by another large OB-fold containing protein, CTC1 [38]–[41]. In contrast to yeast where CST functions in both end protection and telomeric DNA replication [4], vertebrate CST primarily serves to promote telomere replication by stimulating C-strand fill-in and genome-wide replication rescue [42]–[45]. CTC1 and STN1 directly contact the telomerase activator proteins TPP1/POT1 [32], [46], [47]. Recent studies indicate that human CST negatively regulates telomerase by competing with TPP1/POT1 for telomeric DNA binding and by squelching the stimulation of telomerase RAP by TPP1/POT1 [46]. Thus, the interaction of TPP1/POT1 with CST is proposed to terminate G-strand synthesis by telomerase. While the molecular basis for the dynamic exchange between shelterin, telomerase and CST is unknown, shifting interactions between shelterin constituents [48], [49] prompted through posttranslational modification [20], [37], [50], [51] likely control telomere transactions.
Arabidopsis telomeres represent an intriguing blend of features from yeast and vertebrates. Only a subset of shelterin components can be discerned in plants, and although the Arabidopsis CST complex is structurally analogous to mammalian CST, it appears to play a role in chromosome end protection. Loss of any of the Arabidopsis CST subunits elicit dramatic telomere shortening, increased ss telomeric DNA, and chromosomal fusions [38], [39], [41], culminating in stem cell failure [52]. Notably, TEN1 is detected at a significantly smaller fraction of telomeres than CTC1 [39], [41]. In addition, unlike plants lacking STN1 or CTC1, ten1 mutants have higher levels of telomerase enzyme activity overall, and generate longer telomere repeat arrays in vitro, indicating that TEN1 negatively regulates telomerase activity [41].
Arabidopsis harbors two TER genes encoding RNAs that assemble into different RNP complexes with opposing functions. TER1 is a canonical TER subunit required for telomere maintenance, whereas TER2 negatively regulates telomere synthesis by the TER1 RNP in response to DNA damage [53], [54]. Arabidopsis encodes several telomerase accessory factors, but notably the two Est1-like proteins play no obvious role in telomere maintenance and rather are implicated in the regulation of the meiotic cell cycle [55]. POT1a, one of three A. thaliana POT1 paralogs [56]–[58] exhibits properties reminiscent of Est1. POT1a associates with TER1, and localizes to telomeres in S phase [59]. Moreover, plants lacking POT1a are defective in telomere maintenance, and undergo progressive telomere shortening. In addition, pot1a mutants have reduced telomerase activity in vitro [59]. These findings indicate that POT1a positively regulates telomerase enzyme activity and promotes telomere repeat synthesis on chromosome ends.
In this study, we further explore the role of POT1a. We report that POT1a is not required to recruit telomerase to telomeres, but is required for telomerase to maintain telomere tracts. Our biochemical data indicate that POT1a stimulates telomerase enzyme activity, likely by enhancing its RAP. We further show that POT1a directly contacts STN1 and CTC1 in vitro, and its association with STN1 is mutually exclusive of TEN1-STN1 binding. Finally, we demonstrate that CTC1 and STN1, but not TEN1, interact with enzymatically active telomerase in vivo. These findings suggest a model in which POT1a promotes telomere maintenance by activating telomerase at chromosome ends. The data further suggest that the opposing functions of POT1a and TEN1 in telomerase regulation may contribute to the switch from telomerase extendable to the telomerase un-extendable state.
Chromatin immunoprecipitation (ChIP) was used to investigate whether POT1a is needed for telomerase association with telomeres. As expected, the telomerase catalytic subunit TERT [60] could be detected at telomeres in rapidly dividing young wild type seedlings (Fig. 1A). However, there was no significant difference in the level of telomere-bound TERT in pot1a mutants versus wild type (Fig. 1A and C). One possible explanation is that the TERT signal includes telomere-bound TER2 RNP. Since POT1a does not interact with TER2 [54], loss of this protein is not expected to perturb the alternative telomerase RNP. To address this possibility, we generated plants doubly deficient in POT1a and TER2. ChIP assays performed on pot1a ter2 mutants showed the same level of telomere-bound TERT as in wild type plants (Fig. 1A and C). We conclude POT1a is not required for TERT recruitment to telomeres.
If POT1a is not required for telomerase's association with chromosome ends, how does it promote telomere maintenance? One possibility is that POT1a directly modulates telomerase enzyme activity. The conventional telomere repeat amplification protocol (TRAP) assay shows an ∼13 fold decrease in telomerase activity in pot1a relative to wild type extracts [59]. This change in enzyme activity is not due to altered expression of TERT and TER1 transcripts or genes previously shown to inhibit telomerase activity such as TER2 and TEN1 (Fig. S1). Attempts to develop a direct primer extension assay in Arabidopsis have been unsuccessful thus far. To obtain a more accurate gauge of the distribution and quantity of the products of Arabidopsis telomerase, we used a modified version of the TRAP assay, telomerase processivity TRAP (TP-TRAP), developed to provide an indication of mammalian telomerase RAP [41], [61]. Pilot reactions with an oligonucleotide bearing five telomere repeats yielded a discrete band of the expected size (Fig. S2), indicating that the PCR amplification step of TP-TRAP gives a reliable assessment of the length of a telomere repeat array generated in the PCR reaction.
TP-TRAP performed with wild type Arabidopsis extract generated a broad distribution of elongation products, including high molecular weight species corresponding to the addition of at least 15 TTTAGGG repeats (Fig. 2A and C). As expected, extract from ten1 mutants, but not stn1 or ctc1 mutants, generated slightly longer products than wild type (Fig. 2A and S3) [41], supporting the conclusion TEN1 negatively regulates telomerase activity and further that this is a unique property of this CST subunit. The TP-TRAP results for pot1a mutants were markedly different and showed a dramatic reduction in high molecular weight products relative to wild type (Fig. 2C). While standard TRAP assays show a general decrease in telomerase activity in pot1a mutants (Fig. 2B), the TP-TRAP indicated that the defect lies in the production of long arrays of telomere repeats (Fig. 2C). The primer is in vast excess over telomerase in TP-TRAP reactions as in conventional TRAP and the direct primer extension assay. Consequently, long products are unlikely to be generated by telomerase dissociation and rebinding the same primer molecule. The data are consistent with the notion that POT1a stimulates RAP.
To determine if the decreased telomerase activity associated with pot1a mutants is specific to the TER1 RNP complex, we performed TP-TRAP on ter2 seedling extracts. The product profiles were nearly identical to wild type (Fig. 3A), indicating the TER1 RNP efficiently synthesizes telomeric DNA in wild type plants. We confirmed that POT1a modulates the TER1 RNP by analyzing pot1a ter2 mutants. Long products were reduced in the double mutants, but not to the same extent as pot1a (Fig. 3A). In agreement with previous results showing that TER2 negatively regulates TER1 RNP [54], quantitative TRAP (qTRAP) revealed a higher level of telomerase activity in ter2 mutants relative to wild type (Fig. 3B), which could explain why the TP-TRAP and qTRAP signal is higher in pot1a ter2 than pot1a (Fig. 3A and B). Since the TER1 RNP is the only functional telomerase complex in pot1a ter2 mutants, the data indicate POT1a distinctly modulates this complex.
In both yeast and vertebrates, CST plays a key role in controlling G-overhang access to telomerase and DNA Pol-α [4], [7], [46]. To test whether telomerase acts in concert with CST for telomere maintenance, we used a genetic approach. As expected, ctc1 and stn1 mutants exhibited severe morphological aberrancies including irregular phyllotaxy, fasciated stems, and reduced fertility (Fig. 4A and C, and S4A; [38], [39]). These phenotypes were even more pronounced when telomerase was inactivated in stn1 and ctc1 plants (Fig. 4A and S4A). Progeny lacking CTC1/STN1 and TERT were rarely recovered, and when they were, double mutants arrested in a dwarf vegetative state without producing germline tissue (Fig. 4A and Fig. S4A). Telomere length was examined using Terminal Restriction Fragment (TRF) analysis or Primer Extension Telomere Repeat Amplification (PETRA) when insufficient material was available for TRF. Consistent with previous studies [38], [39], stn1 or ctc1 mutants displayed shorter, more heterogeneous telomere tracts than wild type plants. In contrast, while telomeres in tert mutants consisted of a discrete, homogeneous population of bands shorter than wild type (Fig. 4B and Fig. S4B) [62]. The telomeres of plants lacking either CTC1 or STN1 and telomerase were dramatically shorter with some telomeres dipping below the critical threshold of 1 kb (Fig. 4B and Fig. S4B), which triggers telomere fusions [63]. We conclude telomerase is capable of extending telomeres devoid of CTC1 or STN1 to partially alleviate their dysfunction. However, given the very severe telomere deprotection phenotype associated with the loss of CST, these epistasis experiments do not rule out the possibility that STN1 or CTC1 engage telomerase and modulate its activity in vivo.
To determine if POT1a is required for telomerase to mitigate telomere defects in STN1/CTC1 deficient plants, we evaluated pot1a ctc1 and pot1a stn1 double mutants. We were unable to recover viable pot1a ctc1 mutants. However, stn1 pot1a mutants exhibited similar morphological defects as stn1 tert plants (Fig. 4C). In addition, molecular analysis revealed the same type of telomere aberrations (Fig. 4D). Thus, the absence of POT1a renders stn1 mutants incapable of employing telomerase as a recovery mechanism (Fig. 4B). These findings support the conclusion that POT1a is required to activate telomerase at chromosome ends.
Recent studies show that human POT1 and mouse POT1b bind CTC1 and STN1 [46], [47], [64]. Additional contacts between TPP1 and CTC1 and TPP1 and STN1 have been observed [46], [64], [65]. Therefore, we asked if POT1a binds individual CST subunits in vitro via co-immunoprecipitation assays using rabbit reticulocyte lysate (RRL) expressed proteins. We were unable to express intact full length CTC1, and so we employed an amino-terminal deletion construct (CTC1ΔN) that was sufficient to bind STN1 and the DNA Pol α subunit, ICU2 [4], [39]. POT1a was tagged with T7 on its amino terminus and immunoprecipitation (IP) was performed using T7 antibody-conjugated agarose beads. Binding was assessed by the ability of POT1a to co-precipitate 35S-methionine labeled CTC1ΔN, STN1, or TEN1. We detected POT1a binding to CTC1ΔN and STN1, but no interaction between TEN1 and POT1a was observed (Fig. 5A).
Since TEN1 and STN1 form a heterodimer, we considered the possibility that POT1a might compete with TEN1 for STN1 binding. We first tested if STN1 can simultaneously bind POT1a and TEN1. TEN1 was T7 tagged, and incubated with labeled STN1 (Fig. 5B, lane 4), POT1a (Fig. 5B, lane 6) or both proteins (Fig. 5B, lane 2) followed by IP. In the reaction containing STN1 and POT1a, only STN1 was detected in the TEN1 IP (Fig. 5B, lane 2). Because TEN1 does not bind POT1a (Fig. 5A and Fig. 5B, lane 6), this result argues that STN1 binding to TEN1 and POT1a is mutually exclusive.
Next, we asked whether POT1a can compete with TEN1 for STN1 binding in vitro. We expressed and purified E. coli TEN1 protein as well as the first OB-fold of POT1a (POT1a OB1), which is sufficient for POT1a-STN1 interaction in vitro (Fig. S5 and Fig. 5C, lane 5). A competition assay was performed by incubating TEN1 with RRL-expressed [35S]-methionine labeled STN1 in the presence of increasing amounts of POT1a OB1. Following TEN1 IP, E. coli-expressed proteins (TEN1 and POT1a OB1) were monitored by coomassie stain (Fig. 5C top) and STN1 by autoradiography (Fig. 5C bottom). As expected, TEN1 pulled down STN1 (Fig 5C, lane 6). At an equal molar ratio of POT1a OB1 to TEN1, the TEN1-STN1 interaction persisted (Fig. 5C, lane 8). However, a ten-fold excess of POT1a OB1 significantly reduced STN1 in the TEN1 IP (Fig. 5C, lane 9). In contrast, 50-fold excess bovine serum albumin did not dislodge STN1 from TEN1 (Fig. 5C lane 7). Because E. coli POT1a OB1 directly binds STN1 (Fig. 5C, lane 5), these data support the conclusion that STN1 binding to POT1a and TEN1 is mutually exclusive. However, because excess POT1a OB1 is required to disrupt the STN1-TEN1 interaction, the data indicate that STN1 has a higher affinity for TEN1 than POT1a OB1.
The discovery of in vitro interactions between POT1a with STN1 and CTC1 raised the possibility that these CST components associate with enzymatically active telomerase in vivo (Fig. 6). To test this idea, we generated a STN1 antibody that could be used for IP-TRAP. Western blot analysis confirmed that the antibody specifically recognizes STN1 (Fig. 6B). IP-TRAP using TERT antibody as a control revealed abundant telomerase activity (Fig. 6A). Strikingly, IP-TRAP with STN1 antibody gave a similar result. Western blot analysis verified that STN1 was precipitated in the reaction (Fig. 6B). Telomerase activity was not detected in an IP with pre-immune sera and was removed by RNaseA treatment, indicating that the STN1 interaction with telomerase was specific. Importantly, STN1 protein was present in the TERT IP (Fig. 6B), confirming the association of these molecules in vivo. IP of a transgenic CTC1-CFP protein also pulled down active telomerase as well as POT1a (Fig. S6). These findings indicate that both STN1 and CTC1 are associated with enzymatically active telomerase in vivo.
We asked if POT1a was essential for the STN1-telomerase interaction by repeating the STN1 IP-TRAP experiment in a pot1a mutant. Telomerase activity and TERT were detected in the STN1 IP of pot1a extracts (Fig. 6A and B). As expected, telomerase activity was visibly decreased in this background ([59]; Fig. 6B). These data indicate that telomerase can associate with STN1 in the absence of POT1a. The data also support the conclusion that POT1a is not necessary for telomerase localization to telomeres, but is required to promote the full activation of telomerase.
Finally, we performed IP-TRAP with our TEN1 antibody to test if TEN1 is associated with active telomerase. In marked contrast to STN1 and CTC1, telomerase activity was not observed in the TEN1 pull down (Fig. 6C). Moreover, TEN1 protein could not be detected in the TERT IP (Fig. 6D). We conclude that TEN1 is not associated with enzymatically active telomerase in vivo, consistent with its role as a negative regulator of telomerase activity.
Telomere accessibility to telomerase is tightly regulated during the cell cycle. Whereas aspects of telomerase recruitment are similar in yeast and vertebrates, many questions remain unanswered, in part because the specific proteins that mediate these interactions are not well conserved [29]. In this study, we investigated how the interplay between POT1a and CST in Arabidopsis promotes telomere maintenance. Like the budding yeast recruitment factor Est1 [21], [22], [25], [27], POT1a directly contacts the canonical TER, TER1 [53], and is required for robust telomerase activity in vitro and telomere maintenance in vivo [59]. However, unlike Est1 [66], we found that POT1a is not necessary for the telomere localization of TERT. The TERT interaction with telomeres was also unperturbed in plants doubly deficient in POT1a and TER2, indicating TERT is not tethered to telomeres through the TER2 RNP. How telomerase is recruited to chromosome ends in the absence of POT1a is unclear. In yeast, Ku provides an alternative route for telomerase recruitment in G1 [66]. However, Ku inhibits telomere synthesis in plants [67], [68], and thus this mechanism is not used to dock telomerase at Arabidopsis telomeres. The TRF-like protein AtTRB1 was recently shown to interact with telomeres and to contact TERT, suggesting that it might be involved in telomerase recruitment [69]. Another potential telomerase recruitment factor is HOT1, which stimulates telomerase recruitment in mammals through contacts with telomeric DNA and the telomerase RNP independent of shelterin [70]. Notably, Arabidopsis has a putative HOT1 ortholog, but lacks several of the core shelterin components, including TPP1, which is implicated in recruiting vertebrate telomerase [33], [36].
Although POT1a is not required for telomerase recruitment, it is required for the enzyme to extend telomere tracts in vivo ([59]; this study). Our data indicate POT1a directly stimulates telomerase catalysis. Using a modified version of the TRAP assay to gauge the length of telomerase products, we discovered that POT1a is necessary for the synthesis of long telomere repeat arrays. An attractive model is that POT1a promotes telomerase RAP, as shown for other telomerase-associated OB-fold bearing proteins such as human TPP1 and Tetrahymena Teb1 [32], [35], [71]. However, in the absence of a direct primer extension assay for Arabidopsis telomerase, we cannot exclude the possibility that POT1a affects some other parameter of telomerase enzymology (e.g. nucleotide addition processivity, nucleotide binding affinity or affinity for the DNA primer).
Once telomerase binds the telomere, how is its activity controlled? CST has a central role to play in this regard, but precisely how it interfaces with telomerase and whether this association stimulates or represses telomerase differs in yeast and vertebrates. Our analysis indicates that CST is not required to recruit Arabidopsis telomerase to chromosome ends. We found that telomerase can act on telomeres lacking CTC1 or STN1, partially alleviating the telomere dysfunction and the aberrant morphological defects associated with these mutations. Importantly, telomere extension in CTC1 and STN1 deficient plants is dependent upon POT1a, supporting the conclusion that POT1a is required to promote telomere maintenance.
In mammals, CST interaction with POT1 orthologs is linked to telomerase termination [46] and G-overhang maturation [47]. In contrast, we find that STN1 and CTC1 like POT1a are associated with enzymatically active telomerase in Arabidopsis [59]. Our experiments do not distinguish whether these telomerase interactions occur on or off the telomere. Nevertheless, since CTC1 can be detected at Arabidopsis telomeres even in cells arrested in G1 (Surovtseva et al 2009), we postulate that telomerase associates with CTC1 and STN1 on the G-overhang during S phase to facilitate telomere repeat incorporation (see below).
We found a direct interaction between POT1a with both STN1 and CTC1, but not TEN1 in vitro. Our data indicate that STN1 interaction with POT1a and TEN1 is mutually exclusive. FurthermoreTEN1 unlike STN1 and POT1a is not associated with active telomerase in vivo. These observations are consistent with a role for TEN1 in negative regulation of telomerase enzyme activity [41]. Intriguingly, TEN1 may only transiently associate with Arabidopsis telomeres. CTC1 can be detected at ∼50% of the Arabidopsis chromosome ends [39]. Since only half of the Arabidopsis telomeres carry G-overhangs [72], essentially all of the G-overhangs are bound by CTC1. In contrast, TEN1 can only be detected at 11% of the telomeres [41], implying that it dynamically binds telomeres and does not function exclusively in the context of a trimeric CST complex.
Altogether, our data suggest a model in which POT1a facilitates telomere maintenance in two ways: by promoting the switch from the un-extendable to the extendable state and by stimulating telomerase enzyme activity (Fig. 7). In S phase, telomerase holoenzyme is recruited to the G-overhang through an unknown mechanism. The enzyme associates with CTC1 and STN1 through contacts with POT1a, and POT1a stimulates G-strand synthesis. One attractive hypothesis is that mobilization of POT1a to the chromosome terminus triggers the exchange of the telomerase negative regulator TEN1 from STN1 as part of the switch to the telomerase extendable state. Although our in vitro data indicate that STN1 has a higher affinity for TEN1 than POT1a OB1, additional contacts by other regions of POT1a or between POT1a and CTC1 may stabilize its interaction with STN1. Furthermore, shifting telomerase-CST interactions are likely to be governed by cell cycle specific posttranslational modifications such as those described for yeast Est1 and CST, as well as human TPP1 [19], [20], [37]. Once the G-strand is extended, telomerase action is terminated, perhaps with the assistance of TEN1. This clears the way for conventional replication machinery and processing enzymes to complete telomere replication and return the telomere to its fully protected un-extendable state. Although additional studies are needed to precisely delineate the telomere-telomerase interface and its control during telomere replication, our findings underscore the highly dynamic nature of telomerase-telomere transactions and suggest that modulation of telomerase enzyme activity at the chromosome terminus contributes to the bimodal switch in telomere states.
Plants were housed in growth chambers with a 16 hr photoperiod at 22°C. stn1-1, ctc1-1, tert, pot1a-1 and ten1-3 mutants were used for crosses and genotyped as described [38], [39], [41], [59]. pot1a ter2 crosses were generated from homozygous parents. F1 progeny was planted for selection by genotyping. F3 seedlings were used for ChIP assays and pTRAP. In vivo experiments examining telomerase activity, protein interactions, or gene expression were either performed in juvenile seedlings or flowers, which both exhibit high levels of telomerase activity. For telomere length analysis, wild type controls were segregated from heterozygous parents to ensure that changes reflect mutations in the target genes and not stochastic variation [73].
Approximately 4–6 grams of Arabidopsis seven day-old seedlings were used for each genotype. The protocol was adapted from [74] with minor changes. Sonication was performed on ice after crosslinking and nuclei extraction using (Fisher Scientific) with 4 cycles of 15 sec on and 1 min off per sample at 40% amplification. Immunoprecipitation (IP) was performed using rabbit anti-TERT antibody and Protein-A agarose/salmon sperm DNA beads (Millipore). Eluted DNA was subjected to Southern dot blotting using a telomeric [32P] 5′ end-labeled oligonucleotide probe. Stripping and rDNA hybridization performed as previously described using a combination of 5S (5′-TTGCAGAATCCCGTGAACCATCGAGT-3′) and 18S (5′-TGGAGCCTGCGGCTTAATTTGACTCA-3′) rDNA oligo-probes [59]. Quantification was performed on at least three independent biological replicates using Quantity One software (Bio-Rad).
Constructs for E. coli expression of TEN1 and POT1a OB1 were cloned in pET28a vector (Novagen). The POT1a OB1 domain was cloned from the POT1a start codon to residue 158. Four amino acids (SISS) were added to the C-terminus to increase protein solubility. Affinity column purification was achieved using Ni-NTA agarose resin (Qiagen) from BL21 DE3 lysates. Protein was eluted in imidazole buffer and dialyzed overnight. POT1a OB1 was further purified using a Sephadex G-75 (GE Healthcare) size exclusion column. TEN1 and POT1a OB1 protein fractions were analyzed for homogeneity on coomassie stained SDS-PAGE gels and verified by mass spectrometry. Proteins were expressed in rabbit reticulocyte lysate (RRL) (Promega) as indicated according to the manufacturer's instructions with [S35] Met (Perkin-Elmer) to label the protein expressed from pCITE4a, and in some cases pET28a.
POT1a, STN1, TEN1, and CTC1ΔN cDNA were cloned into pET28a (T7-tag fusion) and pCITE4a vectors (Novagen). Details for POT1a OB1, OB1+2, and C-terminus constructs are previously described [53]. Co-IP with the RRL-expressed proteins was performed as described [75]. Competition assays were performed by incubating E. coli TEN1 protein with RRL-expressed STN1, and various amounts of E. coli POT1a OB1 or BSA. Equal loading for STN1 was achieved by evenly dividing a single master mix of RRL-expressed protein among the samples. Pull downs were performed by IP of TEN1 using purified TEN1 antibody [41] and protein-A agarose beads (Pierce). Complexes were washed 10× with buffer W300 [75] and eluted by boiling for 5 min in SDS loading dye. Samples were resolved on 12% SDS-PAGE gels followed by coomassie staining and then dried for analysis by autoradiography.
Extracts from ∼5 grams of wild type and pot1a seedling tissue were prepared as previously described [76] and pre-cleared using protein-A agarose beads (Pierce) with gentle rocking at 4°C for 1 h. IP was performed by adding 15 µg of affinity purified TERT, STN1, TEN1 or anti-GFP (Abcam) antibody (or pre-immune sera) overnight with gentle rocking at 4°C. Anti-rabbit STN1 antibody was raised from E. coli expressed and purified MBP-STN1 antigen. Protein-A agarose beads were added the following day for 2 hrs followed by 5× washes with buffer W300 [75], and 2× washes with buffer TMG [75]. IP samples were left in a final 50∶50 slurry in buffer TMG.
DNA from whole plants was extracted as described [77]. TRF analysis was performed using 50 µg of DNA digested with Tru1I (Fermentas) and hybridized with a [P32] 5′ end–labeled (TTTAGGG)4 probe [76]. Blots were developed using a Pharos FX Plus Molecular Imager (Bio-Rad) and data were analyzed with Quantity One software (Bio-Rad). Primer extension telomere repeat amplification (PETRA) was performed as described [63]. 2 µg of DNA was used per reaction for telomere extension, followed by PCR amplification. PETRA products were separated on an agarose gel and subjected to Southern blotting using the same telomeric probe mentioned above.
Protein for Telomere Repeat Amplification Protocol (TRAP) assays were extracted from 5 day-old seedlings and reactions were conducted as described [76]. TRAP assays on STN1, TEN1, CTC1-CFP, or TERT IP samples were performed by using 1 µl of the final IP slurry. The telomerase processivity TRAP (TP-TRAP) protocol was adapted from [61] and performed as previously described [41]. Briefly, TP-TRAP entails telomerase extension of a substrate primer followed by the first round of PCR with a 1RPgg primer to incorporate a unique sequence tag on telomerase products. A primer complementary to the tagged region (2RP) is added for the second PCR step followed by 33× cycles of PCR. Relative telomerase activity was measured by Quantitative TRAP (qTRAP) via SYBR Geen (Bio-RAD) qPCR after primer extension as discussed [78].
Fifty micrograms of wild type, stn1, and pot1a extracts were used for input samples. IP samples were boiled for 5 min in SDS loading dye. Samples were run on a 12% SDS-PAGE gel followed by protein gel blotting. Proteins were transferred overnight at 4°C onto a polyvinylidene difluoride (PVDF) membrane, followed by 2 hrs of blocking using 6% non-fat dried milk dissolved in 1× TBS-T (50 mM Tris, 150 mM NaCl, 0.1% Tween-20). Rabbit anti-STN1 antibody was diluted 1∶5000 in TBS-T and incubated with the protein blot for 4 hrs followed by 3× washes with TBS-T. Secondary anti-rabbit horseradish peroxidase was diluted 1∶7500 in TBS-T and incubated with the protein blot for 2 hrs, followed by 3× washes with TBS-T. Final detection was performed using an ECL prime protein blotting kit (GE Healthcare). Western blotting was performed as described for CTC1-CFP and POT1a [59] and TEN1 [41].
RNA was extracted from 5 day-old seedlings (Omega Bio-tek) followed by DNase I digestion (Zymogen) for 30 min at room temperature. RNA was phenol: chloroform extracted followed by EtOH precipitation. 1 µg of RNA was reverse transcribed (Quanta Supermix), then diluted 1∶4 using thousand-fold diluted yeast tRNAs. 1 µl of cDNA was used for qRT-PCR using CFX Connect Real-Time System (Bio-Rad) in triplicate. Quantification is from three biological replicates and normalized to wild type for each gene expression.
|
10.1371/journal.pntd.0005126 | Serological and Virological Evidence of Crimean-Congo Haemorrhagic Fever Virus Circulation in the Human Population of Borno State, Northeastern Nigeria | Despite several studies on the seroprevalence of antibodies against Crimean-Congo Haemorrhagic Fever virus (CCHFV) from humans and cattle in Nigeria, detailed investigation looking at IgG and IgM have not been reported. Additionally, there have been no confirmed cases of human CCHFV infection reported from Nigeria.
Samples from sera (n = 1189) collected from four Local Government Areas in Borno State (Askira/Uba, Damboa, Jere and Maiduguri) were assessed for the presence of IgG and IgM antibodies. The positivity rates for IgG and IgM were 10.6% and 3.5%, respectively. Additionally, sera from undiagnosed febrile patients (n = 380) were assessed by RT-PCR assay for the presence of CCHFV RNA. One positive sample was characterised by further by next generation sequencing (NGS) resulting in complete S, M and L segment sequences.
This article provides evidence for the continued exposure of the human population of Nigeria to CCHFV. The genomic analysis provides the first published evidence of a human case of CCHFV in Nigeria and its phylogenetic context.
| Crimean-Congo haemorrhagic fever (CCHF) is an acute tick-borne zoonotic disease. The causative agent, CCHF virus (CCHFV), has the most extensive geographical distribution of the medically important tick-borne viral diseases with a distribution over much of Asia, the Middle East, Africa and expanding areas of south-eastern Europe. Whilst the main route of human infection with CCHFV is predominantly through tick bite, it can also be spread via bodily fluids and it has a reputation for causing nosocomial outbreaks in healthcare settings. Whilst CCHFV has been identified in ticks from Nigeria since 1970, there is scarce information on human infection. Within this report, the prevalence of CCHFV-reactive antibodies has been assessed in human sera providing evidence of continued circulation of the virus in the human population of Borno state, Nigeria. Additionally, in one sample the presence of viral RNA was detected which allowed a full sequence of the CCHFV to be obtained. This is the first report of CCHFV being associated in a human case from Nigeria and the full genetic characterisation of the virus being completed. The evidence within supports the hypothesis that CCHFV is endemic in Nigeria and should be considered as an aetiological agent in febrile patients.
| Crimean-Congo haemorrhagic fever (CCHF) is a disease caused by the CCHF virus (CCHFV) a member of the Nairovirus genus, family Bunyaviridae [1]. The disease was first reported in the Crimean peninsula in the mid-1940s, after a large outbreak of severe haemorrhagic fever, described as Crimean haemorrhagic fever (CHF), occurred with a case fatality rate of >30% [2]. The virus aetiology of CHF was not determined until the late 1960s, when it was subsequently shown to be antigenically identical to the Congo virus identified in Africa in 1967 [3,4]. The naming of the virus went through a series of steps before the name Crimean-Congo haemorrhagic fever virus was agreed in 1979 [5]. The disease is now endemic in many countries in Europe, Asia and Africa [1]. In nature, CCHFV is maintained predominantly in Ixodid tick vectors, particularly ticks of the genus Hyalomma [6]. Whilst tick bite is the most common route of CCHFV infection, person-to-person transmission can occur via direct exposure to blood or other secretions [6]. Direct zoonotic transmission from viremic animal hosts is also possible [7].
The first report of CCHFV in Nigeria occurred in 1970, when it was identified in various tick species, including Hyalomma spp. collected from market animals, and hedgehogs [8]. Interestingly, very few cases of CCHF have been recorded in Africa [9]; the majority are described from South Africa [10]. The risk of CCHF in several African countries is poorly defined and infection with CCHFV is often undiagnosed or unreported in these regions [11]. Importantly, CCHFV is a notorious cause of nosocomial infections especially when undiagnosed, and the virus presents a significant risk to health care workers [12–16].
A previous study of CCHFV conducted within Borno State, Nigeria reported a seroprevalence of 2.4% (7 out of 297 individuals) where the reservoir, vectors and intermediate hosts abound in the area [3]. In the present study, seroprevalance was expanded across different Local Government Areas (LGAs) of Borno State and samples from patients with undiagnosed febrile illness were assessed for the presence of CCHFV RNA.
Blood samples tested in this study were previously taken for the laboratory diagnosis of malaria, typhoid or hepatitis and were classified as clinical specimens. No samples were collected specifically for this work, thus ethical approval for the study design was not required. Samples were anonymised within Nigeria, so investigators were only supplied with sequentially numbered samples. For the sample that was identified as positive for CCHFV RNA, the local team in Nigeria were able to provide more information on the background of the case. The patient described here is anonymous. Blood samples were collected and stored within the University of Maiduguri Teaching Hospital, Nigeria. Samples for testing at Public Health England, UK were sent in accordance with national guidelines for both Nigeria and the UK.
Samples for serology testing were randomly selected from a cross section of humans in both rural and urban populations. Patients generally presented with febrile illnesses, and were screened for common aetiological agents, such as malaria and typhoid. They were associated with different occupational groupings which included abattoir workers, students, civil servants, and unemployed people, and they were of different age groups and genders. They presented to different clinics during 2010–2014. Samples were collected from 4 out of 27 LGAs within Borno State, namely: Askira/Uba, Damboa, Jere, and Maiduguri (Fig 1). Samples were heat treated at 56°C for 30 minutes prior to testing. The nature of the sampling sites is listed in Table 1. Within Askira/Uba is the town of Lassa where the first case of Lassa fever was reported in 1969 [17]. An abattoir within Maiduguri city, Maiduguri Metropolitan Abattoir (MMA), serves as the major animal slaughterhouse for the region. Animals (camel, cattle, goat and sheep) are brought from within the State (Borno), from neighbouring states and also from neighbouring countries such as Cameroon, Chad, Niger, Sudan and Central Africa.
For molecular analysis, 380 serum samples from febrile patients in the acute phase of illness (fever and/or headache) were tested. These samples were from patients had previously been tested and found negative for both malaria and typhoid.
Initial serology testing was conducted in Nigeria using an in-house ELISA to detect IgG and IgM antibodies against recombinant CCHFV nucleoprotein [18]. A selection of samples was shipped to the UK for confirmatory analysis using both the in-house ELISA and a commercially-available assay system (Vector-Best, Russia). For the latter, the manufacturer’s instructions supplied with the kits were followed.
Extraction of RNA was performed using the MagnaPure 96 small volume RNA kit (Roche). Plates were loaded onto the MagnaPure 96 automated extraction robot and RNA was eluted in 60μl nuclease free water. Target amplification was performed using primers to the CCHFV S segment [19] with a Superscript® Platinum One-Step III qRT-PCR Kit (Life Technologies). Amplification was performed using the ABi 7500 (Applied Biosystems) at the following cycling conditions: 50°C for 10 minutes, 95°C for 2 minutes followed by 40 cycles of 95°C for 10 seconds and 60°C for 40 seconds; and a cooling cycle of 40°C for 30 seconds. Temperature cycling was set to maximum ramp speed and data were acquired and analysed using the ABi 7500 on-board software with an automatically selected threshold.
The isolated RNA was treated with DNAse I (Life Technologies) following manufacturer’s instructions. After DNAse inactivation, the sample was cleaned up using a Zymo Clean and Concentrator column and eluted into 6 μl H2O.
A 5μl aliquot of the DNase I-treated RNA was used to prepare cDNA using the Ovation® RNA-Seq System V2 (NuGen) following manufacturer’s instructions, with the exception that RNA was denatured for 5 min at 85°C prior to first strand synthesis. cDNA was purified using a MinElute Reaction Cleanup Kit (Qiagen) and eluted in 15 μl H2O. Final cDNA concentration was determined using the QuBit broad-range double-stranded DNA assay (Life Technologies).
An Illumina sequencing library was prepared using the Nextera XT V2 kit with 1.5 ng of cDNA as input, following manufacturer’s instructions. Indices were selected using the Illumina experiment manager software. Fragment size analysis was performed using a bioanalyser (Agilent Technologies). The KAPA library quantification kit for NGS (Kapa Biosystems) was used for quantification. The prepared sequencing library was run on an Iluminia MiSeq by the PHE Genomics Services Unit (GSDU).
Reads were trimmed to remove adaptors and low quality bases, to achieve an average phred score of Q30 across the read, using trimmomatic [20]. Trimmed reads were taxonomically assigned using Kraken [21] (ver.0.10.4-beta) populated with bacterial, viral and archaeal genomes and a representative yeast genome, from RefSeq (ver. 66) with the addition of 141 viral GenBank sequences.
Following Kraken assignment, viral reads were extracted from the fastq files using seq_select_by_id [22], and assembled using SPADES (ver. 3.1.1) [23]. All reads were then used to scaffold these contigs using SSPACE (ver. 1.0.5) [24]. Contigs larger than 1 kb were aligned to the CCHFV reference sequence (NC_005301.3, NC_005300.2 and NC_005302.1 for L, M and S segments, respectively), using Mauve Contig Mover (ver. 1.0.0) [25].
Reads were mapped to both assembled contigs and the CCHFV reference (NC_005301.3, NC_005300.2 and NC_005302.1) using BWA (ver. 0.7.5) [26]. Consensus genome sequence was produced at a minimum depth of five reads using an in-house script. All of the above was performed using a local instance of the Galaxy Project [27–29]. BAM files were visualised using tablet [30].
Phylogenetic analyses were performed using MEGA 6 [31]. Trees were precomputed using the Neighbour-Joining method [32], then evolutionary history and distances were inferred by the Maximum Likelihood method. Maximum Likelihood phylogenetic trees were generated for the open reading frames of the partial L, M and S sequences recovered from next-generation sequencing. All positions containing gaps and missing data were eliminated from the analysis.
Of the 1,189 sera from the 4 LGAs tested, 126 were positive for CCHFV IgG giving an overall seroprevalence of 10.6%, while 42 (3.5%) and 7 (0.6%) were seropositive for IgM or IgG+IgM antibodies, respectively (Table 2). The study shows that the prevalence of IgG was higher in rural (15%) than in urban (8.9%) areas; conversely the incidence of IgM was slightly higher in urban (3.9%) than in rural (2.7%) areas.
Of 380 samples assessed for the presence of CCHF viral RNA by RT-PCR, a single sample (ID: N428) was positive with a cycle threshold (CT) value of 29.42 and a slope formation typically seen for a positive sample. The RNA from this sample was used for further characterisation.
Kraken taxanomic analysis placed 0.43% of reads as the species CCHFV. For confirmation, these reads were assembled into 136 contigs larger than 200bp. A nucleotide BLAST of these contigs agreed with the Kraken taxonomic analysis in identifying 12 contigs with typically greater than 95% identity with CCHF Sudan ABI_2009. Using Mauve contig mover the 136 contigs were aligned to the CCHFV reference genome where it was found that an assembled 8 contigs aligned to 93.26% of the viral genome. In order to assess the support for these assembled contigs, reads were mapped back to contigs and gave an average coverage of 61.39, 73.38, 34.71, 21.12, 16.38, 22.85, 198.4 and 115.17 fold across each contig respectively.
In parallel, reads were mapped to the CCHFV reference genome (NC_005301.3, NC_005300.2 and NC_005302.1) which gave 94.54% coverage of the full genome at a minimum depth of 5 bases (99.47% coverage for the L segment, 89.73% for the M segment and 94.43% for the S segment) and an average depth of coverage across the genome of 73.3 fold (with an average depth of 66.2 reads for the L segment, 71.2 for the M segment and 82.5 for the S segment). A consensus genome sequence was produced at a minimum depth of five reads. Sequences were submitted to GenBank with accession numbers KX238956-KX238958. Near full genome coverage was seen for the sample when mapped to CCHFV (Table 3). These mapping data are illustrated in Fig 2.
The genome assembly included three segments; the L Segment [GenBank KX238956], M Segment [GenBank KX238957] and S Segment [GenBank KX238958]. Phylogeny was inferred using the maximum likelihood method based on the Tamura-Nei model [33] and confidence assessed with the Bootstrap Test with 1000 resamplings. Phylogenetic analysis clustered the S segment in the Africa 3 phylogenetic group (Fig 3). The S segment open-reading frame showed close homology with a previous isolate of CCHFV from Nigeria (IbAr10200), as well as isolates from Mauritania (ArD39554) and South Africa (SPU415/85 and SPU128/61/7). The M and the L clustered closely with the Sudan AB1-2009 isolate and the Nigeria IbAr10200 isolate (Fig 4).
Our serological results demonstrate the circulation of CCHFV in different LGAs of Borno State, Nigeria. The average positivity rates were 10.6% for IgG responses and 3.5% for IgM responses. A previous study in Borno State using 297 samples collected from patients attending health facilities between September 2011 and February 2012 showed a prevalence rate for IgG antibodies against CCHFV of 2.4% using a similar ELISA test to that used in this study [34]. The difference in IgG levels may be due to different study populations, for example our report was conducted on samples collected from 4 LGAs whereas the earlier report was conducted on samples from 10 LGAs [34]. In 1974, it was reported that 9.6% of 250 sera collected in Nigeria had neutralisation activity specific to CCHFV [35]. As neutralisation activity was not performed in our study, cross-reactivity with other nairoviruses might be considered. The similarity in antibody responses between our study and that conducted in 1974 indicates that despite the 38 year time period, similar frequencies of human CCHFV infection remain in Nigeria. This is consistent with recent evidence consensus of a moderate level of CCHFV in Nigeria [22]. Given the expanding population of working Nigerians [36], there may be a greater likelihood of tick / CCHFV exposure in the future.
Although the data are focused on Borno State, we speculate that CCHFV is circulating in neighbouring countries which share common borders (Cameroon, Chad and Niger). Furthermore these borders are porous, and unrestricted human and animal movements are common throughout the year. Since the natural lifecycle of the Hyalomma tick involves feeding on cattle, CCHFV infectivity of cattle can help support data on human cases. In northern Nigeria, around 25.7% of sera collected from cattle showed antibody responses to CCHFV as tested by agar diffusion precipitation tests [37].
For molecular testing, sera samples from undiagnosed febrile patients were assessed for the presence of CCHFV using RT-PCR. Of the 380 samples tested, only a single sample showed a positive signal. This sample (N428) was obtained from a 15 year old female patient who was resident in old Maiduguri, a settlement almost at the outskirt of Maiduguri city. She was admitted in March 2012 into the female medical ward of the University of Maiduguri Teaching Hospital, a tertiary health facility in northeastern Nigeria with a 6 day history of fever, body pain, bloody diarrhoea and epistaxis. NGS of RNA from sample N428 identified the presence of CCHFV. Phylogenetic characterisation of the viral S segment sequence demonstrated that it belonged to the Africa 3 clade, in congruity with reports for other viruses isolated from Nigeria and similar geographies [6,19]. The patient went on to make a full recovery.
Unfortunately, since samples were heat-treated, virus isolation could not be performed. This is a common issue, exacerbated by difficulties and delays in transportation to appropriate containment facilities [38]. Previously only partial genomic sequence has been attainable in such cases [38,39]. However, the advent and use of NGS technologies here has enabled the efficient and rapid, characterisation of a clinical strain of CCHFV from Nigeria. Analysis of the genetic relationship between this CCHFV and previously characterised isolates shows close homology to the IbAr10200 strain, isolated from Hyalomma excavatum ticks in Sokoto, Nigeria in 1966 [40]. In addition to being a tick virus, this strain has also been passaged multiple times in the laboratory. It is interesting that despite being isolated 50 years ago from a tick, there is remarkable similarity in these Nigerian viruses. Analogous observations have been made in the past [41,42], illustrating that CCHFV can be genetically very stable over long periods of time, while on other occasions there is vast genetic variation between strains, even between those sequenced from similar locations [43]. Such observations may point to the broader ecological conditions which support the virus-host environment; highly changeable ecologies resulting in more opportunities for CCHF viruses to exploit new sequence space, whereas stable ecological conditions would restrict diversity. Whilst there has been strong serological evidence for CCHFV circulation in humans and cattle in Nigeria in the past, our data are the first to detect and directly sequence viral RNA in a human sample. Thus, this is the first report of human CCHFV infection in Nigeria. Importantly our work highlights that CCHFV should be considered as a potential cause of febrile illness in patients within the region and that further studies on the risk of CCHFV infection and how such risks could be reduced should be considered.
|
10.1371/journal.pcbi.1005648 | mRNA/protein sequence complementarity and its determinants: The impact of affinity scales | It has recently been demonstrated that the nucleobase-density profiles of mRNA coding sequences are related in a complementary manner to the nucleobase-affinity profiles of their cognate protein sequences. Based on this, it has been proposed that cognate mRNA/protein pairs may bind in a co-aligned manner, especially if unstructured. Here, we study the dependence of mRNA/protein sequence complementarity on the properties of the nucleobase/amino-acid affinity scales used. Specifically, we sample the space of randomly generated scales by employing a Monte Carlo strategy with a fitness function that depends directly on the level of complementarity. For model organisms representing all three domains of life, we show that even short searches reproducibly converge upon highly optimized scales, implying that the topology of the underlying fitness landscape is decidedly funnel-like. Furthermore, the optimized scales, generated without any consideration of the physicochemical attributes of nucleobases or amino acids, resemble closely the nucleobase/amino-acid binding affinity scales obtained from experimental structures of RNA-protein complexes. This provides support for the claim that mRNA/protein sequence complementarity may indeed be related to binding between the two. Finally, we characterize suboptimal scales and show that intermediate-to-high complementarity can be reached by substantially diverse scales, but with select amino acids contributing disproportionally. Our results expose the dependence of cognate mRNA/protein sequence complementarity on the properties of the underlying nucleobase/amino-acid affinity scales and provide quantitative constraints that any physical scales need to satisfy for the complementarity to hold.
| Messenger RNAs and proteins, two essential types of biopolymers, have recently been shown to exhibit closely related, complementary physicochemical properties. Specifically, density profiles of certain groups in messenger RNA sequences directly match the affinity profiles for precisely those groups in protein sequences they encode. Based on this, it has been suggested that these molecules may interact with each other specifically and in a co-aligned fashion, especially when unstructured. Here, we explore different amino-acid scales used in the above analysis to assess which of their properties dictate the observed matching. Specifically, we define the constraints that need to be satisfied by physical scales for the complementarity to hold and show that the previously derived nucleobase/amino-acid affinity scales indeed satisfy these constraints. As a whole, our work provides a quantitative foundation for understanding the putative messenger RNA/protein complementarity with implications in different areas of RNA/protein biology including transcription, translation, splicing and viral assembly.
| The relationship between mRNAs and the proteins they encode is one of the key defining characteristics of life at the molecular level [1–3]. While this relationship has primarily been studied in the context of biological information transfer, less attention has been paid to a potential link between the physicochemical properties of the two biopolymers. Recently, however, we have reported an unexpectedly strong connection between the nucleobase-density profiles of mRNA coding sequences and the nucleobase-affinity profiles of their cognate proteins [4–6]. For example, purine (PUR) density profiles of E. coli mRNA coding sequences match their cognate protein’s guanine (GUA) affinity profiles with an average Pearson correlation coefficient of -0.76 (note the negative values for R indicate matching as a result of the standard definition of binding affinities) [4,5]. As illustrated in Fig 1A, the protein GUA-affinity profiles in this analysis were calculated by weighting their sequences with the GUA-affinity values for individual amino acids, which in turn were derived from known 3D structures of RNA/protein complexes by using a knowledge-based formalism [4]. In addition, we have also studied other affinity scales derived by diverse experimental and theoretical approaches: 1) a chromatographically determined scale of amino-acid affinity for pyrimidine (PYR) mimetics pyridines [7], 2) a computationally derived variant of the same scale [8], 3) absolute binding free energy scales between nucleobases and amino-acid sidechain analogs in different solvents [9], and 4) affinity scales obtained from simulated partitioning experiments using realistic RNA nucleobases [10,11]. The consensus of these studies has been that the mRNA regions rich in a particular nucleobase or a type of nucleobases (PUR or PYR) tend to encode the protein regions with a pronounced affinity for precisely those or similar bases. This novel finding is well illustrated by the mRNA PUR-density profile and its cognate protein’s GUA-affinity profile of a typical, representative mRNA/protein pair in E. coli whose Pearson correlation coefficient (-0.76) equals that of the mean of the entire E. coli distribution (Fig 1A). Importantly, the only exception to the above rule was seen in the case of adenine (ADE). Namely, protein regions with a high affinity for the purine base ADE tend to be encoded by mRNA regions rich in PYR bases [5,6].
Although robust and of potentially far-reaching implications, the above findings still lack a definitive explanation. We have suggested two possibilities: one, concerning the reasons for the observed complementarity, and another one, concerning its implications. First, the above observations are consistent with a possibility that genetic encoding may have arisen from binding, as postulated by the stereochemical hypothesis of the origin of the genetic code [2,12–13]. This hypothesis dates back to 1960s, but we believe our results provide the strongest-yet evidence for it. Importantly, however, our results shift the focus towards analyzing the binding in the context of longer biopolymers and not just isolated codons and amino acids [4–6,14]. Second, the compositional mirroring between mRNAs and their cognate proteins at the level of primary sequences supports the possibility of complementary, co-aligned binding between the two even in modern systems, especially if they are unstructured [4–6,14]. On the side of proteins, this pertains to both intrinsically disordered proteins as well as the unfolded states of otherwise folded proteins, such as during translation or upon thermal/chemical stress. While we do not exclude the possibility of interactions in the structured states of either partner [15], this also more likely involves mRNA stretches that are not base-paired. Overall, the potential implications of this interpretation concern different facets of RNA/protein biology including translation control, structure of ribonucleoprotein particles, the behavior of non-membrane-bound cellular compartments, viral capsid assembly and others [5]. More than two decades ago, Kyrpides and Ouzounis proposed that cognate mRNA/protein interactions may be an ancient mechanism for auto-regulation of mRNA stability [16]. Our findings now provide mechanistic details behind such a possibility. Here it should also be mentioned that the opposite behavior of ADE suggests that there may have been at least two major phases in the code’s development, the more recent of which introduced ADE to modulate in a negative manner the complementarity engendered by other bases [5].
The statistical significance of the cognate mRNA/protein sequence complementarity has already been probed by different tests involving randomizations of the genetic code table, shuffling of both the affinity scales and the primary mRNA/protein sequences and analyzing the behavior of other physically realistic amino-acid property scales in the same context [4,6,14]. However, these approaches, although valid and necessary, have thus far not included a systematic exploration of the whole space of nucleobase/amino-acid affinity scales. On the other hand, several important questions can only be answered with an in-depth knowledge of the properties of the space of affinity scales and their influence on the observed complementarity. A key open problem in this regard concerns the uniqueness of the affinity scales that yield a high degree of cognate sequence-profile matching. Is there an optimal scale that results in maximized matching between mRNA nucleobase-density profiles and their cognate protein sequence profiles or are there multiple scales that produce the same or similar levels of matching over complete proteomes? How many different scales are there that show intermediate-to-high levels of matching? Finally, are there specific amino acids whose affinities for different nucleobases dominate the matching? It is easily imaginable that there are amino acids that exhibit similar affinities to different nucleobases, while the specificity i.e. the sequence complementarity, is dictated by a select few. To address these questions, one needs to consider not only the already published, physically realistic scales, but also those that do not produce strongly matching profiles. In other words, one would like to sample the space of randomly-generated scales using a fitness function that is related to the degree of cognate mRNA/protein sequence complementarity engendered by those scales.
Here, we present a Monte Carlo (MC) search method that satisfies the above criteria and explores the space of nucleobase/amino-acid affinity scales, while at the same time assessing their impact on the degree of cognate complementarity (Fig 1A and 1B). More specifically, our MC searches start from a uniform scale with identical weights for all 20 amino acids and evolve through a succession of random perturbations, which are accepted or rejected according to a fitness function. The latter, in turn, is related to the degree of complementarity as captured by the proteome-average Pearson correlation coefficient <R> between the cognate sequence profiles (see Materials and Methods for details). In other words, we search for amino-acid scales (i.e. 20-element linear arrays of amino-acid weights), which result in a given level of proteome-wide average matching between mRNA nucleobase-density profiles and the cognate protein profiles generated by weighting their sequences using these scales. Notably, our procedure is completely computational and does not impose any physicochemical constraints on the sampled scales. As a consequence, it provides an unbiased, detailed characterization of the space of amino-acid scales and their effect on the cognate mRNA/protein sequence complementarity. Finally, there are ongoing efforts in our and other laboratories to test the hypothesis that compositional sequence complementarity between mRNAs and their cognate proteins implies binding between them. The primary aim of the present work, however, is to assess the impact of affinity scales on such sequence complementarity, which remains a fact even in the absence of experimental verification of the hypothesis that it implies binding [4–6,9,11].
Complete annotated proteomes of Escherichia coli, Methanocaldococcus jannaschii and Saccharomyces cerevisiae along with the corresponding mRNA coding sequences were analyzed. The protein sequence data was extracted from the UniProtKB database with the maximal-protein-evidence-level set to 4, including only reviewed Swiss-Prot entries for the analysis [17–19]. Coding sequences for each protein were extracted using the ‘Cross-references’ section of each entry in the UniProtKB. Of all the entries, the first one satisfying the length criterion of mRNA length = 3 x protein length + 3 was selected and the sequence downloaded from the European Nucleotide Archive Database. All sequences containing non-canonical amino acids or nucleobases were disregarded in the analysis. The complete sets of mRNA/protein data used in this study are included in S1 Dataset. Note that in the present study our analysis is reserved for primary sequences of complete mRNA coding sequences and their cognate proteins only, without consideration of higher-order structural organization. This not only enables a direct 1-to-1 mapping between mRNA and protein sequences, but also allows for a full exploration of the complementarity hypothesis in the unstructured context. Analysis of the influence of structure on the side of protein has been published elsewhere [15], while an analogous analysis on the side mRNAs will be the topic of our future work.
The mRNA nucleobase-density profiles and the corresponding protein nucleobase-affinity profiles were compared by calculating the linear Pearson correlation coefficients R between them. Prior to calculation, the sequences were smoothed via a window-averaging procedure with a 63-nucleotide window for mRNAs and a 21-residue window for proteins as used before [4–6,20]. Importantly, sequence profile comparison is largely insensitive to the size of the averaging window, as shown previously [4]. Furthermore, the scale values obtained during the simulation may result in any arbitrary value including negative numbers. Once the simulation is finished, the resulting scales are rescaled so that the values are in the range of [0, 1] i.e. the lowest value is set to 0, the highest to 1 and the rest are rescaled accordingly. This is done in order to distinguish between truly different scales and the rescaled versions of the same scale. Importantly, this procedure has no impact on our analysis since both window-averaging and calculation of Pearson correlation coefficients are invariant with respect to linear rescaling.
The simulations were carried out using a combination of a C++ program for calculating the proteome-average correlation coefficients between mRNA and protein sequence profiles [6] and a Python script for implementing the Monte Carlo (MC) search. At each step in a given MC simulation, anywhere between 1 and 4 randomly chosen scale values were perturbed by a randomly chosen offset. A simulated annealing procedure was implemented in order to vary the size of the offsets from a randomly chosen value between [-0.1, 0.1] in the beginning of the simulation to a value between [-0.01, 0.01] in the end, with a linear ramp between the two. A given MC move i.e. a given scale, is accepted according to a zero-temperature, downhill Metropolis criterion: if a new scale results in a lower average Pearson R across the proteome (<R>) as compared to the scale it was derived from, it is accepted, and it is rejected otherwise. Affinity scales were generated individually for each of the four RNA nucleobase (uracil—URA, cytosine—CYT, adenine—ADE and guanine—GUA) as well as PUR (preference for both ADE and GUA). Given that the mRNA PYR fraction in a given stretch is directly related to the PUR fraction (%PYR = 1—%PUR), the PYR scales are by definition the inverses of the PUR scales and were for this reason not explicitly included in our analysis. Finally, the number of steps and the speed of simulated annealing both influence the final result. The number of steps has been chosen to be at least 3 times the number of moves necessary to reach a stable minimum.
The MC approach does not produce interaction scales for a given level of matching but the other way around–only once a scale has been created, its level of matching is calculated. In order to generate the full landscape, we chose those scales that showed the closest value of matching to the target value of the reported Pearson R. In the construction of the landscape, we include individual scales that are within +/-0.01 in Pearson R from each reported value of <R>. Considering the very fast evolution of interaction scales, for some low values of <R> we did not obtain a full set of 1000 scales that would match this criterion, which were therefore not included.
Data analysis was performed using the R statistical programming language. Calculations were performed using custom software written in C++ and Python. Plotting and data visualization was performed in R and Python, while figures were generated in Gimp and Inkscape.
The MC-optimized scales were compared with the corresponding physically realistic knowledge-based scale by calculating the Pearson correlation coefficient R between them. The significance of the obtained correlation coefficients was ascertained by a randomization procedure whereby the reported p-values correspond to the fraction of a set of 106 scales with randomly chosen values exhibiting a more negative Pearson R than the MC-generated scales (a one-tailed significance test). Two-tailed p-values were calculated by multiplying the initial p-value by 2 if below 0.5, or as 1—p-value otherwise. Combined p-values were calculated according to Fisher’s method based on two-tailed p-values [21,22]. The p-values were calculated from their respective Χ² distributions utilizing Microsoft Excel’s function CHIDIST().
Optimized scales, which lead to a close matching between mRNA nucleobase-density profiles and sequence-weighted cognate protein profiles, could be identified in an extremely low number of MC moves. For example, the number of MC moves required to reach a proteome-wide average matching of mRNA PUR-density profiles with <R> ≤ -0.86 in E. coli is only 322 ± 66.5 (standard deviation) as evaluated over 1000 independent MC simulations initiated with the system time as a random seed for each run (S1 Fig). Importantly, for all the combinations tested, the optimal scales appear to be extremely well defined and emerge robustly at the end of all independent MC simulations (scales with average weights are given in S1 Table). Also, the equivalent optimized scales for the three organisms studied are highly similar to each other with pairwise Pearson Rs in all cases exceeding 0.86 (S1 Table). The minor deviation between the scales of different organisms can be explained by the codon usage bias as well as different amino-acid composition of the respective organisms. For example, the difference between CYTE. Coli and CYTM. Jannaschii can be traced back to the different choice of codons for arginine.
In Fig 2A and 2B, we trace the evolution of the average and the standard deviation of the normalized values of individual amino-acid weights corresponding to a scale that was optimized for matching the mRNA PUR-density profiles in E. coli. The mean weights corresponding to the majority of amino acids, as obtained from 1000 independent repetitions, exhibit well-defined ranks starting already with low levels of matching and attain their definitive ranks already at approximately <R> = -0.5 (Fig 2A). In this particular case, the extreme weights correspond reproducibly to Phe on the high side and Glu and Lys on the low side. The reproducibility of optimal scales is attested by the extremely low standard deviations of amino-acid weights at the extremely high values of <R> (Fig 2B). Importantly, although a sharp drop in standard deviation towards high levels of matching is observed, a clear trend for the specific values is achieved only very late in the simulations i.e. for the most extreme values of <R> only. Expectedly, the highest diversity of scale values is obtained for <R> ~ 0, but a substantial variability is retained even for intermediate-to-high values of matching: e.g. for <R> between -0.4 and -0.6, the standard deviations remain close to 0.25 in normalized units (Fig 2B). For comparison, a uniform random distribution between 0 and 1 results in a standard deviation of 1/√12 ~ 0.29.
The sequences of mRNAs and their cognate proteins are, of course, linked by the universal genetic code. Therefore, suitable amino-acid scales for weighting protein sequences to match their cognate mRNA’s density profiles correspond, for a particular nucleobase, to the relative fractions of that nucleobase in the respective codons, weighted by the codon usage bias. For example, the scale derived in such a way results in an average matching of <R> = -0.86 for the mRNA PUR-density profiles in E. coli and the cognate protein sequences weighted by the average, usage-bias-weighted PUR content of the respective codons. Equivalent levels of matching were obtained for all nucleobases and organisms (Table 1). Interestingly, the scale composed of the average MC-optimized weights derived from E. coli mRNA PUR-density profiles correlates with the scale derived from codon PUR fractions with a Pearson R = 0.997 (S1 Table). Similar results were obtained for all other nucleobases and in all other organisms, albeit with slightly lower levels of correlation in some cases (S1 Table).
The main advantage of the MC approach, besides its efficiency, is that it also provides thorough sampling of suboptimal scales. This has allowed us to explore the development of scale properties in relation to the degree of mRNA/protein sequence complementarity. When looking at the distribution of specific values for a given level of <R> over the whole set of mRNAs and proteins, strong differences between the final optimized values and the intermediate values are identified. In Fig 2C and 2D, we show the complete distributions of weights for different amino acids at two different levels of average matching in E. coli (<R> = -0.80 and <R> = -0.86) for the PUR-density scales. Here, from each independent simulation, the scale resulting in <R> closest to the target value was selected. Note that the <R> values of all selected scales round to the reported target value at the second decimal place. As can be seen, only the most extremely optimized scales i.e. those with <R> = -0.86, exhibit well-defined weights for the majority of amino acids. For example, although an average Pearson <R> of -0.80 can be considered a high level of matching, most scale weights are still broadly distributed at that level (Fig 2C). Importantly, different amino acids exhibit distributions of different widths at a given value of <R>, with some converging to tighter distributions earlier than others. For example, the weights for Phe, Lys and Glu attain their final values early on and exhibit standard deviations that are lower than for any other amino acids at most values of <R>. In general, the distinct behavior of the optimized weights corresponding to individual amino acids is also seen for other scales and in other organisms (S2 Fig).
How many distinctly different scales are able to perform similarly well when it comes to the sequence profile matching of mRNAs with their cognate proteins? To separate scales at a given level of matching into several subsets, we have applied a hierarchical clustering algorithm, which has allowed us to build dendrograms for each level of <R>. As a natural distance measure between scales in this clustering approach, we have used 1—R, with lower numbers signifying higher similarity between two given scales and vice versa. In Fig 3A, we show two such dendrograms capturing the diversity of scales obtained by matching mRNA PUR-density profiles in E. coli at <R> = -0.8 and <R> = -0.86. To build a landscape of affinity scales, we have cut these dendrograms at specified distance cutoffs and have reported the number of clusters at a given cutoff as exemplified in Fig 3A for E. coli. The landscape is presented in Fig 3B with the values given representing the upper cutoff for the distance (1—R) of two scales in one cluster. From this landscape, it is clear that only at the highest levels of correlation between mRNA nucleobase profiles and nucleobase-affinity profiles of their cognate proteins do the affinity scales show a very similar structure. At lower levels, a very diverse set of scales is able to perform comparably well i.e. the lower values of matching can be attributed to a wider class of interaction scales.
Previously, several complete interaction scales for nucleobase/amino-acid interactions have been reported from different groups [23–26]. While some of these scales focus on the general affinity of amino acids for RNAs, but do not differentiate between specific nucleobases, other scales report on the specific propensities of the 20 amino acids for each of the four RNA nucleobases. The latter, for example, include scales derived in different ways including chromatographic experiments [7], absolute binding free energy calculations [9], classical and quantum-mechanical interaction enthalpy calculations [26], knowledge-based analysis of nucleobase/amino-acid contacts derived from X-ray and NMR structures [6] and simulated partitioning experiments [10,11]. Here, we focus on the knowledge-based scales as they are arguably the most relevant proxies for the nucleobase/amino-acid affinities at the realistic RNA/protein interfaces. In Fig 4A, we show the Pearson Rs between each of MC-derived scale for the E. coli dataset with each of the physical, knowledge-based scales derived by Polyansky and Zagrovic [6]. Below the correlation coefficients, we list the p-value capturing the statistical significance for the specific correlations. Importantly, three out of five pairs of the corresponding scales (GUA, CYT, PUR) exhibit Pearson Rs > 0.5 and p-values ≤ 0.024 each, suggesting high statistical significance. Moreover, the URA scales also exhibit a positive correlation coefficient (R = 0.30, p-value = 0.098), albeit not as strong as the other three pairs. Interestingly, the knowledge-based ADE-affinity scale displays a strong anti-correlation with its MC-generated counterpart. This anti-correlating behavior is also seen if one focuses on PURs only. Namely, although ADE is a purine base, its affinity scale correlates inversely with the values for MC-generated PUR scale. On the other hand, the GUA-affinity knowledge-based scale shows by far the strongest correlation with the generated PUR scales: the relative preferences of amino acids when it comes to interaction with GUA in RNA-protein structures show values very similar to those obtained by our MC procedure, which only considers the matching between mRNA PUR-density profiles and appropriately weighted cognate protein profiles. The two correlate with a Pearson R of 0.85 and no major outliers (Fig 4B).
We have also calculated the combined p-values from the two-tailed values given in Fig 4 for two different combinations of entries in the table. In the case of the Fisher method for combining p-values, the individual tests need to be independent from each other [21,22]. Arguably, the closest subset to this requirement is the combination of the four individual diagonal elements, entailing a comparison between the corresponding scales for individual bases. This set results in a highly significant combined p-value of 1.7 x 10−4. Moreover, when including all combinations of URA, CYT, ADE GUA and PUR scales, a significance level of 8.3 x 10−11 is reached. Here, it should also be noted that there exist other suboptimal scales, which correlate better with the knowledge-based scales than do the optimal scales. For example, in the case of the E.coli PUR scale vs. the knowledge-based GUA scale, the highest correlation achieved between the two is R = 0.94, which is obtained for a suboptimal scale that itself results in an average proteome-wide correlation of <R> = -0.79. A complete analysis of suboptimal scales and their relationships with knowledge-based scales is presented in S2 Table. In general, the fact that our optimization scheme produces scales that are similar to the physically realistic nucleobase/amino-acid binding affinity scales shows that our sampling is thorough and, more importantly, suggests that compositional matching of mRNA and protein sequence profiles may indeed be related to binding between them.
In addition, a total of 544 different one-dimensional scales [27,28], capturing different physicochemical properties of amino acids including size, hydrophobicity or interaction propensities, have been compared with the scales derived in this work by calculating the Pearson R between them. The result is visualized in Fig 5 in the case of the E. coli PUR scale. Here, it should be noted that the sign of correlation in this comparison depends only on the definition of a given scale and does not carry additional significance: for example, a hydrophilicity scale may be defined as a hydrophobicity scale and result in the same absolute result, but with an inverted sign. In general, the majority of the scales do not correlate closely with the MC-derived scales, but most of those that do are indeed in some way related to RNA/protein interactions or, interestingly, protein structural disorder. For example, the strongest correlation (R = 0.85) is obtained for the knowledge-based GUA-affinity scale [6], while the Woese pyridine affinity scale [29] ranks among the top 3% of all scales and exhibits an R = -0.63 with the MC-derived PUR-based scale. Parenthetically, a potential explanation for the observed lower density of Pearson Rs around 0 (Fig 5) may be that approximately 1/3 of all physical scales examined are related to amino-acid hydrophobicity [4]. Since our optimized PUR scale in general correlates negatively with hydrophobicity scales i.e. positively with hydrophilicity scales, this could create a somewhat lower density of Pearson Rs around 0.
In the present work, we have sampled the space of amino-acid scales using as a fitness function the proteome-average matching between the mRNA nucleobase density profiles and the scale-weighted sequence profiles of cognate proteins. Importantly, our framework disregards all available biological information except the sequences of the two biopolymers and approaches the amino-acid scales from an abstract perspective as arrays of 20 numerical weights that can be chosen arbitrarily. This provides the benefit that the physicochemical interpretation of the scales does not need to be given a priori. Rather, the features of the fitness landscape of scales provide the constraints that any physical scales need to fulfill in order to be consistent with cognate mRNA/protein sequence complementarity. At the same time, effects like codon-usage bias are included by default. Utilizing this method, we have shown that it is possible to identify scales, which are highly optimized to lead to pronounced complementarity, and that our method is highly efficient at doing so.
Notably, it was not a priori clear that a simple MC search would result in unique optimal solutions for the matching problem. A remarkable result that virtually one single scale is the sole result of numerous independent simulations suggests that the intrinsic features of the underlying fitness landscape guide the development of these affinity scales towards a narrow range of values. Against our expectations, especially considering the vast combinatorial space the scales reside in, evolution of highly optimized scales could be performed in less than 200 MC moves (S1 Fig). This result on its own provides some pertinent information about the space the scales reside in. Namely, such a rapid and reproducible convergence can only be explained by a landscape that is shaped like a funnel. In this picture, the continuous downhill slope guides the search reproducibly towards the final optimized scale. Although simulated annealing was included in the MC approach, this would not suffice to reach the exact same minimum in every MC run: if there existed another local minimum separated by a significant barrier, we would also have sampled it. A potential criticism of this interpretation could be that we start all our MC runs from the same scale (all weights equal to 0), which in principle could bias the final outcome. However, the MC runs initiated with different random number seeds decorrelate rapidly from each other in a few steps (as seen in the standard deviations in Fig 2B), only to converge again at the end of the runs. This, in effect, suggests that regardless of the starting point on the landscape, the MC searches reproducibly end up in the same minimum.
Our MC procedure has also resulted in a large number of suboptimal scales at all values of the average Pearson R between ~0.3 and -1. We have applied a hierarchical clustering algorithm to this data set and identified clusters of mutually similar scales. As reported in Fig 3, the development of the number of clusters as a function of the degree of matching exhibits a funnel-like shape. The important point to make is that the funnel is shallow and wide even up to the <R> values of -0.7 or higher. This means that there exist many rather different scales, which could lead to suboptimal, yet still relatively high levels of proteome-average complementary matching. This carries significant implications for our previous investigations of the complementarity hypothesis. As a case in point, our first report on the hypothesis involved a computationally derived scale of amino-acid affinity for PYR mimetics, which had led to a value of <R> of -0.74 across the human proteome [4]. This was interpreted as a strong signal of putative complementary interactions between mRNAs and their cognate proteins. On the other hand, our present analysis shows that there exist over 200 clusters of scales, whereby members of different clusters exhibit a correlation of at most 0.8 or less, all leading to a proteome-average correlation better than -0.74. In other words, the PYR-mimetic affinity scale used in our original study is by no means unique in its ability to lead to relatively high matching. While these findings do call for caution, it should be emphasized that they in no way contradict our previous interpretations. Namely, our present study only enumerates the list of possible scales that could lead to high complementarity. What is more, this list represents only a minor fraction of the space of all possible scales as indicated by our present and previous randomization studies [4,6,14]. The matching, in other words, is not a consequence of just any scale, but rather it can be achieved by only a select few, however mutually different from each other they may be.
The centerpiece of the present study is the comparison of the computationally-derived optimal scales with the published knowledge-based scales derived from structural data [6]. On the one hand, the computational scales are derived from a singular requirement that, when protein sequences are weighted by them, they yield profiles that resemble the cognate mRNA nucleobase-density profiles. Conversely, the knowledge-based scales are derived from the contact statistics of nucleobases and amino-acid side chains at the RNA-protein interfaces. They, therefore, report on the intrinsic binding preferences of the two sets of monomers. It is remarkable that Pearson correlation coefficients of up to 0.85 with highly significant p-values can be achieved between the two (Fig 4). The same can be said for the high overall p-values resulting from combining multiple scales. In other words, we start here from a simple computational exercise in which we search for scales that when applied to protein sequences yield profiles that match their cognate mRNA’s nucleobase density profiles. The fact that the scales obtained in this way resemble the nucleobase/amino-acid binding affinity scales strongly suggests that profile matching and binding indeed may be related, as put forth by the complementarity hypothesis. However, the rather weak correlation with some other affinity scales shall not be ignored. A question remains as to why protein affinity profiles for GUA strongly match the mRNA PUR-density profiles, but not as well the mRNA GUA-density profiles. In line with this, why are the protein ADE-affinity profiles inversely related to the mRNA PUR-density profiles? It has been suggested that the ADE and URA nucleobases may be newer additions to biological systems, while the GUA and CYT may have been the very first nucleobases adopted [5]. If both the matching behavior as postulated by the stereochemical hypothesis and the assumed timeline of RNA evolution hold true, this would suggest that the usage of ADE as an anti-matching base could have served a biologically important purpose. Namely, the presence of ADE has the potential to negatively regulate the level of complementarity and, therefore, the strength of binding between cognate partners, as previously suggested [5].
The present work accounts only for sequence data on both the mRNA and the protein side, with no secondary or higher-order structure elements being considered. This, of course, does not rule out structured mRNAs and proteins as interaction partners, but certainly limits the generality of the current work. It should, however, be noted that the binding between unstructured RNAs and intrinsically disordered proteins belongs to an important, large class of RNA-protein interactions and, moreover, provides a relevant context in which to look for cognate interactions [30–36]. In this regard, it may be potentially informative that some of the closest physically realistic scales to the optimal MC scales derived herein are linked with protein disorder (Fig 5) [37–43]. Finally, our results suggest that the degree of cognate mRNA-protein complementarity is heavily determined by the intrinsic binding affinities of just a handful of nucleobase/amino-acid pairs. For example, the opposite behavior of Phe and Glu/Lys define a large fraction of the PUR-density/PUR-affinity matching. At the same time, the nucleobase-binding preferences of other amino acids are much more ambiguous and diverse, even at relatively high levels of matching. The main biological implication of this is that it defines constraints on the mechanism by which genetic encoding could have evolved from binding, as proposed by the stereochemical hypothesis and our generalization of it. Specifically, the establishment of a coding relationship between codons and amino acids, which would be a consequence of complementary binding between cognate mRNA/protein pairs, is possible only for a narrow set of nucleobase-affinity values for several key residues, as defined by our study. Future research should provide information about this and other related open questions.
|
10.1371/journal.pbio.1002396 | An Ancient Fingerprint Indicates the Common Ancestry of Rossmann-Fold Enzymes Utilizing Different Ribose-Based Cofactors | Nucleoside-based cofactors are presumed to have preceded proteins. The Rossmann fold is one of the most ancient and functionally diverse protein folds, and most Rossmann enzymes utilize nucleoside-based cofactors. We analyzed an omnipresent Rossmann ribose-binding interaction: a carboxylate side chain at the tip of the second β-strand (β2-Asp/Glu). We identified a canonical motif, defined by the β2-topology and unique geometry. The latter relates to the interaction being bidentate (both ribose hydroxyls interacting with the carboxylate oxygens), to the angle between the carboxylate and the ribose, and to the ribose’s ring configuration. We found that this canonical motif exhibits hallmarks of divergence rather than convergence. It is uniquely found in Rossmann enzymes that use different cofactors, primarily SAM (S-adenosyl methionine), NAD (nicotinamide adenine dinucleotide), and FAD (flavin adenine dinucleotide). Ribose-carboxylate bidentate interactions in other folds are not only rare but also have a different topology and geometry. We further show that the canonical geometry is not dictated by a physical constraint—geometries found in noncanonical interactions have similar calculated bond energies. Overall, these data indicate the divergence of several major Rossmann-fold enzyme classes, with different cofactors and catalytic chemistries, from a common pre-LUCA (last universal common ancestor) ancestor that possessed the β2-Asp/Glu motif.
| Common descent is the hallmark of Darwinian evolution. Homology of biological traits, and particularly of protein sequences and structures, serves as an indication for divergence from a common ancestor and a means of assigning phylogenetic relationships. However, because of shared functional demands and chemical-physical constraints, proteins that evolved independently of one another often converge on very similar molecular traits, including structure and sequence. We tested the widely accepted hypothesis of common ancestry of several major enzyme classes, comprising hundreds of different families and using different cofactors and catalytic chemistries. Although they share the same overall architecture—the Rossmann fold—these enzymes show no significant sequence homology across different classes. We describe an analysis based on the omnipresence of a single residue across these classes: an acidic aspartate or glutamate residue that binds ribose, the common denominator of the different cofactors used by these enzymes. We show that Rossmann enzymes possess a unique interaction geometry that represents a fingerprint of common ancestry rather than an outcome of molecular constraint. We thus provide the first systematic test of divergence versus convergence of a highly abundant protein motif and assign common descent in one of the most ancient and functionally diverse protein folds.
| Nucleoside-based cofactors are widely abundant and are likely to have appeared well before proteins [1–3]. The early protein forms may have therefore evolved to bind and function with nucleoside-based cofactors [4]. However, tracing motifs that relate to the earliest stages of protein-cofactor evolution is a challenge [5]. Omnipresent cofactor-binding motifs, such as the P-loop (phosphate-binding loop or Walker A motif), are considered fingerprints of the earliest precursors of modern proteins [5]. However, in general, abundance of a trait per se (in terms of number of species and their distribution in the tree of life) is not sufficient to indicate common ancestry, as convergence of sequence and structure is a feasible alternative. The more minimal a motif is in terms of the number of amino acids, the more likely it is to be the outcome of convergent evolution—namely, to have evolved independently, along separate lineages, yet ended up with the same molecular solution [6]. In fact, there is ample evidence for convergence, both of structural architectures (folds) and of binding and catalytic motifs. Folds such as β-propellers, for example, have emerged in parallel many times [7–10]. Artificial proteins belonging to the most ancient folds are computationally designed with sequences that bear no relation to natural proteins [8,9]. Omnipresent catalytic motifs such as the Asp/Glu dyads of glycosyl hydrolase and transferases are seen in >50 different folds [11] and with no significant sequence homology beyond the dyad itself. Such motifs have probably emerged independently, and their conserved geometry is due to physicochemical constraints dictated by a shared function. In fact, when it comes to binding and catalytic motifs, convergence is probably as dominant as divergence [12]. Overall, differentiating divergent from convergent evolution remains a crucial, largely unresolved dilemma in evolutionary biology in general and in protein evolution in particular [13–16].
Our study focuses on the Rossmann fold. By virtue of catalyzing >300 different enzymatic reactions [17], the Rossmann fold is one of the most widely occurring protein folds [18–21] and is accordingly well represented in the presumed set of proteins that existed in the last universal common ancestor (LUCA) [20,22,23]. Belonging to the general class of β/α proteins, the Rossmann fold comprises two tandem repeats. Each repeat comprises three consecutive strands forming a parallel pleated sheet and two connecting α-helices [24–26]. The strand order along the core β-sheet is 3-2-1–4-5-6, although modifications of the last strand are often seen (Fig 1). Rossmann-fold enzyme families are also characterized by their use of cofactors [20,27,28] and in particular of nucleoside-containing cofactors that were present in the presumed “RNA world,” prior to the emergence of proteins [1,2]. Rossmann-fold enzymes therefore comprise a clear example of the evolutionary link between cofactors and their utilizing enzymes. Indications for pre-LUCA evolutionary links in the Rossmann fold have been noted that relate to nucleoside binding and the shared fold [19,29]. Shared nucleoside binding motifs have also been described upon the identification of the Rossmann fold and at later stages (e.g., [6,30–39]). Specifically, nicotinamide adenine dinucleotide (NAD)- and flavin adenine dinucleotide (FAD)-utilizing enzymes share a Gly-rich loop that resides between H1 and β1 and interacts with the cofactors’ phosphate moieties [19,40,41], and the hydroxyls of the cofactors’ ribose moiety typically interact with a Glu/Asp at the tip of β2 (β2-Asp/Glu; Fig 1) [42,43]. Sequence homology can obviously be detected between NAD and nicotinamide adenine dinucleotide phosphate (NADP) enzymes and may span over to FAD enzymes, specifically in relation to the above two motifs [44,45]. However, the sequence homology with other Rossmann classes such as S-adenosyl methionine (SAM)-dependent methyltransferases is much less clear [36,44]. The ribose-binding Glu/Asp at the tip of β2 has also been detected in methyltransferases [42,43]. However, the Gly-rich motif is not apparent in SAM-utilizing Rossmann enzymes, possibly because SAM does not contain phosphate groups. Consequently, some sequence-based classifiers, including those using sensitive homology detectors such as CATH (Class Architecture Topology Homologous superfamilies), define these classes as separate superfamilies [46]. However, based amongst other considerations on the shared β2-Asp/Glu motif, other classifiers such as ECOD (Evolutionary Classification of Protein Domains) [30] or Interpro [47] classify all three classes (NAD(P), FAD, and SAM-dependent Rossmann enzymes) in the same homology group [31,32,35,38,39].
Overall, a common fold [20] and the shared binding motif (the ribose β2-Asp/Glu interaction) are highly suggestive of a common Rossmann ancestor and specifically of common ancestry of NAD-, FAD-, and SAM-utilizing enzymes [30,34,38]. Indeed, these three classes (and a few additional ones addressed below) are all present in the presumed LUCA [48,49]. However, so far, there has been no attempt, to our knowledge, to examine whether these shared features are indeed a hallmark of common descent [39]. Such a systematic analysis is crucial in view of convergence being common and especially because the shared binding motif comprises a single residue.
We were initially interested in engineering the SAM-binding site of DNA methyltransferases—a Rossmann-fold enzyme superfamily. Our attention was focused on the adenosine group that appears in nearly all of the key enzymatic cofactors. In this context, we were searching for a highly conserved interaction that is critical to adenosine binding and could be modified. However, our analysis indicated that none of the residues that interact with the adenine ring are conserved in all DNA methyltransferases. In contrast, we observed that a Glu residue that interacts with the ribose is entirely conserved. We first observed that the carboxylate-ribose interaction is completely conserved in SAM-dependent methyltransferases, including DNA, RNA, protein, and small molecule methyltransferases. We realized that conservation does not simply concern an active-site Asp/Glu that interacts with SAM [42,43] but primarily relates to a bidentate interaction with the ribose’s 2ʹ and 3ʹ hydroxyls with an unusually narrow distribution of H-bond distances and angles. Distinctly, the interacting Asp/Glu is at the tip of the Rossmann’s second beta strand (β2) (Fig 2A; S1 Fig and S2 Fig). Further, although the β2-Asp/Glu was described as a characteristic of Rossmann NAD dehydrogenases [44], its bidentate nature has not been described as such.
A wider examination that further included NAD- and FAD-dependent oxidreductases was performed (see Methods and S3 Fig). This analysis confirmed that, as suggested earlier [40,41,50], the ribose-interacting Asp/Glu is also widely spread in these two enzyme classes. However, to our knowledge, the prevalence of this Asp/Glu interaction across NAD/FAD oxidoreductases, as well as SAM-dependent methyltransferases, and the geometrical conservation of the bidentate interaction with the bound ribose have not been previously noted. We therefore defined a new canonical Rossmann motif based on four criteria: (i) a tight, bidentate interaction exists between a carboxylate side chain and the ribose’s 2ʹ and 3ʹ-hydroxyls; (ii) the ribose’s furanose ring conformation is in an envelope form, mainly the E1 and 2E conformations (S4 Fig: see also S1 Text); (iii) the angle the ribose and the interacting carboxylate (hereafter the ribose–carboxylate angle α; defined in Fig 2B) is 90°–140°; and (iv) the interacting Glu/Asp is located at the tip of the β2 strand of the Rossmann fold (Fig 2A).
A systematic analysis identified the above motif features as being unique to the Rossmann fold. All nonredundant PDB structures containing ribose ligands were downloaded (Table 1; n = 2,949; S5 Fig). Of these, ~30% were found to have a carboxylate side chain that is within interacting distance (≤3.4 Å) of both the 2ʹ and 3ʹ hydroxyls of the ribose (n = 811). These structures were then categorized by the angle α (Fig 2B). The secondary structural element to which the interacting Glu/Asp residue belongs was also classified, as well as the fold (using Structural Classification of Proteins [SCOP] and/or CATH annotations). This analysis indicated that the canonical bidentate interaction underlies enzyme families and superfamilies that possess a Rossmann fold. Specifically, the canonical interaction was found in 54% of the structures classified as a Rossmann fold (Table 1). These structures were manually examined, and the order of their β-strands was found to fit the Rossmann-fold topology. Further, ≥96% of the examined Rossmann enzymes have their ribose rings in the 2E or E1 configuration (discussed below). Only 8% of the structures belonging to the Rossmann fold possessed noncanonical interactions—namely, bidentate interactions with α < 90° or > 140° and/or with the interacting Glu/Asp not being located at the tip of a β strand. Conversely, in enzymes belonging to non-Rossmann folds, monodentate or no Asp/Glu interactions are the rule (91%). Further, when bidentate interactions are present in non-Rossmann proteins, they almost never meet the canonical criteria, namely the canonical angle and the interacting Glu/Asp being at the tip of a β-strand. Indeed, amongst non-Rossmann enzymes, only 1.7% exhibit bidentate interactions that meet the canonical criteria versus 6% that exhibit bidentate interactions that do not meet the canonical criteria; Fig 2A–2C, S6 Fig).
One notable example showing how unique the canonical motif is to the Rossmann fold is the P-loop nucleoside-triphosphatase (NTPase) fold (CATH annotation 3.40.50.300; SCOP superfamily c.37.1, P-loop containing nucleoside triphosphate hydrolase). This fold also belongs to the class of β/α proteins. Overall, its topology is highly similar to the Rossmann fold, except that the order of strands within its core β-sheet is 2-3-1–4-5-6. Thus, the location of β2, where the canonical Rossmann Asp/Glu ribose-binding residue appears (Fig 1), is shifted relative to the Rossmann topology. We found that none of the structures belonging to the P-loop NTPases superfamily (CATH Family 3.40.50.300; n = 210) contains the canonical carboxylate-ribose interaction. Further, as discussed below, the mode of nucleoside binding in P-loop NTPases differs fundamentally from the one observed in the Rossmann fold.
Nearly half of the structures (279/578) in our original dataset were found to have the canonical carboxylate-ribose interaction but had no SCOP or CATH category (Table 1). We manually examined all 279 structures and found that 271 of these structures have a Rossmann, or Rossmann-like, topology, as defined above, and with the interacting Glu/Asp located at the tip of β2 (S5 and S6 Tables, S7 Fig). In fact, 108 out of the 279 structures that were not annotated in the CATH version v3.5.0 used to make our dataset are annotated in the current version (v.4.0.0; in which the number of annotated domains is larger by 36%). This “blind test” indicates that the applied criteria are sufficient not only to identify the canonical motif in Rossmann enzymes but also to rigorously identify a Rossmann enzyme merely by the existence of this canonical motif.
NAD-utilizing enzymes provide another indication for divergence from a common adenosine-binding ancestor. The cofactor NAD contains two riboses, one attached to adenosine and the other to nicotinamide. However, in the 259 available structures of NAD-dependent enzymes, only bidentate carboxylate-ribose interaction was found with the ribose. Among the NAD enzymes annotated as Rossmann, 145 structures out of 155 fit the canonical criteria with respect to the interaction with the adenosine’s ribose (S7 Table). Only four structures possess an additional bidentate interaction with NAD’s nicotinamide ribose. Of these four, two are annotated as Rossmann folds. Both these structures have one canonical interaction at the tip of β2 binding the adenosine ribose, as do the 145 other NAD Rossmann-fold enzymes. The nicotinamide riboses, however, interact with Glu residues located not at the tip of β2, and these bidentate interactions exhibit noncanonical geometries (Fig 3A and S8 Fig). The variability of the ribose-carboxylate angles and topology (Asp/Glu locations other than β2) and the sporadic presence (4/155 indicating appearance in recently evolved lineages) are all consistent with emergence by convergence. In contrast, the prevalence (145/155) and conservation of both geometry and topology of the interaction with the adenosine’s ribose most likely indicates divergence from a primordial ancestor of the Rossmann fold.
A motif that has been retained for ≥3.7 billion y of evolution is likely to be functionally important. Indeed, the contribution of the Glu/Asp interaction in NAD- and FAD-utilizing enzymes is widely recorded (published data listed in S8 Table) [51,52]. However, we could not find reports describing the experimental examination of its role in SAM-utilizing enzymes. To this end, we examined a typical bacterial mC5 DNA methyltransferase, M.HaeIII, in which Glu29 interacts with the SAM cofactor with the canonical motif geometry (Fig 4), as do nearly all other Rossmann methyltransferases (Table 1). Methylation activity was completely lost upon replacement of Glu29, including conservative replacements such as Gln, or Asp, and dropped by up to 450-fold in terms of kcat/KM in the Glu29Thr and Ala mutants (Fig 4, S8 Table). Overall, it appears that the canonical bidentate interaction have an important contribution to cofactor binding in the three classes of Rossmann enzymes in which it prevails, namely in NAD-, FAD-, and SAM-utilizing enzymes. However, the effects of mutations seemed to differ; for example, in glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (NAD dependent) and sarcosine oxidase (FAD dependent), the conservative D to E mutations reduced kcat/KM by ≤10-fold, whereas in M.HaeIII (SAM dependent), activity was completely lost. Thus, in all three enzymes, relatively conservative exchanges such as D to A or D to N resulted in up to 90-fold losses, yet the loss of activity observed for the SAM-dependent M.HaeIII was generally higher. The contribution of the bidentate interaction to SAM binding is probably higher than in the case of NAD and FAD because in the latter two, the Asp/Glu bidentate interaction is further away from the reaction center.
Is the highly conserved geometry of the Rossmann bidentate motif the outcome of chance or of necessity [54]? Namely, does the canonical geometry comprise the most optimal mode of ribose binding, or is it just one out of several options? Evolution of the Rossmann fold and cofactor binding implies that a single solution was selected at the ancestral stage, presumably owing at least in part to its favorable binding energy, and has been conserved ever since. Indeed, a scenario of divergence typically follows from the existence of several possible solutions; in particular, divergence of the bidentate carboxylate interaction geometries would seem to imply that there are multiple such geometries of similar energy. Convergence, on the other hand, is compatible with a scenario whereby the bidentate interaction geometry seen in existing proteins is the only optimal one or even the only possible one.
We can illustrate the above line of reasoning by considering the dihedral angles (ω) of the peptide bonds in proteins. The distribution of ω along >200,000 peptide bonds in known protein structures is narrow, with a clear maximum at planarity (>97% of bonds within ω = 180 ± 10°). This distribution corresponds to a single optimum value of 180° [55]. The planarity of the peptide bond therefore relates to a physical constraint that dictates all protein structures, rather than to a trait that diverged from the very first peptide. Another example mentioned in the introduction is the Asp/Glu dyads seen in glycosydases of many different folds, whereby the intercarboxylate distances are highly conserved within two categories of retaining glycosidases (5.5 Å) and inverting ones (10 Å) [11].
The favorable contribution of the bidentate carboxylate interaction to binding of vicinal-diols (as are the 2ʹ, 3ʹ hydroxyls of ribose) was indicated in small-molecule structures (S9 Fig) and by quantum mechanical calculations [56]. In the present work, we carried out new calculations to examine how energetically favorable is the geometry of the canonical interaction, and specifically how the energy of this interaction changes with the ribose-carboxylate angle (α) and ribose ring configuration. We performed quantum mechanical calculations designed to produce energy profiles of the different furanose configurations of ribose and of the ribose-carboxylate interaction angle (α) [57]. For this purpose, density functional theory electronic structure calculations with the Solvation Model based on Density (SMD) solvation model were used to study the ribose-carboxylate interaction in model systems in which the structures were energy minimized as a function of the ribose-carboxylate angle α (Fig 5; the energy calculations are described in detail in the S1 Text). The quantum mechanical calculations were performed on two models systems, M1 and M2, defined in Fig 5. After conformational searches, we identified the lowest-energy structures of model M2 (dubbed g-a, g-t, and t-t) and those for M1 (dubbed 2E-endo and 3E-exo). The lowest-energy structure obtained for M1 is 2E-endo, and for M2, it is t-t. Both 2E-endo and t-t exhibit a similar endo conformation, with respective α values of 132° and 129° and a similar envelope form for the ribose ring (2E for 2E-endo and E1 for t-t). The relative energy was accordingly plotted against the angle α (Fig 5A for model M1 and Fig 5B for model M2), indicating the lowest-energy structure for each value of α. These plots show that the bidentate interaction presents an angle optimum of ~130°. This optimum clearly overlaps the canonical Rossmann angle (Fig 2B). Further, the vast majority of Rossmann enzymes possess a ribose ring in a 2E or E1 configuration (96% of 263 PDB structures analyzed; see S1 Text) and an endo conformation (100% of 263 structures; see S1 Text), thus matching their modeled counterparts, 2E-endo and t-t.
However, beyond the canonical optimum, the potential energy surface for the carboxylate-bidentate interaction is relatively flat, with several minima. The only angles that appear to be highly disfavored are the edges, i.e., close to 0° and 180°, and these regions are also unoccupied in natural proteins (Fig 2B). Energy minima corresponding to the 3E-exo configuration for M1, and the g-a configuration for M2, are seen in α range of 10°–37° (Fig 5). According to our calculations, the endo configuration is more stable than the exo, by about 1 kcal/mol for model M1 and by only 0.1 kcal/mol for model M2. These differences are relatively small—an energy difference of 0.55 kcal/mol (the average difference for M1 and M2) corresponds to ~2.5-fold difference in affinity. For comparison, as indicated by the effects of mutations of the canonical Asp/Glu, the contribution of this interaction in Rossmann enzymes of different classes differs by well over 10-fold (see the above section and S8 Table).
The model structures that correspond to the alternative energy minima are seen in typical noncanonical interactions (Fig 2C, carboxyl side chains in variable greens). One characteristic example can be seen in Fig 3A, with the angles of the noncanonical interactions being 16°, far off the canonical range (90°–140°) and within the second predicted minimum (Fig 5). This alternative minimum corresponds to an exo disposition and has the ribose ring in the 3E for 3E-exo and in 2E for g-t. This mode is clearly seen in enzyme structures with the interaction angle in the range of 14° to 43° (Fig 2B and Fig 3), whereby the interaction corresponds to an exo configuration and the furanose conformation of the ribose is scattered among several possibilities (see S1 Text). Another example is human phosphoglyceraldehyde kinase where Glu344, located at the tip of β4, not β2, interacts with the ADP ribose in a bidentate manner, with the angle being 57° (S10 Fig).
Overall, the computations indicate that the canonical interaction is an intrinsically favorable mode for binding of ribose. It also corresponds to a furanose ring configuration that is the most energetically favored irrespective of the protein binding pocket and additional interactions, e.g., with the nucleoside’s base. However, the canonical interaction is only one out of at least two, if not more, favorable modes of bonding. Indeed, a wide distribution of interaction angles (Fig 2B) is seen in non-Rossmann ribose-binding proteins and predominantly in noncanonical interactions in Rossmann enzymes.
The utility of the carboxylate-ribose bidentate interaction, and its appearance in numerous protein families belonging to different folds and binding different cofactors, suggest that it arose independently, i.e., by convergent evolution. This is not surprising in view of the simplicity of this motif—a single carboxylate side chain aligned against the ribose hydroxyls. However, the statistics of occurrence clearly support the hypothesis of divergence. The canonical interaction is >30 times more frequent in Rossmann enzymes (54%) compared to non-Rossmann ones (1.7%). In contrast, the occurrence of noncanonical bidentate interactions in Rossmann and non-Rossmann proteins is nearly identical (8% and 6%, respectively; Table 1). Thus, whilst convergence to the canonical geometry and/or topology did occur, as exemplified in Fig 3B, its frequency of occurrence is not only lower but is also independent of the fold. The distinct features of convergence are apparent, including within Rossmann enzymes.
The distinct geometry of this motif in Rossmann enzymes may also provide a new means for automated classifications, as indicated by our manual examination of the structures with no CATH or SCOP annotations. The presence of an Asp/Glu at the loop connecting the second β-strand and the following helix is insufficient to distinguish between Rossmann from non-Rossmann enzymes (as previously noted [37,39] and also indicated by our data). However, when the carboxylate-ribose angle criterion is added, prediction accuracy increases to 97% (the false positive rate is 8/279).
The ancient origins of the ribose–(Asp/Glu-β2) motif and the claim for divergent evolution are also supported by the role of this motif in the switch of cofactor specificity of dehydrogenases. NADP-dependent dehydrogenases seem to have diverged from NAD-dependent enzymes [58], probably along multiple lineages. NADP differs from NAD in the 3ʹ-hydroxyl of the adenosine ribose being phosphorylated. Thus, binding of NADP is a priori excluded because of the negatively charged Glu/Asp that interacts with the unmodified ribose hydroxyls in NAD dehydrogenases. Indeed, the replacement of the β2-Asp/Glu is a prerequisite for the switch in specificity to NADP (S11 Fig) [59,60]. Thus, loss of the canonical Glu/Asp underlines the evolution of orthogonal, NADP-dependent dehydrogenases.
The existence of alternative ribose-binding modes with binding energies that are similar to that of the canonical Rossmann mode (Fig 5) and the accordingly wide distribution of binding modes of the noncanonical interactions (as reflected by the interaction angle α; Fig 2B) also support the hypothesis that the canonical Rossmann motif is the outcome of common ancestry and not of convergent evolution. Many structural features are the outcome of strict biophysical constraints, namely of one geometry being highly favored (a deep-well potential energy surface). The negative constraints (steric clashes, loss of resonance energy, etc.) are most dominant in dictating deep-well potentials. This is, for example, the case with the planarity of amide bonds [55]. In contrast, the multiminima potential energy surface for the carboxylate-ribose interaction indicates strong constraints acting only at the edges (around 0° and 180°; Fig 5). This suggests that the conservation of the interaction angle in Rossmann enzymes relates to their divergence from a common ancestor in which this angle was dictated by various factors, including but not limited to the favorable ribose-carboxylate interaction.
Common ancestry is the hallmark of Darwinian evolution. Our data support the notion of a primordial Rossmann ancestor in which binding of an adenosine-based cofactor was mediated by the ribose-β2-Asp/Glu interaction, alongside the Gly-loop that resides at the tip of the first strand (β1) (Fig 6, S13 Fig) [24,30,36,39]. The Gly-rich motif binds the phosphate groups of NAD/FAD/adenosine-5ʹ-triphosphate (ATP) (typically, GxGxxG) [5,61]. This motif is also recognizable in methyltransferases, although with low sequence identity because, unlike NAD- and FAD-dependent enzymes, their cofactor, SAM, does not contain a phosphate group (Fig 6). The minimal postulated ancestor therefore spans the Rossmann fold's first two strands and the connecting helix (β1-H1-β2) and includes the Gly-rich and ribose-β2-Asp/Glu interaction (Fig 7A) [40,62]. Our analysis supports a postulated pre-LUCA ancestor that underlined the divergence of at least three major enzyme classes: methyltransferases, NAD(P) and FAD oxireductases [29], and the many superfamilies belonging to these two classes, as well as the divergence of other enzyme families using other adenosine-based cofactors such as ATP (Fig 6). The Gly-rich loop and the ribose-β2-Asp/Glu motif was the keystone of this primordial ancestor [40,62]. Such keystone elements may relate to earlier precursors, possibly shorter polypeptides that contained these binding motifs [5,40,41,43,45] and from which the Rossmann ancestor evolved via a series of duplications, recombination, and fusions [63,64].
For the study of the individual enzyme classes, all structures belonging to SAM-dependent methyltransferases (SCOP category c.66.1), NAD(P)-binding Rossmann-fold domains (c.2.1), and FAD/NAD-linked oxidoreductases (c.3.1.5) were downloaded from SCOP (v.1.75). Redundant structures of the same protein in which the PDB code was the same for the first three letters/digits and the Glu/Asp residue number was identical were removed. Structures with <2.5 Å resolution were further considered, resulting in 55 methyltransferase (c.66.1) and 315 oxidoreductase (c.2.1 and c.3.15) enzyme domains that were assigned as Rossmann by SCOP (a flowchart describing this analysis is available as S3 Fig). For the systematic analysis of all ribose-binding proteins, we first identified 66 ribose-containing ligands (S2 Table) for which ≥10 nonredundant structures are available in the PDB. We excluded ligands that are part of polynucleotides such as RNA or DNA. All PDB structures that have ribose-containing ligands and <2.5 Å resolution were downloaded, and 80% sequence redundancy was removed with cd-hit [71]. The final dataset comprised 2,949 structures (Table 1) comprising 210 P-loop NTPase structures, 2,313 structures containing ligands with one ribose ring, and 426 structures with ligands such as NAD or FAD that contain two riboses (a flowchart describing this analysis is available as S5 Fig). The four structures with NAD ligands and two bidentate interactions were analyzed separately.
We calculated the distances, angles, and dihedral angles of atoms of interest using the PDB coordinates and custom Perl-scripts. For all retrieved PDB structures, the first chain in the asymmetric unit containing the cofactor was extracted. A random sample indicated that the variability in the distances and angles between different molecules in the asymmetric unit is low, and hence, an arbitrary choice of the first chain containing the cofactor is representative (S1 Text; average standard deviation for the distance is 0.074 Å, and for α is 2.2°). First, all residues that bind the ribose ligands were determined using CSU, and based on whether there is an Asp/Glu residue in the vicinity of the 2’, and 3’-OH of the ribose (≤4 Å). Then, we further characterized the ribose-Asp/Glu interaction and defined four binding modes: canonical bidentate, noncanonical bidentate, monodentate, or “no Asp/Glu interaction.”
The canonical bidentate interaction was defined by four criteria:
Noncanonical bidentate interaction was assigned to structures meeting criterion (i), namely structures with a bidentate interaction yet with the plane angle being <90° or >140° and the interacting Asp/Glu not located at the tip of a β-strand.
Monodentate interactions were assigned to structures with a single putative H-bond interaction between an Asp/Glu carboxylate and either the 2ʹ or the 3ʹ-hydroxyl groups. A more generous cutoff distance of ≤4 Å was taken here than for the bidentate interactions (≤3.4 Å) because the latter, and especially the canonical bidentate interactions, tend to be much tighter (average distance = 2.7 Å; S2B Fig). Finally, no Glu/Asp interaction was ascribed to structures where no carboxylate was found within 4 Å of either the 2ʹ or the 3ʹ-hydroxyl groups of the bound ribose.
When available, we retrieved the CATH and SCOP classification for the PDB structures in our dataset. Assignments of Rossmann fold were derived from CATH topology 3.40.50 (CATH_v3.5.0, version date: 20.09.2013, was used for this analysis). However, as explained in the main text, we separately analyzed superfamily 3.40.50.300, the P-loop containing nucleotide triphosphate hydrolases that are usually not considered as Rossmann. For SCOP, categories c.66.1, c.2.1, c.3.1, and c.4.1 were assigned as Rossmann. Including both CATH and SCOP databases significantly increased the fraction of structures with annotated fold (e.g., for structures containing one ribose ligands, the CATH database assigns 207 proteins as Rossmann, and addition of SCOP added another 85). About 46% of structures had neither a CATH nor a SCOP annotation (1,354/2,949). We therefore manually inspected a randomly chosen subset of the structures that possess the canonical interaction. We confirmed these as belonging to the Rossmann fold by identifying the canonical 3-2-1-4-5-6 topology of β-strands, or as Rossmann-like by identifying structures in which the last β strand (β6) is missing (S5 Table).
A variant of M.HaeIII containing four stabilizing mutations and with wild-type-like activity was the starting point for generating the Glu29 mutants [72]. The pASK-IBA3+vector (IBA, ampicillin resistance) plasmid containing the gene for the stabilized M.HaeIII was used as a template for PCR amplification. Mutants in position 29 were constructed by site-directed mutagenesis. The Glu codon was replaced with the Gln codon (CAA), Thr codon (ACC), Leu codon (CTG), Asp codon (GAT), Trp codon (TGG), Ala codon (GCG), Val codon (GTG), or Ser codon (AGC). The mutant encoding plasmids were transformed into E. coli MC1061, [mcrA0 relA1mcrB1 hsdR2 (r-m+; in which DNA methylation is not toxic) bearing the GroEL/ES encoding plasmid pGro7 (chloramphenicol resistance; Takara) to assist the folding of compromised mutants [72]. Transformants were selected by growth in the presence of ampicillin and chloramphenicol. The methyltransferase activity was tested by treatment of the extracted plasmid with the cognate restriction enzyme, HaeIII. The level of plasmid protection by virtue of methylation by M.HaeIII was determined by gel analysis. Bacteria were grown with no inducer or under induction (0.2 μg/ml anhydrotetracycline) and with 0.05% arabinose for induction of GroEL/ES expression. Wild-type M.HaeIII gave full protection even when basally expressed (no inducer). Time-dependent in vitro methylation assays were performed with purified enzyme variants (0.1–8 μM) essentially as described [73], using H3-labeled SAM (0.1–8 μM) and DNA substrate carrying nine methylation GGCC sites per molecule at 2.5 nM.
We carried out quantum mechanical electronic structure calculations on models M1 and M2 (S1 Text) by using the M06-2X/6-31+G(d,p) [74,75] model chemistry including the effect of aqueous solvent by using the SMD solvation model [76]. All electronic structure calculations were performed with Gaussian09 [77]. We performed an exhaustive conformational search for model M1 (Fig 4A). Starting from the lowest-energy optimized structures obtained with model M1, namely 2E-endo and 3E-exo, we carried out a relaxed potential energy surface scan along the coordinate defined by α (see Fig 5A). In the scan, all degrees of freedom were optimized with the exception of the angle α. This was accomplished by interfacing the Gaussian 09 program49 with a utility program we wrote that allows a constraint on the angle between two vectors. For model M2 (Fig 5B), after carrying out a conformational analysis of the molecule of adenosine and an analysis to find the best conformations that lead to a double hydrogen bond with a molecule of acetate, three fully optimized structures of model M2, denoted as g-t, g-a, and t-t, were found. These structures were taken as initial geometries to explore the potential energy surface (PES). The PES was explored by a combination of successive relaxed energy minimization scans along two angles and a dihedral angle that equals to perform a scan along the angle α (see S1 Text).
|
10.1371/journal.pgen.1004920 | Century-scale Methylome Stability in a Recently Diverged Arabidopsis thaliana Lineage | There has been much excitement about the possibility that exposure to specific environments can induce an ecological memory in the form of whole-sale, genome-wide epigenetic changes that are maintained over many generations. In the model plant Arabidopsis thaliana, numerous heritable DNA methylation differences have been identified in greenhouse-grown isogenic lines, but it remains unknown how natural, highly variable environments affect the rate and spectrum of such changes. Here we present detailed methylome analyses in a geographically dispersed A. thaliana population that constitutes a collection of near-isogenic lines, diverged for at least a century from a common ancestor. Methylome variation largely reflected genetic distance, and was in many aspects similar to that of lines raised in uniform conditions. Thus, even when plants are grown in varying and diverse natural sites, genome-wide epigenetic variation accumulates mostly in a clock-like manner, and epigenetic divergence thus parallels the pattern of genome-wide DNA sequence divergence.
| It continues to be hotly debated to what extent environmentally induced epigenetic change is stably inherited and thereby contributes to short-term adaptation. It has been shown before that natural Arabidopsis thaliana lines differ substantially in their methylation profiles. How much of this is independent of genetic changes remains, however, unclear, especially given that there is very little conservation of methylation between species, simply because the methylated sequences themselves, mostly repeats, are not conserved over millions of years. On the other hand, there is no doubt that artificially induced epialleles can contribute to phenotypic variation. To investigate whether epigenetic differentiation, at least in the short term, proceeds very differently from genetic variation, and whether genome-wide epigenetic fingerprints can be used to uncover local adaptation, we have taken advantage of a near-clonal North American A. thaliana population that has diverged under natural conditions for at least a century. We found that both patterns and rates of methylome variation were in many aspects similar to those of lines grown in stable environments, which suggests that environment-induced changes are only minor contributors to durable genome-wide heritable epigenetic variation.
| Differences in DNA methylation and other epigenetic marks between individuals can be due to genetic variation, stochastic events or environmental factors. Epigenetic marks such as DNA methylation are dynamic; they can be turned over during mitosis and meiosis or altered by chromatin remodeling or upon gene silencing caused by RNA-directed DNA methylation (RdDM). Moreover, changes in DNA sequence or structure caused by, for instance, transposable element (TE) insertion, can induce secondary epigenetic effects at the concerned locus [1], [2], or, via processes such as RdDM, even at distant loci [3]–[5]. The high degree of sequence variation, including insertions/deletions (indels), copy number variants (CNVs) and rearrangements among natural accessions in A. thaliana provides ample opportunities for linked epigenetic variation, and the genomes of A. thaliana accessions from around the globe are rife with differentially methylated regions (DMRs) [6]–[10], but it remains unclear how many of these cannot be explained by closely linked genetic mutations and thus are pure epimutations [11] that occur in the absence of any genetic differences.
The seemingly spontaneous occurrence of heritable DNA methylation differences has been documented for wild-type A. thaliana isogenic lines grown for several years in a stable greenhouse environment [12], [13]. Truly spontaneous switches in methylation state are most likely the consequence of incorrect replication or erroneous establishment of the methylation pattern during DNA replication [14]–[16]. A potential amplifier of stochastic noise is the complex and diverse population of small RNAs that are at the core of RdDM [17] and that serve as epigenetic memory between generations. The exact composition of small RNAs at silenced loci can vary considerably between individuals [13], and stochastic inter-individual variation has been invoked to explain differences in remethylation, either after development-dependent or induced demethylation of the genome [18], [19]. Such epigenetic variants can contribute to phenotypic variation within species, and epigenetic variation in otherwise isogenic individuals has been shown to affect ecologically relevant phenotypes in A. thaliana [20]–[22].
In addition to these spontaneous epigenetic changes, the environment can induce demethylation or de novo methylation in plants, for example after pathogen attack [23]. Recently, it has been proposed that repeated exposure to specific environmental conditions can lead to epigenetic differences that can also be transmitted across generations, constituting a form of ecological memory [24]–[27]. The responsiveness of the epigenome to external stimuli and its putative memory effect have moved it also into the focus of attention for epidemiological and chronic disease studies in animals [28], [29]. How the rate of trans-generational reversion among induced epivariants with phenotypic effects compares to the strength of natural selection, which in turn determines whether natural selection can affect the population frequency of epivariants, is largely unknown [30]–[33].
To assess whether a variable and fluctuating environment is likely to have long-lasting effects in the absence of large-scale genetic variation, we have analyzed a lineage of recently diverged A. thaliana accessions collected across North America. Using a new technique for the identification of differential methylation, we found that in a population of thirteen accessions originating from eight different locations and diverged for more than one hundred generations, only 3% of the genome had undergone a change in methylation state. Notably, epimutations at the DNA methylation level did not accumulate at higher rates in the wild as they did in a benign greenhouse environment. Using genetic mutations as a timer, we demonstrate that accumulation of methylation differences was non-linear, corroborating our previous hypothesis that shifts in methylation states are generally only partially stable, and that reversions to the initial state are frequent [12], [34]. Many methylation variants that segregated in the natural North American lineage could also be detected in the greenhouse-grown population, indicating that similar forces determined spontaneous methylation variation, independently of environment and genetic background. Population structure could be inferred from differences in methylation states, and the pairwise degree of methylation polymorphism was linked to the degree of genetic distance. Together, these results suggest that the environment makes only a small contribution to durable, trans-generationally inherited epigenetic variation at the whole-genome scale.
Previous studies of isogenic mutation accumulation (MA) lines raised in uniform greenhouse conditions identified many apparently spontaneously occurring pure epimutations [12], [13]. To determine whether variable and fluctuating environments in the absence of large-scale genetic variation substantially alter the genome-wide DNA methylation landscape over the long term, we analyzed a lineage of recently diverged A. thaliana accessions collected across North America. Different from the native range of the species in Eurasia, where nearly isogenic individuals are generally only found at single sites, about half of all North American individuals appear to be identical when genotyped at 139 genome-wide markers [35]. We selected 13 individuals of this lineage, called haplogroup-1 (HPG1), from locations in Michigan, Illinois and on Long Island, including pairs from four sites (Fig. 1A, S1 Table). Seeds of the accessions had been originally collected between 2002 and 2006 during the spring season, from plants at the end of their life cycle. Because rapid flowering in the greenhouse was dependent on an extended cold treatment, or vernalization, we conclude that the parental plants had germinated in autumn of the previous year and overwintered as rosettes. Climate data from the nearest respective weather station confirmed that precipitation and temperature regimes had varied considerably between sites in the growing season preceding collection (S1-S2 Fig.).
Whole-genome sequencing of pools of eight to ten siblings from each accession identified a shared set of 670,979 single nucleotide polymorphisms (SNPs) and 170,998 structural variants (SVs) relative to the Col-0 reference genome, which were then used to build a HPG1 pseudo reference genome (SOM: Genome analysis of HPG1 individuals; S2-S3 Table; S3 Fig.). Only 1,354 SNPs and 521 SVs segregated in this population (S4 Table, S4-S5 Fig.), confirming that the 13 strains were indeed closely related. Segregating SNPs were noticeably more strongly biased towards GC→AT transitions than shared SNPs, especially in TEs, although the bias was not as extreme as in the greenhouse-grown MA lines (Fig. 1B) [36]. A phylogenetic network and STRUCTURE analysis based on the segregating polymorphisms reflected the geographic origin of the accessions (Fig. 1A, C; S6 Fig.). Three of the pairs of accessions from the same site were closely related, and were responsible for many alleles with a frequency of 2 in the sampled population (Fig. 1D). If the spontaneous genetic mutation rate is similar to that seen in the greenhouse [36], the HPG1 accessions would be 15 to 384 generations separated from each other. With a generation time of one year, their most recent common ancestor would have lived about two centuries ago, which is consistent with A. thaliana having been introduced to North America during colonization by European settlers [37]. This is also in line with the fact that in several US herbarium collections, A. thaliana specimens from the mid-19th century can be found, among these specimens from the Eastern Seaboard and the Upper Midwest. We conclude that the HPG1 accessions constitute a near-isogenic population that should be ideal for the study of heritable epigenetic variants that arise in the absence of large-scale genetic change under natural growth conditions. Because we observed only a weak positive correlation between genetic distance and phenotypic differences in the greenhouse (S7 Fig.), we also infer that life history differences on their own should have little effect on the epigenetic landscape.
To assess the long-term heritable fraction of DNA methylation polymorphisms in the HPG1 lineage, we grew plants under controlled conditions for two generations after collection at the natural sites, before performing whole methylome bisulfite sequencing on two pools of 8–10 individuals per accession (S5 Table). We sequenced pools to reduce inter-individual methylation variation and fluctuations in methylation rate caused by stochastic coverage or read sampling bias. After mapping reads to the HPG1 pseudo reference genome, we first investigated epigenetic variation at the single-cytosine level. There were 535,483 unique differentially methylated positions (DMPs), with an average of 147,975 DMPs between any pair of accessions (SD = 23,745); thus, 86% of methylated cytosines accessible to our analyses were stably methylated across all HPG1 accessions. The vast majority of variable sites (97%) were detected in the CG context (CG-DMPs). As we have discussed previously [12], this can be largely attributed to the lower average CHG and CHH methylation rates at individual sites compared to CG methylation, whereby differences in methylation rates are smaller and statistical tests of differential methylation fail more often for CHG and CHH sites.. Additionally, stable silencing-associated methylation of repeats and TEs, elements rich in CHG and CHH sites, may contribute to this pattern. That only about 2% of all covered cytosines were differentially methylated in the relatively uniform HPG1 population contrasted with a previous epigenomic study, in which most cytosines in the genome were found to be differentially methylated in 140 genetically divergent accessions [10]. Fewer than 10% of all cytosines in the genome were never methylated across these 140 accessions, although most methylation events were confined to single strains (S9 Table of ref. [10]). To make our data more comparable to this other study [10], we identified DMPs of each HPG1 accession against the Col-0 reference genome. On average we found 383,237 DMPs per accession, affecting a total of 1,046,892 unique sites. We estimated that we would have detected 3.6 million DMPs, if we had sequenced 140 accessions from the HPG1 lineage (see Materials and Methods; S8 Fig.). The considerably larger number of DMPs in the 140 accessions [10] is likely due both to different methodology and to the higher degree of genetic variation between the analyzed accessions. For example, Schmitz and colleagues [10] did not directly test for differential methylation at individual sites nor did they apply multiple testing correction, which might contribute to the high number of CHH-DMPs reported in that study.
Using the geographic outlier LISET-036 as a reference strain, we found that 61% of CG-DMPs as well as 36% of the small number of CHG- and CHH-DMPs were present in at least two independent accessions (S9A Fig.), many of them shared between accessions from the same site. As is typical for A. thaliana [38], most methylated positions clustered around the centromere and localized to TEs and intergenic regions (Fig. 2A; S9B Fig.). In contrast, differential methylation in the CG context was over-represented on chromosome arms, localizing predominantly to coding sequences (Fig. 2A; S9B Fig.), similar to what we had previously observed in the greenhouse-grown MA lines [12].
We asked whether DMPs had accumulated more quickly in natural environments than in the greenhouse, using DNA mutations in the HPG1 and MA populations as a molecular clock (SOM: Estimating DMP accumulation rates). Our null hypothesis was that a variable and highly fluctuating natural environment increases the rate of heritable methylation changes. In contrast to this expectation, DMPs appear to have accumulated in sub-linear fashion in both the HPG1 and MA populations [12] (Fig. 2B) – with similar trends for DMPs in all three contexts – and the number of DMPs did not increase more rapidly in the HPG1 than in the MA lines. The steeper initial increase relative to SNP differences as well as the broader distribution of MA line differences relative to HPG1 differences were most likely the result of having compared individual plants in the MA experiment [12], rather than pools of siblings, as in the HPG1 experiment. The effect of pooling individuals, as shown by simulation (S10 Fig.), and a potentially higher genetic mutation rate in the wild than in the greenhouse, for example because of increased stress [39], could lead to a slight underestimation of the true HPG1 epimutation rate, but it remains unlikely that it greatly exceeds the one of the MA lines (SOM: Estimating DMP accumulation rates).
Because it is unclear whether variation at individual methylated cytosines has any consequences in plants, we next focused on differentially methylated regions (DMRs) in the HPG1 population. A limitation of previous plant methylome studies using short read sequencing has been that these relied on integration over methylated or single differentially methylated sites, or on the analysis of fixed sliding windows along the genome to identify DMRs. What appears intuitively to be more appropriate is to first identify regions that are methylated in individual strains (SOM: Differentially methylated regions) [40], and to test only these for differential methylation. We therefore adapted a Hidden Markov Model (HMM), which had been developed for segmentation of animal methylation data [41], to the more complex DNA methylation patterns in plants. We identified on average 32,529 methylated regions (MRs) per strain (median length 122 bp), with the unified set across all strains covering almost a quarter of the HPG1 reference genome, 22.6 Mb (Fig. 2A, C; S11A Fig.; S6 Table). MRs overlapping with coding regions were over-represented in genes responsible for basic cellular processes (p-value <<0.001), in agreement with gene body methylation being a hallmark of constitutively expressed genes [42]. Only 1% of mCHH and 2% of mCHG positions were outside of methylated regions (Fig. 2D), consistent with the dense CHH and CHG methylation found in repeats and silenced TEs [38]. Compared to mCGs within methylated regions, mCGs in unmethylated space localized almost exclusively to genes (94%), were spaced much farther apart, and were separated by many more unmethylated loci (Fig. 2E; S11B-C Fig.). This explains why sparsely methylated genes were under-represented in HMM-determined methylated regions, even though gene body methylation accounts for a large fraction of methylated CG sites. The accuracy of our MR detection method was well supported by independent methods (SOM: Validation of methylated regions).
Using the unified set of MRs, we tested all pairs of accessions for differential methylation, identifying 4,821 DMRs with an average length of 159 bp (S12 Fig.; S11A Fig.; S7 Table). Of the total methylated genome space, only 3% were identified as being differentially methylated, indicating that the heritable methylation patterns had remained largely stable in this set of geographically dispersed accessions. Indeed, 91% of genic and 98% of the TE sequence space were devoid of DMRs. Of the DMRs, 3,199 were classified as highly differentially methylated (hDMRs; S8 Table), i.e. they had a more than three-fold change in methylation rate and were longer than 50 bp. The DMR allele frequency spectrum was similar to that of variably methylated single sites (Fig. 2F). Most DMRs and hDMRs showed statistically significant methylation variation in only one cytosine context, often CG (Fig. 2G), even though DMRs were dominated by CHG and CHH methylation (Fig. 2D, S13 Fig.). Different from individual sites (DMPs), the densities for DMRs and hDMRs were highest in centromeric and pericentromeric regions, and overlapped more often with TEs than with genes (Fig. 2A, C). Relative to all methylated regions, genic regions were two-fold overrepresented in the genome sequence covered by DMRs, and three-fold in hDMRs (Fig. 2C). Currently, we do not know whether this simply reflects the greater power of detecting differential methylation at the typically more highly methylated CG sites compared to CHG or CHH sites, or whether this reflects actual biology.
DNA methylation in gene bodies has been proposed to exclude H2A.Z deposition and thereby stabilize gene expression levels [42]. We therefore asked what impact differential methylation had on transcriptional activity. We identified 269 differentially expressed genes across all possible pairwise combinations (S9-S10 Table), most of which were found in more than one comparison. When we clustered accessions by differentially expressed genes, closely related pairs were placed together (S14 Fig.). We identified 28 differentially expressed genes that overlapped with an hDMR either in their coding or 1 kb upstream region, but the relationship between methylation and expression was variable (S11 Table). By visual examination, we found not more than five instances of demethylation that were associated with increased expression; examples are shown in S15 Fig.
With the caveat that there are uncertainties about the genetic mutation rate in the wild, and therefore how the number of SNPs relates to the number of generations since the last common ancestor, there was no evidence for faster accumulation of variably methylated sites in the HPG1 population, nor for very different epimutation rates among HPG1 lines (Fig. 2B). Importantly, the overlap of differential methylation between the two populations was much greater than expected by chance: the probability of a random mC site in the MA population of being variably methylated in the HPG1 population was only 7%, but it was 41% among sites that were also variably methylated in the MA population – a six-fold enrichment (four-fold enrichment in the reciprocal comparison; Fig. 3A). In other words, almost half of the DMPs in the MA lines were also polymorphic in the HPG1 lines, and almost a third of HPG1 DMPs were also variably methylated in the MA population. These shared DMPs were more heavily biased towards the chromosome arms and towards genic sequences than population-specific epimutations (S16A-S16B Fig.). Conversely, DMPs unique to one population were more likely to be unmethylated throughout the other population when compared to random methylated sites (Fig. 3A), as one might expect for sites that sporadically gain methylation.
DMPs unique to the HPG1 lineage appeared to be less frequent in the pericentromere compared to MA- line-specific DMPs (S16A Fig.), which was also reflected in an apparently higher epimutation frequency in the MA lines for these regions (S16B Fig.). We therefore investigated whether the annotation spectrum differed between these two classes of differentially methylated sites. Even though MA-specific DMPs were more often found in TEs compared to HPG1-specific DMPs, this bias was also observed for all cytosines accessible to our methylome analyses (S16C Fig.), and can therefore be explained by a more accurate read mapping and better TE annotation in the Col-0 reference compared to the HPG1 pseudo-reference genome. Indeed, except for chromosome 4, the average sequencing depth in the pericentromere was higher in the MA lines (S16B Fig.).
DMPs distinguishing MA lines that were separated from each other by only a few generations were more frequently variably methylated in the HPG1 lineage than DMPs identified between distant MA lines (S17 Fig.). We interpret this observation as an indication of privileged sites that are more labile and therefore more likely to have already changed in status after a small number of generations.
We used the methods implemented for the HPG1 population to detect DMRs also in the MA strains. Similar to variable single positions, or DMPs, the overlap between 2,523 DMRs of the MA lines that we could map to the HPG1 methylated genome space with the 4,821 DMRs of the HPG1 accessions was greater than expected and highly significant (Ζ-score = 32.9; 100,000 permutations). HPG1 DMRs were four-fold more likely to coincide with MA DMRs than with a random methylated region from this set (Fig. 3B). We observed similar degrees of overlap independently of sequence context. Shared DMRs between both lineages were, in contrast to shared DMPs, not biased towards genic regions (S18 Fig.). Differentially methylated regions in the HPG1 lineage, however, overlapped with genic sequences more often than MA DMRs (S18 Fig.), which might again be explained by the different efficiencies in mapping to repetitive sequences and TEs (S16B Fig.).
We next wanted to know how this short-term variation compared to methylation variation across much deeper splits. To this end, we identified variably methylated regions between a randomly chosen MA line and a randomly chosen HPG1 line; these DMRs, which differentiate distantly related accessions, were also enriched in each of the two sets of within-population DMRs (MA or HPG1) (Fig. 3C). Finally, we compared DMRs found in the HPG1 population to DMRs that had been identified with a different method among 140 natural accessions from the global range of the species [10] (Fig. 3D). Although only 9,994, less than one fifth, of the variable regions from the global accessions were covered by methylated regions in the HPG1 strains, the overlap of DMRs was highly significant (Ζ-score = 19.8; 100,000 permutations). Together, the high recurrence of differentially methylated sites and regions from different datasets points to the same loci being inherently biased towards undergoing changes in DNA methylation independently of genetic background and growth environment.
To explore potential sources of such lability, we compared variation in the HPG1 lines to that caused by mutations in various components of epigenetic silencing pathways [43]. Almost all variable sites and regions in CG-methylated parts of the HPG1 genome were hypomethylated in mutants deficient in DNA methylation maintenance, most notably in the met1 single and the vim123 triple mutants (S19 Fig.). This is consistent with polymorphic methylation arising primarily because of errors in the maintenance of symmetrical CG methylation during DNA replication. Hypermethylated sites in the rdd triple mutant, which shows impaired demethylation, were also found slightly more often within variably methylated regions of all contexts (S19D Fig.).
To quantify how many methylation differences were co-segregating with genome-wide genetic changes at both linked and unlinked sites, we estimated heritability for each highly differentially methylated region by applying a linear mixed model-based method. We used segregating sequence variants with complete information as genotypic data and average methylation rates of hDMRs with complete information as phenotypes. The median heritability of all hDMRs was 0.41 (mean 0.44), which means that genetic variance across the entire genome contributed less than half of methylation variance (Fig. 4A). hDMRs in the HPG1 strains that were not methylated in the greenhouse-grown MA lines had a higher median heritability, 0.48, than HPG1 hDMRs also found among MA DMRs (0.29), which held true for all sequence contexts (Fig. 4A; S20 Fig.). Regions of highly differential methylation found only in the HPG1 population, especially those in unmethylated regions of the MA lines, were thus more likely to be linked to whole-genome sequence variation than hDMRs found in both populations. For 19% of all hDMRs (21% CG-hDMRs, 14% CHG-hDMRs, 7% CHH-hDMRs), the whole-genome genotype explained more than 90% of their methylation differences (with a standard error of at most 0.1). Of these hDMRs, half had a heritability of greater than 0.99. That 6.7% of the sequence space of these heritable hDMRs still overlapped with MA DMRs (versus 9.4% for the less heritable hDMRs) was in agreement with the hypothesis that there are regions that vary highly in their methylation status independently of genetic background.
To identify genetic variants that potentially directly cause methylation changes in their local genomic neighborhood, we focused on variably methylated regions that were within 1 kb of segregating SNPs or indels. Of 191 such DMRs, only three showed a systematic correlation with nearby sequence polymorphisms. We noticed, however, that coding regions with structural variants larger than 20 bp that distinguished the MA and HPG1 populations were more likely to be methylated in both lineages than non-polymorphic coding regions (Fig. 4B). Consequently, DMPs unique to the HPG1 lines were on average closer to insertions or deletions than DMPs shared between the HPG1 and MA populations (Fig. 4C).
Lastly, we asked whether the genome-wide methylation pattern reflected genetic relatedness, i.e., population structure. Hierarchical clustering by methylation rates of variable sites and regions grouped strains by sampling location (Fig. 4D, E). This result was largely independent of the sequence or the annotation context of these loci, and not seen with sites that our statistical tests had identified as stably methylated (S21 Fig.). That variably methylated regions grouped the accessions similar to DMPs, albeit with less confidence (shorter branch lengths; S21 Fig.), suggested that our DMR calling algorithm was conservative. Methylation data thus paralleled similarity between accessions at the genetic level, in agreement with the interpretation that methylation differences primarily reflect the number of generations since the last common ancestor.
We have tested the hypothesis that accumulation of epigenetic variation under natural conditions proceeds over the short term in a very different manner than the clock-like behavior of genetic variation [24]–[27]. To this end, we have taken advantage of a unique natural experiment, the A. thaliana HPG1 lineage, which has likely diverged for at least a century throughout North America. Our analyses have revealed little evidence for broad-scale and durable epigenetic differentiation that might have been induced by the variable and fluctuating environmental conditions experienced by the HPG1 accessions since they separated from each other. While the exact conditions these plants have been subjected to since their separation from a common ancestor remain unknown, the time scale and diversity of geographic provenance are strong indicators of the variability of the environment between the different sampling sites, supported by temperature and precipitation data from nearby weather monitoring stations. The general analytical framework enabled by the HPG1 lineage – nearly isogenic lines grown for more than a century under variable and fluctuating conditions – could not have been achieved in a controlled greenhouse experiment.
Studies of epiRIL populations have shown that pure epialleles can be stably transmitted across several generations [5], [19], but how often this is the case for environmentally induced epigenetic changes has been heavily debated [33], [44]–[46]. The recent excitement about the transmission of induced epigenetic variants comes from such variants having been proposed to be more often adaptive than random genetic mutations [24]–[26]. Contrary to the expectations discussed above, we found that epimutation rates under natural growth conditions at different sites did not differ substantially from those observed in a controlled greenhouse environment, with polymorphisms accumulating sub-linearly in both situations, apparently because of frequent reversions. Note that we grew the HPG1 plants under controlled conditions for two generations after sampling at the natural site, to reduce the range of epigenetic variation to the long-term heritable fraction. Given that the environment can induce acute methylation changes [23], [47], it is likely that we would have observed greater epigenetic variation, if we had sampled field-grown individuals directly. However, most of such variation induced during ontogeny does not appear to be heritable, as we did not find evidence for it after two extra generations in the greenhouse. Additional studies that directly compare plants grown outdoors to their progeny grown in a stable and controlled environment will help to further clarify this issue.
We found that positions of differential methylation in the HPG1 population are more likely to overlap with DMPs detected between closely related MA lines than between more distantly related MA lines. This observation supports the hypothesis that there are different classes of polymorphic sites. One of these includes ‘high lability’ sites that are independent of the genetic background, that change with a high epimutation rate, and that are therefore more likely to appear in each population. Another class of DMPs comprises more stable sites that gain or lose methylation more slowly and that therefore are less likely to be shared between different populations.
Differences between accessions in terms of DNA methylation recapitulated their genetic relatedness, further corroborating our hypothesis that heritable epigenetic variants arise predominantly as a function of time rather than as a consequence of rapid local adaptation. Epigenetic divergence thus does not become uncoupled from genetic divergence when plants grow in varying environments, nor does the rate of epimutation noticeably increase. A minor fraction of heritable epigenetic variants may be related to habitat, which could be responsible for LISET-036 being epigenetically a slight outlier (Fig. 4E), even though it is not any more genetically diverged from the most recent common ancestor of HPG1 than other lines. Such local epigenetic footprints could also explain fluctuations in epimutation frequency between the MA and HPG1 lineages. Subtle adaptive changes at a limited number of loci would go unnoticed in the present analysis of genome-wide patterns and can therefore not be excluded. However, on a genome-wide scale there was little indication of adaptive change: neither were LISET-036 specific regions of differential methylation in and near genes enriched for GO terms with an obvious connection to environmental adaptation, nor were there overlapping differentially expressed genes (S22 Fig., SOM: Analysis of LISET-036 specific hDMRs). In combination with the general lack of correlation between differential methylation and changes in gene expression, our findings suggest that epigenetic changes in nature are mostly neutral, and thus comparable to genetic mutations. We point out that an annual species such as A. thaliana might be differently disposed to record environmental signals in its epigenome compared to more long-lived species. From an evolutionary perspective, in perennial species the advantage of epigenetically mediated local adaptation to changing conditions could be more pronounced, and future studies are warranted to address this question.
Because of the near-isogenic background of the HPG1 accessions, we were also able to gauge how much of epigenetic variation is either caused by, or stably co-segregates with genetic differences. HPG1-specific highly differentially methylated regions were more often linked to genotype variation than regions that were variably methylated in both the HPG1 and MA populations. This suggests that heritable hDMRs can, to a certain extent, be considered facilitated epigenetic changes [11].
Both differentially methylated regions and positions are over-represented in genes, but TEs and intergenic regions contain many variable regions and only very few variable single sites. Altogether our data indicate that both variably and constitutively methylated positions in genes are typically separated by many unmethylated sites and that a large fraction of these is therefore not classified as being (differentially) methylated. Variability of DNA methylation in plant genes thus mainly affects single, sparsely distributed cytosines, the biological relevance of which remains unclear.
Our comparisons between MA laboratory strains and natural HPG1 accessions have revealed that loci of variable methylation overlapped much more between the two groups than expected by chance, despite these populations having experienced very different environments that also differ greatly in their uniformity, and despite completely different genetic backgrounds. The observation that changes at many sites and loci are independent of the genetic background and geographic provenance suggests that spontaneous switches in methylation predominantly reflect intrinsic properties of the DNA methylation and gene silencing machinery, with the CG maintenance system seemingly being the most error-prone. Our most important finding is probably that DNA methylation is highly stable across dozens, if not hundreds of generations of growth in natural habitats; 97% of the total methylated genome space was not contained in a DMR. The stark contrast to published data, which describes more than 90% of cytosines in the genome as variably methylated in a set of 140 divergent natural accessions [10], can be explained both by the low amount of genetic divergence among the HPG1 accessions and by methodological differences. For future studies, we recommend the application of non-permissive statistical tests in the analysis of differential methylation. The overall stability of methylation presented here is in accordance with the high similarity of methylation in evolutionarily conserved gene sequences [48]. It contrasts, however, with our recent report showing that over longer evolutionary distances that separate species in the same genus or closely related genera, there is very little conservation of global DNA methylation, simply because the sequences that are typically methylated are much more evolutionarily fluid than non-methylated sites [47]. In summary, we propose that the stability of DNA methylation first and foremost depends on the stability of the underlying genetic sequence and that heritable polymorphisms that arise in response to specific growth conditions appear to be much less frequent than those that arise spontaneously. These conclusions are of importance when considering epimutations as a potential evolutionary force.
Accessions [35] were collected in the field at locations indicated in S1 Table. Seeds had been bulked in the Bergelson lab at the University of Chicago before starting the experiment. Plants were then grown at the Max Planck Institute in Tübingen on soil in long-day conditions (23°C, 16 h light, 8 h dark) after seeds had been stratified at 4°C for 6 days in short-day conditions (8 h light, 16 h dark). We grew one plant of each accession under these conditions; seeds of that parental plant were then used for all experiments. Eight plants of the same accession were grown per pot in a randomized setup. All accessions used in this paper have been added to the 1001 Genomes project (http://1001genomes.org) and have been submitted to the stock center.
DNA was extracted from rosette leaves pooled from eight to ten individual adult plants. Plant material was flash-frozen in liquid nitrogen and ground in a mortar. The ground tissue was resuspended in Nuclei Extraction Buffer (10 mM Tris-HCl pH 9.5, 10 mM EDTA, 100 mM KCl, 0.5 M sucrose, 0.1 mM spermine, 0.4 mM spermidine, 0.1% β-mercaptoethanol). After cell lysis in nuclei extraction buffer containing 10% Triton-X-100, nuclei were pelleted by centrifugation at 2000 g for 120 s. Genomic DNA was extracted using the Qiagen Plant DNeasy kit (Qiagen GmbH, Hilden, Germany). Total RNA was extracted from rosette leaves pooled from eight to ten individual adult plants using the Qiagen Plant RNeasy Kit (Qiagen GmbH, Hilden, Germany). Residual DNA was eliminated by DNaseI (Thermo Fisher Scientific, Waltham, MA, USA) treatment.
DNA libraries for genomic and bisulfite sequencing were generated as described previously [12]. Libraries for RNA sequencing were prepared from 1 µg of total RNA using the TruSeq RNA sample prep kit from Illumina (Illumina) according to the manufacturer's protocol.
All sequencing was performed on an Illumina GAII instrument. Genomic and bisulfite-converted libraries were sequenced with 2×101 bp paired-end reads. For bisulfite sequencing, conventional A. thaliana DNA genomic libraries were analyzed in control lanes. Transcriptome libraries were sequenced with 101 bp single end reads. Four libraries with different indexing adapters were pooled in one lane; no control lane was used. For image analysis and base calling, we used the Illumina OLB software version 1.8.
The SHORE pipeline v0.9.0 [49] was used to trim and quality-filter the reads. Reads with more than 2 (or 5) bases in the first 12 (or 25) positions with a base quality score of less than 4 were discarded. Reads were trimmed to the right-most occurrence of two adjacent bases with quality values equal to or greater than 5. Trimmed reads shorter than 40 bases and reads with more than 10% (of the read length) of ambiguous bases were discarded.
Reads were aligned against the Arabidopsis thaliana genome sequence version TAIR9 in iteration 1 and against updated “Haplogroup 1-like” genomes in further iterations. The mapping tool GenomeMapper v0.4.5s [50] was used, allowing for up to 10% mismatches and 7% single-base-pair gaps along the read length to achieve high coverage. All alignments with the least amount of mismatches for each read were reported. A paired-end correction method was applied to discard repetitive reads by comparing the distance between reads and their partner to the average distance between all read pairs. Reads with abnormal distances (differing by more than two standard deviations) were removed if there was at least one other alignment of this read in a concordant distance to its partner. The command line arguments used for SHORE are listed in S1 File.
Genetic variants were called in an iterative approach. In each step, SNPs and structural variants common to all strains were detected and incorporated into a new reference genome. The thus refined “HPG1-like” genomes served as the reference sequence in the subsequent iterations (S3 Fig.). We performed three iterations to call segregating variants and built two reference genomes to retrieve common polymorphisms. The steps performed in each iteration are described in the following paragraphs.
Base counts on all positions were retrieved by SHORE v0.9.0 [49] and a score was assigned to each site and variant (SNP or small indel of up to 7% of read length) depending on different sequence and alignment-related features. Each feature was compared to three different empirical thresholds associated with three different penalties (40%, 20% and 5% reduction of the score, initial score: 40). They can be found in S13 Table.
For comparisons across lines, positions were accepted if at most one intermediate penalty on their score was applicable to at least one strain (score ≥32). In this case, the threshold for the other strains was lowered, accepting at most one high and two intermediate penalties (score ≥15). In this way, information from other strains was used to assess sites from the focal strain under the assumption of mostly conserved variation, allowing the analysis of additional sites.
Only sites sufficiently covered (≥5x) and with accepted base calls in at least half of all strains (≥7 out of 13) were processed further. Variable alleles with a frequency of 100% were classified as "common" and variants with a lower frequency as "segregating".
Additional SNPs were called using the targeted de novo assembly approach described below.
Although a plethora of SV detection tools have been developed, the predicted variants show little overlap between each other on the same data sets. Furthermore, the false positive rate of many methods can be drastic [51]. Hence, rather than taking the intersection of the output from different tools, which would yield only a small number of SVs, we combined different tools and applied a stringent evaluation routine to identify as many true SVs as possible. Since SVs common to all strains should be incorporated into a new reference, only methods that identify SVs on a base pair level could be used. Currently, there are four different SV detection strategies (based on depth of coverage, paired-end mapping, split read alignments or short read assembly, respectively). Only tools based on split read alignments and assemblies are capable of pinpointing SV breakpoints down to the exact base pair. Programs that were used include Pindel v2.4t [52], DELLY v0.0.9 [53], SV-M v0.1 [54] and a custom local de novo assembly pipeline targeted towards sequencing gaps (described below). We reported deletions and insertions from all tools, and additionally inversions from Pindel. DELLY combines split read alignments with the identification of discordant paired-end mappings. Thus, our SV calling made use of three out of four currently available methodologies.
Reads for DELLY were mapped using BWA v0.6.2 [55] against the TAIR9 Col-0 reference genome to produce a BAM file as DELLY's input format.
The arguments for the command line calls of all tools are listed in S1 File.
While using a re-sequencing strategy, there are regions without read coverage (“sequencing gaps”) because either the underlying sequence is being deleted in the newly sequenced strain, or highly divergent to the reference sequence, or present in the focal strain, but not represented in the read set. To access sequences in the first two classes, a local de novo assembly method was developed.
Insertion breakpoints or small deletions, however, can mostly not be detected by zero coverage due to reads ranging with a few base pairs into or beyond the structural variants. Therefore, we defined a “core read region” as the read sequence without the first and last 10 nucleotides. To be able to assemble the latter cases, the definition of “sequencing gaps” was extended from zero-covered regions to stretches not spanned by a single read's core region.
All reads aligned to the surrounding 100 nucleotides of such newly defined sequencing gaps as well as the unmappable reads from the re-sequencing approach together with their potential mapped partners constituted the assembly read set. Two assembly tools were used to generate contigs, SOAPdenovo2 v2.04 [56] and Velvet v1.2.0 [57] (command line arguments in S1 File). Contigs shorter than 200 bp were discarded. To map the remaining contigs of each assembler against the iteration-specific reference genome, their first and last 100 bp were aligned with GenomeMapper v0.4.5s [50], accepting a maximal edit distance of 10. If both contig ends mapped uniquely within 5,000 bp, the thus framed region on the reference was aligned to the contig using a global sequence alignment algorithm after Needleman-Wunsch (‘needle’ from the EMBOSS v6.3.1 package). In addition, non-mapping contigs were aligned with blastn (from the BLAST v2.2.23 package) [58] to yield even more variants.
All differences between contig and reference sequences were parsed (including SNPs, small indels and SVs) for each assembly tool. Only identical variants retrieved from both assemblers were selected.
For each strain, all variants from the SV tools and the de novo assemblies were consolidated (S3A Fig.) and positioned to consistent locations to be comparable using the tool Dindel v1.01 [59]. In the case of contradicting or overlapping variants, identical variants (having the same coordinates and length after re-positioning) predicted by a majority of tools were chosen and the rest discarded, or all were discarded if there was no majority.
Despite sequencing errors or cross-mapping artifacts of the re-sequencing approach, genomic regions covered by reads are generally trusted. Chances of long-range variations in the inner 50% of a mapped read's sequence (“inner core region” of a read) are assumed to be low, since gaps would deteriorate the alignment capability towards the ends of the read.
Therefore, we filtered out variants from the consolidated variant set spanning a genomic region already covered by at least one inner core region of a mapped read of the corresponding strain (S3A Fig.), assuming homozygosity throughout the genome. This “core read criterion” had to be fulfilled at each genomic position spanned by the variant.
Variants passing the core read filter in all strains were classified as common variants and were incorporated into the reference sequence of the previous iteration, thus replacing the reference allele. Segregating variants, which could not be detected in all strains, were additionally built into the reference in separate “haplotype regions” (or “branches” of the reference sequence) to eventually be able to assess whether reads preferentially mapped to the reference or the alternative haplotype sequence (S3A Fig.). Linked variant haplotypes of a strain (distance between consecutive variants ≤107 bp, the maximal possible span of a read on the reference) as well as identical haplotype regions among strains were merged into one branch sequence.
For each strain, all reads were re-mapped to this new reference sequence yielding read counts at the variant site on each branch (rb) and at the corresponding site on the reference haplotype sequence (rref) (S3A Fig.). Here, the read count of a site was defined as the number of inner core regions spanning the site. To increase certainty of variant calling and to rule out heterozygosity, the read count of the major allele was tested against a binomial distribution that assumed 95% allele frequency out of a total of rb+rref observations, i.e. sole presence of either the branch or the reference haplotype (if 100% had been assumed, it would not yield a distribution). The null hypothesis of homozygosity was rejected after P value correction by Storey's method [60] for q values below 0.05.
The same variant could be part of several different haplotypes and thus, could be included into different branch sequences. Reads supporting this variant would map at multiple locations in the reference. Therefore, we allowed all aligned rather than only unique reads to contribute to read counts and omitted the paired-end correction procedure.
We followed a similar “population-aware” approach to prefer commonalities among strains as was used for the SNP calling for labeling variants as being common or segregating. Here, variable sites with accumulated coverage over both branch and reference sequence of ≤3x were marked as “missing data”. If there was at least one haplotype in a strain with a q value above 0.05, it was assumed to be present in the population. If the test on the same haplotype failed in another strain, but the absolute read count of the haplotype sequence exceeded the alternative haplotype read count by ≥2-fold, then this haplotype was considered present in the corresponding strain as well.
We classified variants where at least 7 out of 13 strains did not show missing data as ‘common’ if the branched haplotype was present in all strains, as ‘not present’ if the reference haplotype was present in all strains, or into ‘segregating’ if there was support for both haplotypes.
To combine common variants identified by the described stepwise algorithm into potentially less evolutionary events, we aligned 200 bp around each variant of the last iteration's genome back to the TAIR9 Col-0 reference genome using a global alignment strategy (‘needle’ from the EMBOSS v6.3.1 package).
In total, we found 842,103 common and 2,017 segregating polymorphisms without removing linked loci compared to Col-0 after two iterations, to which the different tools contributed to different extent depending on the variant type (S3C Fig.).
Genomic and bisulfite sequencing were performed as described in ref. [12].
The procedure followed one described [12], except that we aligned reads against the HPG1-like as well as against the Col-0 reference genome sequences. Command line arguments for SHORE are listed in S1 File.
We performed whole methylome bisulfite sequencing to an average depth of 18x per strand (S5 Table) on two pools consisting of 8-10 individuals per accession. We followed the same procedures as described [12] to retrieve statistically significantly methylated positions. Here, we restricted the set of analyzed positions to cytosine sites with a minimum coverage of 3 reads and sufficient quality score (Q25) in at least half of all strains (i.e. ≥7), that is, 21 million positions in total. Out of those, we identified 3.8 million methylated cytosines in at least one strain by applying a false discovery rate (FDR) threshold at 5%, and between 2,120,310 and 2,927,447 methylated sites per strain (S5 Table). False methylation rates retrieved from read mapping against the chloroplast sequence can be found in S5 Table. Using the HPG1 pseudo reference genome instead of the Col-0 reference genome increased the number of cytosines sufficiently covered for statistical analysis by 5% on average, and the number of positions called as methylated by 7% (S5 Table).
We performed the same methods as in ref [12] to obtain DMPs. First, cytosine positions were tested for statistical difference between both replicates of a sample using Fisher's exact test and a 5% FDR threshold. Because individual samples consisted of a pool of several plants, the number of DMPs between replicates was negligible (between 0 and 161). After excluding them, we applied Fisher's exact test on the 3.8 million cytosine sites methylated in at least one strain for all pairwise strain comparisons. Using the same P value correction scheme as in Becker et al., we identified 535,483 DMPs across all 13 strains.
Using the model developed in ref [61], a beta prior distribution was estimated that determined the non-ancestral frequency for each variable site. We assumed the methylation state in Col-0 to be ancestral, which resulted in beta distribution parameters of a = 0.029 and b = 0.644, corresponding to a mean non-ancestral DMP frequency of 0.043 and a corresponding standard deviation of 0.157. These were then used to estimate the fraction of common DMPs that were expected to be found by sequencing a given number of methylomes. Based on the formula presented in supporting section 3 of ref [61], we estimated the total number of DMPs in the population:
For Nind = 13 (the number of accessions in this study) and Δ(1) = 1,046,892 (the total number of DMPs versus the Col-0 reference), we estimated a total number of possible DMPs in the population of N = 59,770,415, which is close to the 43 million cytosines in the A. thaliana genome. Given such an estimate for N, the Δ function can be evaluated numerically to estimate the number of DMPs we would have detected had we analysed the same number of accessions as in ref [10] (S8 Fig.).
The value of an approach that defines methylated regions (MRs) before identifying differentially methylated regions (DMRs) has been demonstrated before with a Hidden Markov Model (HMM) method developed for the analysis of methylated-DNA-immunoprecipitation followed by array hybridization (MeDIP-chip) [40]. An HMM based on next-generation sequencing data was also applied to segment the maize genome, which is much more highly methylated than the A. thaliana genome, into hypo- and hypermethylated regions [62]. We modified the HMM implementation from Molaro and colleagues [41] based solely on within-genome variation in methylation rate. It assumes that the number of methylation-supporting reads at each cytosine follows a beta binomial distribution and that distributions over positions within and between methylated regions will differ from each other, providing a way to distinguish them. Thus, the model learns methylation rate distributions for both an unmethylated and a methylated state for each sequence context separately (CG, CHG and CHH) while simultaneously estimating transition probabilities between the two states from genome-wide data. On the trained model, the most probable path of the HMM along the genome is then used to define regions of high and low methylation. The method of Molaro and colleagues was designed for calling MRs in human samples, where the vast majority of methylated cytosines are in a CG context. In plants, however, one observes considerable methylation in all three contexts (CG, CHG and CHH), each with a different methylation rate distribution. Hence, we extended the HMM by learning the parameters of three different beta binomial distributions per state, one for each context. Additionally, in contrast to humans, only the minority of cytosines in the CG context is methylated, as are cytosines in the other contexts. Hence, methylation rates were inverted to find hypermethylated, rather than hypomethylated regions as in the original HMM implementation.
Apart from these changes and some final filtering steps (see below), we followed the same computational steps as described by Molaro and colleagues [41]: The describing parameters of the – in our case – six distributions (determining the emission probabilities) and the transition probabilities between states were iteratively trained (using the Baum-Welch algorithm) from methylation rates of all cytosines in the corresponding context throughout the genome. After each iteration, all cytosines were probabilistically classified into the most likely state via Posterior Decoding, given the trained model. After training of the HMM, i.e. after maximally 30 iterations or when convergence criteria were met, consecutive stretches of high methylation state were scored, in our case by the sum of all contained methylation rates. Next, P values were computed by testing the scores against an empirical distribution of scores obtained by random permutation of all cytosines throughout the genome. After FDR calculation, consecutive stretches in high state with an FDR <0.05 are defined as methylated regions (MRs).
The HMM was run on all genome-wide cytosines, independent of their coverage. Methylation rates were obtained using accumulated read counts from the strain replicates, resulting in one segmentation of the genome per strain. Gaps of at least 50 bp without a covered C position within a high methylation state automatically led to the end of the high methylation segment. Positions with a methylation rate below 10% at the start or end of highly methylated regions (until the first position with a rate larger than 10%), were assigned to the preceding or subsequent low methylation region, respectively.
The method to identify MRs yielded 13 different segmentations of the genome, one for each strain. We selected regions being in different or highly methylated states between strains and statistically tested them for differential methylation (including FDR calculation). To obtain epiallele frequencies, we clustered strains into groups based on their pairwise comparisons and statistically tested the groupings against each other. Regions that showed statistically significant methylation differences between at least two sets of strains were identified as DMRs. Finally, because of the sensitivity of the statistical test, we empirically filtered DMRs for strong signals and defined highly differentially methylated regions (hDMRs). All these steps are described in depth in the following.
We defined a breakpoint set containing the start and end coordinates of all predicted methylated regions. Each combination of coordinates in this set defined a segment to perform the test for differential methylation in all pairwise comparisons of the strains, if at least one strain was in a high methylation state throughout this whole segment (S12A Fig.). To also detect quantitative differences rather than solely presence/absence methylation, we also compared entirely methylated regions in more than one strain to each other.
Because of the sheer number of such regions, we applied the following greedy filter criteria: Regions were discarded from any pairwise comparison if less than 2 strains contained at least 10 cytosines covered by at least 3 reads each (accumulated over strain replicates) in this region (S12A Fig. (a)). Regions were discarded from any pairwise comparison if the reciprocal overlap of this region to at least one previously tested region was more than or equal to 70% (S12A Fig. (b)). This was done to prevent “similar” regions to be tested twice. Pairwise tests of a region were not performed if both strains were in low methylation state throughout the whole region (S12A Fig. (c)). Strains were excluded from pairwise comparisons in a region if the number of positions covered by at least 3 reads each was less than half of the maximum number of such positions of all strains in the same region (S12A Fig. (d)). This prevented comparing regions with unbalanced coverage to each other, e.g. a strain with 10 data points against another one with only 2.
These filters reduced the set of regions to test from ∼2.5 million to ∼230,000 per pairwise comparison.
We designed a statistical test for differential methylation between two strains for a given region. The test assumes that the number of methylated and unmethylated read counts per position along a region follows a beta binomial distribution – similar to the HMM in MR calling. More precisely, there are 3 distributions for each sequence context and for each strain. Using gradient-based numerical maximum likelihood optimization, we fitted the parameters for each beta binomial distribution on the available read count data of the region in the respective strain. This was done a) for each of the two strains separately (while taking strain replicates into account), resulting in (two times three) strain-specific beta binomial distributions, and b) for the read counts of both strains including their replicates together, resulting in (three) common beta binomial distributions. In this way, we obtained each distribution's mean µ and standard deviation σ. We selected only regions for potential DMRs, whose intervals [µ1 – 2σ1, µ1 + 2σ1] for strain 1 and [µ2 – 2σ2, µ2 + 2σ2] for strain 2 did not overlap.
To further corroborate statistical significance, we computed P values by calculating the ratio of the strain-specific and the common log likelihoods of the available read count data using the corresponding beta binomial distributions and by testing it against a chi-squared distribution (with 6 degrees of freedom). Let sample S have NSc cytosines in context c in total and CScp reads at position p in context c, from which xScp are methylated, then we compute:
After correction for multiple testing using Storey's method [60], an FDR threshold of 0.01 defined statistically different methylated regions (DMRs) between two strains.
Additionally, this method allowed calling differential methylation in a region for each context separately by computing P values as described above without summing over the contexts (c = 1, 2 or 3). We termed resulting DMRs CG-DMRs if the methylation at only CG sites within this region was statistically significantly different, and similarly CHG-DMRs and CHH-DMRs.
For 13 strains there are at maximum 78 pairwise comparisons per region. To summarize pairwise comparisons and obtain epiallele frequencies, we assigned strains into differentially methylated groups. To achieve such clustering, we constructed a graph for each region where strains were represented as vertices and connected to other strains by an edge if the region was identified as a DMR between them (S12B Fig.). We assume that strains within a group are then similarly methylated. The task is to find the smallest number of groups of vertices so that no two strains within a group are connected by an edge.
We set up a custom algorithm, which iteratively solves the “vertex coloring problem” for an increasing number of different colors, starting with two and quitting once all strains could be successfully assigned a color (S12B Fig.). In each iteration, strains were processed in descendent order of their degree (i.e. number of edges it is connected to). Each strain was assigned to all possible colors that did not invoke a collision. Subsequently, the algorithm continued recursively to assign the color of the next strain.
Each strain had 3 context-dependent means of its beta binomial distributions per region (termed strain means from now on). We roughly approximated each group's mean methylation values (group means) as the mean values of all strain means within a group. The grouping diversity describes the accumulated absolute differences between the strain means and their respective group means divided by the number of strains. As an example, consider S12B Fig. For simplicity, it only displays methylation rates for one out of three contexts. In the real data, the respective values were accumulated over all three contexts. The group mean for the blue strains in the example is (89+90+90+93+87)/5 = 89.8% and for the white strains 52%. The grouping diversity of the clustering shown here would be (from strains A to K): (|56–52|+|59–52|+|64–52|+|89–89.8|+|41–52|+|93–89.8|+|90–89.8|+|45–52|+|47–52|+|90–89.8|+|45–52|+|87–89.8|)/11 = 2.84.
If there was more than one possible grouping of the strains, we chose the one with the lowest grouping diversity. A strain with no edges (i.e. which is not statistically differentially methylated to any other strain) was assigned into the group to which the accumulated absolute difference between its strain mean and the group mean was lowest. In the example of S12B Fig., strain L is grouped to the blue strains because its mean methylation value (81%) is closer to the blue group mean (90%) than to the white one (52%).
This procedure summarized the ∼221,000 DMRs of all pairwise strain comparisons into 11,323 DMRs between groups of strains.
Once grouped, the same statistical test as for differential methylation between two strains was used to test groups of strains. Beta binomial distributions were approximated using the read counts of all strains in a group as if they were replicate data. This procedure identified 10,645 groups of regions showing significantly different methylation. Because the method used for the selection of the regions to perform the differential test can result in overlapping regions, DMRs can still overlap each other. From sets of overlapping DMRs, the non-overlapping DMR(s) with the lowest ‘grouping diversity’ was (were) retained, resulting in 4,821 final DMRs. For the vast majority of DMRs (98%), strains were classified into two groups, i.e. there are only two epialleles.
Our sensitive statistical test classified as differential some regions with low variance and only subtle methylation difference; we therefore defined as highly differentially methylated regions (hDMRs) with potentially greater biological relevance all DMRs that were longer than 50 bp and that showed a more-than-three-fold difference in methylation rate in at least one sequence context, when considering at least five cytosines of that context (S12 Fig.). In addition, the overall methylation rate of the DMR in the more highly methylated strain had to be greater than 20%. Of 3,909 size-filtered DMRs, 3,199 (80%) were classified as hDMRs (S8 Table). The grouping of hDMRs yielded similar epiallele frequencies as for the DMPs (54% with frequency larger than 1; Fig. 2F).
The data from Stroud and colleagues [43] contain position-wise methylation rates for each sample. We defined a single site as methylated in wild type (WT) if both Col-0 samples Col_WA034L3 and Col_WB023L8 had a methylation rate of 10% or higher, and if at least one of them is more than 20% methylated. We declared a site in a mutant sample as having ‘lost’ methylation where the wild type was methylated and the mutant showed a methylation rate of less than 10%. In contrast, a ‘gained’ methylation site had less than 10% methylation in at least one of the WT samples and more than 20% methylation in the mutant. To assess if epigenetic variation in the HPG1 lines is enriched at sites affected by impaired methylation machinery, for each mutant, we constructed a set of positions, which were methylated in WT, covered in the mutant sample (i.e. present with a rate in the mutant sample file), and which were covered in the HPG1 and MA populations. A site was considered covered in a population when more than half of the strains showed a high quality and a more than 3-fold covered base call (see ‘Determination of methylated sites’ or [12]). For those positions and different subsets thereof, the fractions of sites with gained or lost methylation in the mutant compared to the wild type samples were plotted in S19 Fig.
For each differentially methylated region, we considered a linear mixed model to estimate the proportion of variance that is attributable to genetic effects (heritability) and its standard error. The approach is similar to variance component models used in GWAS, e.g. refs. [63], [64]. Briefly, we considered the log average methylation rate of DMRs as phenotype and assessed the variance explained by genotype using a Kinship model constructed from all segregating genetic variants. We considered only DMRs and genetic polymorphisms that had no missing data in all accessions.
We identified non-synonymous SNPs using SHOREmap_annotate [65] and excluded them from population structure analyses. We ran STRUCTURE v.2.3.4 [66] with K = 2 to K = 9 with a burn-in of 50,000 and 200,000 chains for 10 repetitions and determined the best K value using the ΔK method [67]. The phylogenetic network was generated using SplitsTree v.4.12.3 [68].
We used the TAIR10 annotation for genes, exons, introns and untranslated regions; transposon annotation was done according to [69]. Positions and regions were hierarchically assigned to annotated elements in the order CDS> intron> 5′ UTR> 3′ UTR> transposon> intergenic space. We defined as intergenic positions and regions those that were not annotated as either CDS, intron, UTR or transposon.
Positions were associated to the corresponding element when they were contained within the boundaries of that element. (D)MRs were associated to a class of element if they overlapped with that class of element; a (D)MR could only be associated to one class of element. When summing up basepairs of an element class covered by (D)MRs, the number of basepairs of a (D)MR overlapping with that class of element were considered. In that case the space covered by a (D)MR could be assigned to different classes of elements, while each basepair of the (D)MR could be assigned to only one class.
We tested for significant overlap of DMRs using multovl version 1.2 (Campus Science Support Facilities GmbH (CSF), Vienna, Austria). We reduced the genome space to the basepair space covered by MRs identified in at least one HPG1 accession. DMRs were considered in the analysis if their start and end positions were contained within the MR space. DMRs that only partially overlapped with the MR space were trimmed to the overlapping part. Overlap between DMRs from different datasets was analyzed by running 100,000 permutations of both DMR sets within the MR basepair space. multovl commands are listed in S1 File.
Reads were processed in the same way as genomic reads, except that trimming was performed from both read ends. Filtered reads were then mapped to the TAIR9 version of the Arabidopsis thaliana (http://www.arabidopsis.org) genome using Tophat version 2.0.8 with Bowtie version 2.1.0 [70], [71]. Coverage search and microexon search were activated. The command lines for Tophat are listed in S1 File.
For quantification of gene expression we used Cufflinks version 2.0.2[72]. We ran a Reference Annotation Based Transcript assembly (RABT) using the TAIR10 gene annotation (ftp://ftp.arabidopsis.org/home/tair/Genes/TAIR10_genome_release/TAIR10_gff3/) supplied with the most recent transposable element annotation [69] Fragment bias correction, multi-read correction and upper quartile normalization were enabled; transcripts of each sample were merged using Cuffmerge version 2.0.2, with RABT enabled. For detection of differential gene expression we ran Cuffdiff version 2.0.2 on the merged transcripts; FDR was set to <0.05 and the minimum number of alignments per transcripts was 10. Fragment bias correction, multi-read correction and upper quartile normalization were enabled. The command lines for the Cufflinks pipeline are listed in S1 File. Analysis and graphical display of differential gene expression data was done using the cummeRbund package version 2.0.0 under R version 3.0.1.
When not mentioned otherwise in the corresponding paragraph, graphical displays were generated using R version 3.0.1 (www.r-project.org). Circular display of genomic information in Fig. 2A was rendered using Circos version 0.63 [73].
Leaf area was determined using the automated IPK LemnaTec System and the IAP analysis pipeline [74]. Plants were grown in a controlled-environment growth-chamber in an alpha-lattice design with eight replicates and three blocks per replicate, taking into account the structural constraints of the LemnaTec system. Each block consisted of eight carriers, each carrying six plants of one line. Stratification for 2 days at 6°C was followed by cultivation at 20/18°C, 60/75% relative humidity in a 16/8 h day/night cycle. Plants were watered and imaged daily until 21 days after sowing (DAS). Adjusted means were calculated using REML in Genstat 14th Edition, with genotype and time of germination as fixed effects, and replicate|block as random effects.
Local temperature and liquid precipitation data was calculated from National Climatic Data Center (NCDC) Global Summary of Day (GSOD) data. Collection locations were matched to the closest weather station with <5% missing data for five years prior to the collection date. Cumulative liquid precipitation was calculated each year starting from January 1.
The DNA and RNA sequencing data have been deposited at the European Nucleotide Archive under accession number PRJEB5287 and PRJEB5331. A GBrowse instance for DNA methylation and transcriptome data is available at http://gbrowse.weigelworld.org/fgb2/gbrowse/ath_methyl_haplotype1/. DNA methylation data, MR coordinates and genetic variant information have also been uploaded to the genome browser of the EPIC consortium (https://www.plant-epigenome.org/; https://genomevolution.org/wiki/index.php/EPIC-CoGe) and can be accessed at http://genomevolution.org/r/939v. The software of our methylation pipeline can be downloaded at http://sourceforge.net/projects/methpipeline.
|
10.1371/journal.pcbi.1007177 | Biophysics and population size constrains speciation in an evolutionary model of developmental system drift | Developmental system drift is a likely mechanism for the origin of hybrid incompatibilities between closely related species. We examine here the detailed mechanistic basis of hybrid incompatibilities between two allopatric lineages, for a genotype-phenotype map of developmental system drift under stabilising selection, where an organismal phenotype is conserved, but the underlying molecular phenotypes and genotype can drift. This leads to number of emergent phenomenon not obtainable by modelling genotype or phenotype alone. Our results show that: 1) speciation is more rapid at smaller population sizes with a characteristic, Orr-like, power law, but at large population sizes slow, characterised by a sub-diffusive growth law; 2) the molecular phenotypes under weakest selection contribute to the earliest incompatibilities; and 3) pair-wise incompatibilities dominate over higher order, contrary to previous predictions that the latter should dominate. The population size effect we find is consistent with previous results on allopatric divergence of transcription factor-DNA binding, where smaller populations have common ancestors with a larger drift load because genetic drift favours phenotypes which have a larger number of genotypes (higher sequence entropy) over more fit phenotypes which have far fewer genotypes; this means less substitutions are required in either lineage before incompatibilities arise. Overall, our results indicate that biophysics and population size provide a much stronger constraint to speciation than suggested by previous models, and point to a general mechanistic principle of how incompatibilities arise the under stabilising selection for an organismal phenotype.
| The process of speciation is of fundamental importance to the field of evolution as it is intimately connected to understanding the immense bio-diversity of life. There is still relatively little understanding of the underlying genetic mechanisms that give rise to hybrid incompatibilities with results suggesting that divergence in transcription factor DNA binding and gene expression play an important role. A key finding from the field of evo-devo is that organismal phenotypes show developmental system drift, where species maintain the same phenotype, but diverge in developmental pathways; this is an important potential source of hybrid incompatibilities. Here, we explore a theoretical framework to understand how incompatibilities arise due to developmental system drift, using a tractable biophysically inspired genotype-phenotype for spatial gene expression. Modelling the evolution of phenotypes in this way has the key advantage that it mirrors how selection works in nature, i.e. that selection acts on phenotypes, but variation (mutation) arise at the level of genotypes. This results, as we demonstrate, in a number of non-trivial and testable predictions concerning speciation due to developmental system drift, which would not be obtainable by modelling evolution of genotypes or phenotypes alone.
| The detailed genetic mechanisms by which non-interbreeding species diverge is still poorly understood. Darwin, inspired by John Herschel, called it that “mystery of mysteries” [1]; he struggled to understand how natural selection could give rise to hybrid inviability or infertility between populations without producing such incompatibilities within the populations. A solution to this problem was conceived independently by Dobzhansky, Muller and Bateson in which cross-mating would combine alleles at different loci that are incompatible due to epistatic interactions (Dobzhansky Muller incompatibilities, DMI) [2–4]. Consider a common ancestor with alleles ab across two loci, which after a period of allopatric divergence give rise to two lineages which have fixed genotypes Ab and aB, respectively. Interbreeding between these two populations would result in the heterozygotic hybrid genotype Aa|Bb, combining the potentially incompatible A and B, a combination that could not arise in either population mating separately.
Assuming that any combination of alleles that have not been “tested” by the process of evolution represents a potential incompatibility, Orr predicted that the number of incompatibilities between pairs of alleles in a sufficiently large genome would increase with the number of substitutions separating the two lineages (K) as K (K−1) ∼ K2 [5]. Similarly, the number of untested combinations involving n loci would increase as ∼ Kn, suggesting that, with evolutionary time, potential incompatibilities would become increasingly dominated by more complex epistatic interactions [5]. This would occur, firstly, because there are a larger number of combinations, and secondly, because there are more ways for separate lineages to evolve around incompatible genotypes when there is a larger number of loci. It is unclear, however, how the simplistic assumptions of this model fare with increased biological realism. Not all possible untested hybrids are equally likely to result in real incompatibilities. Selection acting on each separate lineage affects the substitutions that occur and their likely contributions to reproductive isolation. In particular, evolutionary constraints have a strong effect on the development of more complex DMIs, making it uncertain whether their role is as important as suggested by Orr’s combinatorial argument. This highlights the need for considering more realistic models that better capture the salient aspects of the underlying biology, whilst remaining sufficiently simple for tractable evolutionary modelling and simulation.
In recent years a form of epistasis has been described in a number of organisms whereby closely related species have similar organismal phenotypes but are produced by very different developmental mechanisms [6–9]. This cryptic “developmental system drift” [10, 11] could be an important source of hybrid incompatibilities that cause reproductive isolation [12, 13]. Developmental system drift is an example of a more general characteristic of biological systems where many genotypes can correspond to the same phenotype; this redundancy of the mapping from genotype to phenotype results in a number of non-trivial behaviours which do not arise on fitness landscapes which consider evolution of phenotypes or genotypes independently [14–25]. The degree of redundancy can be represented as the “sequence entropy”, corresponding to the log of the number of genotypes corresponding to a given phenotype, in analogy to the similar expression in statistical mechanics [17, 25–27].
To explore the role of developmental system drift on speciation, we examine the growth of Dobzhansky-Muller incompatibilities using a simple genotype-phenotype map that models the development of spatial patterning of gene expression. The model, introduced by [17] allows for cryptic genetic variation and changes in molecular phenotypes while maintaining organismal phenotype under stabilising selection. In addition, we introduce a novel computational method to decompose hybrid DMIs so we can examine the behaviour of the fundamental pair-wise and higher order incompatibilities. We show that including biologically relevant elements gives rise to a number of novel phenomenon that could not arise with models based only on the fitness of genotypes or phenotypes. Our results show that small populations develop hybrid incompatibilities more quickly, due to the pressure of sequence entropy in small populations meaning the common ancestor harbours on average a larger drift load. For large populations, we find hybrid incompatibilities arise more slowly, with a growth law characteristic of a sub-diffusion of the hybrid binding energies, indicative of kinetic traps in the molecular substitution process due to roughness to the fitness landscape [17]. Strikingly, we find that for moderate population sizes it is the molecular phenotypes under weakest selection that give rise to earliest incompatibilities, since in the common ancestor they are more likely to be already maladapted. Finally, we find that unlike Orr’s prediction that complex DMIs should be abundant, pair-wise interactions between loci dominate the growth of DMIs, showing that biophysics provides a stronger constraint than pure combinatorics.
The genotype-phenotype map we use is a modification of the one described in [17]. The evolutionary task set for the gene regulation module is to turn an exponentially decaying morphogen gradient (M) across a field of cells in an embryo into a sharp step function profile of a downstream transcription factor T with its transition at the mid-point of the embryo, as shown in Fig 1. This is accomplished by having the morphogen and an RNA Polymerase R bind to two adjacent non-overlapping binding sites in the cis-regulatory region (C) region of the transcription factor, the promoter P, and a single binding site B adjacent to it; transcription occurs whenever the polymerase binds to the promoter, although both proteins can bind to both binding sites dependent on their binding affinities. Binding to the regulatory region is cooperative due to stabilising interactions between the two proteins when bound at the two adjacent sites. The sequences of M and R at the DNA binding sites are represented by binary strings of length ℓpd = 10. The corresponding DNA binding sites B and P are also represented by binary strings of the same length. Interactions between a pair of proteins are similarly represented by binary strings of length ℓpp = 5. We assume an exponential morphogen concentration profile [M](x, α), as a function of the position of embryonic cells, x; the decay rate of the morphogen α is represented as a continuous variable, with a relative probability of mutation corresponding to an effective string of length ℓα = 10 bases. This results in a genome G, of total length ℓG = 60. Protein-DNA and protein-protein binding strengths are determined by the number of mismatches between corresponding strings on the two interacting molecules, where for protein-DNA binding the cost of a mismatch is ϵpd = 2kBT and for protein-protein interactions ϵpp = 1kBT, where kBT is Boltzmann’s constant multiplied by room temperature (298K). We assume that there is a fixed concentration [R] of polymerase, in each cell. We then follow [28] and assume that the concentration of the transcription factor in a cell at position x ([T](x)) is simply proportional to the probability of the polymerase being bound to the promoter, where this is calculated using standard methods of equilibrium statistical mechanics allowing for all configurations of protein species bound at these two binding sites, as well as none being bound (see Methods for details). The fitness contribution F of the overall patterning phenotype ranges from 0 to κF depending on how well expression of the transcription factor is confined to the anterior half of the embryo, as shown in Fig 1 (bottom left), where κF is a measure of the relative contribution of this trait to the fitness of the organism. We define a population-scaled fitness contribution 2NeκF, where Ne is the effective population size; for 2Ne κF < 1 the effects of selection are weak, and are conversely strong when 2NκF > 1. We also assume that there is a boundary at F = F*, below which the organism is unviable. We simulate evolution as continuous time Markov process. After evolving a single population for a given number of generations, we form two replicates of the population that evolve independently, representing the process of allopatric speciation. At various time points following this imposed isolation, we consider the fitness and viability of various outcrossings between the two populations. A DMI occurs when the fitness contribution of a particular hybrid drops below F*.
The properties of a similar genotype-phenotype map have been previously explored [17]. An important property of this genotype-phenotype map is that only a single mechanism of patterning is found, in which the polymerase (R) binds with intermediate affinity to the promoter (P) but with high affinity to the morphogen (M), while the morphogen binds to the morphogen binding site (B) only above a critical morphogen concentration. This results in a spatial switch once the morphogen falls below this concentration; evolution then fine tunes the relationship between the protein-DNA binding energies, the protein-protein binding energy and the steepness of the morphogen gradient α to turn off transcription at the mid-point of the embryo. Despite a single global solution there are many different combinations of the protein-DNA and protein-protein binding energies and α that give good patterning, and each of these correspond to many possible genotypes (G). Of the different possible binding energies, we find that E M B , E R P , E ˜ R M (binding energy of M to B, R to P, and R to M, respectively, calculated by the equivalent of Eq 3 in the Methods) are under strong selection, whilst the other possible binding energies are essentially neutral with weak selective effects. At large population sizes it is found that the evolutionary dynamics exhibits what is known as quenched-disorder in statistical physics, where energy phenotypes that are less constrained take different random values between independent evolutionary runs with no further substitutions; this indicates an underlying roughness to the fitness landscape and that these weakly selected traits are trapped in a local optimum [17].
A key property determining the rate at which incompatibilities arise is the distribution of common ancestor phenotypes as a function of the population-scaled fitness contribution 2Ne κF, as shown in Fig 2. For a given value of κF, we see that for large population sizes (2NκF ≫ 10) the distribution is what we would expect from conventional evolutionary theory on a fitness landscape with a fitness maximum. In contrast, as the population size is decreased, we find the distribution shifts to lower fitness values to the point when selection is weak (2NκF ≤ 1) the distribution is poised at the inviability boundary. This effect arises due to genetic drift at low population sizes pushing populations towards marginally fit phenotypes that correspond to the largest number of genotypes, that is, with the largest sequence entropy.
Our genome is composed of 4 loci: 1) the R locus corresponding to the polymerase sequence, 2) the Morphogen (M) locus, 3) the C locus which corresponds to the sequences for the cis-regulatory region of the transcription factor and 4) the α locus, which is the morphogen gradient steepness α. Hybrids between the two lineages are constructed by independent reassortment of these loci assuming complete linkage within each locus and no linkage between them. We define a hybrid genotype by a 4 letter string where each letter corresponds to one of the loci defined above and takes one of two cases correspond to whether the allele is from the 1st line or 2nd line; for example, the hybrid rMCa corresponds to R locus having an allele from the 1st lineage, M locus with the allele from the 2nd lineage, the transcription factor (cis-regulatory) C locus from the 2nd lineage and α locus the allele from the 1st lineage. Note that the underlying sequence of each hybrid changes as different substitutions are accepted in each lineage; the notation only refers to alleles fixed at any point in time. We can represent all combinations of the four loci drawn from the two parents (RMCA, RMCa, RMcA, etc.) as points on a four-dimensional Boolean hypercube. In total there are 24 − 2 = 14 hybrids.
In Fig 3, we plot a typical time series of how the fitness of two different hybrids (Rmca(a & b) and RMcA(c & d) changes over a divergence time μt separating a pair of lineages, for 2NκF = 1 (a & c) and 2NκF = 10 (b & d), where μ = ℓGμ0 is the mutation rate for all base pairs in all loci. 2NκF > 1 indicates strong selection, whilst 2NκF ≤ 1 indicates weak selection where genetic drift dominates (For reference, in human populations it has been estimated that ≈ 20 − 30% of mutations are weakly selected [29, 30], compared to in Drosophila < 10% [30].). We see that the fitness of hybrids generally decreases in a stochastic fashion; when the log-fitness of a hybrid drops below the threshold F* (indicated by the dashed line), a DMI arises as is indicated by a vertical log-fitness line (F = −∞) for that hybrid. As can be seen in Fig 3, at any given time a changing subset of the fourteen possible hybrids might be incompatible.
There is still very little understood about the underlying genetic basis that gives rise to reproductive isolation between lineages. Gene expression divergence is thought to be a strong determinant of the differences between species [33–37] with a growing body of evidence for their direct role in speciation, particularly through transcription factors [38–42]. In particular, transcription factors mediate the control of gene expression and the ultimately body plans of organisms through complicated gene regulatory networks. Studies in drosophila and nematodes have shown in closely diverged species conserved body plans yet plasticity in the underlying molecular architectures between them [7–11, 43].
This evidence suggests that divergence in the regulatory networks controlling gene expression and body plans in allopatric lineages is likely to be an important mechanism of speciation in many higher organisms, yet we do not yet have a quantitative evolutionary framework to model such processes, which can then be used to make predictions. A key challenge is to link changes at the genetic level, where mutations arise, to their outcomes on changes in phenotype, where selection actually acts. Although in the past century we have made great strides in our understanding of evolution through the modern synthesis of Fisher, Wright, Kimura and many others [44–48], this has focussed on the evolution of either genotypes or phenotypes separately. However, in recent years there has been increasing attention on the development of more realistic fitness landscapes, and how this affects the evolutionary process. For example, in biologically realistic systems, often many genotypes correspond to the same phenotype; this redundancy of the mapping from genotype to phenotype results in a number of non-trivial and emergent behaviours which do not arise on fitness landscapes which consider evolution of phenotypes or genotypes independently [14–24].
One important and well explored example is the evolution of transcription factor DNA binding [15, 16, 22, 26, 27, 49, 50], where the genotype-phenotype map from sequence to binding affinity can be explicitly enumerated under simplifying assumptions [51, 52]. These investigations show that for small populations dominated by genetic drift, evolution does not optimise fitness. Rather, there is a trade off between the high fitness of a small number of sequences that bind well and the exponentially larger number of sequences that bind less well. The result is the maximisation of a combination of fitness and the number of sequences that correspond to that phenotype. We can take advantage of analogies with statistical mechanics and represent this combination as the “free fitness”, where the log of the number of sequences is the “sequence entropy” of a phenotype [15, 26, 27, 53]. In this formulation, the effective population size is analogous to an inverse temperature for a physical system connected to a heat-bath, where decreasing population size increases the effect of drift and the importance of sequence entropy relative to fitness.
When the free fitness framework is applied to the role of transcription factor DNA binding in allopatric speciation, our previous work gave rise to a simple prediction: incompatibilities arise more quickly for smaller, drift-dominated, populations [26, 27], supporting previous computational studies by Tulchinsky et al. [54], that showed decreased hybrid fitness for smaller populations. This can be understood as a result of the greater importance of sequence entropy for small populations, resulting in common ancestors with a higher drift load, which are therefore closer to incompatible regions. As a result, fewer substitutions are required for the development of hybrid incompatibilities [26, 27]. Conversely, those transcription factor binding site pairs under weaker selection, at a fixed population size, will give rise to incompatibilities more quickly, as they are more susceptible to drift and in the common ancestor will have a larger drift load.
In this paper, we examine speciation in a more realistic genotype phenotype map. For the first time we examine how incompatibilities arise in allopatry for a simple evolutionary model of developmental system drift, where a higher level organismal spatial patterning phenotype is maintained by stabilising selection, whilst the underlying molecular binding energy phenotypes and ultimately the sequences that determine them, the genotype, are allowed to drift in the evolutionary simulations. Earlier analyses of this model demonstrated the evolution of a number of non-trivial features such as a balance between fitness and sequence entropy deciding the course of evolution at small population sizes and a roughness to the fitness landscape for phenotypes which have high fitness [17]. Importantly here, unlike in previous works [26, 27] we do not directly select for high binding affinity, but only on the organismal level phenotype, but as we discuss, we find the same population size dependence, as well as a number of other novel phenomenon to the speciation process, which would not be obtainable by modelling selection only at the level of phenotypes or genotypes. The results show that biophysics and effective population size provide a much stronger constraint than previous simple modelling of the dynamics of hybrid incompatibilities would suggest [5, 31].
A key result we find is that small populations are characterised by a power law growth of incompatibilities with time, vs large populations a diffusive law (discussed below). Thus we suggest that empirical evidence of power law growth in incompatibilities is a signature of allopatric speciation at small population sizes. The Orr model of the growth of hybrid incompatibilities predicts that incompatibilities grow as a power law of the divergence time between allopatric lineages [5, 31], where the exponent represents the number of genes participating in the interaction (e.g. 2 for a 2-way incompatibility). The results of our model also yield this prediction, but only when populations sizes are sufficiently small. There is, however, an alternative model for the power law behaviour to the combinatoric argument made by Orr. As argued in [27], at small population sizes, where genetic drift is dominant and there is a large drift load, common ancestor populations are poised close to the incompatibility boundary and the growth of DMIs at short times is determined by the likelihood that a few critical substitutions arrive quickly, which is given by a Poisson process; if the critical number of substitutions is K* then for short times we would expect P I ( t ) ∼ ( μ t ) K * and so given that at least n substitutions are needed for a n-way incompatibility, we would expect K* ≥ n. In this paper, we introduced a new method to decompose DMIs into their fundamental pair-wise, 3- and 4-way incompatibilities, and find that for more complex incompatibilities (more loci involved) the larger the exponent of their power law growth. However, we find the exponents we measure for 3- and 4-way incompatibilities are smaller than the predicted exponents of 3 and 4 respectively. We suggest this could be due, as shown in the S1 Text, to the greater number of higher order DMIs arising just by chance, leading to an overestimation of 3- and 4-way DMIs at short times, where at short times a smaller exponent corresponds to a larger number (i.e. τn−1 > τn for τ < 1, where τ is some dimensionless timescale).
Examining the growth of incompatibilities at large population sizes, we see there is a characteristic negative curvature on a log-log plot, predicted theoretically by [26], indicating that, as the number of substitutions needed for incompatibilities is large, the changes in the hybrid traits can be approximated by a diffusion process. However, we find that a simple model of diffusion does not fit the simulation data well; instead a model of sub-diffusion, that would arise if there are a number of kinetic traps giving a broad distribution of substitution times, does fit the data well. This is consistent with the finding that the genotype-phenotype map has a rough fitness landscape, which is only revealed at sufficiently large population sizes [17]. However, it is not clear whether we would observe quenched disorder and sub-diffusive behaviour with more realistic biophysical models that include a larger alphabet size with 20 amino acids and 4 nucleotides.
We also find that incompatibilities arise more rapidly in smaller populations, which is an emergent effect due to the genotype-phenotype map, giving a bias in degeneracy of different phenotypes; lower fitness and less sharp patterning organismal phenotypes have many more sequences than higher fitness, sharper patterning, phenotypes. In smaller drift-dominated populations, this means there is bias towards phenotypes of small sequence entropy (log degeneracy) that counteracts the tendency of natural selection to favour phenotypes of high fitness. Consequently, the common ancestor in small populations is more likely to be slightly maladapted and less substitutions are needed before hybrids develop incompatibilities. These predictions are consistent with empirical evidence for an inverse correlation of speciation rates with effective population size; the net rates of diversification from phylogenetic trees [55–57] indicate smaller populations speciate more quickly, as well from inferred times for post-zygotic isolation to arise [58–60], where for example mammals and cichlids, which have effective population sizes of order 104 [61– 64], develop reproductive isolation more quickly than birds, which have effective populations sizes of order 106 [65]. This model and the similar results obtained for transcription factor DNA binding [26, 27] provide a robust explanation of how stabilising selection can give rise to this population size effect in speciation, which do not require passing through fitness valleys as do models based on the founder effect [66–69].
However, the results for this genotype-phenotype map for developmental system drift are particularly noteworthy compared to the previous results on transcription factor DNA binding, as they are obtained without directly selecting for high affinity, low sequence entropy, binding phenotypes; here we only select on the organismal spatial patterning phenotype, but nonetheless we find small populations develop hybrid incompatibilities more quickly through a similar mechanism of the interplay between fitness, sequence entropy and populations size. Although studies with more complex genotype-phenotype maps will be required, we suggest this points towards a broad principle, where the specificity required of a phenotype to be functional and of high fitness will mean that it will be coded by fewer genotypes. For example, in simple models of protein stability, the empirical observation that all proteins tend to be marginally stable, can be explained by the fact that as the stability of a protein is increased the number of sequences that give that stability decreases rapidly [19]; assuming high fitness corresponds to maximum stability, this phenotype is highly specified, as only a few sequences will meet the requirement that all inter-chain interactions in the protein are favourable.
Another property that emerges from this model not obtainable by simply modelling transcription factor DNA binding is that certain molecular phenotypes are more important than others in giving rise to incompatibilities. One particular feature of this model is that the selective constraints on the different molecular binding energy phenotypes emerge through the evolutionary process of stabilising selection on the organismal phenotype, and are not specified by the model. Most strikingly, and counterintuitively, the model predicts that molecular phenotypes that are under the weakest selective constraints (but not strictly neutral) dominate by giving rise to the earliest incompatibilities for intermediate and large population sizes. Remarkably, here this emerges as a consequence of stabilising selection on the organismal phenotype and not due to selection imposed for good binding affinity as in previous works [27].
We note that these results have been obtained by changing the population size whilst keeping the strength of selection on the organismal trait κF fixed. It would be tempting to use these results to suggest that overall those traits in a genome under weakest selection would give rise to the earliest incompatibilities and hence dominate allopatric speciation. However, the question of the how the dynamics of hybrid incompatibilities changes as the strength of selection changes is a subtle one, which we leave to future work; in this model a reduction in κF has the effect of changing the phenotypic regions of incompatibility, with non-trivial consequences. It should also be noted, the role of sequence entropy in giving a strong population size dependence to the rate of reproductive isolation; if we consider only a peaked phenotypic landscape without a sequence entropy bias, a reduction in population size would only lead to a broadening of the phenotypic distribution and not a change in the mean of the distribution, and so a much weaker effect, as the common ancestor will still be typically far from incompatible regions. On the other hand with strong (exponential) degeneracy biases, the mean phenotype of the common ancestor changes strongly giving the large population size effect seen, which is as demonstrated in Fig 2.
Another finding of significance is that pair-wise or 2-way DMIs dominate compared with higher order DMIs (3- and 4-way in this model with 4 loci). This is in contrast to Orr’s theoretical argument that predicts a very specific ratio of 2-way: 3-way: 4-way DMIs, equal to 12: 24: 14, which assumes that the fraction of viable paths from the common ancestor to the current day species increases as we consider higher order DMIs [5]. This argument partly rests on the assumption that the number of inviable genotypes remains fixed as a larger number of loci are considered, which would seem a very strong assumption. Despite its simplicity, the genotype-phenotype map in this paper has many of the key features required for higher levels of epistasis, with protein-DNA binding, protein-protein binding and control of the morphogen steepness, all interacting in a non-linear fashion to produce a single gene expression patterning phenotype and so there is clearly the potential for the Orr prediction to be verified; in contrast, we find the converse and our results show there is no bias towards 3-way DMIs, in fact showing instead that the ratio of 2-way to 3-way DMIs is at short times many orders of magnitude larger. This suggests, in this simple, but still relatively complex model, that biophysical constraints are far more important than a purely combinatorial argument would suggest. Evidence could be obtained from more detailed studies similar to [6, 70], where a power law with an exponent greater than 2 would indicate higher-order DMIs are dominant; currently this evidence suggests a quadratic growth law, however, a study with more time-points or species-pairs would provide more confidence. An alternative approach would be to look for linkage disequilibrium between unlinked regions of hybrid genomes, such as was found with hybrids of two species of swordtail fish [71], and though computationally challenging and requiring large numbers of parallel datasets, compare this against evidence for pervasiveness of higher order epistasis. Although recent results of [72], would seem to contradict our conclusions, their finding of extensive complex epistasis relates to higher order interactions between sites within a single locus, coding for protein stability or enzymatic activity, whereas our work relates to epistasis between multiple loci. Similarly, the results of hybrid incompatibilities within RNA molecules [73], which show a ‘spiralling complexity’ of DMIs would appear to be of limited biological relevance to allopatric speciation in higher organisms, as these are related to epistasis within a single locus, which are unlikely to segregate into different recombinants in a hybrid.
Finally, for small populations we find clustering in the behaviours of growth of different types of DMIs, in particular, 3-way DMIs (S1 Text), which can be explained by the different sequence entropy constraints on different molecular phenotypes. This degeneracy is then lifted at larger population sizes and each n-way DMI takes on a different identity in their pattern of growth; this has strong analogy to physical phenomenon in statistical physics where constraints of symmetry dominate at large temperatures, in a regime where noise is important, but at smaller temperatures this symmetry is broken.
A clear future direction to investigate would be the effect of multiple transcription factors binding to enhancer regions to control gene expression [74–76] in large gene networks, where there is potential scope for complex epistasis across many loci coding for a large number of transcription factors. However, as our results show, despite the possibility and a prior expectation of a larger number of triplet interactions, pair-wise interactions dominate; for complex transcriptional control, if pair-wise interactions between proteins, and proteins and DNA dominate, for example in determining the binding affinity of transcriptional complexes, then our conclusions would hold. As we broaden the scope to large gene regulatory networks, there is no strong and direct empirical evidence for pervasive higher order epistasis in their function, which could give rise to higher order incompatibilities being dominant [77]. Specifically, although there is evidence that higher order incompatibilities have arisen in natural populations [78–81], nonetheless a survey of these findings suggest there is no evidence for their dominance [81] as would be predicted by Orr and would be consistent with our findings that point towards biophysics providing a stronger constraint.
Overall, our results point to a basic principle, where developmental system drift or cryptic variation [7, 10, 11, 43], play a key role in speciation; basic body plans or phenotypes are conserved, but co-evolution of the components and loci of complicated gene regulatory networks can change differently in different lineages, giving incompatibilities that grow in allopatry. Here, we suggest a universal mechanism, where the rate of growth of incompatibilities is controlled by the drift load, or distribution of phenotypic values, of the common ancestor, which in turn is determined by a balance between selection pushing populations towards phenotypes of higher fitness and genetic drift pushing them towards phenotypes that are more numerous (higher sequence entropy); this basic mechanism would predict in general that traits under weaker selection will dominant the initial development of reproductive isolation. In particular, although in principle more complicated regulation could give rise to more complex patterns of epistasis [5], our findings suggest that more simple, pair-wise, incompatibilities dominate the development of reproductive isolation between allopatric lineages under stabilising selection.
We model the binding energies of proteins to DNA using the “two-state” approximation [51, 52], which assumes that the binding energy of each amino acid-nucleotide interaction at the binding interface is additive and to a good approximation controlled by the number of mismatches, which each have the same penalty in binding affinity. The various protein-DNA binding energies in the main text are then given by the Hamming distance between the respective sequences. We assume these energies E are measured relative the background free energy of specific and non-specific binding all other sites in the genome, such that the probability of a given transcription factor being bound to a single site is p = 1/(1+e(E−μ)/kBT), where μ is the chemical potential (∼log(concentration)) of the TF [17, 26, 82]. For example, the binding energy between the morphogen (M) and the first binding site (B) is given by
E M B = ϵ p d ρ ( g m , g B ) (3)
where ρ(gm, gB) is the Hamming distance between the protein binding sequence (gm) for the morphogen and the sequence for a first regulatory binding site (gB), where ϵpd is the cost in energy for each mismatch. We assume ϵpd = 2kBT as a typical value for the mis-match energy, which are found to be in the range 1−3kBT [51, 52]. Similarly the co-operative protein-protein binding energy, for example between RNAP and the morphogen is E ˜ R M is
E ˜ R M = ϵ p p ( ℓ p p - ρ ( g R , g M ˘ ) ) (4)
where gR is the sequence involved in protein-protein interactions for the polymerase, and g M ˘ represents the equivalent binary sequence for the morphogen, flipped about its centre, which mimics the chirality of real proteins and prevents the co-operative stability from always being maximum for homo-dimers. The parameter ϵpp is the stability added for each favourable hydrophobic interaction between amino acids, which we assume to be ϵpp = −kBT. Given ℓpp = 5 this gives interactions consistent with typical literature values of −2 to −7kBT for hydrophobic interactions between proteins [28, 83].
The morphogen concentration profile is approximately exponential; the exact profile we use is
[ M ] ( x ) = [ M 0 ] cosh ( α ( x - L ) ) sinh ( α L ) (5)
where this arises from solving the reaction-diffusion equation with reflecting boundary conditions and is valid for all α.
To calculate this probability, we use the canonical ensemble of statistical mechanics, for which the partition function Z is most simply expressed in terms of a spin-like variable, which represents the occupation of each binding site, σ j = { 0 , R , M },
Z = ∑ σ P ∑ σ B e - ( E σ P P + E σ B B - μ σ P - μ σ B + E ˜ σ P σ B ) / k B T
with E 0 j = 0, E ˜ i i ′ = 0, for either i = 0 or i′ = 0 and μ0 = 0, where μ σ j = k B T ln [ σ j ] represents the chemical potentials of the protein species at the jth binding site with [σj] being the concentration of species σj and 0 represents a free binding site. Formally this construction is known as a 3−state Pott’s model. So given a ‘genome’ G = [gR, gr, gM, gm gP, gB] from which the protein-DNA and protein-protein binding energies are calculated (Eij and E ˜ i i ′, respectively), pRP is given schematically by
pRP=p(σB−R↱)=1Z((0−R↱)+(M−R↱)+(R−R↱)) (6)
where, for example,
( M - R ↱ ) ≡ e - ( E R P + E M B - μ R - μ M + E ˜ R M ) / k B T
is the Boltzmann factor for co-operative binding of the morphogen and RNAP to R. Note that we ignore in the partition sum protein-protein binding when not bound to DNA, since these co-operative binding energies tend to be relatively weak compared to DNA binding.
We use a kinetic Monte Carlo scheme to simulate the evolutionary process for the genome G and α on two independent lineages, assuming a fixed effective population size of N, and that we are in the regime of small effective population size (ℓG μ0N ≪ 1, where μ0 is the base-pair mutation rate). This means the population is represented by a single fixed sequence (or number for α) for all of the loci at each time-point, where effectively mutations are either instantly fixed or eliminated. Specifically, we use the Gillespie algorithm [84], to simulate evolution as a continuous time Markov process; at each step of the simulation the rate of fixation of all one-step mutations from the currently fixed alleles (wild type) is calculated, and one of these mutations is selected randomly in proportion to their relative rate. Time is then progressed by K−1 ln(u), where K is the sum of the rates of all one-step mutants and u is a random number drawn independently between 0 and 1, which ensures the times at which substitutions occur is Poisson distributed, as we would be expected for a random substitution process. The rates are based upon the Kimura probability of fixation [85]:
k = μ 0 N 1 - e - 2 δ F 1 - e - 2 N δ F ≈ μ 0 2 N δ F 1 - e - 2 N δ F , (7)
where δF is the change of fitness of a mutation at a particular location, given by fitness function defined in the main text, and μ0N is the rate at which mutations arise in the population; the latter approximation in Eq 7 assumes δF ≪ 1. Note that although in the simulations we use the full form for the fixation probability, fitness effects are typically small (δF ≪ 1) in the simulations, so the substitution rates only depend on the population-scaled fitness changes 2NδF which, for a given mutation, is proportional to 2NκF. Finally, we allow continuous ‘mutations’ in the morphogen steepness parameter α, chosen from a Gaussian distribution of standard deviation δα = 0.5 and assign it an 10 effective base-pairs, which are used when assigning relative weight in the kinetic Monte Carlo scheme, where the total number of base-pairs is ℓG = 60.
We determine the Malthusian or log fitness of the spatial gene regulation, from the resulting concentration profile [TF](x) by use of a functional that promotes expression of theTF in the anterior half, whilst penalising expression in the posterior half, with truncation selection below a critical value F*:
F = { κ F ln ( W ) if κ F ln ( W ) > F * - ∞ if κ F ln ( W ) < F * (8)
where,
W [ [ T F ] ( x ) ] = ∫ 0 L / 2 [ T F ] ( x ) d x - ∫ L / 2 L [ T F ] ( x ) d x L 2 max x { [ T F ] ( x ) } . (9)
where we use a value of F* = −1.6 × 10−3, which corresponds a value of W ≈ 0 . 2, when κF = 10−3. Strictly, an inviability on a lineage or a hybrid should correspond to W = 0 or F* = −∞, however, these values were chosen to so that a reasonable number of incompatibilities arise in a simulation; for comparison the typical maximum of the integral W ≈ 0 . 6. In this paper, we explore how the changing the population scaled strength of selection (2NκF) affects the rate of reproductive isolation, by keeping κF fixed and varyingN accordingly. Note that although here the exact form of the fitness is slightly different to the one used in [17], the qualitative behaviour is the same (S1 Text).
The speciation simulations consist of two replicate simulations starting with the same common ancestor and with the same fitness function. We draw the common ancestor from the equilibrium distribution for G and α. To do this we start from a random initial genome, and run one long simulation for 100,000 substitutions for a fixed scaled population size 2NκF, in order to effectively equilibrate the system (typically 10,000 substitutions are required to adapt to an ensemble of fit states). This represents a reference equilibrium state; different random draws from the equilibrium distribution then consist of running the simulation for a further 100 substitutions.
Given a pattern of hybrid incompatibilities, for example, as shown in Fig 4, if there is a 2-way DMI (e.g. between C and α loci, which we denote ICa), then all four hybrid-genotypes containing this DMI (e.g. RMCa, RmCa, rMCa, rmCa) are inviable; these genotypes define a two-dimensional subspace (or face) of the hypercube. Similarly, the points (e.g. rmcA, RmcA, which we denote ImcA) containing a 3-way DMI form a one-dimensional subspace (or line), while a 4-way DMI takes up only a single point in the four-dimensional hypercube. These different 2-way, 3-way and 4-way DMIs are the fundamental incompatibility types which we seek to explain the pattern of hybrid inviable genotypes observed, for example, as in Fig 4a.
However, this decomposition is hugely underdetermined, as there are only 24 − 2 = 14 possible hybrids (not including the well-adapted genotypes of line 1 and line 2) and a total of Imax = 3L + 1 − 2L+1 = 50, different fundamental incompatibilities, for L = 4 loci. This arises as the total number of n−point DMIs is ( 2 n - 2 ) ( L n ), as there are ( L n ) combinations ofN loci amongst L total loci and then considering a binary choice of alleles across both lines, there are a total of 2n allelic combinations or states, 2 of which are the fit allelic combinations where all alleles come from one lineage or the other giving 2n − 2. For example, between each pair of loci there are 22 − 2 = 2 mismatching combinations of alleles (e.g. rM and Rm) that could give DMIs and ( L 2 ) = L ( L - 1 ) / 2 = 6 pairwise interactions. A similar argument would give a total of 24 3-way DMIs as there are 23 − 2 = 6 mismatching combinations of alleles at 3 loci (e.g., excluding rmc and RMC) and ( L 3 ) = 4 3-way interactions and similarly, 14 ( L 4 ) = 14 for 4-way interactions. In total, the max number of DMIs is I m a x = ∑ n = 2 L ( 2 n - 2 ) ( L n ) = 3 L + 1 - 2 L + 1, which for L = 4 loci is Imax = 50.
The approach we take is to find only those combinations of fundamental DMIs that have the smallest total number that can explain the pattern of hybrid incompatibilities, which from a Bayesian perspective would have the smallest Occam factors [86]; for instance, as shown in Fig 4b the list of 6 incompatible hybrid genotypes rmCa, rMCa,RmCa,RMCa,Rmca,RMca, shown by red crosses, can be explained most parsimoniously by three different minimal combinations of fundamental DMIs, each with only 2 DMIs (see main text).
|
10.1371/journal.pntd.0002440 | A Systematic Review on the Epidemiological Data of Erythema Nodosum Leprosum, a Type 2 Leprosy Reaction | Erythema Nodosum Leprosum (ENL) is a humoral immunological response in leprosy that leads to inflammatory skin nodules which may result in nerve and organ damage, and may occur years after antibiotic treatment. Multiple episodes are frequent and suppression requires high doses of immunosuppressive drugs. Global occurrence is unknown.
Systematic review of evidence on ENL incidence resulted in 65 papers, predominantly from India (24) and Brazil (9), and inclusive of four reviews. Average incidences are based on cumulative incidence and size of study populations (n>100). In field-based studies 653/54,737 (1.2%) of all leprosy cases, 194/4,279 (4.5%) of MB cases, and 86/560 (15.4%) of LL cases develop ENL. Some studies found a range of 1–8 per 100 person-years-at-risk (PYAR) amongst MB cases. Hospital samples indicate that 2,393/17,513 (13.7%) of MB cases develop ENL. Regional differences could not be confirmed. Multiple ENL episodes occurred in 39 to 77% of ENL patients, with an average of 2.6. Some studies find a peak in ENL incidence in the first year of treatment, others during the second and third year after starting MDT. The main risk factor for ENL is a high bacteriological index.
Few studies reported on ENL as a primary outcome, and definitions of ENL differed between studies. Although, in this review averages are presented, accurate data on global and regional ENL incidence is lacking. Large prospective studies or accurate surveillance data would be required to clarify this. Health staff needs to be aware of late reactions, as new ENL may develop as late as five years after MDT completion, and reoccurrences up to 8 years afterwards.
| This systematic review addresses an underpublicized and yet highly significant leprosy topic. Erythema Nodosum Leprosum (ENL) is a serious complication in multi-bacillary (MB) leprosy that may lead to severe disability. Inflammatory skin nodules may result in nerve and organ damage and require high doses of immunosuppressive drugs. ENL can occur long after patients are released from antibiotic treatment. Frequency and severity of ENL is unknown; this review confirms the lack of accurate data at global, regional, and national levels. Available data indicates that ENL incidence ranges between 0.7–4.6% of all MB cases and late reoccurrence up to 8 years after release from treatment. ENL episodes often reoccur, with an average of 2.6 times. The main risk factor for ENL is a high bacteriological index. Additionally, data indicate a wide variation of ENL occurrence between and within countries. The conclusions demonstrate the need for increased awareness about ENL, in research, patient surveillance, and in programme management.
| Erythema Nodosum Leprosum (ENL), the main symptom of a type-2 reaction in leprosy, is caused by a humoral immune response to Mycobacterium Leprae [1]. Patients develop fever and tender/painful subcutaneous nodules, often in the face or extensor surfaces of the limbs [2]–[4]. ENL may also damage nerves, skin, eyes, and testes, and involves systemic illness including fever, weight loss and pain [5], all of which result in extreme discomfort. The majority of patients develop multiple episodes of ENL. Severe cases require the use of potent immunosuppressants, and the steroid-induced side effects may increase mortality and morbidity [3], [6]. Furthermore, the limited use of teratogenic thalidomide presents another challenge [5]. The economic impact of ENL is unknown, but likely to be considerable.
ENL is confined to leprosy patients classified as BL or LL (Ridley-Jopling), comprising the multi-bacillary (MB) patient group, as defined by WHO. In 1981 this concerned patients with a bacteriological index (BI) of 2 or more, changing to any positive skin smear in 1988. In 1995 this was widened further; MB comprising any patients with more than five skin lesions [7]. The proportion of MB cases among new leprosy patients varies between countries and is increasing [8], [9]. Global incidence of MB leprosy was 139,125 in 2009, and is decreasing [8]. ENL may occur before, during or after antibiotic treatment, and several years later [10]. It can occur as a single acute episode, but frequently develops into a chronic condition with recurrent episodes [3], [5]. Immune responses causing ENL are triggered by high loads of fragmented bacilli in skin tissue [11].
Although adequate surveillance systems are used to estimate global leprosy prevalence and inform drug supply, this is not available for estimating incidence, frequency and severity of ENL [12]. Geographic variation in ENL prevalence complicates accurate estimations [13], and hampers logistics in drug supplies. For this reason, a systematic literature review was conducted to determine global incidence of ENL, inclusive of incidences of recurrent and severe ENL and contributing factors.
A systematic literature search was conducted in January 2011 in five databases (Pubmed (MEDLINE), EMBASE, LILACS, SCOPUS, Scielo, and Ajol). Keywords used were: <lepro* OR lepra* OR hansen*, Erythema Nodosum OR ENL OR (type 2)>, AND <incidence OR prevalence OR cohort>. Reference lists of included studies were checked and national leprosy control managers and leading leprologists were asked for additional (un-)published articles.
Studies, published after 1980, presenting data on incidence or prevalence of ENL were selected. Focus was on papers in English, whereas Portuguese, Spanish or French studies were included after Google-translation. No separate search was conducted on adverse events and risk factors, but information was retrieved from the included studies. A distinction was made between acute and chronic ENL as well as severe and mild forms [2]. We included all studies reporting on the onset of ENL. The following forms of ENL were included: single acute episodes, multiple acute episodes, and chronic ENL (ENL lasting for more than 6 months, in either single or multiple episodes) [2].
Data extraction regarding onset, risk factors, severity and reoccurrence of ENL was completed by the first author and co-reviewed by the second author. A structured form was designed to retrieve data on the setting (country, region, place studied, other characteristics), methods (study period, design, sampling, data sources, representativeness), study design and characteristics (sample size, population, leprosy classification (Ridley-Jopling), inclusion criteria, ethnicity, gender, age group, other (health) characteristics and study variables (follow-up time, loss to follow up, and MDT-, ENL-, or other treatment, serious adverse events). Evidence was graded according to the Oxford Centre for Evidence Based Medicine guidelines [14].
Depending on availability, incidence rates of ENL are presented in person years at risk (PYAR). Where proportions or actual numbers of patients developing ENL were reported, ENL incidence is based on the proportion of persons at risk (i.e. total number of leprosy cases, MB cases or specific Ridley-Jopling classifications). We considered cases MB as reported in the articles. Occurrence is only presented when sample sizes exceeded 100 at risk (MB) population, for field and hospital studies separately. The average incidence of ENL was calculated taking all different sample sizes together.
The search resulted in 914 records (Figure 1). Scanning the references and consultation with experts resulted in an additional 10 papers. 65 papers met the inclusion criteria. Four literature reviews were analysed separately [2], [12], [15], [16]. One relevant workshop report was included [17].
The majority of studies were from India (24) and Brazil (9), the two countries with the highest incidence of new leprosy cases [8]. Table 1 summarises the characteristics of included studies. Approximately one third of the studies included a minimum of 300 persons at risk for ENL and another third between 100 and 300 persons. 23 studies had sample sizes below 100 persons at risk [10], [18]–[39]. Studies were either cross-sectional or retrospective cohort analyses. Less than half of them aimed specifically at ENL occurrence. The majority reported ENL frequency while their main focus was on clinical or epidemiological aspects of leprosy.
Only five studies reported ENL incidence rates in person years at risk (PYAR). Follow up varied between 2 and 7 years. Incidence rates ranged from 1 to 8 per 100 PYAR [40], [41] among MB leprosy patients (figure 2).
Six prospective [30], [41]–[45] and five retrospective studies [17], [40], [46]–[48] gathered data from a control programme and most accurately reflected ENL occurrence.
Table 2 demonstrates that cumulative ENL incidence varied from 0.2% among all leprosy patients in an Indian study [49] and up to 4.6% in a Chinese study [48], with an average of 1.2%. ENL incidence among MB cases varied from 1.0% in a one year cross-sectional Indonesian study [46] to 8.9% in an Indian cohort [47], with an average of 4.5%. From the latter study, it was not clear if referral cases were included, which may explain the relatively high percentage. Three prospective studies were from the ALERT leprosy control services [41], [42], [45]. Interestingly, cumulative ENL incidence was 2.5% among MB cases after an average follow-up of 2.5 years [45], whereas after 10 years this was doubled [41].
Table 3 indicates the cumulative ENL incidence in 28 studies (>300 patients), ranging from 2–28.9% of MB cases. Calculation from studies with at least 100 patients reveals that on average 13.7% of MB cases developed ENL. In four studies this was more than 30% [50]–[53]. Studies with largest population sizes indicated lower cumulative incidence rates.
Sixteen studies reported ENL occurrence for the Ridley-Jopling classifications (Figure 3). Findings differed widely between countries. Among the four field studies [41], [42], [44], [47] ENL for LL leprosy ranged from 11.1% [42] to 26% [44] with an average of 86/560 (15,4%). For BL cases this varied from 2.7% [42] to 5.1% [47], on average 51/1231 (4,1%). In hospital based studies higher proportions were found, in Brazil up to 56.4% [52] and in India a range of 24.2 [54] to 50.9% [55].
ENL reoccurrence was disproportionately higher in hospital-based studies. Multiple episodes were found in 39% [56] to 77.3% [50] of ENL patients, with an average of 2.6 episodes. Various studies reported 24% of all ENL cases having more than four episodes: the longer the follow-up the more episodes were recorded. Three larger studies (>100 ENL cases, see Table 4) found a range from 49% [57] to 64.3% [58]. Similar ranges were found in field based studies: 44 to 63% of all ENL cases have multiple ENL episodes [41], [44], [45].
There was discrepancy in the average number of ENL episodes, as is evident in the following findings. In a cohort from Zaire [59] there was an average of 1.8 episodes, compared to 3.2 episodes (CI 2.7–3.5), in a study from India [60]. A Thai cohort revealed that ENL episodes often occurred more than 4 times [50]. A large hospital study in India reported that 23.5% of reoccurring cases (15.1% of all ENL cases) had four or more episodes [58]. Similar proportions were found in a Brazilian cohort [52], whereas other studies in India [47] and Nepal [57] found four or more episodes among 5 and 7% of ENL patients respectively. In Ethiopia, almost one third of ENL patients developed a chronic condition lasting more than 2 years [41]. Episodes lasted from 14 days [19] to 26.1 weeks [61]. Total ENL episodes and ENL-free intervals in India found an average of 18.5 months (CI 15.4–21.5) [60].
Six studies distinguished between mild and severe ENL, finding that 30–50% of ENL cases are (moderate to) severe. They represented 0.7–2.0% of all MB leprosy patients and 0.7% of all newly detected cases [46], [62]. However, descriptions of severity differed between the studies. Shortened MDT duration (12 months) almost doubled the incidence of moderate to severe ENL [61], [63]. Poor referral practices leave some severe reactions under-diagnosed [40], while hospital figures misrepresent the field situation [47].
Findings on the onset of ENL differ. Most studies indicated that the incidence of ENL during MDT was at least twice as high than at the time of the initial diagnosis [37], [42], [44], [50], [64], [65]. ENL incidence was highest in the first year of MDT [17], [37], [42], [44], [57], [58], [64]. There were a few exceptions, a from the Philippines (10 year follow-up) [43], [61] and India (13 years follow-up) [58] where most ENL was diagnosed during the second and third year after starting MDT, as was the case in Ethiopia [41].
A study conducted in an Indian hospital found 3% of MB patients developed ENL two years after completing MDT (follow-up 74 months) [58]. Longer term follow up showed ENL three [37], five [66], seven [41], or even eight years after MDT [58]. Similar findings (ENL occurring 5–7 years later) were reported in India [17].
Multiple studies [22], [23], [52], [57], [58], [60], [62] reported a correlation between the bacteriological index (BI) and ENL up to a 8.6 (CI 2.3–32) times higher risk when having a BI of six [41]. Discrepancies are evident Nepali patients with a BI>4+ had a 39% higher risk of ENL (OR; 1.39 (CI 1.11–1.76) adjusted for age) [57] and in India a BI≥4 was associated with an Odds Ratio of 5.2 (2.1–12.9) [60]. Inherent to BI, lepromatous leprosy is a significant risk factor [58], [67]. An Ethiopian study found a 9.6 times higher ENL incidence among LL patients compared to BL or BB (X2 = 18.7, p<0.005) [42]. Odds ratios for the prevalence of ENL in LL as compared to BL varied from 2.8 (1.59–5.2; adjusted for age and BI) [57] to 8.4 (CI 4.6–15.4) [60]. LL cases have higher chances to suffer multiple rather than single ENL episodes (OR 2.94, p = 0.052) [57]. This finding was disputed, however, by a controlled clinical trial conducted in India, which reported no such differences [55].
It has been claimed that the risk of developing ENL has decreased since introducing MDT [42], [51], [54], [57], due to the ENL suppressant effect of clofazimine [22], [51], [68]. A recent multi-country cohort study indicated more severe and longer-lasting episodes of ENL among patients who received 12 as compared to 24 months of MDT, although ENL frequency as such was similar [61], [63].The Bombay Leprosy Project had similar findings: 55.9% and 35.8% of cases receiving 12 and 24 months MDT respectively had a type1 or 2 reaction [17].
Gender is generally not a risk factor for ENL [41], [52], [55], [57], [60], [62], [63]. Some studies appear to challenge this, as a large hospital study in India found a male predominance [69], and a large Indian cohort reported a higher risk for women [58]. These differences, however, may be due to differences in health seeking behaviour [69].
Seemingly, age is not a risk factor for ENL [41], [50], [58], [60], [63], although a Nepali cohort indicated decreased risk for those older than 40 (adjusted OR 0.69, CI 0.5–0.94) [57], and a higher ENL incidence was seen in patients diagnosed with leprosy in their adolescence, but these findings are not supported elsewhere [50].
Pregnancy and lactation appears to be a significant precipitating factor for severe and recurrent ENL [54]. Additionally, hormonal changes are implicated in a study from India: 62% of 32 ENL in women were associated with pregnancy or lactation and 21% with menopause [69]. A major Ethiopian study among pregnant leprosy patients found an increased ENL incidence (22% among BL and 59% among LL patients). Some episodes continued until 15 months after delivery [24].
Minimal evidence has been published regarding co-morbidities as risk factors for ENL, with the exception of HIV that suggested a 5.3 times higher risk for developing ENL (RR 5.3, CI 1.0–2.8). However, numbers (n = 10) were too low to be conclusive [41]. A recent review concluded there is no reliable data on the effect of HIV [13]. In other studies, malaria and tuberculosis were reported to trigger ENL [24], [54].
Presenting a comprehensive overview of the epidemiological data on ENL incidence, was difficult due to lack of available and reliable data. Furthermore, few studies reported ENL as a primary outcome. Findings were complicated by the inconsistency in case definitions of ENL. Additionally, much of the data drawn on in this review was prior to the WHO-MDT era, asserting that 50% of LL patients and 25% of BL patients developed ENL in the course of the disease [12], [70]. This review establishes that prevalence rates are highly variable, in field cohorts up to 26% LL and 5.1% BL patients, and 37% in a hospital sample of MB patients. In an effort to overcome the difficulty of variations in ENL occurrence, average incidences were calculated in field based populations for all leprosy cases (1.2%) and for MB leprosy cases (4.5%). In hospital samples these percentages were higher. This review could not confirm any regional differences and found differences between and within countries.
Few comprehensive prospective studies reported ENL incidence in terms of person years at risk and controlled for confounding factors. Estimates presented in this paper should therefore be taken with caution. We underline the lack of reliable epidemiological data due to the absence of a universally-accepted set of norms and standardized nomenclature as well as lack of awareness and recording [52]. Standardized definitions should be set globally and would facilitate the collection of better quality data. Well-designed field studies to ascertain this have been called for [71]. All findings considered, the authors are of the opinion that if national estimates are needed (e.g. for estimating local needs for clofazimine to treat severe ENL), this is best done on the basis of local evidence and indications by experienced programme and clinical staff.
Alarmingly, ENL reoccurs, and often more than four times, in almost a half of initial ENL reported episodes. Multiple episodes were found in 39–77.3% of ENL patients. Calculations indicate an average of 2.6 episodes per ENL patient. Episodes of ENL peak during MDT, but also occur up to 7–8 years after release from treatment [65]. Therefore, it is imperative that both patients and health workers are on the alert for development of late episodes of ENL [17], [60]. It is of major concern that leprosy control programmes do usually not advocate standardized follow-up [65].
The main risk factors for developing ENL are related to a high bacteriological index and a BL/LL classification in the Ridley-Jopling spectrum [13]. The ENL-suppressive effect of clofazimine, within the MDT regimen, is generally acknowledged [68]. More severe and longer-lasting ENL episodes occur in shorter duration MDT-course (12 months as opposed to 24).
There was no conclusive evidence for co-morbidities or age as risk factors. Possible precipitating factors for ENL included hormonal changes occurring in pregnancy, lactation, menopause, and puberty. Additional findings suggest that intercurrent infection, vaccination and psychological stress, are implicated (Pfaltzgraff and Ramu in Clinical Leprosy) [70]. This appears to be supported by empirical evidence only, and was not confirmed by this literature review. This may be explained by the lack of large prospective studies and relatively low incidence of ENL and co-morbidities. Perhaps the analysis of large existing data sets (e.g. BANDS, AMFES, INFIR, Brazil, possibly other countries) may help in identifying precipitating factors. Prospective studies would be required to elucidate hormonal and genetic risk factors [12].
Most of the literature regarding ENL occurrence was descriptive data, and only a few studies had an adequate sample of patients. Characteristics of cases and populations, definitions, outcomes and procedures were not always systematically described, making a statistical meta-analysis impossible.
To what extent study samples reflected the leprosy population at large was often difficult to assess, as distinction between field and hospital based studies was not clear in each publication. Higher ENL rates were found in hospital based studies, although it is not known how many severe ENL cases actually arrive in referral clinics. In the hospital based studies the population size of which these cases are drawn is not known. Field based studies often only report patients with ENL who actually seek help. Only few appropriate prospective studies could be found that are representative for the most peripheral level.
The majority of publications lacked both a clear case definition of ENL and a clear description of the diagnostic procedure. Both may vary between settings and studies. Only a few studies make a distinction between mild and severe ENL [60]–[63], and mild ENL may have been overlooked and thus incidence rates underestimated.
Considering the limited evidence and the significant differences in ENL rates, country specific data should be interpreted with great caution. The wide range in cumulative incidence and variation of ENL found in this review is most likely explained in terms of duration of treatment and follow-up of the subjects. Furthermore, the widening definition of MB leprosy since 1981 [4], [72] would have decreased rates of ENL. LL patients would be the most appropriate risk group for ENL to report on, especially in research papers. In this study, however, MB was the most common denominator in the articles that were identified. Ideally, future studies on ENL should report incidence in person years at risk, both for MB and Ridley-Jopling classification.
None of the studies included in this review looked at explicitly at the social and medical costs related to ENL.
This review provides a systematic overview of available evidence regarding ENL occurrence. Wide ranges were found between and within different countries. Despite these limitations, a global average incidence was calculated. This review has established that reliable data on ENL occurrence is lacking, and could only be obtained through large comprehensive prospective studies or data obtained from accurate ENL surveillance. Furthermore, studies investigating risk and precipitating factors for ENL would be useful in diagnosis and prevention.
|
10.1371/journal.pgen.1002157 | Expression of Arf Tumor Suppressor in Spermatogonia Facilitates Meiotic Progression in Male Germ Cells | The mammalian Cdkn2a (Ink4a-Arf) locus encodes two tumor suppressor proteins (p16Ink4a and p19Arf) that respectively enforce the anti-proliferative functions of the retinoblastoma protein (Rb) and the p53 transcription factor in response to oncogenic stress. Although p19Arf is not normally detected in tissues of young adult mice, a notable exception occurs in the male germ line, where Arf is expressed in spermatogonia, but not in meiotic spermatocytes arising from them. Unlike other contexts in which the induction of Arf potently inhibits cell proliferation, expression of p19Arf in spermatogonia does not interfere with mitotic cell division. Instead, inactivation of Arf triggers germ cell–autonomous, p53-dependent apoptosis of primary spermatocytes in late meiotic prophase, resulting in reduced sperm production. Arf deficiency also causes premature, elevated, and persistent accumulation of the phosphorylated histone variant H2AX, reduces numbers of chromosome-associated complexes of Rad51 and Dmc1 recombinases during meiotic prophase, and yields incompletely synapsed autosomes during pachynema. Inactivation of Ink4a increases the fraction of spermatogonia in S-phase and restores sperm numbers in Ink4a-Arf doubly deficient mice but does not abrogate γ-H2AX accumulation in spermatocytes or p53-dependent apoptosis resulting from Arf inactivation. Thus, as opposed to its canonical role as a tumor suppressor in inducing p53-dependent senescence or apoptosis, Arf expression in spermatogonia instead initiates a salutary feed-forward program that prevents p53-dependent apoptosis, contributing to the survival of meiotic male germ cells.
| The intimately linked Arf and Ink4a genes, encoded in part by overlapping reading frames within the Cdkn2a locus, are induced by oncogenic stress, activating the p53 and Rb tumor suppressors, respectively, to inhibit proliferation of incipient cancer cells. As such, expression of the p19Arf and p16Ink4a proteins is undetected in most normal mouse tissues. However, p19Arf is physiologically expressed in mitotically dividing spermatogonia, the progenitor cells that differentiate to form meiotic spermatocytes in which Arf expression is extinguished. We show that, instead of provoking cell cycle arrest or death, Arf expression in spermatogonia facilitates survival of their meiotic progeny, ensuring production of normal numbers of mature sperm. When Arf is ablated, meiotic defects ensue, along with p53-dependent cell death of spermatocytes, indicating an unexpected role of p53 in monitoring meiotic progression. Surprisingly, it is the absence of p19Arf rather than its induction that enforces p53 expression in this setting. Co-inactivation of Ink4a compensates for Arf loss by fueling proliferation of spermatogonial progenitors, but does not correct meiotic defects triggered by Arf loss. Although the Arf and Ink4a tumor suppressors are expected to restrain cellular self-renewal, Arf plays an unexpected role in male germ cells by facilitating their proper meiotic progression.
| The Cdkn2a-Cdkn2b gene cluster (also designated Ink4-Arf) encodes two polypeptide inhibitors (p16Ink4a and p15Ink4b) of cyclin D-dependent kinases (Cdk4 and Cdk6), as well as a third protein (p19Arf) that antagonizes the Mdm2 ubiquitin E3 ligase to activate p53 [1]. Although the Ink4a and Ink4b genes likely arose through gene duplication, the structure of the Ink4-Arf gene cluster is highly unusual, as major portions of the p16Ink4a and p19Arf proteins are encoded by alternative reading frames of a shared exon [2]. Induction of p16Ink4a and p15Ink4b prevents the phosphorylation of the retinoblastoma protein (Rb), thereby maintaining Rb in its growth suppressive state and preventing entry into the DNA synthetic (S) phase of the cell division cycle. In contrast, p19Arf expression elicits a p53-dependent transcription program that either enforces cell cycle arrest or triggers apoptosis, depending on cell type, physiologic setting, and collateral modulating signals [1]. The Ink4-Arf genes prevent cell proliferation by implementing Rb- and p53-dependent programs that enforce cellular senescence and inhibit tissue regeneration as animals age, but their intimate genetic linkage facilitates their coordinate repression in embryonic and adult tissue stem cells, thereby allowing self-renewal [3], [4]. Deleterious growth-promoting stimuli conveyed by activated oncogenes induce Ink4-Arf gene expression and engage both p53 and Rb to counter untoward cellular proliferation. Not surprisingly, bi-allelic deletion of the Ink4-Arf gene cluster abrogates this form of tumor suppression and is one of the more frequent events in human cancer.
Despite its canonical role as an inducer of p53 in response to oncogene signaling, Arf also has p53-independent tumor suppressive activity. Deletion of Arf together with Mdm2 and p53 expands the spectrum and decreases the latency of cancers that spontaneously arise in mice lacking p53, p53 and Mdm2, or Arf alone [5]. Although highly basic p19Arf (∼20% arginine) has been reported to physically interact with more than 25 different proteins other than Mdm2, the role of p19Arf, if any, in regulating the functions of these putative “target” proteins remains controversial [6]. Indeed, numerous reports that p19Arf regulates such diverse processes as ribosomal biosynthesis, transcription, DNA repair, apoptosis and autophagy in a p53-independent manner have generally relied on experiments performed with cultured cells but have not been buttressed by more extensive in vivo analyses.
Although the Ink4-Arf locus is not detectably expressed under most normal physiologic conditions, eye and male germ cell development provide notable exceptions [7]. Arf is required for early postnatal regression of the hyaloid vasculature in the vitreous, so that Arf-null mice form a retrolenticular mass predominantly composed of pericytes; the abnormal accumulation of these cells disrupts the retina and lens and leads to blindness [8]. Arf inactivation also results in a significant reduction of sperm production through as yet poorly defined mechanisms, although young male mice remain fertile [9]. In contrast, Arf-null females have no discernable reproductive defects.
Spermatogenesis involves a stereotyped sequence of mitotic and meiotic divisions followed by sperm differentiation [10]. In mice, male germ cell progenitors (gonocytes) renew in the testis between days 1–7 postpartum (P1–P7) and generate spermatogonia that line the basement membranes of developing seminiferous tubules [11], [12]. At P7–P10, spermatogonia divide to form preleptotene spermatocytes that detach from the basement membrane, are displaced toward the tubular lumen, and enter meiosis-I as primary leptotene spermatocytes. During the extended prophase of meiosis-I, homologous pairs of maternal and paternal chromosomes align to form synaptonemal complexes and exchange genetic information through homologous recombination [13]. Meiosis-I is completed by P18, and is followed rapidly by meiosis-II, and by spermiogenesis (P19–P35), after which the first mature spermatozoa enter the epididymis. As spermatogenesis continues throughout life, spermatogonia within mature seminiferous tubules remain localized on the peripheral tubular basement membrane, whereas spermatocytes, spermatids, and mature sperm are arranged in a sequential order from the periphery towards the lumen [10].
Intriguingly, p19Arf is transiently expressed in mitotically dividing spermatogonia, but not in the meiotic cells that arise from them [9]. Here, we provide genetic evidence demonstrating that Arf expression initiates a germ cell autonomous program that protects meiotic spermatocytes from undergoing p53-dependent elimination. This physiologic function of p19Arf directly contrasts with its role as a tumor suppressor in inducing p53.
Lineage tracing experiments in the mouse previously revealed that all viable male germ cells are derived from spermatogonial progenitors in which transient Arf expression neither inhibits proliferation nor subsequent meiotic commitment [9]. Underscoring these findings, expression of p19Arf in young adult mice is observed in all types of spermatogonia, but not in Sox9-expressing Sertoli cells on the tubular basement membrane or in DAPI-stained intratubular spermatocytes, spermatids, or sperm (Figure 1A). The fact that p19Arf is not detected in cells that have detached from the basement membrane implies that Arf expression is extinguished at or near the primary spermatocyte stage of germ cell differentiation. Consistent with this interpretation, the Arf protein does not co-localize with Dmc1 [9], a meiotic recombinase expressed in leptotene spermatocytes. In the mature testis, spermatogenesis occurs in waves along the length of the seminiferous tubules, so that cross sections capture tubules in which dividing spermatogonia are in synchronous phases of the cell cycle. When five month-old mice injected intraperitoneally with BrdU were sacrificed two hours later, dual immunofluorescence analysis revealed that many cells on the tubular basement membrane that had synthesized DNA also expressed p19Arf (Figure 1B). Similarly, at P12 when the number of mitotically cycling progenitors exceeds those of more differentiated germ cells, p19Arf was co-expressed with cyclin D1, a G1 phase marker of proliferating spermatogonia [14] (Figure 1C), and strikingly, was detected during all stages of mitosis (Figure 1D, 1E). Therefore, in spermatogonia, p19Arf is expressed throughout the cell division cycle without interfering with proliferation.
Total body weights of age-matched wild-type, Arf-null, Ink4a-null, and Ink4a-Arf double-null mice are equivalent, but testis weights of Arf-null animals were reduced relative to those of wild-type controls (Figure 2A), and this was associated with a significant reduction in numbers of mature sperm by the time Arf-null mice were two months old (Figure 2B). Nonetheless, young Arf-null males remain fertile, and despite the widespread use of independently derived Arf-null strains by us and others, there is no suggestion that young fertile males produce reduced litter sizes. Hence, defects in spermatogenesis were not previously appreciated.
Knock-in of a cDNA encoding Cre recombinase in place of the first Arf exon creates a functionally null Arf allele that expresses Cre in lieu of p19Arf under the control of the Arf promoter. Crossing ArfCre/+ females to homozygous males containing Arf alleles flanked by LoxP sites (“floxed” alleles) specifically results in the inactivation of Arf function in the testis of compound heterozygous ArfCre/Flox male offspring. Although penetrance of Cre expression is not complete, more than 90% of spermatogonia in the seminiferous tubules of P21 mice had no detectable anti-p19Arf fluorescence signals [9]. Overall, while p19Arf was detected in the testes of haplo-insufficient ArfCre/+ mice, any residual levels of the protein in ArfCre/Flox testes were too low to be detected by immunoblotting analysis (representative data illustrated in Figure 3), confirming significant Cre-mediated Arf deletion in this setting. We therefore used this “targeted” deletion approach to compare the Arf loss-of-function phenotypes of ArfCre/Flox males with those of Arf−/− males. Analysis of testis weights revealed no differences between those of ArfCre/Flox mice and wild-type controls (Figure 2C). However, the sperm counts of ArfCre/Flox animals were reduced to levels approaching those of Arf−/− males (Figure 2D). Notably, the Arf-Cre or Arf-Flox alleles alone had no significant effects in limiting sperm production unless coexpressed in compound heterozygotes. Therefore, tissue-restricted effects of Arf inactivation independently recapitulated those seen in mice that completely lack Arf function.
Hormone signaling networks are involved in the proper control of spermatogenesis. Key regulatory gonadotrophins include luteinizing hormone (LH) and follicle-stimulating hormone (FSH) secreted by the anterior pituitary gland, and testosterone produced by testicular interstitial Leydig cells. The considerable day-to-day and even hour-to-hour variation over a 30-fold range in plasma testosterone levels in age-matched mice of a single strain precluded accurate measurements of strain-specific differences, even in a relatively large sample size (Figure 4) [15]. Importantly, however, no discernable defects in pituitary or Leydig cell development have been observed in Arf-null, Ink4a-null, or doubly-deficient mice, and no significant differences were observed in the ranges of serum FSH and LH among all genotypes examined (Figure 4). These findings suggest that spermatogenesis defects in Arf-deficient mice are not a secondary consequence of hormonal imbalances.
Unlike Arf-null males, those lacking functional Ink4a instead exhibit increased testis weights and produce higher numbers of sperm than wild type mice (Figure 2A and 2B). Cdk4, the major target of p16Ink4a protein inhibition in the adult testis, is expressed at maximal levels at the earlier stages of spermatogenesis, where spermatogonia undergoing mitotic cell divisions predominate [16], [17], and Cdk4 inactivation compromises male fertility [18], [19]. We therefore quantified the in vivo incorporation of BrdU in spermatogonia of young adult wild-type, Arf-null, Ink4a-null, and Ink4a-Arf-null mice by counting stained cells that had entered S phase during a two-hour pulse. The S phase fractions of wild-type and Arf-null spermatogonia did not differ from each other (Figure 5), implying that the failure of Arf-null mice to produce normal numbers of sperm reflects a loss of meiotic cells or their immediate progeny rather than spermatogonia. In contrast, we observed a significant two-fold increase (p<0.0001, Student's t-test) in the relative number of S phase spermatogonia from both strains that lack Ink4a (Figure 5). Looking only at testis weights and sperm counts in Ink4a-Arf double-null animals, Ink4a inactivation appears to compensate for loss of Arf function (Figure 2A and 2B), presumably by fueling the production of a greater number of mitotic progenitors. Together, the consequences of these two independent loss-of-function effects rebalance testis size and sperm output in the doubly null strain. In this sense, these two “tumor suppressor” genes play opposing physiologic roles in male germ cell development.
Testes from two month-old Arf-null mice exhibited a significant increase in the numbers of apoptotic (TUNEL-positive) cells when compared to age-matched wild-type controls (Figure 6A and 6E). The vast majority of apoptotic cells are spermatocytes as judged by the topological relation of TUNEL-positive cells to the expression of the meiosis-specific strand-exchange protein Dmc1, which is expressed during early prophase-I (Figure 6A). Notably, however, intratubular TUNEL-positive cells were not stained with antibodies to Dmc1, implying that Arf-null cells die during a later stage of germ cell development after Dmc1 expression is greatly diminished. To examine this issue further, we conducted TUNEL staining of meiotic chromosome spreads. Characteristic stages of prophase during meiosis-I can be marked by staining chromosomes with antibodies to synaptonemal complex proteins, such as the axial element component SYCP3 [20], and by several ancillary criteria (see Materials and Methods). Unlike pachytene cells from wild-type mice, those from the Arf-null strain exhibited considerable TUNEL staining (Figure 6C) with a concomitant reduction in the fraction of Arf-null diplotene spermatocytes (Figure 6F) that correlated with decreased sperm production (Figure 2B). Inactivation of Ink4a alone did not trigger spermatocyte apoptosis nor limit apoptosis in the Arf-null background (Figure 6E) reinforcing the conclusion that the two closely linked genes play fundamentally different roles within the male germline.
Arf−/−; p53−/− doubly-deficient males are even more susceptible to spontaneous tumor development than mice lacking either Arf or p53 alone [5]. However, the young tumor-free males produce sufficient viable sperm to remain fertile. Inactivation of p53 restored testis weights and sperm production in Arf-null males (Figure 2A and 2B), prevented the apoptotic elimination of Arf-null germ cells (Figure 6D and 6E), and restored the number of diplotene spermatocytes (Figure 6F). Accordingly, higher levels of p53 were detected in whole testis lysates from Arf-null mice as compared to those in wild-type mice (Figure 7). Thus, in direct contrast to the role of p19Arf in triggering a p53 response following abnormal hyperproliferative stress in somatic cells, it is instead the absence of Arf expression in spermatogonial progenitors that impairs the fidelity of meiotic progression and ultimately leads to p53-dependent elimination of Arf-null primary spermatocytes.
Histone H2AX is phosphorylated at serine-139 in response to DNA strand breaks caused by ionizing radiation [21], UV irradiation [22], replication stress [23], [24], failure of nucleotide excision repair [25], and at the leptotene stage of meiosis prior to synaptonemal complex formation [26]. H2AX phosphorylation also occurs in spermatocytes in a DNA damage-independent manner during formation of the sex body, a heterochromatic sub-nuclear domain encompassing the nonhomologous parts of the X and Y chromosomes [27]. Remarkably, staining of testis sections and immunoblotting of whole testis lysates revealed a profound increase in global γ-H2AX levels when Arf was inactivated (Figure 8A and 8B). Again, inactivation of Ink4a neither recapitulated nor ameliorated this Arf-null defect (Figure 8A and 8B). Germ cells at the periphery of Arf-null seminiferous tubules exhibited the greatest increase in γ-H2AX staining, suggesting that more immature cells were the ones most affected (Figure 8A). Microscopic quantification revealed that the number of γ-H2AX-positive spermatogonia, as well as the number of γ-H2AX foci per cell, were increased ∼2-fold when Arf was inactivated (Figure 9), but the greatest increase in γ-H2AX staining was observed in primary Arf-null spermatocytes (Figure 8A and 8B; see below). The increased γ-H2AX in meiotic cells is especially striking because these cells do not normally express p19Arf when the gene is present (Figure 1).
Meiotic double-strand DNA breaks are induced by the Spo11 transesterase and its accessory factors, which are loaded onto chromatin during the final pre-meiotic S-phase [13]. Autosomal γ-H2AX staining is normally observed during the leptotene and zygotene phases of meiosis-I, which are the early stages at which chromatids undergo DNA scission as a prelude to homologous recombination. In contrast, γ-H2AX foci are not normally detected by early pachytene (except in the sex body) once homologous synapsis is complete (Figure 8C, top panels) [28]. Chromosome spreads from meiotic primary spermatocytes from Arf-null males revealed that 150 of 382 individually enumerated pachytene cells (39%) exhibited persistent autosomal γ-H2AX foci in addition to normal sex body staining, whereas very few such cells (6.7%) were detected at the diplotene stage (Figure 8C, bottom panels). It could be that disappearance of γ-H2AX is delayed until diplonema, or that cells with aberrantly elevated γ-H2AX are preferentially eliminated. The latter interpretation is supported by the prophase I apoptosis and depletion of diplotene cells observed in Arf-null mice (Figure 6). Therefore, in the absence of Arf, γ-H2AX accumulates to higher levels than normal starting in the least mature spermatogonia, continuing into meiotic prophase I, and persisting past the time when it would normally disappear from autosomes. Although inactivation of p53 suppresses the increased apoptosis of Arf-null spermatocytes, γ-H2AX persists in cells lacking both of these genes (Figure 8A, lower right panel). In meiotic chromosome spreads from p53; Arf double-null mice, 20.5% (34 of 166) of diplotene spermatocytes display persistant autosomal γ-H2AX immunostaining as compared to 6.7% (5 of 74) of singly Arf-null and less than 1% (1 of 149) of wild-type diplotene spermatocytes. These data underscore the fact that the accumulation of γ-H2AX in Arf-null spermatocytes is p53-independent, whereas the elimination of defective spermatocytes that retain γ-H2AX inappropriately is p53-dependent.
DNA double-strand breaks induced early in prophase I by Spo11 serve as substrates for the strand exchange proteins Rad51 and meiosis-specific Dmc1, which are required for double strand break repair during homologous recombination. Foci of staining using antibodies to Dmc1 (Figure 10A) and Rad51 (Figure 10B) were readily observed in zygotene spermatocytes from wild-type mice (left panels) but were fewer and less prominent in their Arf-null counterparts (right panels). The number and average fluorescence intensities of foci in 100 zygotene cells of each genotype were determined using commercial imaging software. In wild-type zygotene spermatocytes, the frequency of Dmc1 and Rad51 foci peaked at 100–125 per cell (Figure 10C and 10D, respectively, blue bars) and exhibited a broad distribution of relative intensities over a ∼10-fold range (Figure 10E and 10F, blue bars). In contrast, both the number and intensities of Dmc1/Rad51 foci were significantly reduced in Arf-null cells (average number of foci ± S.D.: 103±46 Dmc1 foci in wild-type vs. 44±34 in Arf−/−; 108±45 Rad51 foci in wild-type vs. 45±33 in Arf−/−; average relative intensities ± S.D.: 10239±4959 Dmc1 foci in wild-type vs. 6611±3005 in Arf−/−; 10980±5783 Rad51 foci in wild-type vs. 5637±3340 in Arf−/− N = 100, p<0.0001, Student's t-test; Figure 10C–10F). An accumulation of Arf-null spermatocytes in zygonema (Figure 6F) suggests that there may be a delay at this stage before progression to pachytene.
Pachytene cells are normally characterized by well developed synaptonemal complexes that stretch the length of autosome axes and by knob-like accumulation of SYCP3 at telomeres (Figure 11A). However, 34% of Arf-null cells exhibited defects in synapsis (quantified in Figure 11D), including forked terminal structures and interstitial bubbles on autosomes (Figure 11B) and complete asynapsis of sex chromosomes (arrow, Figure 11C). In addition, interrupted regions of SYCP3 staining (denoted by arrowheads in Figure 11C) were more frequently observed in meiotic chromosome spreads from Arf-null cells versus those in wild-type cells (191 versus 53 such segments, respectively, in 300 pachytene cells of each genotype). Because synapsis was complete in the majority of Arf-null pachytene cells, we could not distinguish whether the observed defects arose from regions in which synaptonemal complexes did not form at all, or where complexes had formed but subsequently disassembled. Taken together, Arf-deficiency results in a series of abnormalities during prophase I that include reduced loading of the Rad51 and Dmc1 recombinases, defects in synapsis, elevated and persistent γ-H2AX expression, and p53-dependent apoptosis, ultimately associated with diminished production of mature sperm.
With few exceptions, the Arf tumor suppressor is not expressed in normal tissues of healthy mice but is induced by abnormally sustained and elevated thresholds of proliferative signals, activating a p53 response that opposes the deleterious effects of oncogene activation. Notably, p53 responds to a much wider range of Arf-independent signal transduction cascades triggered by many other forms of cellular stress, including acute DNA damage, to which the Arf promoter does not respond [6]. By converging on p53, these different signaling pathways inhibit cell cycle progression or trigger apoptosis, acting to suppress tumor formation.
We now document a physiological role of Arf in mouse male germ cell development that is distinct from its tumor suppressive functions in key respects. First, Arf is expressed in spermatogonia, but not in the primary spermatocytes that arise from them. Expression of p19Arf neither arrests spermatogonial mitotic progression nor triggers their p53-dependent apoptosis. However, the absence of Arf expression in spermatogonia leads to p53-dependent apoptosis of spermatocytes before they exit meiosis-I. The defect in spermatogenesis is germ cell autonomous and results in a significant reduction in sperm counts by the time Arf-null mice are two months old, although residual sperm production maintains fertility in young males. Thus, expression of Arf in mitotic progenitor cells enhances the survival of their meiotic progeny in which Arf expression is normally extinguished. These features indicate that Arf expression initiates a salutary, feed-forward program that facilitates meiotic progression. Indeed, although Arf and Ink4a are widely viewed to convey tumor suppressive functions that coordinate the activities of the p53 and Rb signaling “pathways,” inactivation of Arf and Ink4a in the testes leads to opposing outcomes. Disruption of Ink4a increases the mitotic activity of spermatogonial progenitors to enhance sperm output and, in this respect, compensates for Arf loss of function without eliminating the cellular defects that arise in the Arf-null setting. In short, loss of Ink4a increases the spermatogonial pool size, but without Arf expression, spermatocytes undergo increased apoptosis, returning the number of mature sperm to normal levels.
Homologous recombination during meiosis exchanges genetic information between maternally and paternally derived chromosomes and also guides proper segregation of chromosome pairs to maintain correct chromosome numbers in gametes [13]. During meiosis, in contrast to mitotically diving cells, homologous chromosomes are favored over sister chromatids as templates for recombinational DNA repair. Double-strand DNA breaks are formed by the topoisomerase-II-related transesterase Spo11. This process activates the Atm kinase and leads to phosphorylation of the H2AX histone variant near sites of strand breakage during early prophase I. Binding of the RecA family strand exchange proteins, Rad51 and meiosis-specific Dmc1, to Spo11-induced DNA ends generates filaments that search for and invade homologous duplex DNA molecules, leading to pairing of homologous chromosomes. Loading of Rad51 and Dmc1 is normally reversed by early pachytene when chromosomes are fully synapsed, after which γ-H2AX foci are no longer detected.
In the Arf-null setting, a modest but significant increase in γ-H2AX staining was first detected in the least mature spermatogonia, and primary spermatocytes displayed accentuated signals that persisted inappropriately into the pachytene stage. Arf-null cells also formed fewer Dmc1/Rad51 foci at zygotene and exhibited focal regions of asynapsis at pachytene. Aberrant Arf-null spermatocytes underwent apoptosis at pachytene, resulting in the emergence of fewer diplotene cells and a significant reduction in sperm output. Importantly, Arf−/−; p53−/− double-null pachytene cells also exhibited persistent γ-H2AX staining, but these cells escaped elimination. Thus, apoptosis was p53-dependent, but aberrant γ-H2AX accumulation was not.
Although the underlying mechanisms remain unknown, we consider here two plausible interpretations of this apoptotic arrest. First, it may be that reduced Rad51/Dmc1 focus formation and persistent γ-H2AX staining in Arf-null male germ cells connote a defect in DNA repair that then activates p53 through Arf-independent but Atm/Atr-dependent signaling pathways. In this scenario, Spo11-induced DSBs would form at normal levels but Rad51/Dmc1 loading would be impaired such that some DNA damage would persist into pachytene. This might conceivably involve the p53-independent ability of p19Arf to promote the sumoylation of numerous target proteins by inhibiting the SUMO2/3 protease Senp3 [29]–[31]. SUMO2/3 accumulates at sites of DNA damage in mammalian cells [32], [33], and various aspects of DNA repair are regulated by the SUMO conjugation pathway [34]. There is fragmentary evidence that absence of p19Arf compromises nucleotide excision repair in cultured cells [35], [36] raising the possibility that Arf may play an as yet undefined role in promoting homologous recombination. All meiotic mutants that cannot properly synapse homologous chromosomes arrest during pachytene [37], and accompanying defects in sex body formation and failure to properly silence transcription of the sex chromosomes during prophase is itself sufficient to eliminate pachytene cells [27], [38]. However, spermatocytes can also undergo apoptosis in direct response to unrepaired Spo11-induced breaks even if sex body formation is normal [27], [39]. Where tested, spermatocyte apoptosis in meiotic mutants with chromosome synapsis errors has been found to be p53-independent [40]–[42]. Moreover, Spo11-dependent activated phospho-p53 can be transiently detected from leptonema and zygonema in wild-type male mice, and in Drosophila, p53 activity is prolonged in cells defective for meiotic repair [43]. Thus, it remains a formal possibility that meiotic recombination defects can trigger p53-dependent apoptosis.
A second, alternative interpretation rests on the idea that the earlier and less profound accumulation of γ-H2AX in Arf-null spermatogonia might be a symptom of an underlying defect affecting chromatin structure or Atm/Atr signaling. The appearance of γ-H2AX reflects chromatin modifications that flank sites of DNA damage rather than strand breaks themselves, so the kinetics of γ-H2AX formation and dissolution do not necessarily coincide with the appearance and repair of DNA damage [44], [45]. Moreover, aberrant Atm/Atr signaling is itself sufficient to activate p53, whether triggered by DNA breaks or not [46]. Thus, it may be that Arf deficiency causes inappropriate Atm/Atr signaling that provokes p53-dependent apoptosis in a DNA damage-independent manner. In this view, the observed meiotic prophase defects in Arf-null spermatocytes may possibly be a separate downstream consequence of this earlier anomaly, and may not be the cause of apoptosis. Regardless of which interpretation is correct, it is important to note that our findings provide strong evidence that p53-dependent monitoring promotes proper meiotic maturation, in addition to the previously documented p53-independent pathway(s). Whatever the underlying mechanisms, the role of Arf in male germ cell development contrasts with the general paradigm of p19Arf acting as an activator of p53. Instead, it is the absence of Arf in spermatogonia that consequently leads to p53-dependent apoptosis of spermatocytes.
No human or non-human primates were studied. All animal work with mice was performed under established guidelines and supervision by the St. Jude Children's Research Hospital's Institutional Animal Care and Use Committee (IACUC), as required by the United States Animal Welfare Act and NIH policy to ensure proper care and use of laboratory animals for research. Experiments were undertaken in an accredited facility of the Association for Assessment of Laboratory Animal Care under the supervision of trained veterinary personnel and in strict compliance with Howard Hughes Medical Institute, St. Jude Children's Research Hospital, and NIH institutional guidelines. The latter include detailed protocol submission and review of all animal care, monitoring, and experimental procedures prior to initiation of any experiments. Ongoing protocols for animal research not necessitating interim amendments are minimally subjected to annual review by the IACUC. All persons involved in the use of animals have read and understand all implications of pertinent protocols, have received training in the execution of relevant animal-related procedures prior to participation in the protocol, and have participated in educational or training programs deemed necessary by the IACUC or the Animal Resources Center personnel. Studies reported herein did not unnecessarily duplicate previous research, and were undertaken only because suitable non-animal models were unavailable. The number of animals used was consistent with good statistical design. Anesthesia, analgesia and tranquilization were used to relieve pain and distress in accordance with the IACUC recommendations.
Arf-null [47], Arf-GFP [7], Arf-Flox and Arf-Cre mice [9] were generated in the Sherr laboratory. Mouse strains deficient for Ink4a [48] and Ink4a-Arf [49] were generously provided by R.A. DePinho (Dana Farber Cancer Center). All genetically engineered mice were backcrossed nine or more times onto a C57Bl/6 background to create isogenic strains. C57Bl/6 mice deficient for p53 were purchased from Jackson Laboratories (Stock Number 2101). ArfGFP/GFP mice were crossed to p53+/− mice, and compound heterozygotes were interbred to generate ArfGFP/GFP;p53−/− mice functionally null for both genes. ArfCre/+ females were interbred with ArfFlox/Flox males to generate ArfCre/Flox mice.
Caudal epididymides were harvested before dissection of the testes. For each male mouse, two cauda were minced into 1 ml of Dulbecco's modified Eagle's medium (DMEM) containing 25 mM HEPES buffer (pH 7.5) and 4 mg/ml bovine serum albumin and incubated at 37°C for 20 minutes. Suspensions of sperm were fixed at a 1∶25 dilution in 10% formalin and counted on a hemocytometer. All sperm counts were performed between 1:00–3:00 PM. Dissected testes were weighed in pairs.
Mice were euthanized by CO2 asphyxiation, and testes were removed and fixed overnight at 4°C in 4% paraformaldehyde followed by saturation in 30% sucrose at 4°C overnight. Tissues were embedded in TBS Tissue Freezing Medium (Fisher Scientific, Pittsburg PA), and sliced with a HM500M Cryostat (Microm International, Walldorf, Germany) into 10 µm sections. Fixed and frozen samples were sectioned and subjected to antigen retrieval in 0.1 M Na citrate buffer, pH 6.0, followed by one hour incubation at room temperature in a blocking solution of 10% normal goat serum (NGS), 0.1% Triton-X 100 in phosphate-buffered saline (PBS), and then by overnight incubation at 4°C in primary antibodies diluted in 3% NGS, 0.1% Triton-X 100 in PBS. Antibodies were directed to p19Arf [rat monoclonal immunoglobulin 5C3-1 [50], Sox9 (Millipore AB5535, 1∶1000), BrdU (Santa Cruz sc32323, 1∶100), cyclin D1 (Santa Cruz 72-13G, 1∶750), Dmc1 (Santa Cruz H-100, 1∶750), γ-H2AX (Cell Signaling 2577, 1∶200), and SUMO2/3 (Cell Signaling 18H8, 1∶300). Slides were washed three times in PBS, and then incubated for 1 hour at room temperature in 3% NGS, 0.1% Triton-X 100 in PBS containing the relevant secondary antibodies conjugated to Ig-Alexa Fluor 555 or Ig-Alexa Fluor 488 (1∶500 dilutions; Invitrogen). Slides were washed three times in PBS and mounted with Vectashield (Vector Labs) containing 4′-6-diamidino-2-phenylindol (DAPI). TUNEL assays were performed using an in situ cell death detection kit (TMR red, Roche) following the manufacturer's protocol. Images of tissue sections were photographed using a Zeiss Axioscope fluorescence microscope and assembled using Zeiss Axiovision software.
Testes were decapsulated and minced in 5 ml of DMEM per testis and transferred to a 15 ml Falcon tube. After further dissociation of the tubules by pipeting up and down, large pieces were allowed to settle to the bottom of the tube by gravity for 10 minutes on ice. One ml of the supernatant, containing a suspension of spermatocytes, was transferred to a 1.5 ml Eppendorf tube and centrifuged for five minutes at 5800× g. The pellet was resuspended in 40 µl of a 0.1 M sucrose solution, and 20 µl of spermatocyte suspension was applied evenly to a slide containing a thin layer of 1% paraformaldehyde (pH 9.2) containing 0.1% Triton X-100. Slides were allowed to dry for two hours at room temperature in a closed humidity chamber before rinsing in Photo-flo (Kodak 1464510, diluted 1∶250 in doubly distilled H2O) and air dried at room temperature. For immunofluorescence, slides were incubated in PTBG (0.2% bovine serum albumin, 0.2% gelatin, 0.05% Tween 20 in PBS) for 10 minutes with shaking. Primary antibodies were diluted in PTBG, applied to the slide, and covered with parafilm before incubation overnight at 4°C in a humidity chamber. Antibodies were directed to SYCP3 (Santa Cruz G-3, 1∶500) to mark the synaptonemal axial element [20], to γ-H2AX (Cell Signaling 2577, 1∶500) to identify sex body formation and sites of DNA damage, and to Rad51 (Calbiochem Ab-1, 1∶500) and Dmc1 (Santa Cruz H-100, 1∶750) to demonstrate formation of complexes required for DNA strand exchange during homologous recombination. Slides were washed three times in PTBG at room temperature for 3 minutes with shaking. Secondary antibodies, also diluted in PTBG, were applied to slides which were covered with parafilm and incubated at 37°C for one hour in a humidity chamber. Slides were washed three times in PTBG for 3 minute intervals in the dark with shaking and mounted with Vectashield (Vector Labs) containing DAPI. Surface spread spermatocytes were visualized by a Marianas spinning-disc confocal microscope, and images were assembled and analyzed using Slidebook 5.0 SDC software (Intelligent Imaging Innovations, Denver CO). Meiotic spreads from three adult mice (age three months) were analyzed. One hundred spermatocytes were scored each from mouse.
Distinct staining patterns allow for classification of each stage of meiotic prophase [51], [52]. Leptotene cells were categorized by short stretches of axial elements accompanied by intense γ-H2AX staining throughout the nucleus and the absence of a distinct sex body. Zygotene cells also display intense γ-H2AX staining throughout the nucleus and lack a sex body, but can be distinguished by longer stretches of SYCP3 staining, some of which are synapsed. Pachytene cells have fully formed and synapsed axes that appear as thick, continuous SYCP3-stained threads, while displaying intense γ-H2AX staining only in the sex body. Dmc1 and Rad51 foci are normally present at leptotene and zygotene, and largely disappear by pachytene. Diplotene cells have γ-H2AX localized only to the sex body, but fully formed axes are desynapsing and chiasmata are visible.
As previously described [53], detergent lysates were prepared, and protein concentration was quantified by bicinchoninic acid assay (Pierce). Samples (25–75 µg protein per lane) were electrophoretically separated on 4% to 12% Bis-Tris NuPAGE gels (Invitrogen), transferred to polyvinylidene fluoride membranes (Millipore), and detected using antibodies to γ-H2AX (Cell Signaling S139, 1∶500), p19Arf (5C3-1; Bertwistle et al. 2004b), p53 (Cell Signaling 1C12, 1∶500), and actin (Santa Cruz C-11, 1∶500) to control for protein loading.
|
10.1371/journal.pntd.0003993 | Comparative Analysis of Transcriptional Profiles of Adult Schistosoma japonicum from Different Laboratory Animals and the Natural Host, Water Buffalo | Schistosomiasis is one of the most widely distributed parasitic diseases in the world. Schistosoma japonicum, a zoonotic parasite with a wide range of mammalian hosts, is one of the major pathogens of this disease. Although numerous studies on schistosomiasis japonica have been performed using laboratory animal models, systematic comparative analysis of whole-genome expression profiles in parasites from different laboratory animals and nature mammalian hosts is lacking to date.
Adult schistosomes were obtained from laboratory animals BALB/c mice, C57BL/6 mice, New Zealand white rabbits and the natural host, water buffaloes. The gene expression profiles of schistosomes from these animals were obtained and compared by genome-wide oligonucleotide microarray analysis. The results revealed that the gene expression profiles of schistosomes from different laboratory animals and buffaloes were highly consistent (r>0.98) genome-wide. Meanwhile, a total of 450 genes were identified to be differentially expressed in schistosomes which can be clustered into six groups. Pathway analysis revealed that these genes were mainly involved in multiple signal transduction pathways, amino acid, energy, nucleotide and lipid metabolism. We also identified a group of 1,540 abundantly and stably expressed gene products in adult worms, including a panel of 179 Schistosoma- or Platyhelminthes-specific genes that may be essential for parasitism and may be regarded as novel potential anti-parasite intervention targets for future research.
This study provides a comprehensive database of gene expression profiles of schistosomes derived from different laboratory animals and water buffaloes. An expanded number of genes potentially affecting the development of schistosomes in different animals were identified. These findings lay the foundation for schistosomiasis research in different laboratory animals and natural hosts at the transcriptional level and provide a valuable resource for screening anti-schistosomal intervention targets.
| The zoonotic parasite Schistosoma japonicum is one of the major pathogens of schistosomiasis and can parasitize a wide range of mammals. Although numerous schistosome transcriptional profiling studies have been performed using laboratory animal models, the differences in the global gene expression profiles of worms from different laboratory animals and natural mammalian hosts have not been characterized. Therefore, we studied the gene expression profiles of adult worms from BALB/c mice, C57BL/6 mice, rabbits and buffaloes using a transcriptomics approach. Our results indicate that, although the expression profiles of adult worms from different mammals are generally similar, hundreds of genes are differentially expressed, which were mainly involved in various signal transduction pathways, amino acid, energy, nucleotide and lipid metabolism. Numerous abundantly and stably expressed genes in adults were identified, including some genes that are only found in blood flukes or expanded within the phylum Platyhelminthes and may be important for parasitism. Our data provide a basis for schistosomiasis research in different mammalian hosts at the transcriptional level as well as a valuable resource for the screening of anti-schistosomal intervention targets.
| Schistosomiasis is one of the most serious parasitic diseases and affects more than 200 million people globally, based on conservative estimates [1,2]. Schistosomiasis is caused mainly by three Schistosoma species: Schistosoma japonicum, Schistosoma mansoni, and Schistosoma haematobium. S. japonicum is endemic in Asia, principally China and the Philippines, whereas S. mansoni and S. haematobium are distributed in Africa and the Middle East. S. mansoni is also prevalent in South America [3]. Schistosomes have a complex developmental life cycle and exhibit sexual dimorphism. The life cycle comprises seven morphologically discrete stages (i.e., egg, miracidium, mother sporocyst, daughter sporocyst, cercaria, juvenile schistosomulum, and adult worm), complicating the control and prevention of schistosomiasis [4]. In contrast to the other two Schistosoma species known to infect humans, S. japonicum is a true zoonotic parasite that utilizes a wide range of mammalians as definitive hosts, including bovines, mice, rabbits, goats, pigs, and dogs [5,6].
Previous studies have indicated that the susceptibility of different mammalian hosts to S. japonicum infection varies; water buffaloes, rats, horses, and pigs are less susceptible to infection than mice, rabbits, yellow cattle, and goats [5]. Schistosomes derived from different mammalian hosts also exhibit visible changes in morphology, such as the length and width of worms, tegument, sucker and gynoecophorus [7,8]. In the last decade, a number of gene expression profiling studies of schistosomes have been conducted using various analytical approaches, and the findings have tremendously facilitated improved understanding of the molecular basis of schistosome developmental biology, host-parasite interactions and the pathogenesis of schistosomiasis (see more in reviews [9–11]). The majority of schistosomes used in those studies were isolated from laboratory animals such as BALB/c mice, C57BL/6 mice, and New Zealand white rabbits.
Recently, numerous differentially expressed genes that may influence parasite survival and development were identified by comparative proteomic analyses [12] and microarray analysis [13] of schistosomula from susceptible BALB/c mice, less susceptible Wistar rats and resistant reed voles. Yang et al. comparatively analyzed the gene expression profiles of S. japonicum derived from natural reservoir host yellow cattle, goats and water buffaloes using microarrays, and suggested that the gene expression patterns of some genes in schistosomes in natural hosts and laboratory animals may be diverse [7,8]. Data obtained using laboratory animals may not be completely transferable to natural hosts. A comparison of the gene expression profile of S. japonicum in its natural reservoir hosts and laboratory animals will undoubtedly facilitate our understanding of parasite biology. Previous epidemiological studies have revealed that bovines, particularly water buffaloes, are the major natural reservoir of S. japonicum and play a vital role in schistosomiasis transmission in China [14–17]. Therefore, we performed comparative analyses of the gene expression profiles of S. japonicum from the natural reservoir host, water buffalo, with those from laboratory animal mice and rabbits using a genome-wide microarray approach. Our results will be of significance for the screening of anti-schistosome targets and vaccine candidates using laboratory animals to further facilitate the control of schistosomiasis in natural reservoir hosts in endemic areas.
All procedures performed on animals in this study were conducted following animal husbandry guidelines of the Chinese Academy of Medical Sciences and with permission from the Experimental Animal Committee with the Ethical Clearance Number IPB-2011-6.
S. japonicum-infected Oncomelania hupensis were provided by the Hunan Institute of Parasitic Diseases, Yueyang, China. Laboratory BALB/c mice (nine mice), C57BL/6 mice (nine mice), New Zealand white rabbits (three rabbits) and natural reservoir host water buffaloes (three buffaloes) were infected with 40–400 freshly released cercariae through the upper back using the cover glass method [18]. Adult worms were perfused out the hepatic portal vein of infected animals at approximately 7 weeks, washed briefly with PBS and soaked immediately in RNAlater Solution (Ambion, CA, USA) and stored at -80°C until total RNA was extracted.
Total RNA was extracted from worms from different animals using the RNeasy Mini kit (QIAGEN), and contaminating genomic DNA was removed from RNA samples using a DNA-free kit (Ambion, CA, USA). The quantity and quality of the RNA samples were assessed using a ND-1000 spectrophotometer (NanoDrop Technologies, Wilmington, DE) and denaturing agarose gel electrophoresis.
Schistosome genome-wide microarrays were used to analyze the gene expression profiles of S. japonicum derived from different laboratory animals and the natural host, water buffalo. The design and construction of the microarray and the methods used in microarray hybridization and feature extraction have been previously reported [19–21]. The full details of this schistosome microarray design have been deposited in the Gene Expression Omnibus public database (http://www.ncbi.nlm.nih.gov/geo) under platform accession number GPL18617. Briefly, a total of 21,861 target sequences (20,194 sequences derived from S. japonicum and 1667 sequences derived from S. mansoni) were provided to Roche NimbleGen for array design. For each sequence, 3 or 4 60-mer oligonucleotide probes were designed. Probes with random sequences were printed as negative control (background signal) and eight spike-RNA probes from the intergenic sequence of yeast as hybridization controls. The microarray was manufactured by Roche NimbleGen. Microarrays were printed in a 12×135 K feature format. cDNA was labeled with a fluorescent dye (Cy3-dCTP) using cRNA Amplification and Labeling Kit (CapitalBio, Beijing, China). Hybridization was performed using three biological replicates for all samples at CapitalBio. Procedures of array hybridization, washing, scanning, and data acquisition were carried out according to the NimbleGen Arrays user’s guide. The arrays were scanned using MS200 scanner (NimbleGen Systems) with 2 μm resolution, and NimbleScan software (NimbleGen) was used to extract fluorescent intensity raw data from the scanned images. The normalized expression data of genes was generated using the Robust Multichip Average (RMA) algorithm [22–24], which consisted of three steps: a background adjustment, quantile normalization and final summarization. The outlier probes were identified and the contribution of outlier probes was reduced in the reported gene expression level, which has been demonstrated to improve the sensitivity and reproducibility of microarray results. Then, the expression value of a gene is a weighted average of all probes when both background correction and quantile normalization were performed. Raw data and normalized gene level data from the arrays have been deposited at the public database Gene Expression Omnibus under the accession number GSE65327. Genes were considered differentially expressed by expression fold-change (FC) ≥2 between any two compared worm samples and p<0.05 (Student’s t-test). The stably and abundantly expressed genes among the adult worms from the four mammalian hosts were extracted using coefficient of variation (CV) ≤0.1 and fluorescence intensity ≥10,000 on the basis of retrieval of a panel of well characterized Schistosoma genes of which the intensity values were greater than 10,000. Hierarchical clustering analysis of selected genes was performed to generate heat maps using the software Cluster 3.0 [25] and Heatmap Builder 1.0 [26].
A subset of genes with different expression patterns were selected for further validation using real-time PCR as previously described [19]. Real-time PCR primers were designed using Primer Express 3.0 software (Applied Biosystems, Foster City, USA) (S1 Table). For each sample, 1 μg of total RNA was used to synthesize the first strand cDNA using a Reverse Transcriptase Kit (Invitrogen, CA, USA) with oligo (dT) 12–18 primers (Invitrogen) in a final volume of 20 μl. PCR reactions were performed in technological triplicate on a 7300 Real-Time PCR system (Applied Biosystems) using SYBR Green QPCR Master Mix (Agilent Technologies, USA) according to the manufacturer’s instructions. The 26S proteasome non-ATPase regulatory subunit 4 gene (PSMD4, GenBank Accession No. FN320595), which has been validated as a reliable reference gene in transcriptomic analysis of S. japonicum, was employed as a reference gene in the real-time PCR analysis [19]. The relative expression level of each gene was analyzed using SDS 1.4 software (Applied Biosystems).
Gene sequences were functionally annotated using Blast2GO [27], and the output provided combined graphics for three categories of gene ontology (GO) terms: biological processes, molecular functions, and cellular components. The Kyoto Encyclopedia of Genes and Genomes (KEGG) automated annotation server was used to assign pathway-based functional orthology to differentially expressed genes [28]. Signal peptides were predicted using the program SignalP 4.1 server employing both the neural network and hidden Markov model [29], and transmembrane helices were predicted using TMHMM 2.0 [30].
SjVAL genes were identified using the BLASTp and InterProScan [31] algorithms. Initially, the sequences of S. mansoni venom-allergen-like proteins (SmVALs) [32] were used for BLASTp searches of S. japonicum predicted protein sequence database [33] and the non-redundant protein sequence database [34,35] of the National Center for Biotechnology Information (NCBI) (e-value cut-off: 10−5). CD-HIT v4.5.4 software (http://www.bioinformatics.org/cd-hit/) was then used to eliminate redundant protein sequences from the obtained putative SjVALs using the criteria of 95% identity and 80% coverage between two sequences. Finally, all remaining protein sequences were confirmed by the presence of SCP/TAPS-representative protein domains (IPR014044) using InterProScan [31]. A phylogenetic relationship tree was built using the full-length amino acid sequences of Schistosoma VALs in the following steps. First, sequences were aligned using ClustalX [36] and then refined manually, and a phylogenetic tree was finally generated using MEGA 5.0 software [37] by the neighbor-joining (NJ) method (the bootstrap test was performed with 1000 replicates).
Accession numbers for the S. japonicum sequences used in the alignment are: CAX74321.1, CAX74107.1, CAX73316.1, CAX78430.1, CAX72962.1, CAX76177.1, CAX73488.1, CAX78439.1, CAX78429.1, CAX78435.1, AAP06001.1, AAW25717.1, AAW25499.1, AAW25247.1, AAW25007.1, AAW27353.1, Sjp_0038950, Sjp_0038960, Sjp_0083700 and Sjp_0112690. Accession numbers for the S. mansoni sequences used in the alignment are: AAY43180.1 (SmVAL1), AAY43181.1 (SmVAL2), AAZ04923.2 (SmVAL3), AAY43182.1 (SmVAL4), ABB88846.2 (SmVAL5), CCD74794.1 (SmVAL6), AAZ04924.1 (SmVAL7), ABW98681.1 (SmVAL8), ABB88845.1 (SmVAL9), ABO09814.2 (SmVAL10), ABA54555.1 (SmVAL11), ABB88844.1 (SmVAL12), ABB88843.1 (SmVAL13), ABO09815.1 (SmVAL14), CCD80670.1 (SmVAL15), CCD74792.1 (SmVAL16), CCD74934.1 (SmVAL17), CCD80318.1 (SmVAL18), CCD80317.1 (SmVAL19), CCD80812.1 (SmVAL20), CCD80564.1 (SmVAL21), CCD59744.1 (SmVAL22), CCD80666.1 (SmVAL24), CCD80667.1 (SmVAL25), CCD80638.1 (SmVAL26), CCD80648.1 (SmVAL27) and CCD80636.1 (SmVAL28).
This study compared the transcriptional profiles of adult worms isolated from experimental animals (BALB/c mice, C57BL/6 mice and New Zealand white rabbits) and the natural host (water buffaloes) using genome-wide microarray analyses and combinatorial bioinformatics. First, we assessed a common correlation of all the genes within arrays between biological replicates of schistosomes from the same host to assess the quality of the biological replicates. As expected, the correlation coefficient (r) between biological replicates was >0.99, demonstrating a strong consistency between the biological replicates of worms from the same host (Table 1). Similar results were obtained by correlation analyses of biological replicates for the global expression profiles of schistosomes from water buffaloes, yellow cattle and goats [7,8]. We then investigated the transcriptional profile correlation between worm samples from different hosts using an average of three biological replicates. The results demonstrated that the transcriptional profiles between worms from different hosts were also highly similar as a whole (r>0.98, Table 2).
A series of pairwise comparisons between schistosomes isolated from the four mammalian hosts was performed to identify differentially expressed genes. Using FC≥2 and p<0.05, subsets of differentially expressed gene products were identified: 180 (buffalo vs. C57BL/6), 116 (buffalo vs. BALB/c), 97 (buffalo vs. rabbit), 263 (rabbit vs. C57BL/6), 172 (rabbit vs. BALB/c) and 67 (C57BL/6 vs. BALB/c). The distributions of the up- and down-regulated genes between the paired comparisons are displayed as scatter plots (Fig 1 and S2 Table). It is understandable that adult worms from C57BL/6 and BALB/c mice shared more similar transcriptional profiles than those from rabbits and buffaloes due to the evolutionary distance between the two animal species. Previous studies have shown that fewer differentially expressed genes were identified in schistosomes between water buffalo and yellow cattle (69 genes) than between water buffalo and goat (485 genes) [7,8].
After integrating these subsets of differentially expressed genes, a total of 450 non-redundant genes were used for hierarchical clustering to enable collective visualization by gene and array (See supplementary S3 Table for details). As expected, these genes were clustered into four subgroups, of which genes of the biological replicates were clustered together, and branched by the host origination of the samples (Fig 2). The result further demonstrated high consistency between the biological replicates. The heat map demonstrated that these differentially expressed genes clustered into two major transcriptional patterns: cluster I, genes that were up-regulated in schistosomes isolated from C57BL/6 mice and BALB/c mice and down-regulated in schistosomes isolated from buffaloes and rabbits; and cluster II, genes that were down-regulated in schistosomes isolated from C57BL/6 mice and BALB/c mice and up-regulated in schistosomes isolated from buffaloes and rabbits (Fig 2). Meanwhile, several other clusters of differentially expressed genes were also identified by hierarchical clustering: cluster III, genes highly expressed in the worms from rabbits; cluster IV, genes up-regulated in worms from buffaloes; cluster V, genes up-regulated in worms from C57BL/6 mice; and cluster VI, genes highly expressed in worms from BALB/c mice (Fig 2). A list of representative differentially expressed genes from each cluster is displayed in detail as a heat map (Fig 3).
We also compared our results with previous studies on gene expression profiles of S. japonicum derived from the natural mammalian hosts, buffaloes, cattle and goats [7,8]. Some of these differentially expressed genes in adult schistosomes from the different hosts have been identified in previous studies, such as CNUS0000098059, CNUS0000105021, CNUS0000096235, CNUS0000102644, FN318955 and FN313838. Meanwhile, a number of differentially expressed genes identified by previous studies were not identified in our study since the fluorescence intensity values of these genes were below the cut-off value (S4 Table).
To validate the microarray transcription data, a subset of ten genes with various biological functions and different expression patterns was selected for real-time PCR validation. The real-time PCR results well matched the microarray data (Fig 4) with a significant correlation factor of 0.943 (Spearman’s Rho, p<0.0001, n = 40), thereby validating the microarray results.
GO analysis was performed to summarize and explore the major GO categories of the differentially expressed genes which may be susceptible to host environments. A total of 174 gene sequences were annotated with GO terms in three independent categories: biological processes (161 gene sequences), molecular functions (170 gene sequences), and cellular components (50 gene sequences) (Fig 5 and S3 Table). The biological processes analysis revealed that the predominant genes were involved in response to metabolic processes, including primary metabolic process, organic substance metabolic process, cellular metabolic process and nitrogen compound metabolic process. For molecular functions, the majority of genes were annotated with binding activity such as organic cyclic compound binding activities, heterocyclic compound binding activity, and ion binding activity. Notably, genes coding for proteins associated with membrane accounted for the major portion of annotated genes in the cellular components analysis.
Comparative genomics analysis of well characterized signaling pathways between schistosome and mammalian host indicates that schistosome genome encodes various growth factors, receptors and other critical components to regulate numerous cellular processes during tissue development and organogenesis. These components share high sequence similarity with mammalian orthologues, implying that schistosomes may utilize host components as developmental signals besides their own pathways [33]. The KEGG pathway analysis showed that many of these differentially expressed genes were involved in signaling transduction pathways, such as calcium signaling pathway, cAMP signaling pathway, PI3K-Akt signaling pathway, Rap1 signaling pathway, Ras signaling pathway, ErbB signaling pathway and MAPK signaling pathway (Table 3). Previous studies has also found that genes involved in ErbB signaling pathway and MAPK signaling pathway were differentially expressed in worms from buffaloes and goats [8]. Moreover, the S. japonicum genome reveals that schistosomes are not able to de novo synthesize fatty acids, sterols, purines, essential amino acids, which must be acquired from their hosts. Analysis of the KEGG pathways assigned to metabolic process indicated that these differentially expressed genes mainly participated in amino acid, energy, nucleotide and lipid metabolism (Table 3), which is consistent with the findings of Yang et al. [7]. All the results above further prove that schistosomes can exploit host nutrients and signaling pathways for growth and development, and host environments can affect the survival and development of the parasites. For instance, human tissue factor pathway inhibitor (TFPI), which acts as a plasma Kunitz-type serine protease inhibitor, is an anti-coagulation protein that plays an important role in the regulation of the blood coagulation cascade [38]. Interestingly, the S. japonicum TFPI gene (CAX69506.1) was significantly up-regulated in the intra-mammalian stage of the parasite life cycle and variously expressed in schistosomes from different hosts (S1 Fig). It will be engrossing to investigate the function of the TFPI gene in schistosomes in connection with the blood parasitic environment. Notably, one trematode eggshell synthesis protein gene (TES, pfam08034), AAW25913.1, was significantly up-regulated in worms from C57BL/6 mice, buffaloes and rabbits comparing with those from BALB/c mice. The trematode eggshell synthesis protein genes have been identified in several trematode parasites, which are crucial for eggshell synthesis, a key step for determining the quality and quantity of eggs laid [39]. Annexins are a multigene family of calcium-dependent phospholipid-binding proteins, and members of this family have been identified in major eukaryotic phyla [40]. In humans, annexins interact with various cell-membrane components by forming networks on the membrane surface that are involved in the regulation of membrane organization, cell differentiation and migration, intracellular signaling by enzyme modulation and ion fluxes [40–42]. Although annexins lack signal peptides for secretion, some extracellular members have been identified that act as receptors for serum proteases on the endothelium as well as inhibitors of neutrophil migration and blood coagulation [40]. In addition, some human annexin isoforms are involved in immunoregulatory functions such as the resolution of inflammation [43]. In S. mansoni, three annexin genes (Smp_074140, Smp_074150 and Smp_077720) have been identified by proteomic analysis as the membrane-associated constituents of the tegument surface [44]. Of the annexins, Smp_077720 is significantly up-regulated in the transition from free-living cercaria to parasitic schistosomulum and adult worm and binds to the tegument surface membranes in a calcium-dependent manner [45]. We identified three annexin domain-containing protein genes (AAX26603.2, CAX70178.1 and AAP06415.1) that were differentially expressed in schistosomes from the four different mammalian hosts. AAX26603.2 and CAX70178.1 were significantly up-regulated in worms from BALB/c and C57BL/6 mice compared with worms from buffaloes and rabbits, but the opposite expression pattern was observed for AAP06415.1 (Fig 3). This result indicates that these genes are either under different regulatory mechanisms or that the encoded proteins are functionally different. In addition, a gene coding for a putative endoribonuclease (AAX27316.2) containing the conserved endoribonuclease XendoU domain was significantly up-regulated in schistosomes from mice compared with those from buffaloes and rabbits. XendoU, which was first identified in Xenopus laevis, is a U-specific, metal ion-dependent endoribonuclease and is involved in the processing of intron-encoded small nucleolar RNAs (snoRNA) [46–48]. Schwarz et al. recently determined that the calcium-dependent endoribonuclease XendoU promotes endoplasmic reticulum network formation through local RNA degradation [49].
Notably, we observed that many of these differentially expressed genes were annotated as hypothetical proteins. Sequence alignment analysis revealed that these genes shared no sequence similarity to any sequence present in the non-redundant protein sequence (nr) databases of NCBI at a cut-off E value of 10−5 (except for Schistosoma species). Thus these Schistosoma-specific genes are likely specialised for parasitism by schistosomes, although their functions are unknown. Here, we identified a panel of four Schistosoma-specific genes (AAW25097.1, AAW27175.1, AAW24713.1 and CAX69761.1) that encoded a putative signal peptide and were overexpressed in worms from mice compared with those from buffaloes and rabbits. These four genes also exhibited a similar expression pattern among different developmental stages: significant down-regulation in the transition from free-living cercaria to parasitic schistosomulum and adult worm, and no expression in the egg stage (worms isolated from rabbits) (Fig 6). The intriguing expression patterns of these Schistosoma-specific genes spark interests in further characterization of their function and their potential contribution to successful parasitism.
Using microarray fluorescent intensity ≥10,000 as the cut-off value, 1,540 gene products that exhibited the lowest coefficient of variation (0.1) in expression among the worms isolated from different hosts were obtained (S5 Table). A list of selected genes with various functions is presented in Fig 7. The gene annotation results revealed that many of these stably and abundantly expressed genes are conventional housekeeping genes, some of which have been proved to be constitutively expressed across the schistosome lifecycle, such as eukaryotic translation factors, ribosomal proteins, histone proteins, tubulins, proteasome subunits, and NADH dehydrogenase subunits [19, 50]. The genes encoding S. japonicum 26S proteasome non-ATPase regulatory subunit 4 (PSMD4) and NADH dehydrogenase (ubiquinone) flavoprotein 2 (NDUFV2) have been extensively used as references for real-time PCR analysis [19,20,51,52]. Comparative genomics analysis indicated that a substantial proportion of these schistosome genes share various sequence identity with their homologous counterparts in H. sapiens. Importantly, crystal structures of some of the human proteins are available, providing a foundation for the future screening of compounds that specifically target schistosome proteins based on structural disparity (S5 Table). Some of the genes identified in the present study have been previously characterized as important functional genes for schistosome and potential anti-schistosome targets, although some of these genes were developmentally regulated during the lifecycle. For example, some genes encoding schistosome proteases (legumain, cathepsin B, C, D and L) that were up-regulated from cercariae to adult worms play key roles in obtaining nutrients from the host [20,53]. The schistosome tegumental aquaporin gene, which is important for parasite osmotic regulation, is most highly expressed during the intravascular life stages [54], and the schistosome thioredoxin glutathione reductase gene has been validated as a potential drug target for schistosomiasis chemotherapy [55,56].
The schistosome tegument is a dynamic host-interactive surface that is involved in nutrition, immune evasion/modulation, excretion, osmoregulation, sensory reception, and signal transduction [54,57]. The most promising schistosome vaccine candidates are proteins located on the surface of the worms, such as the tegument proteins TSP-2 (tetraspanin protein) and Sm29 [58–60]. In S. mansoni, TSP-2, which plays a prominent part in the parasite tegument development and maturation, seems to be an effective vaccine antigen against the blood fluke [61,62]. A set of 107 genes with putative signal peptides and 276 genes with putative transmembrane helices were identified by sequence interrogation with SignalP 4.1 and TMHMM 2.0. These proteins may be secreted or surface-exposed and thus capable of interacting with the external environment. More importantly, we determined that a significant number of these stably expressed genes in adults were peculiar to Schistosoma species or the phylum Platyhelminthes and have no identity to genes in other organisms. Encouragingly, one of these Schistosoma- or phylum Platyhelminthes-specific genes (SjSP-13) has been identified as a novel biomarker for immunodiagnosis of S. japonicum infection exhibiting extremely high sensitivity (90.4%) and specificity (98.9%), with nearly no cross-reactivity with other fluke infections [63]. Therefore, Schistosoma- or phylum Platyhelminthes-specific genes, particularly secreted and transmembrane proteins, may be regarded as novel potential anti-parasite drug targets or vaccine candidates for future studies.
Structurally related members of the sperm-coating protein/Tpx-1/Ag5/PR-1/Sc7 (SCP/TAPS; Pfam: PF00188) family have been characterized in a wide range of eukaryotes, including parasites [64]. Parasitic helminth SCP/TAPS proteins have been proposed to play important biological roles in the transition from the free-living to the parasitic stage during the invasion of the mammalian host [64,65]. In S. mansoni, this family is termed venom-allergen-like proteins (SmVALs) and comprises at least 28 members [32]. SmVALs can be divided into two groups and exhibit various gene expression patterns throughout the entire life-cycle, including genes exclusively expressed in stages involved in intermediate host invasion or definitive host invasion and ubiquitously expressed genes [32]. Because of their potential functional classification, expression patterns and localization, SmVALs have been proposed as potential drug targets and vaccine candidates, and some members have been well characterized recently [66–68]. For instance, Farias et al. have demonstrated that SmVAL4 and SmVAL26 stimulate differential allergic responses in a murine model of airway inflammation [66]; and in a recent study, egg-derived SmVAL9 was found to carry an N-linked glycan containing a schistosome-specific difucosyl element and function as an immunogenic target during chronic murine schistosomiasis [68]. In our study, 20 non-redundant S. japonicum venom-allergen-like protein (SjVAL) genes were identified in the genome. In order to illuminate the phylogenetic relationships of VALs in the two Schistosoma species, we constructed a phylogenetic tree using all the full-length protein sequences, except SmVAL23 of which conserved domain was incomplete. The result revealed that all the Schistosoma VALs could also be divided into two groups: sequences in group one contain the SCP_euk conserved domain (cd05380) and only a partial sequence had homologous counterparts between the two Schistosoma species; sequences in group two contain the SCP_GAPR-1_like conserved domain(s) (cd05382) and the majority had homologous counterparts (Fig 8A). Microarray analysis revealed that ten SjVAL genes were transcribed in adult worms, including four genes (AAW25717.1, AAW27353.1, CAX74321.1 and Sjp_0038950) that were differentially transcribed in worms from the four hosts and six genes (AAP06001.1, Sjp_0038960, AAW25247.1, AAW25499.1, AAW25007.1 and CAX74107.1) that were stably expressed in worms from the four hosts (Fig 8B). All four differentially expressed SjVAL genes belonged to group one, and three encoded proteins with a putative signal peptide, which may be excreted/secreted proteins and interact with their immediate environment. Moreover, further expression analysis, based on our recently released S. japonicum microarray data for cercaria, schistosomulum, adult worm and egg (GEO accession number: GSE57143) [20] indicated that these ten genes exhibited diverse expression patterns in the four developmental stages (Fig 8C). It suggests that the majority of Schistosoma VALs, if not all, may be susceptible to the environment in the life cycle implying a role in function of host adaptation. In S. mansoni, six SmVAL genes were demonstrated to be highly expressed in the cercaria by microarray analysis of larval stages associated with infection of the mammalian host, implying that the functions of these enigmatic genes are mostly associated with entry into the mammalian host [69]. Specifically, the transcription of gene Sjp_0038960 is restricted to the schistosomulum and adult stages, suggesting that this gene is important for schistosome parasitism in the mammalian host. Notably, only one gene, AAW25499.1, was constantly expressed among the four developmental stages, indicating that this gene may play a fundamental role in the schistosome life cycle. Indeed, AAW25499.1 was also stably expresses in worms from different mammalian hosts.
In conclusion, we systematically compared the gene expression profiles of schistosomes from laboratory animal hosts (BALB/c mice, C57BL/6 mice and rabbits) and the natural host (water buffaloes). The global transcriptional profiles of schistosomes from the four different hosts were generally coincident with each other. Meanwhile, a panel of differentially expressed genes was identified, which mainly involved in signal transduction and metabolism processes. A set of Schistosoma- or phylum Platyhelminthes-specific genes were differentially or stably expressed in schistosomes from the four different hosts and should be targeted in future hypothesis-driven functional studies. The findings of this study provide a rational basis for schistosomiasis research in different laboratory animals and natural mammalian host at the transcriptional level and a valuable resource for the screening of anti-schistosomal intervention targets.
|
10.1371/journal.pgen.1002443 | A Genome-Wide Analysis of Promoter-Mediated Phenotypic Noise in Escherichia coli | Gene expression is subject to random perturbations that lead to fluctuations in the rate of protein production. As a consequence, for any given protein, genetically identical organisms living in a constant environment will contain different amounts of that particular protein, resulting in different phenotypes. This phenomenon is known as “phenotypic noise.” In bacterial systems, previous studies have shown that, for specific genes, both transcriptional and translational processes affect phenotypic noise. Here, we focus on how the promoter regions of genes affect noise and ask whether levels of promoter-mediated noise are correlated with genes' functional attributes, using data for over 60% of all promoters in Escherichia coli. We find that essential genes and genes with a high degree of evolutionary conservation have promoters that confer low levels of noise. We also find that the level of noise cannot be attributed to the evolutionary time that different genes have spent in the genome of E. coli. In contrast to previous results in eukaryotes, we find no association between promoter-mediated noise and gene expression plasticity. These results are consistent with the hypothesis that, in bacteria, natural selection can act to reduce gene expression noise and that some of this noise is controlled through the sequence of the promoter region alone.
| Many biological processes in a cell involve small numbers of molecules and therefore fluctuate over time. As a consequence, genetically identical cells that live in the same environment differ from each other in many phenotypic traits, including the expression level of different genes. Our aim was to identify types of genes with particularly low or high levels of variation (“noise”) and to understand molecular and evolutionary factors that determine noise level. Working with the bacterium E. coli, we analyzed the expression—at the single cell level—of more than 1,500 different genes. We found particularly low levels of noise in genes that E. coli needs to live and genes that this bacterium shares with many related taxa. This suggests that cellular functions that are particularly important for this organism evolved towards low levels of variation. In contrast to previous results with yeast, we find that genes that change their expression levels in response to environmental signals do not have high levels of noise. This suggests that there may be fundamental differences in how noise is controlled in bacteria and eukaryotes.
| The phenotype of an individual is often considered to be a product of the individual's genotype and the environment in which it lives. However, significant phenotypic differences may exist between genetically identical individuals living in a homogeneous environment [1]–[7]. In the absence of genotypic differences or environmental cues, these differences often arise from random molecular processes during protein expression and development. In these cases, such variation is termed phenotypic noise. Although differences between individuals that are due to phenotypic noise are not encoded genetically, the level of phenotypic noise in a given gene may be subject to genetic control. One fundamental question is whether natural selection acts to control or promote phenotypic noise, and how organisms achieve this control.
It is well established that selection acts strongly on mean expression level [8]–[12]. Additionally, there is good evidence that selection can also act on the variation of gene expression, that is, on the level of phenotypic noise. Many studies with bacteria and other microorganisms have identified genes with exceptionally high levels of phenotypic noise, and several studies have provided possible adaptive explanations. Both theoretical [13]–[17] and empirical studies [18]–[21] have shown that increased noise and bistable gene expression can allow organisms to persist in fluctuating environments, and that selection may thus in some cases increase phenotypic noise. Other studies have shown that it can promote the formation of specialized subpopulations that engage in division of labor [5], [22].
However, there have been fewer studies on general patterns of gene expression noise, for example, across functional groups of genes. The best-established connection, and the only connection established for both eukaryotes and bacteria, is between mean expression level and variation in expression: strongly expressed genes have high levels of variation across cells [23], [24]. However, mean expression level does not fully determine variation: analyses in yeast have shown that when mean expression level is accounted for, gene expression noise exhibits certain strong patterns: for example, there is a positive association between gene expression noise and gene expression plasticity (i.e., variation in gene expression across environments) [24]; genes with TATA boxes exhibit high noise [24]; and those genes most critical for cell functioning exhibit lower levels of variation than other genes that are expressed at the same level [24]–[26]. This latter correlation is consistent with selection acting to decouple variation in expression from mean expression in order to decrease noise in important genes. However, this association is confounded by other correlations, such as the strong relationship between noise and expression plasticity.
There is no data addressing the question of whether functionally important genes exhibit lower levels of noise in bacteria: only one analysis of variation in gene expression has been performed in bacteria [23], which established that genes expressed at higher levels exhibit more extrinsic noise. This raises the question of whether these two properties can be decoupled, for example to lower noise in functionally important genes, even though these genes may be expressed at high levels.
Thus, although there is good evidence in yeast that genes important for cell functioning have lower levels of gene expression noise, the interpretation of this result as evidence of selection acting to decrease noise has been complicated by the association between expression plasticity and noise. Additionally, there have been no analyses of whether the decoupling of mean expression level and variation in expression exhibits such general patterns in bacteria. Here, we investigate this possibility, and whether such decoupling exhibits patterns on a general, genome-wide level.
In contrast to previous studies, which have examined protein expression noise, we carried out a comprehensive analysis of the noise conferred by the promoter regions alone in E. coli. Our goals were three-fold. First, we wanted to test whether the DNA sequence of the promoter region has a substantial and consistent effect on noise. Second, we asked whether differences in noise exhibit discernible patterns, for example across functional categories of genes. Finally, we assessed whether these patterns are consistent with selection acting to preventing or promoting phenotypic noise, or whether other explanations account equally well for the patterns we observe.
We used an E. coli promoter library [27] consisting of 1832 strains, in which each strain carries a low-copy number plasmid (3–5 copies per cell [28], [29]) with an E. coli promoter region inserted upstream of a gene for a fast-folding green fluorescent protein (gfp). This library comprises about 75% of all E. coli promoters. The term ‘promoter region’ refers to the intergenic region between two open reading frames, together with 50–150 nucleotides of both the upstream and downstream open reading frame [27]. The mRNA that is produced consists of a transcriptional fusion between a short region of the 5′ end of the native mRNA, 31 bp that are identical for all promoters, and the open reading frame for GFP. A strong ribosome binding site (RBS) is located immediately upstream of gfp. As the 31 bp preceding the gfp start codon are identical for all constructs, effects from differences in the translation initiation rate should be minimal [30], [31]. Additionally, as approximately 90% or more of the mRNA sequence is identical for each construct, in most cases, differences in mRNA half-lives between constructs are likely to be small. The GFP variant is quite stable, so decreases in protein concentration occur primarily through cell growth and division. For the above reasons, differences in the mean concentration of cellular GFP for different promoters are most likely due to differences in transcription (see Text S1). However, in many instances the promoter region may affect mRNA half-life or translation dynamics, since it contains a fraction of the native open reading frame.
This experimental system removes several mechanisms that are likely to affect protein expression noise in the native context. Among these is the chromosomal context of the gene; the mRNA sequence content, affecting both mRNA half-life and translation; and the amino acid sequence, affecting protein degradation. In fact, the only variable among the constructs is the sequence of the promoter region. By definition, then, the effects on noise that we measure here are due to the promoter sequence alone. This experimental approach thus allows us to investigate whether and how the promoter sequence alone affects noise. Although this promoter-mediated noise contributes only partially to the total noise exhibited by a protein, it may play an important role, which we investigate here; later we use data on protein noise to explore other factors that contribute to affecting protein expression noise.
To quantitatively measure variation in gene expression from each promoter, we grew a clonal population of each strain, and used flow cytometry to measure the GFP concentration in approximately 100'000 individual cells from each population. For each strain, we extracted a small gated subset of cells (Figure S1; see Methods). This gating has the effect of minimizing extrinsic variation due to physiological differences among cells, such as cell cycle timing, slow growth, or other physiological stresses (see Text S1). For each of 1832 strains containing a promoter region from E. coli, we measured the mean and variance in fluorescence. 1522 of these yielded measurements significantly above background (GFP vector lacking a promoter; see Methods). We use the data from these 1522 promoters for the remainder of our analyses.
The mean and variance of fluorescence are highly repeatable measurements; when they were assessed for independent cultures, repeated measurements were extremely accurate (r2 = 0.998 and 0.91, for mean and standard deviation, respectively). This repeatability existed even when the cultures were grown in different laboratories, measured on different flow cytometry machines, and when different methods were used to filter events (r2 = 0.92 and 0.51 for mean and standard deviation, respectively; see Methods and Figure S2). Mean fluorescence levels varied over almost 3 orders of magnitude, qualitatively similar to the variation in mRNA levels observed in other studies [23]. Comparing our data on mean fluorescence level with published quantitative data, we also find that our data set correlates well with measured transcript levels, and is thus likely to capture an important aspect of mRNA transcription (see Text S1).
We find a strong dependence of variation in expression on mean expression level for any particular promoter (Figure 1), as has been observed previously [23], [24], [32]. Because the primary effect of selection on gene expression occurs as stabilizing selection on mean expression level [8]–[11], and mean and variation are closely coupled, we use a metric that decouples variation in expression from mean expression level. Modifying the method outlined by Newman et al. [24] we measured noise as the vertical deviation from a smoothed spline of mean log expression versus the coefficient of variation in log expression for all promoters in the library (see Methods; Figure 1F; Text S1; Dataset S1). When describing our findings, the term ‘phenotypic noise’ or ‘noise’ always refers to this metric in which variation is corrected for mean expression; such a measure allows us to assess whether variation in gene expression is controlled independently of the mean.
We emphasize that we use the term ‘noise’ to refer to relative differences in variation when mean expression level is controlled for. Thus, it is a qualitative measure, and for this reason we emphasize comparative results of relative differences in promoter-mediated variation; also for this reason, we restrict our statistical analyses to non-parametric tests. We refer to this measure as ‘noise’ because it is a reflection of differences between cells that are likely to arise from stochastic events, but it is not a quantitative measure of the frequency or effect of those events. In addition, because we have functional data for genes only, and not promoters, when we refer to the noise of a ‘gene’ or the functional category of a ‘promoter,’ we are referring only to the gene that lies directly downstream of the promoter, unless otherwise specified.
When we calculate this noise metric for the entire library of promoters, we find excellent repeatability, even in different culture conditions. The correlations range from 0.50 (Spearman's rho) when using data from different labs, to 0.58 when using data collected in independent experiments in the same lab (Figure S3). These are lower limit estimates of repeatability, as in each of these comparisons different culture conditions were used (see Methods). The repeatability of the noise metric implies that each promoter sequence has a consistent effect on variation in expression: thus, as suggested above, there are characteristics inherent to each promoter that result in different levels of noise.
Noise in gene expression consists of different components [33], [34], and our experimental system mostly reports one of them, promoter-specific extrinsic noise. Since the promoter-gfp construct resides on a plasmid with several copies, the cellular GFP concentration is the sum of the contributions from individual promoters. Intrinsic noise – variation generated at the level of one single promoter – is therefore decreased. In addition, because the GFP protein has a longer half-life than mRNA, the sensitivity of these noise measurements to intrinsic noise events in transcription is decreased [35]. Finally, fluctuations in plasmid number, which are expected to increase noise in all strains equally, may decrease the sensitivity of this system.
The noise that we measure is thus a qualitative and relative indication of the amount of promoter-specific extrinsic transcriptional noise [33], [34]. If we measure high levels of noise in a protein controlled by a particular promoter, most likely this occurs because transcription from this promoter is controlled by factors (or regulatory networks) having higher noise, or that this promoter is more sensitive to global extrinsic noise factors (e.g. variations in polymerase numbers) than other promoters. This experimental system is thus useful to examine extrinsic promoter-mediated noise on a genome-wide scale, and to ask how the level of extrinsic noise differs among promoters.
Even though, as discussed above, our plasmid-based system only captures some aspects of gene expression, we find that it gives similar results to chromosomally integrated systems in both mean and variation of expression. We measured the mean and variation in expression for nine chromosomally integrated promoter-gfp fusion constructs [36], and found that both the mean and CV correlate well with the values that we find for the plasmid-based system (rho = 0.85, p = 0.006; rho = 0.77, p = 0.016 for mean and CV, respectively; see Text S1 and Figure S4).
Given that the promoter sequence alone has a consistent influence on mRNA expression and noise levels (above; Figure S3), this raises the question of whether these levels of noise systematically differ for different classes, or types, of promoters. One broad division that can be made is between promoters that drive the expression of essential genes and those that drive the expression of non-essential genes (we define a gene as essential if its deletion yields an inviable genotype in rich media [37]). We used data for 118 promoters that lie directly upstream of essential genes or operons [38] that contain at least one essential gene, out of 1456 promoters for whose downstream genes we have information about essentiality. We find that promoters of essential genes exhibit significantly lower levels of noise than other promoters: of the genes with the lowest level of noise (first quartile), 13.1% are essential; of the genes with the highest level of noise (fourth quartile), only 2.9% are essential (p = 1.0e-6, Wilcox rank sum test). This difference is not driven by any mechanisms relating to mean expression levels, since our measure of noise corrects for this. Thus, the promoter regions of genes that are essential in the laboratory environment have evolved such that essential genes have lower noise levels.
Essentiality in the laboratory is an incomplete and potentially biased measure of a gene's importance in the natural environment. We thus also looked at gene conservation, which may capture additional aspects of functional importance [39], [40]. Considering non-essential genes only, we found a negative relationship between noise and functional importance: non-essential genes that have high levels of conservation in the gamma-proteobacteria clade (of which E. coli is a member) have promoters conferring low levels of noise (Spearman's rho = −0.19, p = 7.2e-12, n = 1350; Figure 2 and Figures S5 and S6). Furthermore, this relationship between conservation and expression noise exists within functional categories: it does not depend on broad differences in conservation between genes of different function, for example between genes involved in RNA production (expected to be more conserved and less noisy) versus those involved in carbon metabolism (expected to be less conserved and more noisy; Figure S7).
Together with the above data on essential genes, this suggests that the promoter regions of functionally important genes confer low levels of noise; given that the major effect of promoter sequence on protein level occurs through mediating transcription, this decrease in noise likely occurs through the control of transcriptional processes. The transcriptional regulation of some bacterial genes has been shown to be constructed such that increased noise is a result [41]; the data here suggest that on a genome-wide basis there is a tendency for functionally important genes to be controlled by less noisy transcriptional processes, that this trend extends beyond essential genes to conserved, non-essential genes, and that this trend persists within functional categories of genes.
There are several possible explanations for the low levels of noise observed in essential and highly conserved non-essential genes, two of which we discuss here (we explore a third explanation in the following section; however, this list is not exhaustive). First, it is possible that essentiality and gene conservation are good descriptors of the functional importance of a gene, and that selection has acted to decrease noise in such genes. This has been the explanation put forth in previous analyses. A second possible explanation is that low noise levels are difficult to evolve, and as conserved and essential genes have also spent more evolutionary time in a particular genome than non-conserved genes, selection has had more time to minimize noise in these genes. Either of these explanations could result in conserved and essential genes having lower noise. However, the results of our analysis suggest that the second explanation is less likely, for the following reasons.
First, the correlation between gene conservation and noise exists even for genes that have been acquired very distantly in the past. We looked for an association between functional importance and noise considering only genes acquired before the divergence of the E. coli lineage from alpha-proteobacteria (approximately 2.5 billion years ago [42]). These genes have had ample time for noise minimization. Thus, if the time a gene spends in a particular genome is a strong determinant of noise, there should be no relation between conservation and noise in this set of genes, as all have spent at least 2.5 billion years in the E. coli lineage. However, the correlation between conservation and noise within these anciently acquired genes remains strong (Spearman's rho = −0.23, p = 2.8e-4, n = 249). That the amount of noise minimization is related to the level of evolutionary conservation (functional importance) even in anciently acquired genes strongly suggests that the time that a gene spends in an organism has little to do with the level of noise it exhibits.
Second, although horizontally transferred genes are generally enriched for genes of lesser functional importance, many genes important for cell functioning have been horizontally transferred (e.g. some ribosomal genes). Among genes that have been recently horizontally transferred into E. coli [43], strongly conserved genes have lower levels of noise (correlation between noise and conservation: Spearman's rho = −0.22, p = 6.9e-3, n = 221 for genes transferred after the split of E. coli from Haemophilus; Spearman's rho = −0.25 p = 4.8e-4, n = 171, for genes transferred after the split of E. coli from Buchnera). When we consider very recent horizontal gene transfers the negative correlation remains (Spearman's rho = −0.16, p = 0.23, n = 65 for genes transferred after the split of E. coli MG1655 from E. coli CFT073). Although this correlation is not significant, there are only a small number of recently transferred genes, and these vary little in their levels of evolutionary conservation, decreasing the explanatory power of this variable. Given that the nucleotide divergence between MG1655 and CFT073 strains is approximately 2% [44], finding a negative correlation of similar strength (−0.16 vs. −0.19 for the entire data set) is notable.
Thus, the relationship between functional importance and noise does not appear to be related to the time that a gene has spent in a genome. The latter result also implies that the decreased noise observed in functionally important genes, if due to selection, can occur via a small number of mutations. Alternatively, it is possible that features of the promoter that influence noise act independently of the genetic background, so that genes retain characteristic levels of phenotypic noise even when horizontally transferred. We do find some support for this latter hypothesis: promoters of very recently horizontally transferred genes (ORFan genes; e.g. [45]) do not exhibit higher levels of noise than other promoters (Wilcox rank sum, p = 0.69, n = 37).
Our results, showing that functionally important genes exhibit lower gene expression noise, is consistent with the hypothesis that selection has acted to decrease noise in genes important for cell function. However, many other factors may potentially play a role in determining noise. A crucial determinant of noise in gene expression may be in how the gene is regulated: genes that exhibit large expression plasticity, meaning that they can undergo strong repression or activation across different environmental conditions, might be controlled in ways that makes them intrinsically more noisy. A very strong association between expression plasticity and noise has been found previously in yeast [24]–[26].
To investigate whether there is a similar association between noise and expression plasticity in E. coli, we gathered data on changes in gene expression across 240 pairs of environmental conditions [46]. For each pair of conditions, gene expression changes are expressed as the log ratio of expression in one condition relative to a reference condition; the value is positive for genes that increase their expression, and negative for genes that decrease their expression under the respective environmental condition. For each gene, we calculated the median of the absolute values of the expression changes. This value, which we term the expression plasticity, is high for genes whose expression frequently varies between two conditions, and low for genes whose expression is usually constant between two conditions, regardless of whether this occurs through repression or activation, or the nature of the reference condition.
Surprisingly, we found no significant association between noise and expression plasticity in E. coli (Spearman's rho = 0.030, p = 0.27, n = 1354). It is possible that this correlation exists only in some growth conditions, and that these types of conditions are under-represented in the dataset. To account for this possibility, we grouped the condition pairs by their similarity in expression changes into 18 clusters, calculated the median of the absolute values of the expression changes, and again found no significant correlation (Spearman's rho = −0.002, p = 0.94, n = 1354). Performing a similar analysis for yeast yields a significant positive relationship between expression plasticity and noise (data from [47]; unclustered analysis: Spearman's rho = 0.22, p = 7e-26, n = 2479). Although the lack of a correlation in E. coli could be driven by differences in data quality, this is not a likely explanation (see Text S1 and Figure S8).
Together, these data suggest that in yeast, a substantial fraction of gene expression noise might be a consequence of requiring dynamic control of gene expression [26]. However, in E. coli, high gene expression noise is not an unavoidable consequence of genes having high expression plasticity. Further supporting this conclusion is the association between functional importance and expression plasticity in E. coli: essential and conserved genes are the most dynamically regulated: 42% of essential genes are among the most dynamically regulated genes (within the top quartile), while only 13% are among the least dynamically regulated (bottom quartile) (p = 5e-6, Wilcox rank sum for essential versus non-essential genes; Spearman's rho = 0.19, p = 1.1e-11, n = 1209 for the correlation between expression plasticity and conservation). Despite this, promoters of essential and conserved genes exhibit the lowest level of noise. Thus, in E. coli, there does not appear to be a constraint preventing promoters with high expression plasticity from having low noise. In contrast, there is a strong positive correlation between expression plasticity and noise in yeast, suggesting that for many genes, such a constraint may exist. Because essential genes in yeast have low expression plasticity (see Text S1), the previous finding that essential yeast genes exhibit low levels of noise might be a consequence of this association between expression plasticity and noise.
We looked in more detail at how specific functional aspects relate to gene expression noise. We grouped genes according to the categories outlined by MultiFun [48], and found substantial differences between genes having different functional roles (Figure 3). Relatively low levels of noise were exhibited in genes involved in DNA structure (i.e. methylation, bending, and super-coiling) and building block synthesis (biosynthesis of amino acids, nucleotides, cofactors, and fatty acids). Low levels of noise in such housekeeping genes might be expected, given that normal cellular activities are probably compromised if these proteins are too abundant or not abundant enough, as has been suggested previously [49], [50]. We also observed particularly low levels of noise in genes involved in protection (from radiation, cell killing, drug resistance, or for detoxification). Finally, promoters annotated as having binding sites for σ32 (control of heat shock genes) have significantly lower levels of noise; several transcription factors are also associated with low noise (Table 1).
Particularly high levels of noise are primarily found in genes involved in two functional groups: energy metabolism of carbon sources (e.g. glycolysis, the pentose phosphate shunt, fermentation, aerobic respiration), and in adaptation to stress (osmotic pressure, temperature extremes, starvation response, pH response, desiccation, and mechanical, nutritional, or oxidative stress). Finally, promoters with binding sites for σ38 (control of starvation and stationary phase genes) exhibit higher levels of noise than promoters containing binding sites for other sigma factors; several transcription factors were also associated with higher noise levels (Table 1).
As the above analysis implied that high levels of noise are not simply a consequence of having high expression plasticity, the differences in noise between categories is consistent with differential selection (although other factors may also be responsible). For example, one possibility is that some genes exhibit high levels of noise due to an absence of selection (such that drift dominates the evolutionary process), in contrast to the majority of genes in the genome. A second possibility is these genes have experienced selection for high levels of noise. Variation in resource utilization between cells can sometimes increase the growth rate of clonal populations [19], [51] by promoting the utilization of carbon sources that become newly available. Similarly, noise in genes involved in adaptation to stress could allow genotypes to persist under conditions where stressors appear quickly [13], [52], [53]. Alternatively, genes with high noise may also be constrained in their ability to evolve lower noise due to trade-offs with other functions that we have not measured. These results thus generate explicit and testable hypotheses about the possible adaptive functions of increased noise in gene expression.
Our focus until now has been on how the nucleotide sequence of a promoter alone controls phenotypic noise in a plasmid-based context. Noise at the level of protein is possibly controlled through additional mechanisms acting at the post-transcriptional level. To include these mechanisms into our analysis, we used data from a recent study that measured variation in protein numbers between cells for a large number of the protein coding genes in E. coli [23]. This study was based on translational fusions of protein coding genes with YFP in the native chromosomal context. Using approximately 1'000 of these constructs, the authors used microscopy to measure the mean and variation in protein number per cell. This study thus provides us with information on the sum of intrinsic and extrinsic noise that occurs through both transcriptional and post-transcriptional processes.
Using this dataset, we quantified protein expression noise in an analogous manner as for our data, removing genes with very low expression levels and correcting for mean protein expression level. Again, this decouples mean protein expression level from variation in protein expression. We find a significant but weak correlation between protein noise in this dataset and gene expression noise in our own (Spearman's rho = 0.12, p = 0.02, n = 334). A primary reason for this low correlation may be that the noise in protein expression was measured during late exponential phase, while we measured during early exponential phase growth (see Text S1). We find that, similar to the pattern observed for promoter-mediated noise, essential and conserved genes have low protein expression noise (Wilcox rank sum, p = 3e-4, n = 116 essential genes; Spearman's rho = −0.21, p = 7.0e-9, n = 645 non-essential genes). Using variation alone as a metric of noise, without the correction for mean expression level, gives the opposite result: essential genes have significantly higher levels of variation [23], as they are expressed at higher levels, and variation is strongly positively correlated with mean expression. Finally, corroborating the lack of correlation between promoter-mediated noise and expression plasticity, protein expression noise and plasticity exhibit no significant correlation (rho = 0.052, p = 0.16, n = 724).
We find that post-transcriptional processes play a role in controlling protein expression noise: genes with high protein expression noise have slightly higher rates of translation initiation (Spearman's rho = 0.17, p = 3.3e-6, n = 730; computational predictions of ribosomal initiation rates from [30], [54], and slightly longer mRNA half-lives [55] (Spearman's rho = 0.15, p = 4.4e-5, n = 689). This is consistent with the idea that intrinsic noise in post-transcriptional mechanisms has a significant effect on total noise, as theoretical models have suggested [34], [56]–[58]. However, the extent to which the cell actually employs these mechanisms has remained unknown. The data here suggest that these mechanisms affect the noise levels of many genes in E. coli. If this association has occurred through selection, this implies that although these mechanisms are quite costly for the cell [59], the advantage of controlling intrinsic noise outweighs the energetic costs that it imposes.
We have shown here that by using a simple plasmid based system that different promoters consistently confer different levels of phenotypic noise. In particular, we find that functionally important genes have promoters that confer lower levels of gene expression noise, and certain functional categories are enriched or depleted for promoters that confer high noise. The noise metric we use accounts for mean expression level, so these patterns are not due to differences in expression levels between essential and non-essential genes, or to characteristics related indirectly to expression level (for example, systematic differences in cellular stress levels due to GFP). Furthermore, these noise characteristics appear to extend across different growth conditions, as promoter-mediated noise is similar during growth in non-stressful (arabinose and glucose) and stressful (low-levels of antibiotic) conditions (see Figure S3).
We have excluded several confounding factors from the association between noise and functional importance, including the age of the gene and the association with expression plasticity. The lack of association between promoter sequence and expression plasticity is surprising, given the strong relationship that has been observed in yeast [24], and that promoter sequence is a strong determinant of transcript level (see Text S1). The low noise of promoters of functionally important genes is consistent with the hypothesis that natural selection acts to control against variation in proteins that are important for cellular functioning [60]. However, it is important to emphasize that we cannot exclude other factors being responsible for this pattern.
We cannot yet determine the level at which the effects of promoter-mediated noise control extend to the protein level. Processes downstream from transcription may have significant effects on noise, and might sometimes overwhelm the effects arising on the transcriptional level. The association that we find between promoter-mediated noise and protein noise suggests that in many cases, transcriptional noise does correspond with the noise observed further downstream. However, we cannot say how strong this association is for all genes.
As our noise metric largely excludes both intrinsic noise and global extrinsic noise, these results suggest that promoter-mediated noise is systematically reduced in functionally important genes through gene-specific mechanisms. Thus, it seems that the regulatory inputs for these promoters have evolved to minimize noise. This has been shown previously for single regulatory networks [61]; here we show that it also appears to occur for many different genes. In addition to promoter-mediated control of noise, we find that proteins that exhibit low levels of noise have short mRNA half-lives and low rates of translation initiation. Although previous work has shown that variation in expression is strongly positively associated with mean expression level [23], the data here show that these two characters can be uncoupled, so that transcriptional noise can be controlled independently of the mean, and that this uncoupling is stronger for some types of genes (those that are functionally important) than others.
Although it has been hypothesized previously that functionally important genes have been selected to exhibit low levels of noise [62], it has been difficult to unambiguously show this. In particular, it has been difficult to separate the effects of expression plasticity and low noise, as all previous studies connecting noise and functional importance have been in yeast, where this association is quite strong [24]–[26] (see Text S1). The data shown here provide evidence that in E. coli, these two characteristics are unconnected.
In eukaryotes, one of the dominant regulatory mechanisms associated with transcriptionally noisy genes is chromatin structure (noisy genes tend to contain TATA boxes and are frequently regulated by SAGA [21], [24], [63]). A corollary of this is that in yeast there is a strong association between noise and expression plasticity, as dynamically regulated genes are often associated with chromatin remodeling factors. Much of this noise is thought to arise because of the two step process inherent in eukaryotic transcription, in which initial access to the DNA occurs through relaxation of histone binding, followed by transcription factor and polymerase binding [64]. Homologous mechanisms do not exist in bacterial systems; this may fundamentally affect the correlation between noise and expression plasticity. Despite these mechanistic differences, we do find a significant positive correlation between the promoter-mediated noise in E. coli genes and protein expression noise in their S. cerevisiae orthologues (rho = 0.31, p = 0.015, n = 60; Figure 4). Thus, although these organisms might differ in the mechanisms affecting gene expression noise, genes of similar function do exhibit similar levels of noise. However, protein expression noise, as calculated from [23] exhibits no correlation with gene expression noise in S. cerevisiae.
The data presented here show that: (1) For many genes, the promoter region of a gene controls noise in a consistent manner; (2) Functionally important genes are controlled such that noise is decreased; (3) The lower noise observed in functionally important genes does not appear to result from these genes having been present in the genome for a longer period of time; (4) There is no correlation between the noise conferred by a promoter and the expression plasticity of mRNA expression that is controlled through that promoter. In particular, this latter observation implies that there may be fundamental differences between the mechanisms giving rise to phenotypic noise in bacterial versus eukaryotic systems.
All strains have been described previously [27]. Briefly, each strain in the library contains a plasmid with a ‘promoter region’ cloned upstream of a fast-folding GFP. These promoter regions consist of an intergenic region, together with 50–150 bp of the upstream and downstream genes. The inclusion of part of the upstream and downstream open reading frames ensures that the majority of transcriptional control elements are contained in the construct. The library contains all K12 intergenic regions longer than 40 bp. We note that although the system is plasmid based, copy-number variation is relatively low. The plasmid contains an SC101 replication origin, for which segregation is tightly controlled [29]. For this reason variation in plasmid number per cell is expected to be less than under a binomial distribution, although variation in plasmid numbers will contribute additional extrinsic noise.
The strains with chromosomal integrations of the promoter-GFP fusions have been described previously [36]. Briefly, the promoter-GFP fusions were cloned and inserted into the attTn7 locus using a delivery plasmid containing a multiple cloning site surrounded by the terminal repeats of Tn7 [65].
All strains were grown in minimal media (M9) supplemented with 0.2% arabinose. Overnight cultures grown in same media were diluted 1∶500 and allowed to grow to mid-exponential phase at 37°C, shaken at 200 rpm. The cells were incubated with Syto red 62 (Molecular Probes) to stain the chromosome. The filters used for cytometry were 488/530+/−15 for GFP and 633/660+/−10 for the nucleic acid staining. In calculating the repeatability of the noise metric (Figure S3), two additional growth conditions were used: M9 supplemented with 0.2% glucose, and M9 supplemented with 0.2% glucose and 2.5 ng/ml ciprofloxacin.
The data were collected from a culture containing cells in different physiological states and quality. To minimize heterogeneity driven by these processes, we selected a small subset of cells with minimal CV. For the majority of promoters, the CV of the population was minimized between 5,000 and 10,000 cells, although gating had only a minimal effect on CV, decreasing it by 10–20% at most. Larger values than this generally contained cells of differing size and complexity, affecting the variance in fluorescence; smaller values contained too few cells to be a reliable indicator of the population. Thus, for all promoters, fluorescence data for 100,000 cells was collected and this data was subsequently filtered so that the fluorescence data from only 10,000 cells were analyzed further. These data were exported into text files and analyzed using the R statistical framework [66] (the raw data is available at http://mara.unibas.ch/silander.html).
The filtering process occurred in one of two ways. For the majority of the analysis, it occurred as follows: (1) the first 1000 acquisition events were excluded to minimize inaccuracies in fluorescence measurements resulting from sample crossover and initial inaccuracies in measurements that we observed; (2) extreme outliers (all cells with red fluorescence values below ten and GFP values of one or less) were removed; (3) to enrich for cells in similar physiological states and stages of the cell cycle, for each promoter, a kernel density was fitted to the log red fluorescence data (indicative of the amount of nucleic acid in the cell), with Gaussian smoothing in which the density was estimated at 512 points using the method of Silverman for bandwidth selection [67]. The maximum value of this kernel density was determined, and 10,000 cells were selected from a symmetrical interval around this peak (see Dataset S2 for simplified code). This number of cells minimized the variation in GFP signal due to external influences (Figure S2), while still allowing us to sample a large number of cells. The mean, median, and standard deviation for this population of cells were then calculated.
For secondary confirmation of previous measurements, events were filtered on the basis of FSC and SSC alone: (1) again, the first 1000 acquisition events were excluded; (2) extreme outliers (all cells with SSC, FSC or GFP values of one or less) were removed; (3) a bivariate normal was fit to the log FSC and log SSC values, and values outside of two standard deviations were removed (cellular debris); (4) to enrich for cells in similar physiological states and stages of the cell cycle, a 2 d kernel density was fitted to the FSC and SSC data. The maximum value of this kernel density was determined, and 10,000 cells were selected from an elliptical gate around this point, oriented by the covariance between FSC and SSC (Figure S1). This gating was performed using the flowCore package [68]. Again, the mean, median, and standard deviation for this population of cells were calculated.
Several promoters gave rise to distributions that appeared to be either bimodal or have extremely high variances. The promoters having the highest CV (>0.6), and all promoters exhibiting a bimodal expression pattern were reanalyzed by restreaking for single colonies and measuring fluorescence a second time. We found that for all promoters exhibiting bimodal patterns, the bimodality disappeared upon restreaking to obtain a single clone; a previous analysis of protein levels in E. coli cells confirms the rarity of bimodal distributions [23]. We thus concluded that the bimodal distributions were likely due to contamination from a second promoter construct. For this reason, these promoters were removed the analysis. Three samples were removed from the analysis, one on the basis of abnormal DNA staining, and two due to small sample sizes.
We calculated a 95% confidence interval around the mean fluorescence of the empty vectors (containing gfp, but no promoter), and excluded all promoters with a mean fluorescence less than this range from the analysis (below 2.26 fluorescence units). There is thus only a 2.5% chance that the GFP signal for any promoter included in the analysis is due to only to autofluorescence.
Our goal was to define a consistent metric of noise in mRNA expression that enabled comparison of genes with different mean expression levels (in other words, to decouple mean from variation in expression). We thus followed a method similar to that outlined by Newman et al. [24], in which noise is defined as the deviation from a sliding window of the median expression level versus the CV for each promoter. To more robustly estimate the deviation, we defined noise as the vertical deviation from a smoothed spline (6 degrees of freedom) that covered a running median of mean log expression versus CV of log expression (window of 15 data points); a smoothed spline is not subject to the small deviations that a running median is (Figure 1F). For simplicity, we refer to this deviation as noise in gene expression, or noise. We note that noise is homoscedastic across expression levels: mean expression level versus noise or the absolute value of noise gives no significant regression. This is not the case for two related metrics of noise based on vertical deviation from a smooth spline: if log mean expression versus CV of expression or mean log expression versus standard deviation of log expression are used, both result in highly expressed genes having extreme levels of noise (either very high or very low) (Figure 1B, 1C, 1E). In contrast, for the metric of noise we use, genes having very high expression are not more likely to have extreme levels of noise. In addition, there is no significant correlation of noise with mean expression level (rho = −0.035, p = 0.17, n = 1522). Lastly, our results are robust when using similar noise metrics (e.g. vertical deviation from the running median, Euclidean distance from the smoothed spline, or if different spline fits are used; see Text S1). The noise metric is a highly reliable measure; for separate measurements of two independent cultures grown in different growth media yields a Spearman's rho value of 0.58 (p<1e-120; Figure S3).
Data on gene essentiality was taken from the PEC dataset [37]. Promoters were considered essential if they drove the expression of an essential gene or an operon containing an essential gene. For conservation, only the immediate downstream gene was taken into account.
Using data from Ragan et al. (2006), for each gene that appeared to have experienced horizontal transfer, we used the median value of the estimated phyletic depth at which the horizontal transfer occurred. We then selected those genes that had been acquired after the divergence of E. coli from Haemophilus (220 genes), Buchnera (170 genes), or E. coli CFT073 (42 genes), and used these sets to calculate the relationship in recently transferred genes between noise and gene conservation.
We calculated gene conservation using a reciprocal shortest distance strategy [69] to search for putative orthologues of E. coli genes in 105 fully sequenced gamma-proteobacteria or 58 alpha-proteobacteria [70]. We considered genes present in at least 30 out of 58 (>50%) fully sequenced alpha-proteobacterial taxa to have been acquired before the E. coli – alpha-proteobacteria divergence.
Promoters were grouped by functional class according to the gene annotations for the immediate downstream gene, as outlined in MultiFun [48] into eight major categories: metabolism, information transfer, regulation, transport, cell process, cell structure, cellular location, and extra-chromosomal element; each major category contained up to eight subcategories. To test for the enrichment of low or high noise genes, for each major category, each subcategory was tested against the remaining genes in that major category for enrichment of promoters with higher or lower noise using a Wilcox rank sum test.
Data on relative mRNA abundances and half-lives were taken from [55]. Data on relative mRNA expression levels (i.e. expression ratios) for 240 different conditions were taken from the E. coli Gene Expression Database (http://genexpdb.ou.edu/). These conditions were also grouped using hierarchical clustering into 18 clusters in which expression ratios were similar using the Lance-Williams formula as implemented by hclust in the R statistical package.
Data on both operon structure and the binding sites of sigma factors was taken from RegulonDB (http://regulondb.ccg.unam.mx/).
Orthologous genes in yeast were identified using a reciprocal best-hit analysis, with varying e-value cut-offs. The significance of the correlation, although low, is robust to changes in the stringency of the e-value cut-off (we note that as the stringency of this cutoff is increased, the number of orthologues decreases, necessarily decreasing the significance: e-20: rho = 0.2, p = 0.07; e-30: rho = 0.28, p = 0.02; e-40: rho = 0.26, p = 0.06; e-50: rho = 0.25, p = 0.11).
Unless otherwise specified, all categorical comparisons were performed using a non-parametric two-sided Wilcox rank sum test and all reported correlations are non-parametric Spearman rank correlations. The p-values for the Spearman rank correlations were calculated using the default settings of the cor.test() function in R, which uses an asymptotic t approximation.
|
10.1371/journal.pcbi.1004314 | A Neutrophil Phenotype Model for Extracorporeal Treatment of Sepsis | Neutrophils play a central role in eliminating bacterial pathogens, but may also contribute to end-organ damage in sepsis. Interleukin-8 (IL-8), a key modulator of neutrophil function, signals through neutrophil specific surface receptors CXCR-1 and CXCR-2. In this study a mechanistic computational model was used to evaluate and deploy an extracorporeal sepsis treatment which modulates CXCR-1/2 levels. First, a simplified mechanistic computational model of IL-8 mediated activation of CXCR-1/2 receptors was developed, containing 16 ODEs and 43 parameters. Receptor level dynamics and systemic parameters were coupled with multiple neutrophil phenotypes to generate dynamic populations of activated neutrophils which reduce pathogen load, and/or primed neutrophils which cause adverse tissue damage when misdirected. The mathematical model was calibrated using experimental data from baboons administered a two-hour infusion of E coli and followed for a maximum of 28 days. Ensembles of parameters were generated using a Bayesian parallel tempering approach to produce model fits that could recreate experimental outcomes. Stepwise logistic regression identified seven model parameters as key determinants of mortality. Sensitivity analysis showed that parameters controlling the level of killer cell neutrophils affected the overall systemic damage of individuals. To evaluate rescue strategies and provide probabilistic predictions of their impact on mortality, time of onset, duration, and capture efficacy of an extracorporeal device that modulated neutrophil phenotype were explored. Our findings suggest that interventions aiming to modulate phenotypic composition are time sensitive. When introduced between 3–6 hours of infection for a 72 hour duration, the survivor population increased from 31% to 40–80%. Treatment efficacy quickly diminishes if not introduced within 15 hours of infection. Significant harm is possible with treatment durations ranging from 5–24 hours, which may reduce survival to 13%. In severe sepsis, an extracorporeal treatment which modulates CXCR-1/2 levels has therapeutic potential, but also potential for harm. Further development of the computational model will help guide optimal device development and determine which patient populations should be targeted by treatment.
| Sepsis occurs when a patient develops a whole body immune response due to infection. In this condition, white blood cells called neutrophils circulate in an active state, seeking and eliminating invading bacteria. However, when neutrophils are activated, healthy tissue is inadvertently targeted, leading to organ damage and potentially death. Even though sepsis kills millions worldwide, there are still no specific treatments approved in the United States. This may be due to the complexity and diversity of the body’s immune response, which can be managed well using computational modeling. We have developed a computational model to predict how different levels of neutrophil activation impact survival in an overactive inflammatory conditions. The model was utilized to assess the effectiveness of a simulated experimental sepsis treatment which modulates neutrophil populations and activity. This evaluation determined that treatment timing plays a critical role in therapeutic effectiveness. When utilized properly the treatment drastically improves survival, but there is also risk of causing patient harm when introduced at the wrong time. We intend for this computational model to support and guide further development of sepsis treatments and help translate these preliminary results from bench to bedside.
| Sepsis, a systemic inflammatory response due to an infection, affects 900,000 Americans per year and its incidence is expected to increase over the next 10–20 years as the population ages [1]. While it is acknowledged that sepsis is a growing problem, its associated mortality rate has remained persistently high for the last 20 years and is currently near 20% [1–4]. Sepsis is now the leading cause of in-hospital death in the United States, yet there are no FDA approved specific treatments [5]. While understanding of the underlying mechanisms in sepsis has been rapidly improving, translation to clinically effective treatments has proven very challenging [6,7]. Much of this difficulty translating treatments may be the diversity and complexity of individual immune response and patient population [8,9]. These complexities lend themselves well to computational modeling, which can help integrate these complexities into a unified pathophysiological framework and optimize potential treatments [10].
Neutrophils are one of the first responders to sites of inflammation and play a critical role in the innate immune response. When effective, neutrophils migrate from the bloodstream through endothelial walls to the site of inflammation by sensing gradients of chemokines, which bind to neutrophil cell surface receptors. In early stages of sepsis neutrophils potentially play a duplicitous role, both actively fighting the invading pathogen but also contributing to undesirable systemic inflammation, which often leads to multiple organ dysfunction, immune paralysis, or death [11,12]. Neutrophils’ roles in sepsis are well recognized but the dynamics of multiple phenotypes and their impact on treatments is not fully understood. A key chemokine impacting neutrophil behavior and phenotype is interleukin-8 (IL-8). IL-8 signals through functionally distinct surface receptors CXCR-1/2, which are primarily expressed on neutrophils. CXCR-1 is primarily responsible for activating phospholipase D [13], which mediates respiratory burst and other pathogen killing functions. CXCR-2 has been shown to stimulate migratory functions such as chemotaxis and diapedesis [14,15].
The motivation of this work is to use computational modeling of CXCR-1/2 signaling, and the associated dynamics in neutrophil phenotype composition, to explore whether modifying this dynamic could be exploited to favorably impact outcome in sepsis. A population based mechanistic computational model, which incorporates both receptor level dynamics and neutrophil response to pathogen, was developed to explore the mechanisms involved in sepsis progression and calibrated in septic baboons. Furthermore, an experimental extracorporeal treatment which modulates CXCR-1/2 receptor levels was evaluated in silico using the model framework. The computational model described in this manuscript provides a physiologic rationale for neutrophil’s CXCR-1/2 mediated activity in sepsis, delivers insight into the overriding mechanisms involved, and suggests that interventions aiming to modulate phenotypic composition are time sensitive.
Of the 16 baboons subjected to bacterial infusion, 11 (69%) died and 5 (31%) survived, with death occurring within 6 days of bacterial infusion. Based on these two systemic outcomes, a thorough investigation of the model (see Methods section & Fig 1) was completed to identify parameter regimes that explain the dynamics of each group of the responders.
The initial conditions for the state variables of the ODE were fixed to simulate experimental stimulation (Table 1). Among the rate parameters, some were fixed to literature values. These included pathogen growth and decay rates, basal decay rates of naïve neutrophils, CXCR-1/2 internalization and recycling rates and creatinine decay rate (See fixed parameters in Tables 2 and 3). Remaining parameters were estimated by generating parameter ensembles using a Bayesian parallel tempering approach that fit our model to the survivor and non-survivor experimental data sets (see Methods). We conducted the parameter estimation process in two rounds. In round one, the model was fitted to the two data sets separately. By fitting to the two data sets separately, we were able to effectively show that the model was capable of replicating both lethal and non-lethal outcomes through only a change in few parameters. In an attempt to classify the underlying differences, we identified the parameters that were most influential in determining the outcome (survivor or non-survivor) of an individual using stepwise logistic regression. This resulted in a list of seven key parameters. These parameters tend to control the rate at which neutrophils grow and how quickly they can change phenotypes, which play a critical role in determining how quickly and severely the animal will respond to the infection.
Once these differentiating parameters were identified, we put the model through a second round of estimation. In the second round, the model was fit to both data sets simultaneously; allowing only the seven previously identified key parameters to vary between the survivor and non-survivor subpopulations (see Table 3). Additionally, two fixed parameters were allowed to take different values across the two populations to maintain the appropriate initial conditions in creatinine and white blood cell count. This step resulted in two new parameter ensembles that were identical in 28 parameters but varied in nine parameters. This second step enabled us to better crystallize the differences between animals that survived and those that died. These ensembles are biologically more relevant as we expect the animals’ immune responses to be highly similar, with small but important differences indicating susceptibility to a septic insult. Resulting full marginal distributions for each of the 7 parameters were statistically different across survivor and non-survivor populations (Fig 2). The final mean values and the standard deviation of all the estimated parameters are summarized in Tables 2 and 3.
Until now, the focus was on deriving parametric ensembles explaining the mechanism of sepsis progression in each population. In this section, the sensitivity of sepsis-mediated damage to different model parameters (and hence different processes in the network) was evaluated for each population. Area under the damage curve (AUCD) was used as an output metric of cumulative damage from sepsis. The analysis was done in two steps. First the sensitive parameters affecting damage in each population was identified to check if similar parameters were responsible for modulating damage within each population. Next, the two populations were combined to identify the parameters primarily responsible for a switch from a low to a high damage region. Since the model is highly nonlinear, a global sensitivity analysis (GSA) based on variance decomposition was chosen. This method decomposes the total variance in the output into variance and co-variance contributions from each rate parameter and its higher order combinations. To reduce computational cost, a meta-model based approximation was done (See Materials and Methods). The meta-model method called Random Sampling High Dimensional Model Representation (or RS-HDMR), decomposes the output function (AUCD) into a set of component functions that includes the mean followed by first order effects of each parameter and other higher order effects resulting from parameter combinations. The degree of sensitivity of a parameter or its combination with other parameters (as a set) is captured by Sobol’ index which by definition is the fraction of the total output variance attributed to the selected parameter set. To perform GSA, 4000 samples were generated from the parameter distributions of the two ensembles and the dynamics of the damage term was simulated for the survivors and the non-survivors. Fig 6(A) shows the AUCD distributions for each ensemble. As expected, the survivors show lower levels of cumulative damage than the non-survivors. The coefficient of variation was higher for the non-survivors (CV = 1.98) as compared to the survivors (CV = 0.32). When GSA was performed on the survivor and non-survivor samples separately and in combination, it was found that a third order RS-HDMR contributed close to 95% of the variance for both the populations. However, most of the important contributions were from the parameters constituting highly ranked first order indices. Fig 6B and 6C shows the first order and total Sobol’ indices for the first five most sensitive parameters of each population and Fig 6(E) shows the results when both populations are combined. Note that the total Sobol’ index for each parameter, is the sum of first order index and all higher order indices involving that parameter.
For GSA conducted separately on the survivor and non-survivor ensembles, it is found that damage is mainly determined by the decay rate of the killer neutrophils, kNK (direction of influence shown in Fig (6D)). The decay rate of the killer neutrophil controls the rate at which killer neutrophils are removed from the system, and the faster these neutrophils are removed, the lesser the damage. The next important term is the direct damaging effect of the killer neutrophils and this parameter has significant second order interactions with other parameters of the model as seen from the total sensitivity index. The next set of parameters has secondary importance and these parameters are different for the two populations (variance contributions of each parameter in this set is in the range, 1–10%). In survivors, damage is more influenced by the production rate of basal neutrophils and IL-8 in presence of the pathogen. In non-survivors, the effect is more pronounced for damage mediated IL-8 production (a positive feedback component), damage recovery term and killer neutrophil production rate. This indicates that overall damage in non-survivors is more sensitive to the parameters associated with killer cells, IL-8 and damage.
For GSA conducted on the combined population, the decay rate of killer neutrophils remains the most important parameter. Interestingly, the sensitivity value and ranking of three parameters increase relative to the case where the populations are analyzed separately. Among these, the transition rate of naïve neutrophils to the killer phenotype via CXCR1 (kNK−IL8) is the most important parameter. The next two parameters include the decay rate in filter Eq (7) (which determines the delay between pathogen generation and resulting neutrophil entry into circulation during sepsis) followed by the parameter controlling transition rate of killer neutrophil to the dual phenotype by CXCR2. Functional dependence of damage on these three parameters shows that they could be responsible for shift in the population from a low to a high damage region. For example, Fig 6(F) shows the dependence of AUCD on parameter kNK−IL8. Within each population, no particular trend is visible, but relative increase in the transition rate in the non-survivors correlates well with increased damage. Results in Fig 2 showed that the ranges of two of the parameters, kNK−IL8 and kfilter_off were significantly different for the survivors and non-survivors. Results from sensitivity analysis support this prediction and further show that the parameter values correlate well with the transition in observed damage.
Extracorporeal devices are emerging as promising therapies for treatment of sepsis[16–19]. In this instance we propose extracorporeal treatment which directly modulates CXCR-1/2 levels using a bioactive surface which interacts with unbound neutrophil surface receptors upon contact. Such a device, which is currently under development at the University of Pittsburgh, generates targeted and controlled downregulation of neutrophil surface receptors. The dynamics of this device can be analyzed within the framework of the generated computational model to determine its proof of principle in silico and help optimize treatment parameters. The proposed treatment implementation is shown in Fig 7. Specifically, the receptors are allowed to go to a trapped state and become unavailable for activation by IL-8 for the indicated time of treatment. To evaluate the potential of such an immunomodulatory treatment, we next performed an in silico trial by varying (1) the time when the treatment is introduced and removed and (2) the strength of interaction between the trapping device and the unbound neutrophil surface receptors.
This manuscript discusses the development of a mechanistic computational model of IL-8 mediated activation of CXCR-1/2 receptors in baboons which were administered intravenous E. coli. Neutrophil phenotypes, which dictate neutrophil functional response, were generated in silico based on CXCR-1/2 surface receptor levels, linking receptor level dynamics with neutrophil functional response. Parameter ensembles were generated for survivor and non-survivor populations, allowing for in silico observation of sepsis progression. Additionally, an extracorporeal treatment which modulates CXCR-1/2 levels on neutrophils was introduced in silico. This proof of concept evaluation allowed for preliminary device evaluation and optimization of treatment parameters.
To our knowledge, this is the first model describing dynamic interactions of neutrophils which specifically takes into account information sharing between the systemic variables and the receptor levels. The receptor level dynamics of the model function on a rapid time scale, adjusting to systemic IL-8 levels in a matter of minutes. These changes in receptor signaling dictate changes in neutrophil phenotype, which dictates neutrophil function and hence mortality. This link thus provides a valuable mechanistic framework that can be subjected to clinically relevant treatment scenarios. For example, the experimental treatment could be implemented purely on the receptor level. Alternatively, systemic variables such as IL-8 levels or neutrophil phenotype could be modulated to evaluate performance of hemoadsorption or neutrophil sequestration extracorporeal devices.
Application of parallel tempering approach for parameter estimation allowed for the efficient generation of ensembles of parameters and resulted in a model that could fit experimental data well [20], allowing reasonably accurate simulations of the system without making strong claims about the values of single parameters which are notoriously difficult to measure and are likely to vary between individuals. This allows for robust, population-level predictions rather than point predictions of model parameters and model behavior. However, the computed multi-dimensional posterior distribution in parameter space reflects constraints imposed by empirical data, as well as data sparsity and uncertainty. These constraints impose a covariance structure in the posterior distribution such that there is robustness in model behavior, despite large uncertainties in individual parameter values. Learning this structure is likely crucial in building predictive model [21,22]. Yet, the method is making no claim that individual parameter sets in the ensemble represent individuals in a population. At best, an individual could be represented by a smaller ensemble, reflecting uncertainty relating to this particular individual. Yet, it is fair to say that the ensemble is meant to represent uncertainly about a population of individuals, so that simulating the ensemble will provide expected behaviors across a population of individuals, as long as such behaviors are compatible with the empirical data used to generate the ensemble.
One trend that arose in the estimated parameter ensembles was a large difference in the magnitudes of different rate constants, sometimes spanning many orders of magnitude. This is not surprising, due to the inclusion of biological events spanning many time scales, ranging from fast molecular events to cell phenotype transitions and finally to the full duration of infections lasting for days. This suggests that future iterations of the model would benefit from a multiscale approach optimized towards handling these different time scales. Previous efforts [23–25] have worked out approaches that allow for efficient deterministic simulation of fast-scale molecular events, combined with more accurate stochastic simulation of slow-scale or rare events, and such techniques have resulted in impressive results [26,27].
Sensitivity analysis on the parametric ensembles enabled identification of the relative importance of the model parameters to state variables of the model. In general, sensitivity analysis is an important step in systems biology workflows and provides valuable information on model characteristics [28,29]. Most models in the literature resort to a local analysis which is sufficient if the parameters are well defined. For nonlinear dynamic models based on sparse experimental data and for systems which have inherently high parametric uncertainty, a global analysis needs to be done. Global techniques perform combinatorial perturbations of the parameters utilizing samples from the high-dimensional space. Application of meta-modeling approximations via RS-HDMR as was done in this work can significantly reduce the computational cost of sampling requirements for global methods. Additionally, if the sampling process takes into account parameter covariance computed from an ensemble model, biologically relevant sensitivity indices can be obtained. The systematic integration of ensemble modeling and global sensitivity analysis in this work allowed for identification of the parameters that control biological outcomes like sepsis induced tissue damage.
In addition to parameter fits, the behavior of the non-fitted state variables were inspected to check for features relevant to a clinical prognosis. Sepsis progression was analyzed by comparing differences between survivor and non-survivor populations. Neutrophil phenotypes in particular give insight into the differences between survivors and non-survivors. Of importance is the killer neutrophil population, which is highly elevated in the non-survivor population (see Figs 4 & 5). This neutrophil phenotype is associated with neutrophil induced tissue damage in the model. With support from sensitivity analysis, killer neutrophil decay rate, which sets the levels and dynamics of NK, was found to be the most important contributor to total damage in both the populations. Multiple studies support this finding, indicating that non-survivors or those with more severe sepsis experience increased levels of neutrophil induced tissue damage and MPO generation [11,30–33]. Furthermore, the importance of this term is supported by studies on neutrophil apoptosis and lifespan. Research by Taneja [34] and Fialkow [35] determined that neutrophil apoptosis was reduced in cases of severe sepsis, leading to increased lifespan of primed and activated neutrophils. Damage caused by these neutrophils was partially responsible for the progression of sepsis in these severe cases. Upon completion of the combined GSA, kNK−IL8 was also found to be a significant contributor to total damage. Increase of this term leads to preferential generation of the NK neutrophil phenotype, which directly contributes to tissue damage.
On the other hand neutrophils in the migratory phenotype were similar in survivor and non-survivor populations. These findings agree with the data from Cummings et al [32] which found neutrophil’s harvested from septic and non-septic patients migrated to IL-8 at similar levels. Interestingly, survivors and non-survivors had similar levels of neutrophil kill/migrate phenotype, indicating that both ensembles had adequate neutrophil populations to eliminate the source pathogen. Therefore, the additional damage in non-survivors was neutrophil induced resulting from elevated neutrophil killer phenotype levels. The IL-8 mediated killing functions of neutrophils are primarily triggered through CXCR-1 rather than CXCR-2. Modulation of CXCR-1 levels in particular may reduce the killing neutrophil phenotype and reduce neutrophil induced organ damage.
A number of experimental treatments for sepsis and other acute inflammatory diseases have targeted the CXCR-1 receptor with success in animal models [36–38]. However, translation to humans has been difficult for two main reasons [6]. First are inherent species dependent differences between human and animal immune systems that must be recognized and accounted for in pre-clinical studies. Second is the misuse of animal models and misinterpretation of pre-clinical data [39]. The recent debate on the translational fidelity of critical disease mouse models is a prime example where two separate comparisons of the human versus mouse genomic leukocyte responses using the same database resulted in two contradictory conclusions [40,41]. In the case of IL-8 signaling, which is not present in murine models, homologous cytokines and their associated surface receptors must be examined in IL-8’s place [42]. In this context, in silico modeling is an attractive alternative given that it allows preliminary evaluation of experimental human treatments at minimal costs.
Multiple extracorporeal sepsis treatments are currently under investigation with promising results. Blood purification techniques such as hemoadsorption [16,17,43–45] allow for cytokines and other detrimental proteins to be removed directly from the blood during the cytokine storm, curbing the patient’s immune response. Another approach called activated neutrophil sequestration [19,46], selectively removes harmful neutrophil phenotypes from circulation. In this instance we propose extracorporeal treatment which directly modulates CXCR-1/2 levels using a bioactive surface which interacts with unbound neutrophil surface receptors upon contact, resulting in CXCR-1/2 downregulation. This approach is advantageous because no components of blood are removed from circulation, allowing for a healthy immune response after appropriate modulation of neutrophil surface receptors. In addition, all necessary cell-cell interactions are allowed to occur within well-controlled microcirculation of the device. Such a setup also allows treatment to be easily titrated or halted by adjusting blood flow through the device. The dynamics of such a device were analyzed within the framework of the generated ensemble model to determine its proof of principle in silico and to evaluate its benefits in rescuing individuals marked as non-survivors by the parameter ensembles.
When evaluated in silico the proposed extracorporeal CXCR-1/2 modulation device improved mortality from 31% to above 80% when deployed under certain ranges of conditions. This substantial improvement in survival supports the hypothesis that a CXCR-1/2 modulatory device may improve patient outcomes. However, time and length of treatment implementation are critical parameters tied to this success. The importance of quickly beginning sepsis treatment has been well established [47], particularly for antibiotic administration. Our simulations showed a well-defined optimal time for the initiation of treatment, between 3 and 6 hours after the onset of severe infection. Treatment, if started within this time frame, had a high degree of success over a large range of treatment durations and strengths. This window is specific to the animal model under study and will not directly translate to a clinical setting for two main reasons. First, the model was calibrated with experimental data obtained from baboons, and differences between the baboon and human immune systems must be considered. Second, the baboons were exposed to a well-controlled bacterial infusion at a known time point, followed by a predictably quick and strong immune response. In this instance the pathogen load is well controlled and a large portion of the ensemble can therefore be addressed by a single treatment setting. In clinical practice, patients present with varied pathogen loads and they may be in different stages of infection and immune response. So, future experiments will need to combine clinical knowledge with additional data gathering and simulation to obtain treatment timing relevant for human patients.
Clinicians are actively searching for biomarkers to track sepsis disease progression and prescribe treatment [48–50]. Neutrophil phenotype may be a valuable indicator of disease state and individual patient response, but this information is difficult to collect in the clinic. Currently neutrophil phenotype can be evaluated either through functional testing or flow cytometry analysis of critical neutrophil surface receptors. In addition to CXCR-1/2 which are the focus of this model, CD11b, CD88, and CD62L all have roles in dictating neutrophil phenotype [51] and surface receptor expressions vary depending on severity of the inflammatory response. To more readily exploit phenotype data it may be possible to map neutrophil function to easily measurable biomarkers. Using these indirect measures of neutrophil phenotype can guide clinicians to ideal treatment regimens.
In conclusion, the ensemble model presented in this report provided key insights into the progression and mechanisms involved in progression of sepsis. We underline the role of relative abundance of killer, migratory and dual neutrophil phenotypes in deciding survivorship in an animal model. In addition, an in silico extracorporeal treatment which modulates CXCR-1/2 neutrophil surface receptors showed promising results. Further study and collection of experimental data will help further refine both the model and experimental device. Incorporation of data from a diverse patient population and expansion of current ensembles would increase the model’s generalizability, improving the potential for translation. Additional model parameters related to the device such as flow rate, surface area, and form factor could be included, allowing the model to streamline device development.
After general anesthesia, instrumentation and a 30 minute stabilization period, sixteen baboons (Papio ursinus) weighing between 19 and 32 kg were infused with 2 x 109 CFU Escherichia coli per kg over a two-hour period as described previously [52]. Thereafter, antibiotic therapy was delivered (gentamycin 4mg/kg twice a day). Eight animals were placed in an acute study lasting 6 days, while another eight were placed in the chronic study intended to last 28 days. All animals were observed for a 4-hour period after bacteria infusion then 11, 23, 35, 47, 72 hour and 6 days after infusion. Pathogen counts in blood, IL-8, creatinine, white blood cell, neutrophil elastase / α1-PI complex, and other physiologic parameters and biomarkers were gathered at multiple time point. For animals in the chronic study an additional time point was collected at 28 days. At the end of the study period, the baboons were again anesthetized for measurements and thereafter sacrificed with an overdose of pentobarbital. This study was approved by the Institutional Animal Care and Use Committee at Biocon Research Institute and animals were treated according to NIH guidelines.
A simplified mechanistic model of IL-8 mediated activation of CXCR-1/2 receptors and neutrophil response to a pathogen was developed based on available literature information and general knowledge of acute inflammatory response. Receptor level dynamics and systemic parameters were coupled with multiple neutrophil phenotypes to generate dynamic populations of activated neutrophils which reduce pathogen load, and/or primed neutrophils which cause adverse tissue damage when misdirected. Mathematical representation of the interactions detailed in Fig 1 were generated using ordinary differential equation (ODE) framework with the rate of interactions described by mass action kinetics or Hill type kinetics [53,54]. The interactions included in the model gives rise to 16 ODE state variables and 43 rate parameters.
In brief, the model is initiated by a pathogen load, which represents a bacterial inoculation. Presence of pathogen leads to continued growth as well as IL-8 and fMLP cytokine production. IL-8 is generated indirectly from pathogen generation from responding phagocytic mononuclear cells [55]. IL-8 initiates CXCR-1/2 activation in the receptor level, which in turn generates neutrophil phenotype change. Depending on phenotype, neutrophils may cause either pathogen elimination or misdirected tissue damage. A systemic damage indicator represents overall patient health. Increased systemic damage results in further IL-8 generation [56,57], resulting in a positive feedback loop. This simplified system captures the basic functionality of acute IL-8 mediated immune response to pathogen and is capable providing valuable feedback on potential therapeutic treatments modulating these mechanisms. A more detailed description of model equations follows.
The model contains 38 parameters, 13 of which are fixed based on literature data (Table 1). Parameter values were inferred using a Bayesian parallel tempering approach [22,67], which utilizes traditional Markov Chain Monte Carlo (MCMC) methods to sample the Bayesian posterior distribution P(p|y), the probability of parameter set p given data y, given by the Bayes formula
P(p|y)=L(y|p)θ(p)∫L(Y|p)θ(p)
where L(y|p) is the likelihood of observing y for a model with parameters p, θ(p) is the prior distribution, and ∫L(Y | p)θ(p) is the normalizing constant. Additional sampling efficiency is gained by running multiple parallel chains evolving at different temperatures. Higher temperature increases the likelihood of acceptance of proposed steps. This allows the high temperature chains to move more freely through the parameter space, avoiding getting stuck in local minima. This results in more efficient exploration of parameter space [20,68] a method we have applied extensively in parameter estimation of practically unidentifiable complex non-linear models [10,69,70]. This resulted in the creation of parameter ensembles, where each parameter is represented by a posterior distribution, rather than a single value. Free parameters were fit separately to the survivor and non-survivor experimental data sets, resulting in two parameter ensembles representing surviving and non-surviving animals.
In order to better capture the underlying biological differences between animals that survive and those that die following the same challenge we attempted to identify the most important parameters in determining animal fate. After computing ensembles for survivors and non-survivors, we performed regularized logistic regression, forward conditional stepwise logistic regression, and backward conditional logistic regression to identify a subset of parameters that are most indicative of outcome. Predictors consisted of all estimated parameters of both ensembles, and the indicator variable was the source (survivor or non-survivor) of the ensemble. Parameters were selected that were considered significant by all three methods, leaving a set of seven key parameters.
Global Sensitivity analysis was done to determine the independent and correlated contributions of rate parameters on cumulative damage. Area under the damage curve was chosen as the system output. To reduce the computational cost of GSA, Random Sampling High Dimensional Model Representation (RS-HDMR) approach was used [74]. Here, a multivariate output function (eg. AUCD) was approximately represented by weighted optimal expansion functions (called as component functions). The expansion coefficients of these functions were determined by least-squares regression simultaneously from one set of Monte Carlo samples. In general, for input vector, x¯=[x1,x2,…,xn] of rate parameters, in an n–dimensional space, a multivariate output function, f(x¯), is approximated by a sum of terms including the mean (f0) and the component functions (gl). Mathematically,
f(x¯)=f0+∑l=12n−1gl
Here, the index l indicates all possible combinations of the input parameters. In practice, not all component functions are significant and an F-test can be used to determine which component function should be excluded from the expansion [75]. For our work, we evaluated the variance based Sobol’ indices using these component functions. The workflow adopted here starts with generation of Monte Carlo samples of the rate parameters from the ensembles obtained by the parallel tempering approach. Since they come from the ensemble, information on the covariance between the parameter distributions for the population of survivors and non-survivors is retained. Next, a detailed procedure is followed which includes simultaneous construction of all the component functions, removal of non-significant component functions using an F-test ratio score, re-evaluation of component functions and finally evaluation of the Sobol’ sensitivity indices. The first order Sobol’ sensitivity indices which capture the influence of a single parameter (but averaged over the other parameters) are defined as:
Sl=Cov(f(x¯),gl(x¯))σ2,l=1,2,3,….,n
Here, σ2 is the total variance in the output and Cov(•) is the covariance between the output function and each of the first order component functions. For clarity, the component function, gl, is written as a function of x¯ but in reality it is only a function of the input parameter for which it is defined (for example, xl) and not the entire vector. Further, this sensitivity index is a sum of two terms that capture independent (Sla) and correlated contributions (Slb) of the input, which are defined as:
Sla=〈gl(x¯),gl(x¯)〉σ2
and
Slb=∑k=1k≠ln〈gl(x¯),gk(x¯)〉σ2.
The inner products, 〈•〉, are defined as:
〈gk(x¯),gl(x¯)〉=∫x1…∫xnw(x¯)gk(x¯)gl(x¯)dx1…dxn and w(x¯) is the probability density function of the inputs informed by the parameter ensembles. Similar equations can be written for the higher order component functions and sensitivity indices. Further details on the evaluation of the component function for various types of models are given in [74,76,77]. To determine the importance of a given parameter, it is necessary to combine all the important sensitivity indices (all orders) into a total sensitivity index, which for a parameter i can be defined as:
STi=Si+∑j=1j≠inSij+∑j<k=1j,k≠inSijk+…...
For most systems, very high order interactions are negligible and therefore, indices until the third order are sufficient, with most systems requiring only until the second order terms [74]. In this work, we constructed a third order RS-HDMR. All GSA computations were performed using the ExploreHD software (Aerodyne Research Inc., MA, USA).
|
10.1371/journal.pcbi.1003399 | Environmental Influence on the Evolution of Morphological Complexity in Machines | Whether, when, how, and why increased complexity evolves in biological populations is a longstanding open question. In this work we combine a recently developed method for evolving virtual organisms with an information-theoretic metric of morphological complexity in order to investigate how the complexity of morphologies, which are evolved for locomotion, varies across different environments. We first demonstrate that selection for locomotion results in the evolution of organisms with morphologies that increase in complexity over evolutionary time beyond what would be expected due to random chance. This provides evidence that the increase in complexity observed is a result of a driven rather than a passive trend. In subsequent experiments we demonstrate that morphologies having greater complexity evolve in complex environments, when compared to a simple environment when a cost of complexity is imposed. This suggests that in some niches, evolution may act to complexify the body plans of organisms while in other niches selection favors simpler body plans.
| The evolution of complexity, a central issue of evolutionary theory since Darwin's time, remains a controversial topic. One particular question of interest is how the complexity of an organism's body plan (morphology) is influenced by the complexity of the environment in which it evolved. Ideally, it would be desirable to perform investigations on living organisms in which environmental complexity is under experimental control, but our ability to do so in a limited timespan and in a controlled manner is severely constrained. In lieu of such studies, here we employ computer simulations capable of evolving the body plans of virtual organisms to investigate this question in silico. By evolving virtual organisms for locomotion in a variety of environments, we are able to demonstrate that selecting for locomotion causes more complex morphologies to evolve than would be expected solely due to random chance. Moreover, if increased complexity incurs a cost (as it is thought to do in biology), then more complex environments tend to lead to the evolution of more complex body plans than those that evolve in a simpler environment. This result supports the idea that the morphological complexity of organisms is influenced by the complexity of the environments in which they evolve.
| The “arrow of complexity” hypothesis [1] posits that the most complex products of open-ended evolutionary systems tend to increase in complexity over evolutionary time. Whether such a tendency exists is a long standing open question [2]–[6]. While it seems evident that more complex organisms exist today than at the advent of life, simple (single-celled) organisms continue to persist in large numbers, so it is clear that evolution does not guarantee complexity must increase. Moreover, loss of complexity has been observed in many species [7]–[9]. This begs the question: under what circumstances will complexity increase or decrease over evolutionary time? It is likely that particular environmental conditions are more likely to select for increased complexity than others, especially if this complexity comes at a cost.
As argued by proponents of embodied cognition, intelligent behavior emerges from the interplay between an organism's nervous system, morphology, and environment [10]–[14]. Therefore, if the ecological niche of a species remains constant and its body plan is evolutionarily constrained, then the neural system must adapt in order to succeed under this particular set of circumstances. This may be investigated experimentally through the use of evolving robots [15], [16] which stand in for biological organisms. For instance, it has been demonstrated [11], [17] that the complexity of an evolved neural system depends on the particular morphology it is controlling: in a given task environment certain morphologies can readily succeed with simple neural systems, while other morphologies require the discovery of more complex neural systems, or may prevent success altogether.
Another corollary of embodied cognition is that different environments will impose different selection pressures on the nervous systems and/or morphologies of organisms evolving in them. This can be studied by observing how organisms evolve in different environments. For instance, Passy [18] demonstrated that the morphological complexity of benthic colonial diatoms (measured as their fractal dimension) is significantly correlated with the variability of the environmental niches in which they are found. However, the biological evidence for a correlation between environmental and morphological complexity is sparse. This is in part because it is difficult to isolate systems where this may be studied effectively and to develop metrics that quantify morphological and environmental complexity. Ideally, it would be desirable to perform controlled investigations in which environmental complexity is under experimental control. Given enough time and resources it may be possible to carry out these investigations directly on living organisms. However, by performing experiments in silico, it is possible to do so with much greater speed and more precise control over experimental conditions. Specifically, by evolving virtual organisms [19] in physically realistic simulations, it is possible to faithfully model the relevant interactions between organisms and their environments.
Previously, the evolution of complexity has been investigated in silico using an alternative computational model [20]. In that work, populations of computer programs competed among themselves for the energy required to execute their instructions and gained energy by executing specific logic functions. With their system, Lenski et al. were able to demonstrate how complex functional features may evolve and how these features depend on the programs' environment. However, in that system the programs did not have bodies with which to physically interact with their environment. On the contrary, the evolutionary model employed here evolves embodied virtual organisms with evolutionarily determined body plans in physically realistic simulation environments. This provides a testbed for investigating how environment may influence the complexity of evolving physical morphologies.
Using in silico evolution to act on both the morphologies and nervous systems of simulated organisms or robots was first demonstrated by Sims [19], and has since been followed by a number of other studies (e.g. [21]–[32]). These studies employed a variety of experimental techniques, including different genetic encodings, morphological systems (such as branching structures or cellular aggregations), and evolutionary models. However, by constructing morphologies out of a relatively small number of geometric primitives, all of these studies were severely limited in the complexity of the morphologies which they could evolve, and therefore do not offer good test beds for investigating morphological complexity.
Recently, we introduced a new method for evolving virtual organisms that is capable of producing a greater diversity of morphologies than previous systems [33]. By using it to evolve organisms with restricted nervous systems in a variety of environments it was possible to demonstrate how such a system could be used for investigating the relationship between environmental and morphological complexity. Here, the results of [33] are refined and extended to demonstrate that selection for locomotion tends to induce selection pressures favoring more complex morphologies than would be expected solely due to random chance, and is therefore a driven rather than passive trend [3], [6], [34]. In subsequent experiments we employ a multi-objective selection mechanism to select for simplicity in addition to behavioral competency. This selection mechanism filters out morphological complexity that arises due to biases in the underlying evolutionary model or because of genetic drift, and only allows for complexity that confers a selective advantage on the simulated organism. Moreover, this selection mechanism acts to impose a cost on complexity as is thought to occur in biological organisms [35], [36]. Under this regime complex environments tend to induce selection for greater morphological complexity when compared to a simpler environment. This result supports the hypothesis that the environment plays an active role in determining morphological complexity.
In this work organisms are evolved in a variety of simulated environments in order to better understand the role of the environment in shaping morphological complexity. While inspired by the above mentioned studies in which the morphologies and controllers of virtual organisms were also evolved [19], [21]–[32], the system presented here has several advantages which make it better suited for studying the evolution of morphological complexity.
The first advantage relates to the task environments within which organisms evolve. The majority of the studies mentioned above were restricted to evolving for locomotion over flat terrain. While investigating this task has yielded interesting results, it suffers from its simplicity: simple morphologies composed of just a few cuboids or spheres are all that are needed to be successful. Even when more challenging task environments have been explored (e.g. those investigated in [37]), they employed morphologies composed of a small collection of cuboids and therefore the maximum complexity of their evolved morphologies was severely limited. In the current work, a variety of task environments with interesting properties are investigated, and morphologies with greater geometric detail are used, so it is possible to study the evolution of morphological complexity.
Another advantage of the current system is the way in which the genetic material that the evolutionary model acts on is encoded. As has been demonstrated in the past [25], [26], genetic encodings that simulate development to some extent offer demonstrable benefits over those that do not. This is because such encodings tend to produce regularities and symmetries in the phenotype; such patterns in nature are the inevitable result of biological development, which biases the kinds of phenotypes that biological evolution may act on [38]. For this reason, here we employ a particular form of genetic encoding that produces three-dimensional shapes with regular patterns (see Methods for more details) [39]. Each genome generated from this encoding generates a triangular mesh (trimesh) that forms the body plan of the virtual organism. Trimeshes allow evolution to craft morphologies with greater geometric detail compared to other systems in which evolution composes a small number of simple three-dimensional shapes together [19], [21]–[32] (see Figs. 1 and 2 for examples of morphologies evolved with the current system). Finally, populations of these genetic encodings are evolved with a commonly-used evolutionary model which has been demonstrated to be more evolvable than other evolutionary models [40].
The behavior of each virtual organism is simulated in a three-dimensional, physically-realistic virtual environment in order to assess its fitness. Because of the organisms' triangular mesh body plans and the complex environments in which they are evolved, evaluating the fitness of each organism requires considerable time. Moreover, many evolutionary trials were conducted in each of several environments to allow for meaningful statistical analysis. For these reasons all of the experiments were carried out on a 7.1 teraflop supercomputing cluster and required a total of over 100 CPU-years of distributed compute time.
In order to study the relationship between the morphological complexity of the virtual organisms and the task environments within which they evolve, evolutionary trials are conducted in each of 50 different environments. The first environment in which organisms are evolved is composed only of a uniform, flat, high friction ground surface (refer to Fig. 1a). The organisms evolved in this simple environment are considered control cases to compare against organisms evolved in other environments. Subsequent environments are more complex: they all consist of an infinite series of low friction rectangular solids over which an organism must locomote (see below for a characterization of this complexity). These “ice blocks” are constructed such that it is impossible for an organism to gain purchase by moving over their upper surfaces, but must instead reach into the gaps between the blocks to propel themselves forward in some fashion. This requires the evolution of morphologies with appropriate physical forms. Fig. 1 shows a sampling of these environments and virtual organisms that evolved within them.
The icy environments vary according to two parameters: the height of the blocks and the spacing between them. Each of these parameters varies from 0.025 meters to 1.6 meters exponentially for a total of different environments. These two parameters and the their exponential scaling are employed in order to produce a variety of qualitatively different environments that roughly approximate natural surfaces, but yet are also amenable to analysis and efficient simulation. There are certainly many ways in which the environments could be created to more closely approximate natural terrain, and there are many other factors which could influence the complexity of an environment, however the parameterization employed here provides a set of environments within which it is largely possible to evolve organisms capable of successful locomotion with the bare minimum of neural complexity. This allows for isolating the influence of environment on morphological complexity, which is the property of interest in this study (see Conclusions for further discussion).
For each icy environment, 100 evolutionary trials are conducted in that environment and a corresponding 100 evolutionary trials are conducted in the control environment (for a total of evolutionary trials; see Methods for details). Fig. 3 reports the mean distance that the best individuals from each trial traveled (computed across the 100 independent trials) in each icy environment. This figure demonstrates that there is a clear relationship between the environmental parameters and the difficulty of the task. Specifically, moving to the lower right in Fig. 3, where both the spacing and the height of blocks are large, the task becomes increasingly difficult: the organisms all become trapped in the gaps between blocks. Keeping the spacing constant and decreasing the block height (moving left in Fig. 3) gradually eases the task: the organisms are able to navigate over these smaller blocks and displace themselves at least several body lengths. Once the height has been reduced to 0.025 meters the blocks are so short that the environment becomes very similar to flat ground, and in fact distances achieved by organisms in the lower left environments are not significantly different from those of the control environment.
As the spacing between the blocks is reduced (moving upward in Fig. 3) the organisms are no longer able to behave as they would on flat ground, but instead must find ways to move along the tops of the blocks while finding a means of gaining purchase by reaching into the gaps. The height of the blocks loses importance in this part of the parameter space but still has an effect (though opposite to when the spacing is large). Here the general pattern is for taller blocks to make the task easier, because taller blocks provide more voluminous gaps which more easily support a variety of ways to gain purchase. Finally in the top row of Fig. 3, when the spacing is smallest, block height ceases to have much of an impact because however narrow an organism's appendages are they can only reach a short distance into the gaps.
For a better understanding of how the evolved organisms behave in each of these environments it is helpful to observe their behavior. For this purpose, sample videos of evolved organisms are available in the Supplementary Material (Videos S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15).
It is clear that different environments in this parameterization present the evolutionary system with varying degrees of difficulty, but the question now becomes: how does environment influence the evolution of morphological complexity? There are many approaches to quantify the complexity of an evolved morphology. Commonly, the variability of part types such as the number of cell types [41] has been used to measure the morphological complexity of biological organisms. But, the parts under consideration may vary in scale from organelles [42] to limbs [43], and it is unclear what should be considered a part in the current work. More geometric measures describing how space-filling a morphology is could also be employed (see Text S1 and Figure S2). Alternatively, a morphology's surface area to volume ratio could be measured, or its concavity could be computed (e.g. by taking the ratio of a morphology's volume to that of its convex hull). However each of these measures may be deceived by relatively simple body shapes, such as those that are very flat or contain large, simple concavities (e.g. a ‘C’ shape).
Instead, it is useful to think about the complexity of a body plan in information theoretic terms. One commonly used measure of complexity is Shannon's Entropy [44], which measures the uncertainty of a random variable. Recent work [45], [46] has demonstrated how Shannon Entropy can be applied to measure the complexity of a 3D object by considering the curvature of the object as a random variable. This means that in order to have higher complexity it is necessary to have more angles (regions of non-zero curvature) that can not simply be a repeating pattern, exactly what humans would think of as more complex shapes. And in fact, quantifying the complexity of 3D objects in this way has been shown to strongly correlate with human observers' notions of complexity [46].
In this work, the complexity of an organism's morphology is computed as the quantity which is the morphology's entropy of curvature or, in terminology which may be more familiar to biologists, it is the Shannon diversity [47] of the curvature on the organism's exterior (see Methods for details). Does capture the complexity of evolved morphologies? To answer this question, is calculated for all 9800 best-of-trial virtual organisms from all environments (icy and control). Out of those 9800, the five morphologies with the smallest value and the five morphologies with the largest value are selected. Images of these morphologies are shown in Fig. 2. Looking at these two sets of morphologies, those with high values appear more complex than those with low values. In light of this observation and the previous work in this area it is concluded that successfully captures morphological complexity.
Similarly, the concept of entropy may also be applied to characterizing the complexity of an environment. In the current formulation, environments are differentiated by variability in surface friction and terrain elevation. In the flat ground environment both the height of the terrain and the surface friction are uniform throughout, thus conveying zero entropy. On the other hand, in all of the icy environments there is variability in both of these properties. The surface friction is low on the ice blocks, but high on the ground between them. Likewise, the terrain is one height on the blocks and another in the intervening space. Therefore each of the icy environments has non-zero entropies of friction and elevation and so is considered to be more complex than flat ground. However, since each icy environment consists of a uniform series of ice blocks, the relative complexity between these environments is not considered.
Armed with these measures, it is now possible to characterize how different environments influence the morphological complexity of evolving organisms. In order to understand the evolutionary pressures which lead to virtual organisms that are more or less morphologically complex, it is interesting to consider how morphological complexity varies over evolutionary time in different environments, and how these changes correspond to variations in fitness. Towards that goal, Fig. 4 depicts the mean morphological complexity and mean displacement of the current best individual over evolutionary time for each of several icy environments along with a corresponding set of control trials. Here it can be seen that morphological complexity tends to increase over time along with fitness. This means that in these environments selection for locomotion corresponds to an increase in complexity.
However, it is unclear whether this increase of complexity is the result of a passive or a driven trend [3], [6], [34]. Passive trends may result from envelope expansion without any directional bias. For example, if there is a minimum level of complexity necessary for success, but no upper bound, then both the mean and the maximum complexity of the population will increase over time simply due to random variation (what Stephen Jay Gould famously referred to as a “drunkard's walk” [9]). On the other hand, driven trends exhibit a consistent, directional bias. This corresponds to active selection for greater complexity. In this case not only will there be an increase in mean and maximum complexity, but the minimum level of complexity will increase over evolutionary time as well.
When looking only at how morphological complexity varies over evolutionary time it is unclear what change in complexity is due to selection pressure from the environment and what change is due to biases towards increasing complexity within the evolutionary model itself and/or the general tendency of evolutionary systems to produce increasing complexity in the absence of selection [48]. In order to separate the influence of these factors it is useful to compare the evolving populations to a neutral shadow model [49], [50]. For a generational evolutionary model, such as that employed here, a neutral shadow of a given experiment is equivalent to re-running the evolutionary model with the same parameters but with random selection. Fig. 5 shows how the morphological complexity of organisms evolved in flat ground (black), as well as all icy environments (blue), changes over evolutionary time compared to those evolved in 100 independent trials using random selection (purple) in which the only preference is for genomes that produce valid morphologies (so that there exists a morphology for which complexity can be calculated; see Methods). It is known that the evolutionary system employed here [40] has an inherent bias to increase genotypic complexity over evolutionary time. The increasing purple curve in Fig. 5 indicates that there exists a bias to produce more complex morphologies over time as well. In fact, random selection alone produces morphologies that are more complex than those selected in any of the environments investigated. However, this comparison is not entirely fair. At any given generation, individuals in the random selection experiments will be the end product of many more reproduction (mutation and crossover) events than the corresponding individuals evolved for displacement, because under random selection it is unlikely that any individual will persist in the population for very long. Therefore, individuals in the random selection experiments will have had many more opportunities to increase the complexity of their genomes and hence the complexity of their morphologies.
In order to correct for this discrepancy in the number of reproduction events, alternative shadow models are employed. Specifically, neutral shadow models of both the flat ground experiments and a representative icy environment (spacing 0.025, height 0.8) are created, which control for the number of reproduction events leading to the individuals in the current population. In each of the 100 independent trials evolving for locomotion in both of these environments, a record of every reproduction event is kept, and alternative shadow models are created for each trial such that they maintain the same rate of reproduction. These shadow models are detailed in Text S1.
All model alternatives have similar complexity curves (see yellow, green, red and gray lines in Fig. 5) indicating that this shadow formulation is robust to whichever alternative is employed. Qualitatively they both show a much slower increase in morphological complexity (especially early on in evolution) compared to the experiments selecting for displacement, and so contrary to the naïve shadow model, both flat ground and icy environments select for increased morphological complexity beyond what would be expected in a neutral model. This implies that greater morphological complexity is being actively selected for in these environments: there is a driven trend towards increased morphological complexity.
While the results reported so far support the hypothesis that there is a driven trend for increased morphological complexity in all environments, they do not differentiate between the complexities of organisms evolved based on which environment they are evolved in. Specifically, Fig. 5 depicts similar levels of complexity evolving in icy environments as compared to the flat, high friction environment under this regime. In fact, when the morphological complexities of organisms evolved in each of the 49 icy environments are compared with independent sets of trials conducted in the control environment (see Figs. 4 and S1) they do not reflect a consistent relationship between environment and evolved morphological complexity. It is hypothesized that without a cost to becoming more complex the driven trend towards increased morphological complexity will dominate in all of the investigated environments.
On the other hand, it is hypothesized that when complexity does come at a cost–as is thought to occur in biological organisms [35], [36] –there will be greater pressure towards increased morphological complexity in more complex environments. In an an attempt to test this hypothesis, a second set of experiments is conducted which uses Pareto based multi-objective selection [51], [52] to evolve organisms that can locomote in their given task environment and are as simple as possible, therefore imposing a cost on complexity.
As was done for the single-objective experiments, 100 independent trials of a multi-objective model are run in each of the 49 icy environments along with a corresponding 49 independent sets of 100 trials apiece in the high friction, flat ground control environment. By selecting for both maximal displacement and minimal morphological complexity these experiments should evolve organisms that are no more complex than necessary to succeed in their task environment. If indeed more complex environments induce greater selection pressure favoring morphological complexity than simple environments when morphological complexity comes at a cost, then these differences should be observable under this regime.
Comparing the results of these multi-objective experiments, we indeed see that more complex environments tend to select for organisms with greater morphological complexity when compared with organisms evolved in the simple, control environment. Figs. 6–8 show how the morphological complexities of organisms evolved in each of the icy environments under multi-objective selection differs from that of organisms evolved in a corresponding set of trials from the control environment. Since selecting a single representative individual from each trial is not as straightforward as in the single-objective case (see Methods), several different techniques are employed to compare the results of these experiments.
First, for the final Pareto front of each trial in a given environment, the mean morphological complexity is taken. These means (100 from each environment) are compared to the mean morphological complexity in the final Pareto front of each trial from a corresponding set of trials from the control environment. This comparison is depicted in Fig. 6. Fig. 7 presents the same comparison except that it considers the organism with median performance on each Pareto front: the organism with equal number of individuals on the front that displace less and more than it (e.g. the most central point in Fig. 9). Lastly, Fig. 8 shows the same comparison except that it considers the mean complexity of those organisms in the middle half of their respective Pareto fronts. That is, the top quarter of the most complex morphologies (rightmost three points in Fig. 9) and the bottom quarter of most simple morphologies (leftmost three points in Fig. 9) in each front are ignored, and the means are taken across the remaining organisms in each front (which should reduce the influence of any outliers).
While some differences can be observed across these plots, the general pattern is largely consistent (and therefore not an artifact of the particular comparison employed): imposing a cost on complexity results in a multitude of icy environments where significantly more complex morphologies evolve compared to the control environment, and many of these differences are observed at the highest significance level (). This corroborates the hypothesis that the more complex environments induce selection pressure for increased morphological complexity beyond what would evolve in a simpler environment when morphological complexity comes at a cost.
In the lower right of Figs. 6–8, where the environments become too difficult to succeed in (because the organisms get trapped in the large gaps; see Fig. 3), multi-objective selection actually results in the evolution of morphologies that are significantly less complex than those that evolve to locomote on flat ground. The reason for this is that when it is not possible to evolve for greater displacement, the majority of selection bears down on the simplicity objective, and therefore simpler morphologies evolve in these environments under multi-objective selection.
This paper has presented a new method for evolving not only the neural systems but also the body plans of virtual organisms. This system differs from previous work by evolving populations of genetic encodings that produce complex morphologies instantiated in virtual environments as triangular meshes. This methodology opens up the possibility of investigating previously unexplored relationships between evolving organisms and their environments in a systematic manner.
Here, this system was used to investigate how different environments induce differing selection pressures on morphological complexity. By evolving virtual organisms in a number of different task environments and analyzing how an information theoretic measure of morphological complexity varies over evolutionary time, it was demonstrated that not only do all investigated environments actively induce selection pressure favoring greater complexity above and beyond what would be expected in the absence of selection, but that more complex environments in fact induce selection for more complex morphologies then simple environments when a cost is imposed on morphological complexity. Since it is often thought that complexity does incur a cost in biological organisms [35], [36], the differences observed between environments in this regime may be more representative of the selection pressures present in biological systems.
These results have illustrated how the environment may influence the complexity of evolving morphologies. Based on the results presented here it is possible that a similar evolutionary dynamic has been partially responsible for the “arrows of complexity” observed among biological organisms. As organisms have come to occupy more complex niches it is likely that these niches have actively selected for increased morphological complexity. Additionally, it should be possible to leverage this property for evolving more complex artifacts with evolutionary computation systems. However, it is not likely that increased environmental complexity will select for increased morphological complexity in every case where such complexity incurs a cost. While this work has demonstrated that such a relationship can exist, future work is needed to clarify this relationship across different environments, tasks, organisms, evolutionary models, and neural systems.
A number of simplifications were made here which it may be desirable to relax in future work. By constraining the number of morphological components and using very simple neural architectures it was possible to largely bracket the question of neural complexity and focus on one particular aspect of morphological complexity. However, it may be desirable to investigate how many different forms of complexity evolve as a function of environment. For instance, in a recent study [53] we demonstrated that another measure of complexity: “mechanical complexity”, decreased in the same environments that selected for greater morphological complexity. This result lends support to the notion that various forms of complexity may be inversely correlated as discussed in [54], and it also suggests that there is likely a trade-off between the various forms of complexity needed to succeed in a given environment, similar to the trade-off between morphological and neural complexity [11], [17].
To investigate these ideas further it will be interesting to allow for more complex neural architectures, more complex sensorimotor systems, and a greater diversity of materials (including ‘soft’ materials [55]) to study how environments may influence the evolution of sensorial, nuerological, motoric, material, mechanical, and morphological complexity of these various systems. By extending the information theoretic ideas used here for quantifying morphological complexity it is hoped that a ‘common currency of bits’ may be used to investigate these complexity trade-offs in a systematic manner.
The morphologies evolved in this work are encoded by Compositional Pattern Producing Networks (CPPNs) [39]. CPPNs are a form of artificial neural network (ANN) [56] which differ from traditional ANNs in several ways. While each internal node in a traditional ANN typically has the same activation function (such as a sigmoid or a step function), CPPN nodes can take on one of several activation functions from a predefined set. This function set often includes functions that are repetitive, such as or , as well as symmetric functions, such as , thus allowing for motifs seen in natural systems that arise as a result of development: symmetry, repetition, and repetition with variation. Additionally, CPPNs are often used as generative systems to produce other objects of interest, such as images [57], 3D structures [58], [59], robot morphologies [31], [32] or traditional ANNs themselves[60]–[64]. This is in contrast to the typical, direct application of ANNs as robot control architectures or classifiers.
CPPNs act as functions of geometry. Geometric coordinates meaningful to the object being represented are fed as inputs to the CPPN. These input values are passed through the various connections of the CPPN from node to node. Each node aggregates its inputs by taking a weighted sum of the values output by each upstream node (weights are specific to each connection) and outputs the result of applying a particular activation function (specific to that node) to this weighted sum. By passing the inputs through subsequent nodes the activation functions are composed to produce novel outputs while maintaining features of the different functions (hence the “compositional” aspect of CPPNs). Additionally, since these functions are chosen to have desirable properties present across a wide range of natural systems, as discussed above, CPPNs are capable of directly producing structures which in nature require a developmental process. For a more in-depth description of CPPNs, and further discussion of their ability to act as an abstraction of development, the reader is referred to [39].
In this study CPPNs are evolved via CPPN-NEAT [39]. CPPN-NEAT is an extension of the NeuroEvolution of Augmenting Topologies (NEAT) [40] method of neuro-evolution. NEAT is capable of evolving not only connection weights for existing network topologies, but also the network topologies themselves. Its operation is based on a few key ideas. First, the initial population is comprised of minimal networks (those without any internal or hidden nodes), which may then gradually increase in complexity over evolutionary time through structural mutations which add new nodes and links to the network. When a new node or link is created in this manner it is assigned a unique historical marking. These historical markings are inherited during reproduction and allow meaningful crossovers to occur without the use of expensive graph matching procedures. Additionally, these markings are used to divide the population into “species” of similar network topologies. Speciation promotes genotypic diversity and, because competition is primarily intraspecies, novel structural innovations are given time to mature before directly competing with individuals in other species.
CPPN-NEAT extends NEAT to evolve CPPNs. Effectively, this means that since nodes are no longer restricted to having sigmoid activation functions, each node contains an additional parameter which specifies its own activation function. When a new node is added to a network it is assigned a random function from a predefined set (the signed cosine, Gaussian and sigmoid functions are used in the experiments reported here). Additionally, the compatibility distance metric used for speciation is modified to incorporate the number of different activation functions between two networks. In all other respects, CPPN-NEAT behaves the same as NEAT.
NEAT and CPPN-NEAT have successfully evolved ANNs and CPPNs for a variety tasks [40], [57], [59], [61], [62], [65] which makes CPPN-NEAT a good option for evolving the CPPNs used in this study. Moreover, CPPN-NEAT's ability to systematically increase network complexity over evolutionary time as needed should lend itself well to studying how morphologies increase in complexity when evolving inside different environments. For a more thorough description of these algorithms, including additional details of the mechanisms discussed above, please refer to [39], [40].
While previously [31], [32] evolving virtual organisms were constructed out of spherical components, the current study employs a voxel-based method to create morphological components out of triangular meshes (trimeshes) similar to what is done for the creation of 3D shapes in [59]. This process is illustrated in Fig. 10, and is explained in detail below.
First, A regular grid is placed over a region of 3D space which defines the presence of voxel locations. In the current work this region extends from to 1 (inclusive) in each dimension and grid lines are placed at intervals of 0.2. This yields a total of 11 grid lines in each dimension for a total of voxels. A candidate CPPN is iteratively queried with the Cartesian coordinates at every voxel location except for the extrema in each direction. Querying a CPPN at a given location involves resetting all node values, and updating the CPPN for a fixed number of iterations (in this case 10) before the output value is retrieved. This procedure is employed in order to extract consistent output signals from networks with recurrent connections, which may fall into cyclic or chaotic attractors. Previously [32], it was found that allowing recurrent connections in morphology-generating CPPNs increased their evolvability. Voxel locations that exceed a predefined output threshold ( in this case) are considered to contain matter, while those that fall below this threshold are considered to be devoid of matter. All voxels lying on one of the extrema () are given output value so that no matter-containing voxel abuts against the boundary of the grid, and therefore guarantees that the final triangular meshes have completely enclosed surfaces. Once the CPPN has been queried for every voxel location, the Marching Cubes algorithm [66] is employed to create triangular meshes from the underlying voxel data. Specifically, an enclosed triangular mesh is created for each connected voxel component which defines the exterior surface of a single physical shape. These triangular meshes are then sent to the physics simulator where they define the exterior surface of a solid object and are imbued with mass (see Fig. 1 for some examples). This is the first instance of physically simulating evolved, rigid body organisms composed of triangular meshes.
Since the purpose of this study is to investigate how different task environments affect the shapes of evolved morphologies, a number of simplifications are used in order to concentrate on the physical shapes of the evolved organisms and control for other factors that may influence their performance. From the multiple enclosed trimesh components that could be produced when querying a single CPPN, only one of these (the largest in terms of number of triangles) is used in the resulting organism. This single component is copied and reflected across the . The resulting components (the original and its mirror image) are then spread apart by meters and a capsule of this length is placed between them such that it connects their two closest points. The two trimesh components each connect to this capsule by means of a hinge joint. These joints are constructed such that one rotates through the organism's coronal plane while the other rotates through its sagittal plane. Reflecting and copying a single component like this ensures that all organisms have the same mechanical degrees of freedom and ensures that the organisms are all bilaterally symmetric (which should facilitate locomotion) while at the same time it allows for a very large number of different morphologies due to the flexibility of the CPPN representation and trimesh model.
The two mechanical degrees of freedom of each organism are actuated by means of coupled oscillators. Each of the two oscillators is parameterized by several parameters: amplitude, period, and phase shift. These six parameters (three parameters apiece for each of the two joints) are directly encoded in the genome of the evolving organisms as floating point values so that the genome is in actuality a CPPN plus a six dimensional floating point array. These floating point values are recombined and mutated in the same manner as CPPN link weights with mutation magnitudes scaled by the range of values for that parameter. Additionally, crossover on these vectors is possible in all instances of sexual reproduction since every individual contains a vector of the same dimensionality. Values for these parameters are constrained to predefined ranges: amplitude, radians (so that the hinge rotates between and a radians), period simulation time steps (or equivalently of the total evaluation time) and phase shift periods. Each parameter has a mutation probability of 0.1, which was chosen experimentally.
Encoding the control parameters in this fashion is done to keep the controllers as simple as possible so that fitness is primarily dictated by the physical form of the organisms, while at the same time allowing for diverse enough behavior so that the organisms can succeed in the different task environments.
The focus of this study is on how environment influences the evolution of morphological complexity in virtual organisms. Towards this aim a simple task is chosen which can be accomplished with more or less difficulty in a variety of environments. Specifically, as in previous work (e.g. [24], [25], [32], [67], [68]), the task investigated here is to maximize directed displacement in a fixed amount of time, across a range of terrains.
A candidate morphology (triangular mesh) and accompanying set of control parameters are sent to a physics simulator and allowed to act for a fixed number of simulation time steps. (In this work simulations are conducted in the Open Dynamics Engine (http://www.ode.org), a widely used open source, physically realistic simulation environment.) Since trimeshes can be arbitrarily shaped and, unlike spheres, may simultaneously contact the environment at several points, it is necessary to use a much smaller step size than has been used in previous work in order to get physically realistic behavior. Specifically, a step size of 0.001 s is used in this work. Because of this smaller step size a proportionally larger number of time steps are needed to achieve the same effective simulation length. Here organisms are evaluated for time steps.
In this section, the building blocks of computing the entropy of curvature are presented. The reader is referred to [45], [46], [72], [73] for more in-depth discussions of their theoretical underpinnings.
Given a random variable with a probability density function (PDF) , entropy is defined as(3)where is discretized such that where the are specific values of .
Following [45], [46], the random variable on which is calculated is an approximation of the Gaussian curvature of the points on the surface. (The Gaussian curvature of a point is the product of the principal curvatures and of that point [72].) Since the bodies here are built out of triangular meshes the points at which this curvature is non-zero are precisely the vertices of the triangular mesh. Specifically, for each vertex in a trimesh the angular deficit is calculated as(4)where is the internal angle at of each triangle of which is a vertex. This angular deficit is directly proportional to the Gaussian curvature of that point [45], and so here we set for calculating the entropy of curvature. (The relationship between angular deficit and Gaussian curvature can be derived through application of the Gauss-Bonnet theorem [72]; see [73] for more details.)
Following the calculation of for every vertex, a PDF is estimated by placing the values of into discrete bins of uniform width () and counting the number of samples that fall into each bin. This results in a discrete set of probabilities , and Eqn. 3 can be used to arrive at an estimate of entropy that depends on the chosen , denoted here . (see Text S1 for further details.)
The source code used to run the experiments reported in this paper is publicly available at https://github.com/jauerb/CPPN_Trimesh
Additionally, the data files corresponding to the experiments reported in this paper have been made publicly available at http://dx.doi.org/10.6084/m9.figshare.858799
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.